1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Target/TargetOptions.h"
52 #define DEBUG_TYPE "arm-isel"
54 STATISTIC(NumTailCalls, "Number of tail calls");
55 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
56 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
59 EnableARMLongCalls("arm-long-calls", cl::Hidden,
60 cl::desc("Generate calls via indirect call instructions"),
64 ARMInterworking("arm-interworking", cl::Hidden,
65 cl::desc("Enable / disable ARM interworking (for debugging only)"),
69 class ARMCCState : public CCState {
71 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
72 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs,
73 LLVMContext &C, ParmContext PC)
74 : CCState(CC, isVarArg, MF, TM, locs, C) {
75 assert(((PC == Call) || (PC == Prologue)) &&
76 "ARMCCState users must specify whether their context is call"
77 "or prologue generation.");
83 // The APCS parameter registers.
84 static const MCPhysReg GPRArgRegs[] = {
85 ARM::R0, ARM::R1, ARM::R2, ARM::R3
88 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
89 MVT PromotedBitwiseVT) {
90 if (VT != PromotedLdStVT) {
91 setOperationAction(ISD::LOAD, VT, Promote);
92 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
94 setOperationAction(ISD::STORE, VT, Promote);
95 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
98 MVT ElemTy = VT.getVectorElementType();
99 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
100 setOperationAction(ISD::SETCC, VT, Custom);
101 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
102 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
103 if (ElemTy == MVT::i32) {
104 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
105 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
107 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
109 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
110 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
112 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
114 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
115 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
116 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
117 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
118 setOperationAction(ISD::SELECT, VT, Expand);
119 setOperationAction(ISD::SELECT_CC, VT, Expand);
120 setOperationAction(ISD::VSELECT, VT, Expand);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT, Custom);
124 setOperationAction(ISD::SRA, VT, Custom);
125 setOperationAction(ISD::SRL, VT, Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT, Promote);
131 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::OR, VT, Promote);
133 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::XOR, VT, Promote);
135 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT, Expand);
140 setOperationAction(ISD::UDIV, VT, Expand);
141 setOperationAction(ISD::FDIV, VT, Expand);
142 setOperationAction(ISD::SREM, VT, Expand);
143 setOperationAction(ISD::UREM, VT, Expand);
144 setOperationAction(ISD::FREM, VT, Expand);
147 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
148 addRegisterClass(VT, &ARM::DPRRegClass);
149 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
152 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPairRegClass);
154 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
157 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
158 if (TM.getSubtarget<ARMSubtarget>().isTargetMachO())
159 return new TargetLoweringObjectFileMachO();
161 return new ARMElfTargetObjectFile();
164 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
165 : TargetLowering(TM, createTLOF(TM)) {
166 Subtarget = &TM.getSubtarget<ARMSubtarget>();
167 RegInfo = TM.getRegisterInfo();
168 Itins = TM.getInstrItineraryData();
170 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
172 if (Subtarget->isTargetMachO()) {
173 // Uses VFP for Thumb libfuncs if available.
174 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
175 Subtarget->hasARMOps()) {
176 // Single-precision floating-point arithmetic.
177 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
178 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
179 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
180 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
182 // Double-precision floating-point arithmetic.
183 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
184 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
185 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
186 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
188 // Single-precision comparisons.
189 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
190 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
191 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
192 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
193 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
194 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
195 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
196 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
198 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
207 // Double-precision comparisons.
208 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
209 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
210 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
211 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
212 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
213 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
214 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
215 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
217 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
226 // Floating-point to integer conversions.
227 // i64 conversions are done via library routines even when generating VFP
228 // instructions, so use the same ones.
229 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
231 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
232 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
234 // Conversions between floating types.
235 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
236 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
238 // Integer to floating-point conversions.
239 // i64 conversions are done via library routines even when generating VFP
240 // instructions, so use the same ones.
241 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
242 // e.g., __floatunsidf vs. __floatunssidfvfp.
243 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
245 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
246 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
250 // These libcalls are not available in 32-bit.
251 setLibcallName(RTLIB::SHL_I128, 0);
252 setLibcallName(RTLIB::SRL_I128, 0);
253 setLibcallName(RTLIB::SRA_I128, 0);
255 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
256 !Subtarget->isTargetWindows()) {
257 // Double-precision floating-point arithmetic helper functions
258 // RTABI chapter 4.1.2, Table 2
259 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
260 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
261 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
262 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
263 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
264 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
265 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
266 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
268 // Double-precision floating-point comparison helper functions
269 // RTABI chapter 4.1.2, Table 3
270 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
271 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
272 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
273 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
274 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
275 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
276 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
277 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
278 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
279 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
280 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
281 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
282 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
283 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
284 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
285 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
286 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
287 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
295 // Single-precision floating-point arithmetic helper functions
296 // RTABI chapter 4.1.2, Table 4
297 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
298 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
299 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
300 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
301 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
302 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
303 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
304 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
306 // Single-precision floating-point comparison helper functions
307 // RTABI chapter 4.1.2, Table 5
308 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
309 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
310 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
311 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
312 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
313 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
314 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
315 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
316 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
317 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
318 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
319 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
320 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
321 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
322 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
323 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
324 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
325 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
326 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
333 // Floating-point to integer conversions.
334 // RTABI chapter 4.1.2, Table 6
335 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
336 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
337 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
338 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
339 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
340 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
341 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
342 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
343 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
344 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
345 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
346 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
352 // Conversions between floating types.
353 // RTABI chapter 4.1.2, Table 7
354 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
355 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
356 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
359 // Integer to floating-point conversions.
360 // RTABI chapter 4.1.2, Table 8
361 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
362 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
363 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
364 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
365 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
366 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
367 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
368 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
369 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
370 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
378 // Long long helper functions
379 // RTABI chapter 4.2, Table 9
380 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
381 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
382 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
383 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
384 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
391 // Integer division functions
392 // RTABI chapter 4.3.1
393 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
394 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
395 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
396 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
397 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
398 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
399 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
400 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
401 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
402 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
403 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
404 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
411 // RTABI chapter 4.3.4
412 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
413 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
414 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
415 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
416 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
417 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
420 // Use divmod compiler-rt calls for iOS 5.0 and later.
421 if (Subtarget->getTargetTriple().isiOS() &&
422 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
423 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
424 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
427 if (Subtarget->isThumb1Only())
428 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
430 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
431 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
432 !Subtarget->isThumb1Only()) {
433 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
434 if (!Subtarget->isFPOnlySP())
435 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
437 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
440 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
441 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
442 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
443 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
444 setTruncStoreAction((MVT::SimpleValueType)VT,
445 (MVT::SimpleValueType)InnerVT, Expand);
446 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
447 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
448 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
452 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
454 if (Subtarget->hasNEON()) {
455 addDRTypeForNEON(MVT::v2f32);
456 addDRTypeForNEON(MVT::v8i8);
457 addDRTypeForNEON(MVT::v4i16);
458 addDRTypeForNEON(MVT::v2i32);
459 addDRTypeForNEON(MVT::v1i64);
461 addQRTypeForNEON(MVT::v4f32);
462 addQRTypeForNEON(MVT::v2f64);
463 addQRTypeForNEON(MVT::v16i8);
464 addQRTypeForNEON(MVT::v8i16);
465 addQRTypeForNEON(MVT::v4i32);
466 addQRTypeForNEON(MVT::v2i64);
468 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
469 // neither Neon nor VFP support any arithmetic operations on it.
470 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
471 // supported for v4f32.
472 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
473 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
474 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
475 // FIXME: Code duplication: FDIV and FREM are expanded always, see
476 // ARMTargetLowering::addTypeForNEON method for details.
477 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
478 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
479 // FIXME: Create unittest.
480 // In another words, find a way when "copysign" appears in DAG with vector
482 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
483 // FIXME: Code duplication: SETCC has custom operation action, see
484 // ARMTargetLowering::addTypeForNEON method for details.
485 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
486 // FIXME: Create unittest for FNEG and for FABS.
487 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
488 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
489 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
490 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
491 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
492 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
493 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
494 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
495 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
497 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
498 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
499 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
500 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
501 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
502 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
503 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
504 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
505 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
507 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
508 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
510 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
512 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
515 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
517 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
518 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
519 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
520 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
521 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
523 // Mark v2f32 intrinsics.
524 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
525 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
526 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
527 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
528 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
529 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
530 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
531 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
532 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
533 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
534 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
535 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
536 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
537 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
538 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
540 // Neon does not support some operations on v1i64 and v2i64 types.
541 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
542 // Custom handling for some quad-vector types to detect VMULL.
543 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
544 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
545 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
546 // Custom handling for some vector types to avoid expensive expansions
547 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
548 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
549 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
550 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
551 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
552 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
553 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
554 // a destination type that is wider than the source, and nor does
555 // it have a FP_TO_[SU]INT instruction with a narrower destination than
557 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
558 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
559 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
560 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
563 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
565 // NEON does not have single instruction CTPOP for vectors with element
566 // types wider than 8-bits. However, custom lowering can leverage the
567 // v8i8/v16i8 vcnt instruction.
568 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
569 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
570 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
571 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
573 // NEON only has FMA instructions as of VFP4.
574 if (!Subtarget->hasVFP4()) {
575 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
576 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
579 setTargetDAGCombine(ISD::INTRINSIC_VOID);
580 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
581 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
582 setTargetDAGCombine(ISD::SHL);
583 setTargetDAGCombine(ISD::SRL);
584 setTargetDAGCombine(ISD::SRA);
585 setTargetDAGCombine(ISD::SIGN_EXTEND);
586 setTargetDAGCombine(ISD::ZERO_EXTEND);
587 setTargetDAGCombine(ISD::ANY_EXTEND);
588 setTargetDAGCombine(ISD::SELECT_CC);
589 setTargetDAGCombine(ISD::BUILD_VECTOR);
590 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
591 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
592 setTargetDAGCombine(ISD::STORE);
593 setTargetDAGCombine(ISD::FP_TO_SINT);
594 setTargetDAGCombine(ISD::FP_TO_UINT);
595 setTargetDAGCombine(ISD::FDIV);
597 // It is legal to extload from v4i8 to v4i16 or v4i32.
598 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
599 MVT::v4i16, MVT::v2i16,
601 for (unsigned i = 0; i < 6; ++i) {
602 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
603 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
604 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
608 // ARM and Thumb2 support UMLAL/SMLAL.
609 if (!Subtarget->isThumb1Only())
610 setTargetDAGCombine(ISD::ADDC);
613 computeRegisterProperties();
615 // ARM does not have f32 extending load.
616 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
618 // ARM does not have i1 sign extending load.
619 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
621 // ARM supports all 4 flavors of integer indexed load / store.
622 if (!Subtarget->isThumb1Only()) {
623 for (unsigned im = (unsigned)ISD::PRE_INC;
624 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
625 setIndexedLoadAction(im, MVT::i1, Legal);
626 setIndexedLoadAction(im, MVT::i8, Legal);
627 setIndexedLoadAction(im, MVT::i16, Legal);
628 setIndexedLoadAction(im, MVT::i32, Legal);
629 setIndexedStoreAction(im, MVT::i1, Legal);
630 setIndexedStoreAction(im, MVT::i8, Legal);
631 setIndexedStoreAction(im, MVT::i16, Legal);
632 setIndexedStoreAction(im, MVT::i32, Legal);
636 // i64 operation support.
637 setOperationAction(ISD::MUL, MVT::i64, Expand);
638 setOperationAction(ISD::MULHU, MVT::i32, Expand);
639 if (Subtarget->isThumb1Only()) {
640 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
641 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
643 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
644 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
645 setOperationAction(ISD::MULHS, MVT::i32, Expand);
647 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
648 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
649 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
650 setOperationAction(ISD::SRL, MVT::i64, Custom);
651 setOperationAction(ISD::SRA, MVT::i64, Custom);
653 if (!Subtarget->isThumb1Only()) {
654 // FIXME: We should do this for Thumb1 as well.
655 setOperationAction(ISD::ADDC, MVT::i32, Custom);
656 setOperationAction(ISD::ADDE, MVT::i32, Custom);
657 setOperationAction(ISD::SUBC, MVT::i32, Custom);
658 setOperationAction(ISD::SUBE, MVT::i32, Custom);
661 // ARM does not have ROTL.
662 setOperationAction(ISD::ROTL, MVT::i32, Expand);
663 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
664 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
665 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
666 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
668 // These just redirect to CTTZ and CTLZ on ARM.
669 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
670 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
672 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
674 // Only ARMv6 has BSWAP.
675 if (!Subtarget->hasV6Ops())
676 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
678 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
679 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
680 // These are expanded into libcalls if the cpu doesn't have HW divider.
681 setOperationAction(ISD::SDIV, MVT::i32, Expand);
682 setOperationAction(ISD::UDIV, MVT::i32, Expand);
685 // FIXME: Also set divmod for SREM on EABI
686 setOperationAction(ISD::SREM, MVT::i32, Expand);
687 setOperationAction(ISD::UREM, MVT::i32, Expand);
688 // Register based DivRem for AEABI (RTABI 4.2)
689 if (Subtarget->isTargetAEABI()) {
690 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
691 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
692 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
693 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
694 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
695 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
696 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
697 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
699 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
700 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
701 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
702 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
703 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
704 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
705 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
706 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
708 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
709 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
711 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
712 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
715 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
716 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
717 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
718 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
719 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
721 setOperationAction(ISD::TRAP, MVT::Other, Legal);
723 // Use the default implementation.
724 setOperationAction(ISD::VASTART, MVT::Other, Custom);
725 setOperationAction(ISD::VAARG, MVT::Other, Expand);
726 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
727 setOperationAction(ISD::VAEND, MVT::Other, Expand);
728 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
729 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
731 if (!Subtarget->isTargetMachO()) {
732 // Non-MachO platforms may return values in these registers via the
733 // personality function.
734 setExceptionPointerRegister(ARM::R0);
735 setExceptionSelectorRegister(ARM::R1);
738 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
739 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
740 // the default expansion.
741 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
742 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
743 // to ldrex/strex loops already.
744 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
746 // On v8, we have particularly efficient implementations of atomic fences
747 // if they can be combined with nearby atomic loads and stores.
748 if (!Subtarget->hasV8Ops()) {
749 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
750 setInsertFencesForAtomic(true);
753 // If there's anything we can use as a barrier, go through custom lowering
755 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
756 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
758 // Set them all for expansion, which will force libcalls.
759 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
760 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
761 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
762 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
763 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
764 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
765 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
766 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
767 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
768 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
769 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
770 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
771 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
772 // Unordered/Monotonic case.
773 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
774 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
777 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
779 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
780 if (!Subtarget->hasV6Ops()) {
781 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
782 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
784 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
786 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
787 !Subtarget->isThumb1Only()) {
788 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
789 // iff target supports vfp2.
790 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
791 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
794 // We want to custom lower some of our intrinsics.
795 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
796 if (Subtarget->isTargetDarwin()) {
797 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
798 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
799 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
802 setOperationAction(ISD::SETCC, MVT::i32, Expand);
803 setOperationAction(ISD::SETCC, MVT::f32, Expand);
804 setOperationAction(ISD::SETCC, MVT::f64, Expand);
805 setOperationAction(ISD::SELECT, MVT::i32, Custom);
806 setOperationAction(ISD::SELECT, MVT::f32, Custom);
807 setOperationAction(ISD::SELECT, MVT::f64, Custom);
808 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
809 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
810 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
812 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
813 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
814 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
815 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
816 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
818 // We don't support sin/cos/fmod/copysign/pow
819 setOperationAction(ISD::FSIN, MVT::f64, Expand);
820 setOperationAction(ISD::FSIN, MVT::f32, Expand);
821 setOperationAction(ISD::FCOS, MVT::f32, Expand);
822 setOperationAction(ISD::FCOS, MVT::f64, Expand);
823 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
824 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
825 setOperationAction(ISD::FREM, MVT::f64, Expand);
826 setOperationAction(ISD::FREM, MVT::f32, Expand);
827 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
828 !Subtarget->isThumb1Only()) {
829 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
830 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
832 setOperationAction(ISD::FPOW, MVT::f64, Expand);
833 setOperationAction(ISD::FPOW, MVT::f32, Expand);
835 if (!Subtarget->hasVFP4()) {
836 setOperationAction(ISD::FMA, MVT::f64, Expand);
837 setOperationAction(ISD::FMA, MVT::f32, Expand);
840 // Various VFP goodness
841 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
842 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
843 if (Subtarget->hasVFP2()) {
844 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
845 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
846 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
847 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
849 // Special handling for half-precision FP.
850 if (!Subtarget->hasFP16()) {
851 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
852 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
856 // Combine sin / cos into one node or libcall if possible.
857 if (Subtarget->hasSinCos()) {
858 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
859 setLibcallName(RTLIB::SINCOS_F64, "sincos");
860 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
861 // For iOS, we don't want to the normal expansion of a libcall to
862 // sincos. We want to issue a libcall to __sincos_stret.
863 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
864 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
868 // We have target-specific dag combine patterns for the following nodes:
869 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
870 setTargetDAGCombine(ISD::ADD);
871 setTargetDAGCombine(ISD::SUB);
872 setTargetDAGCombine(ISD::MUL);
873 setTargetDAGCombine(ISD::AND);
874 setTargetDAGCombine(ISD::OR);
875 setTargetDAGCombine(ISD::XOR);
877 if (Subtarget->hasV6Ops())
878 setTargetDAGCombine(ISD::SRL);
880 setStackPointerRegisterToSaveRestore(ARM::SP);
882 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
883 !Subtarget->hasVFP2())
884 setSchedulingPreference(Sched::RegPressure);
886 setSchedulingPreference(Sched::Hybrid);
888 //// temporary - rewrite interface to use type
889 MaxStoresPerMemset = 8;
890 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
891 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
892 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
893 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
894 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
896 // On ARM arguments smaller than 4 bytes are extended, so all arguments
897 // are at least 4 bytes aligned.
898 setMinStackArgumentAlignment(4);
900 // Prefer likely predicted branches to selects on out-of-order cores.
901 PredictableSelectIsExpensive = Subtarget->isLikeA9();
903 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
906 // FIXME: It might make sense to define the representative register class as the
907 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
908 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
909 // SPR's representative would be DPR_VFP2. This should work well if register
910 // pressure tracking were modified such that a register use would increment the
911 // pressure of the register class's representative and all of it's super
912 // classes' representatives transitively. We have not implemented this because
913 // of the difficulty prior to coalescing of modeling operand register classes
914 // due to the common occurrence of cross class copies and subregister insertions
916 std::pair<const TargetRegisterClass*, uint8_t>
917 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
918 const TargetRegisterClass *RRC = 0;
920 switch (VT.SimpleTy) {
922 return TargetLowering::findRepresentativeClass(VT);
923 // Use DPR as representative register class for all floating point
924 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
925 // the cost is 1 for both f32 and f64.
926 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
927 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
928 RRC = &ARM::DPRRegClass;
929 // When NEON is used for SP, only half of the register file is available
930 // because operations that define both SP and DP results will be constrained
931 // to the VFP2 class (D0-D15). We currently model this constraint prior to
932 // coalescing by double-counting the SP regs. See the FIXME above.
933 if (Subtarget->useNEONForSinglePrecisionFP())
936 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
937 case MVT::v4f32: case MVT::v2f64:
938 RRC = &ARM::DPRRegClass;
942 RRC = &ARM::DPRRegClass;
946 RRC = &ARM::DPRRegClass;
950 return std::make_pair(RRC, Cost);
953 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
956 case ARMISD::Wrapper: return "ARMISD::Wrapper";
957 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
958 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
959 case ARMISD::CALL: return "ARMISD::CALL";
960 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
961 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
962 case ARMISD::tCALL: return "ARMISD::tCALL";
963 case ARMISD::BRCOND: return "ARMISD::BRCOND";
964 case ARMISD::BR_JT: return "ARMISD::BR_JT";
965 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
966 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
967 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
968 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
969 case ARMISD::CMP: return "ARMISD::CMP";
970 case ARMISD::CMN: return "ARMISD::CMN";
971 case ARMISD::CMPZ: return "ARMISD::CMPZ";
972 case ARMISD::CMPFP: return "ARMISD::CMPFP";
973 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
974 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
975 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
977 case ARMISD::CMOV: return "ARMISD::CMOV";
979 case ARMISD::RBIT: return "ARMISD::RBIT";
981 case ARMISD::FTOSI: return "ARMISD::FTOSI";
982 case ARMISD::FTOUI: return "ARMISD::FTOUI";
983 case ARMISD::SITOF: return "ARMISD::SITOF";
984 case ARMISD::UITOF: return "ARMISD::UITOF";
986 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
987 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
988 case ARMISD::RRX: return "ARMISD::RRX";
990 case ARMISD::ADDC: return "ARMISD::ADDC";
991 case ARMISD::ADDE: return "ARMISD::ADDE";
992 case ARMISD::SUBC: return "ARMISD::SUBC";
993 case ARMISD::SUBE: return "ARMISD::SUBE";
995 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
996 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
998 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
999 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1001 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1003 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1005 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1007 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1009 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1011 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1012 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1013 case ARMISD::VCGE: return "ARMISD::VCGE";
1014 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1015 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1016 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1017 case ARMISD::VCGT: return "ARMISD::VCGT";
1018 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1019 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1020 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1021 case ARMISD::VTST: return "ARMISD::VTST";
1023 case ARMISD::VSHL: return "ARMISD::VSHL";
1024 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1025 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1026 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1027 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1028 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1029 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1030 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1031 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1032 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1033 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1034 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1035 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1036 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1037 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1038 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1039 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1040 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1041 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1042 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1043 case ARMISD::VDUP: return "ARMISD::VDUP";
1044 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1045 case ARMISD::VEXT: return "ARMISD::VEXT";
1046 case ARMISD::VREV64: return "ARMISD::VREV64";
1047 case ARMISD::VREV32: return "ARMISD::VREV32";
1048 case ARMISD::VREV16: return "ARMISD::VREV16";
1049 case ARMISD::VZIP: return "ARMISD::VZIP";
1050 case ARMISD::VUZP: return "ARMISD::VUZP";
1051 case ARMISD::VTRN: return "ARMISD::VTRN";
1052 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1053 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1054 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1055 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1056 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1057 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1058 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1059 case ARMISD::FMAX: return "ARMISD::FMAX";
1060 case ARMISD::FMIN: return "ARMISD::FMIN";
1061 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1062 case ARMISD::VMINNM: return "ARMISD::VMIN";
1063 case ARMISD::BFI: return "ARMISD::BFI";
1064 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1065 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1066 case ARMISD::VBSL: return "ARMISD::VBSL";
1067 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1068 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1069 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1070 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1071 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1072 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1073 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1074 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1075 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1076 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1077 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1078 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1079 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1080 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1081 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1082 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1083 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1084 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1085 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1086 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1090 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1091 if (!VT.isVector()) return getPointerTy();
1092 return VT.changeVectorElementTypeToInteger();
1095 /// getRegClassFor - Return the register class that should be used for the
1096 /// specified value type.
1097 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1098 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1099 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1100 // load / store 4 to 8 consecutive D registers.
1101 if (Subtarget->hasNEON()) {
1102 if (VT == MVT::v4i64)
1103 return &ARM::QQPRRegClass;
1104 if (VT == MVT::v8i64)
1105 return &ARM::QQQQPRRegClass;
1107 return TargetLowering::getRegClassFor(VT);
1110 // Create a fast isel object.
1112 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1113 const TargetLibraryInfo *libInfo) const {
1114 return ARM::createFastISel(funcInfo, libInfo);
1117 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1118 /// be used for loads / stores from the global.
1119 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1120 return (Subtarget->isThumb1Only() ? 127 : 4095);
1123 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1124 unsigned NumVals = N->getNumValues();
1126 return Sched::RegPressure;
1128 for (unsigned i = 0; i != NumVals; ++i) {
1129 EVT VT = N->getValueType(i);
1130 if (VT == MVT::Glue || VT == MVT::Other)
1132 if (VT.isFloatingPoint() || VT.isVector())
1136 if (!N->isMachineOpcode())
1137 return Sched::RegPressure;
1139 // Load are scheduled for latency even if there instruction itinerary
1140 // is not available.
1141 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1142 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1144 if (MCID.getNumDefs() == 0)
1145 return Sched::RegPressure;
1146 if (!Itins->isEmpty() &&
1147 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1150 return Sched::RegPressure;
1153 //===----------------------------------------------------------------------===//
1155 //===----------------------------------------------------------------------===//
1157 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1158 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1160 default: llvm_unreachable("Unknown condition code!");
1161 case ISD::SETNE: return ARMCC::NE;
1162 case ISD::SETEQ: return ARMCC::EQ;
1163 case ISD::SETGT: return ARMCC::GT;
1164 case ISD::SETGE: return ARMCC::GE;
1165 case ISD::SETLT: return ARMCC::LT;
1166 case ISD::SETLE: return ARMCC::LE;
1167 case ISD::SETUGT: return ARMCC::HI;
1168 case ISD::SETUGE: return ARMCC::HS;
1169 case ISD::SETULT: return ARMCC::LO;
1170 case ISD::SETULE: return ARMCC::LS;
1174 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1175 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1176 ARMCC::CondCodes &CondCode2) {
1177 CondCode2 = ARMCC::AL;
1179 default: llvm_unreachable("Unknown FP condition!");
1181 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1183 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1185 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1186 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1187 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1188 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1189 case ISD::SETO: CondCode = ARMCC::VC; break;
1190 case ISD::SETUO: CondCode = ARMCC::VS; break;
1191 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1192 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1193 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1195 case ISD::SETULT: CondCode = ARMCC::LT; break;
1197 case ISD::SETULE: CondCode = ARMCC::LE; break;
1199 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1203 //===----------------------------------------------------------------------===//
1204 // Calling Convention Implementation
1205 //===----------------------------------------------------------------------===//
1207 #include "ARMGenCallingConv.inc"
1209 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1210 /// given CallingConvention value.
1211 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1213 bool isVarArg) const {
1216 llvm_unreachable("Unsupported calling convention");
1217 case CallingConv::Fast:
1218 if (Subtarget->hasVFP2() && !isVarArg) {
1219 if (!Subtarget->isAAPCS_ABI())
1220 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1221 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1222 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1225 case CallingConv::C: {
1226 // Use target triple & subtarget features to do actual dispatch.
1227 if (!Subtarget->isAAPCS_ABI())
1228 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1229 else if (Subtarget->hasVFP2() &&
1230 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1232 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1233 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1235 case CallingConv::ARM_AAPCS_VFP:
1237 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1239 case CallingConv::ARM_AAPCS:
1240 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1241 case CallingConv::ARM_APCS:
1242 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1243 case CallingConv::GHC:
1244 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1248 /// LowerCallResult - Lower the result values of a call into the
1249 /// appropriate copies out of appropriate physical registers.
1251 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1252 CallingConv::ID CallConv, bool isVarArg,
1253 const SmallVectorImpl<ISD::InputArg> &Ins,
1254 SDLoc dl, SelectionDAG &DAG,
1255 SmallVectorImpl<SDValue> &InVals,
1256 bool isThisReturn, SDValue ThisVal) const {
1258 // Assign locations to each value returned by this call.
1259 SmallVector<CCValAssign, 16> RVLocs;
1260 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1261 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1262 CCInfo.AnalyzeCallResult(Ins,
1263 CCAssignFnForNode(CallConv, /* Return*/ true,
1266 // Copy all of the result registers out of their specified physreg.
1267 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1268 CCValAssign VA = RVLocs[i];
1270 // Pass 'this' value directly from the argument to return value, to avoid
1271 // reg unit interference
1272 if (i == 0 && isThisReturn) {
1273 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1274 "unexpected return calling convention register assignment");
1275 InVals.push_back(ThisVal);
1280 if (VA.needsCustom()) {
1281 // Handle f64 or half of a v2f64.
1282 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1284 Chain = Lo.getValue(1);
1285 InFlag = Lo.getValue(2);
1286 VA = RVLocs[++i]; // skip ahead to next loc
1287 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1289 Chain = Hi.getValue(1);
1290 InFlag = Hi.getValue(2);
1291 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1293 if (VA.getLocVT() == MVT::v2f64) {
1294 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1295 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1296 DAG.getConstant(0, MVT::i32));
1298 VA = RVLocs[++i]; // skip ahead to next loc
1299 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1300 Chain = Lo.getValue(1);
1301 InFlag = Lo.getValue(2);
1302 VA = RVLocs[++i]; // skip ahead to next loc
1303 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1304 Chain = Hi.getValue(1);
1305 InFlag = Hi.getValue(2);
1306 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1307 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1308 DAG.getConstant(1, MVT::i32));
1311 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1313 Chain = Val.getValue(1);
1314 InFlag = Val.getValue(2);
1317 switch (VA.getLocInfo()) {
1318 default: llvm_unreachable("Unknown loc info!");
1319 case CCValAssign::Full: break;
1320 case CCValAssign::BCvt:
1321 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1325 InVals.push_back(Val);
1331 /// LowerMemOpCallTo - Store the argument to the stack.
1333 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1334 SDValue StackPtr, SDValue Arg,
1335 SDLoc dl, SelectionDAG &DAG,
1336 const CCValAssign &VA,
1337 ISD::ArgFlagsTy Flags) const {
1338 unsigned LocMemOffset = VA.getLocMemOffset();
1339 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1340 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1341 return DAG.getStore(Chain, dl, Arg, PtrOff,
1342 MachinePointerInfo::getStack(LocMemOffset),
1346 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1347 SDValue Chain, SDValue &Arg,
1348 RegsToPassVector &RegsToPass,
1349 CCValAssign &VA, CCValAssign &NextVA,
1351 SmallVectorImpl<SDValue> &MemOpChains,
1352 ISD::ArgFlagsTy Flags) const {
1354 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1355 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1356 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1358 if (NextVA.isRegLoc())
1359 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1361 assert(NextVA.isMemLoc());
1362 if (StackPtr.getNode() == 0)
1363 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1365 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1371 /// LowerCall - Lowering a call into a callseq_start <-
1372 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1375 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1376 SmallVectorImpl<SDValue> &InVals) const {
1377 SelectionDAG &DAG = CLI.DAG;
1379 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1380 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1381 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1382 SDValue Chain = CLI.Chain;
1383 SDValue Callee = CLI.Callee;
1384 bool &isTailCall = CLI.IsTailCall;
1385 CallingConv::ID CallConv = CLI.CallConv;
1386 bool doesNotRet = CLI.DoesNotReturn;
1387 bool isVarArg = CLI.IsVarArg;
1389 MachineFunction &MF = DAG.getMachineFunction();
1390 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1391 bool isThisReturn = false;
1392 bool isSibCall = false;
1394 // Disable tail calls if they're not supported.
1395 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1399 // Check if it's really possible to do a tail call.
1400 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1401 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1402 Outs, OutVals, Ins, DAG);
1403 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1404 // detected sibcalls.
1411 // Analyze operands of the call, assigning locations to each operand.
1412 SmallVector<CCValAssign, 16> ArgLocs;
1413 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1414 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1415 CCInfo.AnalyzeCallOperands(Outs,
1416 CCAssignFnForNode(CallConv, /* Return*/ false,
1419 // Get a count of how many bytes are to be pushed on the stack.
1420 unsigned NumBytes = CCInfo.getNextStackOffset();
1422 // For tail calls, memory operands are available in our caller's stack.
1426 // Adjust the stack pointer for the new arguments...
1427 // These operations are automatically eliminated by the prolog/epilog pass
1429 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1432 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1434 RegsToPassVector RegsToPass;
1435 SmallVector<SDValue, 8> MemOpChains;
1437 // Walk the register/memloc assignments, inserting copies/loads. In the case
1438 // of tail call optimization, arguments are handled later.
1439 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1441 ++i, ++realArgIdx) {
1442 CCValAssign &VA = ArgLocs[i];
1443 SDValue Arg = OutVals[realArgIdx];
1444 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1445 bool isByVal = Flags.isByVal();
1447 // Promote the value if needed.
1448 switch (VA.getLocInfo()) {
1449 default: llvm_unreachable("Unknown loc info!");
1450 case CCValAssign::Full: break;
1451 case CCValAssign::SExt:
1452 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1454 case CCValAssign::ZExt:
1455 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1457 case CCValAssign::AExt:
1458 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1460 case CCValAssign::BCvt:
1461 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1465 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1466 if (VA.needsCustom()) {
1467 if (VA.getLocVT() == MVT::v2f64) {
1468 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1469 DAG.getConstant(0, MVT::i32));
1470 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1471 DAG.getConstant(1, MVT::i32));
1473 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1474 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1476 VA = ArgLocs[++i]; // skip ahead to next loc
1477 if (VA.isRegLoc()) {
1478 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1479 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1481 assert(VA.isMemLoc());
1483 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1484 dl, DAG, VA, Flags));
1487 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1488 StackPtr, MemOpChains, Flags);
1490 } else if (VA.isRegLoc()) {
1491 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1492 assert(VA.getLocVT() == MVT::i32 &&
1493 "unexpected calling convention register assignment");
1494 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1495 "unexpected use of 'returned'");
1496 isThisReturn = true;
1498 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1499 } else if (isByVal) {
1500 assert(VA.isMemLoc());
1501 unsigned offset = 0;
1503 // True if this byval aggregate will be split between registers
1505 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1506 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1508 if (CurByValIdx < ByValArgsCount) {
1510 unsigned RegBegin, RegEnd;
1511 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1513 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1515 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1516 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1517 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1518 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1519 MachinePointerInfo(),
1520 false, false, false,
1521 DAG.InferPtrAlignment(AddArg));
1522 MemOpChains.push_back(Load.getValue(1));
1523 RegsToPass.push_back(std::make_pair(j, Load));
1526 // If parameter size outsides register area, "offset" value
1527 // helps us to calculate stack slot for remained part properly.
1528 offset = RegEnd - RegBegin;
1530 CCInfo.nextInRegsParam();
1533 if (Flags.getByValSize() > 4*offset) {
1534 unsigned LocMemOffset = VA.getLocMemOffset();
1535 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1536 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1538 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1539 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1540 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1542 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1544 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1545 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1546 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1547 Ops, array_lengthof(Ops)));
1549 } else if (!isSibCall) {
1550 assert(VA.isMemLoc());
1552 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1553 dl, DAG, VA, Flags));
1557 if (!MemOpChains.empty())
1558 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1559 &MemOpChains[0], MemOpChains.size());
1561 // Build a sequence of copy-to-reg nodes chained together with token chain
1562 // and flag operands which copy the outgoing args into the appropriate regs.
1564 // Tail call byval lowering might overwrite argument registers so in case of
1565 // tail call optimization the copies to registers are lowered later.
1567 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1568 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1569 RegsToPass[i].second, InFlag);
1570 InFlag = Chain.getValue(1);
1573 // For tail calls lower the arguments to the 'real' stack slot.
1575 // Force all the incoming stack arguments to be loaded from the stack
1576 // before any new outgoing arguments are stored to the stack, because the
1577 // outgoing stack slots may alias the incoming argument stack slots, and
1578 // the alias isn't otherwise explicit. This is slightly more conservative
1579 // than necessary, because it means that each store effectively depends
1580 // on every argument instead of just those arguments it would clobber.
1582 // Do not flag preceding copytoreg stuff together with the following stuff.
1584 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1585 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1586 RegsToPass[i].second, InFlag);
1587 InFlag = Chain.getValue(1);
1592 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1593 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1594 // node so that legalize doesn't hack it.
1595 bool isDirect = false;
1596 bool isARMFunc = false;
1597 bool isLocalARMFunc = false;
1598 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1600 if (EnableARMLongCalls) {
1601 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1602 && "long-calls with non-static relocation model!");
1603 // Handle a global address or an external symbol. If it's not one of
1604 // those, the target's already in a register, so we don't need to do
1606 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1607 const GlobalValue *GV = G->getGlobal();
1608 // Create a constant pool entry for the callee address
1609 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1610 ARMConstantPoolValue *CPV =
1611 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1613 // Get the address of the callee into a register
1614 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1615 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1616 Callee = DAG.getLoad(getPointerTy(), dl,
1617 DAG.getEntryNode(), CPAddr,
1618 MachinePointerInfo::getConstantPool(),
1619 false, false, false, 0);
1620 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1621 const char *Sym = S->getSymbol();
1623 // Create a constant pool entry for the callee address
1624 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1625 ARMConstantPoolValue *CPV =
1626 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1627 ARMPCLabelIndex, 0);
1628 // Get the address of the callee into a register
1629 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1630 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1631 Callee = DAG.getLoad(getPointerTy(), dl,
1632 DAG.getEntryNode(), CPAddr,
1633 MachinePointerInfo::getConstantPool(),
1634 false, false, false, 0);
1636 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1637 const GlobalValue *GV = G->getGlobal();
1639 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1640 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1641 getTargetMachine().getRelocationModel() != Reloc::Static;
1642 isARMFunc = !Subtarget->isThumb() || isStub;
1643 // ARM call to a local ARM function is predicable.
1644 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1645 // tBX takes a register source operand.
1646 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1647 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1648 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1649 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1651 // On ELF targets for PIC code, direct calls should go through the PLT
1652 unsigned OpFlags = 0;
1653 if (Subtarget->isTargetELF() &&
1654 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1655 OpFlags = ARMII::MO_PLT;
1656 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1658 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1660 bool isStub = Subtarget->isTargetMachO() &&
1661 getTargetMachine().getRelocationModel() != Reloc::Static;
1662 isARMFunc = !Subtarget->isThumb() || isStub;
1663 // tBX takes a register source operand.
1664 const char *Sym = S->getSymbol();
1665 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1666 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1667 ARMConstantPoolValue *CPV =
1668 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1669 ARMPCLabelIndex, 4);
1670 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1671 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1672 Callee = DAG.getLoad(getPointerTy(), dl,
1673 DAG.getEntryNode(), CPAddr,
1674 MachinePointerInfo::getConstantPool(),
1675 false, false, false, 0);
1676 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1677 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1678 getPointerTy(), Callee, PICLabel);
1680 unsigned OpFlags = 0;
1681 // On ELF targets for PIC code, direct calls should go through the PLT
1682 if (Subtarget->isTargetELF() &&
1683 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1684 OpFlags = ARMII::MO_PLT;
1685 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1689 // FIXME: handle tail calls differently.
1691 bool HasMinSizeAttr = Subtarget->isMinSize();
1692 if (Subtarget->isThumb()) {
1693 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1694 CallOpc = ARMISD::CALL_NOLINK;
1696 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1698 if (!isDirect && !Subtarget->hasV5TOps())
1699 CallOpc = ARMISD::CALL_NOLINK;
1700 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1701 // Emit regular call when code size is the priority
1703 // "mov lr, pc; b _foo" to avoid confusing the RSP
1704 CallOpc = ARMISD::CALL_NOLINK;
1706 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1709 std::vector<SDValue> Ops;
1710 Ops.push_back(Chain);
1711 Ops.push_back(Callee);
1713 // Add argument registers to the end of the list so that they are known live
1715 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1716 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1717 RegsToPass[i].second.getValueType()));
1719 // Add a register mask operand representing the call-preserved registers.
1721 const uint32_t *Mask;
1722 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1723 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1725 // For 'this' returns, use the R0-preserving mask if applicable
1726 Mask = ARI->getThisReturnPreservedMask(CallConv);
1728 // Set isThisReturn to false if the calling convention is not one that
1729 // allows 'returned' to be modeled in this way, so LowerCallResult does
1730 // not try to pass 'this' straight through
1731 isThisReturn = false;
1732 Mask = ARI->getCallPreservedMask(CallConv);
1735 Mask = ARI->getCallPreservedMask(CallConv);
1737 assert(Mask && "Missing call preserved mask for calling convention");
1738 Ops.push_back(DAG.getRegisterMask(Mask));
1741 if (InFlag.getNode())
1742 Ops.push_back(InFlag);
1744 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1746 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1748 // Returns a chain and a flag for retval copy to use.
1749 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1750 InFlag = Chain.getValue(1);
1752 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1753 DAG.getIntPtrConstant(0, true), InFlag, dl);
1755 InFlag = Chain.getValue(1);
1757 // Handle result values, copying them out of physregs into vregs that we
1759 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1760 InVals, isThisReturn,
1761 isThisReturn ? OutVals[0] : SDValue());
1764 /// HandleByVal - Every parameter *after* a byval parameter is passed
1765 /// on the stack. Remember the next parameter register to allocate,
1766 /// and then confiscate the rest of the parameter registers to insure
1769 ARMTargetLowering::HandleByVal(
1770 CCState *State, unsigned &size, unsigned Align) const {
1771 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1772 assert((State->getCallOrPrologue() == Prologue ||
1773 State->getCallOrPrologue() == Call) &&
1774 "unhandled ParmContext");
1776 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1777 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1778 unsigned AlignInRegs = Align / 4;
1779 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1780 for (unsigned i = 0; i < Waste; ++i)
1781 reg = State->AllocateReg(GPRArgRegs, 4);
1784 unsigned excess = 4 * (ARM::R4 - reg);
1786 // Special case when NSAA != SP and parameter size greater than size of
1787 // all remained GPR regs. In that case we can't split parameter, we must
1788 // send it to stack. We also must set NCRN to R4, so waste all
1789 // remained registers.
1790 const unsigned NSAAOffset = State->getNextStackOffset();
1791 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1792 while (State->AllocateReg(GPRArgRegs, 4))
1797 // First register for byval parameter is the first register that wasn't
1798 // allocated before this method call, so it would be "reg".
1799 // If parameter is small enough to be saved in range [reg, r4), then
1800 // the end (first after last) register would be reg + param-size-in-regs,
1801 // else parameter would be splitted between registers and stack,
1802 // end register would be r4 in this case.
1803 unsigned ByValRegBegin = reg;
1804 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1805 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1806 // Note, first register is allocated in the beginning of function already,
1807 // allocate remained amount of registers we need.
1808 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1809 State->AllocateReg(GPRArgRegs, 4);
1810 // A byval parameter that is split between registers and memory needs its
1811 // size truncated here.
1812 // In the case where the entire structure fits in registers, we set the
1813 // size in memory to zero.
1822 /// MatchingStackOffset - Return true if the given stack call argument is
1823 /// already available in the same position (relatively) of the caller's
1824 /// incoming argument stack.
1826 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1827 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1828 const TargetInstrInfo *TII) {
1829 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1831 if (Arg.getOpcode() == ISD::CopyFromReg) {
1832 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1833 if (!TargetRegisterInfo::isVirtualRegister(VR))
1835 MachineInstr *Def = MRI->getVRegDef(VR);
1838 if (!Flags.isByVal()) {
1839 if (!TII->isLoadFromStackSlot(Def, FI))
1844 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1845 if (Flags.isByVal())
1846 // ByVal argument is passed in as a pointer but it's now being
1847 // dereferenced. e.g.
1848 // define @foo(%struct.X* %A) {
1849 // tail call @bar(%struct.X* byval %A)
1852 SDValue Ptr = Ld->getBasePtr();
1853 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1856 FI = FINode->getIndex();
1860 assert(FI != INT_MAX);
1861 if (!MFI->isFixedObjectIndex(FI))
1863 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1866 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1867 /// for tail call optimization. Targets which want to do tail call
1868 /// optimization should implement this function.
1870 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1871 CallingConv::ID CalleeCC,
1873 bool isCalleeStructRet,
1874 bool isCallerStructRet,
1875 const SmallVectorImpl<ISD::OutputArg> &Outs,
1876 const SmallVectorImpl<SDValue> &OutVals,
1877 const SmallVectorImpl<ISD::InputArg> &Ins,
1878 SelectionDAG& DAG) const {
1879 const Function *CallerF = DAG.getMachineFunction().getFunction();
1880 CallingConv::ID CallerCC = CallerF->getCallingConv();
1881 bool CCMatch = CallerCC == CalleeCC;
1883 // Look for obvious safe cases to perform tail call optimization that do not
1884 // require ABI changes. This is what gcc calls sibcall.
1886 // Do not sibcall optimize vararg calls unless the call site is not passing
1888 if (isVarArg && !Outs.empty())
1891 // Exception-handling functions need a special set of instructions to indicate
1892 // a return to the hardware. Tail-calling another function would probably
1894 if (CallerF->hasFnAttribute("interrupt"))
1897 // Also avoid sibcall optimization if either caller or callee uses struct
1898 // return semantics.
1899 if (isCalleeStructRet || isCallerStructRet)
1902 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1903 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1904 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1905 // support in the assembler and linker to be used. This would need to be
1906 // fixed to fully support tail calls in Thumb1.
1908 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1909 // LR. This means if we need to reload LR, it takes an extra instructions,
1910 // which outweighs the value of the tail call; but here we don't know yet
1911 // whether LR is going to be used. Probably the right approach is to
1912 // generate the tail call here and turn it back into CALL/RET in
1913 // emitEpilogue if LR is used.
1915 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1916 // but we need to make sure there are enough registers; the only valid
1917 // registers are the 4 used for parameters. We don't currently do this
1919 if (Subtarget->isThumb1Only())
1922 // If the calling conventions do not match, then we'd better make sure the
1923 // results are returned in the same way as what the caller expects.
1925 SmallVector<CCValAssign, 16> RVLocs1;
1926 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1927 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1928 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1930 SmallVector<CCValAssign, 16> RVLocs2;
1931 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1932 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1933 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1935 if (RVLocs1.size() != RVLocs2.size())
1937 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1938 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1940 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1942 if (RVLocs1[i].isRegLoc()) {
1943 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1946 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1952 // If Caller's vararg or byval argument has been split between registers and
1953 // stack, do not perform tail call, since part of the argument is in caller's
1955 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1956 getInfo<ARMFunctionInfo>();
1957 if (AFI_Caller->getArgRegsSaveSize())
1960 // If the callee takes no arguments then go on to check the results of the
1962 if (!Outs.empty()) {
1963 // Check if stack adjustment is needed. For now, do not do this if any
1964 // argument is passed on the stack.
1965 SmallVector<CCValAssign, 16> ArgLocs;
1966 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1967 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1968 CCInfo.AnalyzeCallOperands(Outs,
1969 CCAssignFnForNode(CalleeCC, false, isVarArg));
1970 if (CCInfo.getNextStackOffset()) {
1971 MachineFunction &MF = DAG.getMachineFunction();
1973 // Check if the arguments are already laid out in the right way as
1974 // the caller's fixed stack objects.
1975 MachineFrameInfo *MFI = MF.getFrameInfo();
1976 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1977 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1978 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1980 ++i, ++realArgIdx) {
1981 CCValAssign &VA = ArgLocs[i];
1982 EVT RegVT = VA.getLocVT();
1983 SDValue Arg = OutVals[realArgIdx];
1984 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1985 if (VA.getLocInfo() == CCValAssign::Indirect)
1987 if (VA.needsCustom()) {
1988 // f64 and vector types are split into multiple registers or
1989 // register/stack-slot combinations. The types will not match
1990 // the registers; give up on memory f64 refs until we figure
1991 // out what to do about this.
1994 if (!ArgLocs[++i].isRegLoc())
1996 if (RegVT == MVT::v2f64) {
1997 if (!ArgLocs[++i].isRegLoc())
1999 if (!ArgLocs[++i].isRegLoc())
2002 } else if (!VA.isRegLoc()) {
2003 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2015 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2016 MachineFunction &MF, bool isVarArg,
2017 const SmallVectorImpl<ISD::OutputArg> &Outs,
2018 LLVMContext &Context) const {
2019 SmallVector<CCValAssign, 16> RVLocs;
2020 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2021 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2025 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2026 SDLoc DL, SelectionDAG &DAG) {
2027 const MachineFunction &MF = DAG.getMachineFunction();
2028 const Function *F = MF.getFunction();
2030 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2032 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2033 // version of the "preferred return address". These offsets affect the return
2034 // instruction if this is a return from PL1 without hypervisor extensions.
2035 // IRQ/FIQ: +4 "subs pc, lr, #4"
2036 // SWI: 0 "subs pc, lr, #0"
2037 // ABORT: +4 "subs pc, lr, #4"
2038 // UNDEF: +4/+2 "subs pc, lr, #0"
2039 // UNDEF varies depending on where the exception came from ARM or Thumb
2040 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2043 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2046 else if (IntKind == "SWI" || IntKind == "UNDEF")
2049 report_fatal_error("Unsupported interrupt attribute. If present, value "
2050 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2052 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2054 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other,
2055 RetOps.data(), RetOps.size());
2059 ARMTargetLowering::LowerReturn(SDValue Chain,
2060 CallingConv::ID CallConv, bool isVarArg,
2061 const SmallVectorImpl<ISD::OutputArg> &Outs,
2062 const SmallVectorImpl<SDValue> &OutVals,
2063 SDLoc dl, SelectionDAG &DAG) const {
2065 // CCValAssign - represent the assignment of the return value to a location.
2066 SmallVector<CCValAssign, 16> RVLocs;
2068 // CCState - Info about the registers and stack slots.
2069 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2070 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2072 // Analyze outgoing return values.
2073 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2077 SmallVector<SDValue, 4> RetOps;
2078 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2080 // Copy the result values into the output registers.
2081 for (unsigned i = 0, realRVLocIdx = 0;
2083 ++i, ++realRVLocIdx) {
2084 CCValAssign &VA = RVLocs[i];
2085 assert(VA.isRegLoc() && "Can only return in registers!");
2087 SDValue Arg = OutVals[realRVLocIdx];
2089 switch (VA.getLocInfo()) {
2090 default: llvm_unreachable("Unknown loc info!");
2091 case CCValAssign::Full: break;
2092 case CCValAssign::BCvt:
2093 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2097 if (VA.needsCustom()) {
2098 if (VA.getLocVT() == MVT::v2f64) {
2099 // Extract the first half and return it in two registers.
2100 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2101 DAG.getConstant(0, MVT::i32));
2102 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2103 DAG.getVTList(MVT::i32, MVT::i32), Half);
2105 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2106 Flag = Chain.getValue(1);
2107 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2108 VA = RVLocs[++i]; // skip ahead to next loc
2109 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2110 HalfGPRs.getValue(1), Flag);
2111 Flag = Chain.getValue(1);
2112 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2113 VA = RVLocs[++i]; // skip ahead to next loc
2115 // Extract the 2nd half and fall through to handle it as an f64 value.
2116 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2117 DAG.getConstant(1, MVT::i32));
2119 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2121 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2122 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2123 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2124 Flag = Chain.getValue(1);
2125 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2126 VA = RVLocs[++i]; // skip ahead to next loc
2127 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2130 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2132 // Guarantee that all emitted copies are
2133 // stuck together, avoiding something bad.
2134 Flag = Chain.getValue(1);
2135 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2138 // Update chain and glue.
2141 RetOps.push_back(Flag);
2143 // CPUs which aren't M-class use a special sequence to return from
2144 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2145 // though we use "subs pc, lr, #N").
2147 // M-class CPUs actually use a normal return sequence with a special
2148 // (hardware-provided) value in LR, so the normal code path works.
2149 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2150 !Subtarget->isMClass()) {
2151 if (Subtarget->isThumb1Only())
2152 report_fatal_error("interrupt attribute is not supported in Thumb1");
2153 return LowerInterruptReturn(RetOps, dl, DAG);
2156 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2157 RetOps.data(), RetOps.size());
2160 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2161 if (N->getNumValues() != 1)
2163 if (!N->hasNUsesOfValue(1, 0))
2166 SDValue TCChain = Chain;
2167 SDNode *Copy = *N->use_begin();
2168 if (Copy->getOpcode() == ISD::CopyToReg) {
2169 // If the copy has a glue operand, we conservatively assume it isn't safe to
2170 // perform a tail call.
2171 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2173 TCChain = Copy->getOperand(0);
2174 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2175 SDNode *VMov = Copy;
2176 // f64 returned in a pair of GPRs.
2177 SmallPtrSet<SDNode*, 2> Copies;
2178 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2180 if (UI->getOpcode() != ISD::CopyToReg)
2184 if (Copies.size() > 2)
2187 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2189 SDValue UseChain = UI->getOperand(0);
2190 if (Copies.count(UseChain.getNode()))
2197 } else if (Copy->getOpcode() == ISD::BITCAST) {
2198 // f32 returned in a single GPR.
2199 if (!Copy->hasOneUse())
2201 Copy = *Copy->use_begin();
2202 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2204 TCChain = Copy->getOperand(0);
2209 bool HasRet = false;
2210 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2212 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2213 UI->getOpcode() != ARMISD::INTRET_FLAG)
2225 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2226 if (!Subtarget->supportsTailCall())
2229 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2232 return !Subtarget->isThumb1Only();
2235 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2236 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2237 // one of the above mentioned nodes. It has to be wrapped because otherwise
2238 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2239 // be used to form addressing mode. These wrapped nodes will be selected
2241 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2242 EVT PtrVT = Op.getValueType();
2243 // FIXME there is no actual debug info here
2245 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2247 if (CP->isMachineConstantPoolEntry())
2248 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2249 CP->getAlignment());
2251 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2252 CP->getAlignment());
2253 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2256 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2257 return MachineJumpTableInfo::EK_Inline;
2260 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2261 SelectionDAG &DAG) const {
2262 MachineFunction &MF = DAG.getMachineFunction();
2263 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2264 unsigned ARMPCLabelIndex = 0;
2266 EVT PtrVT = getPointerTy();
2267 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2268 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2270 if (RelocM == Reloc::Static) {
2271 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2273 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2274 ARMPCLabelIndex = AFI->createPICLabelUId();
2275 ARMConstantPoolValue *CPV =
2276 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2277 ARMCP::CPBlockAddress, PCAdj);
2278 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2280 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2281 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2282 MachinePointerInfo::getConstantPool(),
2283 false, false, false, 0);
2284 if (RelocM == Reloc::Static)
2286 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2287 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2290 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2292 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2293 SelectionDAG &DAG) const {
2295 EVT PtrVT = getPointerTy();
2296 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2297 MachineFunction &MF = DAG.getMachineFunction();
2298 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2299 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2300 ARMConstantPoolValue *CPV =
2301 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2302 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2303 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2304 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2305 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2306 MachinePointerInfo::getConstantPool(),
2307 false, false, false, 0);
2308 SDValue Chain = Argument.getValue(1);
2310 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2311 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2313 // call __tls_get_addr.
2316 Entry.Node = Argument;
2317 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2318 Args.push_back(Entry);
2319 // FIXME: is there useful debug info available here?
2320 TargetLowering::CallLoweringInfo CLI(Chain,
2321 (Type *) Type::getInt32Ty(*DAG.getContext()),
2322 false, false, false, false,
2323 0, CallingConv::C, /*isTailCall=*/false,
2324 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2325 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2326 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2327 return CallResult.first;
2330 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2331 // "local exec" model.
2333 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2335 TLSModel::Model model) const {
2336 const GlobalValue *GV = GA->getGlobal();
2339 SDValue Chain = DAG.getEntryNode();
2340 EVT PtrVT = getPointerTy();
2341 // Get the Thread Pointer
2342 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2344 if (model == TLSModel::InitialExec) {
2345 MachineFunction &MF = DAG.getMachineFunction();
2346 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2347 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2348 // Initial exec model.
2349 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2350 ARMConstantPoolValue *CPV =
2351 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2352 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2354 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2355 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2356 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2357 MachinePointerInfo::getConstantPool(),
2358 false, false, false, 0);
2359 Chain = Offset.getValue(1);
2361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2362 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2364 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2365 MachinePointerInfo::getConstantPool(),
2366 false, false, false, 0);
2369 assert(model == TLSModel::LocalExec);
2370 ARMConstantPoolValue *CPV =
2371 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2372 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2373 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2374 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2375 MachinePointerInfo::getConstantPool(),
2376 false, false, false, 0);
2379 // The address of the thread local variable is the add of the thread
2380 // pointer with the offset of the variable.
2381 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2385 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2386 // TODO: implement the "local dynamic" model
2387 assert(Subtarget->isTargetELF() &&
2388 "TLS not implemented for non-ELF targets");
2389 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2391 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2394 case TLSModel::GeneralDynamic:
2395 case TLSModel::LocalDynamic:
2396 return LowerToTLSGeneralDynamicModel(GA, DAG);
2397 case TLSModel::InitialExec:
2398 case TLSModel::LocalExec:
2399 return LowerToTLSExecModels(GA, DAG, model);
2401 llvm_unreachable("bogus TLS model");
2404 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2405 SelectionDAG &DAG) const {
2406 EVT PtrVT = getPointerTy();
2408 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2409 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2410 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2411 ARMConstantPoolValue *CPV =
2412 ARMConstantPoolConstant::Create(GV,
2413 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2414 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2415 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2416 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2418 MachinePointerInfo::getConstantPool(),
2419 false, false, false, 0);
2420 SDValue Chain = Result.getValue(1);
2421 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2422 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2424 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2425 MachinePointerInfo::getGOT(),
2426 false, false, false, 0);
2430 // If we have T2 ops, we can materialize the address directly via movt/movw
2431 // pair. This is always cheaper.
2432 if (Subtarget->useMovt()) {
2434 // FIXME: Once remat is capable of dealing with instructions with register
2435 // operands, expand this into two nodes.
2436 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2437 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2439 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2440 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2441 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2442 MachinePointerInfo::getConstantPool(),
2443 false, false, false, 0);
2447 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2448 SelectionDAG &DAG) const {
2449 EVT PtrVT = getPointerTy();
2451 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2452 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2454 if (Subtarget->useMovt())
2457 // FIXME: Once remat is capable of dealing with instructions with register
2458 // operands, expand this into multiple nodes
2460 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2462 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2463 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2465 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2466 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2467 MachinePointerInfo::getGOT(), false, false, false, 0);
2471 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2472 SelectionDAG &DAG) const {
2473 assert(Subtarget->isTargetELF() &&
2474 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2475 MachineFunction &MF = DAG.getMachineFunction();
2476 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2477 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2478 EVT PtrVT = getPointerTy();
2480 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2481 ARMConstantPoolValue *CPV =
2482 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2483 ARMPCLabelIndex, PCAdj);
2484 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2485 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2486 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2487 MachinePointerInfo::getConstantPool(),
2488 false, false, false, 0);
2489 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2490 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2494 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2496 SDValue Val = DAG.getConstant(0, MVT::i32);
2497 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2498 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2499 Op.getOperand(1), Val);
2503 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2505 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2506 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2510 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2511 const ARMSubtarget *Subtarget) const {
2512 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2515 default: return SDValue(); // Don't custom lower most intrinsics.
2516 case Intrinsic::arm_thread_pointer: {
2517 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2518 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2520 case Intrinsic::eh_sjlj_lsda: {
2521 MachineFunction &MF = DAG.getMachineFunction();
2522 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2523 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2524 EVT PtrVT = getPointerTy();
2525 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2527 unsigned PCAdj = (RelocM != Reloc::PIC_)
2528 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2529 ARMConstantPoolValue *CPV =
2530 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2531 ARMCP::CPLSDA, PCAdj);
2532 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2533 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2535 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2536 MachinePointerInfo::getConstantPool(),
2537 false, false, false, 0);
2539 if (RelocM == Reloc::PIC_) {
2540 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2541 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2545 case Intrinsic::arm_neon_vmulls:
2546 case Intrinsic::arm_neon_vmullu: {
2547 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2548 ? ARMISD::VMULLs : ARMISD::VMULLu;
2549 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2550 Op.getOperand(1), Op.getOperand(2));
2555 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2556 const ARMSubtarget *Subtarget) {
2557 // FIXME: handle "fence singlethread" more efficiently.
2559 if (!Subtarget->hasDataBarrier()) {
2560 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2561 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2563 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2564 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2565 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2566 DAG.getConstant(0, MVT::i32));
2569 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2570 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2571 unsigned Domain = ARM_MB::ISH;
2572 if (Subtarget->isMClass()) {
2573 // Only a full system barrier exists in the M-class architectures.
2574 Domain = ARM_MB::SY;
2575 } else if (Subtarget->isSwift() && Ord == Release) {
2576 // Swift happens to implement ISHST barriers in a way that's compatible with
2577 // Release semantics but weaker than ISH so we'd be fools not to use
2578 // it. Beware: other processors probably don't!
2579 Domain = ARM_MB::ISHST;
2582 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2583 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2584 DAG.getConstant(Domain, MVT::i32));
2587 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2588 const ARMSubtarget *Subtarget) {
2589 // ARM pre v5TE and Thumb1 does not have preload instructions.
2590 if (!(Subtarget->isThumb2() ||
2591 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2592 // Just preserve the chain.
2593 return Op.getOperand(0);
2596 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2598 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2599 // ARMv7 with MP extension has PLDW.
2600 return Op.getOperand(0);
2602 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2603 if (Subtarget->isThumb()) {
2605 isRead = ~isRead & 1;
2606 isData = ~isData & 1;
2609 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2610 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2611 DAG.getConstant(isData, MVT::i32));
2614 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2615 MachineFunction &MF = DAG.getMachineFunction();
2616 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2618 // vastart just stores the address of the VarArgsFrameIndex slot into the
2619 // memory location argument.
2621 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2622 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2623 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2624 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2625 MachinePointerInfo(SV), false, false, 0);
2629 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2630 SDValue &Root, SelectionDAG &DAG,
2632 MachineFunction &MF = DAG.getMachineFunction();
2633 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2635 const TargetRegisterClass *RC;
2636 if (AFI->isThumb1OnlyFunction())
2637 RC = &ARM::tGPRRegClass;
2639 RC = &ARM::GPRRegClass;
2641 // Transform the arguments stored in physical registers into virtual ones.
2642 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2643 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2646 if (NextVA.isMemLoc()) {
2647 MachineFrameInfo *MFI = MF.getFrameInfo();
2648 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2650 // Create load node to retrieve arguments from the stack.
2651 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2652 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2653 MachinePointerInfo::getFixedStack(FI),
2654 false, false, false, 0);
2656 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2657 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2660 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2664 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2665 unsigned InRegsParamRecordIdx,
2667 unsigned &ArgRegsSize,
2668 unsigned &ArgRegsSaveSize)
2671 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2672 unsigned RBegin, REnd;
2673 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2674 NumGPRs = REnd - RBegin;
2676 unsigned int firstUnalloced;
2677 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2678 sizeof(GPRArgRegs) /
2679 sizeof(GPRArgRegs[0]));
2680 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2683 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2684 ArgRegsSize = NumGPRs * 4;
2686 // If parameter is split between stack and GPRs...
2687 if (NumGPRs && Align > 4 &&
2688 (ArgRegsSize < ArgSize ||
2689 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2690 // Add padding for part of param recovered from GPRs. For example,
2691 // if Align == 8, its last byte must be at address K*8 - 1.
2692 // We need to do it, since remained (stack) part of parameter has
2693 // stack alignment, and we need to "attach" "GPRs head" without gaps
2696 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2697 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2699 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2701 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2702 ArgRegsSaveSize = ArgRegsSize + Padding;
2704 // We don't need to extend regs save size for byval parameters if they
2705 // are passed via GPRs only.
2706 ArgRegsSaveSize = ArgRegsSize;
2709 // The remaining GPRs hold either the beginning of variable-argument
2710 // data, or the beginning of an aggregate passed by value (usually
2711 // byval). Either way, we allocate stack slots adjacent to the data
2712 // provided by our caller, and store the unallocated registers there.
2713 // If this is a variadic function, the va_list pointer will begin with
2714 // these values; otherwise, this reassembles a (byval) structure that
2715 // was split between registers and memory.
2716 // Return: The frame index registers were stored into.
2718 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2719 SDLoc dl, SDValue &Chain,
2720 const Value *OrigArg,
2721 unsigned InRegsParamRecordIdx,
2722 unsigned OffsetFromOrigArg,
2726 unsigned ByValStoreOffset,
2727 unsigned TotalArgRegsSaveSize) const {
2729 // Currently, two use-cases possible:
2730 // Case #1. Non-var-args function, and we meet first byval parameter.
2731 // Setup first unallocated register as first byval register;
2732 // eat all remained registers
2733 // (these two actions are performed by HandleByVal method).
2734 // Then, here, we initialize stack frame with
2735 // "store-reg" instructions.
2736 // Case #2. Var-args function, that doesn't contain byval parameters.
2737 // The same: eat all remained unallocated registers,
2738 // initialize stack frame.
2740 MachineFunction &MF = DAG.getMachineFunction();
2741 MachineFrameInfo *MFI = MF.getFrameInfo();
2742 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2743 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2744 unsigned RBegin, REnd;
2745 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2746 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2747 firstRegToSaveIndex = RBegin - ARM::R0;
2748 lastRegToSaveIndex = REnd - ARM::R0;
2750 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2751 (GPRArgRegs, array_lengthof(GPRArgRegs));
2752 lastRegToSaveIndex = 4;
2755 unsigned ArgRegsSize, ArgRegsSaveSize;
2756 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2757 ArgRegsSize, ArgRegsSaveSize);
2759 // Store any by-val regs to their spots on the stack so that they may be
2760 // loaded by deferencing the result of formal parameter pointer or va_next.
2761 // Note: once stack area for byval/varargs registers
2762 // was initialized, it can't be initialized again.
2763 if (ArgRegsSaveSize) {
2764 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2767 assert(AFI->getStoredByValParamsPadding() == 0 &&
2768 "The only parameter may be padded.");
2769 AFI->setStoredByValParamsPadding(Padding);
2772 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2775 (int64_t)TotalArgRegsSaveSize,
2777 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2779 MFI->CreateFixedObject(Padding,
2780 ArgOffset + ByValStoreOffset -
2781 (int64_t)ArgRegsSaveSize,
2785 SmallVector<SDValue, 4> MemOps;
2786 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2787 ++firstRegToSaveIndex, ++i) {
2788 const TargetRegisterClass *RC;
2789 if (AFI->isThumb1OnlyFunction())
2790 RC = &ARM::tGPRRegClass;
2792 RC = &ARM::GPRRegClass;
2794 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2795 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2797 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2798 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2800 MemOps.push_back(Store);
2801 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2802 DAG.getConstant(4, getPointerTy()));
2805 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2807 if (!MemOps.empty())
2808 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2809 &MemOps[0], MemOps.size());
2813 // We cannot allocate a zero-byte object for the first variadic argument,
2814 // so just make up a size.
2817 // This will point to the next argument passed via stack.
2818 return MFI->CreateFixedObject(
2819 ArgSize, ArgOffset, !ForceMutable);
2823 // Setup stack frame, the va_list pointer will start from.
2825 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2826 SDLoc dl, SDValue &Chain,
2828 unsigned TotalArgRegsSaveSize,
2829 bool ForceMutable) const {
2830 MachineFunction &MF = DAG.getMachineFunction();
2831 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2833 // Try to store any remaining integer argument regs
2834 // to their spots on the stack so that they may be loaded by deferencing
2835 // the result of va_next.
2836 // If there is no regs to be stored, just point address after last
2837 // argument passed via stack.
2839 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2840 0, ArgOffset, 0, ForceMutable, 0, TotalArgRegsSaveSize);
2842 AFI->setVarArgsFrameIndex(FrameIndex);
2846 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2847 CallingConv::ID CallConv, bool isVarArg,
2848 const SmallVectorImpl<ISD::InputArg>
2850 SDLoc dl, SelectionDAG &DAG,
2851 SmallVectorImpl<SDValue> &InVals)
2853 MachineFunction &MF = DAG.getMachineFunction();
2854 MachineFrameInfo *MFI = MF.getFrameInfo();
2856 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2858 // Assign locations to all of the incoming arguments.
2859 SmallVector<CCValAssign, 16> ArgLocs;
2860 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2861 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2862 CCInfo.AnalyzeFormalArguments(Ins,
2863 CCAssignFnForNode(CallConv, /* Return*/ false,
2866 SmallVector<SDValue, 16> ArgValues;
2867 int lastInsIndex = -1;
2869 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2870 unsigned CurArgIdx = 0;
2872 // Initially ArgRegsSaveSize is zero.
2873 // Then we increase this value each time we meet byval parameter.
2874 // We also increase this value in case of varargs function.
2875 AFI->setArgRegsSaveSize(0);
2877 unsigned ByValStoreOffset = 0;
2878 unsigned TotalArgRegsSaveSize = 0;
2879 unsigned ArgRegsSaveSizeMaxAlign = 4;
2881 // Calculate the amount of stack space that we need to allocate to store
2882 // byval and variadic arguments that are passed in registers.
2883 // We need to know this before we allocate the first byval or variadic
2884 // argument, as they will be allocated a stack slot below the CFA (Canonical
2885 // Frame Address, the stack pointer at entry to the function).
2886 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2887 CCValAssign &VA = ArgLocs[i];
2888 if (VA.isMemLoc()) {
2889 int index = VA.getValNo();
2890 if (index != lastInsIndex) {
2891 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2892 if (Flags.isByVal()) {
2893 unsigned ExtraArgRegsSize;
2894 unsigned ExtraArgRegsSaveSize;
2895 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2896 Flags.getByValSize(),
2897 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2899 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2900 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2901 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2902 CCInfo.nextInRegsParam();
2904 lastInsIndex = index;
2908 CCInfo.rewindByValRegsInfo();
2911 unsigned ExtraArgRegsSize;
2912 unsigned ExtraArgRegsSaveSize;
2913 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2914 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2915 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2917 // If the arg regs save area contains N-byte aligned values, the
2918 // bottom of it must be at least N-byte aligned.
2919 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2920 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2922 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2923 CCValAssign &VA = ArgLocs[i];
2924 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2925 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2926 // Arguments stored in registers.
2927 if (VA.isRegLoc()) {
2928 EVT RegVT = VA.getLocVT();
2930 if (VA.needsCustom()) {
2931 // f64 and vector types are split up into multiple registers or
2932 // combinations of registers and stack slots.
2933 if (VA.getLocVT() == MVT::v2f64) {
2934 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2936 VA = ArgLocs[++i]; // skip ahead to next loc
2938 if (VA.isMemLoc()) {
2939 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2940 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2941 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2942 MachinePointerInfo::getFixedStack(FI),
2943 false, false, false, 0);
2945 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2948 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2949 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2950 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2951 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2952 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2954 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2957 const TargetRegisterClass *RC;
2959 if (RegVT == MVT::f32)
2960 RC = &ARM::SPRRegClass;
2961 else if (RegVT == MVT::f64)
2962 RC = &ARM::DPRRegClass;
2963 else if (RegVT == MVT::v2f64)
2964 RC = &ARM::QPRRegClass;
2965 else if (RegVT == MVT::i32)
2966 RC = AFI->isThumb1OnlyFunction() ?
2967 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2968 (const TargetRegisterClass*)&ARM::GPRRegClass;
2970 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2972 // Transform the arguments in physical registers into virtual ones.
2973 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2974 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2977 // If this is an 8 or 16-bit value, it is really passed promoted
2978 // to 32 bits. Insert an assert[sz]ext to capture this, then
2979 // truncate to the right size.
2980 switch (VA.getLocInfo()) {
2981 default: llvm_unreachable("Unknown loc info!");
2982 case CCValAssign::Full: break;
2983 case CCValAssign::BCvt:
2984 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2986 case CCValAssign::SExt:
2987 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2988 DAG.getValueType(VA.getValVT()));
2989 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2991 case CCValAssign::ZExt:
2992 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2993 DAG.getValueType(VA.getValVT()));
2994 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2998 InVals.push_back(ArgValue);
3000 } else { // VA.isRegLoc()
3003 assert(VA.isMemLoc());
3004 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3006 int index = ArgLocs[i].getValNo();
3008 // Some Ins[] entries become multiple ArgLoc[] entries.
3009 // Process them only once.
3010 if (index != lastInsIndex)
3012 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3013 // FIXME: For now, all byval parameter objects are marked mutable.
3014 // This can be changed with more analysis.
3015 // In case of tail call optimization mark all arguments mutable.
3016 // Since they could be overwritten by lowering of arguments in case of
3018 if (Flags.isByVal()) {
3019 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3021 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3022 int FrameIndex = StoreByValRegs(
3023 CCInfo, DAG, dl, Chain, CurOrigArg,
3025 Ins[VA.getValNo()].PartOffset,
3026 VA.getLocMemOffset(),
3027 Flags.getByValSize(),
3028 true /*force mutable frames*/,
3030 TotalArgRegsSaveSize);
3031 ByValStoreOffset += Flags.getByValSize();
3032 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3033 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3034 CCInfo.nextInRegsParam();
3036 unsigned FIOffset = VA.getLocMemOffset();
3037 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3040 // Create load nodes to retrieve arguments from the stack.
3041 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3042 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3043 MachinePointerInfo::getFixedStack(FI),
3044 false, false, false, 0));
3046 lastInsIndex = index;
3053 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3054 CCInfo.getNextStackOffset(),
3055 TotalArgRegsSaveSize);
3057 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3062 /// isFloatingPointZero - Return true if this is +0.0.
3063 static bool isFloatingPointZero(SDValue Op) {
3064 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3065 return CFP->getValueAPF().isPosZero();
3066 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3067 // Maybe this has already been legalized into the constant pool?
3068 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3069 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3070 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3071 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3072 return CFP->getValueAPF().isPosZero();
3078 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3079 /// the given operands.
3081 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3082 SDValue &ARMcc, SelectionDAG &DAG,
3084 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3085 unsigned C = RHSC->getZExtValue();
3086 if (!isLegalICmpImmediate(C)) {
3087 // Constant does not fit, try adjusting it by one?
3092 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3093 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3094 RHS = DAG.getConstant(C-1, MVT::i32);
3099 if (C != 0 && isLegalICmpImmediate(C-1)) {
3100 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3101 RHS = DAG.getConstant(C-1, MVT::i32);
3106 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3107 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3108 RHS = DAG.getConstant(C+1, MVT::i32);
3113 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3114 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3115 RHS = DAG.getConstant(C+1, MVT::i32);
3122 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3123 ARMISD::NodeType CompareType;
3126 CompareType = ARMISD::CMP;
3131 CompareType = ARMISD::CMPZ;
3134 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3135 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3138 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3140 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3143 if (!isFloatingPointZero(RHS))
3144 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3146 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3147 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3150 /// duplicateCmp - Glue values can have only one use, so this function
3151 /// duplicates a comparison node.
3153 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3154 unsigned Opc = Cmp.getOpcode();
3156 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3157 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3159 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3160 Cmp = Cmp.getOperand(0);
3161 Opc = Cmp.getOpcode();
3162 if (Opc == ARMISD::CMPFP)
3163 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3165 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3166 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3168 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3171 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3172 SDValue Cond = Op.getOperand(0);
3173 SDValue SelectTrue = Op.getOperand(1);
3174 SDValue SelectFalse = Op.getOperand(2);
3179 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3180 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3182 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3183 const ConstantSDNode *CMOVTrue =
3184 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3185 const ConstantSDNode *CMOVFalse =
3186 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3188 if (CMOVTrue && CMOVFalse) {
3189 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3190 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3194 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3196 False = SelectFalse;
3197 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3202 if (True.getNode() && False.getNode()) {
3203 EVT VT = Op.getValueType();
3204 SDValue ARMcc = Cond.getOperand(2);
3205 SDValue CCR = Cond.getOperand(3);
3206 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3207 assert(True.getValueType() == VT);
3208 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3213 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3214 // undefined bits before doing a full-word comparison with zero.
3215 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3216 DAG.getConstant(1, Cond.getValueType()));
3218 return DAG.getSelectCC(dl, Cond,
3219 DAG.getConstant(0, Cond.getValueType()),
3220 SelectTrue, SelectFalse, ISD::SETNE);
3223 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3224 if (CC == ISD::SETNE)
3226 return ISD::getSetCCInverse(CC, true);
3229 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3230 bool &swpCmpOps, bool &swpVselOps) {
3231 // Start by selecting the GE condition code for opcodes that return true for
3233 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3235 CondCode = ARMCC::GE;
3237 // and GT for opcodes that return false for 'equality'.
3238 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3240 CondCode = ARMCC::GT;
3242 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3243 // to swap the compare operands.
3244 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3248 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3249 // If we have an unordered opcode, we need to swap the operands to the VSEL
3250 // instruction (effectively negating the condition).
3252 // This also has the effect of swapping which one of 'less' or 'greater'
3253 // returns true, so we also swap the compare operands. It also switches
3254 // whether we return true for 'equality', so we compensate by picking the
3255 // opposite condition code to our original choice.
3256 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3257 CC == ISD::SETUGT) {
3258 swpCmpOps = !swpCmpOps;
3259 swpVselOps = !swpVselOps;
3260 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3263 // 'ordered' is 'anything but unordered', so use the VS condition code and
3264 // swap the VSEL operands.
3265 if (CC == ISD::SETO) {
3266 CondCode = ARMCC::VS;
3270 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3271 // code and swap the VSEL operands.
3272 if (CC == ISD::SETUNE) {
3273 CondCode = ARMCC::EQ;
3278 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3279 EVT VT = Op.getValueType();
3280 SDValue LHS = Op.getOperand(0);
3281 SDValue RHS = Op.getOperand(1);
3282 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3283 SDValue TrueVal = Op.getOperand(2);
3284 SDValue FalseVal = Op.getOperand(3);
3287 if (LHS.getValueType() == MVT::i32) {
3288 // Try to generate VSEL on ARMv8.
3289 // The VSEL instruction can't use all the usual ARM condition
3290 // codes: it only has two bits to select the condition code, so it's
3291 // constrained to use only GE, GT, VS and EQ.
3293 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3294 // swap the operands of the previous compare instruction (effectively
3295 // inverting the compare condition, swapping 'less' and 'greater') and
3296 // sometimes need to swap the operands to the VSEL (which inverts the
3297 // condition in the sense of firing whenever the previous condition didn't)
3298 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3299 TrueVal.getValueType() == MVT::f64)) {
3300 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3301 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3302 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3303 CC = getInverseCCForVSEL(CC);
3304 std::swap(TrueVal, FalseVal);
3309 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3310 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3311 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3315 ARMCC::CondCodes CondCode, CondCode2;
3316 FPCCToARMCC(CC, CondCode, CondCode2);
3318 // Try to generate VSEL on ARMv8.
3319 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3320 TrueVal.getValueType() == MVT::f64)) {
3321 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3322 // same operands, as follows:
3323 // c = fcmp [ogt, olt, ugt, ult] a, b
3325 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3326 // handled differently than the original code sequence.
3327 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3329 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3330 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3331 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3332 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3335 bool swpCmpOps = false;
3336 bool swpVselOps = false;
3337 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3339 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3340 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3342 std::swap(LHS, RHS);
3344 std::swap(TrueVal, FalseVal);
3348 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3349 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3350 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3351 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3353 if (CondCode2 != ARMCC::AL) {
3354 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3355 // FIXME: Needs another CMP because flag can have but one use.
3356 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3357 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3358 Result, TrueVal, ARMcc2, CCR, Cmp2);
3363 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3364 /// to morph to an integer compare sequence.
3365 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3366 const ARMSubtarget *Subtarget) {
3367 SDNode *N = Op.getNode();
3368 if (!N->hasOneUse())
3369 // Otherwise it requires moving the value from fp to integer registers.
3371 if (!N->getNumValues())
3373 EVT VT = Op.getValueType();
3374 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3375 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3376 // vmrs are very slow, e.g. cortex-a8.
3379 if (isFloatingPointZero(Op)) {
3383 return ISD::isNormalLoad(N);
3386 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3387 if (isFloatingPointZero(Op))
3388 return DAG.getConstant(0, MVT::i32);
3390 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3391 return DAG.getLoad(MVT::i32, SDLoc(Op),
3392 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3393 Ld->isVolatile(), Ld->isNonTemporal(),
3394 Ld->isInvariant(), Ld->getAlignment());
3396 llvm_unreachable("Unknown VFP cmp argument!");
3399 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3400 SDValue &RetVal1, SDValue &RetVal2) {
3401 if (isFloatingPointZero(Op)) {
3402 RetVal1 = DAG.getConstant(0, MVT::i32);
3403 RetVal2 = DAG.getConstant(0, MVT::i32);
3407 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3408 SDValue Ptr = Ld->getBasePtr();
3409 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3410 Ld->getChain(), Ptr,
3411 Ld->getPointerInfo(),
3412 Ld->isVolatile(), Ld->isNonTemporal(),
3413 Ld->isInvariant(), Ld->getAlignment());
3415 EVT PtrType = Ptr.getValueType();
3416 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3417 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3418 PtrType, Ptr, DAG.getConstant(4, PtrType));
3419 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3420 Ld->getChain(), NewPtr,
3421 Ld->getPointerInfo().getWithOffset(4),
3422 Ld->isVolatile(), Ld->isNonTemporal(),
3423 Ld->isInvariant(), NewAlign);
3427 llvm_unreachable("Unknown VFP cmp argument!");
3430 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3431 /// f32 and even f64 comparisons to integer ones.
3433 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3434 SDValue Chain = Op.getOperand(0);
3435 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3436 SDValue LHS = Op.getOperand(2);
3437 SDValue RHS = Op.getOperand(3);
3438 SDValue Dest = Op.getOperand(4);
3441 bool LHSSeenZero = false;
3442 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3443 bool RHSSeenZero = false;
3444 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3445 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3446 // If unsafe fp math optimization is enabled and there are no other uses of
3447 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3448 // to an integer comparison.
3449 if (CC == ISD::SETOEQ)
3451 else if (CC == ISD::SETUNE)
3454 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3456 if (LHS.getValueType() == MVT::f32) {
3457 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3458 bitcastf32Toi32(LHS, DAG), Mask);
3459 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3460 bitcastf32Toi32(RHS, DAG), Mask);
3461 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3462 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3463 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3464 Chain, Dest, ARMcc, CCR, Cmp);
3469 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3470 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3471 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3472 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3473 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3474 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3475 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3476 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3477 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3483 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3484 SDValue Chain = Op.getOperand(0);
3485 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3486 SDValue LHS = Op.getOperand(2);
3487 SDValue RHS = Op.getOperand(3);
3488 SDValue Dest = Op.getOperand(4);
3491 if (LHS.getValueType() == MVT::i32) {
3493 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3494 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3495 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3496 Chain, Dest, ARMcc, CCR, Cmp);
3499 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3501 if (getTargetMachine().Options.UnsafeFPMath &&
3502 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3503 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3504 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3505 if (Result.getNode())
3509 ARMCC::CondCodes CondCode, CondCode2;
3510 FPCCToARMCC(CC, CondCode, CondCode2);
3512 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3513 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3514 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3515 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3516 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3517 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3518 if (CondCode2 != ARMCC::AL) {
3519 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3520 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3521 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3526 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3527 SDValue Chain = Op.getOperand(0);
3528 SDValue Table = Op.getOperand(1);
3529 SDValue Index = Op.getOperand(2);
3532 EVT PTy = getPointerTy();
3533 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3534 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3535 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3536 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3537 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3538 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3539 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3540 if (Subtarget->isThumb2()) {
3541 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3542 // which does another jump to the destination. This also makes it easier
3543 // to translate it to TBB / TBH later.
3544 // FIXME: This might not work if the function is extremely large.
3545 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3546 Addr, Op.getOperand(2), JTI, UId);
3548 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3549 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3550 MachinePointerInfo::getJumpTable(),
3551 false, false, false, 0);
3552 Chain = Addr.getValue(1);
3553 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3554 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3556 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3557 MachinePointerInfo::getJumpTable(),
3558 false, false, false, 0);
3559 Chain = Addr.getValue(1);
3560 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3564 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3565 EVT VT = Op.getValueType();
3568 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3569 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3571 return DAG.UnrollVectorOp(Op.getNode());
3574 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3575 "Invalid type for custom lowering!");
3576 if (VT != MVT::v4i16)
3577 return DAG.UnrollVectorOp(Op.getNode());
3579 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3580 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3583 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3584 EVT VT = Op.getValueType();
3586 return LowerVectorFP_TO_INT(Op, DAG);
3591 switch (Op.getOpcode()) {
3592 default: llvm_unreachable("Invalid opcode!");
3593 case ISD::FP_TO_SINT:
3594 Opc = ARMISD::FTOSI;
3596 case ISD::FP_TO_UINT:
3597 Opc = ARMISD::FTOUI;
3600 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3601 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3604 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3605 EVT VT = Op.getValueType();
3608 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3609 if (VT.getVectorElementType() == MVT::f32)
3611 return DAG.UnrollVectorOp(Op.getNode());
3614 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3615 "Invalid type for custom lowering!");
3616 if (VT != MVT::v4f32)
3617 return DAG.UnrollVectorOp(Op.getNode());
3621 switch (Op.getOpcode()) {
3622 default: llvm_unreachable("Invalid opcode!");
3623 case ISD::SINT_TO_FP:
3624 CastOpc = ISD::SIGN_EXTEND;
3625 Opc = ISD::SINT_TO_FP;
3627 case ISD::UINT_TO_FP:
3628 CastOpc = ISD::ZERO_EXTEND;
3629 Opc = ISD::UINT_TO_FP;
3633 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3634 return DAG.getNode(Opc, dl, VT, Op);
3637 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3638 EVT VT = Op.getValueType();
3640 return LowerVectorINT_TO_FP(Op, DAG);
3645 switch (Op.getOpcode()) {
3646 default: llvm_unreachable("Invalid opcode!");
3647 case ISD::SINT_TO_FP:
3648 Opc = ARMISD::SITOF;
3650 case ISD::UINT_TO_FP:
3651 Opc = ARMISD::UITOF;
3655 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3656 return DAG.getNode(Opc, dl, VT, Op);
3659 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3660 // Implement fcopysign with a fabs and a conditional fneg.
3661 SDValue Tmp0 = Op.getOperand(0);
3662 SDValue Tmp1 = Op.getOperand(1);
3664 EVT VT = Op.getValueType();
3665 EVT SrcVT = Tmp1.getValueType();
3666 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3667 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3668 bool UseNEON = !InGPR && Subtarget->hasNEON();
3671 // Use VBSL to copy the sign bit.
3672 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3673 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3674 DAG.getTargetConstant(EncodedVal, MVT::i32));
3675 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3677 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3678 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3679 DAG.getConstant(32, MVT::i32));
3680 else /*if (VT == MVT::f32)*/
3681 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3682 if (SrcVT == MVT::f32) {
3683 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3685 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3686 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3687 DAG.getConstant(32, MVT::i32));
3688 } else if (VT == MVT::f32)
3689 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3690 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3691 DAG.getConstant(32, MVT::i32));
3692 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3693 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3695 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3697 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3698 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3699 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3701 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3702 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3703 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3704 if (VT == MVT::f32) {
3705 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3706 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3707 DAG.getConstant(0, MVT::i32));
3709 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3715 // Bitcast operand 1 to i32.
3716 if (SrcVT == MVT::f64)
3717 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3718 &Tmp1, 1).getValue(1);
3719 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3721 // Or in the signbit with integer operations.
3722 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3723 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3724 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3725 if (VT == MVT::f32) {
3726 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3727 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3728 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3729 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3732 // f64: Or the high part with signbit and then combine two parts.
3733 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3735 SDValue Lo = Tmp0.getValue(0);
3736 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3737 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3738 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3741 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3742 MachineFunction &MF = DAG.getMachineFunction();
3743 MachineFrameInfo *MFI = MF.getFrameInfo();
3744 MFI->setReturnAddressIsTaken(true);
3746 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3749 EVT VT = Op.getValueType();
3751 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3753 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3754 SDValue Offset = DAG.getConstant(4, MVT::i32);
3755 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3756 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3757 MachinePointerInfo(), false, false, false, 0);
3760 // Return LR, which contains the return address. Mark it an implicit live-in.
3761 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3762 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3765 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3766 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3767 MFI->setFrameAddressIsTaken(true);
3769 EVT VT = Op.getValueType();
3770 SDLoc dl(Op); // FIXME probably not meaningful
3771 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3772 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetMachO())
3773 ? ARM::R7 : ARM::R11;
3774 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3776 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3777 MachinePointerInfo(),
3778 false, false, false, 0);
3782 /// ExpandBITCAST - If the target supports VFP, this function is called to
3783 /// expand a bit convert where either the source or destination type is i64 to
3784 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3785 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3786 /// vectors), since the legalizer won't know what to do with that.
3787 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3790 SDValue Op = N->getOperand(0);
3792 // This function is only supposed to be called for i64 types, either as the
3793 // source or destination of the bit convert.
3794 EVT SrcVT = Op.getValueType();
3795 EVT DstVT = N->getValueType(0);
3796 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3797 "ExpandBITCAST called for non-i64 type");
3799 // Turn i64->f64 into VMOVDRR.
3800 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3801 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3802 DAG.getConstant(0, MVT::i32));
3803 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3804 DAG.getConstant(1, MVT::i32));
3805 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3806 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3809 // Turn f64->i64 into VMOVRRD.
3810 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3811 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3812 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3813 // Merge the pieces into a single i64 value.
3814 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3820 /// getZeroVector - Returns a vector of specified type with all zero elements.
3821 /// Zero vectors are used to represent vector negation and in those cases
3822 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3823 /// not support i64 elements, so sometimes the zero vectors will need to be
3824 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3826 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3827 assert(VT.isVector() && "Expected a vector type");
3828 // The canonical modified immediate encoding of a zero vector is....0!
3829 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3830 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3831 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3832 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3835 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3836 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3837 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3838 SelectionDAG &DAG) const {
3839 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3840 EVT VT = Op.getValueType();
3841 unsigned VTBits = VT.getSizeInBits();
3843 SDValue ShOpLo = Op.getOperand(0);
3844 SDValue ShOpHi = Op.getOperand(1);
3845 SDValue ShAmt = Op.getOperand(2);
3847 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3849 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3851 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3852 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3853 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3854 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3855 DAG.getConstant(VTBits, MVT::i32));
3856 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3857 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3858 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3860 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3861 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3863 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3864 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3867 SDValue Ops[2] = { Lo, Hi };
3868 return DAG.getMergeValues(Ops, 2, dl);
3871 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3872 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3873 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3874 SelectionDAG &DAG) const {
3875 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3876 EVT VT = Op.getValueType();
3877 unsigned VTBits = VT.getSizeInBits();
3879 SDValue ShOpLo = Op.getOperand(0);
3880 SDValue ShOpHi = Op.getOperand(1);
3881 SDValue ShAmt = Op.getOperand(2);
3884 assert(Op.getOpcode() == ISD::SHL_PARTS);
3885 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3886 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3887 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3888 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3889 DAG.getConstant(VTBits, MVT::i32));
3890 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3891 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3893 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3894 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3895 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3897 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3898 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3901 SDValue Ops[2] = { Lo, Hi };
3902 return DAG.getMergeValues(Ops, 2, dl);
3905 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3906 SelectionDAG &DAG) const {
3907 // The rounding mode is in bits 23:22 of the FPSCR.
3908 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3909 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3910 // so that the shift + and get folded into a bitfield extract.
3912 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3913 DAG.getConstant(Intrinsic::arm_get_fpscr,
3915 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3916 DAG.getConstant(1U << 22, MVT::i32));
3917 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3918 DAG.getConstant(22, MVT::i32));
3919 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3920 DAG.getConstant(3, MVT::i32));
3923 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3924 const ARMSubtarget *ST) {
3925 EVT VT = N->getValueType(0);
3928 if (!ST->hasV6T2Ops())
3931 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3932 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3935 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3936 /// for each 16-bit element from operand, repeated. The basic idea is to
3937 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3939 /// Trace for v4i16:
3940 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3941 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3942 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3943 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3944 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3945 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3946 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3947 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3948 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3949 EVT VT = N->getValueType(0);
3952 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3953 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3954 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3955 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3956 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3957 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3960 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3961 /// bit-count for each 16-bit element from the operand. We need slightly
3962 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3963 /// 64/128-bit registers.
3965 /// Trace for v4i16:
3966 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3967 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3968 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3969 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3970 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3971 EVT VT = N->getValueType(0);
3974 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3975 if (VT.is64BitVector()) {
3976 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3977 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3978 DAG.getIntPtrConstant(0));
3980 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3981 BitCounts, DAG.getIntPtrConstant(0));
3982 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3986 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3987 /// bit-count for each 32-bit element from the operand. The idea here is
3988 /// to split the vector into 16-bit elements, leverage the 16-bit count
3989 /// routine, and then combine the results.
3991 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3992 /// input = [v0 v1 ] (vi: 32-bit elements)
3993 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3994 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3995 /// vrev: N0 = [k1 k0 k3 k2 ]
3997 /// N1 =+[k1 k0 k3 k2 ]
3999 /// N2 =+[k1 k3 k0 k2 ]
4001 /// Extended =+[k1 k3 k0 k2 ]
4003 /// Extracted=+[k1 k3 ]
4005 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4006 EVT VT = N->getValueType(0);
4009 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4011 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4012 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4013 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4014 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4015 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4017 if (VT.is64BitVector()) {
4018 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4019 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4020 DAG.getIntPtrConstant(0));
4022 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4023 DAG.getIntPtrConstant(0));
4024 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4028 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4029 const ARMSubtarget *ST) {
4030 EVT VT = N->getValueType(0);
4032 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4033 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4034 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4035 "Unexpected type for custom ctpop lowering");
4037 if (VT.getVectorElementType() == MVT::i32)
4038 return lowerCTPOP32BitElements(N, DAG);
4040 return lowerCTPOP16BitElements(N, DAG);
4043 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4044 const ARMSubtarget *ST) {
4045 EVT VT = N->getValueType(0);
4051 // Lower vector shifts on NEON to use VSHL.
4052 assert(ST->hasNEON() && "unexpected vector shift");
4054 // Left shifts translate directly to the vshiftu intrinsic.
4055 if (N->getOpcode() == ISD::SHL)
4056 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4057 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4058 N->getOperand(0), N->getOperand(1));
4060 assert((N->getOpcode() == ISD::SRA ||
4061 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4063 // NEON uses the same intrinsics for both left and right shifts. For
4064 // right shifts, the shift amounts are negative, so negate the vector of
4066 EVT ShiftVT = N->getOperand(1).getValueType();
4067 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4068 getZeroVector(ShiftVT, DAG, dl),
4070 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4071 Intrinsic::arm_neon_vshifts :
4072 Intrinsic::arm_neon_vshiftu);
4073 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4074 DAG.getConstant(vshiftInt, MVT::i32),
4075 N->getOperand(0), NegatedCount);
4078 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4079 const ARMSubtarget *ST) {
4080 EVT VT = N->getValueType(0);
4083 // We can get here for a node like i32 = ISD::SHL i32, i64
4087 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4088 "Unknown shift to lower!");
4090 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4091 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4092 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4095 // If we are in thumb mode, we don't have RRX.
4096 if (ST->isThumb1Only()) return SDValue();
4098 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4099 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4100 DAG.getConstant(0, MVT::i32));
4101 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4102 DAG.getConstant(1, MVT::i32));
4104 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4105 // captures the result into a carry flag.
4106 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4107 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4109 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4110 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4112 // Merge the pieces into a single i64 value.
4113 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4116 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4117 SDValue TmpOp0, TmpOp1;
4118 bool Invert = false;
4122 SDValue Op0 = Op.getOperand(0);
4123 SDValue Op1 = Op.getOperand(1);
4124 SDValue CC = Op.getOperand(2);
4125 EVT VT = Op.getValueType();
4126 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4129 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4130 switch (SetCCOpcode) {
4131 default: llvm_unreachable("Illegal FP comparison");
4133 case ISD::SETNE: Invert = true; // Fallthrough
4135 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4137 case ISD::SETLT: Swap = true; // Fallthrough
4139 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4141 case ISD::SETLE: Swap = true; // Fallthrough
4143 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4144 case ISD::SETUGE: Swap = true; // Fallthrough
4145 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4146 case ISD::SETUGT: Swap = true; // Fallthrough
4147 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4148 case ISD::SETUEQ: Invert = true; // Fallthrough
4150 // Expand this to (OLT | OGT).
4154 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4155 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4157 case ISD::SETUO: Invert = true; // Fallthrough
4159 // Expand this to (OLT | OGE).
4163 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4164 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4168 // Integer comparisons.
4169 switch (SetCCOpcode) {
4170 default: llvm_unreachable("Illegal integer comparison");
4171 case ISD::SETNE: Invert = true;
4172 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4173 case ISD::SETLT: Swap = true;
4174 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4175 case ISD::SETLE: Swap = true;
4176 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4177 case ISD::SETULT: Swap = true;
4178 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4179 case ISD::SETULE: Swap = true;
4180 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4183 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4184 if (Opc == ARMISD::VCEQ) {
4187 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4189 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4192 // Ignore bitconvert.
4193 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4194 AndOp = AndOp.getOperand(0);
4196 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4198 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4199 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4206 std::swap(Op0, Op1);
4208 // If one of the operands is a constant vector zero, attempt to fold the
4209 // comparison to a specialized compare-against-zero form.
4211 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4213 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4214 if (Opc == ARMISD::VCGE)
4215 Opc = ARMISD::VCLEZ;
4216 else if (Opc == ARMISD::VCGT)
4217 Opc = ARMISD::VCLTZ;
4222 if (SingleOp.getNode()) {
4225 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4227 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4229 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4231 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4233 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4235 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4238 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4242 Result = DAG.getNOT(dl, Result, VT);
4247 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4248 /// valid vector constant for a NEON instruction with a "modified immediate"
4249 /// operand (e.g., VMOV). If so, return the encoded value.
4250 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4251 unsigned SplatBitSize, SelectionDAG &DAG,
4252 EVT &VT, bool is128Bits, NEONModImmType type) {
4253 unsigned OpCmode, Imm;
4255 // SplatBitSize is set to the smallest size that splats the vector, so a
4256 // zero vector will always have SplatBitSize == 8. However, NEON modified
4257 // immediate instructions others than VMOV do not support the 8-bit encoding
4258 // of a zero vector, and the default encoding of zero is supposed to be the
4263 switch (SplatBitSize) {
4265 if (type != VMOVModImm)
4267 // Any 1-byte value is OK. Op=0, Cmode=1110.
4268 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4271 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4275 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4276 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4277 if ((SplatBits & ~0xff) == 0) {
4278 // Value = 0x00nn: Op=x, Cmode=100x.
4283 if ((SplatBits & ~0xff00) == 0) {
4284 // Value = 0xnn00: Op=x, Cmode=101x.
4286 Imm = SplatBits >> 8;
4292 // NEON's 32-bit VMOV supports splat values where:
4293 // * only one byte is nonzero, or
4294 // * the least significant byte is 0xff and the second byte is nonzero, or
4295 // * the least significant 2 bytes are 0xff and the third is nonzero.
4296 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4297 if ((SplatBits & ~0xff) == 0) {
4298 // Value = 0x000000nn: Op=x, Cmode=000x.
4303 if ((SplatBits & ~0xff00) == 0) {
4304 // Value = 0x0000nn00: Op=x, Cmode=001x.
4306 Imm = SplatBits >> 8;
4309 if ((SplatBits & ~0xff0000) == 0) {
4310 // Value = 0x00nn0000: Op=x, Cmode=010x.
4312 Imm = SplatBits >> 16;
4315 if ((SplatBits & ~0xff000000) == 0) {
4316 // Value = 0xnn000000: Op=x, Cmode=011x.
4318 Imm = SplatBits >> 24;
4322 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4323 if (type == OtherModImm) return SDValue();
4325 if ((SplatBits & ~0xffff) == 0 &&
4326 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4327 // Value = 0x0000nnff: Op=x, Cmode=1100.
4329 Imm = SplatBits >> 8;
4333 if ((SplatBits & ~0xffffff) == 0 &&
4334 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4335 // Value = 0x00nnffff: Op=x, Cmode=1101.
4337 Imm = SplatBits >> 16;
4341 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4342 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4343 // VMOV.I32. A (very) minor optimization would be to replicate the value
4344 // and fall through here to test for a valid 64-bit splat. But, then the
4345 // caller would also need to check and handle the change in size.
4349 if (type != VMOVModImm)
4351 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4352 uint64_t BitMask = 0xff;
4354 unsigned ImmMask = 1;
4356 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4357 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4360 } else if ((SplatBits & BitMask) != 0) {
4366 // Op=1, Cmode=1110.
4368 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4373 llvm_unreachable("unexpected size for isNEONModifiedImm");
4376 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4377 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4380 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4381 const ARMSubtarget *ST) const {
4385 bool IsDouble = Op.getValueType() == MVT::f64;
4386 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4388 // Try splatting with a VMOV.f32...
4389 APFloat FPVal = CFP->getValueAPF();
4390 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4393 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4394 // We have code in place to select a valid ConstantFP already, no need to
4399 // It's a float and we are trying to use NEON operations where
4400 // possible. Lower it to a splat followed by an extract.
4402 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4403 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4405 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4406 DAG.getConstant(0, MVT::i32));
4409 // The rest of our options are NEON only, make sure that's allowed before
4411 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4415 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4417 // It wouldn't really be worth bothering for doubles except for one very
4418 // important value, which does happen to match: 0.0. So make sure we don't do
4420 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4423 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4424 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4426 if (NewVal != SDValue()) {
4428 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4431 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4433 // It's a float: cast and extract a vector element.
4434 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4436 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4437 DAG.getConstant(0, MVT::i32));
4440 // Finally, try a VMVN.i32
4441 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4443 if (NewVal != SDValue()) {
4445 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4448 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4450 // It's a float: cast and extract a vector element.
4451 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4453 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4454 DAG.getConstant(0, MVT::i32));
4460 // check if an VEXT instruction can handle the shuffle mask when the
4461 // vector sources of the shuffle are the same.
4462 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4463 unsigned NumElts = VT.getVectorNumElements();
4465 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4471 // If this is a VEXT shuffle, the immediate value is the index of the first
4472 // element. The other shuffle indices must be the successive elements after
4474 unsigned ExpectedElt = Imm;
4475 for (unsigned i = 1; i < NumElts; ++i) {
4476 // Increment the expected index. If it wraps around, just follow it
4477 // back to index zero and keep going.
4479 if (ExpectedElt == NumElts)
4482 if (M[i] < 0) continue; // ignore UNDEF indices
4483 if (ExpectedElt != static_cast<unsigned>(M[i]))
4491 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4492 bool &ReverseVEXT, unsigned &Imm) {
4493 unsigned NumElts = VT.getVectorNumElements();
4494 ReverseVEXT = false;
4496 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4502 // If this is a VEXT shuffle, the immediate value is the index of the first
4503 // element. The other shuffle indices must be the successive elements after
4505 unsigned ExpectedElt = Imm;
4506 for (unsigned i = 1; i < NumElts; ++i) {
4507 // Increment the expected index. If it wraps around, it may still be
4508 // a VEXT but the source vectors must be swapped.
4510 if (ExpectedElt == NumElts * 2) {
4515 if (M[i] < 0) continue; // ignore UNDEF indices
4516 if (ExpectedElt != static_cast<unsigned>(M[i]))
4520 // Adjust the index value if the source operands will be swapped.
4527 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4528 /// instruction with the specified blocksize. (The order of the elements
4529 /// within each block of the vector is reversed.)
4530 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4531 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4532 "Only possible block sizes for VREV are: 16, 32, 64");
4534 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4538 unsigned NumElts = VT.getVectorNumElements();
4539 unsigned BlockElts = M[0] + 1;
4540 // If the first shuffle index is UNDEF, be optimistic.
4542 BlockElts = BlockSize / EltSz;
4544 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4547 for (unsigned i = 0; i < NumElts; ++i) {
4548 if (M[i] < 0) continue; // ignore UNDEF indices
4549 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4556 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4557 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4558 // range, then 0 is placed into the resulting vector. So pretty much any mask
4559 // of 8 elements can work here.
4560 return VT == MVT::v8i8 && M.size() == 8;
4563 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4564 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4568 unsigned NumElts = VT.getVectorNumElements();
4569 WhichResult = (M[0] == 0 ? 0 : 1);
4570 for (unsigned i = 0; i < NumElts; i += 2) {
4571 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4572 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4578 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4579 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4580 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4581 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4582 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4586 unsigned NumElts = VT.getVectorNumElements();
4587 WhichResult = (M[0] == 0 ? 0 : 1);
4588 for (unsigned i = 0; i < NumElts; i += 2) {
4589 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4590 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4596 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4597 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4601 unsigned NumElts = VT.getVectorNumElements();
4602 WhichResult = (M[0] == 0 ? 0 : 1);
4603 for (unsigned i = 0; i != NumElts; ++i) {
4604 if (M[i] < 0) continue; // ignore UNDEF indices
4605 if ((unsigned) M[i] != 2 * i + WhichResult)
4609 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4610 if (VT.is64BitVector() && EltSz == 32)
4616 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4617 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4618 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4619 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4620 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4624 unsigned Half = VT.getVectorNumElements() / 2;
4625 WhichResult = (M[0] == 0 ? 0 : 1);
4626 for (unsigned j = 0; j != 2; ++j) {
4627 unsigned Idx = WhichResult;
4628 for (unsigned i = 0; i != Half; ++i) {
4629 int MIdx = M[i + j * Half];
4630 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4636 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4637 if (VT.is64BitVector() && EltSz == 32)
4643 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4644 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4648 unsigned NumElts = VT.getVectorNumElements();
4649 WhichResult = (M[0] == 0 ? 0 : 1);
4650 unsigned Idx = WhichResult * NumElts / 2;
4651 for (unsigned i = 0; i != NumElts; i += 2) {
4652 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4653 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4658 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4659 if (VT.is64BitVector() && EltSz == 32)
4665 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4666 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4667 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4668 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4669 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4673 unsigned NumElts = VT.getVectorNumElements();
4674 WhichResult = (M[0] == 0 ? 0 : 1);
4675 unsigned Idx = WhichResult * NumElts / 2;
4676 for (unsigned i = 0; i != NumElts; i += 2) {
4677 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4678 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4683 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4684 if (VT.is64BitVector() && EltSz == 32)
4690 /// \return true if this is a reverse operation on an vector.
4691 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4692 unsigned NumElts = VT.getVectorNumElements();
4693 // Make sure the mask has the right size.
4694 if (NumElts != M.size())
4697 // Look for <15, ..., 3, -1, 1, 0>.
4698 for (unsigned i = 0; i != NumElts; ++i)
4699 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4705 // If N is an integer constant that can be moved into a register in one
4706 // instruction, return an SDValue of such a constant (will become a MOV
4707 // instruction). Otherwise return null.
4708 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4709 const ARMSubtarget *ST, SDLoc dl) {
4711 if (!isa<ConstantSDNode>(N))
4713 Val = cast<ConstantSDNode>(N)->getZExtValue();
4715 if (ST->isThumb1Only()) {
4716 if (Val <= 255 || ~Val <= 255)
4717 return DAG.getConstant(Val, MVT::i32);
4719 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4720 return DAG.getConstant(Val, MVT::i32);
4725 // If this is a case we can't handle, return null and let the default
4726 // expansion code take care of it.
4727 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4728 const ARMSubtarget *ST) const {
4729 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4731 EVT VT = Op.getValueType();
4733 APInt SplatBits, SplatUndef;
4734 unsigned SplatBitSize;
4736 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4737 if (SplatBitSize <= 64) {
4738 // Check if an immediate VMOV works.
4740 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4741 SplatUndef.getZExtValue(), SplatBitSize,
4742 DAG, VmovVT, VT.is128BitVector(),
4744 if (Val.getNode()) {
4745 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4746 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4749 // Try an immediate VMVN.
4750 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4751 Val = isNEONModifiedImm(NegatedImm,
4752 SplatUndef.getZExtValue(), SplatBitSize,
4753 DAG, VmovVT, VT.is128BitVector(),
4755 if (Val.getNode()) {
4756 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4757 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4760 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4761 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4762 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4764 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4765 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4771 // Scan through the operands to see if only one value is used.
4773 // As an optimisation, even if more than one value is used it may be more
4774 // profitable to splat with one value then change some lanes.
4776 // Heuristically we decide to do this if the vector has a "dominant" value,
4777 // defined as splatted to more than half of the lanes.
4778 unsigned NumElts = VT.getVectorNumElements();
4779 bool isOnlyLowElement = true;
4780 bool usesOnlyOneValue = true;
4781 bool hasDominantValue = false;
4782 bool isConstant = true;
4784 // Map of the number of times a particular SDValue appears in the
4786 DenseMap<SDValue, unsigned> ValueCounts;
4788 for (unsigned i = 0; i < NumElts; ++i) {
4789 SDValue V = Op.getOperand(i);
4790 if (V.getOpcode() == ISD::UNDEF)
4793 isOnlyLowElement = false;
4794 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4797 ValueCounts.insert(std::make_pair(V, 0));
4798 unsigned &Count = ValueCounts[V];
4800 // Is this value dominant? (takes up more than half of the lanes)
4801 if (++Count > (NumElts / 2)) {
4802 hasDominantValue = true;
4806 if (ValueCounts.size() != 1)
4807 usesOnlyOneValue = false;
4808 if (!Value.getNode() && ValueCounts.size() > 0)
4809 Value = ValueCounts.begin()->first;
4811 if (ValueCounts.size() == 0)
4812 return DAG.getUNDEF(VT);
4814 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4815 // Keep going if we are hitting this case.
4816 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4817 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4819 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4821 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4822 // i32 and try again.
4823 if (hasDominantValue && EltSize <= 32) {
4827 // If we are VDUPing a value that comes directly from a vector, that will
4828 // cause an unnecessary move to and from a GPR, where instead we could
4829 // just use VDUPLANE. We can only do this if the lane being extracted
4830 // is at a constant index, as the VDUP from lane instructions only have
4831 // constant-index forms.
4832 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4833 isa<ConstantSDNode>(Value->getOperand(1))) {
4834 // We need to create a new undef vector to use for the VDUPLANE if the
4835 // size of the vector from which we get the value is different than the
4836 // size of the vector that we need to create. We will insert the element
4837 // such that the register coalescer will remove unnecessary copies.
4838 if (VT != Value->getOperand(0).getValueType()) {
4839 ConstantSDNode *constIndex;
4840 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4841 assert(constIndex && "The index is not a constant!");
4842 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4843 VT.getVectorNumElements();
4844 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4845 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4846 Value, DAG.getConstant(index, MVT::i32)),
4847 DAG.getConstant(index, MVT::i32));
4849 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4850 Value->getOperand(0), Value->getOperand(1));
4852 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4854 if (!usesOnlyOneValue) {
4855 // The dominant value was splatted as 'N', but we now have to insert
4856 // all differing elements.
4857 for (unsigned I = 0; I < NumElts; ++I) {
4858 if (Op.getOperand(I) == Value)
4860 SmallVector<SDValue, 3> Ops;
4862 Ops.push_back(Op.getOperand(I));
4863 Ops.push_back(DAG.getConstant(I, MVT::i32));
4864 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4869 if (VT.getVectorElementType().isFloatingPoint()) {
4870 SmallVector<SDValue, 8> Ops;
4871 for (unsigned i = 0; i < NumElts; ++i)
4872 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4874 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4875 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4876 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4878 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4880 if (usesOnlyOneValue) {
4881 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4882 if (isConstant && Val.getNode())
4883 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4887 // If all elements are constants and the case above didn't get hit, fall back
4888 // to the default expansion, which will generate a load from the constant
4893 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4895 SDValue shuffle = ReconstructShuffle(Op, DAG);
4896 if (shuffle != SDValue())
4900 // Vectors with 32- or 64-bit elements can be built by directly assigning
4901 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4902 // will be legalized.
4903 if (EltSize >= 32) {
4904 // Do the expansion with floating-point types, since that is what the VFP
4905 // registers are defined to use, and since i64 is not legal.
4906 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4907 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4908 SmallVector<SDValue, 8> Ops;
4909 for (unsigned i = 0; i < NumElts; ++i)
4910 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4911 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4912 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4915 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4916 // know the default expansion would otherwise fall back on something even
4917 // worse. For a vector with one or two non-undef values, that's
4918 // scalar_to_vector for the elements followed by a shuffle (provided the
4919 // shuffle is valid for the target) and materialization element by element
4920 // on the stack followed by a load for everything else.
4921 if (!isConstant && !usesOnlyOneValue) {
4922 SDValue Vec = DAG.getUNDEF(VT);
4923 for (unsigned i = 0 ; i < NumElts; ++i) {
4924 SDValue V = Op.getOperand(i);
4925 if (V.getOpcode() == ISD::UNDEF)
4927 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4928 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4936 // Gather data to see if the operation can be modelled as a
4937 // shuffle in combination with VEXTs.
4938 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4939 SelectionDAG &DAG) const {
4941 EVT VT = Op.getValueType();
4942 unsigned NumElts = VT.getVectorNumElements();
4944 SmallVector<SDValue, 2> SourceVecs;
4945 SmallVector<unsigned, 2> MinElts;
4946 SmallVector<unsigned, 2> MaxElts;
4948 for (unsigned i = 0; i < NumElts; ++i) {
4949 SDValue V = Op.getOperand(i);
4950 if (V.getOpcode() == ISD::UNDEF)
4952 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4953 // A shuffle can only come from building a vector from various
4954 // elements of other vectors.
4956 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4957 VT.getVectorElementType()) {
4958 // This code doesn't know how to handle shuffles where the vector
4959 // element types do not match (this happens because type legalization
4960 // promotes the return type of EXTRACT_VECTOR_ELT).
4961 // FIXME: It might be appropriate to extend this code to handle
4962 // mismatched types.
4966 // Record this extraction against the appropriate vector if possible...
4967 SDValue SourceVec = V.getOperand(0);
4968 // If the element number isn't a constant, we can't effectively
4969 // analyze what's going on.
4970 if (!isa<ConstantSDNode>(V.getOperand(1)))
4972 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4973 bool FoundSource = false;
4974 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4975 if (SourceVecs[j] == SourceVec) {
4976 if (MinElts[j] > EltNo)
4978 if (MaxElts[j] < EltNo)
4985 // Or record a new source if not...
4987 SourceVecs.push_back(SourceVec);
4988 MinElts.push_back(EltNo);
4989 MaxElts.push_back(EltNo);
4993 // Currently only do something sane when at most two source vectors
4995 if (SourceVecs.size() > 2)
4998 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4999 int VEXTOffsets[2] = {0, 0};
5001 // This loop extracts the usage patterns of the source vectors
5002 // and prepares appropriate SDValues for a shuffle if possible.
5003 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5004 if (SourceVecs[i].getValueType() == VT) {
5005 // No VEXT necessary
5006 ShuffleSrcs[i] = SourceVecs[i];
5009 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5010 // It probably isn't worth padding out a smaller vector just to
5011 // break it down again in a shuffle.
5015 // Since only 64-bit and 128-bit vectors are legal on ARM and
5016 // we've eliminated the other cases...
5017 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5018 "unexpected vector sizes in ReconstructShuffle");
5020 if (MaxElts[i] - MinElts[i] >= NumElts) {
5021 // Span too large for a VEXT to cope
5025 if (MinElts[i] >= NumElts) {
5026 // The extraction can just take the second half
5027 VEXTOffsets[i] = NumElts;
5028 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5030 DAG.getIntPtrConstant(NumElts));
5031 } else if (MaxElts[i] < NumElts) {
5032 // The extraction can just take the first half
5034 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5036 DAG.getIntPtrConstant(0));
5038 // An actual VEXT is needed
5039 VEXTOffsets[i] = MinElts[i];
5040 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5042 DAG.getIntPtrConstant(0));
5043 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5045 DAG.getIntPtrConstant(NumElts));
5046 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5047 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5051 SmallVector<int, 8> Mask;
5053 for (unsigned i = 0; i < NumElts; ++i) {
5054 SDValue Entry = Op.getOperand(i);
5055 if (Entry.getOpcode() == ISD::UNDEF) {
5060 SDValue ExtractVec = Entry.getOperand(0);
5061 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5062 .getOperand(1))->getSExtValue();
5063 if (ExtractVec == SourceVecs[0]) {
5064 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5066 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5070 // Final check before we try to produce nonsense...
5071 if (isShuffleMaskLegal(Mask, VT))
5072 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5078 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5079 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5080 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5081 /// are assumed to be legal.
5083 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5085 if (VT.getVectorNumElements() == 4 &&
5086 (VT.is128BitVector() || VT.is64BitVector())) {
5087 unsigned PFIndexes[4];
5088 for (unsigned i = 0; i != 4; ++i) {
5092 PFIndexes[i] = M[i];
5095 // Compute the index in the perfect shuffle table.
5096 unsigned PFTableIndex =
5097 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5098 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5099 unsigned Cost = (PFEntry >> 30);
5106 unsigned Imm, WhichResult;
5108 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5109 return (EltSize >= 32 ||
5110 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5111 isVREVMask(M, VT, 64) ||
5112 isVREVMask(M, VT, 32) ||
5113 isVREVMask(M, VT, 16) ||
5114 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5115 isVTBLMask(M, VT) ||
5116 isVTRNMask(M, VT, WhichResult) ||
5117 isVUZPMask(M, VT, WhichResult) ||
5118 isVZIPMask(M, VT, WhichResult) ||
5119 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5120 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5121 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5122 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5125 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5126 /// the specified operations to build the shuffle.
5127 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5128 SDValue RHS, SelectionDAG &DAG,
5130 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5131 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5132 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5135 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5144 OP_VUZPL, // VUZP, left result
5145 OP_VUZPR, // VUZP, right result
5146 OP_VZIPL, // VZIP, left result
5147 OP_VZIPR, // VZIP, right result
5148 OP_VTRNL, // VTRN, left result
5149 OP_VTRNR // VTRN, right result
5152 if (OpNum == OP_COPY) {
5153 if (LHSID == (1*9+2)*9+3) return LHS;
5154 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5158 SDValue OpLHS, OpRHS;
5159 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5160 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5161 EVT VT = OpLHS.getValueType();
5164 default: llvm_unreachable("Unknown shuffle opcode!");
5166 // VREV divides the vector in half and swaps within the half.
5167 if (VT.getVectorElementType() == MVT::i32 ||
5168 VT.getVectorElementType() == MVT::f32)
5169 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5170 // vrev <4 x i16> -> VREV32
5171 if (VT.getVectorElementType() == MVT::i16)
5172 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5173 // vrev <4 x i8> -> VREV16
5174 assert(VT.getVectorElementType() == MVT::i8);
5175 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5180 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5181 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5185 return DAG.getNode(ARMISD::VEXT, dl, VT,
5187 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5190 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5191 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5194 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5195 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5198 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5199 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5203 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5204 ArrayRef<int> ShuffleMask,
5205 SelectionDAG &DAG) {
5206 // Check to see if we can use the VTBL instruction.
5207 SDValue V1 = Op.getOperand(0);
5208 SDValue V2 = Op.getOperand(1);
5211 SmallVector<SDValue, 8> VTBLMask;
5212 for (ArrayRef<int>::iterator
5213 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5214 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5216 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5217 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5218 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5221 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5222 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5226 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5227 SelectionDAG &DAG) {
5229 SDValue OpLHS = Op.getOperand(0);
5230 EVT VT = OpLHS.getValueType();
5232 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5233 "Expect an v8i16/v16i8 type");
5234 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5235 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5236 // extract the first 8 bytes into the top double word and the last 8 bytes
5237 // into the bottom double word. The v8i16 case is similar.
5238 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5239 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5240 DAG.getConstant(ExtractNum, MVT::i32));
5243 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5244 SDValue V1 = Op.getOperand(0);
5245 SDValue V2 = Op.getOperand(1);
5247 EVT VT = Op.getValueType();
5248 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5250 // Convert shuffles that are directly supported on NEON to target-specific
5251 // DAG nodes, instead of keeping them as shuffles and matching them again
5252 // during code selection. This is more efficient and avoids the possibility
5253 // of inconsistencies between legalization and selection.
5254 // FIXME: floating-point vectors should be canonicalized to integer vectors
5255 // of the same time so that they get CSEd properly.
5256 ArrayRef<int> ShuffleMask = SVN->getMask();
5258 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5259 if (EltSize <= 32) {
5260 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5261 int Lane = SVN->getSplatIndex();
5262 // If this is undef splat, generate it via "just" vdup, if possible.
5263 if (Lane == -1) Lane = 0;
5265 // Test if V1 is a SCALAR_TO_VECTOR.
5266 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5267 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5269 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5270 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5272 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5273 !isa<ConstantSDNode>(V1.getOperand(0))) {
5274 bool IsScalarToVector = true;
5275 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5276 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5277 IsScalarToVector = false;
5280 if (IsScalarToVector)
5281 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5283 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5284 DAG.getConstant(Lane, MVT::i32));
5289 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5292 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5293 DAG.getConstant(Imm, MVT::i32));
5296 if (isVREVMask(ShuffleMask, VT, 64))
5297 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5298 if (isVREVMask(ShuffleMask, VT, 32))
5299 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5300 if (isVREVMask(ShuffleMask, VT, 16))
5301 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5303 if (V2->getOpcode() == ISD::UNDEF &&
5304 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5305 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5306 DAG.getConstant(Imm, MVT::i32));
5309 // Check for Neon shuffles that modify both input vectors in place.
5310 // If both results are used, i.e., if there are two shuffles with the same
5311 // source operands and with masks corresponding to both results of one of
5312 // these operations, DAG memoization will ensure that a single node is
5313 // used for both shuffles.
5314 unsigned WhichResult;
5315 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5316 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5317 V1, V2).getValue(WhichResult);
5318 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5319 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5320 V1, V2).getValue(WhichResult);
5321 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5322 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5323 V1, V2).getValue(WhichResult);
5325 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5326 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5327 V1, V1).getValue(WhichResult);
5328 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5329 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5330 V1, V1).getValue(WhichResult);
5331 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5332 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5333 V1, V1).getValue(WhichResult);
5336 // If the shuffle is not directly supported and it has 4 elements, use
5337 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5338 unsigned NumElts = VT.getVectorNumElements();
5340 unsigned PFIndexes[4];
5341 for (unsigned i = 0; i != 4; ++i) {
5342 if (ShuffleMask[i] < 0)
5345 PFIndexes[i] = ShuffleMask[i];
5348 // Compute the index in the perfect shuffle table.
5349 unsigned PFTableIndex =
5350 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5351 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5352 unsigned Cost = (PFEntry >> 30);
5355 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5358 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5359 if (EltSize >= 32) {
5360 // Do the expansion with floating-point types, since that is what the VFP
5361 // registers are defined to use, and since i64 is not legal.
5362 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5363 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5364 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5365 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5366 SmallVector<SDValue, 8> Ops;
5367 for (unsigned i = 0; i < NumElts; ++i) {
5368 if (ShuffleMask[i] < 0)
5369 Ops.push_back(DAG.getUNDEF(EltVT));
5371 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5372 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5373 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5376 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5377 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5380 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5381 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5383 if (VT == MVT::v8i8) {
5384 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5385 if (NewOp.getNode())
5392 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5393 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5394 SDValue Lane = Op.getOperand(2);
5395 if (!isa<ConstantSDNode>(Lane))
5401 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5402 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5403 SDValue Lane = Op.getOperand(1);
5404 if (!isa<ConstantSDNode>(Lane))
5407 SDValue Vec = Op.getOperand(0);
5408 if (Op.getValueType() == MVT::i32 &&
5409 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5411 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5417 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5418 // The only time a CONCAT_VECTORS operation can have legal types is when
5419 // two 64-bit vectors are concatenated to a 128-bit vector.
5420 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5421 "unexpected CONCAT_VECTORS");
5423 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5424 SDValue Op0 = Op.getOperand(0);
5425 SDValue Op1 = Op.getOperand(1);
5426 if (Op0.getOpcode() != ISD::UNDEF)
5427 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5428 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5429 DAG.getIntPtrConstant(0));
5430 if (Op1.getOpcode() != ISD::UNDEF)
5431 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5432 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5433 DAG.getIntPtrConstant(1));
5434 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5437 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5438 /// element has been zero/sign-extended, depending on the isSigned parameter,
5439 /// from an integer type half its size.
5440 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5442 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5443 EVT VT = N->getValueType(0);
5444 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5445 SDNode *BVN = N->getOperand(0).getNode();
5446 if (BVN->getValueType(0) != MVT::v4i32 ||
5447 BVN->getOpcode() != ISD::BUILD_VECTOR)
5449 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5450 unsigned HiElt = 1 - LoElt;
5451 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5452 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5453 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5454 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5455 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5458 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5459 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5462 if (Hi0->isNullValue() && Hi1->isNullValue())
5468 if (N->getOpcode() != ISD::BUILD_VECTOR)
5471 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5472 SDNode *Elt = N->getOperand(i).getNode();
5473 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5474 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5475 unsigned HalfSize = EltSize / 2;
5477 if (!isIntN(HalfSize, C->getSExtValue()))
5480 if (!isUIntN(HalfSize, C->getZExtValue()))
5491 /// isSignExtended - Check if a node is a vector value that is sign-extended
5492 /// or a constant BUILD_VECTOR with sign-extended elements.
5493 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5494 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5496 if (isExtendedBUILD_VECTOR(N, DAG, true))
5501 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5502 /// or a constant BUILD_VECTOR with zero-extended elements.
5503 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5504 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5506 if (isExtendedBUILD_VECTOR(N, DAG, false))
5511 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5512 if (OrigVT.getSizeInBits() >= 64)
5515 assert(OrigVT.isSimple() && "Expecting a simple value type");
5517 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5518 switch (OrigSimpleTy) {
5519 default: llvm_unreachable("Unexpected Vector Type");
5528 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5529 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5530 /// We insert the required extension here to get the vector to fill a D register.
5531 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5534 unsigned ExtOpcode) {
5535 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5536 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5537 // 64-bits we need to insert a new extension so that it will be 64-bits.
5538 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5539 if (OrigTy.getSizeInBits() >= 64)
5542 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5543 EVT NewVT = getExtensionTo64Bits(OrigTy);
5545 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5548 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5549 /// does not do any sign/zero extension. If the original vector is less
5550 /// than 64 bits, an appropriate extension will be added after the load to
5551 /// reach a total size of 64 bits. We have to add the extension separately
5552 /// because ARM does not have a sign/zero extending load for vectors.
5553 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5554 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5556 // The load already has the right type.
5557 if (ExtendedTy == LD->getMemoryVT())
5558 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5559 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5560 LD->isNonTemporal(), LD->isInvariant(),
5561 LD->getAlignment());
5563 // We need to create a zextload/sextload. We cannot just create a load
5564 // followed by a zext/zext node because LowerMUL is also run during normal
5565 // operation legalization where we can't create illegal types.
5566 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5567 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5568 LD->getMemoryVT(), LD->isVolatile(),
5569 LD->isNonTemporal(), LD->getAlignment());
5572 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5573 /// extending load, or BUILD_VECTOR with extended elements, return the
5574 /// unextended value. The unextended vector should be 64 bits so that it can
5575 /// be used as an operand to a VMULL instruction. If the original vector size
5576 /// before extension is less than 64 bits we add a an extension to resize
5577 /// the vector to 64 bits.
5578 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5579 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5580 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5581 N->getOperand(0)->getValueType(0),
5585 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5586 return SkipLoadExtensionForVMULL(LD, DAG);
5588 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5589 // have been legalized as a BITCAST from v4i32.
5590 if (N->getOpcode() == ISD::BITCAST) {
5591 SDNode *BVN = N->getOperand(0).getNode();
5592 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5593 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5594 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5595 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5596 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5598 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5599 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5600 EVT VT = N->getValueType(0);
5601 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5602 unsigned NumElts = VT.getVectorNumElements();
5603 MVT TruncVT = MVT::getIntegerVT(EltSize);
5604 SmallVector<SDValue, 8> Ops;
5605 for (unsigned i = 0; i != NumElts; ++i) {
5606 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5607 const APInt &CInt = C->getAPIntValue();
5608 // Element types smaller than 32 bits are not legal, so use i32 elements.
5609 // The values are implicitly truncated so sext vs. zext doesn't matter.
5610 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5612 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5613 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5616 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5617 unsigned Opcode = N->getOpcode();
5618 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5619 SDNode *N0 = N->getOperand(0).getNode();
5620 SDNode *N1 = N->getOperand(1).getNode();
5621 return N0->hasOneUse() && N1->hasOneUse() &&
5622 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5627 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5628 unsigned Opcode = N->getOpcode();
5629 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5630 SDNode *N0 = N->getOperand(0).getNode();
5631 SDNode *N1 = N->getOperand(1).getNode();
5632 return N0->hasOneUse() && N1->hasOneUse() &&
5633 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5638 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5639 // Multiplications are only custom-lowered for 128-bit vectors so that
5640 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5641 EVT VT = Op.getValueType();
5642 assert(VT.is128BitVector() && VT.isInteger() &&
5643 "unexpected type for custom-lowering ISD::MUL");
5644 SDNode *N0 = Op.getOperand(0).getNode();
5645 SDNode *N1 = Op.getOperand(1).getNode();
5646 unsigned NewOpc = 0;
5648 bool isN0SExt = isSignExtended(N0, DAG);
5649 bool isN1SExt = isSignExtended(N1, DAG);
5650 if (isN0SExt && isN1SExt)
5651 NewOpc = ARMISD::VMULLs;
5653 bool isN0ZExt = isZeroExtended(N0, DAG);
5654 bool isN1ZExt = isZeroExtended(N1, DAG);
5655 if (isN0ZExt && isN1ZExt)
5656 NewOpc = ARMISD::VMULLu;
5657 else if (isN1SExt || isN1ZExt) {
5658 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5659 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5660 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5661 NewOpc = ARMISD::VMULLs;
5663 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5664 NewOpc = ARMISD::VMULLu;
5666 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5668 NewOpc = ARMISD::VMULLu;
5674 if (VT == MVT::v2i64)
5675 // Fall through to expand this. It is not legal.
5678 // Other vector multiplications are legal.
5683 // Legalize to a VMULL instruction.
5686 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5688 Op0 = SkipExtensionForVMULL(N0, DAG);
5689 assert(Op0.getValueType().is64BitVector() &&
5690 Op1.getValueType().is64BitVector() &&
5691 "unexpected types for extended operands to VMULL");
5692 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5695 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5696 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5703 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5704 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5705 EVT Op1VT = Op1.getValueType();
5706 return DAG.getNode(N0->getOpcode(), DL, VT,
5707 DAG.getNode(NewOpc, DL, VT,
5708 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5709 DAG.getNode(NewOpc, DL, VT,
5710 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5714 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5716 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5717 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5718 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5719 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5720 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5721 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5722 // Get reciprocal estimate.
5723 // float4 recip = vrecpeq_f32(yf);
5724 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5725 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5726 // Because char has a smaller range than uchar, we can actually get away
5727 // without any newton steps. This requires that we use a weird bias
5728 // of 0xb000, however (again, this has been exhaustively tested).
5729 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5730 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5731 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5732 Y = DAG.getConstant(0xb000, MVT::i32);
5733 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5734 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5735 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5736 // Convert back to short.
5737 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5738 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5743 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5745 // Convert to float.
5746 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5747 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5748 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5749 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5750 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5751 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5753 // Use reciprocal estimate and one refinement step.
5754 // float4 recip = vrecpeq_f32(yf);
5755 // recip *= vrecpsq_f32(yf, recip);
5756 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5757 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5758 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5759 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5761 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5762 // Because short has a smaller range than ushort, we can actually get away
5763 // with only a single newton step. This requires that we use a weird bias
5764 // of 89, however (again, this has been exhaustively tested).
5765 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5766 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5767 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5768 N1 = DAG.getConstant(0x89, MVT::i32);
5769 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5770 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5771 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5772 // Convert back to integer and return.
5773 // return vmovn_s32(vcvt_s32_f32(result));
5774 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5775 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5779 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5780 EVT VT = Op.getValueType();
5781 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5782 "unexpected type for custom-lowering ISD::SDIV");
5785 SDValue N0 = Op.getOperand(0);
5786 SDValue N1 = Op.getOperand(1);
5789 if (VT == MVT::v8i8) {
5790 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5791 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5793 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5794 DAG.getIntPtrConstant(4));
5795 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5796 DAG.getIntPtrConstant(4));
5797 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5798 DAG.getIntPtrConstant(0));
5799 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5800 DAG.getIntPtrConstant(0));
5802 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5803 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5805 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5806 N0 = LowerCONCAT_VECTORS(N0, DAG);
5808 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5811 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5814 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5815 EVT VT = Op.getValueType();
5816 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5817 "unexpected type for custom-lowering ISD::UDIV");
5820 SDValue N0 = Op.getOperand(0);
5821 SDValue N1 = Op.getOperand(1);
5824 if (VT == MVT::v8i8) {
5825 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5826 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5828 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5829 DAG.getIntPtrConstant(4));
5830 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5831 DAG.getIntPtrConstant(4));
5832 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5833 DAG.getIntPtrConstant(0));
5834 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5835 DAG.getIntPtrConstant(0));
5837 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5838 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5840 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5841 N0 = LowerCONCAT_VECTORS(N0, DAG);
5843 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5844 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5849 // v4i16 sdiv ... Convert to float.
5850 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5851 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5852 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5853 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5854 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5855 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5857 // Use reciprocal estimate and two refinement steps.
5858 // float4 recip = vrecpeq_f32(yf);
5859 // recip *= vrecpsq_f32(yf, recip);
5860 // recip *= vrecpsq_f32(yf, recip);
5861 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5862 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5863 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5864 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5866 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5867 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5868 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5870 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5871 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5872 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5873 // and that it will never cause us to return an answer too large).
5874 // float4 result = as_float4(as_int4(xf*recip) + 2);
5875 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5876 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5877 N1 = DAG.getConstant(2, MVT::i32);
5878 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5879 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5880 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5881 // Convert back to integer and return.
5882 // return vmovn_u32(vcvt_s32_f32(result));
5883 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5884 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5888 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5889 EVT VT = Op.getNode()->getValueType(0);
5890 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5893 bool ExtraOp = false;
5894 switch (Op.getOpcode()) {
5895 default: llvm_unreachable("Invalid code");
5896 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5897 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5898 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5899 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5903 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5905 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5906 Op.getOperand(1), Op.getOperand(2));
5909 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
5910 assert(Subtarget->isTargetDarwin());
5912 // For iOS, we want to call an alternative entry point: __sincos_stret,
5913 // return values are passed via sret.
5915 SDValue Arg = Op.getOperand(0);
5916 EVT ArgVT = Arg.getValueType();
5917 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
5919 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5920 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5922 // Pair of floats / doubles used to pass the result.
5923 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
5925 // Create stack object for sret.
5926 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
5927 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
5928 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
5929 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
5935 Entry.Ty = RetTy->getPointerTo();
5936 Entry.isSExt = false;
5937 Entry.isZExt = false;
5938 Entry.isSRet = true;
5939 Args.push_back(Entry);
5943 Entry.isSExt = false;
5944 Entry.isZExt = false;
5945 Args.push_back(Entry);
5947 const char *LibcallName = (ArgVT == MVT::f64)
5948 ? "__sincos_stret" : "__sincosf_stret";
5949 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
5952 CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
5953 false, false, false, false, 0,
5954 CallingConv::C, /*isTaillCall=*/false,
5955 /*doesNotRet=*/false, /*isReturnValueUsed*/false,
5956 Callee, Args, DAG, dl);
5957 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
5959 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
5960 MachinePointerInfo(), false, false, false, 0);
5962 // Address of cos field.
5963 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
5964 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
5965 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
5966 MachinePointerInfo(), false, false, false, 0);
5968 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
5969 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
5970 LoadSin.getValue(0), LoadCos.getValue(0));
5973 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5974 // Monotonic load/store is legal for all targets
5975 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5978 // Acquire/Release load/store is not legal for targets without a
5979 // dmb or equivalent available.
5983 static void ReplaceREADCYCLECOUNTER(SDNode *N,
5984 SmallVectorImpl<SDValue> &Results,
5986 const ARMSubtarget *Subtarget) {
5988 SDValue Cycles32, OutChain;
5990 if (Subtarget->hasPerfMon()) {
5991 // Under Power Management extensions, the cycle-count is:
5992 // mrc p15, #0, <Rt>, c9, c13, #0
5993 SDValue Ops[] = { N->getOperand(0), // Chain
5994 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
5995 DAG.getConstant(15, MVT::i32),
5996 DAG.getConstant(0, MVT::i32),
5997 DAG.getConstant(9, MVT::i32),
5998 DAG.getConstant(13, MVT::i32),
5999 DAG.getConstant(0, MVT::i32)
6002 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6003 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6004 array_lengthof(Ops));
6005 OutChain = Cycles32.getValue(1);
6007 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6008 // there are older ARM CPUs that have implementation-specific ways of
6009 // obtaining this information (FIXME!).
6010 Cycles32 = DAG.getConstant(0, MVT::i32);
6011 OutChain = DAG.getEntryNode();
6015 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6016 Cycles32, DAG.getConstant(0, MVT::i32));
6017 Results.push_back(Cycles64);
6018 Results.push_back(OutChain);
6021 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6022 switch (Op.getOpcode()) {
6023 default: llvm_unreachable("Don't know how to custom lower this!");
6024 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6025 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6026 case ISD::GlobalAddress:
6027 return Subtarget->isTargetMachO() ? LowerGlobalAddressDarwin(Op, DAG) :
6028 LowerGlobalAddressELF(Op, DAG);
6029 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6030 case ISD::SELECT: return LowerSELECT(Op, DAG);
6031 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6032 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6033 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6034 case ISD::VASTART: return LowerVASTART(Op, DAG);
6035 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6036 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6037 case ISD::SINT_TO_FP:
6038 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6039 case ISD::FP_TO_SINT:
6040 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6041 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6042 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6043 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6044 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6045 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6046 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6047 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6049 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6052 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6053 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6054 case ISD::SRL_PARTS:
6055 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6056 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6057 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6058 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6059 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6060 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6061 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6062 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6063 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6064 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6065 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6066 case ISD::MUL: return LowerMUL(Op, DAG);
6067 case ISD::SDIV: return LowerSDIV(Op, DAG);
6068 case ISD::UDIV: return LowerUDIV(Op, DAG);
6072 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6073 case ISD::ATOMIC_LOAD:
6074 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6075 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6077 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6081 /// ReplaceNodeResults - Replace the results of node with an illegal result
6082 /// type with new values built out of custom code.
6083 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6084 SmallVectorImpl<SDValue>&Results,
6085 SelectionDAG &DAG) const {
6087 switch (N->getOpcode()) {
6089 llvm_unreachable("Don't know how to custom expand this!");
6091 Res = ExpandBITCAST(N, DAG);
6095 Res = Expand64BitShift(N, DAG, Subtarget);
6097 case ISD::READCYCLECOUNTER:
6098 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6102 Results.push_back(Res);
6105 //===----------------------------------------------------------------------===//
6106 // ARM Scheduler Hooks
6107 //===----------------------------------------------------------------------===//
6109 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6110 /// registers the function context.
6111 void ARMTargetLowering::
6112 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6113 MachineBasicBlock *DispatchBB, int FI) const {
6114 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6115 DebugLoc dl = MI->getDebugLoc();
6116 MachineFunction *MF = MBB->getParent();
6117 MachineRegisterInfo *MRI = &MF->getRegInfo();
6118 MachineConstantPool *MCP = MF->getConstantPool();
6119 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6120 const Function *F = MF->getFunction();
6122 bool isThumb = Subtarget->isThumb();
6123 bool isThumb2 = Subtarget->isThumb2();
6125 unsigned PCLabelId = AFI->createPICLabelUId();
6126 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6127 ARMConstantPoolValue *CPV =
6128 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6129 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6131 const TargetRegisterClass *TRC = isThumb ?
6132 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6133 (const TargetRegisterClass*)&ARM::GPRRegClass;
6135 // Grab constant pool and fixed stack memory operands.
6136 MachineMemOperand *CPMMO =
6137 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6138 MachineMemOperand::MOLoad, 4, 4);
6140 MachineMemOperand *FIMMOSt =
6141 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6142 MachineMemOperand::MOStore, 4, 4);
6144 // Load the address of the dispatch MBB into the jump buffer.
6146 // Incoming value: jbuf
6147 // ldr.n r5, LCPI1_1
6150 // str r5, [$jbuf, #+4] ; &jbuf[1]
6151 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6152 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6153 .addConstantPoolIndex(CPI)
6154 .addMemOperand(CPMMO));
6155 // Set the low bit because of thumb mode.
6156 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6158 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6159 .addReg(NewVReg1, RegState::Kill)
6161 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6162 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6163 .addReg(NewVReg2, RegState::Kill)
6165 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6166 .addReg(NewVReg3, RegState::Kill)
6168 .addImm(36) // &jbuf[1] :: pc
6169 .addMemOperand(FIMMOSt));
6170 } else if (isThumb) {
6171 // Incoming value: jbuf
6172 // ldr.n r1, LCPI1_4
6176 // add r2, $jbuf, #+4 ; &jbuf[1]
6178 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6179 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6180 .addConstantPoolIndex(CPI)
6181 .addMemOperand(CPMMO));
6182 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6183 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6184 .addReg(NewVReg1, RegState::Kill)
6186 // Set the low bit because of thumb mode.
6187 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6188 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6189 .addReg(ARM::CPSR, RegState::Define)
6191 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6192 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6193 .addReg(ARM::CPSR, RegState::Define)
6194 .addReg(NewVReg2, RegState::Kill)
6195 .addReg(NewVReg3, RegState::Kill));
6196 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6197 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6199 .addImm(36)); // &jbuf[1] :: pc
6200 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6201 .addReg(NewVReg4, RegState::Kill)
6202 .addReg(NewVReg5, RegState::Kill)
6204 .addMemOperand(FIMMOSt));
6206 // Incoming value: jbuf
6209 // str r1, [$jbuf, #+4] ; &jbuf[1]
6210 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6211 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6212 .addConstantPoolIndex(CPI)
6214 .addMemOperand(CPMMO));
6215 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6216 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6217 .addReg(NewVReg1, RegState::Kill)
6218 .addImm(PCLabelId));
6219 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6220 .addReg(NewVReg2, RegState::Kill)
6222 .addImm(36) // &jbuf[1] :: pc
6223 .addMemOperand(FIMMOSt));
6227 MachineBasicBlock *ARMTargetLowering::
6228 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6229 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6230 DebugLoc dl = MI->getDebugLoc();
6231 MachineFunction *MF = MBB->getParent();
6232 MachineRegisterInfo *MRI = &MF->getRegInfo();
6233 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6234 MachineFrameInfo *MFI = MF->getFrameInfo();
6235 int FI = MFI->getFunctionContextIndex();
6237 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6238 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6239 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6241 // Get a mapping of the call site numbers to all of the landing pads they're
6243 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6244 unsigned MaxCSNum = 0;
6245 MachineModuleInfo &MMI = MF->getMMI();
6246 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6248 if (!BB->isLandingPad()) continue;
6250 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6252 for (MachineBasicBlock::iterator
6253 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6254 if (!II->isEHLabel()) continue;
6256 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6257 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6259 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6260 for (SmallVectorImpl<unsigned>::iterator
6261 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6262 CSI != CSE; ++CSI) {
6263 CallSiteNumToLPad[*CSI].push_back(BB);
6264 MaxCSNum = std::max(MaxCSNum, *CSI);
6270 // Get an ordered list of the machine basic blocks for the jump table.
6271 std::vector<MachineBasicBlock*> LPadList;
6272 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6273 LPadList.reserve(CallSiteNumToLPad.size());
6274 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6275 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6276 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6277 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6278 LPadList.push_back(*II);
6279 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6283 assert(!LPadList.empty() &&
6284 "No landing pad destinations for the dispatch jump table!");
6286 // Create the jump table and associated information.
6287 MachineJumpTableInfo *JTI =
6288 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6289 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6290 unsigned UId = AFI->createJumpTableUId();
6291 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6293 // Create the MBBs for the dispatch code.
6295 // Shove the dispatch's address into the return slot in the function context.
6296 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6297 DispatchBB->setIsLandingPad();
6299 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6300 unsigned trap_opcode;
6301 if (Subtarget->isThumb())
6302 trap_opcode = ARM::tTRAP;
6304 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6306 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6307 DispatchBB->addSuccessor(TrapBB);
6309 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6310 DispatchBB->addSuccessor(DispContBB);
6313 MF->insert(MF->end(), DispatchBB);
6314 MF->insert(MF->end(), DispContBB);
6315 MF->insert(MF->end(), TrapBB);
6317 // Insert code into the entry block that creates and registers the function
6319 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6321 MachineMemOperand *FIMMOLd =
6322 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6323 MachineMemOperand::MOLoad |
6324 MachineMemOperand::MOVolatile, 4, 4);
6326 MachineInstrBuilder MIB;
6327 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6329 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6330 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6332 // Add a register mask with no preserved registers. This results in all
6333 // registers being marked as clobbered.
6334 MIB.addRegMask(RI.getNoPreservedMask());
6336 unsigned NumLPads = LPadList.size();
6337 if (Subtarget->isThumb2()) {
6338 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6339 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6342 .addMemOperand(FIMMOLd));
6344 if (NumLPads < 256) {
6345 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6347 .addImm(LPadList.size()));
6349 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6350 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6351 .addImm(NumLPads & 0xFFFF));
6353 unsigned VReg2 = VReg1;
6354 if ((NumLPads & 0xFFFF0000) != 0) {
6355 VReg2 = MRI->createVirtualRegister(TRC);
6356 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6358 .addImm(NumLPads >> 16));
6361 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6366 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6371 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6372 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6373 .addJumpTableIndex(MJTI)
6376 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6379 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6380 .addReg(NewVReg3, RegState::Kill)
6382 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6384 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6385 .addReg(NewVReg4, RegState::Kill)
6387 .addJumpTableIndex(MJTI)
6389 } else if (Subtarget->isThumb()) {
6390 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6391 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6394 .addMemOperand(FIMMOLd));
6396 if (NumLPads < 256) {
6397 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6401 MachineConstantPool *ConstantPool = MF->getConstantPool();
6402 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6403 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6405 // MachineConstantPool wants an explicit alignment.
6406 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6408 Align = getDataLayout()->getTypeAllocSize(C->getType());
6409 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6411 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6412 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6413 .addReg(VReg1, RegState::Define)
6414 .addConstantPoolIndex(Idx));
6415 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6420 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6425 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6426 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6427 .addReg(ARM::CPSR, RegState::Define)
6431 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6432 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6433 .addJumpTableIndex(MJTI)
6436 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6437 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6438 .addReg(ARM::CPSR, RegState::Define)
6439 .addReg(NewVReg2, RegState::Kill)
6442 MachineMemOperand *JTMMOLd =
6443 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6444 MachineMemOperand::MOLoad, 4, 4);
6446 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6447 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6448 .addReg(NewVReg4, RegState::Kill)
6450 .addMemOperand(JTMMOLd));
6452 unsigned NewVReg6 = NewVReg5;
6453 if (RelocM == Reloc::PIC_) {
6454 NewVReg6 = MRI->createVirtualRegister(TRC);
6455 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6456 .addReg(ARM::CPSR, RegState::Define)
6457 .addReg(NewVReg5, RegState::Kill)
6461 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6462 .addReg(NewVReg6, RegState::Kill)
6463 .addJumpTableIndex(MJTI)
6466 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6467 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6470 .addMemOperand(FIMMOLd));
6472 if (NumLPads < 256) {
6473 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6476 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6477 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6478 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6479 .addImm(NumLPads & 0xFFFF));
6481 unsigned VReg2 = VReg1;
6482 if ((NumLPads & 0xFFFF0000) != 0) {
6483 VReg2 = MRI->createVirtualRegister(TRC);
6484 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6486 .addImm(NumLPads >> 16));
6489 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6493 MachineConstantPool *ConstantPool = MF->getConstantPool();
6494 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6495 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6497 // MachineConstantPool wants an explicit alignment.
6498 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6500 Align = getDataLayout()->getTypeAllocSize(C->getType());
6501 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6503 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6504 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6505 .addReg(VReg1, RegState::Define)
6506 .addConstantPoolIndex(Idx)
6508 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6510 .addReg(VReg1, RegState::Kill));
6513 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6518 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6520 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6522 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6523 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6524 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6525 .addJumpTableIndex(MJTI)
6528 MachineMemOperand *JTMMOLd =
6529 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6530 MachineMemOperand::MOLoad, 4, 4);
6531 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6533 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6534 .addReg(NewVReg3, RegState::Kill)
6537 .addMemOperand(JTMMOLd));
6539 if (RelocM == Reloc::PIC_) {
6540 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6541 .addReg(NewVReg5, RegState::Kill)
6543 .addJumpTableIndex(MJTI)
6546 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6547 .addReg(NewVReg5, RegState::Kill)
6548 .addJumpTableIndex(MJTI)
6553 // Add the jump table entries as successors to the MBB.
6554 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6555 for (std::vector<MachineBasicBlock*>::iterator
6556 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6557 MachineBasicBlock *CurMBB = *I;
6558 if (SeenMBBs.insert(CurMBB))
6559 DispContBB->addSuccessor(CurMBB);
6562 // N.B. the order the invoke BBs are processed in doesn't matter here.
6563 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6564 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6565 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6566 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6567 MachineBasicBlock *BB = *I;
6569 // Remove the landing pad successor from the invoke block and replace it
6570 // with the new dispatch block.
6571 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6573 while (!Successors.empty()) {
6574 MachineBasicBlock *SMBB = Successors.pop_back_val();
6575 if (SMBB->isLandingPad()) {
6576 BB->removeSuccessor(SMBB);
6577 MBBLPads.push_back(SMBB);
6581 BB->addSuccessor(DispatchBB);
6583 // Find the invoke call and mark all of the callee-saved registers as
6584 // 'implicit defined' so that they're spilled. This prevents code from
6585 // moving instructions to before the EH block, where they will never be
6587 for (MachineBasicBlock::reverse_iterator
6588 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6589 if (!II->isCall()) continue;
6591 DenseMap<unsigned, bool> DefRegs;
6592 for (MachineInstr::mop_iterator
6593 OI = II->operands_begin(), OE = II->operands_end();
6595 if (!OI->isReg()) continue;
6596 DefRegs[OI->getReg()] = true;
6599 MachineInstrBuilder MIB(*MF, &*II);
6601 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6602 unsigned Reg = SavedRegs[i];
6603 if (Subtarget->isThumb2() &&
6604 !ARM::tGPRRegClass.contains(Reg) &&
6605 !ARM::hGPRRegClass.contains(Reg))
6607 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6609 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6612 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6619 // Mark all former landing pads as non-landing pads. The dispatch is the only
6621 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6622 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6623 (*I)->setIsLandingPad(false);
6625 // The instruction is gone now.
6626 MI->eraseFromParent();
6632 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6633 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6634 E = MBB->succ_end(); I != E; ++I)
6637 llvm_unreachable("Expecting a BB with two successors!");
6640 /// Return the load opcode for a given load size. If load size >= 8,
6641 /// neon opcode will be returned.
6642 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6644 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6645 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6647 return LdSize == 4 ? ARM::tLDRi
6648 : LdSize == 2 ? ARM::tLDRHi
6649 : LdSize == 1 ? ARM::tLDRBi : 0;
6651 return LdSize == 4 ? ARM::t2LDR_POST
6652 : LdSize == 2 ? ARM::t2LDRH_POST
6653 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6654 return LdSize == 4 ? ARM::LDR_POST_IMM
6655 : LdSize == 2 ? ARM::LDRH_POST
6656 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6659 /// Return the store opcode for a given store size. If store size >= 8,
6660 /// neon opcode will be returned.
6661 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6663 return StSize == 16 ? ARM::VST1q32wb_fixed
6664 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6666 return StSize == 4 ? ARM::tSTRi
6667 : StSize == 2 ? ARM::tSTRHi
6668 : StSize == 1 ? ARM::tSTRBi : 0;
6670 return StSize == 4 ? ARM::t2STR_POST
6671 : StSize == 2 ? ARM::t2STRH_POST
6672 : StSize == 1 ? ARM::t2STRB_POST : 0;
6673 return StSize == 4 ? ARM::STR_POST_IMM
6674 : StSize == 2 ? ARM::STRH_POST
6675 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6678 /// Emit a post-increment load operation with given size. The instructions
6679 /// will be added to BB at Pos.
6680 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6681 const TargetInstrInfo *TII, DebugLoc dl,
6682 unsigned LdSize, unsigned Data, unsigned AddrIn,
6683 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6684 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6685 assert(LdOpc != 0 && "Should have a load opcode");
6687 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6688 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6690 } else if (IsThumb1) {
6691 // load + update AddrIn
6692 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6693 .addReg(AddrIn).addImm(0));
6694 MachineInstrBuilder MIB =
6695 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6696 MIB = AddDefaultT1CC(MIB);
6697 MIB.addReg(AddrIn).addImm(LdSize);
6698 AddDefaultPred(MIB);
6699 } else if (IsThumb2) {
6700 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6701 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6704 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6705 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6706 .addReg(0).addImm(LdSize));
6710 /// Emit a post-increment store operation with given size. The instructions
6711 /// will be added to BB at Pos.
6712 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6713 const TargetInstrInfo *TII, DebugLoc dl,
6714 unsigned StSize, unsigned Data, unsigned AddrIn,
6715 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6716 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6717 assert(StOpc != 0 && "Should have a store opcode");
6719 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6720 .addReg(AddrIn).addImm(0).addReg(Data));
6721 } else if (IsThumb1) {
6722 // store + update AddrIn
6723 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6724 .addReg(AddrIn).addImm(0));
6725 MachineInstrBuilder MIB =
6726 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6727 MIB = AddDefaultT1CC(MIB);
6728 MIB.addReg(AddrIn).addImm(StSize);
6729 AddDefaultPred(MIB);
6730 } else if (IsThumb2) {
6731 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6732 .addReg(Data).addReg(AddrIn).addImm(StSize));
6734 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6735 .addReg(Data).addReg(AddrIn).addReg(0)
6741 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6742 MachineBasicBlock *BB) const {
6743 // This pseudo instruction has 3 operands: dst, src, size
6744 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6745 // Otherwise, we will generate unrolled scalar copies.
6746 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6747 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6748 MachineFunction::iterator It = BB;
6751 unsigned dest = MI->getOperand(0).getReg();
6752 unsigned src = MI->getOperand(1).getReg();
6753 unsigned SizeVal = MI->getOperand(2).getImm();
6754 unsigned Align = MI->getOperand(3).getImm();
6755 DebugLoc dl = MI->getDebugLoc();
6757 MachineFunction *MF = BB->getParent();
6758 MachineRegisterInfo &MRI = MF->getRegInfo();
6759 unsigned UnitSize = 0;
6760 const TargetRegisterClass *TRC = 0;
6761 const TargetRegisterClass *VecTRC = 0;
6763 bool IsThumb1 = Subtarget->isThumb1Only();
6764 bool IsThumb2 = Subtarget->isThumb2();
6768 } else if (Align & 2) {
6771 // Check whether we can use NEON instructions.
6772 if (!MF->getFunction()->getAttributes().
6773 hasAttribute(AttributeSet::FunctionIndex,
6774 Attribute::NoImplicitFloat) &&
6775 Subtarget->hasNEON()) {
6776 if ((Align % 16 == 0) && SizeVal >= 16)
6778 else if ((Align % 8 == 0) && SizeVal >= 8)
6781 // Can't use NEON instructions.
6786 // Select the correct opcode and register class for unit size load/store
6787 bool IsNeon = UnitSize >= 8;
6788 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
6789 : (const TargetRegisterClass *)&ARM::GPRRegClass;
6791 VecTRC = UnitSize == 16
6792 ? (const TargetRegisterClass *)&ARM::DPairRegClass
6794 ? (const TargetRegisterClass *)&ARM::DPRRegClass
6797 unsigned BytesLeft = SizeVal % UnitSize;
6798 unsigned LoopSize = SizeVal - BytesLeft;
6800 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6801 // Use LDR and STR to copy.
6802 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6803 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6804 unsigned srcIn = src;
6805 unsigned destIn = dest;
6806 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6807 unsigned srcOut = MRI.createVirtualRegister(TRC);
6808 unsigned destOut = MRI.createVirtualRegister(TRC);
6809 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
6810 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
6811 IsThumb1, IsThumb2);
6812 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
6813 IsThumb1, IsThumb2);
6818 // Handle the leftover bytes with LDRB and STRB.
6819 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6820 // [destOut] = STRB_POST(scratch, destIn, 1)
6821 for (unsigned i = 0; i < BytesLeft; i++) {
6822 unsigned srcOut = MRI.createVirtualRegister(TRC);
6823 unsigned destOut = MRI.createVirtualRegister(TRC);
6824 unsigned scratch = MRI.createVirtualRegister(TRC);
6825 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
6826 IsThumb1, IsThumb2);
6827 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
6828 IsThumb1, IsThumb2);
6832 MI->eraseFromParent(); // The instruction is gone now.
6836 // Expand the pseudo op to a loop.
6839 // movw varEnd, # --> with thumb2
6841 // ldrcp varEnd, idx --> without thumb2
6842 // fallthrough --> loopMBB
6844 // PHI varPhi, varEnd, varLoop
6845 // PHI srcPhi, src, srcLoop
6846 // PHI destPhi, dst, destLoop
6847 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6848 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6849 // subs varLoop, varPhi, #UnitSize
6851 // fallthrough --> exitMBB
6853 // epilogue to handle left-over bytes
6854 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6855 // [destOut] = STRB_POST(scratch, destLoop, 1)
6856 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6857 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6858 MF->insert(It, loopMBB);
6859 MF->insert(It, exitMBB);
6861 // Transfer the remainder of BB and its successor edges to exitMBB.
6862 exitMBB->splice(exitMBB->begin(), BB,
6863 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6864 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6866 // Load an immediate to varEnd.
6867 unsigned varEnd = MRI.createVirtualRegister(TRC);
6869 unsigned Vtmp = varEnd;
6870 if ((LoopSize & 0xFFFF0000) != 0)
6871 Vtmp = MRI.createVirtualRegister(TRC);
6872 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
6873 .addImm(LoopSize & 0xFFFF));
6875 if ((LoopSize & 0xFFFF0000) != 0)
6876 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6877 .addReg(Vtmp).addImm(LoopSize >> 16));
6879 MachineConstantPool *ConstantPool = MF->getConstantPool();
6880 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6881 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6883 // MachineConstantPool wants an explicit alignment.
6884 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6886 Align = getDataLayout()->getTypeAllocSize(C->getType());
6887 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6890 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
6891 varEnd, RegState::Define).addConstantPoolIndex(Idx));
6893 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
6894 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
6896 BB->addSuccessor(loopMBB);
6898 // Generate the loop body:
6899 // varPhi = PHI(varLoop, varEnd)
6900 // srcPhi = PHI(srcLoop, src)
6901 // destPhi = PHI(destLoop, dst)
6902 MachineBasicBlock *entryBB = BB;
6904 unsigned varLoop = MRI.createVirtualRegister(TRC);
6905 unsigned varPhi = MRI.createVirtualRegister(TRC);
6906 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6907 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6908 unsigned destLoop = MRI.createVirtualRegister(TRC);
6909 unsigned destPhi = MRI.createVirtualRegister(TRC);
6911 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6912 .addReg(varLoop).addMBB(loopMBB)
6913 .addReg(varEnd).addMBB(entryBB);
6914 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6915 .addReg(srcLoop).addMBB(loopMBB)
6916 .addReg(src).addMBB(entryBB);
6917 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6918 .addReg(destLoop).addMBB(loopMBB)
6919 .addReg(dest).addMBB(entryBB);
6921 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6922 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6923 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
6924 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
6925 IsThumb1, IsThumb2);
6926 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
6927 IsThumb1, IsThumb2);
6929 // Decrement loop variable by UnitSize.
6931 MachineInstrBuilder MIB =
6932 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
6933 MIB = AddDefaultT1CC(MIB);
6934 MIB.addReg(varPhi).addImm(UnitSize);
6935 AddDefaultPred(MIB);
6937 MachineInstrBuilder MIB =
6938 BuildMI(*BB, BB->end(), dl,
6939 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6940 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6941 MIB->getOperand(5).setReg(ARM::CPSR);
6942 MIB->getOperand(5).setIsDef(true);
6944 BuildMI(*BB, BB->end(), dl,
6945 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
6946 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6948 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6949 BB->addSuccessor(loopMBB);
6950 BB->addSuccessor(exitMBB);
6952 // Add epilogue to handle BytesLeft.
6954 MachineInstr *StartOfExit = exitMBB->begin();
6956 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6957 // [destOut] = STRB_POST(scratch, destLoop, 1)
6958 unsigned srcIn = srcLoop;
6959 unsigned destIn = destLoop;
6960 for (unsigned i = 0; i < BytesLeft; i++) {
6961 unsigned srcOut = MRI.createVirtualRegister(TRC);
6962 unsigned destOut = MRI.createVirtualRegister(TRC);
6963 unsigned scratch = MRI.createVirtualRegister(TRC);
6964 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
6965 IsThumb1, IsThumb2);
6966 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
6967 IsThumb1, IsThumb2);
6972 MI->eraseFromParent(); // The instruction is gone now.
6977 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6978 MachineBasicBlock *BB) const {
6979 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6980 DebugLoc dl = MI->getDebugLoc();
6981 bool isThumb2 = Subtarget->isThumb2();
6982 switch (MI->getOpcode()) {
6985 llvm_unreachable("Unexpected instr type to insert");
6987 // The Thumb2 pre-indexed stores have the same MI operands, they just
6988 // define them differently in the .td files from the isel patterns, so
6989 // they need pseudos.
6990 case ARM::t2STR_preidx:
6991 MI->setDesc(TII->get(ARM::t2STR_PRE));
6993 case ARM::t2STRB_preidx:
6994 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6996 case ARM::t2STRH_preidx:
6997 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7000 case ARM::STRi_preidx:
7001 case ARM::STRBi_preidx: {
7002 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7003 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7004 // Decode the offset.
7005 unsigned Offset = MI->getOperand(4).getImm();
7006 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7007 Offset = ARM_AM::getAM2Offset(Offset);
7011 MachineMemOperand *MMO = *MI->memoperands_begin();
7012 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7013 .addOperand(MI->getOperand(0)) // Rn_wb
7014 .addOperand(MI->getOperand(1)) // Rt
7015 .addOperand(MI->getOperand(2)) // Rn
7016 .addImm(Offset) // offset (skip GPR==zero_reg)
7017 .addOperand(MI->getOperand(5)) // pred
7018 .addOperand(MI->getOperand(6))
7019 .addMemOperand(MMO);
7020 MI->eraseFromParent();
7023 case ARM::STRr_preidx:
7024 case ARM::STRBr_preidx:
7025 case ARM::STRH_preidx: {
7027 switch (MI->getOpcode()) {
7028 default: llvm_unreachable("unexpected opcode!");
7029 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7030 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7031 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7033 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7034 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7035 MIB.addOperand(MI->getOperand(i));
7036 MI->eraseFromParent();
7040 case ARM::tMOVCCr_pseudo: {
7041 // To "insert" a SELECT_CC instruction, we actually have to insert the
7042 // diamond control-flow pattern. The incoming instruction knows the
7043 // destination vreg to set, the condition code register to branch on, the
7044 // true/false values to select between, and a branch opcode to use.
7045 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7046 MachineFunction::iterator It = BB;
7052 // cmpTY ccX, r1, r2
7054 // fallthrough --> copy0MBB
7055 MachineBasicBlock *thisMBB = BB;
7056 MachineFunction *F = BB->getParent();
7057 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7058 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7059 F->insert(It, copy0MBB);
7060 F->insert(It, sinkMBB);
7062 // Transfer the remainder of BB and its successor edges to sinkMBB.
7063 sinkMBB->splice(sinkMBB->begin(), BB,
7064 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7065 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7067 BB->addSuccessor(copy0MBB);
7068 BB->addSuccessor(sinkMBB);
7070 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7071 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7074 // %FalseValue = ...
7075 // # fallthrough to sinkMBB
7078 // Update machine-CFG edges
7079 BB->addSuccessor(sinkMBB);
7082 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7085 BuildMI(*BB, BB->begin(), dl,
7086 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7087 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7088 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7090 MI->eraseFromParent(); // The pseudo instruction is gone now.
7095 case ARM::BCCZi64: {
7096 // If there is an unconditional branch to the other successor, remove it.
7097 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7099 // Compare both parts that make up the double comparison separately for
7101 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7103 unsigned LHS1 = MI->getOperand(1).getReg();
7104 unsigned LHS2 = MI->getOperand(2).getReg();
7106 AddDefaultPred(BuildMI(BB, dl,
7107 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7108 .addReg(LHS1).addImm(0));
7109 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7110 .addReg(LHS2).addImm(0)
7111 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7113 unsigned RHS1 = MI->getOperand(3).getReg();
7114 unsigned RHS2 = MI->getOperand(4).getReg();
7115 AddDefaultPred(BuildMI(BB, dl,
7116 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7117 .addReg(LHS1).addReg(RHS1));
7118 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7119 .addReg(LHS2).addReg(RHS2)
7120 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7123 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7124 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7125 if (MI->getOperand(0).getImm() == ARMCC::NE)
7126 std::swap(destMBB, exitMBB);
7128 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7129 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7131 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7133 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7135 MI->eraseFromParent(); // The pseudo instruction is gone now.
7139 case ARM::Int_eh_sjlj_setjmp:
7140 case ARM::Int_eh_sjlj_setjmp_nofp:
7141 case ARM::tInt_eh_sjlj_setjmp:
7142 case ARM::t2Int_eh_sjlj_setjmp:
7143 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7144 EmitSjLjDispatchBlock(MI, BB);
7149 // To insert an ABS instruction, we have to insert the
7150 // diamond control-flow pattern. The incoming instruction knows the
7151 // source vreg to test against 0, the destination vreg to set,
7152 // the condition code register to branch on, the
7153 // true/false values to select between, and a branch opcode to use.
7158 // BCC (branch to SinkBB if V0 >= 0)
7159 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7160 // SinkBB: V1 = PHI(V2, V3)
7161 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7162 MachineFunction::iterator BBI = BB;
7164 MachineFunction *Fn = BB->getParent();
7165 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7166 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7167 Fn->insert(BBI, RSBBB);
7168 Fn->insert(BBI, SinkBB);
7170 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7171 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7172 bool isThumb2 = Subtarget->isThumb2();
7173 MachineRegisterInfo &MRI = Fn->getRegInfo();
7174 // In Thumb mode S must not be specified if source register is the SP or
7175 // PC and if destination register is the SP, so restrict register class
7176 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7177 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7178 (const TargetRegisterClass*)&ARM::GPRRegClass);
7180 // Transfer the remainder of BB and its successor edges to sinkMBB.
7181 SinkBB->splice(SinkBB->begin(), BB,
7182 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7183 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7185 BB->addSuccessor(RSBBB);
7186 BB->addSuccessor(SinkBB);
7188 // fall through to SinkMBB
7189 RSBBB->addSuccessor(SinkBB);
7191 // insert a cmp at the end of BB
7192 AddDefaultPred(BuildMI(BB, dl,
7193 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7194 .addReg(ABSSrcReg).addImm(0));
7196 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7198 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7199 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7201 // insert rsbri in RSBBB
7202 // Note: BCC and rsbri will be converted into predicated rsbmi
7203 // by if-conversion pass
7204 BuildMI(*RSBBB, RSBBB->begin(), dl,
7205 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7206 .addReg(ABSSrcReg, RegState::Kill)
7207 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7209 // insert PHI in SinkBB,
7210 // reuse ABSDstReg to not change uses of ABS instruction
7211 BuildMI(*SinkBB, SinkBB->begin(), dl,
7212 TII->get(ARM::PHI), ABSDstReg)
7213 .addReg(NewRsbDstReg).addMBB(RSBBB)
7214 .addReg(ABSSrcReg).addMBB(BB);
7216 // remove ABS instruction
7217 MI->eraseFromParent();
7219 // return last added BB
7222 case ARM::COPY_STRUCT_BYVAL_I32:
7224 return EmitStructByval(MI, BB);
7228 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7229 SDNode *Node) const {
7230 if (!MI->hasPostISelHook()) {
7231 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7232 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7236 const MCInstrDesc *MCID = &MI->getDesc();
7237 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7238 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7239 // operand is still set to noreg. If needed, set the optional operand's
7240 // register to CPSR, and remove the redundant implicit def.
7242 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7244 // Rename pseudo opcodes.
7245 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7247 const ARMBaseInstrInfo *TII =
7248 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7249 MCID = &TII->get(NewOpc);
7251 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7252 "converted opcode should be the same except for cc_out");
7256 // Add the optional cc_out operand
7257 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7259 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7261 // Any ARM instruction that sets the 's' bit should specify an optional
7262 // "cc_out" operand in the last operand position.
7263 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7264 assert(!NewOpc && "Optional cc_out operand required");
7267 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7268 // since we already have an optional CPSR def.
7269 bool definesCPSR = false;
7270 bool deadCPSR = false;
7271 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7273 const MachineOperand &MO = MI->getOperand(i);
7274 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7278 MI->RemoveOperand(i);
7283 assert(!NewOpc && "Optional cc_out operand required");
7286 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7288 assert(!MI->getOperand(ccOutIdx).getReg() &&
7289 "expect uninitialized optional cc_out operand");
7293 // If this instruction was defined with an optional CPSR def and its dag node
7294 // had a live implicit CPSR def, then activate the optional CPSR def.
7295 MachineOperand &MO = MI->getOperand(ccOutIdx);
7296 MO.setReg(ARM::CPSR);
7300 //===----------------------------------------------------------------------===//
7301 // ARM Optimization Hooks
7302 //===----------------------------------------------------------------------===//
7304 // Helper function that checks if N is a null or all ones constant.
7305 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7306 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7309 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7312 // Return true if N is conditionally 0 or all ones.
7313 // Detects these expressions where cc is an i1 value:
7315 // (select cc 0, y) [AllOnes=0]
7316 // (select cc y, 0) [AllOnes=0]
7317 // (zext cc) [AllOnes=0]
7318 // (sext cc) [AllOnes=0/1]
7319 // (select cc -1, y) [AllOnes=1]
7320 // (select cc y, -1) [AllOnes=1]
7322 // Invert is set when N is the null/all ones constant when CC is false.
7323 // OtherOp is set to the alternative value of N.
7324 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7325 SDValue &CC, bool &Invert,
7327 SelectionDAG &DAG) {
7328 switch (N->getOpcode()) {
7329 default: return false;
7331 CC = N->getOperand(0);
7332 SDValue N1 = N->getOperand(1);
7333 SDValue N2 = N->getOperand(2);
7334 if (isZeroOrAllOnes(N1, AllOnes)) {
7339 if (isZeroOrAllOnes(N2, AllOnes)) {
7346 case ISD::ZERO_EXTEND:
7347 // (zext cc) can never be the all ones value.
7351 case ISD::SIGN_EXTEND: {
7352 EVT VT = N->getValueType(0);
7353 CC = N->getOperand(0);
7354 if (CC.getValueType() != MVT::i1)
7358 // When looking for an AllOnes constant, N is an sext, and the 'other'
7360 OtherOp = DAG.getConstant(0, VT);
7361 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7362 // When looking for a 0 constant, N can be zext or sext.
7363 OtherOp = DAG.getConstant(1, VT);
7365 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7371 // Combine a constant select operand into its use:
7373 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7374 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7375 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7376 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7377 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7379 // The transform is rejected if the select doesn't have a constant operand that
7380 // is null, or all ones when AllOnes is set.
7382 // Also recognize sext/zext from i1:
7384 // (add (zext cc), x) -> (select cc (add x, 1), x)
7385 // (add (sext cc), x) -> (select cc (add x, -1), x)
7387 // These transformations eventually create predicated instructions.
7389 // @param N The node to transform.
7390 // @param Slct The N operand that is a select.
7391 // @param OtherOp The other N operand (x above).
7392 // @param DCI Context.
7393 // @param AllOnes Require the select constant to be all ones instead of null.
7394 // @returns The new node, or SDValue() on failure.
7396 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7397 TargetLowering::DAGCombinerInfo &DCI,
7398 bool AllOnes = false) {
7399 SelectionDAG &DAG = DCI.DAG;
7400 EVT VT = N->getValueType(0);
7401 SDValue NonConstantVal;
7404 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7405 NonConstantVal, DAG))
7408 // Slct is now know to be the desired identity constant when CC is true.
7409 SDValue TrueVal = OtherOp;
7410 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7411 OtherOp, NonConstantVal);
7412 // Unless SwapSelectOps says CC should be false.
7414 std::swap(TrueVal, FalseVal);
7416 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7417 CCOp, TrueVal, FalseVal);
7420 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7422 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7423 TargetLowering::DAGCombinerInfo &DCI) {
7424 SDValue N0 = N->getOperand(0);
7425 SDValue N1 = N->getOperand(1);
7426 if (N0.getNode()->hasOneUse()) {
7427 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7428 if (Result.getNode())
7431 if (N1.getNode()->hasOneUse()) {
7432 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7433 if (Result.getNode())
7439 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7440 // (only after legalization).
7441 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7442 TargetLowering::DAGCombinerInfo &DCI,
7443 const ARMSubtarget *Subtarget) {
7445 // Only perform optimization if after legalize, and if NEON is available. We
7446 // also expected both operands to be BUILD_VECTORs.
7447 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7448 || N0.getOpcode() != ISD::BUILD_VECTOR
7449 || N1.getOpcode() != ISD::BUILD_VECTOR)
7452 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7453 EVT VT = N->getValueType(0);
7454 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7457 // Check that the vector operands are of the right form.
7458 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7459 // operands, where N is the size of the formed vector.
7460 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7461 // index such that we have a pair wise add pattern.
7463 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7464 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7466 SDValue Vec = N0->getOperand(0)->getOperand(0);
7467 SDNode *V = Vec.getNode();
7468 unsigned nextIndex = 0;
7470 // For each operands to the ADD which are BUILD_VECTORs,
7471 // check to see if each of their operands are an EXTRACT_VECTOR with
7472 // the same vector and appropriate index.
7473 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7474 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7475 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7477 SDValue ExtVec0 = N0->getOperand(i);
7478 SDValue ExtVec1 = N1->getOperand(i);
7480 // First operand is the vector, verify its the same.
7481 if (V != ExtVec0->getOperand(0).getNode() ||
7482 V != ExtVec1->getOperand(0).getNode())
7485 // Second is the constant, verify its correct.
7486 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7487 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7489 // For the constant, we want to see all the even or all the odd.
7490 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7491 || C1->getZExtValue() != nextIndex+1)
7500 // Create VPADDL node.
7501 SelectionDAG &DAG = DCI.DAG;
7502 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7504 // Build operand list.
7505 SmallVector<SDValue, 8> Ops;
7506 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7507 TLI.getPointerTy()));
7509 // Input is the vector.
7512 // Get widened type and narrowed type.
7514 unsigned numElem = VT.getVectorNumElements();
7516 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7517 switch (inputLaneType.getSimpleVT().SimpleTy) {
7518 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7519 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7520 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7522 llvm_unreachable("Invalid vector element type for padd optimization.");
7525 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
7526 widenType, &Ops[0], Ops.size());
7527 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7528 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7531 static SDValue findMUL_LOHI(SDValue V) {
7532 if (V->getOpcode() == ISD::UMUL_LOHI ||
7533 V->getOpcode() == ISD::SMUL_LOHI)
7538 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7539 TargetLowering::DAGCombinerInfo &DCI,
7540 const ARMSubtarget *Subtarget) {
7542 if (Subtarget->isThumb1Only()) return SDValue();
7544 // Only perform the checks after legalize when the pattern is available.
7545 if (DCI.isBeforeLegalize()) return SDValue();
7547 // Look for multiply add opportunities.
7548 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7549 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7550 // a glue link from the first add to the second add.
7551 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7552 // a S/UMLAL instruction.
7555 // \ / \ [no multiline comment]
7561 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7562 SDValue AddcOp0 = AddcNode->getOperand(0);
7563 SDValue AddcOp1 = AddcNode->getOperand(1);
7565 // Check if the two operands are from the same mul_lohi node.
7566 if (AddcOp0.getNode() == AddcOp1.getNode())
7569 assert(AddcNode->getNumValues() == 2 &&
7570 AddcNode->getValueType(0) == MVT::i32 &&
7571 "Expect ADDC with two result values. First: i32");
7573 // Check that we have a glued ADDC node.
7574 if (AddcNode->getValueType(1) != MVT::Glue)
7577 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7578 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7579 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7580 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7581 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7584 // Look for the glued ADDE.
7585 SDNode* AddeNode = AddcNode->getGluedUser();
7586 if (AddeNode == NULL)
7589 // Make sure it is really an ADDE.
7590 if (AddeNode->getOpcode() != ISD::ADDE)
7593 assert(AddeNode->getNumOperands() == 3 &&
7594 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7595 "ADDE node has the wrong inputs");
7597 // Check for the triangle shape.
7598 SDValue AddeOp0 = AddeNode->getOperand(0);
7599 SDValue AddeOp1 = AddeNode->getOperand(1);
7601 // Make sure that the ADDE operands are not coming from the same node.
7602 if (AddeOp0.getNode() == AddeOp1.getNode())
7605 // Find the MUL_LOHI node walking up ADDE's operands.
7606 bool IsLeftOperandMUL = false;
7607 SDValue MULOp = findMUL_LOHI(AddeOp0);
7608 if (MULOp == SDValue())
7609 MULOp = findMUL_LOHI(AddeOp1);
7611 IsLeftOperandMUL = true;
7612 if (MULOp == SDValue())
7615 // Figure out the right opcode.
7616 unsigned Opc = MULOp->getOpcode();
7617 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7619 // Figure out the high and low input values to the MLAL node.
7620 SDValue* HiMul = &MULOp;
7621 SDValue* HiAdd = NULL;
7622 SDValue* LoMul = NULL;
7623 SDValue* LowAdd = NULL;
7625 if (IsLeftOperandMUL)
7631 if (AddcOp0->getOpcode() == Opc) {
7635 if (AddcOp1->getOpcode() == Opc) {
7643 if (LoMul->getNode() != HiMul->getNode())
7646 // Create the merged node.
7647 SelectionDAG &DAG = DCI.DAG;
7649 // Build operand list.
7650 SmallVector<SDValue, 8> Ops;
7651 Ops.push_back(LoMul->getOperand(0));
7652 Ops.push_back(LoMul->getOperand(1));
7653 Ops.push_back(*LowAdd);
7654 Ops.push_back(*HiAdd);
7656 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
7657 DAG.getVTList(MVT::i32, MVT::i32),
7658 &Ops[0], Ops.size());
7660 // Replace the ADDs' nodes uses by the MLA node's values.
7661 SDValue HiMLALResult(MLALNode.getNode(), 1);
7662 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7664 SDValue LoMLALResult(MLALNode.getNode(), 0);
7665 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7667 // Return original node to notify the driver to stop replacing.
7668 SDValue resNode(AddcNode, 0);
7672 /// PerformADDCCombine - Target-specific dag combine transform from
7673 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7674 static SDValue PerformADDCCombine(SDNode *N,
7675 TargetLowering::DAGCombinerInfo &DCI,
7676 const ARMSubtarget *Subtarget) {
7678 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7682 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7683 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7684 /// called with the default operands, and if that fails, with commuted
7686 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7687 TargetLowering::DAGCombinerInfo &DCI,
7688 const ARMSubtarget *Subtarget){
7690 // Attempt to create vpaddl for this add.
7691 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7692 if (Result.getNode())
7695 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7696 if (N0.getNode()->hasOneUse()) {
7697 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7698 if (Result.getNode()) return Result;
7703 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7705 static SDValue PerformADDCombine(SDNode *N,
7706 TargetLowering::DAGCombinerInfo &DCI,
7707 const ARMSubtarget *Subtarget) {
7708 SDValue N0 = N->getOperand(0);
7709 SDValue N1 = N->getOperand(1);
7711 // First try with the default operand order.
7712 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7713 if (Result.getNode())
7716 // If that didn't work, try again with the operands commuted.
7717 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7720 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7722 static SDValue PerformSUBCombine(SDNode *N,
7723 TargetLowering::DAGCombinerInfo &DCI) {
7724 SDValue N0 = N->getOperand(0);
7725 SDValue N1 = N->getOperand(1);
7727 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7728 if (N1.getNode()->hasOneUse()) {
7729 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7730 if (Result.getNode()) return Result;
7736 /// PerformVMULCombine
7737 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7738 /// special multiplier accumulator forwarding.
7744 // However, for (A + B) * (A + B),
7751 static SDValue PerformVMULCombine(SDNode *N,
7752 TargetLowering::DAGCombinerInfo &DCI,
7753 const ARMSubtarget *Subtarget) {
7754 if (!Subtarget->hasVMLxForwarding())
7757 SelectionDAG &DAG = DCI.DAG;
7758 SDValue N0 = N->getOperand(0);
7759 SDValue N1 = N->getOperand(1);
7760 unsigned Opcode = N0.getOpcode();
7761 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7762 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7763 Opcode = N1.getOpcode();
7764 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7765 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7773 EVT VT = N->getValueType(0);
7775 SDValue N00 = N0->getOperand(0);
7776 SDValue N01 = N0->getOperand(1);
7777 return DAG.getNode(Opcode, DL, VT,
7778 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7779 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7782 static SDValue PerformMULCombine(SDNode *N,
7783 TargetLowering::DAGCombinerInfo &DCI,
7784 const ARMSubtarget *Subtarget) {
7785 SelectionDAG &DAG = DCI.DAG;
7787 if (Subtarget->isThumb1Only())
7790 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7793 EVT VT = N->getValueType(0);
7794 if (VT.is64BitVector() || VT.is128BitVector())
7795 return PerformVMULCombine(N, DCI, Subtarget);
7799 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7803 int64_t MulAmt = C->getSExtValue();
7804 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
7806 ShiftAmt = ShiftAmt & (32 - 1);
7807 SDValue V = N->getOperand(0);
7811 MulAmt >>= ShiftAmt;
7814 if (isPowerOf2_32(MulAmt - 1)) {
7815 // (mul x, 2^N + 1) => (add (shl x, N), x)
7816 Res = DAG.getNode(ISD::ADD, DL, VT,
7818 DAG.getNode(ISD::SHL, DL, VT,
7820 DAG.getConstant(Log2_32(MulAmt - 1),
7822 } else if (isPowerOf2_32(MulAmt + 1)) {
7823 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7824 Res = DAG.getNode(ISD::SUB, DL, VT,
7825 DAG.getNode(ISD::SHL, DL, VT,
7827 DAG.getConstant(Log2_32(MulAmt + 1),
7833 uint64_t MulAmtAbs = -MulAmt;
7834 if (isPowerOf2_32(MulAmtAbs + 1)) {
7835 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7836 Res = DAG.getNode(ISD::SUB, DL, VT,
7838 DAG.getNode(ISD::SHL, DL, VT,
7840 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7842 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7843 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7844 Res = DAG.getNode(ISD::ADD, DL, VT,
7846 DAG.getNode(ISD::SHL, DL, VT,
7848 DAG.getConstant(Log2_32(MulAmtAbs-1),
7850 Res = DAG.getNode(ISD::SUB, DL, VT,
7851 DAG.getConstant(0, MVT::i32),Res);
7858 Res = DAG.getNode(ISD::SHL, DL, VT,
7859 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7861 // Do not add new nodes to DAG combiner worklist.
7862 DCI.CombineTo(N, Res, false);
7866 static SDValue PerformANDCombine(SDNode *N,
7867 TargetLowering::DAGCombinerInfo &DCI,
7868 const ARMSubtarget *Subtarget) {
7870 // Attempt to use immediate-form VBIC
7871 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7873 EVT VT = N->getValueType(0);
7874 SelectionDAG &DAG = DCI.DAG;
7876 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7879 APInt SplatBits, SplatUndef;
7880 unsigned SplatBitSize;
7883 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7884 if (SplatBitSize <= 64) {
7886 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7887 SplatUndef.getZExtValue(), SplatBitSize,
7888 DAG, VbicVT, VT.is128BitVector(),
7890 if (Val.getNode()) {
7892 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7893 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7894 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7899 if (!Subtarget->isThumb1Only()) {
7900 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7901 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7902 if (Result.getNode())
7909 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7910 static SDValue PerformORCombine(SDNode *N,
7911 TargetLowering::DAGCombinerInfo &DCI,
7912 const ARMSubtarget *Subtarget) {
7913 // Attempt to use immediate-form VORR
7914 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7916 EVT VT = N->getValueType(0);
7917 SelectionDAG &DAG = DCI.DAG;
7919 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7922 APInt SplatBits, SplatUndef;
7923 unsigned SplatBitSize;
7925 if (BVN && Subtarget->hasNEON() &&
7926 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7927 if (SplatBitSize <= 64) {
7929 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7930 SplatUndef.getZExtValue(), SplatBitSize,
7931 DAG, VorrVT, VT.is128BitVector(),
7933 if (Val.getNode()) {
7935 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7936 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7937 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7942 if (!Subtarget->isThumb1Only()) {
7943 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7944 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7945 if (Result.getNode())
7949 // The code below optimizes (or (and X, Y), Z).
7950 // The AND operand needs to have a single user to make these optimizations
7952 SDValue N0 = N->getOperand(0);
7953 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7955 SDValue N1 = N->getOperand(1);
7957 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7958 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7959 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7961 unsigned SplatBitSize;
7964 APInt SplatBits0, SplatBits1;
7965 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7966 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7967 // Ensure that the second operand of both ands are constants
7968 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7969 HasAnyUndefs) && !HasAnyUndefs) {
7970 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7971 HasAnyUndefs) && !HasAnyUndefs) {
7972 // Ensure that the bit width of the constants are the same and that
7973 // the splat arguments are logical inverses as per the pattern we
7974 // are trying to simplify.
7975 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
7976 SplatBits0 == ~SplatBits1) {
7977 // Canonicalize the vector type to make instruction selection
7979 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7980 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7984 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7990 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7993 // BFI is only available on V6T2+
7994 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7998 // 1) or (and A, mask), val => ARMbfi A, val, mask
7999 // iff (val & mask) == val
8001 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8002 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8003 // && mask == ~mask2
8004 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8005 // && ~mask == mask2
8006 // (i.e., copy a bitfield value into another bitfield of the same width)
8011 SDValue N00 = N0.getOperand(0);
8013 // The value and the mask need to be constants so we can verify this is
8014 // actually a bitfield set. If the mask is 0xffff, we can do better
8015 // via a movt instruction, so don't use BFI in that case.
8016 SDValue MaskOp = N0.getOperand(1);
8017 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8020 unsigned Mask = MaskC->getZExtValue();
8024 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8025 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8027 unsigned Val = N1C->getZExtValue();
8028 if ((Val & ~Mask) != Val)
8031 if (ARM::isBitFieldInvertedMask(Mask)) {
8032 Val >>= countTrailingZeros(~Mask);
8034 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8035 DAG.getConstant(Val, MVT::i32),
8036 DAG.getConstant(Mask, MVT::i32));
8038 // Do not add new nodes to DAG combiner worklist.
8039 DCI.CombineTo(N, Res, false);
8042 } else if (N1.getOpcode() == ISD::AND) {
8043 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8044 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8047 unsigned Mask2 = N11C->getZExtValue();
8049 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8051 if (ARM::isBitFieldInvertedMask(Mask) &&
8053 // The pack halfword instruction works better for masks that fit it,
8054 // so use that when it's available.
8055 if (Subtarget->hasT2ExtractPack() &&
8056 (Mask == 0xffff || Mask == 0xffff0000))
8059 unsigned amt = countTrailingZeros(Mask2);
8060 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8061 DAG.getConstant(amt, MVT::i32));
8062 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8063 DAG.getConstant(Mask, MVT::i32));
8064 // Do not add new nodes to DAG combiner worklist.
8065 DCI.CombineTo(N, Res, false);
8067 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8069 // The pack halfword instruction works better for masks that fit it,
8070 // so use that when it's available.
8071 if (Subtarget->hasT2ExtractPack() &&
8072 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8075 unsigned lsb = countTrailingZeros(Mask);
8076 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8077 DAG.getConstant(lsb, MVT::i32));
8078 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8079 DAG.getConstant(Mask2, MVT::i32));
8080 // Do not add new nodes to DAG combiner worklist.
8081 DCI.CombineTo(N, Res, false);
8086 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8087 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8088 ARM::isBitFieldInvertedMask(~Mask)) {
8089 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8090 // where lsb(mask) == #shamt and masked bits of B are known zero.
8091 SDValue ShAmt = N00.getOperand(1);
8092 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8093 unsigned LSB = countTrailingZeros(Mask);
8097 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8098 DAG.getConstant(~Mask, MVT::i32));
8100 // Do not add new nodes to DAG combiner worklist.
8101 DCI.CombineTo(N, Res, false);
8107 static SDValue PerformXORCombine(SDNode *N,
8108 TargetLowering::DAGCombinerInfo &DCI,
8109 const ARMSubtarget *Subtarget) {
8110 EVT VT = N->getValueType(0);
8111 SelectionDAG &DAG = DCI.DAG;
8113 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8116 if (!Subtarget->isThumb1Only()) {
8117 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8118 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8119 if (Result.getNode())
8126 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8127 /// the bits being cleared by the AND are not demanded by the BFI.
8128 static SDValue PerformBFICombine(SDNode *N,
8129 TargetLowering::DAGCombinerInfo &DCI) {
8130 SDValue N1 = N->getOperand(1);
8131 if (N1.getOpcode() == ISD::AND) {
8132 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8135 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8136 unsigned LSB = countTrailingZeros(~InvMask);
8137 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8138 unsigned Mask = (1 << Width)-1;
8139 unsigned Mask2 = N11C->getZExtValue();
8140 if ((Mask & (~Mask2)) == 0)
8141 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8142 N->getOperand(0), N1.getOperand(0),
8148 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8149 /// ARMISD::VMOVRRD.
8150 static SDValue PerformVMOVRRDCombine(SDNode *N,
8151 TargetLowering::DAGCombinerInfo &DCI) {
8152 // vmovrrd(vmovdrr x, y) -> x,y
8153 SDValue InDouble = N->getOperand(0);
8154 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8155 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8157 // vmovrrd(load f64) -> (load i32), (load i32)
8158 SDNode *InNode = InDouble.getNode();
8159 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8160 InNode->getValueType(0) == MVT::f64 &&
8161 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8162 !cast<LoadSDNode>(InNode)->isVolatile()) {
8163 // TODO: Should this be done for non-FrameIndex operands?
8164 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8166 SelectionDAG &DAG = DCI.DAG;
8168 SDValue BasePtr = LD->getBasePtr();
8169 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8170 LD->getPointerInfo(), LD->isVolatile(),
8171 LD->isNonTemporal(), LD->isInvariant(),
8172 LD->getAlignment());
8174 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8175 DAG.getConstant(4, MVT::i32));
8176 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8177 LD->getPointerInfo(), LD->isVolatile(),
8178 LD->isNonTemporal(), LD->isInvariant(),
8179 std::min(4U, LD->getAlignment() / 2));
8181 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8182 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8183 DCI.RemoveFromWorklist(LD);
8191 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8192 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8193 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8194 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8195 SDValue Op0 = N->getOperand(0);
8196 SDValue Op1 = N->getOperand(1);
8197 if (Op0.getOpcode() == ISD::BITCAST)
8198 Op0 = Op0.getOperand(0);
8199 if (Op1.getOpcode() == ISD::BITCAST)
8200 Op1 = Op1.getOperand(0);
8201 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8202 Op0.getNode() == Op1.getNode() &&
8203 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8204 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8205 N->getValueType(0), Op0.getOperand(0));
8209 /// PerformSTORECombine - Target-specific dag combine xforms for
8211 static SDValue PerformSTORECombine(SDNode *N,
8212 TargetLowering::DAGCombinerInfo &DCI) {
8213 StoreSDNode *St = cast<StoreSDNode>(N);
8214 if (St->isVolatile())
8217 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8218 // pack all of the elements in one place. Next, store to memory in fewer
8220 SDValue StVal = St->getValue();
8221 EVT VT = StVal.getValueType();
8222 if (St->isTruncatingStore() && VT.isVector()) {
8223 SelectionDAG &DAG = DCI.DAG;
8224 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8225 EVT StVT = St->getMemoryVT();
8226 unsigned NumElems = VT.getVectorNumElements();
8227 assert(StVT != VT && "Cannot truncate to the same type");
8228 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8229 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8231 // From, To sizes and ElemCount must be pow of two
8232 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8234 // We are going to use the original vector elt for storing.
8235 // Accumulated smaller vector elements must be a multiple of the store size.
8236 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8238 unsigned SizeRatio = FromEltSz / ToEltSz;
8239 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8241 // Create a type on which we perform the shuffle.
8242 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8243 NumElems*SizeRatio);
8244 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8247 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8248 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8249 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8251 // Can't shuffle using an illegal type.
8252 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8254 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8255 DAG.getUNDEF(WideVec.getValueType()),
8257 // At this point all of the data is stored at the bottom of the
8258 // register. We now need to save it to mem.
8260 // Find the largest store unit
8261 MVT StoreType = MVT::i8;
8262 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8263 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8264 MVT Tp = (MVT::SimpleValueType)tp;
8265 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8268 // Didn't find a legal store type.
8269 if (!TLI.isTypeLegal(StoreType))
8272 // Bitcast the original vector into a vector of store-size units
8273 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8274 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8275 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8276 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8277 SmallVector<SDValue, 8> Chains;
8278 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8279 TLI.getPointerTy());
8280 SDValue BasePtr = St->getBasePtr();
8282 // Perform one or more big stores into memory.
8283 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8284 for (unsigned I = 0; I < E; I++) {
8285 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8286 StoreType, ShuffWide,
8287 DAG.getIntPtrConstant(I));
8288 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8289 St->getPointerInfo(), St->isVolatile(),
8290 St->isNonTemporal(), St->getAlignment());
8291 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8293 Chains.push_back(Ch);
8295 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8299 if (!ISD::isNormalStore(St))
8302 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8303 // ARM stores of arguments in the same cache line.
8304 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8305 StVal.getNode()->hasOneUse()) {
8306 SelectionDAG &DAG = DCI.DAG;
8308 SDValue BasePtr = St->getBasePtr();
8309 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8310 StVal.getNode()->getOperand(0), BasePtr,
8311 St->getPointerInfo(), St->isVolatile(),
8312 St->isNonTemporal(), St->getAlignment());
8314 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8315 DAG.getConstant(4, MVT::i32));
8316 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8317 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8318 St->isNonTemporal(),
8319 std::min(4U, St->getAlignment() / 2));
8322 if (StVal.getValueType() != MVT::i64 ||
8323 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8326 // Bitcast an i64 store extracted from a vector to f64.
8327 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8328 SelectionDAG &DAG = DCI.DAG;
8330 SDValue IntVec = StVal.getOperand(0);
8331 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8332 IntVec.getValueType().getVectorNumElements());
8333 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8334 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8335 Vec, StVal.getOperand(1));
8337 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8338 // Make the DAGCombiner fold the bitcasts.
8339 DCI.AddToWorklist(Vec.getNode());
8340 DCI.AddToWorklist(ExtElt.getNode());
8341 DCI.AddToWorklist(V.getNode());
8342 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8343 St->getPointerInfo(), St->isVolatile(),
8344 St->isNonTemporal(), St->getAlignment(),
8348 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8349 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8350 /// i64 vector to have f64 elements, since the value can then be loaded
8351 /// directly into a VFP register.
8352 static bool hasNormalLoadOperand(SDNode *N) {
8353 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8354 for (unsigned i = 0; i < NumElts; ++i) {
8355 SDNode *Elt = N->getOperand(i).getNode();
8356 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8362 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8363 /// ISD::BUILD_VECTOR.
8364 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8365 TargetLowering::DAGCombinerInfo &DCI){
8366 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8367 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8368 // into a pair of GPRs, which is fine when the value is used as a scalar,
8369 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8370 SelectionDAG &DAG = DCI.DAG;
8371 if (N->getNumOperands() == 2) {
8372 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8377 // Load i64 elements as f64 values so that type legalization does not split
8378 // them up into i32 values.
8379 EVT VT = N->getValueType(0);
8380 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8383 SmallVector<SDValue, 8> Ops;
8384 unsigned NumElts = VT.getVectorNumElements();
8385 for (unsigned i = 0; i < NumElts; ++i) {
8386 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8388 // Make the DAGCombiner fold the bitcast.
8389 DCI.AddToWorklist(V.getNode());
8391 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8392 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8393 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8396 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8398 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8399 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8400 // At that time, we may have inserted bitcasts from integer to float.
8401 // If these bitcasts have survived DAGCombine, change the lowering of this
8402 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8403 // force to use floating point types.
8405 // Make sure we can change the type of the vector.
8406 // This is possible iff:
8407 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8408 // 1.1. Vector is used only once.
8409 // 1.2. Use is a bit convert to an integer type.
8410 // 2. The size of its operands are 32-bits (64-bits are not legal).
8411 EVT VT = N->getValueType(0);
8412 EVT EltVT = VT.getVectorElementType();
8414 // Check 1.1. and 2.
8415 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8418 // By construction, the input type must be float.
8419 assert(EltVT == MVT::f32 && "Unexpected type!");
8422 SDNode *Use = *N->use_begin();
8423 if (Use->getOpcode() != ISD::BITCAST ||
8424 Use->getValueType(0).isFloatingPoint())
8427 // Check profitability.
8428 // Model is, if more than half of the relevant operands are bitcast from
8429 // i32, turn the build_vector into a sequence of insert_vector_elt.
8430 // Relevant operands are everything that is not statically
8431 // (i.e., at compile time) bitcasted.
8432 unsigned NumOfBitCastedElts = 0;
8433 unsigned NumElts = VT.getVectorNumElements();
8434 unsigned NumOfRelevantElts = NumElts;
8435 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8436 SDValue Elt = N->getOperand(Idx);
8437 if (Elt->getOpcode() == ISD::BITCAST) {
8438 // Assume only bit cast to i32 will go away.
8439 if (Elt->getOperand(0).getValueType() == MVT::i32)
8440 ++NumOfBitCastedElts;
8441 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8442 // Constants are statically casted, thus do not count them as
8443 // relevant operands.
8444 --NumOfRelevantElts;
8447 // Check if more than half of the elements require a non-free bitcast.
8448 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8451 SelectionDAG &DAG = DCI.DAG;
8452 // Create the new vector type.
8453 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8454 // Check if the type is legal.
8455 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8456 if (!TLI.isTypeLegal(VecVT))
8460 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8461 // => BITCAST INSERT_VECTOR_ELT
8462 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8464 SDValue Vec = DAG.getUNDEF(VecVT);
8466 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8467 SDValue V = N->getOperand(Idx);
8468 if (V.getOpcode() == ISD::UNDEF)
8470 if (V.getOpcode() == ISD::BITCAST &&
8471 V->getOperand(0).getValueType() == MVT::i32)
8472 // Fold obvious case.
8473 V = V.getOperand(0);
8475 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8476 // Make the DAGCombiner fold the bitcasts.
8477 DCI.AddToWorklist(V.getNode());
8479 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8480 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8482 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8483 // Make the DAGCombiner fold the bitcasts.
8484 DCI.AddToWorklist(Vec.getNode());
8488 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8489 /// ISD::INSERT_VECTOR_ELT.
8490 static SDValue PerformInsertEltCombine(SDNode *N,
8491 TargetLowering::DAGCombinerInfo &DCI) {
8492 // Bitcast an i64 load inserted into a vector to f64.
8493 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8494 EVT VT = N->getValueType(0);
8495 SDNode *Elt = N->getOperand(1).getNode();
8496 if (VT.getVectorElementType() != MVT::i64 ||
8497 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8500 SelectionDAG &DAG = DCI.DAG;
8502 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8503 VT.getVectorNumElements());
8504 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8505 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8506 // Make the DAGCombiner fold the bitcasts.
8507 DCI.AddToWorklist(Vec.getNode());
8508 DCI.AddToWorklist(V.getNode());
8509 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8510 Vec, V, N->getOperand(2));
8511 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8514 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8515 /// ISD::VECTOR_SHUFFLE.
8516 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8517 // The LLVM shufflevector instruction does not require the shuffle mask
8518 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8519 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8520 // operands do not match the mask length, they are extended by concatenating
8521 // them with undef vectors. That is probably the right thing for other
8522 // targets, but for NEON it is better to concatenate two double-register
8523 // size vector operands into a single quad-register size vector. Do that
8524 // transformation here:
8525 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8526 // shuffle(concat(v1, v2), undef)
8527 SDValue Op0 = N->getOperand(0);
8528 SDValue Op1 = N->getOperand(1);
8529 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8530 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8531 Op0.getNumOperands() != 2 ||
8532 Op1.getNumOperands() != 2)
8534 SDValue Concat0Op1 = Op0.getOperand(1);
8535 SDValue Concat1Op1 = Op1.getOperand(1);
8536 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8537 Concat1Op1.getOpcode() != ISD::UNDEF)
8539 // Skip the transformation if any of the types are illegal.
8540 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8541 EVT VT = N->getValueType(0);
8542 if (!TLI.isTypeLegal(VT) ||
8543 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8544 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8547 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8548 Op0.getOperand(0), Op1.getOperand(0));
8549 // Translate the shuffle mask.
8550 SmallVector<int, 16> NewMask;
8551 unsigned NumElts = VT.getVectorNumElements();
8552 unsigned HalfElts = NumElts/2;
8553 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8554 for (unsigned n = 0; n < NumElts; ++n) {
8555 int MaskElt = SVN->getMaskElt(n);
8557 if (MaskElt < (int)HalfElts)
8559 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8560 NewElt = HalfElts + MaskElt - NumElts;
8561 NewMask.push_back(NewElt);
8563 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8564 DAG.getUNDEF(VT), NewMask.data());
8567 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8568 /// NEON load/store intrinsics to merge base address updates.
8569 static SDValue CombineBaseUpdate(SDNode *N,
8570 TargetLowering::DAGCombinerInfo &DCI) {
8571 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8574 SelectionDAG &DAG = DCI.DAG;
8575 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8576 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8577 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8578 SDValue Addr = N->getOperand(AddrOpIdx);
8580 // Search for a use of the address operand that is an increment.
8581 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8582 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8584 if (User->getOpcode() != ISD::ADD ||
8585 UI.getUse().getResNo() != Addr.getResNo())
8588 // Check that the add is independent of the load/store. Otherwise, folding
8589 // it would create a cycle.
8590 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8593 // Find the new opcode for the updating load/store.
8595 bool isLaneOp = false;
8596 unsigned NewOpc = 0;
8597 unsigned NumVecs = 0;
8599 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8601 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8602 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8604 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8606 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8608 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8610 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8611 NumVecs = 2; isLaneOp = true; break;
8612 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8613 NumVecs = 3; isLaneOp = true; break;
8614 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8615 NumVecs = 4; isLaneOp = true; break;
8616 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8617 NumVecs = 1; isLoad = false; break;
8618 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8619 NumVecs = 2; isLoad = false; break;
8620 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8621 NumVecs = 3; isLoad = false; break;
8622 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8623 NumVecs = 4; isLoad = false; break;
8624 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8625 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8626 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8627 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8628 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8629 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8633 switch (N->getOpcode()) {
8634 default: llvm_unreachable("unexpected opcode for Neon base update");
8635 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8636 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8637 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8641 // Find the size of memory referenced by the load/store.
8644 VecTy = N->getValueType(0);
8646 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8647 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8649 NumBytes /= VecTy.getVectorNumElements();
8651 // If the increment is a constant, it must match the memory ref size.
8652 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8653 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8654 uint64_t IncVal = CInc->getZExtValue();
8655 if (IncVal != NumBytes)
8657 } else if (NumBytes >= 3 * 16) {
8658 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8659 // separate instructions that make it harder to use a non-constant update.
8663 // Create the new updating load/store node.
8665 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8667 for (n = 0; n < NumResultVecs; ++n)
8669 Tys[n++] = MVT::i32;
8670 Tys[n] = MVT::Other;
8671 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
8672 SmallVector<SDValue, 8> Ops;
8673 Ops.push_back(N->getOperand(0)); // incoming chain
8674 Ops.push_back(N->getOperand(AddrOpIdx));
8676 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8677 Ops.push_back(N->getOperand(i));
8679 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8680 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8681 Ops.data(), Ops.size(),
8682 MemInt->getMemoryVT(),
8683 MemInt->getMemOperand());
8686 std::vector<SDValue> NewResults;
8687 for (unsigned i = 0; i < NumResultVecs; ++i) {
8688 NewResults.push_back(SDValue(UpdN.getNode(), i));
8690 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8691 DCI.CombineTo(N, NewResults);
8692 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8699 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8700 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8701 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8703 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8704 SelectionDAG &DAG = DCI.DAG;
8705 EVT VT = N->getValueType(0);
8706 // vldN-dup instructions only support 64-bit vectors for N > 1.
8707 if (!VT.is64BitVector())
8710 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8711 SDNode *VLD = N->getOperand(0).getNode();
8712 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8714 unsigned NumVecs = 0;
8715 unsigned NewOpc = 0;
8716 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8717 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8719 NewOpc = ARMISD::VLD2DUP;
8720 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8722 NewOpc = ARMISD::VLD3DUP;
8723 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8725 NewOpc = ARMISD::VLD4DUP;
8730 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8731 // numbers match the load.
8732 unsigned VLDLaneNo =
8733 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8734 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8736 // Ignore uses of the chain result.
8737 if (UI.getUse().getResNo() == NumVecs)
8740 if (User->getOpcode() != ARMISD::VDUPLANE ||
8741 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8745 // Create the vldN-dup node.
8748 for (n = 0; n < NumVecs; ++n)
8750 Tys[n] = MVT::Other;
8751 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
8752 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8753 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8754 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
8755 Ops, 2, VLDMemInt->getMemoryVT(),
8756 VLDMemInt->getMemOperand());
8759 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8761 unsigned ResNo = UI.getUse().getResNo();
8762 // Ignore uses of the chain result.
8763 if (ResNo == NumVecs)
8766 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8769 // Now the vldN-lane intrinsic is dead except for its chain result.
8770 // Update uses of the chain.
8771 std::vector<SDValue> VLDDupResults;
8772 for (unsigned n = 0; n < NumVecs; ++n)
8773 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8774 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8775 DCI.CombineTo(VLD, VLDDupResults);
8780 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8781 /// ARMISD::VDUPLANE.
8782 static SDValue PerformVDUPLANECombine(SDNode *N,
8783 TargetLowering::DAGCombinerInfo &DCI) {
8784 SDValue Op = N->getOperand(0);
8786 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8787 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8788 if (CombineVLDDUP(N, DCI))
8789 return SDValue(N, 0);
8791 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8792 // redundant. Ignore bit_converts for now; element sizes are checked below.
8793 while (Op.getOpcode() == ISD::BITCAST)
8794 Op = Op.getOperand(0);
8795 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8798 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8799 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8800 // The canonical VMOV for a zero vector uses a 32-bit element size.
8801 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8803 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8805 EVT VT = N->getValueType(0);
8806 if (EltSize > VT.getVectorElementType().getSizeInBits())
8809 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
8812 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8813 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8814 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8818 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8820 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8825 APFloat APF = C->getValueAPF();
8826 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8827 != APFloat::opOK || !isExact)
8830 c0 = (I == 0) ? cN : c0;
8831 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8838 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8839 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8840 /// when the VMUL has a constant operand that is a power of 2.
8842 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8843 /// vmul.f32 d16, d17, d16
8844 /// vcvt.s32.f32 d16, d16
8846 /// vcvt.s32.f32 d16, d16, #3
8847 static SDValue PerformVCVTCombine(SDNode *N,
8848 TargetLowering::DAGCombinerInfo &DCI,
8849 const ARMSubtarget *Subtarget) {
8850 SelectionDAG &DAG = DCI.DAG;
8851 SDValue Op = N->getOperand(0);
8853 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8854 Op.getOpcode() != ISD::FMUL)
8858 SDValue N0 = Op->getOperand(0);
8859 SDValue ConstVec = Op->getOperand(1);
8860 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8862 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8863 !isConstVecPow2(ConstVec, isSigned, C))
8866 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
8867 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
8868 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
8869 // These instructions only exist converting from f32 to i32. We can handle
8870 // smaller integers by generating an extra truncate, but larger ones would
8875 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8876 Intrinsic::arm_neon_vcvtfp2fxu;
8877 unsigned NumLanes = Op.getValueType().getVectorNumElements();
8878 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8879 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
8880 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8881 DAG.getConstant(Log2_64(C), MVT::i32));
8883 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
8884 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
8889 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8890 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8891 /// when the VDIV has a constant operand that is a power of 2.
8893 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8894 /// vcvt.f32.s32 d16, d16
8895 /// vdiv.f32 d16, d17, d16
8897 /// vcvt.f32.s32 d16, d16, #3
8898 static SDValue PerformVDIVCombine(SDNode *N,
8899 TargetLowering::DAGCombinerInfo &DCI,
8900 const ARMSubtarget *Subtarget) {
8901 SelectionDAG &DAG = DCI.DAG;
8902 SDValue Op = N->getOperand(0);
8903 unsigned OpOpcode = Op.getNode()->getOpcode();
8905 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8906 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8910 SDValue ConstVec = N->getOperand(1);
8911 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8913 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8914 !isConstVecPow2(ConstVec, isSigned, C))
8917 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
8918 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
8919 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
8920 // These instructions only exist converting from i32 to f32. We can handle
8921 // smaller integers by generating an extra extend, but larger ones would
8926 SDValue ConvInput = Op.getOperand(0);
8927 unsigned NumLanes = Op.getValueType().getVectorNumElements();
8928 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
8929 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
8930 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
8933 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8934 Intrinsic::arm_neon_vcvtfxu2fp;
8935 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8937 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8938 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
8941 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8942 /// operand of a vector shift operation, where all the elements of the
8943 /// build_vector must have the same constant integer value.
8944 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8945 // Ignore bit_converts.
8946 while (Op.getOpcode() == ISD::BITCAST)
8947 Op = Op.getOperand(0);
8948 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8949 APInt SplatBits, SplatUndef;
8950 unsigned SplatBitSize;
8952 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8953 HasAnyUndefs, ElementBits) ||
8954 SplatBitSize > ElementBits)
8956 Cnt = SplatBits.getSExtValue();
8960 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8961 /// operand of a vector shift left operation. That value must be in the range:
8962 /// 0 <= Value < ElementBits for a left shift; or
8963 /// 0 <= Value <= ElementBits for a long left shift.
8964 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8965 assert(VT.isVector() && "vector shift count is not a vector type");
8966 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8967 if (! getVShiftImm(Op, ElementBits, Cnt))
8969 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8972 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8973 /// operand of a vector shift right operation. For a shift opcode, the value
8974 /// is positive, but for an intrinsic the value count must be negative. The
8975 /// absolute value must be in the range:
8976 /// 1 <= |Value| <= ElementBits for a right shift; or
8977 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8978 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8980 assert(VT.isVector() && "vector shift count is not a vector type");
8981 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8982 if (! getVShiftImm(Op, ElementBits, Cnt))
8986 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8989 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8990 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8991 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8994 // Don't do anything for most intrinsics.
8997 // Vector shifts: check for immediate versions and lower them.
8998 // Note: This is done during DAG combining instead of DAG legalizing because
8999 // the build_vectors for 64-bit vector element shift counts are generally
9000 // not legal, and it is hard to see their values after they get legalized to
9001 // loads from a constant pool.
9002 case Intrinsic::arm_neon_vshifts:
9003 case Intrinsic::arm_neon_vshiftu:
9004 case Intrinsic::arm_neon_vrshifts:
9005 case Intrinsic::arm_neon_vrshiftu:
9006 case Intrinsic::arm_neon_vrshiftn:
9007 case Intrinsic::arm_neon_vqshifts:
9008 case Intrinsic::arm_neon_vqshiftu:
9009 case Intrinsic::arm_neon_vqshiftsu:
9010 case Intrinsic::arm_neon_vqshiftns:
9011 case Intrinsic::arm_neon_vqshiftnu:
9012 case Intrinsic::arm_neon_vqshiftnsu:
9013 case Intrinsic::arm_neon_vqrshiftns:
9014 case Intrinsic::arm_neon_vqrshiftnu:
9015 case Intrinsic::arm_neon_vqrshiftnsu: {
9016 EVT VT = N->getOperand(1).getValueType();
9018 unsigned VShiftOpc = 0;
9021 case Intrinsic::arm_neon_vshifts:
9022 case Intrinsic::arm_neon_vshiftu:
9023 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9024 VShiftOpc = ARMISD::VSHL;
9027 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9028 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9029 ARMISD::VSHRs : ARMISD::VSHRu);
9034 case Intrinsic::arm_neon_vrshifts:
9035 case Intrinsic::arm_neon_vrshiftu:
9036 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9040 case Intrinsic::arm_neon_vqshifts:
9041 case Intrinsic::arm_neon_vqshiftu:
9042 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9046 case Intrinsic::arm_neon_vqshiftsu:
9047 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9049 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9051 case Intrinsic::arm_neon_vrshiftn:
9052 case Intrinsic::arm_neon_vqshiftns:
9053 case Intrinsic::arm_neon_vqshiftnu:
9054 case Intrinsic::arm_neon_vqshiftnsu:
9055 case Intrinsic::arm_neon_vqrshiftns:
9056 case Intrinsic::arm_neon_vqrshiftnu:
9057 case Intrinsic::arm_neon_vqrshiftnsu:
9058 // Narrowing shifts require an immediate right shift.
9059 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9061 llvm_unreachable("invalid shift count for narrowing vector shift "
9065 llvm_unreachable("unhandled vector shift");
9069 case Intrinsic::arm_neon_vshifts:
9070 case Intrinsic::arm_neon_vshiftu:
9071 // Opcode already set above.
9073 case Intrinsic::arm_neon_vrshifts:
9074 VShiftOpc = ARMISD::VRSHRs; break;
9075 case Intrinsic::arm_neon_vrshiftu:
9076 VShiftOpc = ARMISD::VRSHRu; break;
9077 case Intrinsic::arm_neon_vrshiftn:
9078 VShiftOpc = ARMISD::VRSHRN; break;
9079 case Intrinsic::arm_neon_vqshifts:
9080 VShiftOpc = ARMISD::VQSHLs; break;
9081 case Intrinsic::arm_neon_vqshiftu:
9082 VShiftOpc = ARMISD::VQSHLu; break;
9083 case Intrinsic::arm_neon_vqshiftsu:
9084 VShiftOpc = ARMISD::VQSHLsu; break;
9085 case Intrinsic::arm_neon_vqshiftns:
9086 VShiftOpc = ARMISD::VQSHRNs; break;
9087 case Intrinsic::arm_neon_vqshiftnu:
9088 VShiftOpc = ARMISD::VQSHRNu; break;
9089 case Intrinsic::arm_neon_vqshiftnsu:
9090 VShiftOpc = ARMISD::VQSHRNsu; break;
9091 case Intrinsic::arm_neon_vqrshiftns:
9092 VShiftOpc = ARMISD::VQRSHRNs; break;
9093 case Intrinsic::arm_neon_vqrshiftnu:
9094 VShiftOpc = ARMISD::VQRSHRNu; break;
9095 case Intrinsic::arm_neon_vqrshiftnsu:
9096 VShiftOpc = ARMISD::VQRSHRNsu; break;
9099 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9100 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9103 case Intrinsic::arm_neon_vshiftins: {
9104 EVT VT = N->getOperand(1).getValueType();
9106 unsigned VShiftOpc = 0;
9108 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9109 VShiftOpc = ARMISD::VSLI;
9110 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9111 VShiftOpc = ARMISD::VSRI;
9113 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9116 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9117 N->getOperand(1), N->getOperand(2),
9118 DAG.getConstant(Cnt, MVT::i32));
9121 case Intrinsic::arm_neon_vqrshifts:
9122 case Intrinsic::arm_neon_vqrshiftu:
9123 // No immediate versions of these to check for.
9130 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9131 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9132 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9133 /// vector element shift counts are generally not legal, and it is hard to see
9134 /// their values after they get legalized to loads from a constant pool.
9135 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9136 const ARMSubtarget *ST) {
9137 EVT VT = N->getValueType(0);
9138 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9139 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9140 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9141 SDValue N1 = N->getOperand(1);
9142 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9143 SDValue N0 = N->getOperand(0);
9144 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9145 DAG.MaskedValueIsZero(N0.getOperand(0),
9146 APInt::getHighBitsSet(32, 16)))
9147 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9151 // Nothing to be done for scalar shifts.
9152 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9153 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9156 assert(ST->hasNEON() && "unexpected vector shift");
9159 switch (N->getOpcode()) {
9160 default: llvm_unreachable("unexpected shift opcode");
9163 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9164 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9165 DAG.getConstant(Cnt, MVT::i32));
9170 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9171 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9172 ARMISD::VSHRs : ARMISD::VSHRu);
9173 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9174 DAG.getConstant(Cnt, MVT::i32));
9180 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9181 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9182 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9183 const ARMSubtarget *ST) {
9184 SDValue N0 = N->getOperand(0);
9186 // Check for sign- and zero-extensions of vector extract operations of 8-
9187 // and 16-bit vector elements. NEON supports these directly. They are
9188 // handled during DAG combining because type legalization will promote them
9189 // to 32-bit types and it is messy to recognize the operations after that.
9190 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9191 SDValue Vec = N0.getOperand(0);
9192 SDValue Lane = N0.getOperand(1);
9193 EVT VT = N->getValueType(0);
9194 EVT EltVT = N0.getValueType();
9195 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9197 if (VT == MVT::i32 &&
9198 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9199 TLI.isTypeLegal(Vec.getValueType()) &&
9200 isa<ConstantSDNode>(Lane)) {
9203 switch (N->getOpcode()) {
9204 default: llvm_unreachable("unexpected opcode");
9205 case ISD::SIGN_EXTEND:
9206 Opc = ARMISD::VGETLANEs;
9208 case ISD::ZERO_EXTEND:
9209 case ISD::ANY_EXTEND:
9210 Opc = ARMISD::VGETLANEu;
9213 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9220 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9221 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9222 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9223 const ARMSubtarget *ST) {
9224 // If the target supports NEON, try to use vmax/vmin instructions for f32
9225 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9226 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9227 // a NaN; only do the transformation when it matches that behavior.
9229 // For now only do this when using NEON for FP operations; if using VFP, it
9230 // is not obvious that the benefit outweighs the cost of switching to the
9232 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9233 N->getValueType(0) != MVT::f32)
9236 SDValue CondLHS = N->getOperand(0);
9237 SDValue CondRHS = N->getOperand(1);
9238 SDValue LHS = N->getOperand(2);
9239 SDValue RHS = N->getOperand(3);
9240 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9242 unsigned Opcode = 0;
9244 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9245 IsReversed = false; // x CC y ? x : y
9246 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9247 IsReversed = true ; // x CC y ? y : x
9261 // If LHS is NaN, an ordered comparison will be false and the result will
9262 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9263 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9264 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9265 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9267 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9268 // will return -0, so vmin can only be used for unsafe math or if one of
9269 // the operands is known to be nonzero.
9270 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9271 !DAG.getTarget().Options.UnsafeFPMath &&
9272 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9274 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9283 // If LHS is NaN, an ordered comparison will be false and the result will
9284 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9285 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9286 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9287 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9289 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9290 // will return +0, so vmax can only be used for unsafe math or if one of
9291 // the operands is known to be nonzero.
9292 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9293 !DAG.getTarget().Options.UnsafeFPMath &&
9294 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9296 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9302 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9305 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9307 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9308 SDValue Cmp = N->getOperand(4);
9309 if (Cmp.getOpcode() != ARMISD::CMPZ)
9310 // Only looking at EQ and NE cases.
9313 EVT VT = N->getValueType(0);
9315 SDValue LHS = Cmp.getOperand(0);
9316 SDValue RHS = Cmp.getOperand(1);
9317 SDValue FalseVal = N->getOperand(0);
9318 SDValue TrueVal = N->getOperand(1);
9319 SDValue ARMcc = N->getOperand(2);
9320 ARMCC::CondCodes CC =
9321 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9339 /// FIXME: Turn this into a target neutral optimization?
9341 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9342 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9343 N->getOperand(3), Cmp);
9344 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9346 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9347 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9348 N->getOperand(3), NewCmp);
9351 if (Res.getNode()) {
9352 APInt KnownZero, KnownOne;
9353 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9354 // Capture demanded bits information that would be otherwise lost.
9355 if (KnownZero == 0xfffffffe)
9356 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9357 DAG.getValueType(MVT::i1));
9358 else if (KnownZero == 0xffffff00)
9359 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9360 DAG.getValueType(MVT::i8));
9361 else if (KnownZero == 0xffff0000)
9362 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9363 DAG.getValueType(MVT::i16));
9369 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9370 DAGCombinerInfo &DCI) const {
9371 switch (N->getOpcode()) {
9373 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9374 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9375 case ISD::SUB: return PerformSUBCombine(N, DCI);
9376 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9377 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9378 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9379 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9380 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9381 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9382 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9383 case ISD::STORE: return PerformSTORECombine(N, DCI);
9384 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9385 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9386 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9387 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9388 case ISD::FP_TO_SINT:
9389 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9390 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9391 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9394 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9395 case ISD::SIGN_EXTEND:
9396 case ISD::ZERO_EXTEND:
9397 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9398 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9399 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9400 case ARMISD::VLD2DUP:
9401 case ARMISD::VLD3DUP:
9402 case ARMISD::VLD4DUP:
9403 return CombineBaseUpdate(N, DCI);
9404 case ARMISD::BUILD_VECTOR:
9405 return PerformARMBUILD_VECTORCombine(N, DCI);
9406 case ISD::INTRINSIC_VOID:
9407 case ISD::INTRINSIC_W_CHAIN:
9408 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9409 case Intrinsic::arm_neon_vld1:
9410 case Intrinsic::arm_neon_vld2:
9411 case Intrinsic::arm_neon_vld3:
9412 case Intrinsic::arm_neon_vld4:
9413 case Intrinsic::arm_neon_vld2lane:
9414 case Intrinsic::arm_neon_vld3lane:
9415 case Intrinsic::arm_neon_vld4lane:
9416 case Intrinsic::arm_neon_vst1:
9417 case Intrinsic::arm_neon_vst2:
9418 case Intrinsic::arm_neon_vst3:
9419 case Intrinsic::arm_neon_vst4:
9420 case Intrinsic::arm_neon_vst2lane:
9421 case Intrinsic::arm_neon_vst3lane:
9422 case Intrinsic::arm_neon_vst4lane:
9423 return CombineBaseUpdate(N, DCI);
9431 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9433 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9436 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
9438 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9439 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9441 switch (VT.getSimpleVT().SimpleTy) {
9447 // Unaligned access can use (for example) LRDB, LRDH, LDR
9448 if (AllowsUnaligned) {
9450 *Fast = Subtarget->hasV7Ops();
9457 // For any little-endian targets with neon, we can support unaligned ld/st
9458 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9459 // A big-endian target may also explicitly support unaligned accesses
9460 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9470 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9471 unsigned AlignCheck) {
9472 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9473 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9476 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9477 unsigned DstAlign, unsigned SrcAlign,
9478 bool IsMemset, bool ZeroMemset,
9480 MachineFunction &MF) const {
9481 const Function *F = MF.getFunction();
9483 // See if we can use NEON instructions for this...
9484 if ((!IsMemset || ZeroMemset) &&
9485 Subtarget->hasNEON() &&
9486 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9487 Attribute::NoImplicitFloat)) {
9490 (memOpAlign(SrcAlign, DstAlign, 16) ||
9491 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
9493 } else if (Size >= 8 &&
9494 (memOpAlign(SrcAlign, DstAlign, 8) ||
9495 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
9500 // Lowering to i32/i16 if the size permits.
9506 // Let the target-independent logic figure it out.
9510 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9511 if (Val.getOpcode() != ISD::LOAD)
9514 EVT VT1 = Val.getValueType();
9515 if (!VT1.isSimple() || !VT1.isInteger() ||
9516 !VT2.isSimple() || !VT2.isInteger())
9519 switch (VT1.getSimpleVT().SimpleTy) {
9524 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9531 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9532 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9535 if (!isTypeLegal(EVT::getEVT(Ty1)))
9538 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9540 // Assuming the caller doesn't have a zeroext or signext return parameter,
9541 // truncation all the way down to i1 is valid.
9546 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9551 switch (VT.getSimpleVT().SimpleTy) {
9552 default: return false;
9567 if ((V & (Scale - 1)) != 0)
9570 return V == (V & ((1LL << 5) - 1));
9573 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9574 const ARMSubtarget *Subtarget) {
9581 switch (VT.getSimpleVT().SimpleTy) {
9582 default: return false;
9587 // + imm12 or - imm8
9589 return V == (V & ((1LL << 8) - 1));
9590 return V == (V & ((1LL << 12) - 1));
9593 // Same as ARM mode. FIXME: NEON?
9594 if (!Subtarget->hasVFP2())
9599 return V == (V & ((1LL << 8) - 1));
9603 /// isLegalAddressImmediate - Return true if the integer value can be used
9604 /// as the offset of the target addressing mode for load / store of the
9606 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9607 const ARMSubtarget *Subtarget) {
9614 if (Subtarget->isThumb1Only())
9615 return isLegalT1AddressImmediate(V, VT);
9616 else if (Subtarget->isThumb2())
9617 return isLegalT2AddressImmediate(V, VT, Subtarget);
9622 switch (VT.getSimpleVT().SimpleTy) {
9623 default: return false;
9628 return V == (V & ((1LL << 12) - 1));
9631 return V == (V & ((1LL << 8) - 1));
9634 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9639 return V == (V & ((1LL << 8) - 1));
9643 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9645 int Scale = AM.Scale;
9649 switch (VT.getSimpleVT().SimpleTy) {
9650 default: return false;
9659 return Scale == 2 || Scale == 4 || Scale == 8;
9662 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9666 // Note, we allow "void" uses (basically, uses that aren't loads or
9667 // stores), because arm allows folding a scale into many arithmetic
9668 // operations. This should be made more precise and revisited later.
9670 // Allow r << imm, but the imm has to be a multiple of two.
9671 if (Scale & 1) return false;
9672 return isPowerOf2_32(Scale);
9676 /// isLegalAddressingMode - Return true if the addressing mode represented
9677 /// by AM is legal for this target, for a load/store of the specified type.
9678 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9680 EVT VT = getValueType(Ty, true);
9681 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9684 // Can never fold addr of global into load/store.
9689 case 0: // no scale reg, must be "r+i" or "r", or "i".
9692 if (Subtarget->isThumb1Only())
9696 // ARM doesn't support any R+R*scale+imm addr modes.
9703 if (Subtarget->isThumb2())
9704 return isLegalT2ScaledAddressingMode(AM, VT);
9706 int Scale = AM.Scale;
9707 switch (VT.getSimpleVT().SimpleTy) {
9708 default: return false;
9712 if (Scale < 0) Scale = -Scale;
9716 return isPowerOf2_32(Scale & ~1);
9720 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9725 // Note, we allow "void" uses (basically, uses that aren't loads or
9726 // stores), because arm allows folding a scale into many arithmetic
9727 // operations. This should be made more precise and revisited later.
9729 // Allow r << imm, but the imm has to be a multiple of two.
9730 if (Scale & 1) return false;
9731 return isPowerOf2_32(Scale);
9737 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9738 /// icmp immediate, that is the target has icmp instructions which can compare
9739 /// a register against the immediate without having to materialize the
9740 /// immediate into a register.
9741 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9742 // Thumb2 and ARM modes can use cmn for negative immediates.
9743 if (!Subtarget->isThumb())
9744 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9745 if (Subtarget->isThumb2())
9746 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9747 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9748 return Imm >= 0 && Imm <= 255;
9751 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9752 /// *or sub* immediate, that is the target has add or sub instructions which can
9753 /// add a register with the immediate without having to materialize the
9754 /// immediate into a register.
9755 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9756 // Same encoding for add/sub, just flip the sign.
9757 int64_t AbsImm = llvm::abs64(Imm);
9758 if (!Subtarget->isThumb())
9759 return ARM_AM::getSOImmVal(AbsImm) != -1;
9760 if (Subtarget->isThumb2())
9761 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9762 // Thumb1 only has 8-bit unsigned immediate.
9763 return AbsImm >= 0 && AbsImm <= 255;
9766 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9767 bool isSEXTLoad, SDValue &Base,
9768 SDValue &Offset, bool &isInc,
9769 SelectionDAG &DAG) {
9770 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9773 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9775 Base = Ptr->getOperand(0);
9776 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9777 int RHSC = (int)RHS->getZExtValue();
9778 if (RHSC < 0 && RHSC > -256) {
9779 assert(Ptr->getOpcode() == ISD::ADD);
9781 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9785 isInc = (Ptr->getOpcode() == ISD::ADD);
9786 Offset = Ptr->getOperand(1);
9788 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9790 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9791 int RHSC = (int)RHS->getZExtValue();
9792 if (RHSC < 0 && RHSC > -0x1000) {
9793 assert(Ptr->getOpcode() == ISD::ADD);
9795 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9796 Base = Ptr->getOperand(0);
9801 if (Ptr->getOpcode() == ISD::ADD) {
9803 ARM_AM::ShiftOpc ShOpcVal=
9804 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9805 if (ShOpcVal != ARM_AM::no_shift) {
9806 Base = Ptr->getOperand(1);
9807 Offset = Ptr->getOperand(0);
9809 Base = Ptr->getOperand(0);
9810 Offset = Ptr->getOperand(1);
9815 isInc = (Ptr->getOpcode() == ISD::ADD);
9816 Base = Ptr->getOperand(0);
9817 Offset = Ptr->getOperand(1);
9821 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9825 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9826 bool isSEXTLoad, SDValue &Base,
9827 SDValue &Offset, bool &isInc,
9828 SelectionDAG &DAG) {
9829 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9832 Base = Ptr->getOperand(0);
9833 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9834 int RHSC = (int)RHS->getZExtValue();
9835 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9836 assert(Ptr->getOpcode() == ISD::ADD);
9838 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9840 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9841 isInc = Ptr->getOpcode() == ISD::ADD;
9842 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9850 /// getPreIndexedAddressParts - returns true by value, base pointer and
9851 /// offset pointer and addressing mode by reference if the node's address
9852 /// can be legally represented as pre-indexed load / store address.
9854 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9856 ISD::MemIndexedMode &AM,
9857 SelectionDAG &DAG) const {
9858 if (Subtarget->isThumb1Only())
9863 bool isSEXTLoad = false;
9864 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9865 Ptr = LD->getBasePtr();
9866 VT = LD->getMemoryVT();
9867 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9868 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9869 Ptr = ST->getBasePtr();
9870 VT = ST->getMemoryVT();
9875 bool isLegal = false;
9876 if (Subtarget->isThumb2())
9877 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9878 Offset, isInc, DAG);
9880 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9881 Offset, isInc, DAG);
9885 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9889 /// getPostIndexedAddressParts - returns true by value, base pointer and
9890 /// offset pointer and addressing mode by reference if this node can be
9891 /// combined with a load / store to form a post-indexed load / store.
9892 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9895 ISD::MemIndexedMode &AM,
9896 SelectionDAG &DAG) const {
9897 if (Subtarget->isThumb1Only())
9902 bool isSEXTLoad = false;
9903 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9904 VT = LD->getMemoryVT();
9905 Ptr = LD->getBasePtr();
9906 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9907 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9908 VT = ST->getMemoryVT();
9909 Ptr = ST->getBasePtr();
9914 bool isLegal = false;
9915 if (Subtarget->isThumb2())
9916 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9919 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9925 // Swap base ptr and offset to catch more post-index load / store when
9926 // it's legal. In Thumb2 mode, offset must be an immediate.
9927 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9928 !Subtarget->isThumb2())
9929 std::swap(Base, Offset);
9931 // Post-indexed load / store update the base pointer.
9936 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9940 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9943 const SelectionDAG &DAG,
9944 unsigned Depth) const {
9945 unsigned BitWidth = KnownOne.getBitWidth();
9946 KnownZero = KnownOne = APInt(BitWidth, 0);
9947 switch (Op.getOpcode()) {
9953 // These nodes' second result is a boolean
9954 if (Op.getResNo() == 0)
9956 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
9958 case ARMISD::CMOV: {
9959 // Bits are known zero/one if known on the LHS and RHS.
9960 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9961 if (KnownZero == 0 && KnownOne == 0) return;
9963 APInt KnownZeroRHS, KnownOneRHS;
9964 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9965 KnownZero &= KnownZeroRHS;
9966 KnownOne &= KnownOneRHS;
9969 case ISD::INTRINSIC_W_CHAIN: {
9970 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
9971 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
9974 case Intrinsic::arm_ldaex:
9975 case Intrinsic::arm_ldrex: {
9976 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
9977 unsigned MemBits = VT.getScalarType().getSizeInBits();
9978 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
9986 //===----------------------------------------------------------------------===//
9987 // ARM Inline Assembly Support
9988 //===----------------------------------------------------------------------===//
9990 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9991 // Looking for "rev" which is V6+.
9992 if (!Subtarget->hasV6Ops())
9995 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9996 std::string AsmStr = IA->getAsmString();
9997 SmallVector<StringRef, 4> AsmPieces;
9998 SplitString(AsmStr, AsmPieces, ";\n");
10000 switch (AsmPieces.size()) {
10001 default: return false;
10003 AsmStr = AsmPieces[0];
10005 SplitString(AsmStr, AsmPieces, " \t,");
10008 if (AsmPieces.size() == 3 &&
10009 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10010 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10011 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10012 if (Ty && Ty->getBitWidth() == 32)
10013 return IntrinsicLowering::LowerToByteSwap(CI);
10021 /// getConstraintType - Given a constraint letter, return the type of
10022 /// constraint it is for this target.
10023 ARMTargetLowering::ConstraintType
10024 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10025 if (Constraint.size() == 1) {
10026 switch (Constraint[0]) {
10028 case 'l': return C_RegisterClass;
10029 case 'w': return C_RegisterClass;
10030 case 'h': return C_RegisterClass;
10031 case 'x': return C_RegisterClass;
10032 case 't': return C_RegisterClass;
10033 case 'j': return C_Other; // Constant for movw.
10034 // An address with a single base register. Due to the way we
10035 // currently handle addresses it is the same as an 'r' memory constraint.
10036 case 'Q': return C_Memory;
10038 } else if (Constraint.size() == 2) {
10039 switch (Constraint[0]) {
10041 // All 'U+' constraints are addresses.
10042 case 'U': return C_Memory;
10045 return TargetLowering::getConstraintType(Constraint);
10048 /// Examine constraint type and operand type and determine a weight value.
10049 /// This object must already have been set up with the operand type
10050 /// and the current alternative constraint selected.
10051 TargetLowering::ConstraintWeight
10052 ARMTargetLowering::getSingleConstraintMatchWeight(
10053 AsmOperandInfo &info, const char *constraint) const {
10054 ConstraintWeight weight = CW_Invalid;
10055 Value *CallOperandVal = info.CallOperandVal;
10056 // If we don't have a value, we can't do a match,
10057 // but allow it at the lowest weight.
10058 if (CallOperandVal == NULL)
10060 Type *type = CallOperandVal->getType();
10061 // Look at the constraint type.
10062 switch (*constraint) {
10064 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10067 if (type->isIntegerTy()) {
10068 if (Subtarget->isThumb())
10069 weight = CW_SpecificReg;
10071 weight = CW_Register;
10075 if (type->isFloatingPointTy())
10076 weight = CW_Register;
10082 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10084 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10086 if (Constraint.size() == 1) {
10087 // GCC ARM Constraint Letters
10088 switch (Constraint[0]) {
10089 case 'l': // Low regs or general regs.
10090 if (Subtarget->isThumb())
10091 return RCPair(0U, &ARM::tGPRRegClass);
10092 return RCPair(0U, &ARM::GPRRegClass);
10093 case 'h': // High regs or no regs.
10094 if (Subtarget->isThumb())
10095 return RCPair(0U, &ARM::hGPRRegClass);
10098 return RCPair(0U, &ARM::GPRRegClass);
10100 if (VT == MVT::Other)
10102 if (VT == MVT::f32)
10103 return RCPair(0U, &ARM::SPRRegClass);
10104 if (VT.getSizeInBits() == 64)
10105 return RCPair(0U, &ARM::DPRRegClass);
10106 if (VT.getSizeInBits() == 128)
10107 return RCPair(0U, &ARM::QPRRegClass);
10110 if (VT == MVT::Other)
10112 if (VT == MVT::f32)
10113 return RCPair(0U, &ARM::SPR_8RegClass);
10114 if (VT.getSizeInBits() == 64)
10115 return RCPair(0U, &ARM::DPR_8RegClass);
10116 if (VT.getSizeInBits() == 128)
10117 return RCPair(0U, &ARM::QPR_8RegClass);
10120 if (VT == MVT::f32)
10121 return RCPair(0U, &ARM::SPRRegClass);
10125 if (StringRef("{cc}").equals_lower(Constraint))
10126 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10128 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10131 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10132 /// vector. If it is invalid, don't add anything to Ops.
10133 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10134 std::string &Constraint,
10135 std::vector<SDValue>&Ops,
10136 SelectionDAG &DAG) const {
10137 SDValue Result(0, 0);
10139 // Currently only support length 1 constraints.
10140 if (Constraint.length() != 1) return;
10142 char ConstraintLetter = Constraint[0];
10143 switch (ConstraintLetter) {
10146 case 'I': case 'J': case 'K': case 'L':
10147 case 'M': case 'N': case 'O':
10148 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10152 int64_t CVal64 = C->getSExtValue();
10153 int CVal = (int) CVal64;
10154 // None of these constraints allow values larger than 32 bits. Check
10155 // that the value fits in an int.
10156 if (CVal != CVal64)
10159 switch (ConstraintLetter) {
10161 // Constant suitable for movw, must be between 0 and
10163 if (Subtarget->hasV6T2Ops())
10164 if (CVal >= 0 && CVal <= 65535)
10168 if (Subtarget->isThumb1Only()) {
10169 // This must be a constant between 0 and 255, for ADD
10171 if (CVal >= 0 && CVal <= 255)
10173 } else if (Subtarget->isThumb2()) {
10174 // A constant that can be used as an immediate value in a
10175 // data-processing instruction.
10176 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10179 // A constant that can be used as an immediate value in a
10180 // data-processing instruction.
10181 if (ARM_AM::getSOImmVal(CVal) != -1)
10187 if (Subtarget->isThumb()) { // FIXME thumb2
10188 // This must be a constant between -255 and -1, for negated ADD
10189 // immediates. This can be used in GCC with an "n" modifier that
10190 // prints the negated value, for use with SUB instructions. It is
10191 // not useful otherwise but is implemented for compatibility.
10192 if (CVal >= -255 && CVal <= -1)
10195 // This must be a constant between -4095 and 4095. It is not clear
10196 // what this constraint is intended for. Implemented for
10197 // compatibility with GCC.
10198 if (CVal >= -4095 && CVal <= 4095)
10204 if (Subtarget->isThumb1Only()) {
10205 // A 32-bit value where only one byte has a nonzero value. Exclude
10206 // zero to match GCC. This constraint is used by GCC internally for
10207 // constants that can be loaded with a move/shift combination.
10208 // It is not useful otherwise but is implemented for compatibility.
10209 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10211 } else if (Subtarget->isThumb2()) {
10212 // A constant whose bitwise inverse can be used as an immediate
10213 // value in a data-processing instruction. This can be used in GCC
10214 // with a "B" modifier that prints the inverted value, for use with
10215 // BIC and MVN instructions. It is not useful otherwise but is
10216 // implemented for compatibility.
10217 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10220 // A constant whose bitwise inverse can be used as an immediate
10221 // value in a data-processing instruction. This can be used in GCC
10222 // with a "B" modifier that prints the inverted value, for use with
10223 // BIC and MVN instructions. It is not useful otherwise but is
10224 // implemented for compatibility.
10225 if (ARM_AM::getSOImmVal(~CVal) != -1)
10231 if (Subtarget->isThumb1Only()) {
10232 // This must be a constant between -7 and 7,
10233 // for 3-operand ADD/SUB immediate instructions.
10234 if (CVal >= -7 && CVal < 7)
10236 } else if (Subtarget->isThumb2()) {
10237 // A constant whose negation can be used as an immediate value in a
10238 // data-processing instruction. This can be used in GCC with an "n"
10239 // modifier that prints the negated value, for use with SUB
10240 // instructions. It is not useful otherwise but is implemented for
10242 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10245 // A constant whose negation can be used as an immediate value in a
10246 // data-processing instruction. This can be used in GCC with an "n"
10247 // modifier that prints the negated value, for use with SUB
10248 // instructions. It is not useful otherwise but is implemented for
10250 if (ARM_AM::getSOImmVal(-CVal) != -1)
10256 if (Subtarget->isThumb()) { // FIXME thumb2
10257 // This must be a multiple of 4 between 0 and 1020, for
10258 // ADD sp + immediate.
10259 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10262 // A power of two or a constant between 0 and 32. This is used in
10263 // GCC for the shift amount on shifted register operands, but it is
10264 // useful in general for any shift amounts.
10265 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10271 if (Subtarget->isThumb()) { // FIXME thumb2
10272 // This must be a constant between 0 and 31, for shift amounts.
10273 if (CVal >= 0 && CVal <= 31)
10279 if (Subtarget->isThumb()) { // FIXME thumb2
10280 // This must be a multiple of 4 between -508 and 508, for
10281 // ADD/SUB sp = sp + immediate.
10282 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10287 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10291 if (Result.getNode()) {
10292 Ops.push_back(Result);
10295 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10298 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10299 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10300 unsigned Opcode = Op->getOpcode();
10301 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10302 "Invalid opcode for Div/Rem lowering");
10303 bool isSigned = (Opcode == ISD::SDIVREM);
10304 EVT VT = Op->getValueType(0);
10305 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10308 switch (VT.getSimpleVT().SimpleTy) {
10309 default: llvm_unreachable("Unexpected request for libcall!");
10310 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10311 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10312 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10313 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10316 SDValue InChain = DAG.getEntryNode();
10318 TargetLowering::ArgListTy Args;
10319 TargetLowering::ArgListEntry Entry;
10320 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10321 EVT ArgVT = Op->getOperand(i).getValueType();
10322 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10323 Entry.Node = Op->getOperand(i);
10325 Entry.isSExt = isSigned;
10326 Entry.isZExt = !isSigned;
10327 Args.push_back(Entry);
10330 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10333 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10337 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
10338 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
10339 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
10340 Callee, Args, DAG, dl);
10341 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10343 return CallInfo.first;
10347 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10348 // The ARM target isn't yet aware of offsets.
10352 bool ARM::isBitFieldInvertedMask(unsigned v) {
10353 if (v == 0xffffffff)
10356 // there can be 1's on either or both "outsides", all the "inside"
10357 // bits must be 0's
10358 unsigned TO = CountTrailingOnes_32(v);
10359 unsigned LO = CountLeadingOnes_32(v);
10360 v = (v >> TO) << TO;
10361 v = (v << LO) >> LO;
10365 /// isFPImmLegal - Returns true if the target can instruction select the
10366 /// specified FP immediate natively. If false, the legalizer will
10367 /// materialize the FP immediate as a load from a constant pool.
10368 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10369 if (!Subtarget->hasVFP3())
10371 if (VT == MVT::f32)
10372 return ARM_AM::getFP32Imm(Imm) != -1;
10373 if (VT == MVT::f64)
10374 return ARM_AM::getFP64Imm(Imm) != -1;
10378 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10379 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10380 /// specified in the intrinsic calls.
10381 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10383 unsigned Intrinsic) const {
10384 switch (Intrinsic) {
10385 case Intrinsic::arm_neon_vld1:
10386 case Intrinsic::arm_neon_vld2:
10387 case Intrinsic::arm_neon_vld3:
10388 case Intrinsic::arm_neon_vld4:
10389 case Intrinsic::arm_neon_vld2lane:
10390 case Intrinsic::arm_neon_vld3lane:
10391 case Intrinsic::arm_neon_vld4lane: {
10392 Info.opc = ISD::INTRINSIC_W_CHAIN;
10393 // Conservatively set memVT to the entire set of vectors loaded.
10394 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10395 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10396 Info.ptrVal = I.getArgOperand(0);
10398 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10399 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10400 Info.vol = false; // volatile loads with NEON intrinsics not supported
10401 Info.readMem = true;
10402 Info.writeMem = false;
10405 case Intrinsic::arm_neon_vst1:
10406 case Intrinsic::arm_neon_vst2:
10407 case Intrinsic::arm_neon_vst3:
10408 case Intrinsic::arm_neon_vst4:
10409 case Intrinsic::arm_neon_vst2lane:
10410 case Intrinsic::arm_neon_vst3lane:
10411 case Intrinsic::arm_neon_vst4lane: {
10412 Info.opc = ISD::INTRINSIC_VOID;
10413 // Conservatively set memVT to the entire set of vectors stored.
10414 unsigned NumElts = 0;
10415 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10416 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10417 if (!ArgTy->isVectorTy())
10419 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10421 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10422 Info.ptrVal = I.getArgOperand(0);
10424 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10425 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10426 Info.vol = false; // volatile stores with NEON intrinsics not supported
10427 Info.readMem = false;
10428 Info.writeMem = true;
10431 case Intrinsic::arm_ldaex:
10432 case Intrinsic::arm_ldrex: {
10433 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10434 Info.opc = ISD::INTRINSIC_W_CHAIN;
10435 Info.memVT = MVT::getVT(PtrTy->getElementType());
10436 Info.ptrVal = I.getArgOperand(0);
10438 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10440 Info.readMem = true;
10441 Info.writeMem = false;
10444 case Intrinsic::arm_stlex:
10445 case Intrinsic::arm_strex: {
10446 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10447 Info.opc = ISD::INTRINSIC_W_CHAIN;
10448 Info.memVT = MVT::getVT(PtrTy->getElementType());
10449 Info.ptrVal = I.getArgOperand(1);
10451 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10453 Info.readMem = false;
10454 Info.writeMem = true;
10457 case Intrinsic::arm_stlexd:
10458 case Intrinsic::arm_strexd: {
10459 Info.opc = ISD::INTRINSIC_W_CHAIN;
10460 Info.memVT = MVT::i64;
10461 Info.ptrVal = I.getArgOperand(2);
10465 Info.readMem = false;
10466 Info.writeMem = true;
10469 case Intrinsic::arm_ldaexd:
10470 case Intrinsic::arm_ldrexd: {
10471 Info.opc = ISD::INTRINSIC_W_CHAIN;
10472 Info.memVT = MVT::i64;
10473 Info.ptrVal = I.getArgOperand(0);
10477 Info.readMem = true;
10478 Info.writeMem = false;
10488 /// \brief Returns true if it is beneficial to convert a load of a constant
10489 /// to just the constant itself.
10490 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10492 assert(Ty->isIntegerTy());
10494 unsigned Bits = Ty->getPrimitiveSizeInBits();
10495 if (Bits == 0 || Bits > 32)
10500 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10501 // Loads and stores less than 64-bits are already atomic; ones above that
10502 // are doomed anyway, so defer to the default libcall and blame the OS when
10503 // things go wrong:
10504 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
10505 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 64;
10506 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
10507 return LI->getType()->getPrimitiveSizeInBits() == 64;
10509 // For the real atomic operations, we have ldrex/strex up to 64 bits.
10510 return Inst->getType()->getPrimitiveSizeInBits() <= 64;
10513 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10514 AtomicOrdering Ord) const {
10515 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10516 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10518 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10520 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
10521 // intrinsic must return {i32, i32} and we have to recombine them into a
10522 // single i64 here.
10523 if (ValTy->getPrimitiveSizeInBits() == 64) {
10524 Intrinsic::ID Int =
10525 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
10526 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
10528 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10529 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
10531 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10532 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10533 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10534 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10535 return Builder.CreateOr(
10536 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
10539 Type *Tys[] = { Addr->getType() };
10540 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
10541 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10543 return Builder.CreateTruncOrBitCast(
10544 Builder.CreateCall(Ldrex, Addr),
10545 cast<PointerType>(Addr->getType())->getElementType());
10548 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
10550 AtomicOrdering Ord) const {
10551 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10553 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10555 // Since the intrinsics must have legal type, the i64 intrinsics take two
10556 // parameters: "i32, i32". We must marshal Val into the appropriate form
10557 // before the call.
10558 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
10559 Intrinsic::ID Int =
10560 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
10561 Function *Strex = Intrinsic::getDeclaration(M, Int);
10562 Type *Int32Ty = Type::getInt32Ty(M->getContext());
10564 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
10565 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
10566 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10567 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
10570 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
10571 Type *Tys[] = { Addr->getType() };
10572 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
10574 return Builder.CreateCall2(
10575 Strex, Builder.CreateZExtOrBitCast(
10576 Val, Strex->getFunctionType()->getParamType(0)),