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 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMSubtarget.h"
22 #include "ARMTargetMachine.h"
23 #include "ARMTargetObjectFile.h"
24 #include "MCTargetDesc/ARMAddressingModes.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/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 STATISTIC(NumTailCalls, "Number of tail calls");
53 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
54 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 EnableARMLongCalls("arm-long-calls", cl::Hidden,
58 cl::desc("Generate calls via indirect call instructions"),
62 ARMInterworking("arm-interworking", cl::Hidden,
63 cl::desc("Enable / disable ARM interworking (for debugging only)"),
67 class ARMCCState : public CCState {
69 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
70 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs,
71 LLVMContext &C, ParmContext PC)
72 : CCState(CC, isVarArg, MF, TM, locs, C) {
73 assert(((PC == Call) || (PC == Prologue)) &&
74 "ARMCCState users must specify whether their context is call"
75 "or prologue generation.");
81 // The APCS parameter registers.
82 static const uint16_t GPRArgRegs[] = {
83 ARM::R0, ARM::R1, ARM::R2, ARM::R3
86 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
87 MVT PromotedBitwiseVT) {
88 if (VT != PromotedLdStVT) {
89 setOperationAction(ISD::LOAD, VT, Promote);
90 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
92 setOperationAction(ISD::STORE, VT, Promote);
93 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
96 MVT ElemTy = VT.getVectorElementType();
97 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
98 setOperationAction(ISD::SETCC, VT, Custom);
99 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
100 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
101 if (ElemTy == MVT::i32) {
102 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
103 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
104 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
105 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
107 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
108 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
109 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
110 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
112 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
113 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
114 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
115 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
116 setOperationAction(ISD::SELECT, VT, Expand);
117 setOperationAction(ISD::SELECT_CC, VT, Expand);
118 setOperationAction(ISD::VSELECT, VT, Expand);
119 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
120 if (VT.isInteger()) {
121 setOperationAction(ISD::SHL, VT, Custom);
122 setOperationAction(ISD::SRA, VT, Custom);
123 setOperationAction(ISD::SRL, VT, Custom);
126 // Promote all bit-wise operations.
127 if (VT.isInteger() && VT != PromotedBitwiseVT) {
128 setOperationAction(ISD::AND, VT, Promote);
129 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
130 setOperationAction(ISD::OR, VT, Promote);
131 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::XOR, VT, Promote);
133 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
136 // Neon does not support vector divide/remainder operations.
137 setOperationAction(ISD::SDIV, VT, Expand);
138 setOperationAction(ISD::UDIV, VT, Expand);
139 setOperationAction(ISD::FDIV, VT, Expand);
140 setOperationAction(ISD::SREM, VT, Expand);
141 setOperationAction(ISD::UREM, VT, Expand);
142 setOperationAction(ISD::FREM, VT, Expand);
145 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
146 addRegisterClass(VT, &ARM::DPRRegClass);
147 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
150 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
151 addRegisterClass(VT, &ARM::DPairRegClass);
152 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
155 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
156 if (TM.getSubtarget<ARMSubtarget>().isTargetMachO())
157 return new TargetLoweringObjectFileMachO();
159 return new ARMElfTargetObjectFile();
162 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
163 : TargetLowering(TM, createTLOF(TM)) {
164 Subtarget = &TM.getSubtarget<ARMSubtarget>();
165 RegInfo = TM.getRegisterInfo();
166 Itins = TM.getInstrItineraryData();
168 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
170 if (Subtarget->isTargetMachO()) {
171 // Uses VFP for Thumb libfuncs if available.
172 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
173 Subtarget->hasARMOps()) {
174 // Single-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
176 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
177 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
178 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
180 // Double-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
182 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
183 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
184 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
186 // Single-precision comparisons.
187 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
188 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
189 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
190 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
191 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
192 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
193 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
194 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
196 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
205 // Double-precision comparisons.
206 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
207 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
208 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
209 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
210 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
211 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
212 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
213 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
215 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
224 // Floating-point to integer conversions.
225 // i64 conversions are done via library routines even when generating VFP
226 // instructions, so use the same ones.
227 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
228 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
229 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
232 // Conversions between floating types.
233 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
234 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
236 // Integer to floating-point conversions.
237 // i64 conversions are done via library routines even when generating VFP
238 // instructions, so use the same ones.
239 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
240 // e.g., __floatunsidf vs. __floatunssidfvfp.
241 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
242 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
243 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
248 // These libcalls are not available in 32-bit.
249 setLibcallName(RTLIB::SHL_I128, 0);
250 setLibcallName(RTLIB::SRL_I128, 0);
251 setLibcallName(RTLIB::SRA_I128, 0);
253 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO()) {
254 // Double-precision floating-point arithmetic helper functions
255 // RTABI chapter 4.1.2, Table 2
256 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
257 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
258 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
259 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
260 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
261 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
262 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
263 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
265 // Double-precision floating-point comparison helper functions
266 // RTABI chapter 4.1.2, Table 3
267 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
268 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
269 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
270 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
271 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
272 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
273 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
274 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
275 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
276 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
277 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
278 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
280 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
281 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
282 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
283 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
284 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
285 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
286 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
287 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
292 // Single-precision floating-point arithmetic helper functions
293 // RTABI chapter 4.1.2, Table 4
294 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
295 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
296 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
297 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
298 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
301 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
303 // Single-precision floating-point comparison helper functions
304 // RTABI chapter 4.1.2, Table 5
305 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
306 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
307 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
308 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
309 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
310 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
311 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
312 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
313 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
314 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
315 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
316 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
318 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
319 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
320 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
321 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
322 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
323 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
324 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
325 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
326 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
330 // Floating-point to integer conversions.
331 // RTABI chapter 4.1.2, Table 6
332 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
333 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
334 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
335 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
336 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
337 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
338 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
339 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
340 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
341 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
342 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
343 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
344 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
345 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
346 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
349 // Conversions between floating types.
350 // RTABI chapter 4.1.2, Table 7
351 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
352 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
353 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
356 // Integer to floating-point conversions.
357 // RTABI chapter 4.1.2, Table 8
358 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
359 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
360 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
361 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
362 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
363 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
364 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
365 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
366 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
367 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
368 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
369 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
370 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
375 // Long long helper functions
376 // RTABI chapter 4.2, Table 9
377 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
378 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
379 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
380 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
381 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
388 // Integer division functions
389 // RTABI chapter 4.3.1
390 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
391 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
392 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
393 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
394 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
395 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
396 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
397 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
398 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
399 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
400 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
401 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
402 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
403 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
404 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
408 // RTABI chapter 4.3.4
409 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
410 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
411 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
412 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
417 // Use divmod compiler-rt calls for iOS 5.0 and later.
418 if (Subtarget->getTargetTriple().isiOS() &&
419 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
420 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
421 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
424 if (Subtarget->isThumb1Only())
425 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
427 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
428 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
429 !Subtarget->isThumb1Only()) {
430 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
431 if (!Subtarget->isFPOnlySP())
432 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
434 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
437 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
438 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
439 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
440 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
441 setTruncStoreAction((MVT::SimpleValueType)VT,
442 (MVT::SimpleValueType)InnerVT, Expand);
443 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
444 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
445 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
448 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
449 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
451 if (Subtarget->hasNEON()) {
452 addDRTypeForNEON(MVT::v2f32);
453 addDRTypeForNEON(MVT::v8i8);
454 addDRTypeForNEON(MVT::v4i16);
455 addDRTypeForNEON(MVT::v2i32);
456 addDRTypeForNEON(MVT::v1i64);
458 addQRTypeForNEON(MVT::v4f32);
459 addQRTypeForNEON(MVT::v2f64);
460 addQRTypeForNEON(MVT::v16i8);
461 addQRTypeForNEON(MVT::v8i16);
462 addQRTypeForNEON(MVT::v4i32);
463 addQRTypeForNEON(MVT::v2i64);
465 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
466 // neither Neon nor VFP support any arithmetic operations on it.
467 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
468 // supported for v4f32.
469 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
470 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
471 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
472 // FIXME: Code duplication: FDIV and FREM are expanded always, see
473 // ARMTargetLowering::addTypeForNEON method for details.
474 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
475 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
476 // FIXME: Create unittest.
477 // In another words, find a way when "copysign" appears in DAG with vector
479 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
480 // FIXME: Code duplication: SETCC has custom operation action, see
481 // ARMTargetLowering::addTypeForNEON method for details.
482 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
483 // FIXME: Create unittest for FNEG and for FABS.
484 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
485 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
486 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
487 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
488 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
489 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
490 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
491 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
492 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
493 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
494 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
495 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
496 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
497 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
498 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
499 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
500 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
501 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
502 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
504 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
505 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
506 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
507 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
508 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
509 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
510 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
511 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
512 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
513 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
514 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
515 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
516 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
517 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
518 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
520 // Mark v2f32 intrinsics.
521 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
522 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
523 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
524 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
525 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
526 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
527 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
528 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
529 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
530 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
531 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
532 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
533 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
534 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
535 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
537 // Neon does not support some operations on v1i64 and v2i64 types.
538 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
539 // Custom handling for some quad-vector types to detect VMULL.
540 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
541 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
542 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
543 // Custom handling for some vector types to avoid expensive expansions
544 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
545 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
546 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
547 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
548 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
549 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
550 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
551 // a destination type that is wider than the source, and nor does
552 // it have a FP_TO_[SU]INT instruction with a narrower destination than
554 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
555 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
556 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
557 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
559 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
560 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
562 // NEON does not have single instruction CTPOP for vectors with element
563 // types wider than 8-bits. However, custom lowering can leverage the
564 // v8i8/v16i8 vcnt instruction.
565 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
566 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
567 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
568 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
570 // NEON only has FMA instructions as of VFP4.
571 if (!Subtarget->hasVFP4()) {
572 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
573 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
576 setTargetDAGCombine(ISD::INTRINSIC_VOID);
577 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
578 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
579 setTargetDAGCombine(ISD::SHL);
580 setTargetDAGCombine(ISD::SRL);
581 setTargetDAGCombine(ISD::SRA);
582 setTargetDAGCombine(ISD::SIGN_EXTEND);
583 setTargetDAGCombine(ISD::ZERO_EXTEND);
584 setTargetDAGCombine(ISD::ANY_EXTEND);
585 setTargetDAGCombine(ISD::SELECT_CC);
586 setTargetDAGCombine(ISD::BUILD_VECTOR);
587 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
588 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
589 setTargetDAGCombine(ISD::STORE);
590 setTargetDAGCombine(ISD::FP_TO_SINT);
591 setTargetDAGCombine(ISD::FP_TO_UINT);
592 setTargetDAGCombine(ISD::FDIV);
594 // It is legal to extload from v4i8 to v4i16 or v4i32.
595 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
596 MVT::v4i16, MVT::v2i16,
598 for (unsigned i = 0; i < 6; ++i) {
599 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
600 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
601 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
605 // ARM and Thumb2 support UMLAL/SMLAL.
606 if (!Subtarget->isThumb1Only())
607 setTargetDAGCombine(ISD::ADDC);
610 computeRegisterProperties();
612 // ARM does not have f32 extending load.
613 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
615 // ARM does not have i1 sign extending load.
616 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
618 // ARM supports all 4 flavors of integer indexed load / store.
619 if (!Subtarget->isThumb1Only()) {
620 for (unsigned im = (unsigned)ISD::PRE_INC;
621 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
622 setIndexedLoadAction(im, MVT::i1, Legal);
623 setIndexedLoadAction(im, MVT::i8, Legal);
624 setIndexedLoadAction(im, MVT::i16, Legal);
625 setIndexedLoadAction(im, MVT::i32, Legal);
626 setIndexedStoreAction(im, MVT::i1, Legal);
627 setIndexedStoreAction(im, MVT::i8, Legal);
628 setIndexedStoreAction(im, MVT::i16, Legal);
629 setIndexedStoreAction(im, MVT::i32, Legal);
633 // i64 operation support.
634 setOperationAction(ISD::MUL, MVT::i64, Expand);
635 setOperationAction(ISD::MULHU, MVT::i32, Expand);
636 if (Subtarget->isThumb1Only()) {
637 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
638 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
640 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
641 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
642 setOperationAction(ISD::MULHS, MVT::i32, Expand);
644 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
645 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
646 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
647 setOperationAction(ISD::SRL, MVT::i64, Custom);
648 setOperationAction(ISD::SRA, MVT::i64, Custom);
650 if (!Subtarget->isThumb1Only()) {
651 // FIXME: We should do this for Thumb1 as well.
652 setOperationAction(ISD::ADDC, MVT::i32, Custom);
653 setOperationAction(ISD::ADDE, MVT::i32, Custom);
654 setOperationAction(ISD::SUBC, MVT::i32, Custom);
655 setOperationAction(ISD::SUBE, MVT::i32, Custom);
658 // ARM does not have ROTL.
659 setOperationAction(ISD::ROTL, MVT::i32, Expand);
660 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
661 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
662 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
663 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
665 // These just redirect to CTTZ and CTLZ on ARM.
666 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
667 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
669 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
671 // Only ARMv6 has BSWAP.
672 if (!Subtarget->hasV6Ops())
673 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
675 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
676 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
677 // These are expanded into libcalls if the cpu doesn't have HW divider.
678 setOperationAction(ISD::SDIV, MVT::i32, Expand);
679 setOperationAction(ISD::UDIV, MVT::i32, Expand);
682 // FIXME: Also set divmod for SREM on EABI
683 setOperationAction(ISD::SREM, MVT::i32, Expand);
684 setOperationAction(ISD::UREM, MVT::i32, Expand);
685 // Register based DivRem for AEABI (RTABI 4.2)
686 if (Subtarget->isTargetAEABI()) {
687 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
688 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
689 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
690 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
691 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
692 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
693 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
694 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
696 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
697 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
698 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
699 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
700 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
701 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
702 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
703 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
705 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
706 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
708 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
709 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
712 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
713 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
714 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
715 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
716 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
718 setOperationAction(ISD::TRAP, MVT::Other, Legal);
720 // Use the default implementation.
721 setOperationAction(ISD::VASTART, MVT::Other, Custom);
722 setOperationAction(ISD::VAARG, MVT::Other, Expand);
723 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
724 setOperationAction(ISD::VAEND, MVT::Other, Expand);
725 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
726 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
728 if (!Subtarget->isTargetMachO()) {
729 // Non-MachO platforms may return values in these registers via the
730 // personality function.
731 setExceptionPointerRegister(ARM::R0);
732 setExceptionSelectorRegister(ARM::R1);
735 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
736 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
737 // the default expansion.
738 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
739 // ATOMIC_FENCE needs custom lowering; the other 32-bit ones are legal and
741 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
742 // Custom lowering for 64-bit ops
743 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
744 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
745 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
746 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
747 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
748 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
749 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
750 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
751 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
752 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
753 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
754 // On v8, we have particularly efficient implementations of atomic fences
755 // if they can be combined with nearby atomic loads and stores.
756 if (!Subtarget->hasV8Ops()) {
757 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
758 setInsertFencesForAtomic(true);
760 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
762 // If there's anything we can use as a barrier, go through custom lowering
764 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
765 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
767 // Set them all for expansion, which will force libcalls.
768 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
769 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
770 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
771 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
772 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
773 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
774 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
775 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
776 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
777 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
778 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
779 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
780 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
781 // Unordered/Monotonic case.
782 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
783 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
786 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
788 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
789 if (!Subtarget->hasV6Ops()) {
790 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
791 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
793 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
795 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
796 !Subtarget->isThumb1Only()) {
797 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
798 // iff target supports vfp2.
799 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
800 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
803 // We want to custom lower some of our intrinsics.
804 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
805 if (Subtarget->isTargetDarwin()) {
806 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
807 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
808 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
811 setOperationAction(ISD::SETCC, MVT::i32, Expand);
812 setOperationAction(ISD::SETCC, MVT::f32, Expand);
813 setOperationAction(ISD::SETCC, MVT::f64, Expand);
814 setOperationAction(ISD::SELECT, MVT::i32, Custom);
815 setOperationAction(ISD::SELECT, MVT::f32, Custom);
816 setOperationAction(ISD::SELECT, MVT::f64, Custom);
817 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
818 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
819 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
821 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
822 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
823 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
824 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
825 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
827 // We don't support sin/cos/fmod/copysign/pow
828 setOperationAction(ISD::FSIN, MVT::f64, Expand);
829 setOperationAction(ISD::FSIN, MVT::f32, Expand);
830 setOperationAction(ISD::FCOS, MVT::f32, Expand);
831 setOperationAction(ISD::FCOS, MVT::f64, Expand);
832 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
833 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
834 setOperationAction(ISD::FREM, MVT::f64, Expand);
835 setOperationAction(ISD::FREM, MVT::f32, Expand);
836 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
837 !Subtarget->isThumb1Only()) {
838 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
839 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
841 setOperationAction(ISD::FPOW, MVT::f64, Expand);
842 setOperationAction(ISD::FPOW, MVT::f32, Expand);
844 if (!Subtarget->hasVFP4()) {
845 setOperationAction(ISD::FMA, MVT::f64, Expand);
846 setOperationAction(ISD::FMA, MVT::f32, Expand);
849 // Various VFP goodness
850 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
851 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
852 if (Subtarget->hasVFP2()) {
853 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
854 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
855 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
856 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
858 // Special handling for half-precision FP.
859 if (!Subtarget->hasFP16()) {
860 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
861 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
865 // Combine sin / cos into one node or libcall if possible.
866 if (Subtarget->hasSinCos()) {
867 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
868 setLibcallName(RTLIB::SINCOS_F64, "sincos");
869 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
870 // For iOS, we don't want to the normal expansion of a libcall to
871 // sincos. We want to issue a libcall to __sincos_stret.
872 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
873 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
877 // We have target-specific dag combine patterns for the following nodes:
878 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
879 setTargetDAGCombine(ISD::ADD);
880 setTargetDAGCombine(ISD::SUB);
881 setTargetDAGCombine(ISD::MUL);
882 setTargetDAGCombine(ISD::AND);
883 setTargetDAGCombine(ISD::OR);
884 setTargetDAGCombine(ISD::XOR);
886 if (Subtarget->hasV6Ops())
887 setTargetDAGCombine(ISD::SRL);
889 setStackPointerRegisterToSaveRestore(ARM::SP);
891 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
892 !Subtarget->hasVFP2())
893 setSchedulingPreference(Sched::RegPressure);
895 setSchedulingPreference(Sched::Hybrid);
897 //// temporary - rewrite interface to use type
898 MaxStoresPerMemset = 8;
899 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
900 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
901 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
902 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
903 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
905 // On ARM arguments smaller than 4 bytes are extended, so all arguments
906 // are at least 4 bytes aligned.
907 setMinStackArgumentAlignment(4);
909 // Prefer likely predicted branches to selects on out-of-order cores.
910 PredictableSelectIsExpensive = Subtarget->isLikeA9();
912 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
915 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
916 bool isThumb2, unsigned &LdrOpc,
918 static const unsigned LoadBares[4][2] = {{ARM::LDREXB, ARM::t2LDREXB},
919 {ARM::LDREXH, ARM::t2LDREXH},
920 {ARM::LDREX, ARM::t2LDREX},
921 {ARM::LDREXD, ARM::t2LDREXD}};
922 static const unsigned LoadAcqs[4][2] = {{ARM::LDAEXB, ARM::t2LDAEXB},
923 {ARM::LDAEXH, ARM::t2LDAEXH},
924 {ARM::LDAEX, ARM::t2LDAEX},
925 {ARM::LDAEXD, ARM::t2LDAEXD}};
926 static const unsigned StoreBares[4][2] = {{ARM::STREXB, ARM::t2STREXB},
927 {ARM::STREXH, ARM::t2STREXH},
928 {ARM::STREX, ARM::t2STREX},
929 {ARM::STREXD, ARM::t2STREXD}};
930 static const unsigned StoreRels[4][2] = {{ARM::STLEXB, ARM::t2STLEXB},
931 {ARM::STLEXH, ARM::t2STLEXH},
932 {ARM::STLEX, ARM::t2STLEX},
933 {ARM::STLEXD, ARM::t2STLEXD}};
935 const unsigned (*LoadOps)[2], (*StoreOps)[2];
936 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
941 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
942 StoreOps = StoreRels;
944 StoreOps = StoreBares;
946 assert(isPowerOf2_32(Size) && Size <= 8 &&
947 "unsupported size for atomic binary op!");
949 LdrOpc = LoadOps[Log2_32(Size)][isThumb2];
950 StrOpc = StoreOps[Log2_32(Size)][isThumb2];
953 // FIXME: It might make sense to define the representative register class as the
954 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
955 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
956 // SPR's representative would be DPR_VFP2. This should work well if register
957 // pressure tracking were modified such that a register use would increment the
958 // pressure of the register class's representative and all of it's super
959 // classes' representatives transitively. We have not implemented this because
960 // of the difficulty prior to coalescing of modeling operand register classes
961 // due to the common occurrence of cross class copies and subregister insertions
963 std::pair<const TargetRegisterClass*, uint8_t>
964 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
965 const TargetRegisterClass *RRC = 0;
967 switch (VT.SimpleTy) {
969 return TargetLowering::findRepresentativeClass(VT);
970 // Use DPR as representative register class for all floating point
971 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
972 // the cost is 1 for both f32 and f64.
973 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
974 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
975 RRC = &ARM::DPRRegClass;
976 // When NEON is used for SP, only half of the register file is available
977 // because operations that define both SP and DP results will be constrained
978 // to the VFP2 class (D0-D15). We currently model this constraint prior to
979 // coalescing by double-counting the SP regs. See the FIXME above.
980 if (Subtarget->useNEONForSinglePrecisionFP())
983 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
984 case MVT::v4f32: case MVT::v2f64:
985 RRC = &ARM::DPRRegClass;
989 RRC = &ARM::DPRRegClass;
993 RRC = &ARM::DPRRegClass;
997 return std::make_pair(RRC, Cost);
1000 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1003 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1004 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1005 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1006 case ARMISD::CALL: return "ARMISD::CALL";
1007 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1008 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1009 case ARMISD::tCALL: return "ARMISD::tCALL";
1010 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1011 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1012 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1013 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1014 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1015 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1016 case ARMISD::CMP: return "ARMISD::CMP";
1017 case ARMISD::CMN: return "ARMISD::CMN";
1018 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1019 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1020 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1021 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1022 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1024 case ARMISD::CMOV: return "ARMISD::CMOV";
1026 case ARMISD::RBIT: return "ARMISD::RBIT";
1028 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1029 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1030 case ARMISD::SITOF: return "ARMISD::SITOF";
1031 case ARMISD::UITOF: return "ARMISD::UITOF";
1033 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1034 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1035 case ARMISD::RRX: return "ARMISD::RRX";
1037 case ARMISD::ADDC: return "ARMISD::ADDC";
1038 case ARMISD::ADDE: return "ARMISD::ADDE";
1039 case ARMISD::SUBC: return "ARMISD::SUBC";
1040 case ARMISD::SUBE: return "ARMISD::SUBE";
1042 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1043 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1045 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1046 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1048 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1050 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1052 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1054 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1056 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1058 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1059 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1060 case ARMISD::VCGE: return "ARMISD::VCGE";
1061 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1062 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1063 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1064 case ARMISD::VCGT: return "ARMISD::VCGT";
1065 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1066 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1067 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1068 case ARMISD::VTST: return "ARMISD::VTST";
1070 case ARMISD::VSHL: return "ARMISD::VSHL";
1071 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1072 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1073 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1074 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1075 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1076 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1077 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1078 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1079 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1080 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1081 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1082 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1083 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1084 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1085 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1086 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1087 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1088 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1089 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1090 case ARMISD::VDUP: return "ARMISD::VDUP";
1091 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1092 case ARMISD::VEXT: return "ARMISD::VEXT";
1093 case ARMISD::VREV64: return "ARMISD::VREV64";
1094 case ARMISD::VREV32: return "ARMISD::VREV32";
1095 case ARMISD::VREV16: return "ARMISD::VREV16";
1096 case ARMISD::VZIP: return "ARMISD::VZIP";
1097 case ARMISD::VUZP: return "ARMISD::VUZP";
1098 case ARMISD::VTRN: return "ARMISD::VTRN";
1099 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1100 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1101 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1102 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1103 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1104 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1105 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1106 case ARMISD::FMAX: return "ARMISD::FMAX";
1107 case ARMISD::FMIN: return "ARMISD::FMIN";
1108 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1109 case ARMISD::VMINNM: return "ARMISD::VMIN";
1110 case ARMISD::BFI: return "ARMISD::BFI";
1111 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1112 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1113 case ARMISD::VBSL: return "ARMISD::VBSL";
1114 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1115 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1116 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1117 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1118 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1119 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1120 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1121 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1122 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1123 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1124 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1125 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1126 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1127 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1128 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1129 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1130 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1131 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1132 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1133 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1137 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1138 if (!VT.isVector()) return getPointerTy();
1139 return VT.changeVectorElementTypeToInteger();
1142 /// getRegClassFor - Return the register class that should be used for the
1143 /// specified value type.
1144 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1145 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1146 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1147 // load / store 4 to 8 consecutive D registers.
1148 if (Subtarget->hasNEON()) {
1149 if (VT == MVT::v4i64)
1150 return &ARM::QQPRRegClass;
1151 if (VT == MVT::v8i64)
1152 return &ARM::QQQQPRRegClass;
1154 return TargetLowering::getRegClassFor(VT);
1157 // Create a fast isel object.
1159 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1160 const TargetLibraryInfo *libInfo) const {
1161 return ARM::createFastISel(funcInfo, libInfo);
1164 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1165 /// be used for loads / stores from the global.
1166 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1167 return (Subtarget->isThumb1Only() ? 127 : 4095);
1170 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1171 unsigned NumVals = N->getNumValues();
1173 return Sched::RegPressure;
1175 for (unsigned i = 0; i != NumVals; ++i) {
1176 EVT VT = N->getValueType(i);
1177 if (VT == MVT::Glue || VT == MVT::Other)
1179 if (VT.isFloatingPoint() || VT.isVector())
1183 if (!N->isMachineOpcode())
1184 return Sched::RegPressure;
1186 // Load are scheduled for latency even if there instruction itinerary
1187 // is not available.
1188 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1189 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1191 if (MCID.getNumDefs() == 0)
1192 return Sched::RegPressure;
1193 if (!Itins->isEmpty() &&
1194 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1197 return Sched::RegPressure;
1200 //===----------------------------------------------------------------------===//
1202 //===----------------------------------------------------------------------===//
1204 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1205 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1207 default: llvm_unreachable("Unknown condition code!");
1208 case ISD::SETNE: return ARMCC::NE;
1209 case ISD::SETEQ: return ARMCC::EQ;
1210 case ISD::SETGT: return ARMCC::GT;
1211 case ISD::SETGE: return ARMCC::GE;
1212 case ISD::SETLT: return ARMCC::LT;
1213 case ISD::SETLE: return ARMCC::LE;
1214 case ISD::SETUGT: return ARMCC::HI;
1215 case ISD::SETUGE: return ARMCC::HS;
1216 case ISD::SETULT: return ARMCC::LO;
1217 case ISD::SETULE: return ARMCC::LS;
1221 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1222 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1223 ARMCC::CondCodes &CondCode2) {
1224 CondCode2 = ARMCC::AL;
1226 default: llvm_unreachable("Unknown FP condition!");
1228 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1230 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1232 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1233 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1234 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1235 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1236 case ISD::SETO: CondCode = ARMCC::VC; break;
1237 case ISD::SETUO: CondCode = ARMCC::VS; break;
1238 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1239 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1240 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1242 case ISD::SETULT: CondCode = ARMCC::LT; break;
1244 case ISD::SETULE: CondCode = ARMCC::LE; break;
1246 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1250 //===----------------------------------------------------------------------===//
1251 // Calling Convention Implementation
1252 //===----------------------------------------------------------------------===//
1254 #include "ARMGenCallingConv.inc"
1256 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1257 /// given CallingConvention value.
1258 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1260 bool isVarArg) const {
1263 llvm_unreachable("Unsupported calling convention");
1264 case CallingConv::Fast:
1265 if (Subtarget->hasVFP2() && !isVarArg) {
1266 if (!Subtarget->isAAPCS_ABI())
1267 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1268 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1269 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1272 case CallingConv::C: {
1273 // Use target triple & subtarget features to do actual dispatch.
1274 if (!Subtarget->isAAPCS_ABI())
1275 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1276 else if (Subtarget->hasVFP2() &&
1277 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1279 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1280 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1282 case CallingConv::ARM_AAPCS_VFP:
1284 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1286 case CallingConv::ARM_AAPCS:
1287 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1288 case CallingConv::ARM_APCS:
1289 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1290 case CallingConv::GHC:
1291 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1295 /// LowerCallResult - Lower the result values of a call into the
1296 /// appropriate copies out of appropriate physical registers.
1298 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1299 CallingConv::ID CallConv, bool isVarArg,
1300 const SmallVectorImpl<ISD::InputArg> &Ins,
1301 SDLoc dl, SelectionDAG &DAG,
1302 SmallVectorImpl<SDValue> &InVals,
1303 bool isThisReturn, SDValue ThisVal) const {
1305 // Assign locations to each value returned by this call.
1306 SmallVector<CCValAssign, 16> RVLocs;
1307 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1308 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1309 CCInfo.AnalyzeCallResult(Ins,
1310 CCAssignFnForNode(CallConv, /* Return*/ true,
1313 // Copy all of the result registers out of their specified physreg.
1314 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1315 CCValAssign VA = RVLocs[i];
1317 // Pass 'this' value directly from the argument to return value, to avoid
1318 // reg unit interference
1319 if (i == 0 && isThisReturn) {
1320 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1321 "unexpected return calling convention register assignment");
1322 InVals.push_back(ThisVal);
1327 if (VA.needsCustom()) {
1328 // Handle f64 or half of a v2f64.
1329 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1331 Chain = Lo.getValue(1);
1332 InFlag = Lo.getValue(2);
1333 VA = RVLocs[++i]; // skip ahead to next loc
1334 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1336 Chain = Hi.getValue(1);
1337 InFlag = Hi.getValue(2);
1338 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1340 if (VA.getLocVT() == MVT::v2f64) {
1341 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1342 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1343 DAG.getConstant(0, MVT::i32));
1345 VA = RVLocs[++i]; // skip ahead to next loc
1346 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1347 Chain = Lo.getValue(1);
1348 InFlag = Lo.getValue(2);
1349 VA = RVLocs[++i]; // skip ahead to next loc
1350 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1351 Chain = Hi.getValue(1);
1352 InFlag = Hi.getValue(2);
1353 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1354 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1355 DAG.getConstant(1, MVT::i32));
1358 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1360 Chain = Val.getValue(1);
1361 InFlag = Val.getValue(2);
1364 switch (VA.getLocInfo()) {
1365 default: llvm_unreachable("Unknown loc info!");
1366 case CCValAssign::Full: break;
1367 case CCValAssign::BCvt:
1368 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1372 InVals.push_back(Val);
1378 /// LowerMemOpCallTo - Store the argument to the stack.
1380 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1381 SDValue StackPtr, SDValue Arg,
1382 SDLoc dl, SelectionDAG &DAG,
1383 const CCValAssign &VA,
1384 ISD::ArgFlagsTy Flags) const {
1385 unsigned LocMemOffset = VA.getLocMemOffset();
1386 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1387 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1388 return DAG.getStore(Chain, dl, Arg, PtrOff,
1389 MachinePointerInfo::getStack(LocMemOffset),
1393 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1394 SDValue Chain, SDValue &Arg,
1395 RegsToPassVector &RegsToPass,
1396 CCValAssign &VA, CCValAssign &NextVA,
1398 SmallVectorImpl<SDValue> &MemOpChains,
1399 ISD::ArgFlagsTy Flags) const {
1401 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1402 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1403 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1405 if (NextVA.isRegLoc())
1406 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1408 assert(NextVA.isMemLoc());
1409 if (StackPtr.getNode() == 0)
1410 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1412 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1418 /// LowerCall - Lowering a call into a callseq_start <-
1419 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1422 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1423 SmallVectorImpl<SDValue> &InVals) const {
1424 SelectionDAG &DAG = CLI.DAG;
1426 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1427 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1428 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1429 SDValue Chain = CLI.Chain;
1430 SDValue Callee = CLI.Callee;
1431 bool &isTailCall = CLI.IsTailCall;
1432 CallingConv::ID CallConv = CLI.CallConv;
1433 bool doesNotRet = CLI.DoesNotReturn;
1434 bool isVarArg = CLI.IsVarArg;
1436 MachineFunction &MF = DAG.getMachineFunction();
1437 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1438 bool isThisReturn = false;
1439 bool isSibCall = false;
1441 // Disable tail calls if they're not supported.
1442 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1446 // Check if it's really possible to do a tail call.
1447 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1448 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1449 Outs, OutVals, Ins, DAG);
1450 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1451 // detected sibcalls.
1458 // Analyze operands of the call, assigning locations to each operand.
1459 SmallVector<CCValAssign, 16> ArgLocs;
1460 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1461 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1462 CCInfo.AnalyzeCallOperands(Outs,
1463 CCAssignFnForNode(CallConv, /* Return*/ false,
1466 // Get a count of how many bytes are to be pushed on the stack.
1467 unsigned NumBytes = CCInfo.getNextStackOffset();
1469 // For tail calls, memory operands are available in our caller's stack.
1473 // Adjust the stack pointer for the new arguments...
1474 // These operations are automatically eliminated by the prolog/epilog pass
1476 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1479 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1481 RegsToPassVector RegsToPass;
1482 SmallVector<SDValue, 8> MemOpChains;
1484 // Walk the register/memloc assignments, inserting copies/loads. In the case
1485 // of tail call optimization, arguments are handled later.
1486 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1488 ++i, ++realArgIdx) {
1489 CCValAssign &VA = ArgLocs[i];
1490 SDValue Arg = OutVals[realArgIdx];
1491 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1492 bool isByVal = Flags.isByVal();
1494 // Promote the value if needed.
1495 switch (VA.getLocInfo()) {
1496 default: llvm_unreachable("Unknown loc info!");
1497 case CCValAssign::Full: break;
1498 case CCValAssign::SExt:
1499 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1501 case CCValAssign::ZExt:
1502 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1504 case CCValAssign::AExt:
1505 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1507 case CCValAssign::BCvt:
1508 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1512 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1513 if (VA.needsCustom()) {
1514 if (VA.getLocVT() == MVT::v2f64) {
1515 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1516 DAG.getConstant(0, MVT::i32));
1517 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1518 DAG.getConstant(1, MVT::i32));
1520 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1521 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1523 VA = ArgLocs[++i]; // skip ahead to next loc
1524 if (VA.isRegLoc()) {
1525 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1526 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1528 assert(VA.isMemLoc());
1530 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1531 dl, DAG, VA, Flags));
1534 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1535 StackPtr, MemOpChains, Flags);
1537 } else if (VA.isRegLoc()) {
1538 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1539 assert(VA.getLocVT() == MVT::i32 &&
1540 "unexpected calling convention register assignment");
1541 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1542 "unexpected use of 'returned'");
1543 isThisReturn = true;
1545 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1546 } else if (isByVal) {
1547 assert(VA.isMemLoc());
1548 unsigned offset = 0;
1550 // True if this byval aggregate will be split between registers
1552 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1553 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1555 if (CurByValIdx < ByValArgsCount) {
1557 unsigned RegBegin, RegEnd;
1558 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1560 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1562 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1563 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1564 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1565 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1566 MachinePointerInfo(),
1567 false, false, false,
1568 DAG.InferPtrAlignment(AddArg));
1569 MemOpChains.push_back(Load.getValue(1));
1570 RegsToPass.push_back(std::make_pair(j, Load));
1573 // If parameter size outsides register area, "offset" value
1574 // helps us to calculate stack slot for remained part properly.
1575 offset = RegEnd - RegBegin;
1577 CCInfo.nextInRegsParam();
1580 if (Flags.getByValSize() > 4*offset) {
1581 unsigned LocMemOffset = VA.getLocMemOffset();
1582 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1583 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1585 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1586 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1587 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1589 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1591 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1592 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1593 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1594 Ops, array_lengthof(Ops)));
1596 } else if (!isSibCall) {
1597 assert(VA.isMemLoc());
1599 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1600 dl, DAG, VA, Flags));
1604 if (!MemOpChains.empty())
1605 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1606 &MemOpChains[0], MemOpChains.size());
1608 // Build a sequence of copy-to-reg nodes chained together with token chain
1609 // and flag operands which copy the outgoing args into the appropriate regs.
1611 // Tail call byval lowering might overwrite argument registers so in case of
1612 // tail call optimization the copies to registers are lowered later.
1614 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1615 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1616 RegsToPass[i].second, InFlag);
1617 InFlag = Chain.getValue(1);
1620 // For tail calls lower the arguments to the 'real' stack slot.
1622 // Force all the incoming stack arguments to be loaded from the stack
1623 // before any new outgoing arguments are stored to the stack, because the
1624 // outgoing stack slots may alias the incoming argument stack slots, and
1625 // the alias isn't otherwise explicit. This is slightly more conservative
1626 // than necessary, because it means that each store effectively depends
1627 // on every argument instead of just those arguments it would clobber.
1629 // Do not flag preceding copytoreg stuff together with the following stuff.
1631 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1632 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1633 RegsToPass[i].second, InFlag);
1634 InFlag = Chain.getValue(1);
1639 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1640 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1641 // node so that legalize doesn't hack it.
1642 bool isDirect = false;
1643 bool isARMFunc = false;
1644 bool isLocalARMFunc = false;
1645 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1647 if (EnableARMLongCalls) {
1648 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1649 && "long-calls with non-static relocation model!");
1650 // Handle a global address or an external symbol. If it's not one of
1651 // those, the target's already in a register, so we don't need to do
1653 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1654 const GlobalValue *GV = G->getGlobal();
1655 // Create a constant pool entry for the callee address
1656 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1657 ARMConstantPoolValue *CPV =
1658 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1660 // Get the address of the callee into a register
1661 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1662 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1663 Callee = DAG.getLoad(getPointerTy(), dl,
1664 DAG.getEntryNode(), CPAddr,
1665 MachinePointerInfo::getConstantPool(),
1666 false, false, false, 0);
1667 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1668 const char *Sym = S->getSymbol();
1670 // Create a constant pool entry for the callee address
1671 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1672 ARMConstantPoolValue *CPV =
1673 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1674 ARMPCLabelIndex, 0);
1675 // Get the address of the callee into a register
1676 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1677 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1678 Callee = DAG.getLoad(getPointerTy(), dl,
1679 DAG.getEntryNode(), CPAddr,
1680 MachinePointerInfo::getConstantPool(),
1681 false, false, false, 0);
1683 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1684 const GlobalValue *GV = G->getGlobal();
1686 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1687 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1688 getTargetMachine().getRelocationModel() != Reloc::Static;
1689 isARMFunc = !Subtarget->isThumb() || isStub;
1690 // ARM call to a local ARM function is predicable.
1691 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1692 // tBX takes a register source operand.
1693 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1694 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1695 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1696 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1698 // On ELF targets for PIC code, direct calls should go through the PLT
1699 unsigned OpFlags = 0;
1700 if (Subtarget->isTargetELF() &&
1701 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1702 OpFlags = ARMII::MO_PLT;
1703 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1705 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1707 bool isStub = Subtarget->isTargetMachO() &&
1708 getTargetMachine().getRelocationModel() != Reloc::Static;
1709 isARMFunc = !Subtarget->isThumb() || isStub;
1710 // tBX takes a register source operand.
1711 const char *Sym = S->getSymbol();
1712 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1713 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1714 ARMConstantPoolValue *CPV =
1715 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1716 ARMPCLabelIndex, 4);
1717 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1718 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1719 Callee = DAG.getLoad(getPointerTy(), dl,
1720 DAG.getEntryNode(), CPAddr,
1721 MachinePointerInfo::getConstantPool(),
1722 false, false, false, 0);
1723 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1724 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1725 getPointerTy(), Callee, PICLabel);
1727 unsigned OpFlags = 0;
1728 // On ELF targets for PIC code, direct calls should go through the PLT
1729 if (Subtarget->isTargetELF() &&
1730 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1731 OpFlags = ARMII::MO_PLT;
1732 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1736 // FIXME: handle tail calls differently.
1738 bool HasMinSizeAttr = Subtarget->isMinSize();
1739 if (Subtarget->isThumb()) {
1740 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1741 CallOpc = ARMISD::CALL_NOLINK;
1743 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1745 if (!isDirect && !Subtarget->hasV5TOps())
1746 CallOpc = ARMISD::CALL_NOLINK;
1747 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1748 // Emit regular call when code size is the priority
1750 // "mov lr, pc; b _foo" to avoid confusing the RSP
1751 CallOpc = ARMISD::CALL_NOLINK;
1753 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1756 std::vector<SDValue> Ops;
1757 Ops.push_back(Chain);
1758 Ops.push_back(Callee);
1760 // Add argument registers to the end of the list so that they are known live
1762 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1763 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1764 RegsToPass[i].second.getValueType()));
1766 // Add a register mask operand representing the call-preserved registers.
1768 const uint32_t *Mask;
1769 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1770 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1772 // For 'this' returns, use the R0-preserving mask if applicable
1773 Mask = ARI->getThisReturnPreservedMask(CallConv);
1775 // Set isThisReturn to false if the calling convention is not one that
1776 // allows 'returned' to be modeled in this way, so LowerCallResult does
1777 // not try to pass 'this' straight through
1778 isThisReturn = false;
1779 Mask = ARI->getCallPreservedMask(CallConv);
1782 Mask = ARI->getCallPreservedMask(CallConv);
1784 assert(Mask && "Missing call preserved mask for calling convention");
1785 Ops.push_back(DAG.getRegisterMask(Mask));
1788 if (InFlag.getNode())
1789 Ops.push_back(InFlag);
1791 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1793 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1795 // Returns a chain and a flag for retval copy to use.
1796 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1797 InFlag = Chain.getValue(1);
1799 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1800 DAG.getIntPtrConstant(0, true), InFlag, dl);
1802 InFlag = Chain.getValue(1);
1804 // Handle result values, copying them out of physregs into vregs that we
1806 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1807 InVals, isThisReturn,
1808 isThisReturn ? OutVals[0] : SDValue());
1811 /// HandleByVal - Every parameter *after* a byval parameter is passed
1812 /// on the stack. Remember the next parameter register to allocate,
1813 /// and then confiscate the rest of the parameter registers to insure
1816 ARMTargetLowering::HandleByVal(
1817 CCState *State, unsigned &size, unsigned Align) const {
1818 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1819 assert((State->getCallOrPrologue() == Prologue ||
1820 State->getCallOrPrologue() == Call) &&
1821 "unhandled ParmContext");
1823 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1824 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1825 unsigned AlignInRegs = Align / 4;
1826 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1827 for (unsigned i = 0; i < Waste; ++i)
1828 reg = State->AllocateReg(GPRArgRegs, 4);
1831 unsigned excess = 4 * (ARM::R4 - reg);
1833 // Special case when NSAA != SP and parameter size greater than size of
1834 // all remained GPR regs. In that case we can't split parameter, we must
1835 // send it to stack. We also must set NCRN to R4, so waste all
1836 // remained registers.
1837 const unsigned NSAAOffset = State->getNextStackOffset();
1838 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1839 while (State->AllocateReg(GPRArgRegs, 4))
1844 // First register for byval parameter is the first register that wasn't
1845 // allocated before this method call, so it would be "reg".
1846 // If parameter is small enough to be saved in range [reg, r4), then
1847 // the end (first after last) register would be reg + param-size-in-regs,
1848 // else parameter would be splitted between registers and stack,
1849 // end register would be r4 in this case.
1850 unsigned ByValRegBegin = reg;
1851 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1852 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1853 // Note, first register is allocated in the beginning of function already,
1854 // allocate remained amount of registers we need.
1855 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1856 State->AllocateReg(GPRArgRegs, 4);
1857 // A byval parameter that is split between registers and memory needs its
1858 // size truncated here.
1859 // In the case where the entire structure fits in registers, we set the
1860 // size in memory to zero.
1869 /// MatchingStackOffset - Return true if the given stack call argument is
1870 /// already available in the same position (relatively) of the caller's
1871 /// incoming argument stack.
1873 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1874 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1875 const TargetInstrInfo *TII) {
1876 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1878 if (Arg.getOpcode() == ISD::CopyFromReg) {
1879 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1880 if (!TargetRegisterInfo::isVirtualRegister(VR))
1882 MachineInstr *Def = MRI->getVRegDef(VR);
1885 if (!Flags.isByVal()) {
1886 if (!TII->isLoadFromStackSlot(Def, FI))
1891 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1892 if (Flags.isByVal())
1893 // ByVal argument is passed in as a pointer but it's now being
1894 // dereferenced. e.g.
1895 // define @foo(%struct.X* %A) {
1896 // tail call @bar(%struct.X* byval %A)
1899 SDValue Ptr = Ld->getBasePtr();
1900 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1903 FI = FINode->getIndex();
1907 assert(FI != INT_MAX);
1908 if (!MFI->isFixedObjectIndex(FI))
1910 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1913 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1914 /// for tail call optimization. Targets which want to do tail call
1915 /// optimization should implement this function.
1917 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1918 CallingConv::ID CalleeCC,
1920 bool isCalleeStructRet,
1921 bool isCallerStructRet,
1922 const SmallVectorImpl<ISD::OutputArg> &Outs,
1923 const SmallVectorImpl<SDValue> &OutVals,
1924 const SmallVectorImpl<ISD::InputArg> &Ins,
1925 SelectionDAG& DAG) const {
1926 const Function *CallerF = DAG.getMachineFunction().getFunction();
1927 CallingConv::ID CallerCC = CallerF->getCallingConv();
1928 bool CCMatch = CallerCC == CalleeCC;
1930 // Look for obvious safe cases to perform tail call optimization that do not
1931 // require ABI changes. This is what gcc calls sibcall.
1933 // Do not sibcall optimize vararg calls unless the call site is not passing
1935 if (isVarArg && !Outs.empty())
1938 // Exception-handling functions need a special set of instructions to indicate
1939 // a return to the hardware. Tail-calling another function would probably
1941 if (CallerF->hasFnAttribute("interrupt"))
1944 // Also avoid sibcall optimization if either caller or callee uses struct
1945 // return semantics.
1946 if (isCalleeStructRet || isCallerStructRet)
1949 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1950 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1951 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1952 // support in the assembler and linker to be used. This would need to be
1953 // fixed to fully support tail calls in Thumb1.
1955 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1956 // LR. This means if we need to reload LR, it takes an extra instructions,
1957 // which outweighs the value of the tail call; but here we don't know yet
1958 // whether LR is going to be used. Probably the right approach is to
1959 // generate the tail call here and turn it back into CALL/RET in
1960 // emitEpilogue if LR is used.
1962 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1963 // but we need to make sure there are enough registers; the only valid
1964 // registers are the 4 used for parameters. We don't currently do this
1966 if (Subtarget->isThumb1Only())
1969 // If the calling conventions do not match, then we'd better make sure the
1970 // results are returned in the same way as what the caller expects.
1972 SmallVector<CCValAssign, 16> RVLocs1;
1973 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1974 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1975 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1977 SmallVector<CCValAssign, 16> RVLocs2;
1978 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1979 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1980 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1982 if (RVLocs1.size() != RVLocs2.size())
1984 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1985 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1987 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1989 if (RVLocs1[i].isRegLoc()) {
1990 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1993 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1999 // If Caller's vararg or byval argument has been split between registers and
2000 // stack, do not perform tail call, since part of the argument is in caller's
2002 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2003 getInfo<ARMFunctionInfo>();
2004 if (AFI_Caller->getArgRegsSaveSize())
2007 // If the callee takes no arguments then go on to check the results of the
2009 if (!Outs.empty()) {
2010 // Check if stack adjustment is needed. For now, do not do this if any
2011 // argument is passed on the stack.
2012 SmallVector<CCValAssign, 16> ArgLocs;
2013 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2014 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2015 CCInfo.AnalyzeCallOperands(Outs,
2016 CCAssignFnForNode(CalleeCC, false, isVarArg));
2017 if (CCInfo.getNextStackOffset()) {
2018 MachineFunction &MF = DAG.getMachineFunction();
2020 // Check if the arguments are already laid out in the right way as
2021 // the caller's fixed stack objects.
2022 MachineFrameInfo *MFI = MF.getFrameInfo();
2023 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2024 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2025 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2027 ++i, ++realArgIdx) {
2028 CCValAssign &VA = ArgLocs[i];
2029 EVT RegVT = VA.getLocVT();
2030 SDValue Arg = OutVals[realArgIdx];
2031 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2032 if (VA.getLocInfo() == CCValAssign::Indirect)
2034 if (VA.needsCustom()) {
2035 // f64 and vector types are split into multiple registers or
2036 // register/stack-slot combinations. The types will not match
2037 // the registers; give up on memory f64 refs until we figure
2038 // out what to do about this.
2041 if (!ArgLocs[++i].isRegLoc())
2043 if (RegVT == MVT::v2f64) {
2044 if (!ArgLocs[++i].isRegLoc())
2046 if (!ArgLocs[++i].isRegLoc())
2049 } else if (!VA.isRegLoc()) {
2050 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2062 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2063 MachineFunction &MF, bool isVarArg,
2064 const SmallVectorImpl<ISD::OutputArg> &Outs,
2065 LLVMContext &Context) const {
2066 SmallVector<CCValAssign, 16> RVLocs;
2067 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2068 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2072 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2073 SDLoc DL, SelectionDAG &DAG) {
2074 const MachineFunction &MF = DAG.getMachineFunction();
2075 const Function *F = MF.getFunction();
2077 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2079 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2080 // version of the "preferred return address". These offsets affect the return
2081 // instruction if this is a return from PL1 without hypervisor extensions.
2082 // IRQ/FIQ: +4 "subs pc, lr, #4"
2083 // SWI: 0 "subs pc, lr, #0"
2084 // ABORT: +4 "subs pc, lr, #4"
2085 // UNDEF: +4/+2 "subs pc, lr, #0"
2086 // UNDEF varies depending on where the exception came from ARM or Thumb
2087 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2090 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2093 else if (IntKind == "SWI" || IntKind == "UNDEF")
2096 report_fatal_error("Unsupported interrupt attribute. If present, value "
2097 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2099 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2101 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other,
2102 RetOps.data(), RetOps.size());
2106 ARMTargetLowering::LowerReturn(SDValue Chain,
2107 CallingConv::ID CallConv, bool isVarArg,
2108 const SmallVectorImpl<ISD::OutputArg> &Outs,
2109 const SmallVectorImpl<SDValue> &OutVals,
2110 SDLoc dl, SelectionDAG &DAG) const {
2112 // CCValAssign - represent the assignment of the return value to a location.
2113 SmallVector<CCValAssign, 16> RVLocs;
2115 // CCState - Info about the registers and stack slots.
2116 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2117 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2119 // Analyze outgoing return values.
2120 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2124 SmallVector<SDValue, 4> RetOps;
2125 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2127 // Copy the result values into the output registers.
2128 for (unsigned i = 0, realRVLocIdx = 0;
2130 ++i, ++realRVLocIdx) {
2131 CCValAssign &VA = RVLocs[i];
2132 assert(VA.isRegLoc() && "Can only return in registers!");
2134 SDValue Arg = OutVals[realRVLocIdx];
2136 switch (VA.getLocInfo()) {
2137 default: llvm_unreachable("Unknown loc info!");
2138 case CCValAssign::Full: break;
2139 case CCValAssign::BCvt:
2140 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2144 if (VA.needsCustom()) {
2145 if (VA.getLocVT() == MVT::v2f64) {
2146 // Extract the first half and return it in two registers.
2147 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2148 DAG.getConstant(0, MVT::i32));
2149 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2150 DAG.getVTList(MVT::i32, MVT::i32), Half);
2152 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2153 Flag = Chain.getValue(1);
2154 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2155 VA = RVLocs[++i]; // skip ahead to next loc
2156 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2157 HalfGPRs.getValue(1), Flag);
2158 Flag = Chain.getValue(1);
2159 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2160 VA = RVLocs[++i]; // skip ahead to next loc
2162 // Extract the 2nd half and fall through to handle it as an f64 value.
2163 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2164 DAG.getConstant(1, MVT::i32));
2166 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2168 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2169 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2170 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2171 Flag = Chain.getValue(1);
2172 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2173 VA = RVLocs[++i]; // skip ahead to next loc
2174 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2177 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2179 // Guarantee that all emitted copies are
2180 // stuck together, avoiding something bad.
2181 Flag = Chain.getValue(1);
2182 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2185 // Update chain and glue.
2188 RetOps.push_back(Flag);
2190 // CPUs which aren't M-class use a special sequence to return from
2191 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2192 // though we use "subs pc, lr, #N").
2194 // M-class CPUs actually use a normal return sequence with a special
2195 // (hardware-provided) value in LR, so the normal code path works.
2196 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2197 !Subtarget->isMClass()) {
2198 if (Subtarget->isThumb1Only())
2199 report_fatal_error("interrupt attribute is not supported in Thumb1");
2200 return LowerInterruptReturn(RetOps, dl, DAG);
2203 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2204 RetOps.data(), RetOps.size());
2207 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2208 if (N->getNumValues() != 1)
2210 if (!N->hasNUsesOfValue(1, 0))
2213 SDValue TCChain = Chain;
2214 SDNode *Copy = *N->use_begin();
2215 if (Copy->getOpcode() == ISD::CopyToReg) {
2216 // If the copy has a glue operand, we conservatively assume it isn't safe to
2217 // perform a tail call.
2218 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2220 TCChain = Copy->getOperand(0);
2221 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2222 SDNode *VMov = Copy;
2223 // f64 returned in a pair of GPRs.
2224 SmallPtrSet<SDNode*, 2> Copies;
2225 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2227 if (UI->getOpcode() != ISD::CopyToReg)
2231 if (Copies.size() > 2)
2234 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2236 SDValue UseChain = UI->getOperand(0);
2237 if (Copies.count(UseChain.getNode()))
2244 } else if (Copy->getOpcode() == ISD::BITCAST) {
2245 // f32 returned in a single GPR.
2246 if (!Copy->hasOneUse())
2248 Copy = *Copy->use_begin();
2249 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2251 TCChain = Copy->getOperand(0);
2256 bool HasRet = false;
2257 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2259 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2260 UI->getOpcode() != ARMISD::INTRET_FLAG)
2272 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2273 if (!Subtarget->supportsTailCall())
2276 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2279 return !Subtarget->isThumb1Only();
2282 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2283 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2284 // one of the above mentioned nodes. It has to be wrapped because otherwise
2285 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2286 // be used to form addressing mode. These wrapped nodes will be selected
2288 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2289 EVT PtrVT = Op.getValueType();
2290 // FIXME there is no actual debug info here
2292 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2294 if (CP->isMachineConstantPoolEntry())
2295 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2296 CP->getAlignment());
2298 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2299 CP->getAlignment());
2300 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2303 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2304 return MachineJumpTableInfo::EK_Inline;
2307 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2308 SelectionDAG &DAG) const {
2309 MachineFunction &MF = DAG.getMachineFunction();
2310 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2311 unsigned ARMPCLabelIndex = 0;
2313 EVT PtrVT = getPointerTy();
2314 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2315 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2317 if (RelocM == Reloc::Static) {
2318 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2320 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2321 ARMPCLabelIndex = AFI->createPICLabelUId();
2322 ARMConstantPoolValue *CPV =
2323 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2324 ARMCP::CPBlockAddress, PCAdj);
2325 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2327 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2328 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2329 MachinePointerInfo::getConstantPool(),
2330 false, false, false, 0);
2331 if (RelocM == Reloc::Static)
2333 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2334 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2337 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2339 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2340 SelectionDAG &DAG) const {
2342 EVT PtrVT = getPointerTy();
2343 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2344 MachineFunction &MF = DAG.getMachineFunction();
2345 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2346 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2347 ARMConstantPoolValue *CPV =
2348 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2349 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2350 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2351 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2352 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2353 MachinePointerInfo::getConstantPool(),
2354 false, false, false, 0);
2355 SDValue Chain = Argument.getValue(1);
2357 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2358 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2360 // call __tls_get_addr.
2363 Entry.Node = Argument;
2364 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2365 Args.push_back(Entry);
2366 // FIXME: is there useful debug info available here?
2367 TargetLowering::CallLoweringInfo CLI(Chain,
2368 (Type *) Type::getInt32Ty(*DAG.getContext()),
2369 false, false, false, false,
2370 0, CallingConv::C, /*isTailCall=*/false,
2371 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2372 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2373 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2374 return CallResult.first;
2377 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2378 // "local exec" model.
2380 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2382 TLSModel::Model model) const {
2383 const GlobalValue *GV = GA->getGlobal();
2386 SDValue Chain = DAG.getEntryNode();
2387 EVT PtrVT = getPointerTy();
2388 // Get the Thread Pointer
2389 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2391 if (model == TLSModel::InitialExec) {
2392 MachineFunction &MF = DAG.getMachineFunction();
2393 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2394 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2395 // Initial exec model.
2396 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2397 ARMConstantPoolValue *CPV =
2398 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2399 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2401 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2402 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2403 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2404 MachinePointerInfo::getConstantPool(),
2405 false, false, false, 0);
2406 Chain = Offset.getValue(1);
2408 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2409 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2411 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2412 MachinePointerInfo::getConstantPool(),
2413 false, false, false, 0);
2416 assert(model == TLSModel::LocalExec);
2417 ARMConstantPoolValue *CPV =
2418 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2419 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2420 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2421 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2422 MachinePointerInfo::getConstantPool(),
2423 false, false, false, 0);
2426 // The address of the thread local variable is the add of the thread
2427 // pointer with the offset of the variable.
2428 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2432 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2433 // TODO: implement the "local dynamic" model
2434 assert(Subtarget->isTargetELF() &&
2435 "TLS not implemented for non-ELF targets");
2436 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2438 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2441 case TLSModel::GeneralDynamic:
2442 case TLSModel::LocalDynamic:
2443 return LowerToTLSGeneralDynamicModel(GA, DAG);
2444 case TLSModel::InitialExec:
2445 case TLSModel::LocalExec:
2446 return LowerToTLSExecModels(GA, DAG, model);
2448 llvm_unreachable("bogus TLS model");
2451 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2452 SelectionDAG &DAG) const {
2453 EVT PtrVT = getPointerTy();
2455 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2456 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2457 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2458 ARMConstantPoolValue *CPV =
2459 ARMConstantPoolConstant::Create(GV,
2460 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2461 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2462 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2463 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2465 MachinePointerInfo::getConstantPool(),
2466 false, false, false, 0);
2467 SDValue Chain = Result.getValue(1);
2468 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2469 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2471 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2472 MachinePointerInfo::getGOT(),
2473 false, false, false, 0);
2477 // If we have T2 ops, we can materialize the address directly via movt/movw
2478 // pair. This is always cheaper.
2479 if (Subtarget->useMovt()) {
2481 // FIXME: Once remat is capable of dealing with instructions with register
2482 // operands, expand this into two nodes.
2483 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2484 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2486 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2487 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2488 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2489 MachinePointerInfo::getConstantPool(),
2490 false, false, false, 0);
2494 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2495 SelectionDAG &DAG) const {
2496 EVT PtrVT = getPointerTy();
2498 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2499 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2501 if (Subtarget->useMovt())
2504 // FIXME: Once remat is capable of dealing with instructions with register
2505 // operands, expand this into multiple nodes
2507 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2509 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2510 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2512 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2513 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2514 MachinePointerInfo::getGOT(), false, false, false, 0);
2518 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2519 SelectionDAG &DAG) const {
2520 assert(Subtarget->isTargetELF() &&
2521 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2522 MachineFunction &MF = DAG.getMachineFunction();
2523 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2524 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2525 EVT PtrVT = getPointerTy();
2527 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2528 ARMConstantPoolValue *CPV =
2529 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2530 ARMPCLabelIndex, PCAdj);
2531 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2532 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2533 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2534 MachinePointerInfo::getConstantPool(),
2535 false, false, false, 0);
2536 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2537 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2541 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2543 SDValue Val = DAG.getConstant(0, MVT::i32);
2544 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2545 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2546 Op.getOperand(1), Val);
2550 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2552 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2553 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2557 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2558 const ARMSubtarget *Subtarget) const {
2559 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2562 default: return SDValue(); // Don't custom lower most intrinsics.
2563 case Intrinsic::arm_thread_pointer: {
2564 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2565 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2567 case Intrinsic::eh_sjlj_lsda: {
2568 MachineFunction &MF = DAG.getMachineFunction();
2569 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2570 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2571 EVT PtrVT = getPointerTy();
2572 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2574 unsigned PCAdj = (RelocM != Reloc::PIC_)
2575 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2576 ARMConstantPoolValue *CPV =
2577 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2578 ARMCP::CPLSDA, PCAdj);
2579 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2580 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2582 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2583 MachinePointerInfo::getConstantPool(),
2584 false, false, false, 0);
2586 if (RelocM == Reloc::PIC_) {
2587 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2588 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2592 case Intrinsic::arm_neon_vmulls:
2593 case Intrinsic::arm_neon_vmullu: {
2594 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2595 ? ARMISD::VMULLs : ARMISD::VMULLu;
2596 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2597 Op.getOperand(1), Op.getOperand(2));
2602 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2603 const ARMSubtarget *Subtarget) {
2604 // FIXME: handle "fence singlethread" more efficiently.
2606 if (!Subtarget->hasDataBarrier()) {
2607 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2608 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2610 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2611 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2612 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2613 DAG.getConstant(0, MVT::i32));
2616 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2617 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2618 unsigned Domain = ARM_MB::ISH;
2619 if (Subtarget->isMClass()) {
2620 // Only a full system barrier exists in the M-class architectures.
2621 Domain = ARM_MB::SY;
2622 } else if (Subtarget->isSwift() && Ord == Release) {
2623 // Swift happens to implement ISHST barriers in a way that's compatible with
2624 // Release semantics but weaker than ISH so we'd be fools not to use
2625 // it. Beware: other processors probably don't!
2626 Domain = ARM_MB::ISHST;
2629 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2630 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2631 DAG.getConstant(Domain, MVT::i32));
2634 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2635 const ARMSubtarget *Subtarget) {
2636 // ARM pre v5TE and Thumb1 does not have preload instructions.
2637 if (!(Subtarget->isThumb2() ||
2638 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2639 // Just preserve the chain.
2640 return Op.getOperand(0);
2643 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2645 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2646 // ARMv7 with MP extension has PLDW.
2647 return Op.getOperand(0);
2649 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2650 if (Subtarget->isThumb()) {
2652 isRead = ~isRead & 1;
2653 isData = ~isData & 1;
2656 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2657 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2658 DAG.getConstant(isData, MVT::i32));
2661 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2662 MachineFunction &MF = DAG.getMachineFunction();
2663 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2665 // vastart just stores the address of the VarArgsFrameIndex slot into the
2666 // memory location argument.
2668 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2669 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2670 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2671 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2672 MachinePointerInfo(SV), false, false, 0);
2676 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2677 SDValue &Root, SelectionDAG &DAG,
2679 MachineFunction &MF = DAG.getMachineFunction();
2680 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2682 const TargetRegisterClass *RC;
2683 if (AFI->isThumb1OnlyFunction())
2684 RC = &ARM::tGPRRegClass;
2686 RC = &ARM::GPRRegClass;
2688 // Transform the arguments stored in physical registers into virtual ones.
2689 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2690 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2693 if (NextVA.isMemLoc()) {
2694 MachineFrameInfo *MFI = MF.getFrameInfo();
2695 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2697 // Create load node to retrieve arguments from the stack.
2698 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2699 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2700 MachinePointerInfo::getFixedStack(FI),
2701 false, false, false, 0);
2703 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2704 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2707 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2711 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2712 unsigned InRegsParamRecordIdx,
2714 unsigned &ArgRegsSize,
2715 unsigned &ArgRegsSaveSize)
2718 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2719 unsigned RBegin, REnd;
2720 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2721 NumGPRs = REnd - RBegin;
2723 unsigned int firstUnalloced;
2724 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2725 sizeof(GPRArgRegs) /
2726 sizeof(GPRArgRegs[0]));
2727 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2730 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2731 ArgRegsSize = NumGPRs * 4;
2733 // If parameter is split between stack and GPRs...
2734 if (NumGPRs && Align > 4 &&
2735 (ArgRegsSize < ArgSize ||
2736 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2737 // Add padding for part of param recovered from GPRs. For example,
2738 // if Align == 8, its last byte must be at address K*8 - 1.
2739 // We need to do it, since remained (stack) part of parameter has
2740 // stack alignment, and we need to "attach" "GPRs head" without gaps
2743 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2744 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2746 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2748 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2749 ArgRegsSaveSize = ArgRegsSize + Padding;
2751 // We don't need to extend regs save size for byval parameters if they
2752 // are passed via GPRs only.
2753 ArgRegsSaveSize = ArgRegsSize;
2756 // The remaining GPRs hold either the beginning of variable-argument
2757 // data, or the beginning of an aggregate passed by value (usually
2758 // byval). Either way, we allocate stack slots adjacent to the data
2759 // provided by our caller, and store the unallocated registers there.
2760 // If this is a variadic function, the va_list pointer will begin with
2761 // these values; otherwise, this reassembles a (byval) structure that
2762 // was split between registers and memory.
2763 // Return: The frame index registers were stored into.
2765 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2766 SDLoc dl, SDValue &Chain,
2767 const Value *OrigArg,
2768 unsigned InRegsParamRecordIdx,
2769 unsigned OffsetFromOrigArg,
2773 unsigned ByValStoreOffset,
2774 unsigned TotalArgRegsSaveSize) const {
2776 // Currently, two use-cases possible:
2777 // Case #1. Non-var-args function, and we meet first byval parameter.
2778 // Setup first unallocated register as first byval register;
2779 // eat all remained registers
2780 // (these two actions are performed by HandleByVal method).
2781 // Then, here, we initialize stack frame with
2782 // "store-reg" instructions.
2783 // Case #2. Var-args function, that doesn't contain byval parameters.
2784 // The same: eat all remained unallocated registers,
2785 // initialize stack frame.
2787 MachineFunction &MF = DAG.getMachineFunction();
2788 MachineFrameInfo *MFI = MF.getFrameInfo();
2789 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2790 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2791 unsigned RBegin, REnd;
2792 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2793 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2794 firstRegToSaveIndex = RBegin - ARM::R0;
2795 lastRegToSaveIndex = REnd - ARM::R0;
2797 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2798 (GPRArgRegs, array_lengthof(GPRArgRegs));
2799 lastRegToSaveIndex = 4;
2802 unsigned ArgRegsSize, ArgRegsSaveSize;
2803 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2804 ArgRegsSize, ArgRegsSaveSize);
2806 // Store any by-val regs to their spots on the stack so that they may be
2807 // loaded by deferencing the result of formal parameter pointer or va_next.
2808 // Note: once stack area for byval/varargs registers
2809 // was initialized, it can't be initialized again.
2810 if (ArgRegsSaveSize) {
2811 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2814 assert(AFI->getStoredByValParamsPadding() == 0 &&
2815 "The only parameter may be padded.");
2816 AFI->setStoredByValParamsPadding(Padding);
2819 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2822 (int64_t)TotalArgRegsSaveSize,
2824 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2826 MFI->CreateFixedObject(Padding,
2827 ArgOffset + ByValStoreOffset -
2828 (int64_t)ArgRegsSaveSize,
2832 SmallVector<SDValue, 4> MemOps;
2833 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2834 ++firstRegToSaveIndex, ++i) {
2835 const TargetRegisterClass *RC;
2836 if (AFI->isThumb1OnlyFunction())
2837 RC = &ARM::tGPRRegClass;
2839 RC = &ARM::GPRRegClass;
2841 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2842 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2844 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2845 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2847 MemOps.push_back(Store);
2848 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2849 DAG.getConstant(4, getPointerTy()));
2852 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2854 if (!MemOps.empty())
2855 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2856 &MemOps[0], MemOps.size());
2860 // We cannot allocate a zero-byte object for the first variadic argument,
2861 // so just make up a size.
2864 // This will point to the next argument passed via stack.
2865 return MFI->CreateFixedObject(
2866 ArgSize, ArgOffset, !ForceMutable);
2870 // Setup stack frame, the va_list pointer will start from.
2872 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2873 SDLoc dl, SDValue &Chain,
2875 unsigned TotalArgRegsSaveSize,
2876 bool ForceMutable) const {
2877 MachineFunction &MF = DAG.getMachineFunction();
2878 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2880 // Try to store any remaining integer argument regs
2881 // to their spots on the stack so that they may be loaded by deferencing
2882 // the result of va_next.
2883 // If there is no regs to be stored, just point address after last
2884 // argument passed via stack.
2886 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2887 0, ArgOffset, 0, ForceMutable, 0, TotalArgRegsSaveSize);
2889 AFI->setVarArgsFrameIndex(FrameIndex);
2893 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2894 CallingConv::ID CallConv, bool isVarArg,
2895 const SmallVectorImpl<ISD::InputArg>
2897 SDLoc dl, SelectionDAG &DAG,
2898 SmallVectorImpl<SDValue> &InVals)
2900 MachineFunction &MF = DAG.getMachineFunction();
2901 MachineFrameInfo *MFI = MF.getFrameInfo();
2903 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2905 // Assign locations to all of the incoming arguments.
2906 SmallVector<CCValAssign, 16> ArgLocs;
2907 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2908 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2909 CCInfo.AnalyzeFormalArguments(Ins,
2910 CCAssignFnForNode(CallConv, /* Return*/ false,
2913 SmallVector<SDValue, 16> ArgValues;
2914 int lastInsIndex = -1;
2916 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2917 unsigned CurArgIdx = 0;
2919 // Initially ArgRegsSaveSize is zero.
2920 // Then we increase this value each time we meet byval parameter.
2921 // We also increase this value in case of varargs function.
2922 AFI->setArgRegsSaveSize(0);
2924 unsigned ByValStoreOffset = 0;
2925 unsigned TotalArgRegsSaveSize = 0;
2926 unsigned ArgRegsSaveSizeMaxAlign = 4;
2928 // Calculate the amount of stack space that we need to allocate to store
2929 // byval and variadic arguments that are passed in registers.
2930 // We need to know this before we allocate the first byval or variadic
2931 // argument, as they will be allocated a stack slot below the CFA (Canonical
2932 // Frame Address, the stack pointer at entry to the function).
2933 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2934 CCValAssign &VA = ArgLocs[i];
2935 if (VA.isMemLoc()) {
2936 int index = VA.getValNo();
2937 if (index != lastInsIndex) {
2938 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2939 if (Flags.isByVal()) {
2940 unsigned ExtraArgRegsSize;
2941 unsigned ExtraArgRegsSaveSize;
2942 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2943 Flags.getByValSize(),
2944 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2946 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2947 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2948 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2949 CCInfo.nextInRegsParam();
2951 lastInsIndex = index;
2955 CCInfo.rewindByValRegsInfo();
2958 unsigned ExtraArgRegsSize;
2959 unsigned ExtraArgRegsSaveSize;
2960 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2961 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2962 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2964 // If the arg regs save area contains N-byte aligned values, the
2965 // bottom of it must be at least N-byte aligned.
2966 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2967 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2969 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2970 CCValAssign &VA = ArgLocs[i];
2971 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2972 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2973 // Arguments stored in registers.
2974 if (VA.isRegLoc()) {
2975 EVT RegVT = VA.getLocVT();
2977 if (VA.needsCustom()) {
2978 // f64 and vector types are split up into multiple registers or
2979 // combinations of registers and stack slots.
2980 if (VA.getLocVT() == MVT::v2f64) {
2981 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2983 VA = ArgLocs[++i]; // skip ahead to next loc
2985 if (VA.isMemLoc()) {
2986 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2987 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2988 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2989 MachinePointerInfo::getFixedStack(FI),
2990 false, false, false, 0);
2992 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2995 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2996 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2997 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2998 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2999 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3001 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3004 const TargetRegisterClass *RC;
3006 if (RegVT == MVT::f32)
3007 RC = &ARM::SPRRegClass;
3008 else if (RegVT == MVT::f64)
3009 RC = &ARM::DPRRegClass;
3010 else if (RegVT == MVT::v2f64)
3011 RC = &ARM::QPRRegClass;
3012 else if (RegVT == MVT::i32)
3013 RC = AFI->isThumb1OnlyFunction() ?
3014 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3015 (const TargetRegisterClass*)&ARM::GPRRegClass;
3017 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3019 // Transform the arguments in physical registers into virtual ones.
3020 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3021 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3024 // If this is an 8 or 16-bit value, it is really passed promoted
3025 // to 32 bits. Insert an assert[sz]ext to capture this, then
3026 // truncate to the right size.
3027 switch (VA.getLocInfo()) {
3028 default: llvm_unreachable("Unknown loc info!");
3029 case CCValAssign::Full: break;
3030 case CCValAssign::BCvt:
3031 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3033 case CCValAssign::SExt:
3034 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3035 DAG.getValueType(VA.getValVT()));
3036 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3038 case CCValAssign::ZExt:
3039 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3040 DAG.getValueType(VA.getValVT()));
3041 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3045 InVals.push_back(ArgValue);
3047 } else { // VA.isRegLoc()
3050 assert(VA.isMemLoc());
3051 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3053 int index = ArgLocs[i].getValNo();
3055 // Some Ins[] entries become multiple ArgLoc[] entries.
3056 // Process them only once.
3057 if (index != lastInsIndex)
3059 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3060 // FIXME: For now, all byval parameter objects are marked mutable.
3061 // This can be changed with more analysis.
3062 // In case of tail call optimization mark all arguments mutable.
3063 // Since they could be overwritten by lowering of arguments in case of
3065 if (Flags.isByVal()) {
3066 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3068 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3069 int FrameIndex = StoreByValRegs(
3070 CCInfo, DAG, dl, Chain, CurOrigArg,
3072 Ins[VA.getValNo()].PartOffset,
3073 VA.getLocMemOffset(),
3074 Flags.getByValSize(),
3075 true /*force mutable frames*/,
3077 TotalArgRegsSaveSize);
3078 ByValStoreOffset += Flags.getByValSize();
3079 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3080 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3081 CCInfo.nextInRegsParam();
3083 unsigned FIOffset = VA.getLocMemOffset();
3084 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3087 // Create load nodes to retrieve arguments from the stack.
3088 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3089 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3090 MachinePointerInfo::getFixedStack(FI),
3091 false, false, false, 0));
3093 lastInsIndex = index;
3100 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3101 CCInfo.getNextStackOffset(),
3102 TotalArgRegsSaveSize);
3107 /// isFloatingPointZero - Return true if this is +0.0.
3108 static bool isFloatingPointZero(SDValue Op) {
3109 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3110 return CFP->getValueAPF().isPosZero();
3111 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3112 // Maybe this has already been legalized into the constant pool?
3113 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3114 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3115 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3116 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3117 return CFP->getValueAPF().isPosZero();
3123 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3124 /// the given operands.
3126 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3127 SDValue &ARMcc, SelectionDAG &DAG,
3129 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3130 unsigned C = RHSC->getZExtValue();
3131 if (!isLegalICmpImmediate(C)) {
3132 // Constant does not fit, try adjusting it by one?
3137 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3138 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3139 RHS = DAG.getConstant(C-1, MVT::i32);
3144 if (C != 0 && isLegalICmpImmediate(C-1)) {
3145 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3146 RHS = DAG.getConstant(C-1, MVT::i32);
3151 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3152 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3153 RHS = DAG.getConstant(C+1, MVT::i32);
3158 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3159 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3160 RHS = DAG.getConstant(C+1, MVT::i32);
3167 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3168 ARMISD::NodeType CompareType;
3171 CompareType = ARMISD::CMP;
3176 CompareType = ARMISD::CMPZ;
3179 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3180 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3183 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3185 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3188 if (!isFloatingPointZero(RHS))
3189 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3191 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3192 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3195 /// duplicateCmp - Glue values can have only one use, so this function
3196 /// duplicates a comparison node.
3198 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3199 unsigned Opc = Cmp.getOpcode();
3201 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3202 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3204 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3205 Cmp = Cmp.getOperand(0);
3206 Opc = Cmp.getOpcode();
3207 if (Opc == ARMISD::CMPFP)
3208 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3210 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3211 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3213 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3216 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3217 SDValue Cond = Op.getOperand(0);
3218 SDValue SelectTrue = Op.getOperand(1);
3219 SDValue SelectFalse = Op.getOperand(2);
3224 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3225 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3227 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3228 const ConstantSDNode *CMOVTrue =
3229 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3230 const ConstantSDNode *CMOVFalse =
3231 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3233 if (CMOVTrue && CMOVFalse) {
3234 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3235 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3239 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3241 False = SelectFalse;
3242 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3247 if (True.getNode() && False.getNode()) {
3248 EVT VT = Op.getValueType();
3249 SDValue ARMcc = Cond.getOperand(2);
3250 SDValue CCR = Cond.getOperand(3);
3251 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3252 assert(True.getValueType() == VT);
3253 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3258 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3259 // undefined bits before doing a full-word comparison with zero.
3260 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3261 DAG.getConstant(1, Cond.getValueType()));
3263 return DAG.getSelectCC(dl, Cond,
3264 DAG.getConstant(0, Cond.getValueType()),
3265 SelectTrue, SelectFalse, ISD::SETNE);
3268 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3269 if (CC == ISD::SETNE)
3271 return ISD::getSetCCInverse(CC, true);
3274 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3275 bool &swpCmpOps, bool &swpVselOps) {
3276 // Start by selecting the GE condition code for opcodes that return true for
3278 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3280 CondCode = ARMCC::GE;
3282 // and GT for opcodes that return false for 'equality'.
3283 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3285 CondCode = ARMCC::GT;
3287 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3288 // to swap the compare operands.
3289 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3293 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3294 // If we have an unordered opcode, we need to swap the operands to the VSEL
3295 // instruction (effectively negating the condition).
3297 // This also has the effect of swapping which one of 'less' or 'greater'
3298 // returns true, so we also swap the compare operands. It also switches
3299 // whether we return true for 'equality', so we compensate by picking the
3300 // opposite condition code to our original choice.
3301 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3302 CC == ISD::SETUGT) {
3303 swpCmpOps = !swpCmpOps;
3304 swpVselOps = !swpVselOps;
3305 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3308 // 'ordered' is 'anything but unordered', so use the VS condition code and
3309 // swap the VSEL operands.
3310 if (CC == ISD::SETO) {
3311 CondCode = ARMCC::VS;
3315 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3316 // code and swap the VSEL operands.
3317 if (CC == ISD::SETUNE) {
3318 CondCode = ARMCC::EQ;
3323 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3324 EVT VT = Op.getValueType();
3325 SDValue LHS = Op.getOperand(0);
3326 SDValue RHS = Op.getOperand(1);
3327 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3328 SDValue TrueVal = Op.getOperand(2);
3329 SDValue FalseVal = Op.getOperand(3);
3332 if (LHS.getValueType() == MVT::i32) {
3333 // Try to generate VSEL on ARMv8.
3334 // The VSEL instruction can't use all the usual ARM condition
3335 // codes: it only has two bits to select the condition code, so it's
3336 // constrained to use only GE, GT, VS and EQ.
3338 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3339 // swap the operands of the previous compare instruction (effectively
3340 // inverting the compare condition, swapping 'less' and 'greater') and
3341 // sometimes need to swap the operands to the VSEL (which inverts the
3342 // condition in the sense of firing whenever the previous condition didn't)
3343 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3344 TrueVal.getValueType() == MVT::f64)) {
3345 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3346 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3347 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3348 CC = getInverseCCForVSEL(CC);
3349 std::swap(TrueVal, FalseVal);
3354 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3355 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3356 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3360 ARMCC::CondCodes CondCode, CondCode2;
3361 FPCCToARMCC(CC, CondCode, CondCode2);
3363 // Try to generate VSEL on ARMv8.
3364 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3365 TrueVal.getValueType() == MVT::f64)) {
3366 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3367 // same operands, as follows:
3368 // c = fcmp [ogt, olt, ugt, ult] a, b
3370 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3371 // handled differently than the original code sequence.
3372 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3374 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3375 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3376 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3377 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3380 bool swpCmpOps = false;
3381 bool swpVselOps = false;
3382 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3384 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3385 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3387 std::swap(LHS, RHS);
3389 std::swap(TrueVal, FalseVal);
3393 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3394 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3395 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3396 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3398 if (CondCode2 != ARMCC::AL) {
3399 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3400 // FIXME: Needs another CMP because flag can have but one use.
3401 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3402 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3403 Result, TrueVal, ARMcc2, CCR, Cmp2);
3408 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3409 /// to morph to an integer compare sequence.
3410 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3411 const ARMSubtarget *Subtarget) {
3412 SDNode *N = Op.getNode();
3413 if (!N->hasOneUse())
3414 // Otherwise it requires moving the value from fp to integer registers.
3416 if (!N->getNumValues())
3418 EVT VT = Op.getValueType();
3419 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3420 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3421 // vmrs are very slow, e.g. cortex-a8.
3424 if (isFloatingPointZero(Op)) {
3428 return ISD::isNormalLoad(N);
3431 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3432 if (isFloatingPointZero(Op))
3433 return DAG.getConstant(0, MVT::i32);
3435 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3436 return DAG.getLoad(MVT::i32, SDLoc(Op),
3437 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3438 Ld->isVolatile(), Ld->isNonTemporal(),
3439 Ld->isInvariant(), Ld->getAlignment());
3441 llvm_unreachable("Unknown VFP cmp argument!");
3444 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3445 SDValue &RetVal1, SDValue &RetVal2) {
3446 if (isFloatingPointZero(Op)) {
3447 RetVal1 = DAG.getConstant(0, MVT::i32);
3448 RetVal2 = DAG.getConstant(0, MVT::i32);
3452 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3453 SDValue Ptr = Ld->getBasePtr();
3454 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3455 Ld->getChain(), Ptr,
3456 Ld->getPointerInfo(),
3457 Ld->isVolatile(), Ld->isNonTemporal(),
3458 Ld->isInvariant(), Ld->getAlignment());
3460 EVT PtrType = Ptr.getValueType();
3461 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3462 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3463 PtrType, Ptr, DAG.getConstant(4, PtrType));
3464 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3465 Ld->getChain(), NewPtr,
3466 Ld->getPointerInfo().getWithOffset(4),
3467 Ld->isVolatile(), Ld->isNonTemporal(),
3468 Ld->isInvariant(), NewAlign);
3472 llvm_unreachable("Unknown VFP cmp argument!");
3475 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3476 /// f32 and even f64 comparisons to integer ones.
3478 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3479 SDValue Chain = Op.getOperand(0);
3480 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3481 SDValue LHS = Op.getOperand(2);
3482 SDValue RHS = Op.getOperand(3);
3483 SDValue Dest = Op.getOperand(4);
3486 bool LHSSeenZero = false;
3487 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3488 bool RHSSeenZero = false;
3489 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3490 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3491 // If unsafe fp math optimization is enabled and there are no other uses of
3492 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3493 // to an integer comparison.
3494 if (CC == ISD::SETOEQ)
3496 else if (CC == ISD::SETUNE)
3499 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3501 if (LHS.getValueType() == MVT::f32) {
3502 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3503 bitcastf32Toi32(LHS, DAG), Mask);
3504 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3505 bitcastf32Toi32(RHS, DAG), Mask);
3506 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3507 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3508 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3509 Chain, Dest, ARMcc, CCR, Cmp);
3514 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3515 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3516 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3517 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3518 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3519 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3520 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3521 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3522 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3528 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3529 SDValue Chain = Op.getOperand(0);
3530 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3531 SDValue LHS = Op.getOperand(2);
3532 SDValue RHS = Op.getOperand(3);
3533 SDValue Dest = Op.getOperand(4);
3536 if (LHS.getValueType() == MVT::i32) {
3538 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3539 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3540 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3541 Chain, Dest, ARMcc, CCR, Cmp);
3544 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3546 if (getTargetMachine().Options.UnsafeFPMath &&
3547 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3548 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3549 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3550 if (Result.getNode())
3554 ARMCC::CondCodes CondCode, CondCode2;
3555 FPCCToARMCC(CC, CondCode, CondCode2);
3557 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3558 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3559 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3560 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3561 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3562 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3563 if (CondCode2 != ARMCC::AL) {
3564 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3565 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3566 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3571 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3572 SDValue Chain = Op.getOperand(0);
3573 SDValue Table = Op.getOperand(1);
3574 SDValue Index = Op.getOperand(2);
3577 EVT PTy = getPointerTy();
3578 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3579 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3580 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3581 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3582 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3583 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3584 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3585 if (Subtarget->isThumb2()) {
3586 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3587 // which does another jump to the destination. This also makes it easier
3588 // to translate it to TBB / TBH later.
3589 // FIXME: This might not work if the function is extremely large.
3590 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3591 Addr, Op.getOperand(2), JTI, UId);
3593 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3594 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3595 MachinePointerInfo::getJumpTable(),
3596 false, false, false, 0);
3597 Chain = Addr.getValue(1);
3598 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3599 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3601 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3602 MachinePointerInfo::getJumpTable(),
3603 false, false, false, 0);
3604 Chain = Addr.getValue(1);
3605 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3609 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3610 EVT VT = Op.getValueType();
3613 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3614 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3616 return DAG.UnrollVectorOp(Op.getNode());
3619 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3620 "Invalid type for custom lowering!");
3621 if (VT != MVT::v4i16)
3622 return DAG.UnrollVectorOp(Op.getNode());
3624 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3625 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3628 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3629 EVT VT = Op.getValueType();
3631 return LowerVectorFP_TO_INT(Op, DAG);
3636 switch (Op.getOpcode()) {
3637 default: llvm_unreachable("Invalid opcode!");
3638 case ISD::FP_TO_SINT:
3639 Opc = ARMISD::FTOSI;
3641 case ISD::FP_TO_UINT:
3642 Opc = ARMISD::FTOUI;
3645 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3646 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3649 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3650 EVT VT = Op.getValueType();
3653 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3654 if (VT.getVectorElementType() == MVT::f32)
3656 return DAG.UnrollVectorOp(Op.getNode());
3659 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3660 "Invalid type for custom lowering!");
3661 if (VT != MVT::v4f32)
3662 return DAG.UnrollVectorOp(Op.getNode());
3666 switch (Op.getOpcode()) {
3667 default: llvm_unreachable("Invalid opcode!");
3668 case ISD::SINT_TO_FP:
3669 CastOpc = ISD::SIGN_EXTEND;
3670 Opc = ISD::SINT_TO_FP;
3672 case ISD::UINT_TO_FP:
3673 CastOpc = ISD::ZERO_EXTEND;
3674 Opc = ISD::UINT_TO_FP;
3678 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3679 return DAG.getNode(Opc, dl, VT, Op);
3682 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3683 EVT VT = Op.getValueType();
3685 return LowerVectorINT_TO_FP(Op, DAG);
3690 switch (Op.getOpcode()) {
3691 default: llvm_unreachable("Invalid opcode!");
3692 case ISD::SINT_TO_FP:
3693 Opc = ARMISD::SITOF;
3695 case ISD::UINT_TO_FP:
3696 Opc = ARMISD::UITOF;
3700 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3701 return DAG.getNode(Opc, dl, VT, Op);
3704 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3705 // Implement fcopysign with a fabs and a conditional fneg.
3706 SDValue Tmp0 = Op.getOperand(0);
3707 SDValue Tmp1 = Op.getOperand(1);
3709 EVT VT = Op.getValueType();
3710 EVT SrcVT = Tmp1.getValueType();
3711 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3712 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3713 bool UseNEON = !InGPR && Subtarget->hasNEON();
3716 // Use VBSL to copy the sign bit.
3717 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3718 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3719 DAG.getTargetConstant(EncodedVal, MVT::i32));
3720 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3722 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3723 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3724 DAG.getConstant(32, MVT::i32));
3725 else /*if (VT == MVT::f32)*/
3726 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3727 if (SrcVT == MVT::f32) {
3728 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3730 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3731 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3732 DAG.getConstant(32, MVT::i32));
3733 } else if (VT == MVT::f32)
3734 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3735 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3736 DAG.getConstant(32, MVT::i32));
3737 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3738 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3740 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3742 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3743 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3744 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3746 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3747 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3748 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3749 if (VT == MVT::f32) {
3750 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3751 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3752 DAG.getConstant(0, MVT::i32));
3754 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3760 // Bitcast operand 1 to i32.
3761 if (SrcVT == MVT::f64)
3762 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3763 &Tmp1, 1).getValue(1);
3764 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3766 // Or in the signbit with integer operations.
3767 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3768 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3769 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3770 if (VT == MVT::f32) {
3771 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3772 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3773 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3774 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3777 // f64: Or the high part with signbit and then combine two parts.
3778 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3780 SDValue Lo = Tmp0.getValue(0);
3781 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3782 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3783 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3786 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3787 MachineFunction &MF = DAG.getMachineFunction();
3788 MachineFrameInfo *MFI = MF.getFrameInfo();
3789 MFI->setReturnAddressIsTaken(true);
3791 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3794 EVT VT = Op.getValueType();
3796 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3798 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3799 SDValue Offset = DAG.getConstant(4, MVT::i32);
3800 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3801 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3802 MachinePointerInfo(), false, false, false, 0);
3805 // Return LR, which contains the return address. Mark it an implicit live-in.
3806 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3807 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3810 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3811 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3812 MFI->setFrameAddressIsTaken(true);
3814 EVT VT = Op.getValueType();
3815 SDLoc dl(Op); // FIXME probably not meaningful
3816 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3817 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetMachO())
3818 ? ARM::R7 : ARM::R11;
3819 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3821 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3822 MachinePointerInfo(),
3823 false, false, false, 0);
3827 /// ExpandBITCAST - If the target supports VFP, this function is called to
3828 /// expand a bit convert where either the source or destination type is i64 to
3829 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3830 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3831 /// vectors), since the legalizer won't know what to do with that.
3832 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3833 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3835 SDValue Op = N->getOperand(0);
3837 // This function is only supposed to be called for i64 types, either as the
3838 // source or destination of the bit convert.
3839 EVT SrcVT = Op.getValueType();
3840 EVT DstVT = N->getValueType(0);
3841 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3842 "ExpandBITCAST called for non-i64 type");
3844 // Turn i64->f64 into VMOVDRR.
3845 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3846 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3847 DAG.getConstant(0, MVT::i32));
3848 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3849 DAG.getConstant(1, MVT::i32));
3850 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3851 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3854 // Turn f64->i64 into VMOVRRD.
3855 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3856 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3857 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3858 // Merge the pieces into a single i64 value.
3859 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3865 /// getZeroVector - Returns a vector of specified type with all zero elements.
3866 /// Zero vectors are used to represent vector negation and in those cases
3867 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3868 /// not support i64 elements, so sometimes the zero vectors will need to be
3869 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3871 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3872 assert(VT.isVector() && "Expected a vector type");
3873 // The canonical modified immediate encoding of a zero vector is....0!
3874 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3875 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3876 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3877 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3880 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3881 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3882 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3883 SelectionDAG &DAG) const {
3884 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3885 EVT VT = Op.getValueType();
3886 unsigned VTBits = VT.getSizeInBits();
3888 SDValue ShOpLo = Op.getOperand(0);
3889 SDValue ShOpHi = Op.getOperand(1);
3890 SDValue ShAmt = Op.getOperand(2);
3892 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3894 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3896 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3897 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3898 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3899 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3900 DAG.getConstant(VTBits, MVT::i32));
3901 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3902 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3903 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3905 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3906 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3908 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3909 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3912 SDValue Ops[2] = { Lo, Hi };
3913 return DAG.getMergeValues(Ops, 2, dl);
3916 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3917 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3918 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3919 SelectionDAG &DAG) const {
3920 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3921 EVT VT = Op.getValueType();
3922 unsigned VTBits = VT.getSizeInBits();
3924 SDValue ShOpLo = Op.getOperand(0);
3925 SDValue ShOpHi = Op.getOperand(1);
3926 SDValue ShAmt = Op.getOperand(2);
3929 assert(Op.getOpcode() == ISD::SHL_PARTS);
3930 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3931 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3932 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3933 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3934 DAG.getConstant(VTBits, MVT::i32));
3935 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3936 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3938 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3939 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3940 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3942 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3943 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3946 SDValue Ops[2] = { Lo, Hi };
3947 return DAG.getMergeValues(Ops, 2, dl);
3950 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3951 SelectionDAG &DAG) const {
3952 // The rounding mode is in bits 23:22 of the FPSCR.
3953 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3954 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3955 // so that the shift + and get folded into a bitfield extract.
3957 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3958 DAG.getConstant(Intrinsic::arm_get_fpscr,
3960 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3961 DAG.getConstant(1U << 22, MVT::i32));
3962 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3963 DAG.getConstant(22, MVT::i32));
3964 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3965 DAG.getConstant(3, MVT::i32));
3968 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3969 const ARMSubtarget *ST) {
3970 EVT VT = N->getValueType(0);
3973 if (!ST->hasV6T2Ops())
3976 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3977 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3980 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3981 /// for each 16-bit element from operand, repeated. The basic idea is to
3982 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3984 /// Trace for v4i16:
3985 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3986 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3987 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3988 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3989 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3990 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3991 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3992 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3993 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3994 EVT VT = N->getValueType(0);
3997 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3998 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3999 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4000 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4001 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4002 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4005 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4006 /// bit-count for each 16-bit element from the operand. We need slightly
4007 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4008 /// 64/128-bit registers.
4010 /// Trace for v4i16:
4011 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4012 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4013 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4014 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4015 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4016 EVT VT = N->getValueType(0);
4019 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4020 if (VT.is64BitVector()) {
4021 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4022 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4023 DAG.getIntPtrConstant(0));
4025 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4026 BitCounts, DAG.getIntPtrConstant(0));
4027 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4031 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4032 /// bit-count for each 32-bit element from the operand. The idea here is
4033 /// to split the vector into 16-bit elements, leverage the 16-bit count
4034 /// routine, and then combine the results.
4036 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4037 /// input = [v0 v1 ] (vi: 32-bit elements)
4038 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4039 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4040 /// vrev: N0 = [k1 k0 k3 k2 ]
4042 /// N1 =+[k1 k0 k3 k2 ]
4044 /// N2 =+[k1 k3 k0 k2 ]
4046 /// Extended =+[k1 k3 k0 k2 ]
4048 /// Extracted=+[k1 k3 ]
4050 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4051 EVT VT = N->getValueType(0);
4054 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4056 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4057 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4058 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4059 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4060 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4062 if (VT.is64BitVector()) {
4063 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4064 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4065 DAG.getIntPtrConstant(0));
4067 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4068 DAG.getIntPtrConstant(0));
4069 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4073 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4074 const ARMSubtarget *ST) {
4075 EVT VT = N->getValueType(0);
4077 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4078 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4079 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4080 "Unexpected type for custom ctpop lowering");
4082 if (VT.getVectorElementType() == MVT::i32)
4083 return lowerCTPOP32BitElements(N, DAG);
4085 return lowerCTPOP16BitElements(N, DAG);
4088 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4089 const ARMSubtarget *ST) {
4090 EVT VT = N->getValueType(0);
4096 // Lower vector shifts on NEON to use VSHL.
4097 assert(ST->hasNEON() && "unexpected vector shift");
4099 // Left shifts translate directly to the vshiftu intrinsic.
4100 if (N->getOpcode() == ISD::SHL)
4101 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4102 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4103 N->getOperand(0), N->getOperand(1));
4105 assert((N->getOpcode() == ISD::SRA ||
4106 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4108 // NEON uses the same intrinsics for both left and right shifts. For
4109 // right shifts, the shift amounts are negative, so negate the vector of
4111 EVT ShiftVT = N->getOperand(1).getValueType();
4112 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4113 getZeroVector(ShiftVT, DAG, dl),
4115 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4116 Intrinsic::arm_neon_vshifts :
4117 Intrinsic::arm_neon_vshiftu);
4118 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4119 DAG.getConstant(vshiftInt, MVT::i32),
4120 N->getOperand(0), NegatedCount);
4123 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4124 const ARMSubtarget *ST) {
4125 EVT VT = N->getValueType(0);
4128 // We can get here for a node like i32 = ISD::SHL i32, i64
4132 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4133 "Unknown shift to lower!");
4135 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4136 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4137 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4140 // If we are in thumb mode, we don't have RRX.
4141 if (ST->isThumb1Only()) return SDValue();
4143 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4144 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4145 DAG.getConstant(0, MVT::i32));
4146 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4147 DAG.getConstant(1, MVT::i32));
4149 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4150 // captures the result into a carry flag.
4151 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4152 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4154 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4155 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4157 // Merge the pieces into a single i64 value.
4158 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4161 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4162 SDValue TmpOp0, TmpOp1;
4163 bool Invert = false;
4167 SDValue Op0 = Op.getOperand(0);
4168 SDValue Op1 = Op.getOperand(1);
4169 SDValue CC = Op.getOperand(2);
4170 EVT VT = Op.getValueType();
4171 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4174 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4175 switch (SetCCOpcode) {
4176 default: llvm_unreachable("Illegal FP comparison");
4178 case ISD::SETNE: Invert = true; // Fallthrough
4180 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4182 case ISD::SETLT: Swap = true; // Fallthrough
4184 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4186 case ISD::SETLE: Swap = true; // Fallthrough
4188 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4189 case ISD::SETUGE: Swap = true; // Fallthrough
4190 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4191 case ISD::SETUGT: Swap = true; // Fallthrough
4192 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4193 case ISD::SETUEQ: Invert = true; // Fallthrough
4195 // Expand this to (OLT | OGT).
4199 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4200 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4202 case ISD::SETUO: Invert = true; // Fallthrough
4204 // Expand this to (OLT | OGE).
4208 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4209 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4213 // Integer comparisons.
4214 switch (SetCCOpcode) {
4215 default: llvm_unreachable("Illegal integer comparison");
4216 case ISD::SETNE: Invert = true;
4217 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4218 case ISD::SETLT: Swap = true;
4219 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4220 case ISD::SETLE: Swap = true;
4221 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4222 case ISD::SETULT: Swap = true;
4223 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4224 case ISD::SETULE: Swap = true;
4225 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4228 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4229 if (Opc == ARMISD::VCEQ) {
4232 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4234 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4237 // Ignore bitconvert.
4238 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4239 AndOp = AndOp.getOperand(0);
4241 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4243 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4244 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4251 std::swap(Op0, Op1);
4253 // If one of the operands is a constant vector zero, attempt to fold the
4254 // comparison to a specialized compare-against-zero form.
4256 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4258 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4259 if (Opc == ARMISD::VCGE)
4260 Opc = ARMISD::VCLEZ;
4261 else if (Opc == ARMISD::VCGT)
4262 Opc = ARMISD::VCLTZ;
4267 if (SingleOp.getNode()) {
4270 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4272 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4274 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4276 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4278 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4280 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4283 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4287 Result = DAG.getNOT(dl, Result, VT);
4292 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4293 /// valid vector constant for a NEON instruction with a "modified immediate"
4294 /// operand (e.g., VMOV). If so, return the encoded value.
4295 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4296 unsigned SplatBitSize, SelectionDAG &DAG,
4297 EVT &VT, bool is128Bits, NEONModImmType type) {
4298 unsigned OpCmode, Imm;
4300 // SplatBitSize is set to the smallest size that splats the vector, so a
4301 // zero vector will always have SplatBitSize == 8. However, NEON modified
4302 // immediate instructions others than VMOV do not support the 8-bit encoding
4303 // of a zero vector, and the default encoding of zero is supposed to be the
4308 switch (SplatBitSize) {
4310 if (type != VMOVModImm)
4312 // Any 1-byte value is OK. Op=0, Cmode=1110.
4313 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4316 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4320 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4321 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4322 if ((SplatBits & ~0xff) == 0) {
4323 // Value = 0x00nn: Op=x, Cmode=100x.
4328 if ((SplatBits & ~0xff00) == 0) {
4329 // Value = 0xnn00: Op=x, Cmode=101x.
4331 Imm = SplatBits >> 8;
4337 // NEON's 32-bit VMOV supports splat values where:
4338 // * only one byte is nonzero, or
4339 // * the least significant byte is 0xff and the second byte is nonzero, or
4340 // * the least significant 2 bytes are 0xff and the third is nonzero.
4341 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4342 if ((SplatBits & ~0xff) == 0) {
4343 // Value = 0x000000nn: Op=x, Cmode=000x.
4348 if ((SplatBits & ~0xff00) == 0) {
4349 // Value = 0x0000nn00: Op=x, Cmode=001x.
4351 Imm = SplatBits >> 8;
4354 if ((SplatBits & ~0xff0000) == 0) {
4355 // Value = 0x00nn0000: Op=x, Cmode=010x.
4357 Imm = SplatBits >> 16;
4360 if ((SplatBits & ~0xff000000) == 0) {
4361 // Value = 0xnn000000: Op=x, Cmode=011x.
4363 Imm = SplatBits >> 24;
4367 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4368 if (type == OtherModImm) return SDValue();
4370 if ((SplatBits & ~0xffff) == 0 &&
4371 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4372 // Value = 0x0000nnff: Op=x, Cmode=1100.
4374 Imm = SplatBits >> 8;
4378 if ((SplatBits & ~0xffffff) == 0 &&
4379 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4380 // Value = 0x00nnffff: Op=x, Cmode=1101.
4382 Imm = SplatBits >> 16;
4386 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4387 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4388 // VMOV.I32. A (very) minor optimization would be to replicate the value
4389 // and fall through here to test for a valid 64-bit splat. But, then the
4390 // caller would also need to check and handle the change in size.
4394 if (type != VMOVModImm)
4396 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4397 uint64_t BitMask = 0xff;
4399 unsigned ImmMask = 1;
4401 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4402 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4405 } else if ((SplatBits & BitMask) != 0) {
4411 // Op=1, Cmode=1110.
4413 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4418 llvm_unreachable("unexpected size for isNEONModifiedImm");
4421 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4422 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4425 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4426 const ARMSubtarget *ST) const {
4430 bool IsDouble = Op.getValueType() == MVT::f64;
4431 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4433 // Try splatting with a VMOV.f32...
4434 APFloat FPVal = CFP->getValueAPF();
4435 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4438 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4439 // We have code in place to select a valid ConstantFP already, no need to
4444 // It's a float and we are trying to use NEON operations where
4445 // possible. Lower it to a splat followed by an extract.
4447 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4448 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4450 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4451 DAG.getConstant(0, MVT::i32));
4454 // The rest of our options are NEON only, make sure that's allowed before
4456 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4460 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4462 // It wouldn't really be worth bothering for doubles except for one very
4463 // important value, which does happen to match: 0.0. So make sure we don't do
4465 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4468 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4469 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4471 if (NewVal != SDValue()) {
4473 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4476 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4478 // It's a float: cast and extract a vector element.
4479 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4481 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4482 DAG.getConstant(0, MVT::i32));
4485 // Finally, try a VMVN.i32
4486 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4488 if (NewVal != SDValue()) {
4490 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4493 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4495 // It's a float: cast and extract a vector element.
4496 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4498 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4499 DAG.getConstant(0, MVT::i32));
4505 // check if an VEXT instruction can handle the shuffle mask when the
4506 // vector sources of the shuffle are the same.
4507 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4508 unsigned NumElts = VT.getVectorNumElements();
4510 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4516 // If this is a VEXT shuffle, the immediate value is the index of the first
4517 // element. The other shuffle indices must be the successive elements after
4519 unsigned ExpectedElt = Imm;
4520 for (unsigned i = 1; i < NumElts; ++i) {
4521 // Increment the expected index. If it wraps around, just follow it
4522 // back to index zero and keep going.
4524 if (ExpectedElt == NumElts)
4527 if (M[i] < 0) continue; // ignore UNDEF indices
4528 if (ExpectedElt != static_cast<unsigned>(M[i]))
4536 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4537 bool &ReverseVEXT, unsigned &Imm) {
4538 unsigned NumElts = VT.getVectorNumElements();
4539 ReverseVEXT = false;
4541 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4547 // If this is a VEXT shuffle, the immediate value is the index of the first
4548 // element. The other shuffle indices must be the successive elements after
4550 unsigned ExpectedElt = Imm;
4551 for (unsigned i = 1; i < NumElts; ++i) {
4552 // Increment the expected index. If it wraps around, it may still be
4553 // a VEXT but the source vectors must be swapped.
4555 if (ExpectedElt == NumElts * 2) {
4560 if (M[i] < 0) continue; // ignore UNDEF indices
4561 if (ExpectedElt != static_cast<unsigned>(M[i]))
4565 // Adjust the index value if the source operands will be swapped.
4572 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4573 /// instruction with the specified blocksize. (The order of the elements
4574 /// within each block of the vector is reversed.)
4575 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4576 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4577 "Only possible block sizes for VREV are: 16, 32, 64");
4579 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4583 unsigned NumElts = VT.getVectorNumElements();
4584 unsigned BlockElts = M[0] + 1;
4585 // If the first shuffle index is UNDEF, be optimistic.
4587 BlockElts = BlockSize / EltSz;
4589 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4592 for (unsigned i = 0; i < NumElts; ++i) {
4593 if (M[i] < 0) continue; // ignore UNDEF indices
4594 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4601 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4602 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4603 // range, then 0 is placed into the resulting vector. So pretty much any mask
4604 // of 8 elements can work here.
4605 return VT == MVT::v8i8 && M.size() == 8;
4608 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4609 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4613 unsigned NumElts = VT.getVectorNumElements();
4614 WhichResult = (M[0] == 0 ? 0 : 1);
4615 for (unsigned i = 0; i < NumElts; i += 2) {
4616 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4617 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4623 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4624 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4625 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4626 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4627 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4631 unsigned NumElts = VT.getVectorNumElements();
4632 WhichResult = (M[0] == 0 ? 0 : 1);
4633 for (unsigned i = 0; i < NumElts; i += 2) {
4634 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4635 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4641 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4642 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4646 unsigned NumElts = VT.getVectorNumElements();
4647 WhichResult = (M[0] == 0 ? 0 : 1);
4648 for (unsigned i = 0; i != NumElts; ++i) {
4649 if (M[i] < 0) continue; // ignore UNDEF indices
4650 if ((unsigned) M[i] != 2 * i + WhichResult)
4654 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4655 if (VT.is64BitVector() && EltSz == 32)
4661 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4662 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4663 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4664 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4665 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4669 unsigned Half = VT.getVectorNumElements() / 2;
4670 WhichResult = (M[0] == 0 ? 0 : 1);
4671 for (unsigned j = 0; j != 2; ++j) {
4672 unsigned Idx = WhichResult;
4673 for (unsigned i = 0; i != Half; ++i) {
4674 int MIdx = M[i + j * Half];
4675 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4681 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4682 if (VT.is64BitVector() && EltSz == 32)
4688 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4689 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4693 unsigned NumElts = VT.getVectorNumElements();
4694 WhichResult = (M[0] == 0 ? 0 : 1);
4695 unsigned Idx = WhichResult * NumElts / 2;
4696 for (unsigned i = 0; i != NumElts; i += 2) {
4697 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4698 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4703 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4704 if (VT.is64BitVector() && EltSz == 32)
4710 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4711 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4712 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4713 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4714 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4718 unsigned NumElts = VT.getVectorNumElements();
4719 WhichResult = (M[0] == 0 ? 0 : 1);
4720 unsigned Idx = WhichResult * NumElts / 2;
4721 for (unsigned i = 0; i != NumElts; i += 2) {
4722 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4723 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4728 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4729 if (VT.is64BitVector() && EltSz == 32)
4735 /// \return true if this is a reverse operation on an vector.
4736 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4737 unsigned NumElts = VT.getVectorNumElements();
4738 // Make sure the mask has the right size.
4739 if (NumElts != M.size())
4742 // Look for <15, ..., 3, -1, 1, 0>.
4743 for (unsigned i = 0; i != NumElts; ++i)
4744 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4750 // If N is an integer constant that can be moved into a register in one
4751 // instruction, return an SDValue of such a constant (will become a MOV
4752 // instruction). Otherwise return null.
4753 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4754 const ARMSubtarget *ST, SDLoc dl) {
4756 if (!isa<ConstantSDNode>(N))
4758 Val = cast<ConstantSDNode>(N)->getZExtValue();
4760 if (ST->isThumb1Only()) {
4761 if (Val <= 255 || ~Val <= 255)
4762 return DAG.getConstant(Val, MVT::i32);
4764 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4765 return DAG.getConstant(Val, MVT::i32);
4770 // If this is a case we can't handle, return null and let the default
4771 // expansion code take care of it.
4772 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4773 const ARMSubtarget *ST) const {
4774 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4776 EVT VT = Op.getValueType();
4778 APInt SplatBits, SplatUndef;
4779 unsigned SplatBitSize;
4781 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4782 if (SplatBitSize <= 64) {
4783 // Check if an immediate VMOV works.
4785 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4786 SplatUndef.getZExtValue(), SplatBitSize,
4787 DAG, VmovVT, VT.is128BitVector(),
4789 if (Val.getNode()) {
4790 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4791 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4794 // Try an immediate VMVN.
4795 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4796 Val = isNEONModifiedImm(NegatedImm,
4797 SplatUndef.getZExtValue(), SplatBitSize,
4798 DAG, VmovVT, VT.is128BitVector(),
4800 if (Val.getNode()) {
4801 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4802 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4805 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4806 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4807 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4809 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4810 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4816 // Scan through the operands to see if only one value is used.
4818 // As an optimisation, even if more than one value is used it may be more
4819 // profitable to splat with one value then change some lanes.
4821 // Heuristically we decide to do this if the vector has a "dominant" value,
4822 // defined as splatted to more than half of the lanes.
4823 unsigned NumElts = VT.getVectorNumElements();
4824 bool isOnlyLowElement = true;
4825 bool usesOnlyOneValue = true;
4826 bool hasDominantValue = false;
4827 bool isConstant = true;
4829 // Map of the number of times a particular SDValue appears in the
4831 DenseMap<SDValue, unsigned> ValueCounts;
4833 for (unsigned i = 0; i < NumElts; ++i) {
4834 SDValue V = Op.getOperand(i);
4835 if (V.getOpcode() == ISD::UNDEF)
4838 isOnlyLowElement = false;
4839 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4842 ValueCounts.insert(std::make_pair(V, 0));
4843 unsigned &Count = ValueCounts[V];
4845 // Is this value dominant? (takes up more than half of the lanes)
4846 if (++Count > (NumElts / 2)) {
4847 hasDominantValue = true;
4851 if (ValueCounts.size() != 1)
4852 usesOnlyOneValue = false;
4853 if (!Value.getNode() && ValueCounts.size() > 0)
4854 Value = ValueCounts.begin()->first;
4856 if (ValueCounts.size() == 0)
4857 return DAG.getUNDEF(VT);
4859 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4860 // Keep going if we are hitting this case.
4861 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4862 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4864 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4866 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4867 // i32 and try again.
4868 if (hasDominantValue && EltSize <= 32) {
4872 // If we are VDUPing a value that comes directly from a vector, that will
4873 // cause an unnecessary move to and from a GPR, where instead we could
4874 // just use VDUPLANE. We can only do this if the lane being extracted
4875 // is at a constant index, as the VDUP from lane instructions only have
4876 // constant-index forms.
4877 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4878 isa<ConstantSDNode>(Value->getOperand(1))) {
4879 // We need to create a new undef vector to use for the VDUPLANE if the
4880 // size of the vector from which we get the value is different than the
4881 // size of the vector that we need to create. We will insert the element
4882 // such that the register coalescer will remove unnecessary copies.
4883 if (VT != Value->getOperand(0).getValueType()) {
4884 ConstantSDNode *constIndex;
4885 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4886 assert(constIndex && "The index is not a constant!");
4887 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4888 VT.getVectorNumElements();
4889 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4890 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4891 Value, DAG.getConstant(index, MVT::i32)),
4892 DAG.getConstant(index, MVT::i32));
4894 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4895 Value->getOperand(0), Value->getOperand(1));
4897 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4899 if (!usesOnlyOneValue) {
4900 // The dominant value was splatted as 'N', but we now have to insert
4901 // all differing elements.
4902 for (unsigned I = 0; I < NumElts; ++I) {
4903 if (Op.getOperand(I) == Value)
4905 SmallVector<SDValue, 3> Ops;
4907 Ops.push_back(Op.getOperand(I));
4908 Ops.push_back(DAG.getConstant(I, MVT::i32));
4909 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4914 if (VT.getVectorElementType().isFloatingPoint()) {
4915 SmallVector<SDValue, 8> Ops;
4916 for (unsigned i = 0; i < NumElts; ++i)
4917 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4919 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4920 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4921 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4923 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4925 if (usesOnlyOneValue) {
4926 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4927 if (isConstant && Val.getNode())
4928 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4932 // If all elements are constants and the case above didn't get hit, fall back
4933 // to the default expansion, which will generate a load from the constant
4938 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4940 SDValue shuffle = ReconstructShuffle(Op, DAG);
4941 if (shuffle != SDValue())
4945 // Vectors with 32- or 64-bit elements can be built by directly assigning
4946 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4947 // will be legalized.
4948 if (EltSize >= 32) {
4949 // Do the expansion with floating-point types, since that is what the VFP
4950 // registers are defined to use, and since i64 is not legal.
4951 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4952 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4953 SmallVector<SDValue, 8> Ops;
4954 for (unsigned i = 0; i < NumElts; ++i)
4955 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4956 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4957 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4960 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4961 // know the default expansion would otherwise fall back on something even
4962 // worse. For a vector with one or two non-undef values, that's
4963 // scalar_to_vector for the elements followed by a shuffle (provided the
4964 // shuffle is valid for the target) and materialization element by element
4965 // on the stack followed by a load for everything else.
4966 if (!isConstant && !usesOnlyOneValue) {
4967 SDValue Vec = DAG.getUNDEF(VT);
4968 for (unsigned i = 0 ; i < NumElts; ++i) {
4969 SDValue V = Op.getOperand(i);
4970 if (V.getOpcode() == ISD::UNDEF)
4972 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4973 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4981 // Gather data to see if the operation can be modelled as a
4982 // shuffle in combination with VEXTs.
4983 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4984 SelectionDAG &DAG) const {
4986 EVT VT = Op.getValueType();
4987 unsigned NumElts = VT.getVectorNumElements();
4989 SmallVector<SDValue, 2> SourceVecs;
4990 SmallVector<unsigned, 2> MinElts;
4991 SmallVector<unsigned, 2> MaxElts;
4993 for (unsigned i = 0; i < NumElts; ++i) {
4994 SDValue V = Op.getOperand(i);
4995 if (V.getOpcode() == ISD::UNDEF)
4997 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4998 // A shuffle can only come from building a vector from various
4999 // elements of other vectors.
5001 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5002 VT.getVectorElementType()) {
5003 // This code doesn't know how to handle shuffles where the vector
5004 // element types do not match (this happens because type legalization
5005 // promotes the return type of EXTRACT_VECTOR_ELT).
5006 // FIXME: It might be appropriate to extend this code to handle
5007 // mismatched types.
5011 // Record this extraction against the appropriate vector if possible...
5012 SDValue SourceVec = V.getOperand(0);
5013 // If the element number isn't a constant, we can't effectively
5014 // analyze what's going on.
5015 if (!isa<ConstantSDNode>(V.getOperand(1)))
5017 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5018 bool FoundSource = false;
5019 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5020 if (SourceVecs[j] == SourceVec) {
5021 if (MinElts[j] > EltNo)
5023 if (MaxElts[j] < EltNo)
5030 // Or record a new source if not...
5032 SourceVecs.push_back(SourceVec);
5033 MinElts.push_back(EltNo);
5034 MaxElts.push_back(EltNo);
5038 // Currently only do something sane when at most two source vectors
5040 if (SourceVecs.size() > 2)
5043 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5044 int VEXTOffsets[2] = {0, 0};
5046 // This loop extracts the usage patterns of the source vectors
5047 // and prepares appropriate SDValues for a shuffle if possible.
5048 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5049 if (SourceVecs[i].getValueType() == VT) {
5050 // No VEXT necessary
5051 ShuffleSrcs[i] = SourceVecs[i];
5054 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5055 // It probably isn't worth padding out a smaller vector just to
5056 // break it down again in a shuffle.
5060 // Since only 64-bit and 128-bit vectors are legal on ARM and
5061 // we've eliminated the other cases...
5062 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5063 "unexpected vector sizes in ReconstructShuffle");
5065 if (MaxElts[i] - MinElts[i] >= NumElts) {
5066 // Span too large for a VEXT to cope
5070 if (MinElts[i] >= NumElts) {
5071 // The extraction can just take the second half
5072 VEXTOffsets[i] = NumElts;
5073 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5075 DAG.getIntPtrConstant(NumElts));
5076 } else if (MaxElts[i] < NumElts) {
5077 // The extraction can just take the first half
5079 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5081 DAG.getIntPtrConstant(0));
5083 // An actual VEXT is needed
5084 VEXTOffsets[i] = MinElts[i];
5085 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5087 DAG.getIntPtrConstant(0));
5088 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5090 DAG.getIntPtrConstant(NumElts));
5091 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5092 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5096 SmallVector<int, 8> Mask;
5098 for (unsigned i = 0; i < NumElts; ++i) {
5099 SDValue Entry = Op.getOperand(i);
5100 if (Entry.getOpcode() == ISD::UNDEF) {
5105 SDValue ExtractVec = Entry.getOperand(0);
5106 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5107 .getOperand(1))->getSExtValue();
5108 if (ExtractVec == SourceVecs[0]) {
5109 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5111 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5115 // Final check before we try to produce nonsense...
5116 if (isShuffleMaskLegal(Mask, VT))
5117 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5123 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5124 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5125 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5126 /// are assumed to be legal.
5128 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5130 if (VT.getVectorNumElements() == 4 &&
5131 (VT.is128BitVector() || VT.is64BitVector())) {
5132 unsigned PFIndexes[4];
5133 for (unsigned i = 0; i != 4; ++i) {
5137 PFIndexes[i] = M[i];
5140 // Compute the index in the perfect shuffle table.
5141 unsigned PFTableIndex =
5142 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5143 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5144 unsigned Cost = (PFEntry >> 30);
5151 unsigned Imm, WhichResult;
5153 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5154 return (EltSize >= 32 ||
5155 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5156 isVREVMask(M, VT, 64) ||
5157 isVREVMask(M, VT, 32) ||
5158 isVREVMask(M, VT, 16) ||
5159 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5160 isVTBLMask(M, VT) ||
5161 isVTRNMask(M, VT, WhichResult) ||
5162 isVUZPMask(M, VT, WhichResult) ||
5163 isVZIPMask(M, VT, WhichResult) ||
5164 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5165 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5166 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5167 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5170 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5171 /// the specified operations to build the shuffle.
5172 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5173 SDValue RHS, SelectionDAG &DAG,
5175 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5176 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5177 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5180 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5189 OP_VUZPL, // VUZP, left result
5190 OP_VUZPR, // VUZP, right result
5191 OP_VZIPL, // VZIP, left result
5192 OP_VZIPR, // VZIP, right result
5193 OP_VTRNL, // VTRN, left result
5194 OP_VTRNR // VTRN, right result
5197 if (OpNum == OP_COPY) {
5198 if (LHSID == (1*9+2)*9+3) return LHS;
5199 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5203 SDValue OpLHS, OpRHS;
5204 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5205 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5206 EVT VT = OpLHS.getValueType();
5209 default: llvm_unreachable("Unknown shuffle opcode!");
5211 // VREV divides the vector in half and swaps within the half.
5212 if (VT.getVectorElementType() == MVT::i32 ||
5213 VT.getVectorElementType() == MVT::f32)
5214 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5215 // vrev <4 x i16> -> VREV32
5216 if (VT.getVectorElementType() == MVT::i16)
5217 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5218 // vrev <4 x i8> -> VREV16
5219 assert(VT.getVectorElementType() == MVT::i8);
5220 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5225 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5226 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5230 return DAG.getNode(ARMISD::VEXT, dl, VT,
5232 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5235 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5236 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5239 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5240 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5243 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5244 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5248 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5249 ArrayRef<int> ShuffleMask,
5250 SelectionDAG &DAG) {
5251 // Check to see if we can use the VTBL instruction.
5252 SDValue V1 = Op.getOperand(0);
5253 SDValue V2 = Op.getOperand(1);
5256 SmallVector<SDValue, 8> VTBLMask;
5257 for (ArrayRef<int>::iterator
5258 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5259 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5261 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5262 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5263 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5266 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5267 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5271 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5272 SelectionDAG &DAG) {
5274 SDValue OpLHS = Op.getOperand(0);
5275 EVT VT = OpLHS.getValueType();
5277 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5278 "Expect an v8i16/v16i8 type");
5279 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5280 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5281 // extract the first 8 bytes into the top double word and the last 8 bytes
5282 // into the bottom double word. The v8i16 case is similar.
5283 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5284 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5285 DAG.getConstant(ExtractNum, MVT::i32));
5288 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5289 SDValue V1 = Op.getOperand(0);
5290 SDValue V2 = Op.getOperand(1);
5292 EVT VT = Op.getValueType();
5293 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5295 // Convert shuffles that are directly supported on NEON to target-specific
5296 // DAG nodes, instead of keeping them as shuffles and matching them again
5297 // during code selection. This is more efficient and avoids the possibility
5298 // of inconsistencies between legalization and selection.
5299 // FIXME: floating-point vectors should be canonicalized to integer vectors
5300 // of the same time so that they get CSEd properly.
5301 ArrayRef<int> ShuffleMask = SVN->getMask();
5303 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5304 if (EltSize <= 32) {
5305 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5306 int Lane = SVN->getSplatIndex();
5307 // If this is undef splat, generate it via "just" vdup, if possible.
5308 if (Lane == -1) Lane = 0;
5310 // Test if V1 is a SCALAR_TO_VECTOR.
5311 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5312 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5314 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5315 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5317 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5318 !isa<ConstantSDNode>(V1.getOperand(0))) {
5319 bool IsScalarToVector = true;
5320 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5321 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5322 IsScalarToVector = false;
5325 if (IsScalarToVector)
5326 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5328 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5329 DAG.getConstant(Lane, MVT::i32));
5334 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5337 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5338 DAG.getConstant(Imm, MVT::i32));
5341 if (isVREVMask(ShuffleMask, VT, 64))
5342 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5343 if (isVREVMask(ShuffleMask, VT, 32))
5344 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5345 if (isVREVMask(ShuffleMask, VT, 16))
5346 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5348 if (V2->getOpcode() == ISD::UNDEF &&
5349 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5350 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5351 DAG.getConstant(Imm, MVT::i32));
5354 // Check for Neon shuffles that modify both input vectors in place.
5355 // If both results are used, i.e., if there are two shuffles with the same
5356 // source operands and with masks corresponding to both results of one of
5357 // these operations, DAG memoization will ensure that a single node is
5358 // used for both shuffles.
5359 unsigned WhichResult;
5360 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5361 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5362 V1, V2).getValue(WhichResult);
5363 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5364 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5365 V1, V2).getValue(WhichResult);
5366 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5367 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5368 V1, V2).getValue(WhichResult);
5370 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5371 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5372 V1, V1).getValue(WhichResult);
5373 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5374 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5375 V1, V1).getValue(WhichResult);
5376 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5377 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5378 V1, V1).getValue(WhichResult);
5381 // If the shuffle is not directly supported and it has 4 elements, use
5382 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5383 unsigned NumElts = VT.getVectorNumElements();
5385 unsigned PFIndexes[4];
5386 for (unsigned i = 0; i != 4; ++i) {
5387 if (ShuffleMask[i] < 0)
5390 PFIndexes[i] = ShuffleMask[i];
5393 // Compute the index in the perfect shuffle table.
5394 unsigned PFTableIndex =
5395 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5396 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5397 unsigned Cost = (PFEntry >> 30);
5400 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5403 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5404 if (EltSize >= 32) {
5405 // Do the expansion with floating-point types, since that is what the VFP
5406 // registers are defined to use, and since i64 is not legal.
5407 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5408 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5409 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5410 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5411 SmallVector<SDValue, 8> Ops;
5412 for (unsigned i = 0; i < NumElts; ++i) {
5413 if (ShuffleMask[i] < 0)
5414 Ops.push_back(DAG.getUNDEF(EltVT));
5416 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5417 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5418 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5421 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5422 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5425 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5426 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5428 if (VT == MVT::v8i8) {
5429 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5430 if (NewOp.getNode())
5437 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5438 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5439 SDValue Lane = Op.getOperand(2);
5440 if (!isa<ConstantSDNode>(Lane))
5446 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5447 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5448 SDValue Lane = Op.getOperand(1);
5449 if (!isa<ConstantSDNode>(Lane))
5452 SDValue Vec = Op.getOperand(0);
5453 if (Op.getValueType() == MVT::i32 &&
5454 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5456 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5462 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5463 // The only time a CONCAT_VECTORS operation can have legal types is when
5464 // two 64-bit vectors are concatenated to a 128-bit vector.
5465 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5466 "unexpected CONCAT_VECTORS");
5468 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5469 SDValue Op0 = Op.getOperand(0);
5470 SDValue Op1 = Op.getOperand(1);
5471 if (Op0.getOpcode() != ISD::UNDEF)
5472 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5473 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5474 DAG.getIntPtrConstant(0));
5475 if (Op1.getOpcode() != ISD::UNDEF)
5476 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5477 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5478 DAG.getIntPtrConstant(1));
5479 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5482 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5483 /// element has been zero/sign-extended, depending on the isSigned parameter,
5484 /// from an integer type half its size.
5485 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5487 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5488 EVT VT = N->getValueType(0);
5489 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5490 SDNode *BVN = N->getOperand(0).getNode();
5491 if (BVN->getValueType(0) != MVT::v4i32 ||
5492 BVN->getOpcode() != ISD::BUILD_VECTOR)
5494 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5495 unsigned HiElt = 1 - LoElt;
5496 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5497 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5498 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5499 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5500 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5503 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5504 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5507 if (Hi0->isNullValue() && Hi1->isNullValue())
5513 if (N->getOpcode() != ISD::BUILD_VECTOR)
5516 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5517 SDNode *Elt = N->getOperand(i).getNode();
5518 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5519 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5520 unsigned HalfSize = EltSize / 2;
5522 if (!isIntN(HalfSize, C->getSExtValue()))
5525 if (!isUIntN(HalfSize, C->getZExtValue()))
5536 /// isSignExtended - Check if a node is a vector value that is sign-extended
5537 /// or a constant BUILD_VECTOR with sign-extended elements.
5538 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5539 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5541 if (isExtendedBUILD_VECTOR(N, DAG, true))
5546 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5547 /// or a constant BUILD_VECTOR with zero-extended elements.
5548 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5549 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5551 if (isExtendedBUILD_VECTOR(N, DAG, false))
5556 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5557 if (OrigVT.getSizeInBits() >= 64)
5560 assert(OrigVT.isSimple() && "Expecting a simple value type");
5562 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5563 switch (OrigSimpleTy) {
5564 default: llvm_unreachable("Unexpected Vector Type");
5573 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5574 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5575 /// We insert the required extension here to get the vector to fill a D register.
5576 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5579 unsigned ExtOpcode) {
5580 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5581 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5582 // 64-bits we need to insert a new extension so that it will be 64-bits.
5583 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5584 if (OrigTy.getSizeInBits() >= 64)
5587 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5588 EVT NewVT = getExtensionTo64Bits(OrigTy);
5590 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5593 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5594 /// does not do any sign/zero extension. If the original vector is less
5595 /// than 64 bits, an appropriate extension will be added after the load to
5596 /// reach a total size of 64 bits. We have to add the extension separately
5597 /// because ARM does not have a sign/zero extending load for vectors.
5598 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5599 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5601 // The load already has the right type.
5602 if (ExtendedTy == LD->getMemoryVT())
5603 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5604 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5605 LD->isNonTemporal(), LD->isInvariant(),
5606 LD->getAlignment());
5608 // We need to create a zextload/sextload. We cannot just create a load
5609 // followed by a zext/zext node because LowerMUL is also run during normal
5610 // operation legalization where we can't create illegal types.
5611 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5612 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5613 LD->getMemoryVT(), LD->isVolatile(),
5614 LD->isNonTemporal(), LD->getAlignment());
5617 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5618 /// extending load, or BUILD_VECTOR with extended elements, return the
5619 /// unextended value. The unextended vector should be 64 bits so that it can
5620 /// be used as an operand to a VMULL instruction. If the original vector size
5621 /// before extension is less than 64 bits we add a an extension to resize
5622 /// the vector to 64 bits.
5623 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5624 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5625 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5626 N->getOperand(0)->getValueType(0),
5630 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5631 return SkipLoadExtensionForVMULL(LD, DAG);
5633 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5634 // have been legalized as a BITCAST from v4i32.
5635 if (N->getOpcode() == ISD::BITCAST) {
5636 SDNode *BVN = N->getOperand(0).getNode();
5637 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5638 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5639 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5640 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5641 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5643 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5644 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5645 EVT VT = N->getValueType(0);
5646 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5647 unsigned NumElts = VT.getVectorNumElements();
5648 MVT TruncVT = MVT::getIntegerVT(EltSize);
5649 SmallVector<SDValue, 8> Ops;
5650 for (unsigned i = 0; i != NumElts; ++i) {
5651 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5652 const APInt &CInt = C->getAPIntValue();
5653 // Element types smaller than 32 bits are not legal, so use i32 elements.
5654 // The values are implicitly truncated so sext vs. zext doesn't matter.
5655 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5657 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5658 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5661 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5662 unsigned Opcode = N->getOpcode();
5663 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5664 SDNode *N0 = N->getOperand(0).getNode();
5665 SDNode *N1 = N->getOperand(1).getNode();
5666 return N0->hasOneUse() && N1->hasOneUse() &&
5667 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5672 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5673 unsigned Opcode = N->getOpcode();
5674 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5675 SDNode *N0 = N->getOperand(0).getNode();
5676 SDNode *N1 = N->getOperand(1).getNode();
5677 return N0->hasOneUse() && N1->hasOneUse() &&
5678 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5683 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5684 // Multiplications are only custom-lowered for 128-bit vectors so that
5685 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5686 EVT VT = Op.getValueType();
5687 assert(VT.is128BitVector() && VT.isInteger() &&
5688 "unexpected type for custom-lowering ISD::MUL");
5689 SDNode *N0 = Op.getOperand(0).getNode();
5690 SDNode *N1 = Op.getOperand(1).getNode();
5691 unsigned NewOpc = 0;
5693 bool isN0SExt = isSignExtended(N0, DAG);
5694 bool isN1SExt = isSignExtended(N1, DAG);
5695 if (isN0SExt && isN1SExt)
5696 NewOpc = ARMISD::VMULLs;
5698 bool isN0ZExt = isZeroExtended(N0, DAG);
5699 bool isN1ZExt = isZeroExtended(N1, DAG);
5700 if (isN0ZExt && isN1ZExt)
5701 NewOpc = ARMISD::VMULLu;
5702 else if (isN1SExt || isN1ZExt) {
5703 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5704 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5705 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5706 NewOpc = ARMISD::VMULLs;
5708 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5709 NewOpc = ARMISD::VMULLu;
5711 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5713 NewOpc = ARMISD::VMULLu;
5719 if (VT == MVT::v2i64)
5720 // Fall through to expand this. It is not legal.
5723 // Other vector multiplications are legal.
5728 // Legalize to a VMULL instruction.
5731 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5733 Op0 = SkipExtensionForVMULL(N0, DAG);
5734 assert(Op0.getValueType().is64BitVector() &&
5735 Op1.getValueType().is64BitVector() &&
5736 "unexpected types for extended operands to VMULL");
5737 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5740 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5741 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5748 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5749 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5750 EVT Op1VT = Op1.getValueType();
5751 return DAG.getNode(N0->getOpcode(), DL, VT,
5752 DAG.getNode(NewOpc, DL, VT,
5753 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5754 DAG.getNode(NewOpc, DL, VT,
5755 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5759 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5761 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5762 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5763 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5764 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5765 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5766 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5767 // Get reciprocal estimate.
5768 // float4 recip = vrecpeq_f32(yf);
5769 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5770 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5771 // Because char has a smaller range than uchar, we can actually get away
5772 // without any newton steps. This requires that we use a weird bias
5773 // of 0xb000, however (again, this has been exhaustively tested).
5774 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5775 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5776 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5777 Y = DAG.getConstant(0xb000, MVT::i32);
5778 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5779 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5780 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5781 // Convert back to short.
5782 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5783 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5788 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5790 // Convert to float.
5791 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5792 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5793 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5794 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5795 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5796 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5798 // Use reciprocal estimate and one refinement step.
5799 // float4 recip = vrecpeq_f32(yf);
5800 // recip *= vrecpsq_f32(yf, recip);
5801 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5802 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5803 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5804 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5806 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5807 // Because short has a smaller range than ushort, we can actually get away
5808 // with only a single newton step. This requires that we use a weird bias
5809 // of 89, however (again, this has been exhaustively tested).
5810 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5811 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5812 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5813 N1 = DAG.getConstant(0x89, MVT::i32);
5814 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5815 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5816 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5817 // Convert back to integer and return.
5818 // return vmovn_s32(vcvt_s32_f32(result));
5819 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5820 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5824 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5825 EVT VT = Op.getValueType();
5826 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5827 "unexpected type for custom-lowering ISD::SDIV");
5830 SDValue N0 = Op.getOperand(0);
5831 SDValue N1 = Op.getOperand(1);
5834 if (VT == MVT::v8i8) {
5835 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5836 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5838 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5839 DAG.getIntPtrConstant(4));
5840 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5841 DAG.getIntPtrConstant(4));
5842 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5843 DAG.getIntPtrConstant(0));
5844 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5845 DAG.getIntPtrConstant(0));
5847 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5848 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5850 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5851 N0 = LowerCONCAT_VECTORS(N0, DAG);
5853 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5856 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5859 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5860 EVT VT = Op.getValueType();
5861 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5862 "unexpected type for custom-lowering ISD::UDIV");
5865 SDValue N0 = Op.getOperand(0);
5866 SDValue N1 = Op.getOperand(1);
5869 if (VT == MVT::v8i8) {
5870 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5871 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5873 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5874 DAG.getIntPtrConstant(4));
5875 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5876 DAG.getIntPtrConstant(4));
5877 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5878 DAG.getIntPtrConstant(0));
5879 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5880 DAG.getIntPtrConstant(0));
5882 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5883 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5885 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5886 N0 = LowerCONCAT_VECTORS(N0, DAG);
5888 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5889 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5894 // v4i16 sdiv ... Convert to float.
5895 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5896 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5897 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5898 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5899 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5900 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5902 // Use reciprocal estimate and two refinement steps.
5903 // float4 recip = vrecpeq_f32(yf);
5904 // recip *= vrecpsq_f32(yf, recip);
5905 // recip *= vrecpsq_f32(yf, recip);
5906 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5907 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5908 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5909 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5911 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5912 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5913 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5915 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5916 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5917 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5918 // and that it will never cause us to return an answer too large).
5919 // float4 result = as_float4(as_int4(xf*recip) + 2);
5920 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5921 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5922 N1 = DAG.getConstant(2, MVT::i32);
5923 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5924 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5925 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5926 // Convert back to integer and return.
5927 // return vmovn_u32(vcvt_s32_f32(result));
5928 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5929 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5933 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5934 EVT VT = Op.getNode()->getValueType(0);
5935 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5938 bool ExtraOp = false;
5939 switch (Op.getOpcode()) {
5940 default: llvm_unreachable("Invalid code");
5941 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5942 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5943 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5944 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5948 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5950 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5951 Op.getOperand(1), Op.getOperand(2));
5954 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
5955 assert(Subtarget->isTargetDarwin());
5957 // For iOS, we want to call an alternative entry point: __sincos_stret,
5958 // return values are passed via sret.
5960 SDValue Arg = Op.getOperand(0);
5961 EVT ArgVT = Arg.getValueType();
5962 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
5964 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5965 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5967 // Pair of floats / doubles used to pass the result.
5968 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
5970 // Create stack object for sret.
5971 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
5972 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
5973 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
5974 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
5980 Entry.Ty = RetTy->getPointerTo();
5981 Entry.isSExt = false;
5982 Entry.isZExt = false;
5983 Entry.isSRet = true;
5984 Args.push_back(Entry);
5988 Entry.isSExt = false;
5989 Entry.isZExt = false;
5990 Args.push_back(Entry);
5992 const char *LibcallName = (ArgVT == MVT::f64)
5993 ? "__sincos_stret" : "__sincosf_stret";
5994 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
5997 CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
5998 false, false, false, false, 0,
5999 CallingConv::C, /*isTaillCall=*/false,
6000 /*doesNotRet=*/false, /*isReturnValueUsed*/false,
6001 Callee, Args, DAG, dl);
6002 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6004 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6005 MachinePointerInfo(), false, false, false, 0);
6007 // Address of cos field.
6008 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6009 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6010 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6011 MachinePointerInfo(), false, false, false, 0);
6013 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6014 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6015 LoadSin.getValue(0), LoadCos.getValue(0));
6018 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6019 // Monotonic load/store is legal for all targets
6020 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6023 // Acquire/Release load/store is not legal for targets without a
6024 // dmb or equivalent available.
6029 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
6030 SelectionDAG &DAG) {
6032 assert (Node->getValueType(0) == MVT::i64 &&
6033 "Only know how to expand i64 atomics");
6034 AtomicSDNode *AN = cast<AtomicSDNode>(Node);
6036 SmallVector<SDValue, 6> Ops;
6037 Ops.push_back(Node->getOperand(0)); // Chain
6038 Ops.push_back(Node->getOperand(1)); // Ptr
6039 for(unsigned i=2; i<Node->getNumOperands(); i++) {
6041 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6042 Node->getOperand(i), DAG.getIntPtrConstant(0)));
6044 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6045 Node->getOperand(i), DAG.getIntPtrConstant(1)));
6047 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6048 SDValue Result = DAG.getAtomic(
6049 Node->getOpcode(), dl, MVT::i64, Tys, Ops.data(), Ops.size(),
6050 cast<MemSDNode>(Node)->getMemOperand(), AN->getSuccessOrdering(),
6051 AN->getFailureOrdering(), AN->getSynchScope());
6052 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
6053 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6054 Results.push_back(Result.getValue(2));
6057 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6058 SmallVectorImpl<SDValue> &Results,
6060 const ARMSubtarget *Subtarget) {
6062 SDValue Cycles32, OutChain;
6064 if (Subtarget->hasPerfMon()) {
6065 // Under Power Management extensions, the cycle-count is:
6066 // mrc p15, #0, <Rt>, c9, c13, #0
6067 SDValue Ops[] = { N->getOperand(0), // Chain
6068 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6069 DAG.getConstant(15, MVT::i32),
6070 DAG.getConstant(0, MVT::i32),
6071 DAG.getConstant(9, MVT::i32),
6072 DAG.getConstant(13, MVT::i32),
6073 DAG.getConstant(0, MVT::i32)
6076 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6077 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6078 array_lengthof(Ops));
6079 OutChain = Cycles32.getValue(1);
6081 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6082 // there are older ARM CPUs that have implementation-specific ways of
6083 // obtaining this information (FIXME!).
6084 Cycles32 = DAG.getConstant(0, MVT::i32);
6085 OutChain = DAG.getEntryNode();
6089 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6090 Cycles32, DAG.getConstant(0, MVT::i32));
6091 Results.push_back(Cycles64);
6092 Results.push_back(OutChain);
6095 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6096 switch (Op.getOpcode()) {
6097 default: llvm_unreachable("Don't know how to custom lower this!");
6098 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6099 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6100 case ISD::GlobalAddress:
6101 return Subtarget->isTargetMachO() ? LowerGlobalAddressDarwin(Op, DAG) :
6102 LowerGlobalAddressELF(Op, DAG);
6103 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6104 case ISD::SELECT: return LowerSELECT(Op, DAG);
6105 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6106 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6107 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6108 case ISD::VASTART: return LowerVASTART(Op, DAG);
6109 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6110 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6111 case ISD::SINT_TO_FP:
6112 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6113 case ISD::FP_TO_SINT:
6114 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6115 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6116 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6117 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6118 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6119 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6120 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6121 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6123 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6126 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6127 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6128 case ISD::SRL_PARTS:
6129 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6130 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6131 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6132 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6133 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6134 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6135 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6136 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6137 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6138 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6139 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6140 case ISD::MUL: return LowerMUL(Op, DAG);
6141 case ISD::SDIV: return LowerSDIV(Op, DAG);
6142 case ISD::UDIV: return LowerUDIV(Op, DAG);
6146 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6147 case ISD::ATOMIC_LOAD:
6148 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6149 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6151 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6155 /// ReplaceNodeResults - Replace the results of node with an illegal result
6156 /// type with new values built out of custom code.
6157 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6158 SmallVectorImpl<SDValue>&Results,
6159 SelectionDAG &DAG) const {
6161 switch (N->getOpcode()) {
6163 llvm_unreachable("Don't know how to custom expand this!");
6165 Res = ExpandBITCAST(N, DAG);
6169 Res = Expand64BitShift(N, DAG, Subtarget);
6171 case ISD::READCYCLECOUNTER:
6172 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6174 case ISD::ATOMIC_STORE:
6175 case ISD::ATOMIC_LOAD:
6176 case ISD::ATOMIC_LOAD_ADD:
6177 case ISD::ATOMIC_LOAD_AND:
6178 case ISD::ATOMIC_LOAD_NAND:
6179 case ISD::ATOMIC_LOAD_OR:
6180 case ISD::ATOMIC_LOAD_SUB:
6181 case ISD::ATOMIC_LOAD_XOR:
6182 case ISD::ATOMIC_SWAP:
6183 case ISD::ATOMIC_CMP_SWAP:
6184 case ISD::ATOMIC_LOAD_MIN:
6185 case ISD::ATOMIC_LOAD_UMIN:
6186 case ISD::ATOMIC_LOAD_MAX:
6187 case ISD::ATOMIC_LOAD_UMAX:
6188 ReplaceATOMIC_OP_64(N, Results, DAG);
6192 Results.push_back(Res);
6195 //===----------------------------------------------------------------------===//
6196 // ARM Scheduler Hooks
6197 //===----------------------------------------------------------------------===//
6200 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6201 MachineBasicBlock *BB,
6202 unsigned Size) const {
6203 unsigned dest = MI->getOperand(0).getReg();
6204 unsigned ptr = MI->getOperand(1).getReg();
6205 unsigned oldval = MI->getOperand(2).getReg();
6206 unsigned newval = MI->getOperand(3).getReg();
6207 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6208 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
6209 DebugLoc dl = MI->getDebugLoc();
6210 bool isThumb2 = Subtarget->isThumb2();
6212 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6213 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6214 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6215 (const TargetRegisterClass*)&ARM::GPRRegClass);
6218 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6219 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6220 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6223 unsigned ldrOpc, strOpc;
6224 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6226 MachineFunction *MF = BB->getParent();
6227 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6228 MachineFunction::iterator It = BB;
6229 ++It; // insert the new blocks after the current block
6231 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6232 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6233 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6234 MF->insert(It, loop1MBB);
6235 MF->insert(It, loop2MBB);
6236 MF->insert(It, exitMBB);
6238 // Transfer the remainder of BB and its successor edges to exitMBB.
6239 exitMBB->splice(exitMBB->begin(), BB,
6240 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6241 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6245 // fallthrough --> loop1MBB
6246 BB->addSuccessor(loop1MBB);
6249 // ldrex dest, [ptr]
6253 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6254 if (ldrOpc == ARM::t2LDREX)
6256 AddDefaultPred(MIB);
6257 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6258 .addReg(dest).addReg(oldval));
6259 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6260 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6261 BB->addSuccessor(loop2MBB);
6262 BB->addSuccessor(exitMBB);
6265 // strex scratch, newval, [ptr]
6269 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6270 if (strOpc == ARM::t2STREX)
6272 AddDefaultPred(MIB);
6273 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6274 .addReg(scratch).addImm(0));
6275 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6276 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6277 BB->addSuccessor(loop1MBB);
6278 BB->addSuccessor(exitMBB);
6284 MI->eraseFromParent(); // The instruction is gone now.
6290 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6291 unsigned Size, unsigned BinOpcode) const {
6292 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6293 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6295 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6296 MachineFunction *MF = BB->getParent();
6297 MachineFunction::iterator It = BB;
6300 unsigned dest = MI->getOperand(0).getReg();
6301 unsigned ptr = MI->getOperand(1).getReg();
6302 unsigned incr = MI->getOperand(2).getReg();
6303 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6304 DebugLoc dl = MI->getDebugLoc();
6305 bool isThumb2 = Subtarget->isThumb2();
6307 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6309 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6310 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6311 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6314 unsigned ldrOpc, strOpc;
6315 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6317 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6318 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6319 MF->insert(It, loopMBB);
6320 MF->insert(It, exitMBB);
6322 // Transfer the remainder of BB and its successor edges to exitMBB.
6323 exitMBB->splice(exitMBB->begin(), BB,
6324 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6325 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6327 const TargetRegisterClass *TRC = isThumb2 ?
6328 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6329 (const TargetRegisterClass*)&ARM::GPRRegClass;
6330 unsigned scratch = MRI.createVirtualRegister(TRC);
6331 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6335 // fallthrough --> loopMBB
6336 BB->addSuccessor(loopMBB);
6340 // <binop> scratch2, dest, incr
6341 // strex scratch, scratch2, ptr
6344 // fallthrough --> exitMBB
6346 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6347 if (ldrOpc == ARM::t2LDREX)
6349 AddDefaultPred(MIB);
6351 // operand order needs to go the other way for NAND
6352 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6353 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6354 addReg(incr).addReg(dest)).addReg(0);
6356 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6357 addReg(dest).addReg(incr)).addReg(0);
6360 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6361 if (strOpc == ARM::t2STREX)
6363 AddDefaultPred(MIB);
6364 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6365 .addReg(scratch).addImm(0));
6366 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6367 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6369 BB->addSuccessor(loopMBB);
6370 BB->addSuccessor(exitMBB);
6376 MI->eraseFromParent(); // The instruction is gone now.
6382 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6383 MachineBasicBlock *BB,
6386 ARMCC::CondCodes Cond) const {
6387 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6389 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6390 MachineFunction *MF = BB->getParent();
6391 MachineFunction::iterator It = BB;
6394 unsigned dest = MI->getOperand(0).getReg();
6395 unsigned ptr = MI->getOperand(1).getReg();
6396 unsigned incr = MI->getOperand(2).getReg();
6397 unsigned oldval = dest;
6398 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6399 DebugLoc dl = MI->getDebugLoc();
6400 bool isThumb2 = Subtarget->isThumb2();
6402 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6404 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6405 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6406 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6409 unsigned ldrOpc, strOpc, extendOpc;
6410 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6412 default: llvm_unreachable("unsupported size for AtomicBinaryMinMax!");
6414 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6417 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6424 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6425 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6426 MF->insert(It, loopMBB);
6427 MF->insert(It, exitMBB);
6429 // Transfer the remainder of BB and its successor edges to exitMBB.
6430 exitMBB->splice(exitMBB->begin(), BB,
6431 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6432 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6434 const TargetRegisterClass *TRC = isThumb2 ?
6435 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6436 (const TargetRegisterClass*)&ARM::GPRRegClass;
6437 unsigned scratch = MRI.createVirtualRegister(TRC);
6438 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6442 // fallthrough --> loopMBB
6443 BB->addSuccessor(loopMBB);
6447 // (sign extend dest, if required)
6449 // cmov.cond scratch2, incr, dest
6450 // strex scratch, scratch2, ptr
6453 // fallthrough --> exitMBB
6455 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6456 if (ldrOpc == ARM::t2LDREX)
6458 AddDefaultPred(MIB);
6460 // Sign extend the value, if necessary.
6461 if (signExtend && extendOpc) {
6462 oldval = MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass
6463 : &ARM::GPRnopcRegClass);
6465 MRI.constrainRegClass(dest, &ARM::GPRnopcRegClass);
6466 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6471 // Build compare and cmov instructions.
6472 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6473 .addReg(oldval).addReg(incr));
6474 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6475 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6477 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6478 if (strOpc == ARM::t2STREX)
6480 AddDefaultPred(MIB);
6481 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6482 .addReg(scratch).addImm(0));
6483 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6484 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6486 BB->addSuccessor(loopMBB);
6487 BB->addSuccessor(exitMBB);
6493 MI->eraseFromParent(); // The instruction is gone now.
6499 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6500 unsigned Op1, unsigned Op2,
6501 bool NeedsCarry, bool IsCmpxchg,
6502 bool IsMinMax, ARMCC::CondCodes CC) const {
6503 // This also handles ATOMIC_SWAP and ATOMIC_STORE, indicated by Op1==0.
6504 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6506 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6507 MachineFunction *MF = BB->getParent();
6508 MachineFunction::iterator It = BB;
6511 unsigned destlo = MI->getOperand(0).getReg();
6512 unsigned desthi = MI->getOperand(1).getReg();
6513 unsigned ptr = MI->getOperand(2).getReg();
6514 unsigned vallo = MI->getOperand(3).getReg();
6515 unsigned valhi = MI->getOperand(4).getReg();
6516 AtomicOrdering Ord =
6517 static_cast<AtomicOrdering>(MI->getOperand(IsCmpxchg ? 7 : 5).getImm());
6518 DebugLoc dl = MI->getDebugLoc();
6519 bool isThumb2 = Subtarget->isThumb2();
6521 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6523 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6524 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6525 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6526 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6527 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6530 unsigned ldrOpc, strOpc;
6531 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6533 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6534 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6535 if (IsCmpxchg || IsMinMax)
6536 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6538 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6539 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6541 MF->insert(It, loopMBB);
6542 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6543 if (IsCmpxchg) MF->insert(It, cont2BB);
6544 MF->insert(It, exitMBB);
6546 // Transfer the remainder of BB and its successor edges to exitMBB.
6547 exitMBB->splice(exitMBB->begin(), BB,
6548 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6549 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6551 const TargetRegisterClass *TRC = isThumb2 ?
6552 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6553 (const TargetRegisterClass*)&ARM::GPRRegClass;
6554 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6558 // fallthrough --> loopMBB
6559 BB->addSuccessor(loopMBB);
6562 // ldrexd r2, r3, ptr
6563 // <binopa> r0, r2, incr
6564 // <binopb> r1, r3, incr
6565 // strexd storesuccess, r0, r1, ptr
6566 // cmp storesuccess, #0
6568 // fallthrough --> exitMBB
6573 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6574 .addReg(destlo, RegState::Define)
6575 .addReg(desthi, RegState::Define)
6578 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6579 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6580 .addReg(GPRPair0, RegState::Define)
6582 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6583 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6584 .addReg(GPRPair0, 0, ARM::gsub_0);
6585 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6586 .addReg(GPRPair0, 0, ARM::gsub_1);
6589 unsigned StoreLo, StoreHi;
6592 for (unsigned i = 0; i < 2; i++) {
6593 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6595 .addReg(i == 0 ? destlo : desthi)
6596 .addReg(i == 0 ? vallo : valhi));
6597 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6598 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6599 BB->addSuccessor(exitMBB);
6600 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6601 BB = (i == 0 ? contBB : cont2BB);
6604 // Copy to physregs for strexd
6605 StoreLo = MI->getOperand(5).getReg();
6606 StoreHi = MI->getOperand(6).getReg();
6608 // Perform binary operation
6609 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6610 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6611 .addReg(destlo).addReg(vallo))
6612 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6613 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6614 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6615 .addReg(desthi).addReg(valhi))
6616 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6621 // Copy to physregs for strexd
6626 // Compare and branch to exit block.
6627 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6628 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6629 BB->addSuccessor(exitMBB);
6630 BB->addSuccessor(contBB);
6638 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6639 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6640 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6641 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6643 // Marshal a pair...
6644 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6645 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6646 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6647 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6648 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6651 .addImm(ARM::gsub_0);
6652 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6655 .addImm(ARM::gsub_1);
6658 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6659 .addReg(StorePair).addReg(ptr));
6662 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6663 .addReg(storesuccess).addImm(0));
6664 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6665 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6667 BB->addSuccessor(loopMBB);
6668 BB->addSuccessor(exitMBB);
6674 MI->eraseFromParent(); // The instruction is gone now.
6680 ARMTargetLowering::EmitAtomicLoad64(MachineInstr *MI, MachineBasicBlock *BB) const {
6682 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6684 unsigned destlo = MI->getOperand(0).getReg();
6685 unsigned desthi = MI->getOperand(1).getReg();
6686 unsigned ptr = MI->getOperand(2).getReg();
6687 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6688 DebugLoc dl = MI->getDebugLoc();
6689 bool isThumb2 = Subtarget->isThumb2();
6691 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6693 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6694 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6695 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6697 unsigned ldrOpc, strOpc;
6698 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6700 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(ldrOpc));
6703 MIB.addReg(destlo, RegState::Define)
6704 .addReg(desthi, RegState::Define)
6708 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6709 MIB.addReg(GPRPair0, RegState::Define).addReg(ptr);
6711 // Copy GPRPair0 into dest. (This copy will normally be coalesced.)
6712 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), destlo)
6713 .addReg(GPRPair0, 0, ARM::gsub_0);
6714 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), desthi)
6715 .addReg(GPRPair0, 0, ARM::gsub_1);
6717 AddDefaultPred(MIB);
6719 MI->eraseFromParent(); // The instruction is gone now.
6724 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6725 /// registers the function context.
6726 void ARMTargetLowering::
6727 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6728 MachineBasicBlock *DispatchBB, int FI) const {
6729 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6730 DebugLoc dl = MI->getDebugLoc();
6731 MachineFunction *MF = MBB->getParent();
6732 MachineRegisterInfo *MRI = &MF->getRegInfo();
6733 MachineConstantPool *MCP = MF->getConstantPool();
6734 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6735 const Function *F = MF->getFunction();
6737 bool isThumb = Subtarget->isThumb();
6738 bool isThumb2 = Subtarget->isThumb2();
6740 unsigned PCLabelId = AFI->createPICLabelUId();
6741 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6742 ARMConstantPoolValue *CPV =
6743 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6744 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6746 const TargetRegisterClass *TRC = isThumb ?
6747 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6748 (const TargetRegisterClass*)&ARM::GPRRegClass;
6750 // Grab constant pool and fixed stack memory operands.
6751 MachineMemOperand *CPMMO =
6752 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6753 MachineMemOperand::MOLoad, 4, 4);
6755 MachineMemOperand *FIMMOSt =
6756 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6757 MachineMemOperand::MOStore, 4, 4);
6759 // Load the address of the dispatch MBB into the jump buffer.
6761 // Incoming value: jbuf
6762 // ldr.n r5, LCPI1_1
6765 // str r5, [$jbuf, #+4] ; &jbuf[1]
6766 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6767 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6768 .addConstantPoolIndex(CPI)
6769 .addMemOperand(CPMMO));
6770 // Set the low bit because of thumb mode.
6771 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6773 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6774 .addReg(NewVReg1, RegState::Kill)
6776 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6777 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6778 .addReg(NewVReg2, RegState::Kill)
6780 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6781 .addReg(NewVReg3, RegState::Kill)
6783 .addImm(36) // &jbuf[1] :: pc
6784 .addMemOperand(FIMMOSt));
6785 } else if (isThumb) {
6786 // Incoming value: jbuf
6787 // ldr.n r1, LCPI1_4
6791 // add r2, $jbuf, #+4 ; &jbuf[1]
6793 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6794 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6795 .addConstantPoolIndex(CPI)
6796 .addMemOperand(CPMMO));
6797 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6798 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6799 .addReg(NewVReg1, RegState::Kill)
6801 // Set the low bit because of thumb mode.
6802 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6803 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6804 .addReg(ARM::CPSR, RegState::Define)
6806 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6807 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6808 .addReg(ARM::CPSR, RegState::Define)
6809 .addReg(NewVReg2, RegState::Kill)
6810 .addReg(NewVReg3, RegState::Kill));
6811 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6812 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6814 .addImm(36)); // &jbuf[1] :: pc
6815 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6816 .addReg(NewVReg4, RegState::Kill)
6817 .addReg(NewVReg5, RegState::Kill)
6819 .addMemOperand(FIMMOSt));
6821 // Incoming value: jbuf
6824 // str r1, [$jbuf, #+4] ; &jbuf[1]
6825 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6826 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6827 .addConstantPoolIndex(CPI)
6829 .addMemOperand(CPMMO));
6830 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6831 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6832 .addReg(NewVReg1, RegState::Kill)
6833 .addImm(PCLabelId));
6834 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6835 .addReg(NewVReg2, RegState::Kill)
6837 .addImm(36) // &jbuf[1] :: pc
6838 .addMemOperand(FIMMOSt));
6842 MachineBasicBlock *ARMTargetLowering::
6843 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6844 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6845 DebugLoc dl = MI->getDebugLoc();
6846 MachineFunction *MF = MBB->getParent();
6847 MachineRegisterInfo *MRI = &MF->getRegInfo();
6848 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6849 MachineFrameInfo *MFI = MF->getFrameInfo();
6850 int FI = MFI->getFunctionContextIndex();
6852 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6853 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6854 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6856 // Get a mapping of the call site numbers to all of the landing pads they're
6858 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6859 unsigned MaxCSNum = 0;
6860 MachineModuleInfo &MMI = MF->getMMI();
6861 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6863 if (!BB->isLandingPad()) continue;
6865 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6867 for (MachineBasicBlock::iterator
6868 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6869 if (!II->isEHLabel()) continue;
6871 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6872 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6874 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6875 for (SmallVectorImpl<unsigned>::iterator
6876 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6877 CSI != CSE; ++CSI) {
6878 CallSiteNumToLPad[*CSI].push_back(BB);
6879 MaxCSNum = std::max(MaxCSNum, *CSI);
6885 // Get an ordered list of the machine basic blocks for the jump table.
6886 std::vector<MachineBasicBlock*> LPadList;
6887 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6888 LPadList.reserve(CallSiteNumToLPad.size());
6889 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6890 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6891 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6892 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6893 LPadList.push_back(*II);
6894 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6898 assert(!LPadList.empty() &&
6899 "No landing pad destinations for the dispatch jump table!");
6901 // Create the jump table and associated information.
6902 MachineJumpTableInfo *JTI =
6903 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6904 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6905 unsigned UId = AFI->createJumpTableUId();
6906 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6908 // Create the MBBs for the dispatch code.
6910 // Shove the dispatch's address into the return slot in the function context.
6911 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6912 DispatchBB->setIsLandingPad();
6914 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6915 unsigned trap_opcode;
6916 if (Subtarget->isThumb())
6917 trap_opcode = ARM::tTRAP;
6919 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6921 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6922 DispatchBB->addSuccessor(TrapBB);
6924 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6925 DispatchBB->addSuccessor(DispContBB);
6928 MF->insert(MF->end(), DispatchBB);
6929 MF->insert(MF->end(), DispContBB);
6930 MF->insert(MF->end(), TrapBB);
6932 // Insert code into the entry block that creates and registers the function
6934 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6936 MachineMemOperand *FIMMOLd =
6937 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6938 MachineMemOperand::MOLoad |
6939 MachineMemOperand::MOVolatile, 4, 4);
6941 MachineInstrBuilder MIB;
6942 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6944 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6945 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6947 // Add a register mask with no preserved registers. This results in all
6948 // registers being marked as clobbered.
6949 MIB.addRegMask(RI.getNoPreservedMask());
6951 unsigned NumLPads = LPadList.size();
6952 if (Subtarget->isThumb2()) {
6953 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6954 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6957 .addMemOperand(FIMMOLd));
6959 if (NumLPads < 256) {
6960 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6962 .addImm(LPadList.size()));
6964 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6965 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6966 .addImm(NumLPads & 0xFFFF));
6968 unsigned VReg2 = VReg1;
6969 if ((NumLPads & 0xFFFF0000) != 0) {
6970 VReg2 = MRI->createVirtualRegister(TRC);
6971 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6973 .addImm(NumLPads >> 16));
6976 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6981 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6986 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6987 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6988 .addJumpTableIndex(MJTI)
6991 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6994 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6995 .addReg(NewVReg3, RegState::Kill)
6997 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6999 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7000 .addReg(NewVReg4, RegState::Kill)
7002 .addJumpTableIndex(MJTI)
7004 } else if (Subtarget->isThumb()) {
7005 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7006 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7009 .addMemOperand(FIMMOLd));
7011 if (NumLPads < 256) {
7012 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7016 MachineConstantPool *ConstantPool = MF->getConstantPool();
7017 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7018 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7020 // MachineConstantPool wants an explicit alignment.
7021 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7023 Align = getDataLayout()->getTypeAllocSize(C->getType());
7024 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7026 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7027 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7028 .addReg(VReg1, RegState::Define)
7029 .addConstantPoolIndex(Idx));
7030 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7035 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7040 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7041 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7042 .addReg(ARM::CPSR, RegState::Define)
7046 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7047 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7048 .addJumpTableIndex(MJTI)
7051 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7052 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7053 .addReg(ARM::CPSR, RegState::Define)
7054 .addReg(NewVReg2, RegState::Kill)
7057 MachineMemOperand *JTMMOLd =
7058 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7059 MachineMemOperand::MOLoad, 4, 4);
7061 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7062 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7063 .addReg(NewVReg4, RegState::Kill)
7065 .addMemOperand(JTMMOLd));
7067 unsigned NewVReg6 = NewVReg5;
7068 if (RelocM == Reloc::PIC_) {
7069 NewVReg6 = MRI->createVirtualRegister(TRC);
7070 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7071 .addReg(ARM::CPSR, RegState::Define)
7072 .addReg(NewVReg5, RegState::Kill)
7076 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7077 .addReg(NewVReg6, RegState::Kill)
7078 .addJumpTableIndex(MJTI)
7081 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7082 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7085 .addMemOperand(FIMMOLd));
7087 if (NumLPads < 256) {
7088 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7091 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7092 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7093 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7094 .addImm(NumLPads & 0xFFFF));
7096 unsigned VReg2 = VReg1;
7097 if ((NumLPads & 0xFFFF0000) != 0) {
7098 VReg2 = MRI->createVirtualRegister(TRC);
7099 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7101 .addImm(NumLPads >> 16));
7104 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7108 MachineConstantPool *ConstantPool = MF->getConstantPool();
7109 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7110 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7112 // MachineConstantPool wants an explicit alignment.
7113 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7115 Align = getDataLayout()->getTypeAllocSize(C->getType());
7116 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7118 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7119 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7120 .addReg(VReg1, RegState::Define)
7121 .addConstantPoolIndex(Idx)
7123 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7125 .addReg(VReg1, RegState::Kill));
7128 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7133 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7135 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7137 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7138 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7139 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7140 .addJumpTableIndex(MJTI)
7143 MachineMemOperand *JTMMOLd =
7144 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7145 MachineMemOperand::MOLoad, 4, 4);
7146 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7148 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7149 .addReg(NewVReg3, RegState::Kill)
7152 .addMemOperand(JTMMOLd));
7154 if (RelocM == Reloc::PIC_) {
7155 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7156 .addReg(NewVReg5, RegState::Kill)
7158 .addJumpTableIndex(MJTI)
7161 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7162 .addReg(NewVReg5, RegState::Kill)
7163 .addJumpTableIndex(MJTI)
7168 // Add the jump table entries as successors to the MBB.
7169 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7170 for (std::vector<MachineBasicBlock*>::iterator
7171 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7172 MachineBasicBlock *CurMBB = *I;
7173 if (SeenMBBs.insert(CurMBB))
7174 DispContBB->addSuccessor(CurMBB);
7177 // N.B. the order the invoke BBs are processed in doesn't matter here.
7178 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7179 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7180 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7181 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7182 MachineBasicBlock *BB = *I;
7184 // Remove the landing pad successor from the invoke block and replace it
7185 // with the new dispatch block.
7186 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7188 while (!Successors.empty()) {
7189 MachineBasicBlock *SMBB = Successors.pop_back_val();
7190 if (SMBB->isLandingPad()) {
7191 BB->removeSuccessor(SMBB);
7192 MBBLPads.push_back(SMBB);
7196 BB->addSuccessor(DispatchBB);
7198 // Find the invoke call and mark all of the callee-saved registers as
7199 // 'implicit defined' so that they're spilled. This prevents code from
7200 // moving instructions to before the EH block, where they will never be
7202 for (MachineBasicBlock::reverse_iterator
7203 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7204 if (!II->isCall()) continue;
7206 DenseMap<unsigned, bool> DefRegs;
7207 for (MachineInstr::mop_iterator
7208 OI = II->operands_begin(), OE = II->operands_end();
7210 if (!OI->isReg()) continue;
7211 DefRegs[OI->getReg()] = true;
7214 MachineInstrBuilder MIB(*MF, &*II);
7216 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7217 unsigned Reg = SavedRegs[i];
7218 if (Subtarget->isThumb2() &&
7219 !ARM::tGPRRegClass.contains(Reg) &&
7220 !ARM::hGPRRegClass.contains(Reg))
7222 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7224 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7227 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7234 // Mark all former landing pads as non-landing pads. The dispatch is the only
7236 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7237 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7238 (*I)->setIsLandingPad(false);
7240 // The instruction is gone now.
7241 MI->eraseFromParent();
7247 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7248 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7249 E = MBB->succ_end(); I != E; ++I)
7252 llvm_unreachable("Expecting a BB with two successors!");
7255 /// Return the load opcode for a given load size. If load size >= 8,
7256 /// neon opcode will be returned.
7257 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7259 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7260 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7262 return LdSize == 4 ? ARM::tLDRi
7263 : LdSize == 2 ? ARM::tLDRHi
7264 : LdSize == 1 ? ARM::tLDRBi : 0;
7266 return LdSize == 4 ? ARM::t2LDR_POST
7267 : LdSize == 2 ? ARM::t2LDRH_POST
7268 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7269 return LdSize == 4 ? ARM::LDR_POST_IMM
7270 : LdSize == 2 ? ARM::LDRH_POST
7271 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7274 /// Return the store opcode for a given store size. If store size >= 8,
7275 /// neon opcode will be returned.
7276 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7278 return StSize == 16 ? ARM::VST1q32wb_fixed
7279 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7281 return StSize == 4 ? ARM::tSTRi
7282 : StSize == 2 ? ARM::tSTRHi
7283 : StSize == 1 ? ARM::tSTRBi : 0;
7285 return StSize == 4 ? ARM::t2STR_POST
7286 : StSize == 2 ? ARM::t2STRH_POST
7287 : StSize == 1 ? ARM::t2STRB_POST : 0;
7288 return StSize == 4 ? ARM::STR_POST_IMM
7289 : StSize == 2 ? ARM::STRH_POST
7290 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7293 /// Emit a post-increment load operation with given size. The instructions
7294 /// will be added to BB at Pos.
7295 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7296 const TargetInstrInfo *TII, DebugLoc dl,
7297 unsigned LdSize, unsigned Data, unsigned AddrIn,
7298 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7299 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7300 assert(LdOpc != 0 && "Should have a load opcode");
7302 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7303 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7305 } else if (IsThumb1) {
7306 // load + update AddrIn
7307 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7308 .addReg(AddrIn).addImm(0));
7309 MachineInstrBuilder MIB =
7310 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7311 MIB = AddDefaultT1CC(MIB);
7312 MIB.addReg(AddrIn).addImm(LdSize);
7313 AddDefaultPred(MIB);
7314 } else if (IsThumb2) {
7315 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7316 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7319 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7320 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7321 .addReg(0).addImm(LdSize));
7325 /// Emit a post-increment store operation with given size. The instructions
7326 /// will be added to BB at Pos.
7327 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7328 const TargetInstrInfo *TII, DebugLoc dl,
7329 unsigned StSize, unsigned Data, unsigned AddrIn,
7330 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7331 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7332 assert(StOpc != 0 && "Should have a store opcode");
7334 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7335 .addReg(AddrIn).addImm(0).addReg(Data));
7336 } else if (IsThumb1) {
7337 // store + update AddrIn
7338 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7339 .addReg(AddrIn).addImm(0));
7340 MachineInstrBuilder MIB =
7341 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7342 MIB = AddDefaultT1CC(MIB);
7343 MIB.addReg(AddrIn).addImm(StSize);
7344 AddDefaultPred(MIB);
7345 } else if (IsThumb2) {
7346 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7347 .addReg(Data).addReg(AddrIn).addImm(StSize));
7349 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7350 .addReg(Data).addReg(AddrIn).addReg(0)
7356 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7357 MachineBasicBlock *BB) const {
7358 // This pseudo instruction has 3 operands: dst, src, size
7359 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7360 // Otherwise, we will generate unrolled scalar copies.
7361 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7362 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7363 MachineFunction::iterator It = BB;
7366 unsigned dest = MI->getOperand(0).getReg();
7367 unsigned src = MI->getOperand(1).getReg();
7368 unsigned SizeVal = MI->getOperand(2).getImm();
7369 unsigned Align = MI->getOperand(3).getImm();
7370 DebugLoc dl = MI->getDebugLoc();
7372 MachineFunction *MF = BB->getParent();
7373 MachineRegisterInfo &MRI = MF->getRegInfo();
7374 unsigned UnitSize = 0;
7375 const TargetRegisterClass *TRC = 0;
7376 const TargetRegisterClass *VecTRC = 0;
7378 bool IsThumb1 = Subtarget->isThumb1Only();
7379 bool IsThumb2 = Subtarget->isThumb2();
7383 } else if (Align & 2) {
7386 // Check whether we can use NEON instructions.
7387 if (!MF->getFunction()->getAttributes().
7388 hasAttribute(AttributeSet::FunctionIndex,
7389 Attribute::NoImplicitFloat) &&
7390 Subtarget->hasNEON()) {
7391 if ((Align % 16 == 0) && SizeVal >= 16)
7393 else if ((Align % 8 == 0) && SizeVal >= 8)
7396 // Can't use NEON instructions.
7401 // Select the correct opcode and register class for unit size load/store
7402 bool IsNeon = UnitSize >= 8;
7403 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
7404 : (const TargetRegisterClass *)&ARM::GPRRegClass;
7406 VecTRC = UnitSize == 16
7407 ? (const TargetRegisterClass *)&ARM::DPairRegClass
7409 ? (const TargetRegisterClass *)&ARM::DPRRegClass
7412 unsigned BytesLeft = SizeVal % UnitSize;
7413 unsigned LoopSize = SizeVal - BytesLeft;
7415 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7416 // Use LDR and STR to copy.
7417 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7418 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7419 unsigned srcIn = src;
7420 unsigned destIn = dest;
7421 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7422 unsigned srcOut = MRI.createVirtualRegister(TRC);
7423 unsigned destOut = MRI.createVirtualRegister(TRC);
7424 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7425 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7426 IsThumb1, IsThumb2);
7427 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7428 IsThumb1, IsThumb2);
7433 // Handle the leftover bytes with LDRB and STRB.
7434 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7435 // [destOut] = STRB_POST(scratch, destIn, 1)
7436 for (unsigned i = 0; i < BytesLeft; i++) {
7437 unsigned srcOut = MRI.createVirtualRegister(TRC);
7438 unsigned destOut = MRI.createVirtualRegister(TRC);
7439 unsigned scratch = MRI.createVirtualRegister(TRC);
7440 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7441 IsThumb1, IsThumb2);
7442 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7443 IsThumb1, IsThumb2);
7447 MI->eraseFromParent(); // The instruction is gone now.
7451 // Expand the pseudo op to a loop.
7454 // movw varEnd, # --> with thumb2
7456 // ldrcp varEnd, idx --> without thumb2
7457 // fallthrough --> loopMBB
7459 // PHI varPhi, varEnd, varLoop
7460 // PHI srcPhi, src, srcLoop
7461 // PHI destPhi, dst, destLoop
7462 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7463 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7464 // subs varLoop, varPhi, #UnitSize
7466 // fallthrough --> exitMBB
7468 // epilogue to handle left-over bytes
7469 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7470 // [destOut] = STRB_POST(scratch, destLoop, 1)
7471 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7472 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7473 MF->insert(It, loopMBB);
7474 MF->insert(It, exitMBB);
7476 // Transfer the remainder of BB and its successor edges to exitMBB.
7477 exitMBB->splice(exitMBB->begin(), BB,
7478 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7479 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7481 // Load an immediate to varEnd.
7482 unsigned varEnd = MRI.createVirtualRegister(TRC);
7484 unsigned Vtmp = varEnd;
7485 if ((LoopSize & 0xFFFF0000) != 0)
7486 Vtmp = MRI.createVirtualRegister(TRC);
7487 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7488 .addImm(LoopSize & 0xFFFF));
7490 if ((LoopSize & 0xFFFF0000) != 0)
7491 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7492 .addReg(Vtmp).addImm(LoopSize >> 16));
7494 MachineConstantPool *ConstantPool = MF->getConstantPool();
7495 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7496 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7498 // MachineConstantPool wants an explicit alignment.
7499 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7501 Align = getDataLayout()->getTypeAllocSize(C->getType());
7502 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7505 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7506 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7508 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7509 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7511 BB->addSuccessor(loopMBB);
7513 // Generate the loop body:
7514 // varPhi = PHI(varLoop, varEnd)
7515 // srcPhi = PHI(srcLoop, src)
7516 // destPhi = PHI(destLoop, dst)
7517 MachineBasicBlock *entryBB = BB;
7519 unsigned varLoop = MRI.createVirtualRegister(TRC);
7520 unsigned varPhi = MRI.createVirtualRegister(TRC);
7521 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7522 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7523 unsigned destLoop = MRI.createVirtualRegister(TRC);
7524 unsigned destPhi = MRI.createVirtualRegister(TRC);
7526 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7527 .addReg(varLoop).addMBB(loopMBB)
7528 .addReg(varEnd).addMBB(entryBB);
7529 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7530 .addReg(srcLoop).addMBB(loopMBB)
7531 .addReg(src).addMBB(entryBB);
7532 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7533 .addReg(destLoop).addMBB(loopMBB)
7534 .addReg(dest).addMBB(entryBB);
7536 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7537 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7538 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7539 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7540 IsThumb1, IsThumb2);
7541 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7542 IsThumb1, IsThumb2);
7544 // Decrement loop variable by UnitSize.
7546 MachineInstrBuilder MIB =
7547 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7548 MIB = AddDefaultT1CC(MIB);
7549 MIB.addReg(varPhi).addImm(UnitSize);
7550 AddDefaultPred(MIB);
7552 MachineInstrBuilder MIB =
7553 BuildMI(*BB, BB->end(), dl,
7554 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7555 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7556 MIB->getOperand(5).setReg(ARM::CPSR);
7557 MIB->getOperand(5).setIsDef(true);
7559 BuildMI(*BB, BB->end(), dl,
7560 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7561 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7563 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7564 BB->addSuccessor(loopMBB);
7565 BB->addSuccessor(exitMBB);
7567 // Add epilogue to handle BytesLeft.
7569 MachineInstr *StartOfExit = exitMBB->begin();
7571 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7572 // [destOut] = STRB_POST(scratch, destLoop, 1)
7573 unsigned srcIn = srcLoop;
7574 unsigned destIn = destLoop;
7575 for (unsigned i = 0; i < BytesLeft; i++) {
7576 unsigned srcOut = MRI.createVirtualRegister(TRC);
7577 unsigned destOut = MRI.createVirtualRegister(TRC);
7578 unsigned scratch = MRI.createVirtualRegister(TRC);
7579 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7580 IsThumb1, IsThumb2);
7581 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7582 IsThumb1, IsThumb2);
7587 MI->eraseFromParent(); // The instruction is gone now.
7592 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7593 MachineBasicBlock *BB) const {
7594 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7595 DebugLoc dl = MI->getDebugLoc();
7596 bool isThumb2 = Subtarget->isThumb2();
7597 switch (MI->getOpcode()) {
7600 llvm_unreachable("Unexpected instr type to insert");
7602 // The Thumb2 pre-indexed stores have the same MI operands, they just
7603 // define them differently in the .td files from the isel patterns, so
7604 // they need pseudos.
7605 case ARM::t2STR_preidx:
7606 MI->setDesc(TII->get(ARM::t2STR_PRE));
7608 case ARM::t2STRB_preidx:
7609 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7611 case ARM::t2STRH_preidx:
7612 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7615 case ARM::STRi_preidx:
7616 case ARM::STRBi_preidx: {
7617 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7618 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7619 // Decode the offset.
7620 unsigned Offset = MI->getOperand(4).getImm();
7621 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7622 Offset = ARM_AM::getAM2Offset(Offset);
7626 MachineMemOperand *MMO = *MI->memoperands_begin();
7627 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7628 .addOperand(MI->getOperand(0)) // Rn_wb
7629 .addOperand(MI->getOperand(1)) // Rt
7630 .addOperand(MI->getOperand(2)) // Rn
7631 .addImm(Offset) // offset (skip GPR==zero_reg)
7632 .addOperand(MI->getOperand(5)) // pred
7633 .addOperand(MI->getOperand(6))
7634 .addMemOperand(MMO);
7635 MI->eraseFromParent();
7638 case ARM::STRr_preidx:
7639 case ARM::STRBr_preidx:
7640 case ARM::STRH_preidx: {
7642 switch (MI->getOpcode()) {
7643 default: llvm_unreachable("unexpected opcode!");
7644 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7645 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7646 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7648 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7649 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7650 MIB.addOperand(MI->getOperand(i));
7651 MI->eraseFromParent();
7654 case ARM::ATOMIC_LOAD_ADD_I8:
7655 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7656 case ARM::ATOMIC_LOAD_ADD_I16:
7657 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7658 case ARM::ATOMIC_LOAD_ADD_I32:
7659 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7661 case ARM::ATOMIC_LOAD_AND_I8:
7662 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7663 case ARM::ATOMIC_LOAD_AND_I16:
7664 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7665 case ARM::ATOMIC_LOAD_AND_I32:
7666 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7668 case ARM::ATOMIC_LOAD_OR_I8:
7669 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7670 case ARM::ATOMIC_LOAD_OR_I16:
7671 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7672 case ARM::ATOMIC_LOAD_OR_I32:
7673 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7675 case ARM::ATOMIC_LOAD_XOR_I8:
7676 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7677 case ARM::ATOMIC_LOAD_XOR_I16:
7678 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7679 case ARM::ATOMIC_LOAD_XOR_I32:
7680 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7682 case ARM::ATOMIC_LOAD_NAND_I8:
7683 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7684 case ARM::ATOMIC_LOAD_NAND_I16:
7685 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7686 case ARM::ATOMIC_LOAD_NAND_I32:
7687 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7689 case ARM::ATOMIC_LOAD_SUB_I8:
7690 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7691 case ARM::ATOMIC_LOAD_SUB_I16:
7692 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7693 case ARM::ATOMIC_LOAD_SUB_I32:
7694 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7696 case ARM::ATOMIC_LOAD_MIN_I8:
7697 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7698 case ARM::ATOMIC_LOAD_MIN_I16:
7699 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7700 case ARM::ATOMIC_LOAD_MIN_I32:
7701 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7703 case ARM::ATOMIC_LOAD_MAX_I8:
7704 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7705 case ARM::ATOMIC_LOAD_MAX_I16:
7706 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7707 case ARM::ATOMIC_LOAD_MAX_I32:
7708 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7710 case ARM::ATOMIC_LOAD_UMIN_I8:
7711 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7712 case ARM::ATOMIC_LOAD_UMIN_I16:
7713 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7714 case ARM::ATOMIC_LOAD_UMIN_I32:
7715 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7717 case ARM::ATOMIC_LOAD_UMAX_I8:
7718 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7719 case ARM::ATOMIC_LOAD_UMAX_I16:
7720 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7721 case ARM::ATOMIC_LOAD_UMAX_I32:
7722 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7724 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7725 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7726 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7728 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7729 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7730 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7732 case ARM::ATOMIC_LOAD_I64:
7733 return EmitAtomicLoad64(MI, BB);
7735 case ARM::ATOMIC_LOAD_ADD_I64:
7736 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7737 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7738 /*NeedsCarry*/ true);
7739 case ARM::ATOMIC_LOAD_SUB_I64:
7740 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7741 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7742 /*NeedsCarry*/ true);
7743 case ARM::ATOMIC_LOAD_OR_I64:
7744 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7745 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7746 case ARM::ATOMIC_LOAD_XOR_I64:
7747 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7748 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7749 case ARM::ATOMIC_LOAD_AND_I64:
7750 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7751 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7752 case ARM::ATOMIC_SWAP_I64:
7753 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7754 case ARM::ATOMIC_CMP_SWAP_I64:
7755 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7756 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7757 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7758 case ARM::ATOMIC_LOAD_MIN_I64:
7759 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7760 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7761 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7762 /*IsMinMax*/ true, ARMCC::LT);
7763 case ARM::ATOMIC_LOAD_MAX_I64:
7764 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7765 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7766 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7767 /*IsMinMax*/ true, ARMCC::GE);
7768 case ARM::ATOMIC_LOAD_UMIN_I64:
7769 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7770 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7771 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7772 /*IsMinMax*/ true, ARMCC::LO);
7773 case ARM::ATOMIC_LOAD_UMAX_I64:
7774 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7775 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7776 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7777 /*IsMinMax*/ true, ARMCC::HS);
7779 case ARM::tMOVCCr_pseudo: {
7780 // To "insert" a SELECT_CC instruction, we actually have to insert the
7781 // diamond control-flow pattern. The incoming instruction knows the
7782 // destination vreg to set, the condition code register to branch on, the
7783 // true/false values to select between, and a branch opcode to use.
7784 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7785 MachineFunction::iterator It = BB;
7791 // cmpTY ccX, r1, r2
7793 // fallthrough --> copy0MBB
7794 MachineBasicBlock *thisMBB = BB;
7795 MachineFunction *F = BB->getParent();
7796 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7797 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7798 F->insert(It, copy0MBB);
7799 F->insert(It, sinkMBB);
7801 // Transfer the remainder of BB and its successor edges to sinkMBB.
7802 sinkMBB->splice(sinkMBB->begin(), BB,
7803 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7804 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7806 BB->addSuccessor(copy0MBB);
7807 BB->addSuccessor(sinkMBB);
7809 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7810 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7813 // %FalseValue = ...
7814 // # fallthrough to sinkMBB
7817 // Update machine-CFG edges
7818 BB->addSuccessor(sinkMBB);
7821 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7824 BuildMI(*BB, BB->begin(), dl,
7825 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7826 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7827 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7829 MI->eraseFromParent(); // The pseudo instruction is gone now.
7834 case ARM::BCCZi64: {
7835 // If there is an unconditional branch to the other successor, remove it.
7836 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7838 // Compare both parts that make up the double comparison separately for
7840 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7842 unsigned LHS1 = MI->getOperand(1).getReg();
7843 unsigned LHS2 = MI->getOperand(2).getReg();
7845 AddDefaultPred(BuildMI(BB, dl,
7846 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7847 .addReg(LHS1).addImm(0));
7848 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7849 .addReg(LHS2).addImm(0)
7850 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7852 unsigned RHS1 = MI->getOperand(3).getReg();
7853 unsigned RHS2 = MI->getOperand(4).getReg();
7854 AddDefaultPred(BuildMI(BB, dl,
7855 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7856 .addReg(LHS1).addReg(RHS1));
7857 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7858 .addReg(LHS2).addReg(RHS2)
7859 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7862 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7863 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7864 if (MI->getOperand(0).getImm() == ARMCC::NE)
7865 std::swap(destMBB, exitMBB);
7867 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7868 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7870 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7872 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7874 MI->eraseFromParent(); // The pseudo instruction is gone now.
7878 case ARM::Int_eh_sjlj_setjmp:
7879 case ARM::Int_eh_sjlj_setjmp_nofp:
7880 case ARM::tInt_eh_sjlj_setjmp:
7881 case ARM::t2Int_eh_sjlj_setjmp:
7882 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7883 EmitSjLjDispatchBlock(MI, BB);
7888 // To insert an ABS instruction, we have to insert the
7889 // diamond control-flow pattern. The incoming instruction knows the
7890 // source vreg to test against 0, the destination vreg to set,
7891 // the condition code register to branch on, the
7892 // true/false values to select between, and a branch opcode to use.
7897 // BCC (branch to SinkBB if V0 >= 0)
7898 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7899 // SinkBB: V1 = PHI(V2, V3)
7900 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7901 MachineFunction::iterator BBI = BB;
7903 MachineFunction *Fn = BB->getParent();
7904 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7905 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7906 Fn->insert(BBI, RSBBB);
7907 Fn->insert(BBI, SinkBB);
7909 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7910 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7911 bool isThumb2 = Subtarget->isThumb2();
7912 MachineRegisterInfo &MRI = Fn->getRegInfo();
7913 // In Thumb mode S must not be specified if source register is the SP or
7914 // PC and if destination register is the SP, so restrict register class
7915 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7916 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7917 (const TargetRegisterClass*)&ARM::GPRRegClass);
7919 // Transfer the remainder of BB and its successor edges to sinkMBB.
7920 SinkBB->splice(SinkBB->begin(), BB,
7921 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7922 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7924 BB->addSuccessor(RSBBB);
7925 BB->addSuccessor(SinkBB);
7927 // fall through to SinkMBB
7928 RSBBB->addSuccessor(SinkBB);
7930 // insert a cmp at the end of BB
7931 AddDefaultPred(BuildMI(BB, dl,
7932 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7933 .addReg(ABSSrcReg).addImm(0));
7935 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7937 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7938 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7940 // insert rsbri in RSBBB
7941 // Note: BCC and rsbri will be converted into predicated rsbmi
7942 // by if-conversion pass
7943 BuildMI(*RSBBB, RSBBB->begin(), dl,
7944 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7945 .addReg(ABSSrcReg, RegState::Kill)
7946 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7948 // insert PHI in SinkBB,
7949 // reuse ABSDstReg to not change uses of ABS instruction
7950 BuildMI(*SinkBB, SinkBB->begin(), dl,
7951 TII->get(ARM::PHI), ABSDstReg)
7952 .addReg(NewRsbDstReg).addMBB(RSBBB)
7953 .addReg(ABSSrcReg).addMBB(BB);
7955 // remove ABS instruction
7956 MI->eraseFromParent();
7958 // return last added BB
7961 case ARM::COPY_STRUCT_BYVAL_I32:
7963 return EmitStructByval(MI, BB);
7967 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7968 SDNode *Node) const {
7969 if (!MI->hasPostISelHook()) {
7970 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7971 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7975 const MCInstrDesc *MCID = &MI->getDesc();
7976 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7977 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7978 // operand is still set to noreg. If needed, set the optional operand's
7979 // register to CPSR, and remove the redundant implicit def.
7981 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7983 // Rename pseudo opcodes.
7984 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7986 const ARMBaseInstrInfo *TII =
7987 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7988 MCID = &TII->get(NewOpc);
7990 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7991 "converted opcode should be the same except for cc_out");
7995 // Add the optional cc_out operand
7996 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7998 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8000 // Any ARM instruction that sets the 's' bit should specify an optional
8001 // "cc_out" operand in the last operand position.
8002 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8003 assert(!NewOpc && "Optional cc_out operand required");
8006 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8007 // since we already have an optional CPSR def.
8008 bool definesCPSR = false;
8009 bool deadCPSR = false;
8010 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8012 const MachineOperand &MO = MI->getOperand(i);
8013 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8017 MI->RemoveOperand(i);
8022 assert(!NewOpc && "Optional cc_out operand required");
8025 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8027 assert(!MI->getOperand(ccOutIdx).getReg() &&
8028 "expect uninitialized optional cc_out operand");
8032 // If this instruction was defined with an optional CPSR def and its dag node
8033 // had a live implicit CPSR def, then activate the optional CPSR def.
8034 MachineOperand &MO = MI->getOperand(ccOutIdx);
8035 MO.setReg(ARM::CPSR);
8039 //===----------------------------------------------------------------------===//
8040 // ARM Optimization Hooks
8041 //===----------------------------------------------------------------------===//
8043 // Helper function that checks if N is a null or all ones constant.
8044 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8045 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8048 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8051 // Return true if N is conditionally 0 or all ones.
8052 // Detects these expressions where cc is an i1 value:
8054 // (select cc 0, y) [AllOnes=0]
8055 // (select cc y, 0) [AllOnes=0]
8056 // (zext cc) [AllOnes=0]
8057 // (sext cc) [AllOnes=0/1]
8058 // (select cc -1, y) [AllOnes=1]
8059 // (select cc y, -1) [AllOnes=1]
8061 // Invert is set when N is the null/all ones constant when CC is false.
8062 // OtherOp is set to the alternative value of N.
8063 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8064 SDValue &CC, bool &Invert,
8066 SelectionDAG &DAG) {
8067 switch (N->getOpcode()) {
8068 default: return false;
8070 CC = N->getOperand(0);
8071 SDValue N1 = N->getOperand(1);
8072 SDValue N2 = N->getOperand(2);
8073 if (isZeroOrAllOnes(N1, AllOnes)) {
8078 if (isZeroOrAllOnes(N2, AllOnes)) {
8085 case ISD::ZERO_EXTEND:
8086 // (zext cc) can never be the all ones value.
8090 case ISD::SIGN_EXTEND: {
8091 EVT VT = N->getValueType(0);
8092 CC = N->getOperand(0);
8093 if (CC.getValueType() != MVT::i1)
8097 // When looking for an AllOnes constant, N is an sext, and the 'other'
8099 OtherOp = DAG.getConstant(0, VT);
8100 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8101 // When looking for a 0 constant, N can be zext or sext.
8102 OtherOp = DAG.getConstant(1, VT);
8104 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
8110 // Combine a constant select operand into its use:
8112 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8113 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8114 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8115 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8116 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8118 // The transform is rejected if the select doesn't have a constant operand that
8119 // is null, or all ones when AllOnes is set.
8121 // Also recognize sext/zext from i1:
8123 // (add (zext cc), x) -> (select cc (add x, 1), x)
8124 // (add (sext cc), x) -> (select cc (add x, -1), x)
8126 // These transformations eventually create predicated instructions.
8128 // @param N The node to transform.
8129 // @param Slct The N operand that is a select.
8130 // @param OtherOp The other N operand (x above).
8131 // @param DCI Context.
8132 // @param AllOnes Require the select constant to be all ones instead of null.
8133 // @returns The new node, or SDValue() on failure.
8135 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8136 TargetLowering::DAGCombinerInfo &DCI,
8137 bool AllOnes = false) {
8138 SelectionDAG &DAG = DCI.DAG;
8139 EVT VT = N->getValueType(0);
8140 SDValue NonConstantVal;
8143 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8144 NonConstantVal, DAG))
8147 // Slct is now know to be the desired identity constant when CC is true.
8148 SDValue TrueVal = OtherOp;
8149 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8150 OtherOp, NonConstantVal);
8151 // Unless SwapSelectOps says CC should be false.
8153 std::swap(TrueVal, FalseVal);
8155 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8156 CCOp, TrueVal, FalseVal);
8159 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8161 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8162 TargetLowering::DAGCombinerInfo &DCI) {
8163 SDValue N0 = N->getOperand(0);
8164 SDValue N1 = N->getOperand(1);
8165 if (N0.getNode()->hasOneUse()) {
8166 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8167 if (Result.getNode())
8170 if (N1.getNode()->hasOneUse()) {
8171 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8172 if (Result.getNode())
8178 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8179 // (only after legalization).
8180 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8181 TargetLowering::DAGCombinerInfo &DCI,
8182 const ARMSubtarget *Subtarget) {
8184 // Only perform optimization if after legalize, and if NEON is available. We
8185 // also expected both operands to be BUILD_VECTORs.
8186 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8187 || N0.getOpcode() != ISD::BUILD_VECTOR
8188 || N1.getOpcode() != ISD::BUILD_VECTOR)
8191 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8192 EVT VT = N->getValueType(0);
8193 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8196 // Check that the vector operands are of the right form.
8197 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8198 // operands, where N is the size of the formed vector.
8199 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8200 // index such that we have a pair wise add pattern.
8202 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8203 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8205 SDValue Vec = N0->getOperand(0)->getOperand(0);
8206 SDNode *V = Vec.getNode();
8207 unsigned nextIndex = 0;
8209 // For each operands to the ADD which are BUILD_VECTORs,
8210 // check to see if each of their operands are an EXTRACT_VECTOR with
8211 // the same vector and appropriate index.
8212 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8213 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8214 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8216 SDValue ExtVec0 = N0->getOperand(i);
8217 SDValue ExtVec1 = N1->getOperand(i);
8219 // First operand is the vector, verify its the same.
8220 if (V != ExtVec0->getOperand(0).getNode() ||
8221 V != ExtVec1->getOperand(0).getNode())
8224 // Second is the constant, verify its correct.
8225 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8226 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8228 // For the constant, we want to see all the even or all the odd.
8229 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8230 || C1->getZExtValue() != nextIndex+1)
8239 // Create VPADDL node.
8240 SelectionDAG &DAG = DCI.DAG;
8241 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8243 // Build operand list.
8244 SmallVector<SDValue, 8> Ops;
8245 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8246 TLI.getPointerTy()));
8248 // Input is the vector.
8251 // Get widened type and narrowed type.
8253 unsigned numElem = VT.getVectorNumElements();
8254 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8255 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8256 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8257 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8259 llvm_unreachable("Invalid vector element type for padd optimization.");
8262 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8263 widenType, &Ops[0], Ops.size());
8264 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8267 static SDValue findMUL_LOHI(SDValue V) {
8268 if (V->getOpcode() == ISD::UMUL_LOHI ||
8269 V->getOpcode() == ISD::SMUL_LOHI)
8274 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8275 TargetLowering::DAGCombinerInfo &DCI,
8276 const ARMSubtarget *Subtarget) {
8278 if (Subtarget->isThumb1Only()) return SDValue();
8280 // Only perform the checks after legalize when the pattern is available.
8281 if (DCI.isBeforeLegalize()) return SDValue();
8283 // Look for multiply add opportunities.
8284 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8285 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8286 // a glue link from the first add to the second add.
8287 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8288 // a S/UMLAL instruction.
8291 // \ / \ [no multiline comment]
8297 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8298 SDValue AddcOp0 = AddcNode->getOperand(0);
8299 SDValue AddcOp1 = AddcNode->getOperand(1);
8301 // Check if the two operands are from the same mul_lohi node.
8302 if (AddcOp0.getNode() == AddcOp1.getNode())
8305 assert(AddcNode->getNumValues() == 2 &&
8306 AddcNode->getValueType(0) == MVT::i32 &&
8307 "Expect ADDC with two result values. First: i32");
8309 // Check that we have a glued ADDC node.
8310 if (AddcNode->getValueType(1) != MVT::Glue)
8313 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8314 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8315 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8316 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8317 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8320 // Look for the glued ADDE.
8321 SDNode* AddeNode = AddcNode->getGluedUser();
8322 if (AddeNode == NULL)
8325 // Make sure it is really an ADDE.
8326 if (AddeNode->getOpcode() != ISD::ADDE)
8329 assert(AddeNode->getNumOperands() == 3 &&
8330 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8331 "ADDE node has the wrong inputs");
8333 // Check for the triangle shape.
8334 SDValue AddeOp0 = AddeNode->getOperand(0);
8335 SDValue AddeOp1 = AddeNode->getOperand(1);
8337 // Make sure that the ADDE operands are not coming from the same node.
8338 if (AddeOp0.getNode() == AddeOp1.getNode())
8341 // Find the MUL_LOHI node walking up ADDE's operands.
8342 bool IsLeftOperandMUL = false;
8343 SDValue MULOp = findMUL_LOHI(AddeOp0);
8344 if (MULOp == SDValue())
8345 MULOp = findMUL_LOHI(AddeOp1);
8347 IsLeftOperandMUL = true;
8348 if (MULOp == SDValue())
8351 // Figure out the right opcode.
8352 unsigned Opc = MULOp->getOpcode();
8353 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8355 // Figure out the high and low input values to the MLAL node.
8356 SDValue* HiMul = &MULOp;
8357 SDValue* HiAdd = NULL;
8358 SDValue* LoMul = NULL;
8359 SDValue* LowAdd = NULL;
8361 if (IsLeftOperandMUL)
8367 if (AddcOp0->getOpcode() == Opc) {
8371 if (AddcOp1->getOpcode() == Opc) {
8379 if (LoMul->getNode() != HiMul->getNode())
8382 // Create the merged node.
8383 SelectionDAG &DAG = DCI.DAG;
8385 // Build operand list.
8386 SmallVector<SDValue, 8> Ops;
8387 Ops.push_back(LoMul->getOperand(0));
8388 Ops.push_back(LoMul->getOperand(1));
8389 Ops.push_back(*LowAdd);
8390 Ops.push_back(*HiAdd);
8392 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8393 DAG.getVTList(MVT::i32, MVT::i32),
8394 &Ops[0], Ops.size());
8396 // Replace the ADDs' nodes uses by the MLA node's values.
8397 SDValue HiMLALResult(MLALNode.getNode(), 1);
8398 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8400 SDValue LoMLALResult(MLALNode.getNode(), 0);
8401 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8403 // Return original node to notify the driver to stop replacing.
8404 SDValue resNode(AddcNode, 0);
8408 /// PerformADDCCombine - Target-specific dag combine transform from
8409 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8410 static SDValue PerformADDCCombine(SDNode *N,
8411 TargetLowering::DAGCombinerInfo &DCI,
8412 const ARMSubtarget *Subtarget) {
8414 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8418 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8419 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8420 /// called with the default operands, and if that fails, with commuted
8422 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8423 TargetLowering::DAGCombinerInfo &DCI,
8424 const ARMSubtarget *Subtarget){
8426 // Attempt to create vpaddl for this add.
8427 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8428 if (Result.getNode())
8431 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8432 if (N0.getNode()->hasOneUse()) {
8433 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8434 if (Result.getNode()) return Result;
8439 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8441 static SDValue PerformADDCombine(SDNode *N,
8442 TargetLowering::DAGCombinerInfo &DCI,
8443 const ARMSubtarget *Subtarget) {
8444 SDValue N0 = N->getOperand(0);
8445 SDValue N1 = N->getOperand(1);
8447 // First try with the default operand order.
8448 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8449 if (Result.getNode())
8452 // If that didn't work, try again with the operands commuted.
8453 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8456 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8458 static SDValue PerformSUBCombine(SDNode *N,
8459 TargetLowering::DAGCombinerInfo &DCI) {
8460 SDValue N0 = N->getOperand(0);
8461 SDValue N1 = N->getOperand(1);
8463 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8464 if (N1.getNode()->hasOneUse()) {
8465 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8466 if (Result.getNode()) return Result;
8472 /// PerformVMULCombine
8473 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8474 /// special multiplier accumulator forwarding.
8480 // However, for (A + B) * (A + B),
8487 static SDValue PerformVMULCombine(SDNode *N,
8488 TargetLowering::DAGCombinerInfo &DCI,
8489 const ARMSubtarget *Subtarget) {
8490 if (!Subtarget->hasVMLxForwarding())
8493 SelectionDAG &DAG = DCI.DAG;
8494 SDValue N0 = N->getOperand(0);
8495 SDValue N1 = N->getOperand(1);
8496 unsigned Opcode = N0.getOpcode();
8497 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8498 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8499 Opcode = N1.getOpcode();
8500 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8501 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8509 EVT VT = N->getValueType(0);
8511 SDValue N00 = N0->getOperand(0);
8512 SDValue N01 = N0->getOperand(1);
8513 return DAG.getNode(Opcode, DL, VT,
8514 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8515 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8518 static SDValue PerformMULCombine(SDNode *N,
8519 TargetLowering::DAGCombinerInfo &DCI,
8520 const ARMSubtarget *Subtarget) {
8521 SelectionDAG &DAG = DCI.DAG;
8523 if (Subtarget->isThumb1Only())
8526 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8529 EVT VT = N->getValueType(0);
8530 if (VT.is64BitVector() || VT.is128BitVector())
8531 return PerformVMULCombine(N, DCI, Subtarget);
8535 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8539 int64_t MulAmt = C->getSExtValue();
8540 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8542 ShiftAmt = ShiftAmt & (32 - 1);
8543 SDValue V = N->getOperand(0);
8547 MulAmt >>= ShiftAmt;
8550 if (isPowerOf2_32(MulAmt - 1)) {
8551 // (mul x, 2^N + 1) => (add (shl x, N), x)
8552 Res = DAG.getNode(ISD::ADD, DL, VT,
8554 DAG.getNode(ISD::SHL, DL, VT,
8556 DAG.getConstant(Log2_32(MulAmt - 1),
8558 } else if (isPowerOf2_32(MulAmt + 1)) {
8559 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8560 Res = DAG.getNode(ISD::SUB, DL, VT,
8561 DAG.getNode(ISD::SHL, DL, VT,
8563 DAG.getConstant(Log2_32(MulAmt + 1),
8569 uint64_t MulAmtAbs = -MulAmt;
8570 if (isPowerOf2_32(MulAmtAbs + 1)) {
8571 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8572 Res = DAG.getNode(ISD::SUB, DL, VT,
8574 DAG.getNode(ISD::SHL, DL, VT,
8576 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8578 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8579 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8580 Res = DAG.getNode(ISD::ADD, DL, VT,
8582 DAG.getNode(ISD::SHL, DL, VT,
8584 DAG.getConstant(Log2_32(MulAmtAbs-1),
8586 Res = DAG.getNode(ISD::SUB, DL, VT,
8587 DAG.getConstant(0, MVT::i32),Res);
8594 Res = DAG.getNode(ISD::SHL, DL, VT,
8595 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8597 // Do not add new nodes to DAG combiner worklist.
8598 DCI.CombineTo(N, Res, false);
8602 static SDValue PerformANDCombine(SDNode *N,
8603 TargetLowering::DAGCombinerInfo &DCI,
8604 const ARMSubtarget *Subtarget) {
8606 // Attempt to use immediate-form VBIC
8607 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8609 EVT VT = N->getValueType(0);
8610 SelectionDAG &DAG = DCI.DAG;
8612 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8615 APInt SplatBits, SplatUndef;
8616 unsigned SplatBitSize;
8619 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8620 if (SplatBitSize <= 64) {
8622 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8623 SplatUndef.getZExtValue(), SplatBitSize,
8624 DAG, VbicVT, VT.is128BitVector(),
8626 if (Val.getNode()) {
8628 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8629 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8630 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8635 if (!Subtarget->isThumb1Only()) {
8636 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8637 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8638 if (Result.getNode())
8645 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8646 static SDValue PerformORCombine(SDNode *N,
8647 TargetLowering::DAGCombinerInfo &DCI,
8648 const ARMSubtarget *Subtarget) {
8649 // Attempt to use immediate-form VORR
8650 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8652 EVT VT = N->getValueType(0);
8653 SelectionDAG &DAG = DCI.DAG;
8655 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8658 APInt SplatBits, SplatUndef;
8659 unsigned SplatBitSize;
8661 if (BVN && Subtarget->hasNEON() &&
8662 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8663 if (SplatBitSize <= 64) {
8665 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8666 SplatUndef.getZExtValue(), SplatBitSize,
8667 DAG, VorrVT, VT.is128BitVector(),
8669 if (Val.getNode()) {
8671 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8672 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8673 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8678 if (!Subtarget->isThumb1Only()) {
8679 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8680 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8681 if (Result.getNode())
8685 // The code below optimizes (or (and X, Y), Z).
8686 // The AND operand needs to have a single user to make these optimizations
8688 SDValue N0 = N->getOperand(0);
8689 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8691 SDValue N1 = N->getOperand(1);
8693 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8694 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8695 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8697 unsigned SplatBitSize;
8700 APInt SplatBits0, SplatBits1;
8701 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8702 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8703 // Ensure that the second operand of both ands are constants
8704 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8705 HasAnyUndefs) && !HasAnyUndefs) {
8706 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8707 HasAnyUndefs) && !HasAnyUndefs) {
8708 // Ensure that the bit width of the constants are the same and that
8709 // the splat arguments are logical inverses as per the pattern we
8710 // are trying to simplify.
8711 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8712 SplatBits0 == ~SplatBits1) {
8713 // Canonicalize the vector type to make instruction selection
8715 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8716 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8720 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8726 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8729 // BFI is only available on V6T2+
8730 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8734 // 1) or (and A, mask), val => ARMbfi A, val, mask
8735 // iff (val & mask) == val
8737 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8738 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8739 // && mask == ~mask2
8740 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8741 // && ~mask == mask2
8742 // (i.e., copy a bitfield value into another bitfield of the same width)
8747 SDValue N00 = N0.getOperand(0);
8749 // The value and the mask need to be constants so we can verify this is
8750 // actually a bitfield set. If the mask is 0xffff, we can do better
8751 // via a movt instruction, so don't use BFI in that case.
8752 SDValue MaskOp = N0.getOperand(1);
8753 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8756 unsigned Mask = MaskC->getZExtValue();
8760 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8761 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8763 unsigned Val = N1C->getZExtValue();
8764 if ((Val & ~Mask) != Val)
8767 if (ARM::isBitFieldInvertedMask(Mask)) {
8768 Val >>= countTrailingZeros(~Mask);
8770 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8771 DAG.getConstant(Val, MVT::i32),
8772 DAG.getConstant(Mask, MVT::i32));
8774 // Do not add new nodes to DAG combiner worklist.
8775 DCI.CombineTo(N, Res, false);
8778 } else if (N1.getOpcode() == ISD::AND) {
8779 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8780 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8783 unsigned Mask2 = N11C->getZExtValue();
8785 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8787 if (ARM::isBitFieldInvertedMask(Mask) &&
8789 // The pack halfword instruction works better for masks that fit it,
8790 // so use that when it's available.
8791 if (Subtarget->hasT2ExtractPack() &&
8792 (Mask == 0xffff || Mask == 0xffff0000))
8795 unsigned amt = countTrailingZeros(Mask2);
8796 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8797 DAG.getConstant(amt, MVT::i32));
8798 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8799 DAG.getConstant(Mask, MVT::i32));
8800 // Do not add new nodes to DAG combiner worklist.
8801 DCI.CombineTo(N, Res, false);
8803 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8805 // The pack halfword instruction works better for masks that fit it,
8806 // so use that when it's available.
8807 if (Subtarget->hasT2ExtractPack() &&
8808 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8811 unsigned lsb = countTrailingZeros(Mask);
8812 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8813 DAG.getConstant(lsb, MVT::i32));
8814 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8815 DAG.getConstant(Mask2, MVT::i32));
8816 // Do not add new nodes to DAG combiner worklist.
8817 DCI.CombineTo(N, Res, false);
8822 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8823 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8824 ARM::isBitFieldInvertedMask(~Mask)) {
8825 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8826 // where lsb(mask) == #shamt and masked bits of B are known zero.
8827 SDValue ShAmt = N00.getOperand(1);
8828 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8829 unsigned LSB = countTrailingZeros(Mask);
8833 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8834 DAG.getConstant(~Mask, MVT::i32));
8836 // Do not add new nodes to DAG combiner worklist.
8837 DCI.CombineTo(N, Res, false);
8843 static SDValue PerformXORCombine(SDNode *N,
8844 TargetLowering::DAGCombinerInfo &DCI,
8845 const ARMSubtarget *Subtarget) {
8846 EVT VT = N->getValueType(0);
8847 SelectionDAG &DAG = DCI.DAG;
8849 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8852 if (!Subtarget->isThumb1Only()) {
8853 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8854 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8855 if (Result.getNode())
8862 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8863 /// the bits being cleared by the AND are not demanded by the BFI.
8864 static SDValue PerformBFICombine(SDNode *N,
8865 TargetLowering::DAGCombinerInfo &DCI) {
8866 SDValue N1 = N->getOperand(1);
8867 if (N1.getOpcode() == ISD::AND) {
8868 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8871 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8872 unsigned LSB = countTrailingZeros(~InvMask);
8873 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8874 unsigned Mask = (1 << Width)-1;
8875 unsigned Mask2 = N11C->getZExtValue();
8876 if ((Mask & (~Mask2)) == 0)
8877 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8878 N->getOperand(0), N1.getOperand(0),
8884 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8885 /// ARMISD::VMOVRRD.
8886 static SDValue PerformVMOVRRDCombine(SDNode *N,
8887 TargetLowering::DAGCombinerInfo &DCI) {
8888 // vmovrrd(vmovdrr x, y) -> x,y
8889 SDValue InDouble = N->getOperand(0);
8890 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8891 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8893 // vmovrrd(load f64) -> (load i32), (load i32)
8894 SDNode *InNode = InDouble.getNode();
8895 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8896 InNode->getValueType(0) == MVT::f64 &&
8897 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8898 !cast<LoadSDNode>(InNode)->isVolatile()) {
8899 // TODO: Should this be done for non-FrameIndex operands?
8900 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8902 SelectionDAG &DAG = DCI.DAG;
8904 SDValue BasePtr = LD->getBasePtr();
8905 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8906 LD->getPointerInfo(), LD->isVolatile(),
8907 LD->isNonTemporal(), LD->isInvariant(),
8908 LD->getAlignment());
8910 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8911 DAG.getConstant(4, MVT::i32));
8912 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8913 LD->getPointerInfo(), LD->isVolatile(),
8914 LD->isNonTemporal(), LD->isInvariant(),
8915 std::min(4U, LD->getAlignment() / 2));
8917 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8918 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8919 DCI.RemoveFromWorklist(LD);
8927 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8928 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8929 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8930 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8931 SDValue Op0 = N->getOperand(0);
8932 SDValue Op1 = N->getOperand(1);
8933 if (Op0.getOpcode() == ISD::BITCAST)
8934 Op0 = Op0.getOperand(0);
8935 if (Op1.getOpcode() == ISD::BITCAST)
8936 Op1 = Op1.getOperand(0);
8937 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8938 Op0.getNode() == Op1.getNode() &&
8939 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8940 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8941 N->getValueType(0), Op0.getOperand(0));
8945 /// PerformSTORECombine - Target-specific dag combine xforms for
8947 static SDValue PerformSTORECombine(SDNode *N,
8948 TargetLowering::DAGCombinerInfo &DCI) {
8949 StoreSDNode *St = cast<StoreSDNode>(N);
8950 if (St->isVolatile())
8953 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8954 // pack all of the elements in one place. Next, store to memory in fewer
8956 SDValue StVal = St->getValue();
8957 EVT VT = StVal.getValueType();
8958 if (St->isTruncatingStore() && VT.isVector()) {
8959 SelectionDAG &DAG = DCI.DAG;
8960 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8961 EVT StVT = St->getMemoryVT();
8962 unsigned NumElems = VT.getVectorNumElements();
8963 assert(StVT != VT && "Cannot truncate to the same type");
8964 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8965 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8967 // From, To sizes and ElemCount must be pow of two
8968 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8970 // We are going to use the original vector elt for storing.
8971 // Accumulated smaller vector elements must be a multiple of the store size.
8972 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8974 unsigned SizeRatio = FromEltSz / ToEltSz;
8975 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8977 // Create a type on which we perform the shuffle.
8978 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8979 NumElems*SizeRatio);
8980 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8983 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8984 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8985 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8987 // Can't shuffle using an illegal type.
8988 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8990 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8991 DAG.getUNDEF(WideVec.getValueType()),
8993 // At this point all of the data is stored at the bottom of the
8994 // register. We now need to save it to mem.
8996 // Find the largest store unit
8997 MVT StoreType = MVT::i8;
8998 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8999 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
9000 MVT Tp = (MVT::SimpleValueType)tp;
9001 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9004 // Didn't find a legal store type.
9005 if (!TLI.isTypeLegal(StoreType))
9008 // Bitcast the original vector into a vector of store-size units
9009 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9010 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9011 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9012 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9013 SmallVector<SDValue, 8> Chains;
9014 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9015 TLI.getPointerTy());
9016 SDValue BasePtr = St->getBasePtr();
9018 // Perform one or more big stores into memory.
9019 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9020 for (unsigned I = 0; I < E; I++) {
9021 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9022 StoreType, ShuffWide,
9023 DAG.getIntPtrConstant(I));
9024 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9025 St->getPointerInfo(), St->isVolatile(),
9026 St->isNonTemporal(), St->getAlignment());
9027 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9029 Chains.push_back(Ch);
9031 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
9035 if (!ISD::isNormalStore(St))
9038 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9039 // ARM stores of arguments in the same cache line.
9040 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9041 StVal.getNode()->hasOneUse()) {
9042 SelectionDAG &DAG = DCI.DAG;
9044 SDValue BasePtr = St->getBasePtr();
9045 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9046 StVal.getNode()->getOperand(0), BasePtr,
9047 St->getPointerInfo(), St->isVolatile(),
9048 St->isNonTemporal(), St->getAlignment());
9050 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9051 DAG.getConstant(4, MVT::i32));
9052 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
9053 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9054 St->isNonTemporal(),
9055 std::min(4U, St->getAlignment() / 2));
9058 if (StVal.getValueType() != MVT::i64 ||
9059 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9062 // Bitcast an i64 store extracted from a vector to f64.
9063 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9064 SelectionDAG &DAG = DCI.DAG;
9066 SDValue IntVec = StVal.getOperand(0);
9067 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9068 IntVec.getValueType().getVectorNumElements());
9069 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9070 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9071 Vec, StVal.getOperand(1));
9073 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9074 // Make the DAGCombiner fold the bitcasts.
9075 DCI.AddToWorklist(Vec.getNode());
9076 DCI.AddToWorklist(ExtElt.getNode());
9077 DCI.AddToWorklist(V.getNode());
9078 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9079 St->getPointerInfo(), St->isVolatile(),
9080 St->isNonTemporal(), St->getAlignment(),
9084 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9085 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9086 /// i64 vector to have f64 elements, since the value can then be loaded
9087 /// directly into a VFP register.
9088 static bool hasNormalLoadOperand(SDNode *N) {
9089 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9090 for (unsigned i = 0; i < NumElts; ++i) {
9091 SDNode *Elt = N->getOperand(i).getNode();
9092 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9098 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9099 /// ISD::BUILD_VECTOR.
9100 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9101 TargetLowering::DAGCombinerInfo &DCI){
9102 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9103 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9104 // into a pair of GPRs, which is fine when the value is used as a scalar,
9105 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9106 SelectionDAG &DAG = DCI.DAG;
9107 if (N->getNumOperands() == 2) {
9108 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9113 // Load i64 elements as f64 values so that type legalization does not split
9114 // them up into i32 values.
9115 EVT VT = N->getValueType(0);
9116 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9119 SmallVector<SDValue, 8> Ops;
9120 unsigned NumElts = VT.getVectorNumElements();
9121 for (unsigned i = 0; i < NumElts; ++i) {
9122 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9124 // Make the DAGCombiner fold the bitcast.
9125 DCI.AddToWorklist(V.getNode());
9127 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9128 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
9129 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9132 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9134 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9135 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9136 // At that time, we may have inserted bitcasts from integer to float.
9137 // If these bitcasts have survived DAGCombine, change the lowering of this
9138 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9139 // force to use floating point types.
9141 // Make sure we can change the type of the vector.
9142 // This is possible iff:
9143 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9144 // 1.1. Vector is used only once.
9145 // 1.2. Use is a bit convert to an integer type.
9146 // 2. The size of its operands are 32-bits (64-bits are not legal).
9147 EVT VT = N->getValueType(0);
9148 EVT EltVT = VT.getVectorElementType();
9150 // Check 1.1. and 2.
9151 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9154 // By construction, the input type must be float.
9155 assert(EltVT == MVT::f32 && "Unexpected type!");
9158 SDNode *Use = *N->use_begin();
9159 if (Use->getOpcode() != ISD::BITCAST ||
9160 Use->getValueType(0).isFloatingPoint())
9163 // Check profitability.
9164 // Model is, if more than half of the relevant operands are bitcast from
9165 // i32, turn the build_vector into a sequence of insert_vector_elt.
9166 // Relevant operands are everything that is not statically
9167 // (i.e., at compile time) bitcasted.
9168 unsigned NumOfBitCastedElts = 0;
9169 unsigned NumElts = VT.getVectorNumElements();
9170 unsigned NumOfRelevantElts = NumElts;
9171 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9172 SDValue Elt = N->getOperand(Idx);
9173 if (Elt->getOpcode() == ISD::BITCAST) {
9174 // Assume only bit cast to i32 will go away.
9175 if (Elt->getOperand(0).getValueType() == MVT::i32)
9176 ++NumOfBitCastedElts;
9177 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9178 // Constants are statically casted, thus do not count them as
9179 // relevant operands.
9180 --NumOfRelevantElts;
9183 // Check if more than half of the elements require a non-free bitcast.
9184 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9187 SelectionDAG &DAG = DCI.DAG;
9188 // Create the new vector type.
9189 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9190 // Check if the type is legal.
9191 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9192 if (!TLI.isTypeLegal(VecVT))
9196 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9197 // => BITCAST INSERT_VECTOR_ELT
9198 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9200 SDValue Vec = DAG.getUNDEF(VecVT);
9202 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9203 SDValue V = N->getOperand(Idx);
9204 if (V.getOpcode() == ISD::UNDEF)
9206 if (V.getOpcode() == ISD::BITCAST &&
9207 V->getOperand(0).getValueType() == MVT::i32)
9208 // Fold obvious case.
9209 V = V.getOperand(0);
9211 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9212 // Make the DAGCombiner fold the bitcasts.
9213 DCI.AddToWorklist(V.getNode());
9215 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9216 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9218 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9219 // Make the DAGCombiner fold the bitcasts.
9220 DCI.AddToWorklist(Vec.getNode());
9224 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9225 /// ISD::INSERT_VECTOR_ELT.
9226 static SDValue PerformInsertEltCombine(SDNode *N,
9227 TargetLowering::DAGCombinerInfo &DCI) {
9228 // Bitcast an i64 load inserted into a vector to f64.
9229 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9230 EVT VT = N->getValueType(0);
9231 SDNode *Elt = N->getOperand(1).getNode();
9232 if (VT.getVectorElementType() != MVT::i64 ||
9233 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9236 SelectionDAG &DAG = DCI.DAG;
9238 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9239 VT.getVectorNumElements());
9240 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9241 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9242 // Make the DAGCombiner fold the bitcasts.
9243 DCI.AddToWorklist(Vec.getNode());
9244 DCI.AddToWorklist(V.getNode());
9245 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9246 Vec, V, N->getOperand(2));
9247 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9250 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9251 /// ISD::VECTOR_SHUFFLE.
9252 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9253 // The LLVM shufflevector instruction does not require the shuffle mask
9254 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9255 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9256 // operands do not match the mask length, they are extended by concatenating
9257 // them with undef vectors. That is probably the right thing for other
9258 // targets, but for NEON it is better to concatenate two double-register
9259 // size vector operands into a single quad-register size vector. Do that
9260 // transformation here:
9261 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9262 // shuffle(concat(v1, v2), undef)
9263 SDValue Op0 = N->getOperand(0);
9264 SDValue Op1 = N->getOperand(1);
9265 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9266 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9267 Op0.getNumOperands() != 2 ||
9268 Op1.getNumOperands() != 2)
9270 SDValue Concat0Op1 = Op0.getOperand(1);
9271 SDValue Concat1Op1 = Op1.getOperand(1);
9272 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9273 Concat1Op1.getOpcode() != ISD::UNDEF)
9275 // Skip the transformation if any of the types are illegal.
9276 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9277 EVT VT = N->getValueType(0);
9278 if (!TLI.isTypeLegal(VT) ||
9279 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9280 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9283 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9284 Op0.getOperand(0), Op1.getOperand(0));
9285 // Translate the shuffle mask.
9286 SmallVector<int, 16> NewMask;
9287 unsigned NumElts = VT.getVectorNumElements();
9288 unsigned HalfElts = NumElts/2;
9289 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9290 for (unsigned n = 0; n < NumElts; ++n) {
9291 int MaskElt = SVN->getMaskElt(n);
9293 if (MaskElt < (int)HalfElts)
9295 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9296 NewElt = HalfElts + MaskElt - NumElts;
9297 NewMask.push_back(NewElt);
9299 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9300 DAG.getUNDEF(VT), NewMask.data());
9303 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9304 /// NEON load/store intrinsics to merge base address updates.
9305 static SDValue CombineBaseUpdate(SDNode *N,
9306 TargetLowering::DAGCombinerInfo &DCI) {
9307 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9310 SelectionDAG &DAG = DCI.DAG;
9311 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9312 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9313 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9314 SDValue Addr = N->getOperand(AddrOpIdx);
9316 // Search for a use of the address operand that is an increment.
9317 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9318 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9320 if (User->getOpcode() != ISD::ADD ||
9321 UI.getUse().getResNo() != Addr.getResNo())
9324 // Check that the add is independent of the load/store. Otherwise, folding
9325 // it would create a cycle.
9326 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9329 // Find the new opcode for the updating load/store.
9331 bool isLaneOp = false;
9332 unsigned NewOpc = 0;
9333 unsigned NumVecs = 0;
9335 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9337 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9338 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9340 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9342 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9344 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9346 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9347 NumVecs = 2; isLaneOp = true; break;
9348 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9349 NumVecs = 3; isLaneOp = true; break;
9350 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9351 NumVecs = 4; isLaneOp = true; break;
9352 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9353 NumVecs = 1; isLoad = false; break;
9354 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9355 NumVecs = 2; isLoad = false; break;
9356 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9357 NumVecs = 3; isLoad = false; break;
9358 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9359 NumVecs = 4; isLoad = false; break;
9360 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9361 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9362 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9363 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9364 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9365 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9369 switch (N->getOpcode()) {
9370 default: llvm_unreachable("unexpected opcode for Neon base update");
9371 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9372 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9373 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9377 // Find the size of memory referenced by the load/store.
9380 VecTy = N->getValueType(0);
9382 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9383 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9385 NumBytes /= VecTy.getVectorNumElements();
9387 // If the increment is a constant, it must match the memory ref size.
9388 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9389 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9390 uint64_t IncVal = CInc->getZExtValue();
9391 if (IncVal != NumBytes)
9393 } else if (NumBytes >= 3 * 16) {
9394 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9395 // separate instructions that make it harder to use a non-constant update.
9399 // Create the new updating load/store node.
9401 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9403 for (n = 0; n < NumResultVecs; ++n)
9405 Tys[n++] = MVT::i32;
9406 Tys[n] = MVT::Other;
9407 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9408 SmallVector<SDValue, 8> Ops;
9409 Ops.push_back(N->getOperand(0)); // incoming chain
9410 Ops.push_back(N->getOperand(AddrOpIdx));
9412 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9413 Ops.push_back(N->getOperand(i));
9415 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9416 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9417 Ops.data(), Ops.size(),
9418 MemInt->getMemoryVT(),
9419 MemInt->getMemOperand());
9422 std::vector<SDValue> NewResults;
9423 for (unsigned i = 0; i < NumResultVecs; ++i) {
9424 NewResults.push_back(SDValue(UpdN.getNode(), i));
9426 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9427 DCI.CombineTo(N, NewResults);
9428 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9435 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9436 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9437 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9439 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9440 SelectionDAG &DAG = DCI.DAG;
9441 EVT VT = N->getValueType(0);
9442 // vldN-dup instructions only support 64-bit vectors for N > 1.
9443 if (!VT.is64BitVector())
9446 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9447 SDNode *VLD = N->getOperand(0).getNode();
9448 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9450 unsigned NumVecs = 0;
9451 unsigned NewOpc = 0;
9452 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9453 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9455 NewOpc = ARMISD::VLD2DUP;
9456 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9458 NewOpc = ARMISD::VLD3DUP;
9459 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9461 NewOpc = ARMISD::VLD4DUP;
9466 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9467 // numbers match the load.
9468 unsigned VLDLaneNo =
9469 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9470 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9472 // Ignore uses of the chain result.
9473 if (UI.getUse().getResNo() == NumVecs)
9476 if (User->getOpcode() != ARMISD::VDUPLANE ||
9477 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9481 // Create the vldN-dup node.
9484 for (n = 0; n < NumVecs; ++n)
9486 Tys[n] = MVT::Other;
9487 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9488 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9489 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9490 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9491 Ops, 2, VLDMemInt->getMemoryVT(),
9492 VLDMemInt->getMemOperand());
9495 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9497 unsigned ResNo = UI.getUse().getResNo();
9498 // Ignore uses of the chain result.
9499 if (ResNo == NumVecs)
9502 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9505 // Now the vldN-lane intrinsic is dead except for its chain result.
9506 // Update uses of the chain.
9507 std::vector<SDValue> VLDDupResults;
9508 for (unsigned n = 0; n < NumVecs; ++n)
9509 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9510 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9511 DCI.CombineTo(VLD, VLDDupResults);
9516 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9517 /// ARMISD::VDUPLANE.
9518 static SDValue PerformVDUPLANECombine(SDNode *N,
9519 TargetLowering::DAGCombinerInfo &DCI) {
9520 SDValue Op = N->getOperand(0);
9522 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9523 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9524 if (CombineVLDDUP(N, DCI))
9525 return SDValue(N, 0);
9527 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9528 // redundant. Ignore bit_converts for now; element sizes are checked below.
9529 while (Op.getOpcode() == ISD::BITCAST)
9530 Op = Op.getOperand(0);
9531 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9534 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9535 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9536 // The canonical VMOV for a zero vector uses a 32-bit element size.
9537 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9539 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9541 EVT VT = N->getValueType(0);
9542 if (EltSize > VT.getVectorElementType().getSizeInBits())
9545 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9548 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9549 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9550 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9554 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9556 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9561 APFloat APF = C->getValueAPF();
9562 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9563 != APFloat::opOK || !isExact)
9566 c0 = (I == 0) ? cN : c0;
9567 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9574 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9575 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9576 /// when the VMUL has a constant operand that is a power of 2.
9578 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9579 /// vmul.f32 d16, d17, d16
9580 /// vcvt.s32.f32 d16, d16
9582 /// vcvt.s32.f32 d16, d16, #3
9583 static SDValue PerformVCVTCombine(SDNode *N,
9584 TargetLowering::DAGCombinerInfo &DCI,
9585 const ARMSubtarget *Subtarget) {
9586 SelectionDAG &DAG = DCI.DAG;
9587 SDValue Op = N->getOperand(0);
9589 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9590 Op.getOpcode() != ISD::FMUL)
9594 SDValue N0 = Op->getOperand(0);
9595 SDValue ConstVec = Op->getOperand(1);
9596 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9598 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9599 !isConstVecPow2(ConstVec, isSigned, C))
9602 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9603 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9604 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9605 // These instructions only exist converting from f32 to i32. We can handle
9606 // smaller integers by generating an extra truncate, but larger ones would
9611 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9612 Intrinsic::arm_neon_vcvtfp2fxu;
9613 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9614 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9615 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9616 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9617 DAG.getConstant(Log2_64(C), MVT::i32));
9619 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9620 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9625 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9626 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9627 /// when the VDIV has a constant operand that is a power of 2.
9629 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9630 /// vcvt.f32.s32 d16, d16
9631 /// vdiv.f32 d16, d17, d16
9633 /// vcvt.f32.s32 d16, d16, #3
9634 static SDValue PerformVDIVCombine(SDNode *N,
9635 TargetLowering::DAGCombinerInfo &DCI,
9636 const ARMSubtarget *Subtarget) {
9637 SelectionDAG &DAG = DCI.DAG;
9638 SDValue Op = N->getOperand(0);
9639 unsigned OpOpcode = Op.getNode()->getOpcode();
9641 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9642 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9646 SDValue ConstVec = N->getOperand(1);
9647 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9649 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9650 !isConstVecPow2(ConstVec, isSigned, C))
9653 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9654 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9655 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9656 // These instructions only exist converting from i32 to f32. We can handle
9657 // smaller integers by generating an extra extend, but larger ones would
9662 SDValue ConvInput = Op.getOperand(0);
9663 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9664 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9665 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9666 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9669 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9670 Intrinsic::arm_neon_vcvtfxu2fp;
9671 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9673 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9674 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9677 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9678 /// operand of a vector shift operation, where all the elements of the
9679 /// build_vector must have the same constant integer value.
9680 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9681 // Ignore bit_converts.
9682 while (Op.getOpcode() == ISD::BITCAST)
9683 Op = Op.getOperand(0);
9684 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9685 APInt SplatBits, SplatUndef;
9686 unsigned SplatBitSize;
9688 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9689 HasAnyUndefs, ElementBits) ||
9690 SplatBitSize > ElementBits)
9692 Cnt = SplatBits.getSExtValue();
9696 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9697 /// operand of a vector shift left operation. That value must be in the range:
9698 /// 0 <= Value < ElementBits for a left shift; or
9699 /// 0 <= Value <= ElementBits for a long left shift.
9700 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9701 assert(VT.isVector() && "vector shift count is not a vector type");
9702 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9703 if (! getVShiftImm(Op, ElementBits, Cnt))
9705 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9708 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9709 /// operand of a vector shift right operation. For a shift opcode, the value
9710 /// is positive, but for an intrinsic the value count must be negative. The
9711 /// absolute value must be in the range:
9712 /// 1 <= |Value| <= ElementBits for a right shift; or
9713 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9714 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9716 assert(VT.isVector() && "vector shift count is not a vector type");
9717 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9718 if (! getVShiftImm(Op, ElementBits, Cnt))
9722 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9725 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9726 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9727 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9730 // Don't do anything for most intrinsics.
9733 // Vector shifts: check for immediate versions and lower them.
9734 // Note: This is done during DAG combining instead of DAG legalizing because
9735 // the build_vectors for 64-bit vector element shift counts are generally
9736 // not legal, and it is hard to see their values after they get legalized to
9737 // loads from a constant pool.
9738 case Intrinsic::arm_neon_vshifts:
9739 case Intrinsic::arm_neon_vshiftu:
9740 case Intrinsic::arm_neon_vrshifts:
9741 case Intrinsic::arm_neon_vrshiftu:
9742 case Intrinsic::arm_neon_vrshiftn:
9743 case Intrinsic::arm_neon_vqshifts:
9744 case Intrinsic::arm_neon_vqshiftu:
9745 case Intrinsic::arm_neon_vqshiftsu:
9746 case Intrinsic::arm_neon_vqshiftns:
9747 case Intrinsic::arm_neon_vqshiftnu:
9748 case Intrinsic::arm_neon_vqshiftnsu:
9749 case Intrinsic::arm_neon_vqrshiftns:
9750 case Intrinsic::arm_neon_vqrshiftnu:
9751 case Intrinsic::arm_neon_vqrshiftnsu: {
9752 EVT VT = N->getOperand(1).getValueType();
9754 unsigned VShiftOpc = 0;
9757 case Intrinsic::arm_neon_vshifts:
9758 case Intrinsic::arm_neon_vshiftu:
9759 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9760 VShiftOpc = ARMISD::VSHL;
9763 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9764 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9765 ARMISD::VSHRs : ARMISD::VSHRu);
9770 case Intrinsic::arm_neon_vrshifts:
9771 case Intrinsic::arm_neon_vrshiftu:
9772 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9776 case Intrinsic::arm_neon_vqshifts:
9777 case Intrinsic::arm_neon_vqshiftu:
9778 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9782 case Intrinsic::arm_neon_vqshiftsu:
9783 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9785 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9787 case Intrinsic::arm_neon_vrshiftn:
9788 case Intrinsic::arm_neon_vqshiftns:
9789 case Intrinsic::arm_neon_vqshiftnu:
9790 case Intrinsic::arm_neon_vqshiftnsu:
9791 case Intrinsic::arm_neon_vqrshiftns:
9792 case Intrinsic::arm_neon_vqrshiftnu:
9793 case Intrinsic::arm_neon_vqrshiftnsu:
9794 // Narrowing shifts require an immediate right shift.
9795 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9797 llvm_unreachable("invalid shift count for narrowing vector shift "
9801 llvm_unreachable("unhandled vector shift");
9805 case Intrinsic::arm_neon_vshifts:
9806 case Intrinsic::arm_neon_vshiftu:
9807 // Opcode already set above.
9809 case Intrinsic::arm_neon_vrshifts:
9810 VShiftOpc = ARMISD::VRSHRs; break;
9811 case Intrinsic::arm_neon_vrshiftu:
9812 VShiftOpc = ARMISD::VRSHRu; break;
9813 case Intrinsic::arm_neon_vrshiftn:
9814 VShiftOpc = ARMISD::VRSHRN; break;
9815 case Intrinsic::arm_neon_vqshifts:
9816 VShiftOpc = ARMISD::VQSHLs; break;
9817 case Intrinsic::arm_neon_vqshiftu:
9818 VShiftOpc = ARMISD::VQSHLu; break;
9819 case Intrinsic::arm_neon_vqshiftsu:
9820 VShiftOpc = ARMISD::VQSHLsu; break;
9821 case Intrinsic::arm_neon_vqshiftns:
9822 VShiftOpc = ARMISD::VQSHRNs; break;
9823 case Intrinsic::arm_neon_vqshiftnu:
9824 VShiftOpc = ARMISD::VQSHRNu; break;
9825 case Intrinsic::arm_neon_vqshiftnsu:
9826 VShiftOpc = ARMISD::VQSHRNsu; break;
9827 case Intrinsic::arm_neon_vqrshiftns:
9828 VShiftOpc = ARMISD::VQRSHRNs; break;
9829 case Intrinsic::arm_neon_vqrshiftnu:
9830 VShiftOpc = ARMISD::VQRSHRNu; break;
9831 case Intrinsic::arm_neon_vqrshiftnsu:
9832 VShiftOpc = ARMISD::VQRSHRNsu; break;
9835 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9836 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9839 case Intrinsic::arm_neon_vshiftins: {
9840 EVT VT = N->getOperand(1).getValueType();
9842 unsigned VShiftOpc = 0;
9844 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9845 VShiftOpc = ARMISD::VSLI;
9846 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9847 VShiftOpc = ARMISD::VSRI;
9849 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9852 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9853 N->getOperand(1), N->getOperand(2),
9854 DAG.getConstant(Cnt, MVT::i32));
9857 case Intrinsic::arm_neon_vqrshifts:
9858 case Intrinsic::arm_neon_vqrshiftu:
9859 // No immediate versions of these to check for.
9866 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9867 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9868 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9869 /// vector element shift counts are generally not legal, and it is hard to see
9870 /// their values after they get legalized to loads from a constant pool.
9871 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9872 const ARMSubtarget *ST) {
9873 EVT VT = N->getValueType(0);
9874 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9875 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9876 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9877 SDValue N1 = N->getOperand(1);
9878 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9879 SDValue N0 = N->getOperand(0);
9880 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9881 DAG.MaskedValueIsZero(N0.getOperand(0),
9882 APInt::getHighBitsSet(32, 16)))
9883 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9887 // Nothing to be done for scalar shifts.
9888 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9889 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9892 assert(ST->hasNEON() && "unexpected vector shift");
9895 switch (N->getOpcode()) {
9896 default: llvm_unreachable("unexpected shift opcode");
9899 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9900 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9901 DAG.getConstant(Cnt, MVT::i32));
9906 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9907 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9908 ARMISD::VSHRs : ARMISD::VSHRu);
9909 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9910 DAG.getConstant(Cnt, MVT::i32));
9916 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9917 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9918 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9919 const ARMSubtarget *ST) {
9920 SDValue N0 = N->getOperand(0);
9922 // Check for sign- and zero-extensions of vector extract operations of 8-
9923 // and 16-bit vector elements. NEON supports these directly. They are
9924 // handled during DAG combining because type legalization will promote them
9925 // to 32-bit types and it is messy to recognize the operations after that.
9926 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9927 SDValue Vec = N0.getOperand(0);
9928 SDValue Lane = N0.getOperand(1);
9929 EVT VT = N->getValueType(0);
9930 EVT EltVT = N0.getValueType();
9931 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9933 if (VT == MVT::i32 &&
9934 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9935 TLI.isTypeLegal(Vec.getValueType()) &&
9936 isa<ConstantSDNode>(Lane)) {
9939 switch (N->getOpcode()) {
9940 default: llvm_unreachable("unexpected opcode");
9941 case ISD::SIGN_EXTEND:
9942 Opc = ARMISD::VGETLANEs;
9944 case ISD::ZERO_EXTEND:
9945 case ISD::ANY_EXTEND:
9946 Opc = ARMISD::VGETLANEu;
9949 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9956 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9957 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9958 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9959 const ARMSubtarget *ST) {
9960 // If the target supports NEON, try to use vmax/vmin instructions for f32
9961 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9962 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9963 // a NaN; only do the transformation when it matches that behavior.
9965 // For now only do this when using NEON for FP operations; if using VFP, it
9966 // is not obvious that the benefit outweighs the cost of switching to the
9968 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9969 N->getValueType(0) != MVT::f32)
9972 SDValue CondLHS = N->getOperand(0);
9973 SDValue CondRHS = N->getOperand(1);
9974 SDValue LHS = N->getOperand(2);
9975 SDValue RHS = N->getOperand(3);
9976 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9978 unsigned Opcode = 0;
9980 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9981 IsReversed = false; // x CC y ? x : y
9982 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9983 IsReversed = true ; // x CC y ? y : x
9997 // If LHS is NaN, an ordered comparison will be false and the result will
9998 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9999 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10000 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
10001 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10003 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
10004 // will return -0, so vmin can only be used for unsafe math or if one of
10005 // the operands is known to be nonzero.
10006 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
10007 !DAG.getTarget().Options.UnsafeFPMath &&
10008 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10010 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
10019 // If LHS is NaN, an ordered comparison will be false and the result will
10020 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
10021 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10022 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
10023 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10025 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
10026 // will return +0, so vmax can only be used for unsafe math or if one of
10027 // the operands is known to be nonzero.
10028 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
10029 !DAG.getTarget().Options.UnsafeFPMath &&
10030 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10032 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
10038 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
10041 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10043 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10044 SDValue Cmp = N->getOperand(4);
10045 if (Cmp.getOpcode() != ARMISD::CMPZ)
10046 // Only looking at EQ and NE cases.
10049 EVT VT = N->getValueType(0);
10051 SDValue LHS = Cmp.getOperand(0);
10052 SDValue RHS = Cmp.getOperand(1);
10053 SDValue FalseVal = N->getOperand(0);
10054 SDValue TrueVal = N->getOperand(1);
10055 SDValue ARMcc = N->getOperand(2);
10056 ARMCC::CondCodes CC =
10057 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10075 /// FIXME: Turn this into a target neutral optimization?
10077 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10078 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10079 N->getOperand(3), Cmp);
10080 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10082 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10083 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10084 N->getOperand(3), NewCmp);
10087 if (Res.getNode()) {
10088 APInt KnownZero, KnownOne;
10089 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
10090 // Capture demanded bits information that would be otherwise lost.
10091 if (KnownZero == 0xfffffffe)
10092 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10093 DAG.getValueType(MVT::i1));
10094 else if (KnownZero == 0xffffff00)
10095 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10096 DAG.getValueType(MVT::i8));
10097 else if (KnownZero == 0xffff0000)
10098 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10099 DAG.getValueType(MVT::i16));
10105 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10106 DAGCombinerInfo &DCI) const {
10107 switch (N->getOpcode()) {
10109 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10110 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10111 case ISD::SUB: return PerformSUBCombine(N, DCI);
10112 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10113 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10114 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10115 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10116 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10117 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
10118 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10119 case ISD::STORE: return PerformSTORECombine(N, DCI);
10120 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
10121 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10122 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10123 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10124 case ISD::FP_TO_SINT:
10125 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10126 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10127 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10130 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10131 case ISD::SIGN_EXTEND:
10132 case ISD::ZERO_EXTEND:
10133 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10134 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10135 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10136 case ARMISD::VLD2DUP:
10137 case ARMISD::VLD3DUP:
10138 case ARMISD::VLD4DUP:
10139 return CombineBaseUpdate(N, DCI);
10140 case ARMISD::BUILD_VECTOR:
10141 return PerformARMBUILD_VECTORCombine(N, DCI);
10142 case ISD::INTRINSIC_VOID:
10143 case ISD::INTRINSIC_W_CHAIN:
10144 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10145 case Intrinsic::arm_neon_vld1:
10146 case Intrinsic::arm_neon_vld2:
10147 case Intrinsic::arm_neon_vld3:
10148 case Intrinsic::arm_neon_vld4:
10149 case Intrinsic::arm_neon_vld2lane:
10150 case Intrinsic::arm_neon_vld3lane:
10151 case Intrinsic::arm_neon_vld4lane:
10152 case Intrinsic::arm_neon_vst1:
10153 case Intrinsic::arm_neon_vst2:
10154 case Intrinsic::arm_neon_vst3:
10155 case Intrinsic::arm_neon_vst4:
10156 case Intrinsic::arm_neon_vst2lane:
10157 case Intrinsic::arm_neon_vst3lane:
10158 case Intrinsic::arm_neon_vst4lane:
10159 return CombineBaseUpdate(N, DCI);
10167 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10169 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10172 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
10173 bool *Fast) const {
10174 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10175 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10177 switch (VT.getSimpleVT().SimpleTy) {
10183 // Unaligned access can use (for example) LRDB, LRDH, LDR
10184 if (AllowsUnaligned) {
10186 *Fast = Subtarget->hasV7Ops();
10193 // For any little-endian targets with neon, we can support unaligned ld/st
10194 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10195 // A big-endian target may also explicitly support unaligned accesses
10196 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10206 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10207 unsigned AlignCheck) {
10208 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10209 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10212 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10213 unsigned DstAlign, unsigned SrcAlign,
10214 bool IsMemset, bool ZeroMemset,
10216 MachineFunction &MF) const {
10217 const Function *F = MF.getFunction();
10219 // See if we can use NEON instructions for this...
10220 if ((!IsMemset || ZeroMemset) &&
10221 Subtarget->hasNEON() &&
10222 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10223 Attribute::NoImplicitFloat)) {
10226 (memOpAlign(SrcAlign, DstAlign, 16) ||
10227 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
10229 } else if (Size >= 8 &&
10230 (memOpAlign(SrcAlign, DstAlign, 8) ||
10231 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
10236 // Lowering to i32/i16 if the size permits.
10239 else if (Size >= 2)
10242 // Let the target-independent logic figure it out.
10246 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10247 if (Val.getOpcode() != ISD::LOAD)
10250 EVT VT1 = Val.getValueType();
10251 if (!VT1.isSimple() || !VT1.isInteger() ||
10252 !VT2.isSimple() || !VT2.isInteger())
10255 switch (VT1.getSimpleVT().SimpleTy) {
10260 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10267 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10268 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10271 if (!isTypeLegal(EVT::getEVT(Ty1)))
10274 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10276 // Assuming the caller doesn't have a zeroext or signext return parameter,
10277 // truncation all the way down to i1 is valid.
10282 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10286 unsigned Scale = 1;
10287 switch (VT.getSimpleVT().SimpleTy) {
10288 default: return false;
10303 if ((V & (Scale - 1)) != 0)
10306 return V == (V & ((1LL << 5) - 1));
10309 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10310 const ARMSubtarget *Subtarget) {
10311 bool isNeg = false;
10317 switch (VT.getSimpleVT().SimpleTy) {
10318 default: return false;
10323 // + imm12 or - imm8
10325 return V == (V & ((1LL << 8) - 1));
10326 return V == (V & ((1LL << 12) - 1));
10329 // Same as ARM mode. FIXME: NEON?
10330 if (!Subtarget->hasVFP2())
10335 return V == (V & ((1LL << 8) - 1));
10339 /// isLegalAddressImmediate - Return true if the integer value can be used
10340 /// as the offset of the target addressing mode for load / store of the
10342 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10343 const ARMSubtarget *Subtarget) {
10347 if (!VT.isSimple())
10350 if (Subtarget->isThumb1Only())
10351 return isLegalT1AddressImmediate(V, VT);
10352 else if (Subtarget->isThumb2())
10353 return isLegalT2AddressImmediate(V, VT, Subtarget);
10358 switch (VT.getSimpleVT().SimpleTy) {
10359 default: return false;
10364 return V == (V & ((1LL << 12) - 1));
10367 return V == (V & ((1LL << 8) - 1));
10370 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10375 return V == (V & ((1LL << 8) - 1));
10379 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10381 int Scale = AM.Scale;
10385 switch (VT.getSimpleVT().SimpleTy) {
10386 default: return false;
10394 Scale = Scale & ~1;
10395 return Scale == 2 || Scale == 4 || Scale == 8;
10398 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10402 // Note, we allow "void" uses (basically, uses that aren't loads or
10403 // stores), because arm allows folding a scale into many arithmetic
10404 // operations. This should be made more precise and revisited later.
10406 // Allow r << imm, but the imm has to be a multiple of two.
10407 if (Scale & 1) return false;
10408 return isPowerOf2_32(Scale);
10412 /// isLegalAddressingMode - Return true if the addressing mode represented
10413 /// by AM is legal for this target, for a load/store of the specified type.
10414 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10416 EVT VT = getValueType(Ty, true);
10417 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10420 // Can never fold addr of global into load/store.
10424 switch (AM.Scale) {
10425 case 0: // no scale reg, must be "r+i" or "r", or "i".
10428 if (Subtarget->isThumb1Only())
10432 // ARM doesn't support any R+R*scale+imm addr modes.
10436 if (!VT.isSimple())
10439 if (Subtarget->isThumb2())
10440 return isLegalT2ScaledAddressingMode(AM, VT);
10442 int Scale = AM.Scale;
10443 switch (VT.getSimpleVT().SimpleTy) {
10444 default: return false;
10448 if (Scale < 0) Scale = -Scale;
10452 return isPowerOf2_32(Scale & ~1);
10456 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10461 // Note, we allow "void" uses (basically, uses that aren't loads or
10462 // stores), because arm allows folding a scale into many arithmetic
10463 // operations. This should be made more precise and revisited later.
10465 // Allow r << imm, but the imm has to be a multiple of two.
10466 if (Scale & 1) return false;
10467 return isPowerOf2_32(Scale);
10473 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10474 /// icmp immediate, that is the target has icmp instructions which can compare
10475 /// a register against the immediate without having to materialize the
10476 /// immediate into a register.
10477 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10478 // Thumb2 and ARM modes can use cmn for negative immediates.
10479 if (!Subtarget->isThumb())
10480 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10481 if (Subtarget->isThumb2())
10482 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10483 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10484 return Imm >= 0 && Imm <= 255;
10487 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10488 /// *or sub* immediate, that is the target has add or sub instructions which can
10489 /// add a register with the immediate without having to materialize the
10490 /// immediate into a register.
10491 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10492 // Same encoding for add/sub, just flip the sign.
10493 int64_t AbsImm = llvm::abs64(Imm);
10494 if (!Subtarget->isThumb())
10495 return ARM_AM::getSOImmVal(AbsImm) != -1;
10496 if (Subtarget->isThumb2())
10497 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10498 // Thumb1 only has 8-bit unsigned immediate.
10499 return AbsImm >= 0 && AbsImm <= 255;
10502 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10503 bool isSEXTLoad, SDValue &Base,
10504 SDValue &Offset, bool &isInc,
10505 SelectionDAG &DAG) {
10506 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10509 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10510 // AddressingMode 3
10511 Base = Ptr->getOperand(0);
10512 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10513 int RHSC = (int)RHS->getZExtValue();
10514 if (RHSC < 0 && RHSC > -256) {
10515 assert(Ptr->getOpcode() == ISD::ADD);
10517 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10521 isInc = (Ptr->getOpcode() == ISD::ADD);
10522 Offset = Ptr->getOperand(1);
10524 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10525 // AddressingMode 2
10526 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10527 int RHSC = (int)RHS->getZExtValue();
10528 if (RHSC < 0 && RHSC > -0x1000) {
10529 assert(Ptr->getOpcode() == ISD::ADD);
10531 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10532 Base = Ptr->getOperand(0);
10537 if (Ptr->getOpcode() == ISD::ADD) {
10539 ARM_AM::ShiftOpc ShOpcVal=
10540 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10541 if (ShOpcVal != ARM_AM::no_shift) {
10542 Base = Ptr->getOperand(1);
10543 Offset = Ptr->getOperand(0);
10545 Base = Ptr->getOperand(0);
10546 Offset = Ptr->getOperand(1);
10551 isInc = (Ptr->getOpcode() == ISD::ADD);
10552 Base = Ptr->getOperand(0);
10553 Offset = Ptr->getOperand(1);
10557 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10561 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10562 bool isSEXTLoad, SDValue &Base,
10563 SDValue &Offset, bool &isInc,
10564 SelectionDAG &DAG) {
10565 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10568 Base = Ptr->getOperand(0);
10569 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10570 int RHSC = (int)RHS->getZExtValue();
10571 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10572 assert(Ptr->getOpcode() == ISD::ADD);
10574 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10576 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10577 isInc = Ptr->getOpcode() == ISD::ADD;
10578 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10586 /// getPreIndexedAddressParts - returns true by value, base pointer and
10587 /// offset pointer and addressing mode by reference if the node's address
10588 /// can be legally represented as pre-indexed load / store address.
10590 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10592 ISD::MemIndexedMode &AM,
10593 SelectionDAG &DAG) const {
10594 if (Subtarget->isThumb1Only())
10599 bool isSEXTLoad = false;
10600 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10601 Ptr = LD->getBasePtr();
10602 VT = LD->getMemoryVT();
10603 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10604 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10605 Ptr = ST->getBasePtr();
10606 VT = ST->getMemoryVT();
10611 bool isLegal = false;
10612 if (Subtarget->isThumb2())
10613 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10614 Offset, isInc, DAG);
10616 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10617 Offset, isInc, DAG);
10621 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10625 /// getPostIndexedAddressParts - returns true by value, base pointer and
10626 /// offset pointer and addressing mode by reference if this node can be
10627 /// combined with a load / store to form a post-indexed load / store.
10628 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10631 ISD::MemIndexedMode &AM,
10632 SelectionDAG &DAG) const {
10633 if (Subtarget->isThumb1Only())
10638 bool isSEXTLoad = false;
10639 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10640 VT = LD->getMemoryVT();
10641 Ptr = LD->getBasePtr();
10642 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10643 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10644 VT = ST->getMemoryVT();
10645 Ptr = ST->getBasePtr();
10650 bool isLegal = false;
10651 if (Subtarget->isThumb2())
10652 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10655 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10661 // Swap base ptr and offset to catch more post-index load / store when
10662 // it's legal. In Thumb2 mode, offset must be an immediate.
10663 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10664 !Subtarget->isThumb2())
10665 std::swap(Base, Offset);
10667 // Post-indexed load / store update the base pointer.
10672 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10676 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10679 const SelectionDAG &DAG,
10680 unsigned Depth) const {
10681 unsigned BitWidth = KnownOne.getBitWidth();
10682 KnownZero = KnownOne = APInt(BitWidth, 0);
10683 switch (Op.getOpcode()) {
10689 // These nodes' second result is a boolean
10690 if (Op.getResNo() == 0)
10692 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10694 case ARMISD::CMOV: {
10695 // Bits are known zero/one if known on the LHS and RHS.
10696 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10697 if (KnownZero == 0 && KnownOne == 0) return;
10699 APInt KnownZeroRHS, KnownOneRHS;
10700 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10701 KnownZero &= KnownZeroRHS;
10702 KnownOne &= KnownOneRHS;
10708 //===----------------------------------------------------------------------===//
10709 // ARM Inline Assembly Support
10710 //===----------------------------------------------------------------------===//
10712 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10713 // Looking for "rev" which is V6+.
10714 if (!Subtarget->hasV6Ops())
10717 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10718 std::string AsmStr = IA->getAsmString();
10719 SmallVector<StringRef, 4> AsmPieces;
10720 SplitString(AsmStr, AsmPieces, ";\n");
10722 switch (AsmPieces.size()) {
10723 default: return false;
10725 AsmStr = AsmPieces[0];
10727 SplitString(AsmStr, AsmPieces, " \t,");
10730 if (AsmPieces.size() == 3 &&
10731 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10732 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10733 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10734 if (Ty && Ty->getBitWidth() == 32)
10735 return IntrinsicLowering::LowerToByteSwap(CI);
10743 /// getConstraintType - Given a constraint letter, return the type of
10744 /// constraint it is for this target.
10745 ARMTargetLowering::ConstraintType
10746 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10747 if (Constraint.size() == 1) {
10748 switch (Constraint[0]) {
10750 case 'l': return C_RegisterClass;
10751 case 'w': return C_RegisterClass;
10752 case 'h': return C_RegisterClass;
10753 case 'x': return C_RegisterClass;
10754 case 't': return C_RegisterClass;
10755 case 'j': return C_Other; // Constant for movw.
10756 // An address with a single base register. Due to the way we
10757 // currently handle addresses it is the same as an 'r' memory constraint.
10758 case 'Q': return C_Memory;
10760 } else if (Constraint.size() == 2) {
10761 switch (Constraint[0]) {
10763 // All 'U+' constraints are addresses.
10764 case 'U': return C_Memory;
10767 return TargetLowering::getConstraintType(Constraint);
10770 /// Examine constraint type and operand type and determine a weight value.
10771 /// This object must already have been set up with the operand type
10772 /// and the current alternative constraint selected.
10773 TargetLowering::ConstraintWeight
10774 ARMTargetLowering::getSingleConstraintMatchWeight(
10775 AsmOperandInfo &info, const char *constraint) const {
10776 ConstraintWeight weight = CW_Invalid;
10777 Value *CallOperandVal = info.CallOperandVal;
10778 // If we don't have a value, we can't do a match,
10779 // but allow it at the lowest weight.
10780 if (CallOperandVal == NULL)
10782 Type *type = CallOperandVal->getType();
10783 // Look at the constraint type.
10784 switch (*constraint) {
10786 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10789 if (type->isIntegerTy()) {
10790 if (Subtarget->isThumb())
10791 weight = CW_SpecificReg;
10793 weight = CW_Register;
10797 if (type->isFloatingPointTy())
10798 weight = CW_Register;
10804 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10806 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10808 if (Constraint.size() == 1) {
10809 // GCC ARM Constraint Letters
10810 switch (Constraint[0]) {
10811 case 'l': // Low regs or general regs.
10812 if (Subtarget->isThumb())
10813 return RCPair(0U, &ARM::tGPRRegClass);
10814 return RCPair(0U, &ARM::GPRRegClass);
10815 case 'h': // High regs or no regs.
10816 if (Subtarget->isThumb())
10817 return RCPair(0U, &ARM::hGPRRegClass);
10820 return RCPair(0U, &ARM::GPRRegClass);
10822 if (VT == MVT::Other)
10824 if (VT == MVT::f32)
10825 return RCPair(0U, &ARM::SPRRegClass);
10826 if (VT.getSizeInBits() == 64)
10827 return RCPair(0U, &ARM::DPRRegClass);
10828 if (VT.getSizeInBits() == 128)
10829 return RCPair(0U, &ARM::QPRRegClass);
10832 if (VT == MVT::Other)
10834 if (VT == MVT::f32)
10835 return RCPair(0U, &ARM::SPR_8RegClass);
10836 if (VT.getSizeInBits() == 64)
10837 return RCPair(0U, &ARM::DPR_8RegClass);
10838 if (VT.getSizeInBits() == 128)
10839 return RCPair(0U, &ARM::QPR_8RegClass);
10842 if (VT == MVT::f32)
10843 return RCPair(0U, &ARM::SPRRegClass);
10847 if (StringRef("{cc}").equals_lower(Constraint))
10848 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10850 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10853 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10854 /// vector. If it is invalid, don't add anything to Ops.
10855 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10856 std::string &Constraint,
10857 std::vector<SDValue>&Ops,
10858 SelectionDAG &DAG) const {
10859 SDValue Result(0, 0);
10861 // Currently only support length 1 constraints.
10862 if (Constraint.length() != 1) return;
10864 char ConstraintLetter = Constraint[0];
10865 switch (ConstraintLetter) {
10868 case 'I': case 'J': case 'K': case 'L':
10869 case 'M': case 'N': case 'O':
10870 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10874 int64_t CVal64 = C->getSExtValue();
10875 int CVal = (int) CVal64;
10876 // None of these constraints allow values larger than 32 bits. Check
10877 // that the value fits in an int.
10878 if (CVal != CVal64)
10881 switch (ConstraintLetter) {
10883 // Constant suitable for movw, must be between 0 and
10885 if (Subtarget->hasV6T2Ops())
10886 if (CVal >= 0 && CVal <= 65535)
10890 if (Subtarget->isThumb1Only()) {
10891 // This must be a constant between 0 and 255, for ADD
10893 if (CVal >= 0 && CVal <= 255)
10895 } else if (Subtarget->isThumb2()) {
10896 // A constant that can be used as an immediate value in a
10897 // data-processing instruction.
10898 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10901 // A constant that can be used as an immediate value in a
10902 // data-processing instruction.
10903 if (ARM_AM::getSOImmVal(CVal) != -1)
10909 if (Subtarget->isThumb()) { // FIXME thumb2
10910 // This must be a constant between -255 and -1, for negated ADD
10911 // immediates. This can be used in GCC with an "n" modifier that
10912 // prints the negated value, for use with SUB instructions. It is
10913 // not useful otherwise but is implemented for compatibility.
10914 if (CVal >= -255 && CVal <= -1)
10917 // This must be a constant between -4095 and 4095. It is not clear
10918 // what this constraint is intended for. Implemented for
10919 // compatibility with GCC.
10920 if (CVal >= -4095 && CVal <= 4095)
10926 if (Subtarget->isThumb1Only()) {
10927 // A 32-bit value where only one byte has a nonzero value. Exclude
10928 // zero to match GCC. This constraint is used by GCC internally for
10929 // constants that can be loaded with a move/shift combination.
10930 // It is not useful otherwise but is implemented for compatibility.
10931 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10933 } else if (Subtarget->isThumb2()) {
10934 // A constant whose bitwise inverse can be used as an immediate
10935 // value in a data-processing instruction. This can be used in GCC
10936 // with a "B" modifier that prints the inverted value, for use with
10937 // BIC and MVN instructions. It is not useful otherwise but is
10938 // implemented for compatibility.
10939 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10942 // A constant whose bitwise inverse can be used as an immediate
10943 // value in a data-processing instruction. This can be used in GCC
10944 // with a "B" modifier that prints the inverted value, for use with
10945 // BIC and MVN instructions. It is not useful otherwise but is
10946 // implemented for compatibility.
10947 if (ARM_AM::getSOImmVal(~CVal) != -1)
10953 if (Subtarget->isThumb1Only()) {
10954 // This must be a constant between -7 and 7,
10955 // for 3-operand ADD/SUB immediate instructions.
10956 if (CVal >= -7 && CVal < 7)
10958 } else if (Subtarget->isThumb2()) {
10959 // A constant whose negation can be used as an immediate value in a
10960 // data-processing instruction. This can be used in GCC with an "n"
10961 // modifier that prints the negated value, for use with SUB
10962 // instructions. It is not useful otherwise but is implemented for
10964 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10967 // A constant whose negation can be used as an immediate value in a
10968 // data-processing instruction. This can be used in GCC with an "n"
10969 // modifier that prints the negated value, for use with SUB
10970 // instructions. It is not useful otherwise but is implemented for
10972 if (ARM_AM::getSOImmVal(-CVal) != -1)
10978 if (Subtarget->isThumb()) { // FIXME thumb2
10979 // This must be a multiple of 4 between 0 and 1020, for
10980 // ADD sp + immediate.
10981 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10984 // A power of two or a constant between 0 and 32. This is used in
10985 // GCC for the shift amount on shifted register operands, but it is
10986 // useful in general for any shift amounts.
10987 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10993 if (Subtarget->isThumb()) { // FIXME thumb2
10994 // This must be a constant between 0 and 31, for shift amounts.
10995 if (CVal >= 0 && CVal <= 31)
11001 if (Subtarget->isThumb()) { // FIXME thumb2
11002 // This must be a multiple of 4 between -508 and 508, for
11003 // ADD/SUB sp = sp + immediate.
11004 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11009 Result = DAG.getTargetConstant(CVal, Op.getValueType());
11013 if (Result.getNode()) {
11014 Ops.push_back(Result);
11017 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11020 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11021 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
11022 unsigned Opcode = Op->getOpcode();
11023 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11024 "Invalid opcode for Div/Rem lowering");
11025 bool isSigned = (Opcode == ISD::SDIVREM);
11026 EVT VT = Op->getValueType(0);
11027 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11030 switch (VT.getSimpleVT().SimpleTy) {
11031 default: llvm_unreachable("Unexpected request for libcall!");
11032 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11033 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11034 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11035 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11038 SDValue InChain = DAG.getEntryNode();
11040 TargetLowering::ArgListTy Args;
11041 TargetLowering::ArgListEntry Entry;
11042 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
11043 EVT ArgVT = Op->getOperand(i).getValueType();
11044 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
11045 Entry.Node = Op->getOperand(i);
11047 Entry.isSExt = isSigned;
11048 Entry.isZExt = !isSigned;
11049 Args.push_back(Entry);
11052 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11055 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
11059 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
11060 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
11061 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
11062 Callee, Args, DAG, dl);
11063 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11065 return CallInfo.first;
11069 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11070 // The ARM target isn't yet aware of offsets.
11074 bool ARM::isBitFieldInvertedMask(unsigned v) {
11075 if (v == 0xffffffff)
11078 // there can be 1's on either or both "outsides", all the "inside"
11079 // bits must be 0's
11080 unsigned TO = CountTrailingOnes_32(v);
11081 unsigned LO = CountLeadingOnes_32(v);
11082 v = (v >> TO) << TO;
11083 v = (v << LO) >> LO;
11087 /// isFPImmLegal - Returns true if the target can instruction select the
11088 /// specified FP immediate natively. If false, the legalizer will
11089 /// materialize the FP immediate as a load from a constant pool.
11090 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11091 if (!Subtarget->hasVFP3())
11093 if (VT == MVT::f32)
11094 return ARM_AM::getFP32Imm(Imm) != -1;
11095 if (VT == MVT::f64)
11096 return ARM_AM::getFP64Imm(Imm) != -1;
11100 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11101 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11102 /// specified in the intrinsic calls.
11103 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11105 unsigned Intrinsic) const {
11106 switch (Intrinsic) {
11107 case Intrinsic::arm_neon_vld1:
11108 case Intrinsic::arm_neon_vld2:
11109 case Intrinsic::arm_neon_vld3:
11110 case Intrinsic::arm_neon_vld4:
11111 case Intrinsic::arm_neon_vld2lane:
11112 case Intrinsic::arm_neon_vld3lane:
11113 case Intrinsic::arm_neon_vld4lane: {
11114 Info.opc = ISD::INTRINSIC_W_CHAIN;
11115 // Conservatively set memVT to the entire set of vectors loaded.
11116 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11117 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11118 Info.ptrVal = I.getArgOperand(0);
11120 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11121 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11122 Info.vol = false; // volatile loads with NEON intrinsics not supported
11123 Info.readMem = true;
11124 Info.writeMem = false;
11127 case Intrinsic::arm_neon_vst1:
11128 case Intrinsic::arm_neon_vst2:
11129 case Intrinsic::arm_neon_vst3:
11130 case Intrinsic::arm_neon_vst4:
11131 case Intrinsic::arm_neon_vst2lane:
11132 case Intrinsic::arm_neon_vst3lane:
11133 case Intrinsic::arm_neon_vst4lane: {
11134 Info.opc = ISD::INTRINSIC_VOID;
11135 // Conservatively set memVT to the entire set of vectors stored.
11136 unsigned NumElts = 0;
11137 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11138 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11139 if (!ArgTy->isVectorTy())
11141 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11143 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11144 Info.ptrVal = I.getArgOperand(0);
11146 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11147 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11148 Info.vol = false; // volatile stores with NEON intrinsics not supported
11149 Info.readMem = false;
11150 Info.writeMem = true;
11153 case Intrinsic::arm_ldaex:
11154 case Intrinsic::arm_ldrex: {
11155 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11156 Info.opc = ISD::INTRINSIC_W_CHAIN;
11157 Info.memVT = MVT::getVT(PtrTy->getElementType());
11158 Info.ptrVal = I.getArgOperand(0);
11160 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11162 Info.readMem = true;
11163 Info.writeMem = false;
11166 case Intrinsic::arm_stlex:
11167 case Intrinsic::arm_strex: {
11168 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11169 Info.opc = ISD::INTRINSIC_W_CHAIN;
11170 Info.memVT = MVT::getVT(PtrTy->getElementType());
11171 Info.ptrVal = I.getArgOperand(1);
11173 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11175 Info.readMem = false;
11176 Info.writeMem = true;
11179 case Intrinsic::arm_stlexd:
11180 case Intrinsic::arm_strexd: {
11181 Info.opc = ISD::INTRINSIC_W_CHAIN;
11182 Info.memVT = MVT::i64;
11183 Info.ptrVal = I.getArgOperand(2);
11187 Info.readMem = false;
11188 Info.writeMem = true;
11191 case Intrinsic::arm_ldaexd:
11192 case Intrinsic::arm_ldrexd: {
11193 Info.opc = ISD::INTRINSIC_W_CHAIN;
11194 Info.memVT = MVT::i64;
11195 Info.ptrVal = I.getArgOperand(0);
11199 Info.readMem = true;
11200 Info.writeMem = false;
11210 /// \brief Returns true if it is beneficial to convert a load of a constant
11211 /// to just the constant itself.
11212 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11214 assert(Ty->isIntegerTy());
11216 unsigned Bits = Ty->getPrimitiveSizeInBits();
11217 if (Bits == 0 || Bits > 32)