1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Target/TargetOptions.h"
52 #define DEBUG_TYPE "arm-isel"
54 STATISTIC(NumTailCalls, "Number of tail calls");
55 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
56 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
59 EnableARMLongCalls("arm-long-calls", cl::Hidden,
60 cl::desc("Generate calls via indirect call instructions"),
64 ARMInterworking("arm-interworking", cl::Hidden,
65 cl::desc("Enable / disable ARM interworking (for debugging only)"),
69 class ARMCCState : public CCState {
71 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
72 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs,
73 LLVMContext &C, ParmContext PC)
74 : CCState(CC, isVarArg, MF, TM, locs, C) {
75 assert(((PC == Call) || (PC == Prologue)) &&
76 "ARMCCState users must specify whether their context is call"
77 "or prologue generation.");
83 // The APCS parameter registers.
84 static const MCPhysReg GPRArgRegs[] = {
85 ARM::R0, ARM::R1, ARM::R2, ARM::R3
88 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
89 MVT PromotedBitwiseVT) {
90 if (VT != PromotedLdStVT) {
91 setOperationAction(ISD::LOAD, VT, Promote);
92 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
94 setOperationAction(ISD::STORE, VT, Promote);
95 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
98 MVT ElemTy = VT.getVectorElementType();
99 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
100 setOperationAction(ISD::SETCC, VT, Custom);
101 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
102 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
103 if (ElemTy == MVT::i32) {
104 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
105 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
107 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
109 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
110 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
112 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
114 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
115 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
116 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
117 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
118 setOperationAction(ISD::SELECT, VT, Expand);
119 setOperationAction(ISD::SELECT_CC, VT, Expand);
120 setOperationAction(ISD::VSELECT, VT, Expand);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT, Custom);
124 setOperationAction(ISD::SRA, VT, Custom);
125 setOperationAction(ISD::SRL, VT, Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT, Promote);
131 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::OR, VT, Promote);
133 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::XOR, VT, Promote);
135 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT, Expand);
140 setOperationAction(ISD::UDIV, VT, Expand);
141 setOperationAction(ISD::FDIV, VT, Expand);
142 setOperationAction(ISD::SREM, VT, Expand);
143 setOperationAction(ISD::UREM, VT, Expand);
144 setOperationAction(ISD::FREM, VT, Expand);
147 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
148 addRegisterClass(VT, &ARM::DPRRegClass);
149 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
152 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPairRegClass);
154 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
157 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
158 if (TM.getSubtarget<ARMSubtarget>().isTargetMachO())
159 return new TargetLoweringObjectFileMachO();
161 return new ARMElfTargetObjectFile();
164 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
165 : TargetLowering(TM, createTLOF(TM)) {
166 Subtarget = &TM.getSubtarget<ARMSubtarget>();
167 RegInfo = TM.getRegisterInfo();
168 Itins = TM.getInstrItineraryData();
170 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
172 if (Subtarget->isTargetMachO()) {
173 // Uses VFP for Thumb libfuncs if available.
174 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
175 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
176 // Single-precision floating-point arithmetic.
177 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
178 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
179 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
180 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
182 // Double-precision floating-point arithmetic.
183 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
184 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
185 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
186 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
188 // Single-precision comparisons.
189 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
190 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
191 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
192 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
193 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
194 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
195 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
196 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
198 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
207 // Double-precision comparisons.
208 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
209 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
210 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
211 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
212 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
213 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
214 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
215 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
217 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
226 // Floating-point to integer conversions.
227 // i64 conversions are done via library routines even when generating VFP
228 // instructions, so use the same ones.
229 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
231 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
232 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
234 // Conversions between floating types.
235 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
236 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
238 // Integer to floating-point conversions.
239 // i64 conversions are done via library routines even when generating VFP
240 // instructions, so use the same ones.
241 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
242 // e.g., __floatunsidf vs. __floatunssidfvfp.
243 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
245 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
246 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
250 // These libcalls are not available in 32-bit.
251 setLibcallName(RTLIB::SHL_I128, 0);
252 setLibcallName(RTLIB::SRL_I128, 0);
253 setLibcallName(RTLIB::SRA_I128, 0);
255 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
256 !Subtarget->isTargetWindows()) {
257 // Double-precision floating-point arithmetic helper functions
258 // RTABI chapter 4.1.2, Table 2
259 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
260 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
261 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
262 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
263 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
264 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
265 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
266 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
268 // Double-precision floating-point comparison helper functions
269 // RTABI chapter 4.1.2, Table 3
270 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
271 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
272 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
273 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
274 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
275 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
276 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
277 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
278 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
279 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
280 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
281 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
282 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
283 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
284 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
285 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
286 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
287 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
288 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
295 // Single-precision floating-point arithmetic helper functions
296 // RTABI chapter 4.1.2, Table 4
297 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
298 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
299 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
300 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
301 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
302 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
303 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
304 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
306 // Single-precision floating-point comparison helper functions
307 // RTABI chapter 4.1.2, Table 5
308 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
309 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
310 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
311 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
312 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
313 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
314 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
315 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
316 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
317 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
318 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
319 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
320 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
321 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
322 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
323 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
324 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
325 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
326 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
333 // Floating-point to integer conversions.
334 // RTABI chapter 4.1.2, Table 6
335 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
336 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
337 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
338 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
339 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
340 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
341 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
342 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
343 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
344 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
345 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
346 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
352 // Conversions between floating types.
353 // RTABI chapter 4.1.2, Table 7
354 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
355 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
356 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
359 // Integer to floating-point conversions.
360 // RTABI chapter 4.1.2, Table 8
361 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
362 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
363 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
364 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
365 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
366 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
367 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
368 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
369 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
370 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
371 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
378 // Long long helper functions
379 // RTABI chapter 4.2, Table 9
380 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
381 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
382 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
383 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
384 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
391 // Integer division functions
392 // RTABI chapter 4.3.1
393 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
394 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
395 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
396 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
397 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
398 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
399 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
400 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
401 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
402 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
403 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
404 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
411 // RTABI chapter 4.3.4
412 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
413 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
414 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
415 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
416 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
417 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
420 // Use divmod compiler-rt calls for iOS 5.0 and later.
421 if (Subtarget->getTargetTriple().isiOS() &&
422 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
423 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
424 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
427 if (Subtarget->isThumb1Only())
428 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
430 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
431 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
432 !Subtarget->isThumb1Only()) {
433 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
434 if (!Subtarget->isFPOnlySP())
435 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
437 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
440 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
441 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
442 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
443 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
444 setTruncStoreAction((MVT::SimpleValueType)VT,
445 (MVT::SimpleValueType)InnerVT, Expand);
446 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
447 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
448 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
452 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
454 if (Subtarget->hasNEON()) {
455 addDRTypeForNEON(MVT::v2f32);
456 addDRTypeForNEON(MVT::v8i8);
457 addDRTypeForNEON(MVT::v4i16);
458 addDRTypeForNEON(MVT::v2i32);
459 addDRTypeForNEON(MVT::v1i64);
461 addQRTypeForNEON(MVT::v4f32);
462 addQRTypeForNEON(MVT::v2f64);
463 addQRTypeForNEON(MVT::v16i8);
464 addQRTypeForNEON(MVT::v8i16);
465 addQRTypeForNEON(MVT::v4i32);
466 addQRTypeForNEON(MVT::v2i64);
468 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
469 // neither Neon nor VFP support any arithmetic operations on it.
470 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
471 // supported for v4f32.
472 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
473 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
474 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
475 // FIXME: Code duplication: FDIV and FREM are expanded always, see
476 // ARMTargetLowering::addTypeForNEON method for details.
477 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
478 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
479 // FIXME: Create unittest.
480 // In another words, find a way when "copysign" appears in DAG with vector
482 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
483 // FIXME: Code duplication: SETCC has custom operation action, see
484 // ARMTargetLowering::addTypeForNEON method for details.
485 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
486 // FIXME: Create unittest for FNEG and for FABS.
487 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
488 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
489 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
490 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
491 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
492 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
493 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
494 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
495 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
497 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
498 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
499 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
500 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
501 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
502 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
503 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
504 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
505 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
507 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
508 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
510 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
512 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
515 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
517 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
518 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
519 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
520 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
521 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
523 // Mark v2f32 intrinsics.
524 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
525 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
526 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
527 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
528 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
529 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
530 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
531 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
532 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
533 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
534 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
535 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
536 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
537 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
538 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
540 // Neon does not support some operations on v1i64 and v2i64 types.
541 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
542 // Custom handling for some quad-vector types to detect VMULL.
543 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
544 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
545 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
546 // Custom handling for some vector types to avoid expensive expansions
547 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
548 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
549 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
550 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
551 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
552 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
553 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
554 // a destination type that is wider than the source, and nor does
555 // it have a FP_TO_[SU]INT instruction with a narrower destination than
557 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
558 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
559 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
560 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
563 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
565 // NEON does not have single instruction CTPOP for vectors with element
566 // types wider than 8-bits. However, custom lowering can leverage the
567 // v8i8/v16i8 vcnt instruction.
568 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
569 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
570 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
571 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
573 // NEON only has FMA instructions as of VFP4.
574 if (!Subtarget->hasVFP4()) {
575 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
576 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
579 setTargetDAGCombine(ISD::INTRINSIC_VOID);
580 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
581 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
582 setTargetDAGCombine(ISD::SHL);
583 setTargetDAGCombine(ISD::SRL);
584 setTargetDAGCombine(ISD::SRA);
585 setTargetDAGCombine(ISD::SIGN_EXTEND);
586 setTargetDAGCombine(ISD::ZERO_EXTEND);
587 setTargetDAGCombine(ISD::ANY_EXTEND);
588 setTargetDAGCombine(ISD::SELECT_CC);
589 setTargetDAGCombine(ISD::BUILD_VECTOR);
590 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
591 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
592 setTargetDAGCombine(ISD::STORE);
593 setTargetDAGCombine(ISD::FP_TO_SINT);
594 setTargetDAGCombine(ISD::FP_TO_UINT);
595 setTargetDAGCombine(ISD::FDIV);
597 // It is legal to extload from v4i8 to v4i16 or v4i32.
598 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
599 MVT::v4i16, MVT::v2i16,
601 for (unsigned i = 0; i < 6; ++i) {
602 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
603 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
604 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
608 // ARM and Thumb2 support UMLAL/SMLAL.
609 if (!Subtarget->isThumb1Only())
610 setTargetDAGCombine(ISD::ADDC);
613 computeRegisterProperties();
615 // ARM does not have f32 extending load.
616 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
618 // ARM does not have i1 sign extending load.
619 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
621 // ARM supports all 4 flavors of integer indexed load / store.
622 if (!Subtarget->isThumb1Only()) {
623 for (unsigned im = (unsigned)ISD::PRE_INC;
624 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
625 setIndexedLoadAction(im, MVT::i1, Legal);
626 setIndexedLoadAction(im, MVT::i8, Legal);
627 setIndexedLoadAction(im, MVT::i16, Legal);
628 setIndexedLoadAction(im, MVT::i32, Legal);
629 setIndexedStoreAction(im, MVT::i1, Legal);
630 setIndexedStoreAction(im, MVT::i8, Legal);
631 setIndexedStoreAction(im, MVT::i16, Legal);
632 setIndexedStoreAction(im, MVT::i32, Legal);
636 // i64 operation support.
637 setOperationAction(ISD::MUL, MVT::i64, Expand);
638 setOperationAction(ISD::MULHU, MVT::i32, Expand);
639 if (Subtarget->isThumb1Only()) {
640 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
641 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
643 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
644 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
645 setOperationAction(ISD::MULHS, MVT::i32, Expand);
647 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
648 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
649 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
650 setOperationAction(ISD::SRL, MVT::i64, Custom);
651 setOperationAction(ISD::SRA, MVT::i64, Custom);
653 if (!Subtarget->isThumb1Only()) {
654 // FIXME: We should do this for Thumb1 as well.
655 setOperationAction(ISD::ADDC, MVT::i32, Custom);
656 setOperationAction(ISD::ADDE, MVT::i32, Custom);
657 setOperationAction(ISD::SUBC, MVT::i32, Custom);
658 setOperationAction(ISD::SUBE, MVT::i32, Custom);
661 // ARM does not have ROTL.
662 setOperationAction(ISD::ROTL, MVT::i32, Expand);
663 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
664 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
665 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
666 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
668 // These just redirect to CTTZ and CTLZ on ARM.
669 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
670 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
672 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
674 // Only ARMv6 has BSWAP.
675 if (!Subtarget->hasV6Ops())
676 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
678 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
679 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
680 // These are expanded into libcalls if the cpu doesn't have HW divider.
681 setOperationAction(ISD::SDIV, MVT::i32, Expand);
682 setOperationAction(ISD::UDIV, MVT::i32, Expand);
685 // FIXME: Also set divmod for SREM on EABI
686 setOperationAction(ISD::SREM, MVT::i32, Expand);
687 setOperationAction(ISD::UREM, MVT::i32, Expand);
688 // Register based DivRem for AEABI (RTABI 4.2)
689 if (Subtarget->isTargetAEABI()) {
690 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
691 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
692 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
693 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
694 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
695 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
696 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
697 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
699 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
700 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
701 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
702 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
703 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
704 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
705 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
706 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
708 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
709 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
711 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
712 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
715 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
716 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
717 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
718 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
719 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
721 setOperationAction(ISD::TRAP, MVT::Other, Legal);
723 // Use the default implementation.
724 setOperationAction(ISD::VASTART, MVT::Other, Custom);
725 setOperationAction(ISD::VAARG, MVT::Other, Expand);
726 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
727 setOperationAction(ISD::VAEND, MVT::Other, Expand);
728 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
729 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
731 if (!Subtarget->isTargetMachO()) {
732 // Non-MachO platforms may return values in these registers via the
733 // personality function.
734 setExceptionPointerRegister(ARM::R0);
735 setExceptionSelectorRegister(ARM::R1);
738 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
739 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
740 // the default expansion.
741 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
742 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
743 // to ldrex/strex loops already.
744 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
746 // On v8, we have particularly efficient implementations of atomic fences
747 // if they can be combined with nearby atomic loads and stores.
748 if (!Subtarget->hasV8Ops()) {
749 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
750 setInsertFencesForAtomic(true);
753 // If there's anything we can use as a barrier, go through custom lowering
755 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
756 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
758 // Set them all for expansion, which will force libcalls.
759 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
760 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
761 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
762 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
763 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
764 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
765 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
766 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
767 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
768 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
769 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
770 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
771 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
772 // Unordered/Monotonic case.
773 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
774 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
777 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
779 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
780 if (!Subtarget->hasV6Ops()) {
781 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
782 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
784 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
786 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
787 !Subtarget->isThumb1Only()) {
788 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
789 // iff target supports vfp2.
790 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
791 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
794 // We want to custom lower some of our intrinsics.
795 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
796 if (Subtarget->isTargetDarwin()) {
797 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
798 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
799 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
802 setOperationAction(ISD::SETCC, MVT::i32, Expand);
803 setOperationAction(ISD::SETCC, MVT::f32, Expand);
804 setOperationAction(ISD::SETCC, MVT::f64, Expand);
805 setOperationAction(ISD::SELECT, MVT::i32, Custom);
806 setOperationAction(ISD::SELECT, MVT::f32, Custom);
807 setOperationAction(ISD::SELECT, MVT::f64, Custom);
808 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
809 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
810 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
812 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
813 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
814 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
815 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
816 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
818 // We don't support sin/cos/fmod/copysign/pow
819 setOperationAction(ISD::FSIN, MVT::f64, Expand);
820 setOperationAction(ISD::FSIN, MVT::f32, Expand);
821 setOperationAction(ISD::FCOS, MVT::f32, Expand);
822 setOperationAction(ISD::FCOS, MVT::f64, Expand);
823 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
824 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
825 setOperationAction(ISD::FREM, MVT::f64, Expand);
826 setOperationAction(ISD::FREM, MVT::f32, Expand);
827 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
828 !Subtarget->isThumb1Only()) {
829 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
830 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
832 setOperationAction(ISD::FPOW, MVT::f64, Expand);
833 setOperationAction(ISD::FPOW, MVT::f32, Expand);
835 if (!Subtarget->hasVFP4()) {
836 setOperationAction(ISD::FMA, MVT::f64, Expand);
837 setOperationAction(ISD::FMA, MVT::f32, Expand);
840 // Various VFP goodness
841 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
842 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
843 if (Subtarget->hasVFP2()) {
844 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
845 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
846 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
847 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
849 // Special handling for half-precision FP.
850 if (!Subtarget->hasFP16()) {
851 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
852 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
856 // Combine sin / cos into one node or libcall if possible.
857 if (Subtarget->hasSinCos()) {
858 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
859 setLibcallName(RTLIB::SINCOS_F64, "sincos");
860 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
861 // For iOS, we don't want to the normal expansion of a libcall to
862 // sincos. We want to issue a libcall to __sincos_stret.
863 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
864 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
868 // We have target-specific dag combine patterns for the following nodes:
869 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
870 setTargetDAGCombine(ISD::ADD);
871 setTargetDAGCombine(ISD::SUB);
872 setTargetDAGCombine(ISD::MUL);
873 setTargetDAGCombine(ISD::AND);
874 setTargetDAGCombine(ISD::OR);
875 setTargetDAGCombine(ISD::XOR);
877 if (Subtarget->hasV6Ops())
878 setTargetDAGCombine(ISD::SRL);
880 setStackPointerRegisterToSaveRestore(ARM::SP);
882 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
883 !Subtarget->hasVFP2())
884 setSchedulingPreference(Sched::RegPressure);
886 setSchedulingPreference(Sched::Hybrid);
888 //// temporary - rewrite interface to use type
889 MaxStoresPerMemset = 8;
890 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
891 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
892 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
893 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
894 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
896 // On ARM arguments smaller than 4 bytes are extended, so all arguments
897 // are at least 4 bytes aligned.
898 setMinStackArgumentAlignment(4);
900 // Prefer likely predicted branches to selects on out-of-order cores.
901 PredictableSelectIsExpensive = Subtarget->isLikeA9();
903 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
906 // FIXME: It might make sense to define the representative register class as the
907 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
908 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
909 // SPR's representative would be DPR_VFP2. This should work well if register
910 // pressure tracking were modified such that a register use would increment the
911 // pressure of the register class's representative and all of it's super
912 // classes' representatives transitively. We have not implemented this because
913 // of the difficulty prior to coalescing of modeling operand register classes
914 // due to the common occurrence of cross class copies and subregister insertions
916 std::pair<const TargetRegisterClass*, uint8_t>
917 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
918 const TargetRegisterClass *RRC = 0;
920 switch (VT.SimpleTy) {
922 return TargetLowering::findRepresentativeClass(VT);
923 // Use DPR as representative register class for all floating point
924 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
925 // the cost is 1 for both f32 and f64.
926 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
927 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
928 RRC = &ARM::DPRRegClass;
929 // When NEON is used for SP, only half of the register file is available
930 // because operations that define both SP and DP results will be constrained
931 // to the VFP2 class (D0-D15). We currently model this constraint prior to
932 // coalescing by double-counting the SP regs. See the FIXME above.
933 if (Subtarget->useNEONForSinglePrecisionFP())
936 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
937 case MVT::v4f32: case MVT::v2f64:
938 RRC = &ARM::DPRRegClass;
942 RRC = &ARM::DPRRegClass;
946 RRC = &ARM::DPRRegClass;
950 return std::make_pair(RRC, Cost);
953 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
956 case ARMISD::Wrapper: return "ARMISD::Wrapper";
957 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
958 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
959 case ARMISD::CALL: return "ARMISD::CALL";
960 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
961 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
962 case ARMISD::tCALL: return "ARMISD::tCALL";
963 case ARMISD::BRCOND: return "ARMISD::BRCOND";
964 case ARMISD::BR_JT: return "ARMISD::BR_JT";
965 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
966 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
967 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
968 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
969 case ARMISD::CMP: return "ARMISD::CMP";
970 case ARMISD::CMN: return "ARMISD::CMN";
971 case ARMISD::CMPZ: return "ARMISD::CMPZ";
972 case ARMISD::CMPFP: return "ARMISD::CMPFP";
973 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
974 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
975 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
977 case ARMISD::CMOV: return "ARMISD::CMOV";
979 case ARMISD::RBIT: return "ARMISD::RBIT";
981 case ARMISD::FTOSI: return "ARMISD::FTOSI";
982 case ARMISD::FTOUI: return "ARMISD::FTOUI";
983 case ARMISD::SITOF: return "ARMISD::SITOF";
984 case ARMISD::UITOF: return "ARMISD::UITOF";
986 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
987 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
988 case ARMISD::RRX: return "ARMISD::RRX";
990 case ARMISD::ADDC: return "ARMISD::ADDC";
991 case ARMISD::ADDE: return "ARMISD::ADDE";
992 case ARMISD::SUBC: return "ARMISD::SUBC";
993 case ARMISD::SUBE: return "ARMISD::SUBE";
995 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
996 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
998 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
999 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1001 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1003 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1005 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1007 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1009 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1011 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1012 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1013 case ARMISD::VCGE: return "ARMISD::VCGE";
1014 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1015 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1016 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1017 case ARMISD::VCGT: return "ARMISD::VCGT";
1018 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1019 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1020 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1021 case ARMISD::VTST: return "ARMISD::VTST";
1023 case ARMISD::VSHL: return "ARMISD::VSHL";
1024 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1025 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1026 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1027 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1028 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1029 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1030 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1031 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1032 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1033 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1034 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1035 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1036 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1037 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1038 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1039 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1040 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1041 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1042 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1043 case ARMISD::VDUP: return "ARMISD::VDUP";
1044 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1045 case ARMISD::VEXT: return "ARMISD::VEXT";
1046 case ARMISD::VREV64: return "ARMISD::VREV64";
1047 case ARMISD::VREV32: return "ARMISD::VREV32";
1048 case ARMISD::VREV16: return "ARMISD::VREV16";
1049 case ARMISD::VZIP: return "ARMISD::VZIP";
1050 case ARMISD::VUZP: return "ARMISD::VUZP";
1051 case ARMISD::VTRN: return "ARMISD::VTRN";
1052 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1053 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1054 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1055 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1056 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1057 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1058 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1059 case ARMISD::FMAX: return "ARMISD::FMAX";
1060 case ARMISD::FMIN: return "ARMISD::FMIN";
1061 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1062 case ARMISD::VMINNM: return "ARMISD::VMIN";
1063 case ARMISD::BFI: return "ARMISD::BFI";
1064 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1065 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1066 case ARMISD::VBSL: return "ARMISD::VBSL";
1067 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1068 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1069 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1070 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1071 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1072 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1073 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1074 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1075 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1076 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1077 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1078 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1079 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1080 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1081 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1082 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1083 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1084 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1085 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1086 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1090 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1091 if (!VT.isVector()) return getPointerTy();
1092 return VT.changeVectorElementTypeToInteger();
1095 /// getRegClassFor - Return the register class that should be used for the
1096 /// specified value type.
1097 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1098 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1099 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1100 // load / store 4 to 8 consecutive D registers.
1101 if (Subtarget->hasNEON()) {
1102 if (VT == MVT::v4i64)
1103 return &ARM::QQPRRegClass;
1104 if (VT == MVT::v8i64)
1105 return &ARM::QQQQPRRegClass;
1107 return TargetLowering::getRegClassFor(VT);
1110 // Create a fast isel object.
1112 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1113 const TargetLibraryInfo *libInfo) const {
1114 return ARM::createFastISel(funcInfo, libInfo);
1117 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1118 /// be used for loads / stores from the global.
1119 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1120 return (Subtarget->isThumb1Only() ? 127 : 4095);
1123 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1124 unsigned NumVals = N->getNumValues();
1126 return Sched::RegPressure;
1128 for (unsigned i = 0; i != NumVals; ++i) {
1129 EVT VT = N->getValueType(i);
1130 if (VT == MVT::Glue || VT == MVT::Other)
1132 if (VT.isFloatingPoint() || VT.isVector())
1136 if (!N->isMachineOpcode())
1137 return Sched::RegPressure;
1139 // Load are scheduled for latency even if there instruction itinerary
1140 // is not available.
1141 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1142 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1144 if (MCID.getNumDefs() == 0)
1145 return Sched::RegPressure;
1146 if (!Itins->isEmpty() &&
1147 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1150 return Sched::RegPressure;
1153 //===----------------------------------------------------------------------===//
1155 //===----------------------------------------------------------------------===//
1157 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1158 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1160 default: llvm_unreachable("Unknown condition code!");
1161 case ISD::SETNE: return ARMCC::NE;
1162 case ISD::SETEQ: return ARMCC::EQ;
1163 case ISD::SETGT: return ARMCC::GT;
1164 case ISD::SETGE: return ARMCC::GE;
1165 case ISD::SETLT: return ARMCC::LT;
1166 case ISD::SETLE: return ARMCC::LE;
1167 case ISD::SETUGT: return ARMCC::HI;
1168 case ISD::SETUGE: return ARMCC::HS;
1169 case ISD::SETULT: return ARMCC::LO;
1170 case ISD::SETULE: return ARMCC::LS;
1174 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1175 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1176 ARMCC::CondCodes &CondCode2) {
1177 CondCode2 = ARMCC::AL;
1179 default: llvm_unreachable("Unknown FP condition!");
1181 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1183 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1185 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1186 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1187 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1188 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1189 case ISD::SETO: CondCode = ARMCC::VC; break;
1190 case ISD::SETUO: CondCode = ARMCC::VS; break;
1191 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1192 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1193 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1195 case ISD::SETULT: CondCode = ARMCC::LT; break;
1197 case ISD::SETULE: CondCode = ARMCC::LE; break;
1199 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1203 //===----------------------------------------------------------------------===//
1204 // Calling Convention Implementation
1205 //===----------------------------------------------------------------------===//
1207 #include "ARMGenCallingConv.inc"
1209 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1210 /// given CallingConvention value.
1211 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1213 bool isVarArg) const {
1216 llvm_unreachable("Unsupported calling convention");
1217 case CallingConv::Fast:
1218 if (Subtarget->hasVFP2() && !isVarArg) {
1219 if (!Subtarget->isAAPCS_ABI())
1220 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1221 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1222 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1225 case CallingConv::C: {
1226 // Use target triple & subtarget features to do actual dispatch.
1227 if (!Subtarget->isAAPCS_ABI())
1228 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1229 else if (Subtarget->hasVFP2() &&
1230 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1232 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1233 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1235 case CallingConv::ARM_AAPCS_VFP:
1237 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1239 case CallingConv::ARM_AAPCS:
1240 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1241 case CallingConv::ARM_APCS:
1242 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1243 case CallingConv::GHC:
1244 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1248 /// LowerCallResult - Lower the result values of a call into the
1249 /// appropriate copies out of appropriate physical registers.
1251 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1252 CallingConv::ID CallConv, bool isVarArg,
1253 const SmallVectorImpl<ISD::InputArg> &Ins,
1254 SDLoc dl, SelectionDAG &DAG,
1255 SmallVectorImpl<SDValue> &InVals,
1256 bool isThisReturn, SDValue ThisVal) const {
1258 // Assign locations to each value returned by this call.
1259 SmallVector<CCValAssign, 16> RVLocs;
1260 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1261 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1262 CCInfo.AnalyzeCallResult(Ins,
1263 CCAssignFnForNode(CallConv, /* Return*/ true,
1266 // Copy all of the result registers out of their specified physreg.
1267 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1268 CCValAssign VA = RVLocs[i];
1270 // Pass 'this' value directly from the argument to return value, to avoid
1271 // reg unit interference
1272 if (i == 0 && isThisReturn) {
1273 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1274 "unexpected return calling convention register assignment");
1275 InVals.push_back(ThisVal);
1280 if (VA.needsCustom()) {
1281 // Handle f64 or half of a v2f64.
1282 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1284 Chain = Lo.getValue(1);
1285 InFlag = Lo.getValue(2);
1286 VA = RVLocs[++i]; // skip ahead to next loc
1287 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1289 Chain = Hi.getValue(1);
1290 InFlag = Hi.getValue(2);
1291 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1293 if (VA.getLocVT() == MVT::v2f64) {
1294 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1295 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1296 DAG.getConstant(0, MVT::i32));
1298 VA = RVLocs[++i]; // skip ahead to next loc
1299 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1300 Chain = Lo.getValue(1);
1301 InFlag = Lo.getValue(2);
1302 VA = RVLocs[++i]; // skip ahead to next loc
1303 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1304 Chain = Hi.getValue(1);
1305 InFlag = Hi.getValue(2);
1306 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1307 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1308 DAG.getConstant(1, MVT::i32));
1311 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1313 Chain = Val.getValue(1);
1314 InFlag = Val.getValue(2);
1317 switch (VA.getLocInfo()) {
1318 default: llvm_unreachable("Unknown loc info!");
1319 case CCValAssign::Full: break;
1320 case CCValAssign::BCvt:
1321 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1325 InVals.push_back(Val);
1331 /// LowerMemOpCallTo - Store the argument to the stack.
1333 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1334 SDValue StackPtr, SDValue Arg,
1335 SDLoc dl, SelectionDAG &DAG,
1336 const CCValAssign &VA,
1337 ISD::ArgFlagsTy Flags) const {
1338 unsigned LocMemOffset = VA.getLocMemOffset();
1339 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1340 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1341 return DAG.getStore(Chain, dl, Arg, PtrOff,
1342 MachinePointerInfo::getStack(LocMemOffset),
1346 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1347 SDValue Chain, SDValue &Arg,
1348 RegsToPassVector &RegsToPass,
1349 CCValAssign &VA, CCValAssign &NextVA,
1351 SmallVectorImpl<SDValue> &MemOpChains,
1352 ISD::ArgFlagsTy Flags) const {
1354 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1355 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1356 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1358 if (NextVA.isRegLoc())
1359 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1361 assert(NextVA.isMemLoc());
1362 if (StackPtr.getNode() == 0)
1363 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1365 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1371 /// LowerCall - Lowering a call into a callseq_start <-
1372 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1375 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1376 SmallVectorImpl<SDValue> &InVals) const {
1377 SelectionDAG &DAG = CLI.DAG;
1379 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1380 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1381 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1382 SDValue Chain = CLI.Chain;
1383 SDValue Callee = CLI.Callee;
1384 bool &isTailCall = CLI.IsTailCall;
1385 CallingConv::ID CallConv = CLI.CallConv;
1386 bool doesNotRet = CLI.DoesNotReturn;
1387 bool isVarArg = CLI.IsVarArg;
1389 MachineFunction &MF = DAG.getMachineFunction();
1390 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1391 bool isThisReturn = false;
1392 bool isSibCall = false;
1394 // Disable tail calls if they're not supported.
1395 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1399 // Check if it's really possible to do a tail call.
1400 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1401 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1402 Outs, OutVals, Ins, DAG);
1403 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1404 report_fatal_error("failed to perform tail call elimination on a call "
1405 "site marked musttail");
1406 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1407 // detected sibcalls.
1414 // Analyze operands of the call, assigning locations to each operand.
1415 SmallVector<CCValAssign, 16> ArgLocs;
1416 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1417 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1418 CCInfo.AnalyzeCallOperands(Outs,
1419 CCAssignFnForNode(CallConv, /* Return*/ false,
1422 // Get a count of how many bytes are to be pushed on the stack.
1423 unsigned NumBytes = CCInfo.getNextStackOffset();
1425 // For tail calls, memory operands are available in our caller's stack.
1429 // Adjust the stack pointer for the new arguments...
1430 // These operations are automatically eliminated by the prolog/epilog pass
1432 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1435 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1437 RegsToPassVector RegsToPass;
1438 SmallVector<SDValue, 8> MemOpChains;
1440 // Walk the register/memloc assignments, inserting copies/loads. In the case
1441 // of tail call optimization, arguments are handled later.
1442 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1444 ++i, ++realArgIdx) {
1445 CCValAssign &VA = ArgLocs[i];
1446 SDValue Arg = OutVals[realArgIdx];
1447 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1448 bool isByVal = Flags.isByVal();
1450 // Promote the value if needed.
1451 switch (VA.getLocInfo()) {
1452 default: llvm_unreachable("Unknown loc info!");
1453 case CCValAssign::Full: break;
1454 case CCValAssign::SExt:
1455 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1457 case CCValAssign::ZExt:
1458 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1460 case CCValAssign::AExt:
1461 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1463 case CCValAssign::BCvt:
1464 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1468 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1469 if (VA.needsCustom()) {
1470 if (VA.getLocVT() == MVT::v2f64) {
1471 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1472 DAG.getConstant(0, MVT::i32));
1473 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1474 DAG.getConstant(1, MVT::i32));
1476 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1477 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1479 VA = ArgLocs[++i]; // skip ahead to next loc
1480 if (VA.isRegLoc()) {
1481 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1482 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1484 assert(VA.isMemLoc());
1486 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1487 dl, DAG, VA, Flags));
1490 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1491 StackPtr, MemOpChains, Flags);
1493 } else if (VA.isRegLoc()) {
1494 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1495 assert(VA.getLocVT() == MVT::i32 &&
1496 "unexpected calling convention register assignment");
1497 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1498 "unexpected use of 'returned'");
1499 isThisReturn = true;
1501 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1502 } else if (isByVal) {
1503 assert(VA.isMemLoc());
1504 unsigned offset = 0;
1506 // True if this byval aggregate will be split between registers
1508 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1509 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1511 if (CurByValIdx < ByValArgsCount) {
1513 unsigned RegBegin, RegEnd;
1514 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1516 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1518 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1519 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1520 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1521 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1522 MachinePointerInfo(),
1523 false, false, false,
1524 DAG.InferPtrAlignment(AddArg));
1525 MemOpChains.push_back(Load.getValue(1));
1526 RegsToPass.push_back(std::make_pair(j, Load));
1529 // If parameter size outsides register area, "offset" value
1530 // helps us to calculate stack slot for remained part properly.
1531 offset = RegEnd - RegBegin;
1533 CCInfo.nextInRegsParam();
1536 if (Flags.getByValSize() > 4*offset) {
1537 unsigned LocMemOffset = VA.getLocMemOffset();
1538 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1539 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1541 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1542 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1543 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1545 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1547 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1548 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1549 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1550 Ops, array_lengthof(Ops)));
1552 } else if (!isSibCall) {
1553 assert(VA.isMemLoc());
1555 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1556 dl, DAG, VA, Flags));
1560 if (!MemOpChains.empty())
1561 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1562 &MemOpChains[0], MemOpChains.size());
1564 // Build a sequence of copy-to-reg nodes chained together with token chain
1565 // and flag operands which copy the outgoing args into the appropriate regs.
1567 // Tail call byval lowering might overwrite argument registers so in case of
1568 // tail call optimization the copies to registers are lowered later.
1570 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1571 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1572 RegsToPass[i].second, InFlag);
1573 InFlag = Chain.getValue(1);
1576 // For tail calls lower the arguments to the 'real' stack slot.
1578 // Force all the incoming stack arguments to be loaded from the stack
1579 // before any new outgoing arguments are stored to the stack, because the
1580 // outgoing stack slots may alias the incoming argument stack slots, and
1581 // the alias isn't otherwise explicit. This is slightly more conservative
1582 // than necessary, because it means that each store effectively depends
1583 // on every argument instead of just those arguments it would clobber.
1585 // Do not flag preceding copytoreg stuff together with the following stuff.
1587 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1588 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1589 RegsToPass[i].second, InFlag);
1590 InFlag = Chain.getValue(1);
1595 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1596 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1597 // node so that legalize doesn't hack it.
1598 bool isDirect = false;
1599 bool isARMFunc = false;
1600 bool isLocalARMFunc = false;
1601 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1603 if (EnableARMLongCalls) {
1604 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1605 && "long-calls with non-static relocation model!");
1606 // Handle a global address or an external symbol. If it's not one of
1607 // those, the target's already in a register, so we don't need to do
1609 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1610 const GlobalValue *GV = G->getGlobal();
1611 // Create a constant pool entry for the callee address
1612 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1613 ARMConstantPoolValue *CPV =
1614 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1616 // Get the address of the callee into a register
1617 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1618 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1619 Callee = DAG.getLoad(getPointerTy(), dl,
1620 DAG.getEntryNode(), CPAddr,
1621 MachinePointerInfo::getConstantPool(),
1622 false, false, false, 0);
1623 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1624 const char *Sym = S->getSymbol();
1626 // Create a constant pool entry for the callee address
1627 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1628 ARMConstantPoolValue *CPV =
1629 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1630 ARMPCLabelIndex, 0);
1631 // Get the address of the callee into a register
1632 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1633 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1634 Callee = DAG.getLoad(getPointerTy(), dl,
1635 DAG.getEntryNode(), CPAddr,
1636 MachinePointerInfo::getConstantPool(),
1637 false, false, false, 0);
1639 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1640 const GlobalValue *GV = G->getGlobal();
1642 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1643 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1644 getTargetMachine().getRelocationModel() != Reloc::Static;
1645 isARMFunc = !Subtarget->isThumb() || isStub;
1646 // ARM call to a local ARM function is predicable.
1647 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1648 // tBX takes a register source operand.
1649 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1650 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1651 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1652 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1654 // On ELF targets for PIC code, direct calls should go through the PLT
1655 unsigned OpFlags = 0;
1656 if (Subtarget->isTargetELF() &&
1657 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1658 OpFlags = ARMII::MO_PLT;
1659 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1661 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1663 bool isStub = Subtarget->isTargetMachO() &&
1664 getTargetMachine().getRelocationModel() != Reloc::Static;
1665 isARMFunc = !Subtarget->isThumb() || isStub;
1666 // tBX takes a register source operand.
1667 const char *Sym = S->getSymbol();
1668 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1669 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1670 ARMConstantPoolValue *CPV =
1671 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1672 ARMPCLabelIndex, 4);
1673 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1674 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1675 Callee = DAG.getLoad(getPointerTy(), dl,
1676 DAG.getEntryNode(), CPAddr,
1677 MachinePointerInfo::getConstantPool(),
1678 false, false, false, 0);
1679 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1680 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1681 getPointerTy(), Callee, PICLabel);
1683 unsigned OpFlags = 0;
1684 // On ELF targets for PIC code, direct calls should go through the PLT
1685 if (Subtarget->isTargetELF() &&
1686 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1687 OpFlags = ARMII::MO_PLT;
1688 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1692 // FIXME: handle tail calls differently.
1694 bool HasMinSizeAttr = Subtarget->isMinSize();
1695 if (Subtarget->isThumb()) {
1696 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1697 CallOpc = ARMISD::CALL_NOLINK;
1699 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1701 if (!isDirect && !Subtarget->hasV5TOps())
1702 CallOpc = ARMISD::CALL_NOLINK;
1703 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1704 // Emit regular call when code size is the priority
1706 // "mov lr, pc; b _foo" to avoid confusing the RSP
1707 CallOpc = ARMISD::CALL_NOLINK;
1709 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1712 std::vector<SDValue> Ops;
1713 Ops.push_back(Chain);
1714 Ops.push_back(Callee);
1716 // Add argument registers to the end of the list so that they are known live
1718 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1719 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1720 RegsToPass[i].second.getValueType()));
1722 // Add a register mask operand representing the call-preserved registers.
1724 const uint32_t *Mask;
1725 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1726 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1728 // For 'this' returns, use the R0-preserving mask if applicable
1729 Mask = ARI->getThisReturnPreservedMask(CallConv);
1731 // Set isThisReturn to false if the calling convention is not one that
1732 // allows 'returned' to be modeled in this way, so LowerCallResult does
1733 // not try to pass 'this' straight through
1734 isThisReturn = false;
1735 Mask = ARI->getCallPreservedMask(CallConv);
1738 Mask = ARI->getCallPreservedMask(CallConv);
1740 assert(Mask && "Missing call preserved mask for calling convention");
1741 Ops.push_back(DAG.getRegisterMask(Mask));
1744 if (InFlag.getNode())
1745 Ops.push_back(InFlag);
1747 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1749 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1751 // Returns a chain and a flag for retval copy to use.
1752 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1753 InFlag = Chain.getValue(1);
1755 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1756 DAG.getIntPtrConstant(0, true), InFlag, dl);
1758 InFlag = Chain.getValue(1);
1760 // Handle result values, copying them out of physregs into vregs that we
1762 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1763 InVals, isThisReturn,
1764 isThisReturn ? OutVals[0] : SDValue());
1767 /// HandleByVal - Every parameter *after* a byval parameter is passed
1768 /// on the stack. Remember the next parameter register to allocate,
1769 /// and then confiscate the rest of the parameter registers to insure
1772 ARMTargetLowering::HandleByVal(
1773 CCState *State, unsigned &size, unsigned Align) const {
1774 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1775 assert((State->getCallOrPrologue() == Prologue ||
1776 State->getCallOrPrologue() == Call) &&
1777 "unhandled ParmContext");
1779 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1780 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1781 unsigned AlignInRegs = Align / 4;
1782 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1783 for (unsigned i = 0; i < Waste; ++i)
1784 reg = State->AllocateReg(GPRArgRegs, 4);
1787 unsigned excess = 4 * (ARM::R4 - reg);
1789 // Special case when NSAA != SP and parameter size greater than size of
1790 // all remained GPR regs. In that case we can't split parameter, we must
1791 // send it to stack. We also must set NCRN to R4, so waste all
1792 // remained registers.
1793 const unsigned NSAAOffset = State->getNextStackOffset();
1794 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1795 while (State->AllocateReg(GPRArgRegs, 4))
1800 // First register for byval parameter is the first register that wasn't
1801 // allocated before this method call, so it would be "reg".
1802 // If parameter is small enough to be saved in range [reg, r4), then
1803 // the end (first after last) register would be reg + param-size-in-regs,
1804 // else parameter would be splitted between registers and stack,
1805 // end register would be r4 in this case.
1806 unsigned ByValRegBegin = reg;
1807 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1808 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1809 // Note, first register is allocated in the beginning of function already,
1810 // allocate remained amount of registers we need.
1811 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1812 State->AllocateReg(GPRArgRegs, 4);
1813 // A byval parameter that is split between registers and memory needs its
1814 // size truncated here.
1815 // In the case where the entire structure fits in registers, we set the
1816 // size in memory to zero.
1825 /// MatchingStackOffset - Return true if the given stack call argument is
1826 /// already available in the same position (relatively) of the caller's
1827 /// incoming argument stack.
1829 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1830 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1831 const TargetInstrInfo *TII) {
1832 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1834 if (Arg.getOpcode() == ISD::CopyFromReg) {
1835 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1836 if (!TargetRegisterInfo::isVirtualRegister(VR))
1838 MachineInstr *Def = MRI->getVRegDef(VR);
1841 if (!Flags.isByVal()) {
1842 if (!TII->isLoadFromStackSlot(Def, FI))
1847 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1848 if (Flags.isByVal())
1849 // ByVal argument is passed in as a pointer but it's now being
1850 // dereferenced. e.g.
1851 // define @foo(%struct.X* %A) {
1852 // tail call @bar(%struct.X* byval %A)
1855 SDValue Ptr = Ld->getBasePtr();
1856 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1859 FI = FINode->getIndex();
1863 assert(FI != INT_MAX);
1864 if (!MFI->isFixedObjectIndex(FI))
1866 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1869 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1870 /// for tail call optimization. Targets which want to do tail call
1871 /// optimization should implement this function.
1873 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1874 CallingConv::ID CalleeCC,
1876 bool isCalleeStructRet,
1877 bool isCallerStructRet,
1878 const SmallVectorImpl<ISD::OutputArg> &Outs,
1879 const SmallVectorImpl<SDValue> &OutVals,
1880 const SmallVectorImpl<ISD::InputArg> &Ins,
1881 SelectionDAG& DAG) const {
1882 const Function *CallerF = DAG.getMachineFunction().getFunction();
1883 CallingConv::ID CallerCC = CallerF->getCallingConv();
1884 bool CCMatch = CallerCC == CalleeCC;
1886 // Look for obvious safe cases to perform tail call optimization that do not
1887 // require ABI changes. This is what gcc calls sibcall.
1889 // Do not sibcall optimize vararg calls unless the call site is not passing
1891 if (isVarArg && !Outs.empty())
1894 // Exception-handling functions need a special set of instructions to indicate
1895 // a return to the hardware. Tail-calling another function would probably
1897 if (CallerF->hasFnAttribute("interrupt"))
1900 // Also avoid sibcall optimization if either caller or callee uses struct
1901 // return semantics.
1902 if (isCalleeStructRet || isCallerStructRet)
1905 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1906 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1907 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1908 // support in the assembler and linker to be used. This would need to be
1909 // fixed to fully support tail calls in Thumb1.
1911 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1912 // LR. This means if we need to reload LR, it takes an extra instructions,
1913 // which outweighs the value of the tail call; but here we don't know yet
1914 // whether LR is going to be used. Probably the right approach is to
1915 // generate the tail call here and turn it back into CALL/RET in
1916 // emitEpilogue if LR is used.
1918 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1919 // but we need to make sure there are enough registers; the only valid
1920 // registers are the 4 used for parameters. We don't currently do this
1922 if (Subtarget->isThumb1Only())
1925 // If the calling conventions do not match, then we'd better make sure the
1926 // results are returned in the same way as what the caller expects.
1928 SmallVector<CCValAssign, 16> RVLocs1;
1929 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1930 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1931 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1933 SmallVector<CCValAssign, 16> RVLocs2;
1934 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1935 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1936 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1938 if (RVLocs1.size() != RVLocs2.size())
1940 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1941 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1943 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1945 if (RVLocs1[i].isRegLoc()) {
1946 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1949 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1955 // If Caller's vararg or byval argument has been split between registers and
1956 // stack, do not perform tail call, since part of the argument is in caller's
1958 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1959 getInfo<ARMFunctionInfo>();
1960 if (AFI_Caller->getArgRegsSaveSize())
1963 // If the callee takes no arguments then go on to check the results of the
1965 if (!Outs.empty()) {
1966 // Check if stack adjustment is needed. For now, do not do this if any
1967 // argument is passed on the stack.
1968 SmallVector<CCValAssign, 16> ArgLocs;
1969 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1970 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1971 CCInfo.AnalyzeCallOperands(Outs,
1972 CCAssignFnForNode(CalleeCC, false, isVarArg));
1973 if (CCInfo.getNextStackOffset()) {
1974 MachineFunction &MF = DAG.getMachineFunction();
1976 // Check if the arguments are already laid out in the right way as
1977 // the caller's fixed stack objects.
1978 MachineFrameInfo *MFI = MF.getFrameInfo();
1979 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1980 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1981 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1983 ++i, ++realArgIdx) {
1984 CCValAssign &VA = ArgLocs[i];
1985 EVT RegVT = VA.getLocVT();
1986 SDValue Arg = OutVals[realArgIdx];
1987 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1988 if (VA.getLocInfo() == CCValAssign::Indirect)
1990 if (VA.needsCustom()) {
1991 // f64 and vector types are split into multiple registers or
1992 // register/stack-slot combinations. The types will not match
1993 // the registers; give up on memory f64 refs until we figure
1994 // out what to do about this.
1997 if (!ArgLocs[++i].isRegLoc())
1999 if (RegVT == MVT::v2f64) {
2000 if (!ArgLocs[++i].isRegLoc())
2002 if (!ArgLocs[++i].isRegLoc())
2005 } else if (!VA.isRegLoc()) {
2006 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2018 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2019 MachineFunction &MF, bool isVarArg,
2020 const SmallVectorImpl<ISD::OutputArg> &Outs,
2021 LLVMContext &Context) const {
2022 SmallVector<CCValAssign, 16> RVLocs;
2023 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2024 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2028 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2029 SDLoc DL, SelectionDAG &DAG) {
2030 const MachineFunction &MF = DAG.getMachineFunction();
2031 const Function *F = MF.getFunction();
2033 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2035 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2036 // version of the "preferred return address". These offsets affect the return
2037 // instruction if this is a return from PL1 without hypervisor extensions.
2038 // IRQ/FIQ: +4 "subs pc, lr, #4"
2039 // SWI: 0 "subs pc, lr, #0"
2040 // ABORT: +4 "subs pc, lr, #4"
2041 // UNDEF: +4/+2 "subs pc, lr, #0"
2042 // UNDEF varies depending on where the exception came from ARM or Thumb
2043 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2046 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2049 else if (IntKind == "SWI" || IntKind == "UNDEF")
2052 report_fatal_error("Unsupported interrupt attribute. If present, value "
2053 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2055 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2057 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other,
2058 RetOps.data(), RetOps.size());
2062 ARMTargetLowering::LowerReturn(SDValue Chain,
2063 CallingConv::ID CallConv, bool isVarArg,
2064 const SmallVectorImpl<ISD::OutputArg> &Outs,
2065 const SmallVectorImpl<SDValue> &OutVals,
2066 SDLoc dl, SelectionDAG &DAG) const {
2068 // CCValAssign - represent the assignment of the return value to a location.
2069 SmallVector<CCValAssign, 16> RVLocs;
2071 // CCState - Info about the registers and stack slots.
2072 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2073 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2075 // Analyze outgoing return values.
2076 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2080 SmallVector<SDValue, 4> RetOps;
2081 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2083 // Copy the result values into the output registers.
2084 for (unsigned i = 0, realRVLocIdx = 0;
2086 ++i, ++realRVLocIdx) {
2087 CCValAssign &VA = RVLocs[i];
2088 assert(VA.isRegLoc() && "Can only return in registers!");
2090 SDValue Arg = OutVals[realRVLocIdx];
2092 switch (VA.getLocInfo()) {
2093 default: llvm_unreachable("Unknown loc info!");
2094 case CCValAssign::Full: break;
2095 case CCValAssign::BCvt:
2096 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2100 if (VA.needsCustom()) {
2101 if (VA.getLocVT() == MVT::v2f64) {
2102 // Extract the first half and return it in two registers.
2103 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2104 DAG.getConstant(0, MVT::i32));
2105 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2106 DAG.getVTList(MVT::i32, MVT::i32), Half);
2108 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2109 Flag = Chain.getValue(1);
2110 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2111 VA = RVLocs[++i]; // skip ahead to next loc
2112 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2113 HalfGPRs.getValue(1), Flag);
2114 Flag = Chain.getValue(1);
2115 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2116 VA = RVLocs[++i]; // skip ahead to next loc
2118 // Extract the 2nd half and fall through to handle it as an f64 value.
2119 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2120 DAG.getConstant(1, MVT::i32));
2122 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2124 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2125 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2126 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2127 Flag = Chain.getValue(1);
2128 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2129 VA = RVLocs[++i]; // skip ahead to next loc
2130 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2133 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2135 // Guarantee that all emitted copies are
2136 // stuck together, avoiding something bad.
2137 Flag = Chain.getValue(1);
2138 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2141 // Update chain and glue.
2144 RetOps.push_back(Flag);
2146 // CPUs which aren't M-class use a special sequence to return from
2147 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2148 // though we use "subs pc, lr, #N").
2150 // M-class CPUs actually use a normal return sequence with a special
2151 // (hardware-provided) value in LR, so the normal code path works.
2152 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2153 !Subtarget->isMClass()) {
2154 if (Subtarget->isThumb1Only())
2155 report_fatal_error("interrupt attribute is not supported in Thumb1");
2156 return LowerInterruptReturn(RetOps, dl, DAG);
2159 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2160 RetOps.data(), RetOps.size());
2163 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2164 if (N->getNumValues() != 1)
2166 if (!N->hasNUsesOfValue(1, 0))
2169 SDValue TCChain = Chain;
2170 SDNode *Copy = *N->use_begin();
2171 if (Copy->getOpcode() == ISD::CopyToReg) {
2172 // If the copy has a glue operand, we conservatively assume it isn't safe to
2173 // perform a tail call.
2174 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2176 TCChain = Copy->getOperand(0);
2177 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2178 SDNode *VMov = Copy;
2179 // f64 returned in a pair of GPRs.
2180 SmallPtrSet<SDNode*, 2> Copies;
2181 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2183 if (UI->getOpcode() != ISD::CopyToReg)
2187 if (Copies.size() > 2)
2190 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2192 SDValue UseChain = UI->getOperand(0);
2193 if (Copies.count(UseChain.getNode()))
2200 } else if (Copy->getOpcode() == ISD::BITCAST) {
2201 // f32 returned in a single GPR.
2202 if (!Copy->hasOneUse())
2204 Copy = *Copy->use_begin();
2205 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2207 TCChain = Copy->getOperand(0);
2212 bool HasRet = false;
2213 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2215 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2216 UI->getOpcode() != ARMISD::INTRET_FLAG)
2228 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2229 if (!Subtarget->supportsTailCall())
2232 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2235 return !Subtarget->isThumb1Only();
2238 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2239 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2240 // one of the above mentioned nodes. It has to be wrapped because otherwise
2241 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2242 // be used to form addressing mode. These wrapped nodes will be selected
2244 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2245 EVT PtrVT = Op.getValueType();
2246 // FIXME there is no actual debug info here
2248 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2250 if (CP->isMachineConstantPoolEntry())
2251 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2252 CP->getAlignment());
2254 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2255 CP->getAlignment());
2256 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2259 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2260 return MachineJumpTableInfo::EK_Inline;
2263 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2264 SelectionDAG &DAG) const {
2265 MachineFunction &MF = DAG.getMachineFunction();
2266 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2267 unsigned ARMPCLabelIndex = 0;
2269 EVT PtrVT = getPointerTy();
2270 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2271 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2273 if (RelocM == Reloc::Static) {
2274 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2276 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2277 ARMPCLabelIndex = AFI->createPICLabelUId();
2278 ARMConstantPoolValue *CPV =
2279 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2280 ARMCP::CPBlockAddress, PCAdj);
2281 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2283 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2284 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2285 MachinePointerInfo::getConstantPool(),
2286 false, false, false, 0);
2287 if (RelocM == Reloc::Static)
2289 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2290 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2293 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2295 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2296 SelectionDAG &DAG) const {
2298 EVT PtrVT = getPointerTy();
2299 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2300 MachineFunction &MF = DAG.getMachineFunction();
2301 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2302 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2303 ARMConstantPoolValue *CPV =
2304 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2305 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2306 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2307 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2308 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2309 MachinePointerInfo::getConstantPool(),
2310 false, false, false, 0);
2311 SDValue Chain = Argument.getValue(1);
2313 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2314 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2316 // call __tls_get_addr.
2319 Entry.Node = Argument;
2320 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2321 Args.push_back(Entry);
2322 // FIXME: is there useful debug info available here?
2323 TargetLowering::CallLoweringInfo CLI(Chain,
2324 (Type *) Type::getInt32Ty(*DAG.getContext()),
2325 false, false, false, false,
2326 0, CallingConv::C, /*isTailCall=*/false,
2327 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2328 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2329 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2330 return CallResult.first;
2333 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2334 // "local exec" model.
2336 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2338 TLSModel::Model model) const {
2339 const GlobalValue *GV = GA->getGlobal();
2342 SDValue Chain = DAG.getEntryNode();
2343 EVT PtrVT = getPointerTy();
2344 // Get the Thread Pointer
2345 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2347 if (model == TLSModel::InitialExec) {
2348 MachineFunction &MF = DAG.getMachineFunction();
2349 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2350 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2351 // Initial exec model.
2352 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2353 ARMConstantPoolValue *CPV =
2354 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2355 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2357 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2358 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2359 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2360 MachinePointerInfo::getConstantPool(),
2361 false, false, false, 0);
2362 Chain = Offset.getValue(1);
2364 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2365 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2367 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2368 MachinePointerInfo::getConstantPool(),
2369 false, false, false, 0);
2372 assert(model == TLSModel::LocalExec);
2373 ARMConstantPoolValue *CPV =
2374 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2375 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2376 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2377 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2378 MachinePointerInfo::getConstantPool(),
2379 false, false, false, 0);
2382 // The address of the thread local variable is the add of the thread
2383 // pointer with the offset of the variable.
2384 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2388 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2389 // TODO: implement the "local dynamic" model
2390 assert(Subtarget->isTargetELF() &&
2391 "TLS not implemented for non-ELF targets");
2392 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2394 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2397 case TLSModel::GeneralDynamic:
2398 case TLSModel::LocalDynamic:
2399 return LowerToTLSGeneralDynamicModel(GA, DAG);
2400 case TLSModel::InitialExec:
2401 case TLSModel::LocalExec:
2402 return LowerToTLSExecModels(GA, DAG, model);
2404 llvm_unreachable("bogus TLS model");
2407 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2408 SelectionDAG &DAG) const {
2409 EVT PtrVT = getPointerTy();
2411 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2412 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2413 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2414 ARMConstantPoolValue *CPV =
2415 ARMConstantPoolConstant::Create(GV,
2416 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2417 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2418 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2419 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2421 MachinePointerInfo::getConstantPool(),
2422 false, false, false, 0);
2423 SDValue Chain = Result.getValue(1);
2424 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2425 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2427 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2428 MachinePointerInfo::getGOT(),
2429 false, false, false, 0);
2433 // If we have T2 ops, we can materialize the address directly via movt/movw
2434 // pair. This is always cheaper.
2435 if (Subtarget->useMovt()) {
2437 // FIXME: Once remat is capable of dealing with instructions with register
2438 // operands, expand this into two nodes.
2439 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2440 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2442 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2443 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2444 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2445 MachinePointerInfo::getConstantPool(),
2446 false, false, false, 0);
2450 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2451 SelectionDAG &DAG) const {
2452 EVT PtrVT = getPointerTy();
2454 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2455 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2457 if (Subtarget->useMovt())
2460 // FIXME: Once remat is capable of dealing with instructions with register
2461 // operands, expand this into multiple nodes
2463 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2465 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2466 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2468 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2469 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2470 MachinePointerInfo::getGOT(), false, false, false, 0);
2474 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2475 SelectionDAG &DAG) const {
2476 assert(Subtarget->isTargetELF() &&
2477 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2478 MachineFunction &MF = DAG.getMachineFunction();
2479 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2480 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2481 EVT PtrVT = getPointerTy();
2483 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2484 ARMConstantPoolValue *CPV =
2485 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2486 ARMPCLabelIndex, PCAdj);
2487 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2488 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2489 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2490 MachinePointerInfo::getConstantPool(),
2491 false, false, false, 0);
2492 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2493 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2497 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2499 SDValue Val = DAG.getConstant(0, MVT::i32);
2500 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2501 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2502 Op.getOperand(1), Val);
2506 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2508 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2509 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2513 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2514 const ARMSubtarget *Subtarget) const {
2515 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2518 default: return SDValue(); // Don't custom lower most intrinsics.
2519 case Intrinsic::arm_thread_pointer: {
2520 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2521 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2523 case Intrinsic::eh_sjlj_lsda: {
2524 MachineFunction &MF = DAG.getMachineFunction();
2525 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2526 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2527 EVT PtrVT = getPointerTy();
2528 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2530 unsigned PCAdj = (RelocM != Reloc::PIC_)
2531 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2532 ARMConstantPoolValue *CPV =
2533 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2534 ARMCP::CPLSDA, PCAdj);
2535 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2536 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2538 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2539 MachinePointerInfo::getConstantPool(),
2540 false, false, false, 0);
2542 if (RelocM == Reloc::PIC_) {
2543 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2544 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2548 case Intrinsic::arm_neon_vmulls:
2549 case Intrinsic::arm_neon_vmullu: {
2550 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2551 ? ARMISD::VMULLs : ARMISD::VMULLu;
2552 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2553 Op.getOperand(1), Op.getOperand(2));
2558 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2559 const ARMSubtarget *Subtarget) {
2560 // FIXME: handle "fence singlethread" more efficiently.
2562 if (!Subtarget->hasDataBarrier()) {
2563 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2564 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2566 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2567 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2568 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2569 DAG.getConstant(0, MVT::i32));
2572 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2573 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2574 unsigned Domain = ARM_MB::ISH;
2575 if (Subtarget->isMClass()) {
2576 // Only a full system barrier exists in the M-class architectures.
2577 Domain = ARM_MB::SY;
2578 } else if (Subtarget->isSwift() && Ord == Release) {
2579 // Swift happens to implement ISHST barriers in a way that's compatible with
2580 // Release semantics but weaker than ISH so we'd be fools not to use
2581 // it. Beware: other processors probably don't!
2582 Domain = ARM_MB::ISHST;
2585 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2586 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2587 DAG.getConstant(Domain, MVT::i32));
2590 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2591 const ARMSubtarget *Subtarget) {
2592 // ARM pre v5TE and Thumb1 does not have preload instructions.
2593 if (!(Subtarget->isThumb2() ||
2594 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2595 // Just preserve the chain.
2596 return Op.getOperand(0);
2599 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2601 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2602 // ARMv7 with MP extension has PLDW.
2603 return Op.getOperand(0);
2605 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2606 if (Subtarget->isThumb()) {
2608 isRead = ~isRead & 1;
2609 isData = ~isData & 1;
2612 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2613 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2614 DAG.getConstant(isData, MVT::i32));
2617 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2618 MachineFunction &MF = DAG.getMachineFunction();
2619 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2621 // vastart just stores the address of the VarArgsFrameIndex slot into the
2622 // memory location argument.
2624 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2625 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2626 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2627 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2628 MachinePointerInfo(SV), false, false, 0);
2632 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2633 SDValue &Root, SelectionDAG &DAG,
2635 MachineFunction &MF = DAG.getMachineFunction();
2636 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2638 const TargetRegisterClass *RC;
2639 if (AFI->isThumb1OnlyFunction())
2640 RC = &ARM::tGPRRegClass;
2642 RC = &ARM::GPRRegClass;
2644 // Transform the arguments stored in physical registers into virtual ones.
2645 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2646 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2649 if (NextVA.isMemLoc()) {
2650 MachineFrameInfo *MFI = MF.getFrameInfo();
2651 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2653 // Create load node to retrieve arguments from the stack.
2654 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2655 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2656 MachinePointerInfo::getFixedStack(FI),
2657 false, false, false, 0);
2659 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2660 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2663 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2667 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2668 unsigned InRegsParamRecordIdx,
2670 unsigned &ArgRegsSize,
2671 unsigned &ArgRegsSaveSize)
2674 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2675 unsigned RBegin, REnd;
2676 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2677 NumGPRs = REnd - RBegin;
2679 unsigned int firstUnalloced;
2680 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2681 sizeof(GPRArgRegs) /
2682 sizeof(GPRArgRegs[0]));
2683 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2686 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2687 ArgRegsSize = NumGPRs * 4;
2689 // If parameter is split between stack and GPRs...
2690 if (NumGPRs && Align > 4 &&
2691 (ArgRegsSize < ArgSize ||
2692 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2693 // Add padding for part of param recovered from GPRs. For example,
2694 // if Align == 8, its last byte must be at address K*8 - 1.
2695 // We need to do it, since remained (stack) part of parameter has
2696 // stack alignment, and we need to "attach" "GPRs head" without gaps
2699 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2700 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2702 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2704 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2705 ArgRegsSaveSize = ArgRegsSize + Padding;
2707 // We don't need to extend regs save size for byval parameters if they
2708 // are passed via GPRs only.
2709 ArgRegsSaveSize = ArgRegsSize;
2712 // The remaining GPRs hold either the beginning of variable-argument
2713 // data, or the beginning of an aggregate passed by value (usually
2714 // byval). Either way, we allocate stack slots adjacent to the data
2715 // provided by our caller, and store the unallocated registers there.
2716 // If this is a variadic function, the va_list pointer will begin with
2717 // these values; otherwise, this reassembles a (byval) structure that
2718 // was split between registers and memory.
2719 // Return: The frame index registers were stored into.
2721 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2722 SDLoc dl, SDValue &Chain,
2723 const Value *OrigArg,
2724 unsigned InRegsParamRecordIdx,
2725 unsigned OffsetFromOrigArg,
2729 unsigned ByValStoreOffset,
2730 unsigned TotalArgRegsSaveSize) const {
2732 // Currently, two use-cases possible:
2733 // Case #1. Non-var-args function, and we meet first byval parameter.
2734 // Setup first unallocated register as first byval register;
2735 // eat all remained registers
2736 // (these two actions are performed by HandleByVal method).
2737 // Then, here, we initialize stack frame with
2738 // "store-reg" instructions.
2739 // Case #2. Var-args function, that doesn't contain byval parameters.
2740 // The same: eat all remained unallocated registers,
2741 // initialize stack frame.
2743 MachineFunction &MF = DAG.getMachineFunction();
2744 MachineFrameInfo *MFI = MF.getFrameInfo();
2745 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2746 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2747 unsigned RBegin, REnd;
2748 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2749 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2750 firstRegToSaveIndex = RBegin - ARM::R0;
2751 lastRegToSaveIndex = REnd - ARM::R0;
2753 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2754 (GPRArgRegs, array_lengthof(GPRArgRegs));
2755 lastRegToSaveIndex = 4;
2758 unsigned ArgRegsSize, ArgRegsSaveSize;
2759 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2760 ArgRegsSize, ArgRegsSaveSize);
2762 // Store any by-val regs to their spots on the stack so that they may be
2763 // loaded by deferencing the result of formal parameter pointer or va_next.
2764 // Note: once stack area for byval/varargs registers
2765 // was initialized, it can't be initialized again.
2766 if (ArgRegsSaveSize) {
2767 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2770 assert(AFI->getStoredByValParamsPadding() == 0 &&
2771 "The only parameter may be padded.");
2772 AFI->setStoredByValParamsPadding(Padding);
2775 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2778 (int64_t)TotalArgRegsSaveSize,
2780 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2782 MFI->CreateFixedObject(Padding,
2783 ArgOffset + ByValStoreOffset -
2784 (int64_t)ArgRegsSaveSize,
2788 SmallVector<SDValue, 4> MemOps;
2789 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2790 ++firstRegToSaveIndex, ++i) {
2791 const TargetRegisterClass *RC;
2792 if (AFI->isThumb1OnlyFunction())
2793 RC = &ARM::tGPRRegClass;
2795 RC = &ARM::GPRRegClass;
2797 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2798 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2800 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2801 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2803 MemOps.push_back(Store);
2804 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2805 DAG.getConstant(4, getPointerTy()));
2808 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2810 if (!MemOps.empty())
2811 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2812 &MemOps[0], MemOps.size());
2816 // We cannot allocate a zero-byte object for the first variadic argument,
2817 // so just make up a size.
2820 // This will point to the next argument passed via stack.
2821 return MFI->CreateFixedObject(
2822 ArgSize, ArgOffset, !ForceMutable);
2826 // Setup stack frame, the va_list pointer will start from.
2828 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2829 SDLoc dl, SDValue &Chain,
2831 unsigned TotalArgRegsSaveSize,
2832 bool ForceMutable) const {
2833 MachineFunction &MF = DAG.getMachineFunction();
2834 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2836 // Try to store any remaining integer argument regs
2837 // to their spots on the stack so that they may be loaded by deferencing
2838 // the result of va_next.
2839 // If there is no regs to be stored, just point address after last
2840 // argument passed via stack.
2842 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2843 0, ArgOffset, 0, ForceMutable, 0, TotalArgRegsSaveSize);
2845 AFI->setVarArgsFrameIndex(FrameIndex);
2849 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2850 CallingConv::ID CallConv, bool isVarArg,
2851 const SmallVectorImpl<ISD::InputArg>
2853 SDLoc dl, SelectionDAG &DAG,
2854 SmallVectorImpl<SDValue> &InVals)
2856 MachineFunction &MF = DAG.getMachineFunction();
2857 MachineFrameInfo *MFI = MF.getFrameInfo();
2859 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2861 // Assign locations to all of the incoming arguments.
2862 SmallVector<CCValAssign, 16> ArgLocs;
2863 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2864 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2865 CCInfo.AnalyzeFormalArguments(Ins,
2866 CCAssignFnForNode(CallConv, /* Return*/ false,
2869 SmallVector<SDValue, 16> ArgValues;
2870 int lastInsIndex = -1;
2872 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2873 unsigned CurArgIdx = 0;
2875 // Initially ArgRegsSaveSize is zero.
2876 // Then we increase this value each time we meet byval parameter.
2877 // We also increase this value in case of varargs function.
2878 AFI->setArgRegsSaveSize(0);
2880 unsigned ByValStoreOffset = 0;
2881 unsigned TotalArgRegsSaveSize = 0;
2882 unsigned ArgRegsSaveSizeMaxAlign = 4;
2884 // Calculate the amount of stack space that we need to allocate to store
2885 // byval and variadic arguments that are passed in registers.
2886 // We need to know this before we allocate the first byval or variadic
2887 // argument, as they will be allocated a stack slot below the CFA (Canonical
2888 // Frame Address, the stack pointer at entry to the function).
2889 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2890 CCValAssign &VA = ArgLocs[i];
2891 if (VA.isMemLoc()) {
2892 int index = VA.getValNo();
2893 if (index != lastInsIndex) {
2894 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2895 if (Flags.isByVal()) {
2896 unsigned ExtraArgRegsSize;
2897 unsigned ExtraArgRegsSaveSize;
2898 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2899 Flags.getByValSize(),
2900 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2902 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2903 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2904 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2905 CCInfo.nextInRegsParam();
2907 lastInsIndex = index;
2911 CCInfo.rewindByValRegsInfo();
2914 unsigned ExtraArgRegsSize;
2915 unsigned ExtraArgRegsSaveSize;
2916 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2917 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2918 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2920 // If the arg regs save area contains N-byte aligned values, the
2921 // bottom of it must be at least N-byte aligned.
2922 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2923 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2925 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2926 CCValAssign &VA = ArgLocs[i];
2927 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2928 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2929 // Arguments stored in registers.
2930 if (VA.isRegLoc()) {
2931 EVT RegVT = VA.getLocVT();
2933 if (VA.needsCustom()) {
2934 // f64 and vector types are split up into multiple registers or
2935 // combinations of registers and stack slots.
2936 if (VA.getLocVT() == MVT::v2f64) {
2937 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2939 VA = ArgLocs[++i]; // skip ahead to next loc
2941 if (VA.isMemLoc()) {
2942 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2943 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2944 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2945 MachinePointerInfo::getFixedStack(FI),
2946 false, false, false, 0);
2948 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2951 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2952 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2953 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2954 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2955 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2957 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2960 const TargetRegisterClass *RC;
2962 if (RegVT == MVT::f32)
2963 RC = &ARM::SPRRegClass;
2964 else if (RegVT == MVT::f64)
2965 RC = &ARM::DPRRegClass;
2966 else if (RegVT == MVT::v2f64)
2967 RC = &ARM::QPRRegClass;
2968 else if (RegVT == MVT::i32)
2969 RC = AFI->isThumb1OnlyFunction() ?
2970 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2971 (const TargetRegisterClass*)&ARM::GPRRegClass;
2973 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2975 // Transform the arguments in physical registers into virtual ones.
2976 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2977 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2980 // If this is an 8 or 16-bit value, it is really passed promoted
2981 // to 32 bits. Insert an assert[sz]ext to capture this, then
2982 // truncate to the right size.
2983 switch (VA.getLocInfo()) {
2984 default: llvm_unreachable("Unknown loc info!");
2985 case CCValAssign::Full: break;
2986 case CCValAssign::BCvt:
2987 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2989 case CCValAssign::SExt:
2990 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2991 DAG.getValueType(VA.getValVT()));
2992 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2994 case CCValAssign::ZExt:
2995 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2996 DAG.getValueType(VA.getValVT()));
2997 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3001 InVals.push_back(ArgValue);
3003 } else { // VA.isRegLoc()
3006 assert(VA.isMemLoc());
3007 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3009 int index = ArgLocs[i].getValNo();
3011 // Some Ins[] entries become multiple ArgLoc[] entries.
3012 // Process them only once.
3013 if (index != lastInsIndex)
3015 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3016 // FIXME: For now, all byval parameter objects are marked mutable.
3017 // This can be changed with more analysis.
3018 // In case of tail call optimization mark all arguments mutable.
3019 // Since they could be overwritten by lowering of arguments in case of
3021 if (Flags.isByVal()) {
3022 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3024 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3025 int FrameIndex = StoreByValRegs(
3026 CCInfo, DAG, dl, Chain, CurOrigArg,
3028 Ins[VA.getValNo()].PartOffset,
3029 VA.getLocMemOffset(),
3030 Flags.getByValSize(),
3031 true /*force mutable frames*/,
3033 TotalArgRegsSaveSize);
3034 ByValStoreOffset += Flags.getByValSize();
3035 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3036 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3037 CCInfo.nextInRegsParam();
3039 unsigned FIOffset = VA.getLocMemOffset();
3040 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3043 // Create load nodes to retrieve arguments from the stack.
3044 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3045 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3046 MachinePointerInfo::getFixedStack(FI),
3047 false, false, false, 0));
3049 lastInsIndex = index;
3056 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3057 CCInfo.getNextStackOffset(),
3058 TotalArgRegsSaveSize);
3060 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3065 /// isFloatingPointZero - Return true if this is +0.0.
3066 static bool isFloatingPointZero(SDValue Op) {
3067 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3068 return CFP->getValueAPF().isPosZero();
3069 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3070 // Maybe this has already been legalized into the constant pool?
3071 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3072 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3073 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3074 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3075 return CFP->getValueAPF().isPosZero();
3081 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3082 /// the given operands.
3084 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3085 SDValue &ARMcc, SelectionDAG &DAG,
3087 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3088 unsigned C = RHSC->getZExtValue();
3089 if (!isLegalICmpImmediate(C)) {
3090 // Constant does not fit, try adjusting it by one?
3095 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3096 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3097 RHS = DAG.getConstant(C-1, MVT::i32);
3102 if (C != 0 && isLegalICmpImmediate(C-1)) {
3103 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3104 RHS = DAG.getConstant(C-1, MVT::i32);
3109 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3110 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3111 RHS = DAG.getConstant(C+1, MVT::i32);
3116 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3117 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3118 RHS = DAG.getConstant(C+1, MVT::i32);
3125 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3126 ARMISD::NodeType CompareType;
3129 CompareType = ARMISD::CMP;
3134 CompareType = ARMISD::CMPZ;
3137 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3138 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3141 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3143 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3146 if (!isFloatingPointZero(RHS))
3147 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3149 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3150 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3153 /// duplicateCmp - Glue values can have only one use, so this function
3154 /// duplicates a comparison node.
3156 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3157 unsigned Opc = Cmp.getOpcode();
3159 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3160 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3162 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3163 Cmp = Cmp.getOperand(0);
3164 Opc = Cmp.getOpcode();
3165 if (Opc == ARMISD::CMPFP)
3166 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3168 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3169 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3171 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3174 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3175 SDValue Cond = Op.getOperand(0);
3176 SDValue SelectTrue = Op.getOperand(1);
3177 SDValue SelectFalse = Op.getOperand(2);
3182 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3183 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3185 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3186 const ConstantSDNode *CMOVTrue =
3187 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3188 const ConstantSDNode *CMOVFalse =
3189 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3191 if (CMOVTrue && CMOVFalse) {
3192 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3193 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3197 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3199 False = SelectFalse;
3200 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3205 if (True.getNode() && False.getNode()) {
3206 EVT VT = Op.getValueType();
3207 SDValue ARMcc = Cond.getOperand(2);
3208 SDValue CCR = Cond.getOperand(3);
3209 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3210 assert(True.getValueType() == VT);
3211 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3216 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3217 // undefined bits before doing a full-word comparison with zero.
3218 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3219 DAG.getConstant(1, Cond.getValueType()));
3221 return DAG.getSelectCC(dl, Cond,
3222 DAG.getConstant(0, Cond.getValueType()),
3223 SelectTrue, SelectFalse, ISD::SETNE);
3226 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3227 if (CC == ISD::SETNE)
3229 return ISD::getSetCCInverse(CC, true);
3232 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3233 bool &swpCmpOps, bool &swpVselOps) {
3234 // Start by selecting the GE condition code for opcodes that return true for
3236 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3238 CondCode = ARMCC::GE;
3240 // and GT for opcodes that return false for 'equality'.
3241 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3243 CondCode = ARMCC::GT;
3245 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3246 // to swap the compare operands.
3247 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3251 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3252 // If we have an unordered opcode, we need to swap the operands to the VSEL
3253 // instruction (effectively negating the condition).
3255 // This also has the effect of swapping which one of 'less' or 'greater'
3256 // returns true, so we also swap the compare operands. It also switches
3257 // whether we return true for 'equality', so we compensate by picking the
3258 // opposite condition code to our original choice.
3259 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3260 CC == ISD::SETUGT) {
3261 swpCmpOps = !swpCmpOps;
3262 swpVselOps = !swpVselOps;
3263 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3266 // 'ordered' is 'anything but unordered', so use the VS condition code and
3267 // swap the VSEL operands.
3268 if (CC == ISD::SETO) {
3269 CondCode = ARMCC::VS;
3273 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3274 // code and swap the VSEL operands.
3275 if (CC == ISD::SETUNE) {
3276 CondCode = ARMCC::EQ;
3281 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3282 EVT VT = Op.getValueType();
3283 SDValue LHS = Op.getOperand(0);
3284 SDValue RHS = Op.getOperand(1);
3285 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3286 SDValue TrueVal = Op.getOperand(2);
3287 SDValue FalseVal = Op.getOperand(3);
3290 if (LHS.getValueType() == MVT::i32) {
3291 // Try to generate VSEL on ARMv8.
3292 // The VSEL instruction can't use all the usual ARM condition
3293 // codes: it only has two bits to select the condition code, so it's
3294 // constrained to use only GE, GT, VS and EQ.
3296 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3297 // swap the operands of the previous compare instruction (effectively
3298 // inverting the compare condition, swapping 'less' and 'greater') and
3299 // sometimes need to swap the operands to the VSEL (which inverts the
3300 // condition in the sense of firing whenever the previous condition didn't)
3301 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3302 TrueVal.getValueType() == MVT::f64)) {
3303 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3304 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3305 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3306 CC = getInverseCCForVSEL(CC);
3307 std::swap(TrueVal, FalseVal);
3312 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3313 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3314 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3318 ARMCC::CondCodes CondCode, CondCode2;
3319 FPCCToARMCC(CC, CondCode, CondCode2);
3321 // Try to generate VSEL on ARMv8.
3322 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3323 TrueVal.getValueType() == MVT::f64)) {
3324 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3325 // same operands, as follows:
3326 // c = fcmp [ogt, olt, ugt, ult] a, b
3328 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3329 // handled differently than the original code sequence.
3330 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3332 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3333 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3334 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3335 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3338 bool swpCmpOps = false;
3339 bool swpVselOps = false;
3340 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3342 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3343 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3345 std::swap(LHS, RHS);
3347 std::swap(TrueVal, FalseVal);
3351 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3352 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3353 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3354 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3356 if (CondCode2 != ARMCC::AL) {
3357 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3358 // FIXME: Needs another CMP because flag can have but one use.
3359 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3360 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3361 Result, TrueVal, ARMcc2, CCR, Cmp2);
3366 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3367 /// to morph to an integer compare sequence.
3368 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3369 const ARMSubtarget *Subtarget) {
3370 SDNode *N = Op.getNode();
3371 if (!N->hasOneUse())
3372 // Otherwise it requires moving the value from fp to integer registers.
3374 if (!N->getNumValues())
3376 EVT VT = Op.getValueType();
3377 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3378 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3379 // vmrs are very slow, e.g. cortex-a8.
3382 if (isFloatingPointZero(Op)) {
3386 return ISD::isNormalLoad(N);
3389 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3390 if (isFloatingPointZero(Op))
3391 return DAG.getConstant(0, MVT::i32);
3393 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3394 return DAG.getLoad(MVT::i32, SDLoc(Op),
3395 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3396 Ld->isVolatile(), Ld->isNonTemporal(),
3397 Ld->isInvariant(), Ld->getAlignment());
3399 llvm_unreachable("Unknown VFP cmp argument!");
3402 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3403 SDValue &RetVal1, SDValue &RetVal2) {
3404 if (isFloatingPointZero(Op)) {
3405 RetVal1 = DAG.getConstant(0, MVT::i32);
3406 RetVal2 = DAG.getConstant(0, MVT::i32);
3410 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3411 SDValue Ptr = Ld->getBasePtr();
3412 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3413 Ld->getChain(), Ptr,
3414 Ld->getPointerInfo(),
3415 Ld->isVolatile(), Ld->isNonTemporal(),
3416 Ld->isInvariant(), Ld->getAlignment());
3418 EVT PtrType = Ptr.getValueType();
3419 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3420 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3421 PtrType, Ptr, DAG.getConstant(4, PtrType));
3422 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3423 Ld->getChain(), NewPtr,
3424 Ld->getPointerInfo().getWithOffset(4),
3425 Ld->isVolatile(), Ld->isNonTemporal(),
3426 Ld->isInvariant(), NewAlign);
3430 llvm_unreachable("Unknown VFP cmp argument!");
3433 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3434 /// f32 and even f64 comparisons to integer ones.
3436 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3437 SDValue Chain = Op.getOperand(0);
3438 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3439 SDValue LHS = Op.getOperand(2);
3440 SDValue RHS = Op.getOperand(3);
3441 SDValue Dest = Op.getOperand(4);
3444 bool LHSSeenZero = false;
3445 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3446 bool RHSSeenZero = false;
3447 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3448 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3449 // If unsafe fp math optimization is enabled and there are no other uses of
3450 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3451 // to an integer comparison.
3452 if (CC == ISD::SETOEQ)
3454 else if (CC == ISD::SETUNE)
3457 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3459 if (LHS.getValueType() == MVT::f32) {
3460 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3461 bitcastf32Toi32(LHS, DAG), Mask);
3462 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3463 bitcastf32Toi32(RHS, DAG), Mask);
3464 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3465 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3466 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3467 Chain, Dest, ARMcc, CCR, Cmp);
3472 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3473 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3474 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3475 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3476 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3477 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3478 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3479 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3480 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3486 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3487 SDValue Chain = Op.getOperand(0);
3488 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3489 SDValue LHS = Op.getOperand(2);
3490 SDValue RHS = Op.getOperand(3);
3491 SDValue Dest = Op.getOperand(4);
3494 if (LHS.getValueType() == MVT::i32) {
3496 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3497 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3498 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3499 Chain, Dest, ARMcc, CCR, Cmp);
3502 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3504 if (getTargetMachine().Options.UnsafeFPMath &&
3505 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3506 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3507 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3508 if (Result.getNode())
3512 ARMCC::CondCodes CondCode, CondCode2;
3513 FPCCToARMCC(CC, CondCode, CondCode2);
3515 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3516 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3517 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3518 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3519 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3520 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3521 if (CondCode2 != ARMCC::AL) {
3522 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3523 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3524 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3529 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3530 SDValue Chain = Op.getOperand(0);
3531 SDValue Table = Op.getOperand(1);
3532 SDValue Index = Op.getOperand(2);
3535 EVT PTy = getPointerTy();
3536 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3537 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3538 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3539 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3540 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3541 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3542 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3543 if (Subtarget->isThumb2()) {
3544 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3545 // which does another jump to the destination. This also makes it easier
3546 // to translate it to TBB / TBH later.
3547 // FIXME: This might not work if the function is extremely large.
3548 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3549 Addr, Op.getOperand(2), JTI, UId);
3551 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3552 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3553 MachinePointerInfo::getJumpTable(),
3554 false, false, false, 0);
3555 Chain = Addr.getValue(1);
3556 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3557 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3559 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3560 MachinePointerInfo::getJumpTable(),
3561 false, false, false, 0);
3562 Chain = Addr.getValue(1);
3563 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3567 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3568 EVT VT = Op.getValueType();
3571 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3572 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3574 return DAG.UnrollVectorOp(Op.getNode());
3577 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3578 "Invalid type for custom lowering!");
3579 if (VT != MVT::v4i16)
3580 return DAG.UnrollVectorOp(Op.getNode());
3582 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3583 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3586 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3587 EVT VT = Op.getValueType();
3589 return LowerVectorFP_TO_INT(Op, DAG);
3594 switch (Op.getOpcode()) {
3595 default: llvm_unreachable("Invalid opcode!");
3596 case ISD::FP_TO_SINT:
3597 Opc = ARMISD::FTOSI;
3599 case ISD::FP_TO_UINT:
3600 Opc = ARMISD::FTOUI;
3603 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3604 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3607 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3608 EVT VT = Op.getValueType();
3611 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3612 if (VT.getVectorElementType() == MVT::f32)
3614 return DAG.UnrollVectorOp(Op.getNode());
3617 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3618 "Invalid type for custom lowering!");
3619 if (VT != MVT::v4f32)
3620 return DAG.UnrollVectorOp(Op.getNode());
3624 switch (Op.getOpcode()) {
3625 default: llvm_unreachable("Invalid opcode!");
3626 case ISD::SINT_TO_FP:
3627 CastOpc = ISD::SIGN_EXTEND;
3628 Opc = ISD::SINT_TO_FP;
3630 case ISD::UINT_TO_FP:
3631 CastOpc = ISD::ZERO_EXTEND;
3632 Opc = ISD::UINT_TO_FP;
3636 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3637 return DAG.getNode(Opc, dl, VT, Op);
3640 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3641 EVT VT = Op.getValueType();
3643 return LowerVectorINT_TO_FP(Op, DAG);
3648 switch (Op.getOpcode()) {
3649 default: llvm_unreachable("Invalid opcode!");
3650 case ISD::SINT_TO_FP:
3651 Opc = ARMISD::SITOF;
3653 case ISD::UINT_TO_FP:
3654 Opc = ARMISD::UITOF;
3658 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3659 return DAG.getNode(Opc, dl, VT, Op);
3662 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3663 // Implement fcopysign with a fabs and a conditional fneg.
3664 SDValue Tmp0 = Op.getOperand(0);
3665 SDValue Tmp1 = Op.getOperand(1);
3667 EVT VT = Op.getValueType();
3668 EVT SrcVT = Tmp1.getValueType();
3669 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3670 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3671 bool UseNEON = !InGPR && Subtarget->hasNEON();
3674 // Use VBSL to copy the sign bit.
3675 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3676 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3677 DAG.getTargetConstant(EncodedVal, MVT::i32));
3678 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3680 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3681 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3682 DAG.getConstant(32, MVT::i32));
3683 else /*if (VT == MVT::f32)*/
3684 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3685 if (SrcVT == MVT::f32) {
3686 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3688 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3689 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3690 DAG.getConstant(32, MVT::i32));
3691 } else if (VT == MVT::f32)
3692 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3693 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3694 DAG.getConstant(32, MVT::i32));
3695 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3696 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3698 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3700 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3701 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3702 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3704 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3705 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3706 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3707 if (VT == MVT::f32) {
3708 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3709 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3710 DAG.getConstant(0, MVT::i32));
3712 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3718 // Bitcast operand 1 to i32.
3719 if (SrcVT == MVT::f64)
3720 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3721 &Tmp1, 1).getValue(1);
3722 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3724 // Or in the signbit with integer operations.
3725 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3726 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3727 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3728 if (VT == MVT::f32) {
3729 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3730 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3731 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3732 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3735 // f64: Or the high part with signbit and then combine two parts.
3736 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3738 SDValue Lo = Tmp0.getValue(0);
3739 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3740 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3741 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3744 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3745 MachineFunction &MF = DAG.getMachineFunction();
3746 MachineFrameInfo *MFI = MF.getFrameInfo();
3747 MFI->setReturnAddressIsTaken(true);
3749 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3752 EVT VT = Op.getValueType();
3754 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3756 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3757 SDValue Offset = DAG.getConstant(4, MVT::i32);
3758 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3759 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3760 MachinePointerInfo(), false, false, false, 0);
3763 // Return LR, which contains the return address. Mark it an implicit live-in.
3764 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3765 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3768 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3769 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3770 MFI->setFrameAddressIsTaken(true);
3772 EVT VT = Op.getValueType();
3773 SDLoc dl(Op); // FIXME probably not meaningful
3774 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3775 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetMachO())
3776 ? ARM::R7 : ARM::R11;
3777 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3779 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3780 MachinePointerInfo(),
3781 false, false, false, 0);
3785 /// ExpandBITCAST - If the target supports VFP, this function is called to
3786 /// expand a bit convert where either the source or destination type is i64 to
3787 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3788 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3789 /// vectors), since the legalizer won't know what to do with that.
3790 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3791 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3793 SDValue Op = N->getOperand(0);
3795 // This function is only supposed to be called for i64 types, either as the
3796 // source or destination of the bit convert.
3797 EVT SrcVT = Op.getValueType();
3798 EVT DstVT = N->getValueType(0);
3799 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3800 "ExpandBITCAST called for non-i64 type");
3802 // Turn i64->f64 into VMOVDRR.
3803 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3804 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3805 DAG.getConstant(0, MVT::i32));
3806 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3807 DAG.getConstant(1, MVT::i32));
3808 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3809 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3812 // Turn f64->i64 into VMOVRRD.
3813 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3814 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3815 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3816 // Merge the pieces into a single i64 value.
3817 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3823 /// getZeroVector - Returns a vector of specified type with all zero elements.
3824 /// Zero vectors are used to represent vector negation and in those cases
3825 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3826 /// not support i64 elements, so sometimes the zero vectors will need to be
3827 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3829 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3830 assert(VT.isVector() && "Expected a vector type");
3831 // The canonical modified immediate encoding of a zero vector is....0!
3832 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3833 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3834 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3835 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3838 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3839 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3840 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3841 SelectionDAG &DAG) const {
3842 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3843 EVT VT = Op.getValueType();
3844 unsigned VTBits = VT.getSizeInBits();
3846 SDValue ShOpLo = Op.getOperand(0);
3847 SDValue ShOpHi = Op.getOperand(1);
3848 SDValue ShAmt = Op.getOperand(2);
3850 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3852 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3854 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3855 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3856 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3857 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3858 DAG.getConstant(VTBits, MVT::i32));
3859 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3860 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3861 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3863 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3864 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3866 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3867 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3870 SDValue Ops[2] = { Lo, Hi };
3871 return DAG.getMergeValues(Ops, 2, dl);
3874 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3875 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3876 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3877 SelectionDAG &DAG) const {
3878 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3879 EVT VT = Op.getValueType();
3880 unsigned VTBits = VT.getSizeInBits();
3882 SDValue ShOpLo = Op.getOperand(0);
3883 SDValue ShOpHi = Op.getOperand(1);
3884 SDValue ShAmt = Op.getOperand(2);
3887 assert(Op.getOpcode() == ISD::SHL_PARTS);
3888 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3889 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3890 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3891 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3892 DAG.getConstant(VTBits, MVT::i32));
3893 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3894 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3896 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3897 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3898 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3900 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3901 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3904 SDValue Ops[2] = { Lo, Hi };
3905 return DAG.getMergeValues(Ops, 2, dl);
3908 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3909 SelectionDAG &DAG) const {
3910 // The rounding mode is in bits 23:22 of the FPSCR.
3911 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3912 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3913 // so that the shift + and get folded into a bitfield extract.
3915 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3916 DAG.getConstant(Intrinsic::arm_get_fpscr,
3918 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3919 DAG.getConstant(1U << 22, MVT::i32));
3920 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3921 DAG.getConstant(22, MVT::i32));
3922 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3923 DAG.getConstant(3, MVT::i32));
3926 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3927 const ARMSubtarget *ST) {
3928 EVT VT = N->getValueType(0);
3931 if (!ST->hasV6T2Ops())
3934 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3935 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3938 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3939 /// for each 16-bit element from operand, repeated. The basic idea is to
3940 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3942 /// Trace for v4i16:
3943 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3944 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3945 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3946 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3947 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3948 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3949 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3950 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3951 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3952 EVT VT = N->getValueType(0);
3955 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3956 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3957 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3958 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3959 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3960 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3963 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3964 /// bit-count for each 16-bit element from the operand. We need slightly
3965 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3966 /// 64/128-bit registers.
3968 /// Trace for v4i16:
3969 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3970 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3971 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3972 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3973 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3974 EVT VT = N->getValueType(0);
3977 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3978 if (VT.is64BitVector()) {
3979 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3980 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3981 DAG.getIntPtrConstant(0));
3983 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3984 BitCounts, DAG.getIntPtrConstant(0));
3985 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3989 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3990 /// bit-count for each 32-bit element from the operand. The idea here is
3991 /// to split the vector into 16-bit elements, leverage the 16-bit count
3992 /// routine, and then combine the results.
3994 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3995 /// input = [v0 v1 ] (vi: 32-bit elements)
3996 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3997 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3998 /// vrev: N0 = [k1 k0 k3 k2 ]
4000 /// N1 =+[k1 k0 k3 k2 ]
4002 /// N2 =+[k1 k3 k0 k2 ]
4004 /// Extended =+[k1 k3 k0 k2 ]
4006 /// Extracted=+[k1 k3 ]
4008 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4009 EVT VT = N->getValueType(0);
4012 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4014 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4015 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4016 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4017 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4018 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4020 if (VT.is64BitVector()) {
4021 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4022 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4023 DAG.getIntPtrConstant(0));
4025 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4026 DAG.getIntPtrConstant(0));
4027 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4031 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4032 const ARMSubtarget *ST) {
4033 EVT VT = N->getValueType(0);
4035 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4036 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4037 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4038 "Unexpected type for custom ctpop lowering");
4040 if (VT.getVectorElementType() == MVT::i32)
4041 return lowerCTPOP32BitElements(N, DAG);
4043 return lowerCTPOP16BitElements(N, DAG);
4046 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4047 const ARMSubtarget *ST) {
4048 EVT VT = N->getValueType(0);
4054 // Lower vector shifts on NEON to use VSHL.
4055 assert(ST->hasNEON() && "unexpected vector shift");
4057 // Left shifts translate directly to the vshiftu intrinsic.
4058 if (N->getOpcode() == ISD::SHL)
4059 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4060 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4061 N->getOperand(0), N->getOperand(1));
4063 assert((N->getOpcode() == ISD::SRA ||
4064 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4066 // NEON uses the same intrinsics for both left and right shifts. For
4067 // right shifts, the shift amounts are negative, so negate the vector of
4069 EVT ShiftVT = N->getOperand(1).getValueType();
4070 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4071 getZeroVector(ShiftVT, DAG, dl),
4073 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4074 Intrinsic::arm_neon_vshifts :
4075 Intrinsic::arm_neon_vshiftu);
4076 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4077 DAG.getConstant(vshiftInt, MVT::i32),
4078 N->getOperand(0), NegatedCount);
4081 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4082 const ARMSubtarget *ST) {
4083 EVT VT = N->getValueType(0);
4086 // We can get here for a node like i32 = ISD::SHL i32, i64
4090 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4091 "Unknown shift to lower!");
4093 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4094 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4095 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4098 // If we are in thumb mode, we don't have RRX.
4099 if (ST->isThumb1Only()) return SDValue();
4101 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4102 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4103 DAG.getConstant(0, MVT::i32));
4104 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4105 DAG.getConstant(1, MVT::i32));
4107 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4108 // captures the result into a carry flag.
4109 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4110 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4112 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4113 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4115 // Merge the pieces into a single i64 value.
4116 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4119 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4120 SDValue TmpOp0, TmpOp1;
4121 bool Invert = false;
4125 SDValue Op0 = Op.getOperand(0);
4126 SDValue Op1 = Op.getOperand(1);
4127 SDValue CC = Op.getOperand(2);
4128 EVT VT = Op.getValueType();
4129 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4132 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4133 switch (SetCCOpcode) {
4134 default: llvm_unreachable("Illegal FP comparison");
4136 case ISD::SETNE: Invert = true; // Fallthrough
4138 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4140 case ISD::SETLT: Swap = true; // Fallthrough
4142 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4144 case ISD::SETLE: Swap = true; // Fallthrough
4146 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4147 case ISD::SETUGE: Swap = true; // Fallthrough
4148 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4149 case ISD::SETUGT: Swap = true; // Fallthrough
4150 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4151 case ISD::SETUEQ: Invert = true; // Fallthrough
4153 // Expand this to (OLT | OGT).
4157 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4158 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4160 case ISD::SETUO: Invert = true; // Fallthrough
4162 // Expand this to (OLT | OGE).
4166 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4167 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4171 // Integer comparisons.
4172 switch (SetCCOpcode) {
4173 default: llvm_unreachable("Illegal integer comparison");
4174 case ISD::SETNE: Invert = true;
4175 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4176 case ISD::SETLT: Swap = true;
4177 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4178 case ISD::SETLE: Swap = true;
4179 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4180 case ISD::SETULT: Swap = true;
4181 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4182 case ISD::SETULE: Swap = true;
4183 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4186 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4187 if (Opc == ARMISD::VCEQ) {
4190 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4192 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4195 // Ignore bitconvert.
4196 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4197 AndOp = AndOp.getOperand(0);
4199 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4201 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4202 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4209 std::swap(Op0, Op1);
4211 // If one of the operands is a constant vector zero, attempt to fold the
4212 // comparison to a specialized compare-against-zero form.
4214 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4216 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4217 if (Opc == ARMISD::VCGE)
4218 Opc = ARMISD::VCLEZ;
4219 else if (Opc == ARMISD::VCGT)
4220 Opc = ARMISD::VCLTZ;
4225 if (SingleOp.getNode()) {
4228 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4230 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4232 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4234 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4236 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4238 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4241 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4245 Result = DAG.getNOT(dl, Result, VT);
4250 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4251 /// valid vector constant for a NEON instruction with a "modified immediate"
4252 /// operand (e.g., VMOV). If so, return the encoded value.
4253 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4254 unsigned SplatBitSize, SelectionDAG &DAG,
4255 EVT &VT, bool is128Bits, NEONModImmType type) {
4256 unsigned OpCmode, Imm;
4258 // SplatBitSize is set to the smallest size that splats the vector, so a
4259 // zero vector will always have SplatBitSize == 8. However, NEON modified
4260 // immediate instructions others than VMOV do not support the 8-bit encoding
4261 // of a zero vector, and the default encoding of zero is supposed to be the
4266 switch (SplatBitSize) {
4268 if (type != VMOVModImm)
4270 // Any 1-byte value is OK. Op=0, Cmode=1110.
4271 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4274 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4278 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4279 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4280 if ((SplatBits & ~0xff) == 0) {
4281 // Value = 0x00nn: Op=x, Cmode=100x.
4286 if ((SplatBits & ~0xff00) == 0) {
4287 // Value = 0xnn00: Op=x, Cmode=101x.
4289 Imm = SplatBits >> 8;
4295 // NEON's 32-bit VMOV supports splat values where:
4296 // * only one byte is nonzero, or
4297 // * the least significant byte is 0xff and the second byte is nonzero, or
4298 // * the least significant 2 bytes are 0xff and the third is nonzero.
4299 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4300 if ((SplatBits & ~0xff) == 0) {
4301 // Value = 0x000000nn: Op=x, Cmode=000x.
4306 if ((SplatBits & ~0xff00) == 0) {
4307 // Value = 0x0000nn00: Op=x, Cmode=001x.
4309 Imm = SplatBits >> 8;
4312 if ((SplatBits & ~0xff0000) == 0) {
4313 // Value = 0x00nn0000: Op=x, Cmode=010x.
4315 Imm = SplatBits >> 16;
4318 if ((SplatBits & ~0xff000000) == 0) {
4319 // Value = 0xnn000000: Op=x, Cmode=011x.
4321 Imm = SplatBits >> 24;
4325 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4326 if (type == OtherModImm) return SDValue();
4328 if ((SplatBits & ~0xffff) == 0 &&
4329 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4330 // Value = 0x0000nnff: Op=x, Cmode=1100.
4332 Imm = SplatBits >> 8;
4336 if ((SplatBits & ~0xffffff) == 0 &&
4337 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4338 // Value = 0x00nnffff: Op=x, Cmode=1101.
4340 Imm = SplatBits >> 16;
4344 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4345 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4346 // VMOV.I32. A (very) minor optimization would be to replicate the value
4347 // and fall through here to test for a valid 64-bit splat. But, then the
4348 // caller would also need to check and handle the change in size.
4352 if (type != VMOVModImm)
4354 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4355 uint64_t BitMask = 0xff;
4357 unsigned ImmMask = 1;
4359 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4360 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4363 } else if ((SplatBits & BitMask) != 0) {
4369 // Op=1, Cmode=1110.
4371 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4376 llvm_unreachable("unexpected size for isNEONModifiedImm");
4379 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4380 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4383 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4384 const ARMSubtarget *ST) const {
4388 bool IsDouble = Op.getValueType() == MVT::f64;
4389 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4391 // Try splatting with a VMOV.f32...
4392 APFloat FPVal = CFP->getValueAPF();
4393 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4396 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4397 // We have code in place to select a valid ConstantFP already, no need to
4402 // It's a float and we are trying to use NEON operations where
4403 // possible. Lower it to a splat followed by an extract.
4405 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4406 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4408 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4409 DAG.getConstant(0, MVT::i32));
4412 // The rest of our options are NEON only, make sure that's allowed before
4414 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4418 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4420 // It wouldn't really be worth bothering for doubles except for one very
4421 // important value, which does happen to match: 0.0. So make sure we don't do
4423 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4426 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4427 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4429 if (NewVal != SDValue()) {
4431 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4434 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4436 // It's a float: cast and extract a vector element.
4437 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4439 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4440 DAG.getConstant(0, MVT::i32));
4443 // Finally, try a VMVN.i32
4444 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4446 if (NewVal != SDValue()) {
4448 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4451 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4453 // It's a float: cast and extract a vector element.
4454 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4456 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4457 DAG.getConstant(0, MVT::i32));
4463 // check if an VEXT instruction can handle the shuffle mask when the
4464 // vector sources of the shuffle are the same.
4465 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4466 unsigned NumElts = VT.getVectorNumElements();
4468 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4474 // If this is a VEXT shuffle, the immediate value is the index of the first
4475 // element. The other shuffle indices must be the successive elements after
4477 unsigned ExpectedElt = Imm;
4478 for (unsigned i = 1; i < NumElts; ++i) {
4479 // Increment the expected index. If it wraps around, just follow it
4480 // back to index zero and keep going.
4482 if (ExpectedElt == NumElts)
4485 if (M[i] < 0) continue; // ignore UNDEF indices
4486 if (ExpectedElt != static_cast<unsigned>(M[i]))
4494 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4495 bool &ReverseVEXT, unsigned &Imm) {
4496 unsigned NumElts = VT.getVectorNumElements();
4497 ReverseVEXT = false;
4499 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4505 // If this is a VEXT shuffle, the immediate value is the index of the first
4506 // element. The other shuffle indices must be the successive elements after
4508 unsigned ExpectedElt = Imm;
4509 for (unsigned i = 1; i < NumElts; ++i) {
4510 // Increment the expected index. If it wraps around, it may still be
4511 // a VEXT but the source vectors must be swapped.
4513 if (ExpectedElt == NumElts * 2) {
4518 if (M[i] < 0) continue; // ignore UNDEF indices
4519 if (ExpectedElt != static_cast<unsigned>(M[i]))
4523 // Adjust the index value if the source operands will be swapped.
4530 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4531 /// instruction with the specified blocksize. (The order of the elements
4532 /// within each block of the vector is reversed.)
4533 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4534 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4535 "Only possible block sizes for VREV are: 16, 32, 64");
4537 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4541 unsigned NumElts = VT.getVectorNumElements();
4542 unsigned BlockElts = M[0] + 1;
4543 // If the first shuffle index is UNDEF, be optimistic.
4545 BlockElts = BlockSize / EltSz;
4547 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4550 for (unsigned i = 0; i < NumElts; ++i) {
4551 if (M[i] < 0) continue; // ignore UNDEF indices
4552 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4559 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4560 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4561 // range, then 0 is placed into the resulting vector. So pretty much any mask
4562 // of 8 elements can work here.
4563 return VT == MVT::v8i8 && M.size() == 8;
4566 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4567 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4571 unsigned NumElts = VT.getVectorNumElements();
4572 WhichResult = (M[0] == 0 ? 0 : 1);
4573 for (unsigned i = 0; i < NumElts; i += 2) {
4574 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4575 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4581 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4582 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4583 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4584 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4585 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4589 unsigned NumElts = VT.getVectorNumElements();
4590 WhichResult = (M[0] == 0 ? 0 : 1);
4591 for (unsigned i = 0; i < NumElts; i += 2) {
4592 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4593 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4599 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4600 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4604 unsigned NumElts = VT.getVectorNumElements();
4605 WhichResult = (M[0] == 0 ? 0 : 1);
4606 for (unsigned i = 0; i != NumElts; ++i) {
4607 if (M[i] < 0) continue; // ignore UNDEF indices
4608 if ((unsigned) M[i] != 2 * i + WhichResult)
4612 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4613 if (VT.is64BitVector() && EltSz == 32)
4619 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4620 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4621 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4622 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4623 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4627 unsigned Half = VT.getVectorNumElements() / 2;
4628 WhichResult = (M[0] == 0 ? 0 : 1);
4629 for (unsigned j = 0; j != 2; ++j) {
4630 unsigned Idx = WhichResult;
4631 for (unsigned i = 0; i != Half; ++i) {
4632 int MIdx = M[i + j * Half];
4633 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4639 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4640 if (VT.is64BitVector() && EltSz == 32)
4646 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4647 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4651 unsigned NumElts = VT.getVectorNumElements();
4652 WhichResult = (M[0] == 0 ? 0 : 1);
4653 unsigned Idx = WhichResult * NumElts / 2;
4654 for (unsigned i = 0; i != NumElts; i += 2) {
4655 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4656 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4661 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4662 if (VT.is64BitVector() && EltSz == 32)
4668 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4669 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4670 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4671 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4672 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4676 unsigned NumElts = VT.getVectorNumElements();
4677 WhichResult = (M[0] == 0 ? 0 : 1);
4678 unsigned Idx = WhichResult * NumElts / 2;
4679 for (unsigned i = 0; i != NumElts; i += 2) {
4680 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4681 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4686 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4687 if (VT.is64BitVector() && EltSz == 32)
4693 /// \return true if this is a reverse operation on an vector.
4694 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4695 unsigned NumElts = VT.getVectorNumElements();
4696 // Make sure the mask has the right size.
4697 if (NumElts != M.size())
4700 // Look for <15, ..., 3, -1, 1, 0>.
4701 for (unsigned i = 0; i != NumElts; ++i)
4702 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4708 // If N is an integer constant that can be moved into a register in one
4709 // instruction, return an SDValue of such a constant (will become a MOV
4710 // instruction). Otherwise return null.
4711 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4712 const ARMSubtarget *ST, SDLoc dl) {
4714 if (!isa<ConstantSDNode>(N))
4716 Val = cast<ConstantSDNode>(N)->getZExtValue();
4718 if (ST->isThumb1Only()) {
4719 if (Val <= 255 || ~Val <= 255)
4720 return DAG.getConstant(Val, MVT::i32);
4722 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4723 return DAG.getConstant(Val, MVT::i32);
4728 // If this is a case we can't handle, return null and let the default
4729 // expansion code take care of it.
4730 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4731 const ARMSubtarget *ST) const {
4732 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4734 EVT VT = Op.getValueType();
4736 APInt SplatBits, SplatUndef;
4737 unsigned SplatBitSize;
4739 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4740 if (SplatBitSize <= 64) {
4741 // Check if an immediate VMOV works.
4743 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4744 SplatUndef.getZExtValue(), SplatBitSize,
4745 DAG, VmovVT, VT.is128BitVector(),
4747 if (Val.getNode()) {
4748 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4749 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4752 // Try an immediate VMVN.
4753 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4754 Val = isNEONModifiedImm(NegatedImm,
4755 SplatUndef.getZExtValue(), SplatBitSize,
4756 DAG, VmovVT, VT.is128BitVector(),
4758 if (Val.getNode()) {
4759 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4760 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4763 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4764 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4765 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4767 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4768 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4774 // Scan through the operands to see if only one value is used.
4776 // As an optimisation, even if more than one value is used it may be more
4777 // profitable to splat with one value then change some lanes.
4779 // Heuristically we decide to do this if the vector has a "dominant" value,
4780 // defined as splatted to more than half of the lanes.
4781 unsigned NumElts = VT.getVectorNumElements();
4782 bool isOnlyLowElement = true;
4783 bool usesOnlyOneValue = true;
4784 bool hasDominantValue = false;
4785 bool isConstant = true;
4787 // Map of the number of times a particular SDValue appears in the
4789 DenseMap<SDValue, unsigned> ValueCounts;
4791 for (unsigned i = 0; i < NumElts; ++i) {
4792 SDValue V = Op.getOperand(i);
4793 if (V.getOpcode() == ISD::UNDEF)
4796 isOnlyLowElement = false;
4797 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4800 ValueCounts.insert(std::make_pair(V, 0));
4801 unsigned &Count = ValueCounts[V];
4803 // Is this value dominant? (takes up more than half of the lanes)
4804 if (++Count > (NumElts / 2)) {
4805 hasDominantValue = true;
4809 if (ValueCounts.size() != 1)
4810 usesOnlyOneValue = false;
4811 if (!Value.getNode() && ValueCounts.size() > 0)
4812 Value = ValueCounts.begin()->first;
4814 if (ValueCounts.size() == 0)
4815 return DAG.getUNDEF(VT);
4817 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4818 // Keep going if we are hitting this case.
4819 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4820 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4822 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4824 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4825 // i32 and try again.
4826 if (hasDominantValue && EltSize <= 32) {
4830 // If we are VDUPing a value that comes directly from a vector, that will
4831 // cause an unnecessary move to and from a GPR, where instead we could
4832 // just use VDUPLANE. We can only do this if the lane being extracted
4833 // is at a constant index, as the VDUP from lane instructions only have
4834 // constant-index forms.
4835 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4836 isa<ConstantSDNode>(Value->getOperand(1))) {
4837 // We need to create a new undef vector to use for the VDUPLANE if the
4838 // size of the vector from which we get the value is different than the
4839 // size of the vector that we need to create. We will insert the element
4840 // such that the register coalescer will remove unnecessary copies.
4841 if (VT != Value->getOperand(0).getValueType()) {
4842 ConstantSDNode *constIndex;
4843 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4844 assert(constIndex && "The index is not a constant!");
4845 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4846 VT.getVectorNumElements();
4847 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4848 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4849 Value, DAG.getConstant(index, MVT::i32)),
4850 DAG.getConstant(index, MVT::i32));
4852 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4853 Value->getOperand(0), Value->getOperand(1));
4855 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4857 if (!usesOnlyOneValue) {
4858 // The dominant value was splatted as 'N', but we now have to insert
4859 // all differing elements.
4860 for (unsigned I = 0; I < NumElts; ++I) {
4861 if (Op.getOperand(I) == Value)
4863 SmallVector<SDValue, 3> Ops;
4865 Ops.push_back(Op.getOperand(I));
4866 Ops.push_back(DAG.getConstant(I, MVT::i32));
4867 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4872 if (VT.getVectorElementType().isFloatingPoint()) {
4873 SmallVector<SDValue, 8> Ops;
4874 for (unsigned i = 0; i < NumElts; ++i)
4875 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4877 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4878 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4879 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4881 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4883 if (usesOnlyOneValue) {
4884 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4885 if (isConstant && Val.getNode())
4886 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4890 // If all elements are constants and the case above didn't get hit, fall back
4891 // to the default expansion, which will generate a load from the constant
4896 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4898 SDValue shuffle = ReconstructShuffle(Op, DAG);
4899 if (shuffle != SDValue())
4903 // Vectors with 32- or 64-bit elements can be built by directly assigning
4904 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4905 // will be legalized.
4906 if (EltSize >= 32) {
4907 // Do the expansion with floating-point types, since that is what the VFP
4908 // registers are defined to use, and since i64 is not legal.
4909 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4910 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4911 SmallVector<SDValue, 8> Ops;
4912 for (unsigned i = 0; i < NumElts; ++i)
4913 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4914 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4915 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4918 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4919 // know the default expansion would otherwise fall back on something even
4920 // worse. For a vector with one or two non-undef values, that's
4921 // scalar_to_vector for the elements followed by a shuffle (provided the
4922 // shuffle is valid for the target) and materialization element by element
4923 // on the stack followed by a load for everything else.
4924 if (!isConstant && !usesOnlyOneValue) {
4925 SDValue Vec = DAG.getUNDEF(VT);
4926 for (unsigned i = 0 ; i < NumElts; ++i) {
4927 SDValue V = Op.getOperand(i);
4928 if (V.getOpcode() == ISD::UNDEF)
4930 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4931 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4939 // Gather data to see if the operation can be modelled as a
4940 // shuffle in combination with VEXTs.
4941 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4942 SelectionDAG &DAG) const {
4944 EVT VT = Op.getValueType();
4945 unsigned NumElts = VT.getVectorNumElements();
4947 SmallVector<SDValue, 2> SourceVecs;
4948 SmallVector<unsigned, 2> MinElts;
4949 SmallVector<unsigned, 2> MaxElts;
4951 for (unsigned i = 0; i < NumElts; ++i) {
4952 SDValue V = Op.getOperand(i);
4953 if (V.getOpcode() == ISD::UNDEF)
4955 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4956 // A shuffle can only come from building a vector from various
4957 // elements of other vectors.
4959 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4960 VT.getVectorElementType()) {
4961 // This code doesn't know how to handle shuffles where the vector
4962 // element types do not match (this happens because type legalization
4963 // promotes the return type of EXTRACT_VECTOR_ELT).
4964 // FIXME: It might be appropriate to extend this code to handle
4965 // mismatched types.
4969 // Record this extraction against the appropriate vector if possible...
4970 SDValue SourceVec = V.getOperand(0);
4971 // If the element number isn't a constant, we can't effectively
4972 // analyze what's going on.
4973 if (!isa<ConstantSDNode>(V.getOperand(1)))
4975 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4976 bool FoundSource = false;
4977 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4978 if (SourceVecs[j] == SourceVec) {
4979 if (MinElts[j] > EltNo)
4981 if (MaxElts[j] < EltNo)
4988 // Or record a new source if not...
4990 SourceVecs.push_back(SourceVec);
4991 MinElts.push_back(EltNo);
4992 MaxElts.push_back(EltNo);
4996 // Currently only do something sane when at most two source vectors
4998 if (SourceVecs.size() > 2)
5001 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5002 int VEXTOffsets[2] = {0, 0};
5004 // This loop extracts the usage patterns of the source vectors
5005 // and prepares appropriate SDValues for a shuffle if possible.
5006 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5007 if (SourceVecs[i].getValueType() == VT) {
5008 // No VEXT necessary
5009 ShuffleSrcs[i] = SourceVecs[i];
5012 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5013 // It probably isn't worth padding out a smaller vector just to
5014 // break it down again in a shuffle.
5018 // Since only 64-bit and 128-bit vectors are legal on ARM and
5019 // we've eliminated the other cases...
5020 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5021 "unexpected vector sizes in ReconstructShuffle");
5023 if (MaxElts[i] - MinElts[i] >= NumElts) {
5024 // Span too large for a VEXT to cope
5028 if (MinElts[i] >= NumElts) {
5029 // The extraction can just take the second half
5030 VEXTOffsets[i] = NumElts;
5031 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5033 DAG.getIntPtrConstant(NumElts));
5034 } else if (MaxElts[i] < NumElts) {
5035 // The extraction can just take the first half
5037 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5039 DAG.getIntPtrConstant(0));
5041 // An actual VEXT is needed
5042 VEXTOffsets[i] = MinElts[i];
5043 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5045 DAG.getIntPtrConstant(0));
5046 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5048 DAG.getIntPtrConstant(NumElts));
5049 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5050 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5054 SmallVector<int, 8> Mask;
5056 for (unsigned i = 0; i < NumElts; ++i) {
5057 SDValue Entry = Op.getOperand(i);
5058 if (Entry.getOpcode() == ISD::UNDEF) {
5063 SDValue ExtractVec = Entry.getOperand(0);
5064 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5065 .getOperand(1))->getSExtValue();
5066 if (ExtractVec == SourceVecs[0]) {
5067 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5069 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5073 // Final check before we try to produce nonsense...
5074 if (isShuffleMaskLegal(Mask, VT))
5075 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5081 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5082 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5083 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5084 /// are assumed to be legal.
5086 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5088 if (VT.getVectorNumElements() == 4 &&
5089 (VT.is128BitVector() || VT.is64BitVector())) {
5090 unsigned PFIndexes[4];
5091 for (unsigned i = 0; i != 4; ++i) {
5095 PFIndexes[i] = M[i];
5098 // Compute the index in the perfect shuffle table.
5099 unsigned PFTableIndex =
5100 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5101 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5102 unsigned Cost = (PFEntry >> 30);
5109 unsigned Imm, WhichResult;
5111 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5112 return (EltSize >= 32 ||
5113 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5114 isVREVMask(M, VT, 64) ||
5115 isVREVMask(M, VT, 32) ||
5116 isVREVMask(M, VT, 16) ||
5117 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5118 isVTBLMask(M, VT) ||
5119 isVTRNMask(M, VT, WhichResult) ||
5120 isVUZPMask(M, VT, WhichResult) ||
5121 isVZIPMask(M, VT, WhichResult) ||
5122 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5123 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5124 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5125 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5128 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5129 /// the specified operations to build the shuffle.
5130 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5131 SDValue RHS, SelectionDAG &DAG,
5133 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5134 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5135 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5138 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5147 OP_VUZPL, // VUZP, left result
5148 OP_VUZPR, // VUZP, right result
5149 OP_VZIPL, // VZIP, left result
5150 OP_VZIPR, // VZIP, right result
5151 OP_VTRNL, // VTRN, left result
5152 OP_VTRNR // VTRN, right result
5155 if (OpNum == OP_COPY) {
5156 if (LHSID == (1*9+2)*9+3) return LHS;
5157 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5161 SDValue OpLHS, OpRHS;
5162 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5163 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5164 EVT VT = OpLHS.getValueType();
5167 default: llvm_unreachable("Unknown shuffle opcode!");
5169 // VREV divides the vector in half and swaps within the half.
5170 if (VT.getVectorElementType() == MVT::i32 ||
5171 VT.getVectorElementType() == MVT::f32)
5172 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5173 // vrev <4 x i16> -> VREV32
5174 if (VT.getVectorElementType() == MVT::i16)
5175 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5176 // vrev <4 x i8> -> VREV16
5177 assert(VT.getVectorElementType() == MVT::i8);
5178 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5183 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5184 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5188 return DAG.getNode(ARMISD::VEXT, dl, VT,
5190 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5193 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5194 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5197 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5198 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5201 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5202 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5206 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5207 ArrayRef<int> ShuffleMask,
5208 SelectionDAG &DAG) {
5209 // Check to see if we can use the VTBL instruction.
5210 SDValue V1 = Op.getOperand(0);
5211 SDValue V2 = Op.getOperand(1);
5214 SmallVector<SDValue, 8> VTBLMask;
5215 for (ArrayRef<int>::iterator
5216 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5217 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5219 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5220 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5221 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5224 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5225 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5229 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5230 SelectionDAG &DAG) {
5232 SDValue OpLHS = Op.getOperand(0);
5233 EVT VT = OpLHS.getValueType();
5235 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5236 "Expect an v8i16/v16i8 type");
5237 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5238 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5239 // extract the first 8 bytes into the top double word and the last 8 bytes
5240 // into the bottom double word. The v8i16 case is similar.
5241 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5242 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5243 DAG.getConstant(ExtractNum, MVT::i32));
5246 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5247 SDValue V1 = Op.getOperand(0);
5248 SDValue V2 = Op.getOperand(1);
5250 EVT VT = Op.getValueType();
5251 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5253 // Convert shuffles that are directly supported on NEON to target-specific
5254 // DAG nodes, instead of keeping them as shuffles and matching them again
5255 // during code selection. This is more efficient and avoids the possibility
5256 // of inconsistencies between legalization and selection.
5257 // FIXME: floating-point vectors should be canonicalized to integer vectors
5258 // of the same time so that they get CSEd properly.
5259 ArrayRef<int> ShuffleMask = SVN->getMask();
5261 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5262 if (EltSize <= 32) {
5263 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5264 int Lane = SVN->getSplatIndex();
5265 // If this is undef splat, generate it via "just" vdup, if possible.
5266 if (Lane == -1) Lane = 0;
5268 // Test if V1 is a SCALAR_TO_VECTOR.
5269 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5270 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5272 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5273 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5275 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5276 !isa<ConstantSDNode>(V1.getOperand(0))) {
5277 bool IsScalarToVector = true;
5278 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5279 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5280 IsScalarToVector = false;
5283 if (IsScalarToVector)
5284 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5286 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5287 DAG.getConstant(Lane, MVT::i32));
5292 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5295 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5296 DAG.getConstant(Imm, MVT::i32));
5299 if (isVREVMask(ShuffleMask, VT, 64))
5300 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5301 if (isVREVMask(ShuffleMask, VT, 32))
5302 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5303 if (isVREVMask(ShuffleMask, VT, 16))
5304 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5306 if (V2->getOpcode() == ISD::UNDEF &&
5307 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5308 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5309 DAG.getConstant(Imm, MVT::i32));
5312 // Check for Neon shuffles that modify both input vectors in place.
5313 // If both results are used, i.e., if there are two shuffles with the same
5314 // source operands and with masks corresponding to both results of one of
5315 // these operations, DAG memoization will ensure that a single node is
5316 // used for both shuffles.
5317 unsigned WhichResult;
5318 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5319 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5320 V1, V2).getValue(WhichResult);
5321 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5322 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5323 V1, V2).getValue(WhichResult);
5324 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5325 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5326 V1, V2).getValue(WhichResult);
5328 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5329 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5330 V1, V1).getValue(WhichResult);
5331 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5332 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5333 V1, V1).getValue(WhichResult);
5334 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5335 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5336 V1, V1).getValue(WhichResult);
5339 // If the shuffle is not directly supported and it has 4 elements, use
5340 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5341 unsigned NumElts = VT.getVectorNumElements();
5343 unsigned PFIndexes[4];
5344 for (unsigned i = 0; i != 4; ++i) {
5345 if (ShuffleMask[i] < 0)
5348 PFIndexes[i] = ShuffleMask[i];
5351 // Compute the index in the perfect shuffle table.
5352 unsigned PFTableIndex =
5353 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5354 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5355 unsigned Cost = (PFEntry >> 30);
5358 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5361 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5362 if (EltSize >= 32) {
5363 // Do the expansion with floating-point types, since that is what the VFP
5364 // registers are defined to use, and since i64 is not legal.
5365 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5366 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5367 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5368 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5369 SmallVector<SDValue, 8> Ops;
5370 for (unsigned i = 0; i < NumElts; ++i) {
5371 if (ShuffleMask[i] < 0)
5372 Ops.push_back(DAG.getUNDEF(EltVT));
5374 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5375 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5376 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5379 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5380 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5383 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5384 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5386 if (VT == MVT::v8i8) {
5387 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5388 if (NewOp.getNode())
5395 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5396 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5397 SDValue Lane = Op.getOperand(2);
5398 if (!isa<ConstantSDNode>(Lane))
5404 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5405 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5406 SDValue Lane = Op.getOperand(1);
5407 if (!isa<ConstantSDNode>(Lane))
5410 SDValue Vec = Op.getOperand(0);
5411 if (Op.getValueType() == MVT::i32 &&
5412 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5414 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5420 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5421 // The only time a CONCAT_VECTORS operation can have legal types is when
5422 // two 64-bit vectors are concatenated to a 128-bit vector.
5423 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5424 "unexpected CONCAT_VECTORS");
5426 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5427 SDValue Op0 = Op.getOperand(0);
5428 SDValue Op1 = Op.getOperand(1);
5429 if (Op0.getOpcode() != ISD::UNDEF)
5430 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5431 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5432 DAG.getIntPtrConstant(0));
5433 if (Op1.getOpcode() != ISD::UNDEF)
5434 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5435 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5436 DAG.getIntPtrConstant(1));
5437 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5440 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5441 /// element has been zero/sign-extended, depending on the isSigned parameter,
5442 /// from an integer type half its size.
5443 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5445 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5446 EVT VT = N->getValueType(0);
5447 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5448 SDNode *BVN = N->getOperand(0).getNode();
5449 if (BVN->getValueType(0) != MVT::v4i32 ||
5450 BVN->getOpcode() != ISD::BUILD_VECTOR)
5452 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5453 unsigned HiElt = 1 - LoElt;
5454 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5455 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5456 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5457 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5458 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5461 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5462 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5465 if (Hi0->isNullValue() && Hi1->isNullValue())
5471 if (N->getOpcode() != ISD::BUILD_VECTOR)
5474 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5475 SDNode *Elt = N->getOperand(i).getNode();
5476 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5477 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5478 unsigned HalfSize = EltSize / 2;
5480 if (!isIntN(HalfSize, C->getSExtValue()))
5483 if (!isUIntN(HalfSize, C->getZExtValue()))
5494 /// isSignExtended - Check if a node is a vector value that is sign-extended
5495 /// or a constant BUILD_VECTOR with sign-extended elements.
5496 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5497 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5499 if (isExtendedBUILD_VECTOR(N, DAG, true))
5504 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5505 /// or a constant BUILD_VECTOR with zero-extended elements.
5506 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5507 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5509 if (isExtendedBUILD_VECTOR(N, DAG, false))
5514 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5515 if (OrigVT.getSizeInBits() >= 64)
5518 assert(OrigVT.isSimple() && "Expecting a simple value type");
5520 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5521 switch (OrigSimpleTy) {
5522 default: llvm_unreachable("Unexpected Vector Type");
5531 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5532 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5533 /// We insert the required extension here to get the vector to fill a D register.
5534 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5537 unsigned ExtOpcode) {
5538 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5539 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5540 // 64-bits we need to insert a new extension so that it will be 64-bits.
5541 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5542 if (OrigTy.getSizeInBits() >= 64)
5545 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5546 EVT NewVT = getExtensionTo64Bits(OrigTy);
5548 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5551 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5552 /// does not do any sign/zero extension. If the original vector is less
5553 /// than 64 bits, an appropriate extension will be added after the load to
5554 /// reach a total size of 64 bits. We have to add the extension separately
5555 /// because ARM does not have a sign/zero extending load for vectors.
5556 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5557 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5559 // The load already has the right type.
5560 if (ExtendedTy == LD->getMemoryVT())
5561 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5562 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5563 LD->isNonTemporal(), LD->isInvariant(),
5564 LD->getAlignment());
5566 // We need to create a zextload/sextload. We cannot just create a load
5567 // followed by a zext/zext node because LowerMUL is also run during normal
5568 // operation legalization where we can't create illegal types.
5569 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5570 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5571 LD->getMemoryVT(), LD->isVolatile(),
5572 LD->isNonTemporal(), LD->getAlignment());
5575 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5576 /// extending load, or BUILD_VECTOR with extended elements, return the
5577 /// unextended value. The unextended vector should be 64 bits so that it can
5578 /// be used as an operand to a VMULL instruction. If the original vector size
5579 /// before extension is less than 64 bits we add a an extension to resize
5580 /// the vector to 64 bits.
5581 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5582 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5583 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5584 N->getOperand(0)->getValueType(0),
5588 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5589 return SkipLoadExtensionForVMULL(LD, DAG);
5591 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5592 // have been legalized as a BITCAST from v4i32.
5593 if (N->getOpcode() == ISD::BITCAST) {
5594 SDNode *BVN = N->getOperand(0).getNode();
5595 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5596 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5597 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5598 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5599 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5601 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5602 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5603 EVT VT = N->getValueType(0);
5604 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5605 unsigned NumElts = VT.getVectorNumElements();
5606 MVT TruncVT = MVT::getIntegerVT(EltSize);
5607 SmallVector<SDValue, 8> Ops;
5608 for (unsigned i = 0; i != NumElts; ++i) {
5609 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5610 const APInt &CInt = C->getAPIntValue();
5611 // Element types smaller than 32 bits are not legal, so use i32 elements.
5612 // The values are implicitly truncated so sext vs. zext doesn't matter.
5613 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5615 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5616 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5619 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5620 unsigned Opcode = N->getOpcode();
5621 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5622 SDNode *N0 = N->getOperand(0).getNode();
5623 SDNode *N1 = N->getOperand(1).getNode();
5624 return N0->hasOneUse() && N1->hasOneUse() &&
5625 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5630 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5631 unsigned Opcode = N->getOpcode();
5632 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5633 SDNode *N0 = N->getOperand(0).getNode();
5634 SDNode *N1 = N->getOperand(1).getNode();
5635 return N0->hasOneUse() && N1->hasOneUse() &&
5636 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5641 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5642 // Multiplications are only custom-lowered for 128-bit vectors so that
5643 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5644 EVT VT = Op.getValueType();
5645 assert(VT.is128BitVector() && VT.isInteger() &&
5646 "unexpected type for custom-lowering ISD::MUL");
5647 SDNode *N0 = Op.getOperand(0).getNode();
5648 SDNode *N1 = Op.getOperand(1).getNode();
5649 unsigned NewOpc = 0;
5651 bool isN0SExt = isSignExtended(N0, DAG);
5652 bool isN1SExt = isSignExtended(N1, DAG);
5653 if (isN0SExt && isN1SExt)
5654 NewOpc = ARMISD::VMULLs;
5656 bool isN0ZExt = isZeroExtended(N0, DAG);
5657 bool isN1ZExt = isZeroExtended(N1, DAG);
5658 if (isN0ZExt && isN1ZExt)
5659 NewOpc = ARMISD::VMULLu;
5660 else if (isN1SExt || isN1ZExt) {
5661 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5662 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5663 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5664 NewOpc = ARMISD::VMULLs;
5666 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5667 NewOpc = ARMISD::VMULLu;
5669 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5671 NewOpc = ARMISD::VMULLu;
5677 if (VT == MVT::v2i64)
5678 // Fall through to expand this. It is not legal.
5681 // Other vector multiplications are legal.
5686 // Legalize to a VMULL instruction.
5689 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5691 Op0 = SkipExtensionForVMULL(N0, DAG);
5692 assert(Op0.getValueType().is64BitVector() &&
5693 Op1.getValueType().is64BitVector() &&
5694 "unexpected types for extended operands to VMULL");
5695 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5698 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5699 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5706 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5707 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5708 EVT Op1VT = Op1.getValueType();
5709 return DAG.getNode(N0->getOpcode(), DL, VT,
5710 DAG.getNode(NewOpc, DL, VT,
5711 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5712 DAG.getNode(NewOpc, DL, VT,
5713 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5717 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5719 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5720 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5721 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5722 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5723 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5724 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5725 // Get reciprocal estimate.
5726 // float4 recip = vrecpeq_f32(yf);
5727 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5728 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5729 // Because char has a smaller range than uchar, we can actually get away
5730 // without any newton steps. This requires that we use a weird bias
5731 // of 0xb000, however (again, this has been exhaustively tested).
5732 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5733 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5734 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5735 Y = DAG.getConstant(0xb000, MVT::i32);
5736 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5737 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5738 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5739 // Convert back to short.
5740 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5741 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5746 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5748 // Convert to float.
5749 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5750 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5751 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5752 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5753 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5754 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5756 // Use reciprocal estimate and one refinement step.
5757 // float4 recip = vrecpeq_f32(yf);
5758 // recip *= vrecpsq_f32(yf, recip);
5759 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5760 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5761 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5762 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5764 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5765 // Because short has a smaller range than ushort, we can actually get away
5766 // with only a single newton step. This requires that we use a weird bias
5767 // of 89, however (again, this has been exhaustively tested).
5768 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5769 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5770 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5771 N1 = DAG.getConstant(0x89, MVT::i32);
5772 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5773 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5774 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5775 // Convert back to integer and return.
5776 // return vmovn_s32(vcvt_s32_f32(result));
5777 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5778 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5782 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5783 EVT VT = Op.getValueType();
5784 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5785 "unexpected type for custom-lowering ISD::SDIV");
5788 SDValue N0 = Op.getOperand(0);
5789 SDValue N1 = Op.getOperand(1);
5792 if (VT == MVT::v8i8) {
5793 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5794 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5796 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5797 DAG.getIntPtrConstant(4));
5798 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5799 DAG.getIntPtrConstant(4));
5800 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5801 DAG.getIntPtrConstant(0));
5802 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5803 DAG.getIntPtrConstant(0));
5805 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5806 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5808 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5809 N0 = LowerCONCAT_VECTORS(N0, DAG);
5811 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5814 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5817 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5818 EVT VT = Op.getValueType();
5819 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5820 "unexpected type for custom-lowering ISD::UDIV");
5823 SDValue N0 = Op.getOperand(0);
5824 SDValue N1 = Op.getOperand(1);
5827 if (VT == MVT::v8i8) {
5828 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5829 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5831 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5832 DAG.getIntPtrConstant(4));
5833 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5834 DAG.getIntPtrConstant(4));
5835 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5836 DAG.getIntPtrConstant(0));
5837 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5838 DAG.getIntPtrConstant(0));
5840 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5841 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5843 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5844 N0 = LowerCONCAT_VECTORS(N0, DAG);
5846 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5847 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5852 // v4i16 sdiv ... Convert to float.
5853 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5854 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5855 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5856 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5857 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5858 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5860 // Use reciprocal estimate and two refinement steps.
5861 // float4 recip = vrecpeq_f32(yf);
5862 // recip *= vrecpsq_f32(yf, recip);
5863 // recip *= vrecpsq_f32(yf, recip);
5864 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5865 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5866 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5867 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5869 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5870 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5871 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5873 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5874 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5875 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5876 // and that it will never cause us to return an answer too large).
5877 // float4 result = as_float4(as_int4(xf*recip) + 2);
5878 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5879 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5880 N1 = DAG.getConstant(2, MVT::i32);
5881 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5882 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5883 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5884 // Convert back to integer and return.
5885 // return vmovn_u32(vcvt_s32_f32(result));
5886 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5887 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5891 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5892 EVT VT = Op.getNode()->getValueType(0);
5893 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5896 bool ExtraOp = false;
5897 switch (Op.getOpcode()) {
5898 default: llvm_unreachable("Invalid code");
5899 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5900 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5901 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5902 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5906 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5908 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5909 Op.getOperand(1), Op.getOperand(2));
5912 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
5913 assert(Subtarget->isTargetDarwin());
5915 // For iOS, we want to call an alternative entry point: __sincos_stret,
5916 // return values are passed via sret.
5918 SDValue Arg = Op.getOperand(0);
5919 EVT ArgVT = Arg.getValueType();
5920 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
5922 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5923 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5925 // Pair of floats / doubles used to pass the result.
5926 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
5928 // Create stack object for sret.
5929 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
5930 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
5931 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
5932 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
5938 Entry.Ty = RetTy->getPointerTo();
5939 Entry.isSExt = false;
5940 Entry.isZExt = false;
5941 Entry.isSRet = true;
5942 Args.push_back(Entry);
5946 Entry.isSExt = false;
5947 Entry.isZExt = false;
5948 Args.push_back(Entry);
5950 const char *LibcallName = (ArgVT == MVT::f64)
5951 ? "__sincos_stret" : "__sincosf_stret";
5952 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
5955 CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
5956 false, false, false, false, 0,
5957 CallingConv::C, /*isTaillCall=*/false,
5958 /*doesNotRet=*/false, /*isReturnValueUsed*/false,
5959 Callee, Args, DAG, dl);
5960 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
5962 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
5963 MachinePointerInfo(), false, false, false, 0);
5965 // Address of cos field.
5966 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
5967 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
5968 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
5969 MachinePointerInfo(), false, false, false, 0);
5971 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
5972 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
5973 LoadSin.getValue(0), LoadCos.getValue(0));
5976 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5977 // Monotonic load/store is legal for all targets
5978 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5981 // Acquire/Release load/store is not legal for targets without a
5982 // dmb or equivalent available.
5986 static void ReplaceREADCYCLECOUNTER(SDNode *N,
5987 SmallVectorImpl<SDValue> &Results,
5989 const ARMSubtarget *Subtarget) {
5991 SDValue Cycles32, OutChain;
5993 if (Subtarget->hasPerfMon()) {
5994 // Under Power Management extensions, the cycle-count is:
5995 // mrc p15, #0, <Rt>, c9, c13, #0
5996 SDValue Ops[] = { N->getOperand(0), // Chain
5997 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
5998 DAG.getConstant(15, MVT::i32),
5999 DAG.getConstant(0, MVT::i32),
6000 DAG.getConstant(9, MVT::i32),
6001 DAG.getConstant(13, MVT::i32),
6002 DAG.getConstant(0, MVT::i32)
6005 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6006 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6007 array_lengthof(Ops));
6008 OutChain = Cycles32.getValue(1);
6010 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6011 // there are older ARM CPUs that have implementation-specific ways of
6012 // obtaining this information (FIXME!).
6013 Cycles32 = DAG.getConstant(0, MVT::i32);
6014 OutChain = DAG.getEntryNode();
6018 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6019 Cycles32, DAG.getConstant(0, MVT::i32));
6020 Results.push_back(Cycles64);
6021 Results.push_back(OutChain);
6024 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6025 switch (Op.getOpcode()) {
6026 default: llvm_unreachable("Don't know how to custom lower this!");
6027 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6028 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6029 case ISD::GlobalAddress:
6030 return Subtarget->isTargetMachO() ? LowerGlobalAddressDarwin(Op, DAG) :
6031 LowerGlobalAddressELF(Op, DAG);
6032 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6033 case ISD::SELECT: return LowerSELECT(Op, DAG);
6034 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6035 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6036 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6037 case ISD::VASTART: return LowerVASTART(Op, DAG);
6038 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6039 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6040 case ISD::SINT_TO_FP:
6041 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6042 case ISD::FP_TO_SINT:
6043 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6044 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6045 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6046 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6047 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6048 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6049 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6050 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6052 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6055 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6056 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6057 case ISD::SRL_PARTS:
6058 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6059 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6060 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6061 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6062 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6063 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6064 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6065 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6066 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6067 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6068 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6069 case ISD::MUL: return LowerMUL(Op, DAG);
6070 case ISD::SDIV: return LowerSDIV(Op, DAG);
6071 case ISD::UDIV: return LowerUDIV(Op, DAG);
6075 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6076 case ISD::ATOMIC_LOAD:
6077 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6078 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6080 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6084 /// ReplaceNodeResults - Replace the results of node with an illegal result
6085 /// type with new values built out of custom code.
6086 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6087 SmallVectorImpl<SDValue>&Results,
6088 SelectionDAG &DAG) const {
6090 switch (N->getOpcode()) {
6092 llvm_unreachable("Don't know how to custom expand this!");
6094 Res = ExpandBITCAST(N, DAG);
6098 Res = Expand64BitShift(N, DAG, Subtarget);
6100 case ISD::READCYCLECOUNTER:
6101 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6105 Results.push_back(Res);
6108 //===----------------------------------------------------------------------===//
6109 // ARM Scheduler Hooks
6110 //===----------------------------------------------------------------------===//
6112 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6113 /// registers the function context.
6114 void ARMTargetLowering::
6115 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6116 MachineBasicBlock *DispatchBB, int FI) const {
6117 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6118 DebugLoc dl = MI->getDebugLoc();
6119 MachineFunction *MF = MBB->getParent();
6120 MachineRegisterInfo *MRI = &MF->getRegInfo();
6121 MachineConstantPool *MCP = MF->getConstantPool();
6122 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6123 const Function *F = MF->getFunction();
6125 bool isThumb = Subtarget->isThumb();
6126 bool isThumb2 = Subtarget->isThumb2();
6128 unsigned PCLabelId = AFI->createPICLabelUId();
6129 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6130 ARMConstantPoolValue *CPV =
6131 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6132 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6134 const TargetRegisterClass *TRC = isThumb ?
6135 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6136 (const TargetRegisterClass*)&ARM::GPRRegClass;
6138 // Grab constant pool and fixed stack memory operands.
6139 MachineMemOperand *CPMMO =
6140 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6141 MachineMemOperand::MOLoad, 4, 4);
6143 MachineMemOperand *FIMMOSt =
6144 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6145 MachineMemOperand::MOStore, 4, 4);
6147 // Load the address of the dispatch MBB into the jump buffer.
6149 // Incoming value: jbuf
6150 // ldr.n r5, LCPI1_1
6153 // str r5, [$jbuf, #+4] ; &jbuf[1]
6154 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6155 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6156 .addConstantPoolIndex(CPI)
6157 .addMemOperand(CPMMO));
6158 // Set the low bit because of thumb mode.
6159 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6161 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6162 .addReg(NewVReg1, RegState::Kill)
6164 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6165 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6166 .addReg(NewVReg2, RegState::Kill)
6168 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6169 .addReg(NewVReg3, RegState::Kill)
6171 .addImm(36) // &jbuf[1] :: pc
6172 .addMemOperand(FIMMOSt));
6173 } else if (isThumb) {
6174 // Incoming value: jbuf
6175 // ldr.n r1, LCPI1_4
6179 // add r2, $jbuf, #+4 ; &jbuf[1]
6181 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6182 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6183 .addConstantPoolIndex(CPI)
6184 .addMemOperand(CPMMO));
6185 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6186 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6187 .addReg(NewVReg1, RegState::Kill)
6189 // Set the low bit because of thumb mode.
6190 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6191 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6192 .addReg(ARM::CPSR, RegState::Define)
6194 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6195 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6196 .addReg(ARM::CPSR, RegState::Define)
6197 .addReg(NewVReg2, RegState::Kill)
6198 .addReg(NewVReg3, RegState::Kill));
6199 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6200 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6202 .addImm(36)); // &jbuf[1] :: pc
6203 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6204 .addReg(NewVReg4, RegState::Kill)
6205 .addReg(NewVReg5, RegState::Kill)
6207 .addMemOperand(FIMMOSt));
6209 // Incoming value: jbuf
6212 // str r1, [$jbuf, #+4] ; &jbuf[1]
6213 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6214 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6215 .addConstantPoolIndex(CPI)
6217 .addMemOperand(CPMMO));
6218 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6219 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6220 .addReg(NewVReg1, RegState::Kill)
6221 .addImm(PCLabelId));
6222 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6223 .addReg(NewVReg2, RegState::Kill)
6225 .addImm(36) // &jbuf[1] :: pc
6226 .addMemOperand(FIMMOSt));
6230 MachineBasicBlock *ARMTargetLowering::
6231 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6232 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6233 DebugLoc dl = MI->getDebugLoc();
6234 MachineFunction *MF = MBB->getParent();
6235 MachineRegisterInfo *MRI = &MF->getRegInfo();
6236 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6237 MachineFrameInfo *MFI = MF->getFrameInfo();
6238 int FI = MFI->getFunctionContextIndex();
6240 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6241 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6242 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6244 // Get a mapping of the call site numbers to all of the landing pads they're
6246 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6247 unsigned MaxCSNum = 0;
6248 MachineModuleInfo &MMI = MF->getMMI();
6249 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6251 if (!BB->isLandingPad()) continue;
6253 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6255 for (MachineBasicBlock::iterator
6256 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6257 if (!II->isEHLabel()) continue;
6259 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6260 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6262 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6263 for (SmallVectorImpl<unsigned>::iterator
6264 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6265 CSI != CSE; ++CSI) {
6266 CallSiteNumToLPad[*CSI].push_back(BB);
6267 MaxCSNum = std::max(MaxCSNum, *CSI);
6273 // Get an ordered list of the machine basic blocks for the jump table.
6274 std::vector<MachineBasicBlock*> LPadList;
6275 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6276 LPadList.reserve(CallSiteNumToLPad.size());
6277 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6278 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6279 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6280 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6281 LPadList.push_back(*II);
6282 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6286 assert(!LPadList.empty() &&
6287 "No landing pad destinations for the dispatch jump table!");
6289 // Create the jump table and associated information.
6290 MachineJumpTableInfo *JTI =
6291 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6292 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6293 unsigned UId = AFI->createJumpTableUId();
6294 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6296 // Create the MBBs for the dispatch code.
6298 // Shove the dispatch's address into the return slot in the function context.
6299 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6300 DispatchBB->setIsLandingPad();
6302 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6303 unsigned trap_opcode;
6304 if (Subtarget->isThumb())
6305 trap_opcode = ARM::tTRAP;
6307 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6309 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6310 DispatchBB->addSuccessor(TrapBB);
6312 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6313 DispatchBB->addSuccessor(DispContBB);
6316 MF->insert(MF->end(), DispatchBB);
6317 MF->insert(MF->end(), DispContBB);
6318 MF->insert(MF->end(), TrapBB);
6320 // Insert code into the entry block that creates and registers the function
6322 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6324 MachineMemOperand *FIMMOLd =
6325 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6326 MachineMemOperand::MOLoad |
6327 MachineMemOperand::MOVolatile, 4, 4);
6329 MachineInstrBuilder MIB;
6330 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6332 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6333 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6335 // Add a register mask with no preserved registers. This results in all
6336 // registers being marked as clobbered.
6337 MIB.addRegMask(RI.getNoPreservedMask());
6339 unsigned NumLPads = LPadList.size();
6340 if (Subtarget->isThumb2()) {
6341 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6342 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6345 .addMemOperand(FIMMOLd));
6347 if (NumLPads < 256) {
6348 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6350 .addImm(LPadList.size()));
6352 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6353 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6354 .addImm(NumLPads & 0xFFFF));
6356 unsigned VReg2 = VReg1;
6357 if ((NumLPads & 0xFFFF0000) != 0) {
6358 VReg2 = MRI->createVirtualRegister(TRC);
6359 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6361 .addImm(NumLPads >> 16));
6364 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6369 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6374 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6375 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6376 .addJumpTableIndex(MJTI)
6379 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6382 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6383 .addReg(NewVReg3, RegState::Kill)
6385 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6387 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6388 .addReg(NewVReg4, RegState::Kill)
6390 .addJumpTableIndex(MJTI)
6392 } else if (Subtarget->isThumb()) {
6393 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6394 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6397 .addMemOperand(FIMMOLd));
6399 if (NumLPads < 256) {
6400 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6404 MachineConstantPool *ConstantPool = MF->getConstantPool();
6405 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6406 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6408 // MachineConstantPool wants an explicit alignment.
6409 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6411 Align = getDataLayout()->getTypeAllocSize(C->getType());
6412 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6414 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6415 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6416 .addReg(VReg1, RegState::Define)
6417 .addConstantPoolIndex(Idx));
6418 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6423 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6428 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6429 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6430 .addReg(ARM::CPSR, RegState::Define)
6434 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6435 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6436 .addJumpTableIndex(MJTI)
6439 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6440 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6441 .addReg(ARM::CPSR, RegState::Define)
6442 .addReg(NewVReg2, RegState::Kill)
6445 MachineMemOperand *JTMMOLd =
6446 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6447 MachineMemOperand::MOLoad, 4, 4);
6449 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6450 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6451 .addReg(NewVReg4, RegState::Kill)
6453 .addMemOperand(JTMMOLd));
6455 unsigned NewVReg6 = NewVReg5;
6456 if (RelocM == Reloc::PIC_) {
6457 NewVReg6 = MRI->createVirtualRegister(TRC);
6458 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6459 .addReg(ARM::CPSR, RegState::Define)
6460 .addReg(NewVReg5, RegState::Kill)
6464 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6465 .addReg(NewVReg6, RegState::Kill)
6466 .addJumpTableIndex(MJTI)
6469 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6470 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6473 .addMemOperand(FIMMOLd));
6475 if (NumLPads < 256) {
6476 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6479 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6480 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6481 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6482 .addImm(NumLPads & 0xFFFF));
6484 unsigned VReg2 = VReg1;
6485 if ((NumLPads & 0xFFFF0000) != 0) {
6486 VReg2 = MRI->createVirtualRegister(TRC);
6487 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6489 .addImm(NumLPads >> 16));
6492 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6496 MachineConstantPool *ConstantPool = MF->getConstantPool();
6497 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6498 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6500 // MachineConstantPool wants an explicit alignment.
6501 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6503 Align = getDataLayout()->getTypeAllocSize(C->getType());
6504 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6506 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6507 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6508 .addReg(VReg1, RegState::Define)
6509 .addConstantPoolIndex(Idx)
6511 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6513 .addReg(VReg1, RegState::Kill));
6516 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6521 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6523 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6525 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6526 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6527 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6528 .addJumpTableIndex(MJTI)
6531 MachineMemOperand *JTMMOLd =
6532 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6533 MachineMemOperand::MOLoad, 4, 4);
6534 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6536 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6537 .addReg(NewVReg3, RegState::Kill)
6540 .addMemOperand(JTMMOLd));
6542 if (RelocM == Reloc::PIC_) {
6543 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6544 .addReg(NewVReg5, RegState::Kill)
6546 .addJumpTableIndex(MJTI)
6549 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6550 .addReg(NewVReg5, RegState::Kill)
6551 .addJumpTableIndex(MJTI)
6556 // Add the jump table entries as successors to the MBB.
6557 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6558 for (std::vector<MachineBasicBlock*>::iterator
6559 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6560 MachineBasicBlock *CurMBB = *I;
6561 if (SeenMBBs.insert(CurMBB))
6562 DispContBB->addSuccessor(CurMBB);
6565 // N.B. the order the invoke BBs are processed in doesn't matter here.
6566 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6567 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6568 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6569 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6570 MachineBasicBlock *BB = *I;
6572 // Remove the landing pad successor from the invoke block and replace it
6573 // with the new dispatch block.
6574 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6576 while (!Successors.empty()) {
6577 MachineBasicBlock *SMBB = Successors.pop_back_val();
6578 if (SMBB->isLandingPad()) {
6579 BB->removeSuccessor(SMBB);
6580 MBBLPads.push_back(SMBB);
6584 BB->addSuccessor(DispatchBB);
6586 // Find the invoke call and mark all of the callee-saved registers as
6587 // 'implicit defined' so that they're spilled. This prevents code from
6588 // moving instructions to before the EH block, where they will never be
6590 for (MachineBasicBlock::reverse_iterator
6591 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6592 if (!II->isCall()) continue;
6594 DenseMap<unsigned, bool> DefRegs;
6595 for (MachineInstr::mop_iterator
6596 OI = II->operands_begin(), OE = II->operands_end();
6598 if (!OI->isReg()) continue;
6599 DefRegs[OI->getReg()] = true;
6602 MachineInstrBuilder MIB(*MF, &*II);
6604 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6605 unsigned Reg = SavedRegs[i];
6606 if (Subtarget->isThumb2() &&
6607 !ARM::tGPRRegClass.contains(Reg) &&
6608 !ARM::hGPRRegClass.contains(Reg))
6610 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6612 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6615 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6622 // Mark all former landing pads as non-landing pads. The dispatch is the only
6624 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6625 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6626 (*I)->setIsLandingPad(false);
6628 // The instruction is gone now.
6629 MI->eraseFromParent();
6635 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6636 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6637 E = MBB->succ_end(); I != E; ++I)
6640 llvm_unreachable("Expecting a BB with two successors!");
6643 /// Return the load opcode for a given load size. If load size >= 8,
6644 /// neon opcode will be returned.
6645 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6647 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6648 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6650 return LdSize == 4 ? ARM::tLDRi
6651 : LdSize == 2 ? ARM::tLDRHi
6652 : LdSize == 1 ? ARM::tLDRBi : 0;
6654 return LdSize == 4 ? ARM::t2LDR_POST
6655 : LdSize == 2 ? ARM::t2LDRH_POST
6656 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6657 return LdSize == 4 ? ARM::LDR_POST_IMM
6658 : LdSize == 2 ? ARM::LDRH_POST
6659 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6662 /// Return the store opcode for a given store size. If store size >= 8,
6663 /// neon opcode will be returned.
6664 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6666 return StSize == 16 ? ARM::VST1q32wb_fixed
6667 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6669 return StSize == 4 ? ARM::tSTRi
6670 : StSize == 2 ? ARM::tSTRHi
6671 : StSize == 1 ? ARM::tSTRBi : 0;
6673 return StSize == 4 ? ARM::t2STR_POST
6674 : StSize == 2 ? ARM::t2STRH_POST
6675 : StSize == 1 ? ARM::t2STRB_POST : 0;
6676 return StSize == 4 ? ARM::STR_POST_IMM
6677 : StSize == 2 ? ARM::STRH_POST
6678 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6681 /// Emit a post-increment load operation with given size. The instructions
6682 /// will be added to BB at Pos.
6683 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6684 const TargetInstrInfo *TII, DebugLoc dl,
6685 unsigned LdSize, unsigned Data, unsigned AddrIn,
6686 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6687 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6688 assert(LdOpc != 0 && "Should have a load opcode");
6690 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6691 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6693 } else if (IsThumb1) {
6694 // load + update AddrIn
6695 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6696 .addReg(AddrIn).addImm(0));
6697 MachineInstrBuilder MIB =
6698 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6699 MIB = AddDefaultT1CC(MIB);
6700 MIB.addReg(AddrIn).addImm(LdSize);
6701 AddDefaultPred(MIB);
6702 } else if (IsThumb2) {
6703 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6704 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6707 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6708 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6709 .addReg(0).addImm(LdSize));
6713 /// Emit a post-increment store operation with given size. The instructions
6714 /// will be added to BB at Pos.
6715 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6716 const TargetInstrInfo *TII, DebugLoc dl,
6717 unsigned StSize, unsigned Data, unsigned AddrIn,
6718 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6719 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6720 assert(StOpc != 0 && "Should have a store opcode");
6722 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6723 .addReg(AddrIn).addImm(0).addReg(Data));
6724 } else if (IsThumb1) {
6725 // store + update AddrIn
6726 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6727 .addReg(AddrIn).addImm(0));
6728 MachineInstrBuilder MIB =
6729 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6730 MIB = AddDefaultT1CC(MIB);
6731 MIB.addReg(AddrIn).addImm(StSize);
6732 AddDefaultPred(MIB);
6733 } else if (IsThumb2) {
6734 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6735 .addReg(Data).addReg(AddrIn).addImm(StSize));
6737 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6738 .addReg(Data).addReg(AddrIn).addReg(0)
6744 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6745 MachineBasicBlock *BB) const {
6746 // This pseudo instruction has 3 operands: dst, src, size
6747 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6748 // Otherwise, we will generate unrolled scalar copies.
6749 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6750 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6751 MachineFunction::iterator It = BB;
6754 unsigned dest = MI->getOperand(0).getReg();
6755 unsigned src = MI->getOperand(1).getReg();
6756 unsigned SizeVal = MI->getOperand(2).getImm();
6757 unsigned Align = MI->getOperand(3).getImm();
6758 DebugLoc dl = MI->getDebugLoc();
6760 MachineFunction *MF = BB->getParent();
6761 MachineRegisterInfo &MRI = MF->getRegInfo();
6762 unsigned UnitSize = 0;
6763 const TargetRegisterClass *TRC = 0;
6764 const TargetRegisterClass *VecTRC = 0;
6766 bool IsThumb1 = Subtarget->isThumb1Only();
6767 bool IsThumb2 = Subtarget->isThumb2();
6771 } else if (Align & 2) {
6774 // Check whether we can use NEON instructions.
6775 if (!MF->getFunction()->getAttributes().
6776 hasAttribute(AttributeSet::FunctionIndex,
6777 Attribute::NoImplicitFloat) &&
6778 Subtarget->hasNEON()) {
6779 if ((Align % 16 == 0) && SizeVal >= 16)
6781 else if ((Align % 8 == 0) && SizeVal >= 8)
6784 // Can't use NEON instructions.
6789 // Select the correct opcode and register class for unit size load/store
6790 bool IsNeon = UnitSize >= 8;
6791 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
6792 : (const TargetRegisterClass *)&ARM::GPRRegClass;
6794 VecTRC = UnitSize == 16
6795 ? (const TargetRegisterClass *)&ARM::DPairRegClass
6797 ? (const TargetRegisterClass *)&ARM::DPRRegClass
6800 unsigned BytesLeft = SizeVal % UnitSize;
6801 unsigned LoopSize = SizeVal - BytesLeft;
6803 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6804 // Use LDR and STR to copy.
6805 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6806 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6807 unsigned srcIn = src;
6808 unsigned destIn = dest;
6809 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6810 unsigned srcOut = MRI.createVirtualRegister(TRC);
6811 unsigned destOut = MRI.createVirtualRegister(TRC);
6812 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
6813 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
6814 IsThumb1, IsThumb2);
6815 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
6816 IsThumb1, IsThumb2);
6821 // Handle the leftover bytes with LDRB and STRB.
6822 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6823 // [destOut] = STRB_POST(scratch, destIn, 1)
6824 for (unsigned i = 0; i < BytesLeft; i++) {
6825 unsigned srcOut = MRI.createVirtualRegister(TRC);
6826 unsigned destOut = MRI.createVirtualRegister(TRC);
6827 unsigned scratch = MRI.createVirtualRegister(TRC);
6828 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
6829 IsThumb1, IsThumb2);
6830 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
6831 IsThumb1, IsThumb2);
6835 MI->eraseFromParent(); // The instruction is gone now.
6839 // Expand the pseudo op to a loop.
6842 // movw varEnd, # --> with thumb2
6844 // ldrcp varEnd, idx --> without thumb2
6845 // fallthrough --> loopMBB
6847 // PHI varPhi, varEnd, varLoop
6848 // PHI srcPhi, src, srcLoop
6849 // PHI destPhi, dst, destLoop
6850 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6851 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6852 // subs varLoop, varPhi, #UnitSize
6854 // fallthrough --> exitMBB
6856 // epilogue to handle left-over bytes
6857 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6858 // [destOut] = STRB_POST(scratch, destLoop, 1)
6859 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6860 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6861 MF->insert(It, loopMBB);
6862 MF->insert(It, exitMBB);
6864 // Transfer the remainder of BB and its successor edges to exitMBB.
6865 exitMBB->splice(exitMBB->begin(), BB,
6866 std::next(MachineBasicBlock::iterator(MI)), BB->end());
6867 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6869 // Load an immediate to varEnd.
6870 unsigned varEnd = MRI.createVirtualRegister(TRC);
6872 unsigned Vtmp = varEnd;
6873 if ((LoopSize & 0xFFFF0000) != 0)
6874 Vtmp = MRI.createVirtualRegister(TRC);
6875 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
6876 .addImm(LoopSize & 0xFFFF));
6878 if ((LoopSize & 0xFFFF0000) != 0)
6879 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6880 .addReg(Vtmp).addImm(LoopSize >> 16));
6882 MachineConstantPool *ConstantPool = MF->getConstantPool();
6883 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6884 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6886 // MachineConstantPool wants an explicit alignment.
6887 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6889 Align = getDataLayout()->getTypeAllocSize(C->getType());
6890 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6893 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
6894 varEnd, RegState::Define).addConstantPoolIndex(Idx));
6896 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
6897 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
6899 BB->addSuccessor(loopMBB);
6901 // Generate the loop body:
6902 // varPhi = PHI(varLoop, varEnd)
6903 // srcPhi = PHI(srcLoop, src)
6904 // destPhi = PHI(destLoop, dst)
6905 MachineBasicBlock *entryBB = BB;
6907 unsigned varLoop = MRI.createVirtualRegister(TRC);
6908 unsigned varPhi = MRI.createVirtualRegister(TRC);
6909 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6910 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6911 unsigned destLoop = MRI.createVirtualRegister(TRC);
6912 unsigned destPhi = MRI.createVirtualRegister(TRC);
6914 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6915 .addReg(varLoop).addMBB(loopMBB)
6916 .addReg(varEnd).addMBB(entryBB);
6917 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6918 .addReg(srcLoop).addMBB(loopMBB)
6919 .addReg(src).addMBB(entryBB);
6920 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6921 .addReg(destLoop).addMBB(loopMBB)
6922 .addReg(dest).addMBB(entryBB);
6924 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6925 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6926 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
6927 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
6928 IsThumb1, IsThumb2);
6929 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
6930 IsThumb1, IsThumb2);
6932 // Decrement loop variable by UnitSize.
6934 MachineInstrBuilder MIB =
6935 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
6936 MIB = AddDefaultT1CC(MIB);
6937 MIB.addReg(varPhi).addImm(UnitSize);
6938 AddDefaultPred(MIB);
6940 MachineInstrBuilder MIB =
6941 BuildMI(*BB, BB->end(), dl,
6942 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6943 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6944 MIB->getOperand(5).setReg(ARM::CPSR);
6945 MIB->getOperand(5).setIsDef(true);
6947 BuildMI(*BB, BB->end(), dl,
6948 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
6949 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6951 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6952 BB->addSuccessor(loopMBB);
6953 BB->addSuccessor(exitMBB);
6955 // Add epilogue to handle BytesLeft.
6957 MachineInstr *StartOfExit = exitMBB->begin();
6959 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6960 // [destOut] = STRB_POST(scratch, destLoop, 1)
6961 unsigned srcIn = srcLoop;
6962 unsigned destIn = destLoop;
6963 for (unsigned i = 0; i < BytesLeft; i++) {
6964 unsigned srcOut = MRI.createVirtualRegister(TRC);
6965 unsigned destOut = MRI.createVirtualRegister(TRC);
6966 unsigned scratch = MRI.createVirtualRegister(TRC);
6967 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
6968 IsThumb1, IsThumb2);
6969 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
6970 IsThumb1, IsThumb2);
6975 MI->eraseFromParent(); // The instruction is gone now.
6980 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6981 MachineBasicBlock *BB) const {
6982 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6983 DebugLoc dl = MI->getDebugLoc();
6984 bool isThumb2 = Subtarget->isThumb2();
6985 switch (MI->getOpcode()) {
6988 llvm_unreachable("Unexpected instr type to insert");
6990 // The Thumb2 pre-indexed stores have the same MI operands, they just
6991 // define them differently in the .td files from the isel patterns, so
6992 // they need pseudos.
6993 case ARM::t2STR_preidx:
6994 MI->setDesc(TII->get(ARM::t2STR_PRE));
6996 case ARM::t2STRB_preidx:
6997 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6999 case ARM::t2STRH_preidx:
7000 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7003 case ARM::STRi_preidx:
7004 case ARM::STRBi_preidx: {
7005 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7006 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7007 // Decode the offset.
7008 unsigned Offset = MI->getOperand(4).getImm();
7009 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7010 Offset = ARM_AM::getAM2Offset(Offset);
7014 MachineMemOperand *MMO = *MI->memoperands_begin();
7015 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7016 .addOperand(MI->getOperand(0)) // Rn_wb
7017 .addOperand(MI->getOperand(1)) // Rt
7018 .addOperand(MI->getOperand(2)) // Rn
7019 .addImm(Offset) // offset (skip GPR==zero_reg)
7020 .addOperand(MI->getOperand(5)) // pred
7021 .addOperand(MI->getOperand(6))
7022 .addMemOperand(MMO);
7023 MI->eraseFromParent();
7026 case ARM::STRr_preidx:
7027 case ARM::STRBr_preidx:
7028 case ARM::STRH_preidx: {
7030 switch (MI->getOpcode()) {
7031 default: llvm_unreachable("unexpected opcode!");
7032 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7033 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7034 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7036 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7037 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7038 MIB.addOperand(MI->getOperand(i));
7039 MI->eraseFromParent();
7043 case ARM::tMOVCCr_pseudo: {
7044 // To "insert" a SELECT_CC instruction, we actually have to insert the
7045 // diamond control-flow pattern. The incoming instruction knows the
7046 // destination vreg to set, the condition code register to branch on, the
7047 // true/false values to select between, and a branch opcode to use.
7048 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7049 MachineFunction::iterator It = BB;
7055 // cmpTY ccX, r1, r2
7057 // fallthrough --> copy0MBB
7058 MachineBasicBlock *thisMBB = BB;
7059 MachineFunction *F = BB->getParent();
7060 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7061 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7062 F->insert(It, copy0MBB);
7063 F->insert(It, sinkMBB);
7065 // Transfer the remainder of BB and its successor edges to sinkMBB.
7066 sinkMBB->splice(sinkMBB->begin(), BB,
7067 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7068 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7070 BB->addSuccessor(copy0MBB);
7071 BB->addSuccessor(sinkMBB);
7073 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7074 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7077 // %FalseValue = ...
7078 // # fallthrough to sinkMBB
7081 // Update machine-CFG edges
7082 BB->addSuccessor(sinkMBB);
7085 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7088 BuildMI(*BB, BB->begin(), dl,
7089 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7090 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7091 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7093 MI->eraseFromParent(); // The pseudo instruction is gone now.
7098 case ARM::BCCZi64: {
7099 // If there is an unconditional branch to the other successor, remove it.
7100 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7102 // Compare both parts that make up the double comparison separately for
7104 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7106 unsigned LHS1 = MI->getOperand(1).getReg();
7107 unsigned LHS2 = MI->getOperand(2).getReg();
7109 AddDefaultPred(BuildMI(BB, dl,
7110 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7111 .addReg(LHS1).addImm(0));
7112 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7113 .addReg(LHS2).addImm(0)
7114 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7116 unsigned RHS1 = MI->getOperand(3).getReg();
7117 unsigned RHS2 = MI->getOperand(4).getReg();
7118 AddDefaultPred(BuildMI(BB, dl,
7119 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7120 .addReg(LHS1).addReg(RHS1));
7121 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7122 .addReg(LHS2).addReg(RHS2)
7123 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7126 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7127 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7128 if (MI->getOperand(0).getImm() == ARMCC::NE)
7129 std::swap(destMBB, exitMBB);
7131 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7132 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7134 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7136 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7138 MI->eraseFromParent(); // The pseudo instruction is gone now.
7142 case ARM::Int_eh_sjlj_setjmp:
7143 case ARM::Int_eh_sjlj_setjmp_nofp:
7144 case ARM::tInt_eh_sjlj_setjmp:
7145 case ARM::t2Int_eh_sjlj_setjmp:
7146 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7147 EmitSjLjDispatchBlock(MI, BB);
7152 // To insert an ABS instruction, we have to insert the
7153 // diamond control-flow pattern. The incoming instruction knows the
7154 // source vreg to test against 0, the destination vreg to set,
7155 // the condition code register to branch on, the
7156 // true/false values to select between, and a branch opcode to use.
7161 // BCC (branch to SinkBB if V0 >= 0)
7162 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7163 // SinkBB: V1 = PHI(V2, V3)
7164 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7165 MachineFunction::iterator BBI = BB;
7167 MachineFunction *Fn = BB->getParent();
7168 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7169 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7170 Fn->insert(BBI, RSBBB);
7171 Fn->insert(BBI, SinkBB);
7173 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7174 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7175 bool isThumb2 = Subtarget->isThumb2();
7176 MachineRegisterInfo &MRI = Fn->getRegInfo();
7177 // In Thumb mode S must not be specified if source register is the SP or
7178 // PC and if destination register is the SP, so restrict register class
7179 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7180 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7181 (const TargetRegisterClass*)&ARM::GPRRegClass);
7183 // Transfer the remainder of BB and its successor edges to sinkMBB.
7184 SinkBB->splice(SinkBB->begin(), BB,
7185 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7186 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7188 BB->addSuccessor(RSBBB);
7189 BB->addSuccessor(SinkBB);
7191 // fall through to SinkMBB
7192 RSBBB->addSuccessor(SinkBB);
7194 // insert a cmp at the end of BB
7195 AddDefaultPred(BuildMI(BB, dl,
7196 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7197 .addReg(ABSSrcReg).addImm(0));
7199 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7201 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7202 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7204 // insert rsbri in RSBBB
7205 // Note: BCC and rsbri will be converted into predicated rsbmi
7206 // by if-conversion pass
7207 BuildMI(*RSBBB, RSBBB->begin(), dl,
7208 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7209 .addReg(ABSSrcReg, RegState::Kill)
7210 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7212 // insert PHI in SinkBB,
7213 // reuse ABSDstReg to not change uses of ABS instruction
7214 BuildMI(*SinkBB, SinkBB->begin(), dl,
7215 TII->get(ARM::PHI), ABSDstReg)
7216 .addReg(NewRsbDstReg).addMBB(RSBBB)
7217 .addReg(ABSSrcReg).addMBB(BB);
7219 // remove ABS instruction
7220 MI->eraseFromParent();
7222 // return last added BB
7225 case ARM::COPY_STRUCT_BYVAL_I32:
7227 return EmitStructByval(MI, BB);
7231 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7232 SDNode *Node) const {
7233 if (!MI->hasPostISelHook()) {
7234 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7235 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7239 const MCInstrDesc *MCID = &MI->getDesc();
7240 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7241 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7242 // operand is still set to noreg. If needed, set the optional operand's
7243 // register to CPSR, and remove the redundant implicit def.
7245 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7247 // Rename pseudo opcodes.
7248 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7250 const ARMBaseInstrInfo *TII =
7251 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7252 MCID = &TII->get(NewOpc);
7254 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7255 "converted opcode should be the same except for cc_out");
7259 // Add the optional cc_out operand
7260 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7262 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7264 // Any ARM instruction that sets the 's' bit should specify an optional
7265 // "cc_out" operand in the last operand position.
7266 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7267 assert(!NewOpc && "Optional cc_out operand required");
7270 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7271 // since we already have an optional CPSR def.
7272 bool definesCPSR = false;
7273 bool deadCPSR = false;
7274 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7276 const MachineOperand &MO = MI->getOperand(i);
7277 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7281 MI->RemoveOperand(i);
7286 assert(!NewOpc && "Optional cc_out operand required");
7289 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7291 assert(!MI->getOperand(ccOutIdx).getReg() &&
7292 "expect uninitialized optional cc_out operand");
7296 // If this instruction was defined with an optional CPSR def and its dag node
7297 // had a live implicit CPSR def, then activate the optional CPSR def.
7298 MachineOperand &MO = MI->getOperand(ccOutIdx);
7299 MO.setReg(ARM::CPSR);
7303 //===----------------------------------------------------------------------===//
7304 // ARM Optimization Hooks
7305 //===----------------------------------------------------------------------===//
7307 // Helper function that checks if N is a null or all ones constant.
7308 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7309 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7312 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7315 // Return true if N is conditionally 0 or all ones.
7316 // Detects these expressions where cc is an i1 value:
7318 // (select cc 0, y) [AllOnes=0]
7319 // (select cc y, 0) [AllOnes=0]
7320 // (zext cc) [AllOnes=0]
7321 // (sext cc) [AllOnes=0/1]
7322 // (select cc -1, y) [AllOnes=1]
7323 // (select cc y, -1) [AllOnes=1]
7325 // Invert is set when N is the null/all ones constant when CC is false.
7326 // OtherOp is set to the alternative value of N.
7327 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7328 SDValue &CC, bool &Invert,
7330 SelectionDAG &DAG) {
7331 switch (N->getOpcode()) {
7332 default: return false;
7334 CC = N->getOperand(0);
7335 SDValue N1 = N->getOperand(1);
7336 SDValue N2 = N->getOperand(2);
7337 if (isZeroOrAllOnes(N1, AllOnes)) {
7342 if (isZeroOrAllOnes(N2, AllOnes)) {
7349 case ISD::ZERO_EXTEND:
7350 // (zext cc) can never be the all ones value.
7354 case ISD::SIGN_EXTEND: {
7355 EVT VT = N->getValueType(0);
7356 CC = N->getOperand(0);
7357 if (CC.getValueType() != MVT::i1)
7361 // When looking for an AllOnes constant, N is an sext, and the 'other'
7363 OtherOp = DAG.getConstant(0, VT);
7364 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7365 // When looking for a 0 constant, N can be zext or sext.
7366 OtherOp = DAG.getConstant(1, VT);
7368 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7374 // Combine a constant select operand into its use:
7376 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7377 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7378 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7379 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7380 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7382 // The transform is rejected if the select doesn't have a constant operand that
7383 // is null, or all ones when AllOnes is set.
7385 // Also recognize sext/zext from i1:
7387 // (add (zext cc), x) -> (select cc (add x, 1), x)
7388 // (add (sext cc), x) -> (select cc (add x, -1), x)
7390 // These transformations eventually create predicated instructions.
7392 // @param N The node to transform.
7393 // @param Slct The N operand that is a select.
7394 // @param OtherOp The other N operand (x above).
7395 // @param DCI Context.
7396 // @param AllOnes Require the select constant to be all ones instead of null.
7397 // @returns The new node, or SDValue() on failure.
7399 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7400 TargetLowering::DAGCombinerInfo &DCI,
7401 bool AllOnes = false) {
7402 SelectionDAG &DAG = DCI.DAG;
7403 EVT VT = N->getValueType(0);
7404 SDValue NonConstantVal;
7407 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7408 NonConstantVal, DAG))
7411 // Slct is now know to be the desired identity constant when CC is true.
7412 SDValue TrueVal = OtherOp;
7413 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7414 OtherOp, NonConstantVal);
7415 // Unless SwapSelectOps says CC should be false.
7417 std::swap(TrueVal, FalseVal);
7419 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7420 CCOp, TrueVal, FalseVal);
7423 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7425 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7426 TargetLowering::DAGCombinerInfo &DCI) {
7427 SDValue N0 = N->getOperand(0);
7428 SDValue N1 = N->getOperand(1);
7429 if (N0.getNode()->hasOneUse()) {
7430 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7431 if (Result.getNode())
7434 if (N1.getNode()->hasOneUse()) {
7435 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7436 if (Result.getNode())
7442 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7443 // (only after legalization).
7444 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7445 TargetLowering::DAGCombinerInfo &DCI,
7446 const ARMSubtarget *Subtarget) {
7448 // Only perform optimization if after legalize, and if NEON is available. We
7449 // also expected both operands to be BUILD_VECTORs.
7450 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7451 || N0.getOpcode() != ISD::BUILD_VECTOR
7452 || N1.getOpcode() != ISD::BUILD_VECTOR)
7455 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7456 EVT VT = N->getValueType(0);
7457 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7460 // Check that the vector operands are of the right form.
7461 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7462 // operands, where N is the size of the formed vector.
7463 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7464 // index such that we have a pair wise add pattern.
7466 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7467 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7469 SDValue Vec = N0->getOperand(0)->getOperand(0);
7470 SDNode *V = Vec.getNode();
7471 unsigned nextIndex = 0;
7473 // For each operands to the ADD which are BUILD_VECTORs,
7474 // check to see if each of their operands are an EXTRACT_VECTOR with
7475 // the same vector and appropriate index.
7476 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7477 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7478 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7480 SDValue ExtVec0 = N0->getOperand(i);
7481 SDValue ExtVec1 = N1->getOperand(i);
7483 // First operand is the vector, verify its the same.
7484 if (V != ExtVec0->getOperand(0).getNode() ||
7485 V != ExtVec1->getOperand(0).getNode())
7488 // Second is the constant, verify its correct.
7489 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7490 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7492 // For the constant, we want to see all the even or all the odd.
7493 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7494 || C1->getZExtValue() != nextIndex+1)
7503 // Create VPADDL node.
7504 SelectionDAG &DAG = DCI.DAG;
7505 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7507 // Build operand list.
7508 SmallVector<SDValue, 8> Ops;
7509 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7510 TLI.getPointerTy()));
7512 // Input is the vector.
7515 // Get widened type and narrowed type.
7517 unsigned numElem = VT.getVectorNumElements();
7519 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7520 switch (inputLaneType.getSimpleVT().SimpleTy) {
7521 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7522 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7523 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7525 llvm_unreachable("Invalid vector element type for padd optimization.");
7528 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
7529 widenType, &Ops[0], Ops.size());
7530 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7531 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7534 static SDValue findMUL_LOHI(SDValue V) {
7535 if (V->getOpcode() == ISD::UMUL_LOHI ||
7536 V->getOpcode() == ISD::SMUL_LOHI)
7541 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7542 TargetLowering::DAGCombinerInfo &DCI,
7543 const ARMSubtarget *Subtarget) {
7545 if (Subtarget->isThumb1Only()) return SDValue();
7547 // Only perform the checks after legalize when the pattern is available.
7548 if (DCI.isBeforeLegalize()) return SDValue();
7550 // Look for multiply add opportunities.
7551 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7552 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7553 // a glue link from the first add to the second add.
7554 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7555 // a S/UMLAL instruction.
7558 // \ / \ [no multiline comment]
7564 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7565 SDValue AddcOp0 = AddcNode->getOperand(0);
7566 SDValue AddcOp1 = AddcNode->getOperand(1);
7568 // Check if the two operands are from the same mul_lohi node.
7569 if (AddcOp0.getNode() == AddcOp1.getNode())
7572 assert(AddcNode->getNumValues() == 2 &&
7573 AddcNode->getValueType(0) == MVT::i32 &&
7574 "Expect ADDC with two result values. First: i32");
7576 // Check that we have a glued ADDC node.
7577 if (AddcNode->getValueType(1) != MVT::Glue)
7580 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7581 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7582 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7583 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7584 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7587 // Look for the glued ADDE.
7588 SDNode* AddeNode = AddcNode->getGluedUser();
7589 if (AddeNode == NULL)
7592 // Make sure it is really an ADDE.
7593 if (AddeNode->getOpcode() != ISD::ADDE)
7596 assert(AddeNode->getNumOperands() == 3 &&
7597 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7598 "ADDE node has the wrong inputs");
7600 // Check for the triangle shape.
7601 SDValue AddeOp0 = AddeNode->getOperand(0);
7602 SDValue AddeOp1 = AddeNode->getOperand(1);
7604 // Make sure that the ADDE operands are not coming from the same node.
7605 if (AddeOp0.getNode() == AddeOp1.getNode())
7608 // Find the MUL_LOHI node walking up ADDE's operands.
7609 bool IsLeftOperandMUL = false;
7610 SDValue MULOp = findMUL_LOHI(AddeOp0);
7611 if (MULOp == SDValue())
7612 MULOp = findMUL_LOHI(AddeOp1);
7614 IsLeftOperandMUL = true;
7615 if (MULOp == SDValue())
7618 // Figure out the right opcode.
7619 unsigned Opc = MULOp->getOpcode();
7620 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7622 // Figure out the high and low input values to the MLAL node.
7623 SDValue* HiMul = &MULOp;
7624 SDValue* HiAdd = NULL;
7625 SDValue* LoMul = NULL;
7626 SDValue* LowAdd = NULL;
7628 if (IsLeftOperandMUL)
7634 if (AddcOp0->getOpcode() == Opc) {
7638 if (AddcOp1->getOpcode() == Opc) {
7646 if (LoMul->getNode() != HiMul->getNode())
7649 // Create the merged node.
7650 SelectionDAG &DAG = DCI.DAG;
7652 // Build operand list.
7653 SmallVector<SDValue, 8> Ops;
7654 Ops.push_back(LoMul->getOperand(0));
7655 Ops.push_back(LoMul->getOperand(1));
7656 Ops.push_back(*LowAdd);
7657 Ops.push_back(*HiAdd);
7659 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
7660 DAG.getVTList(MVT::i32, MVT::i32),
7661 &Ops[0], Ops.size());
7663 // Replace the ADDs' nodes uses by the MLA node's values.
7664 SDValue HiMLALResult(MLALNode.getNode(), 1);
7665 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7667 SDValue LoMLALResult(MLALNode.getNode(), 0);
7668 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7670 // Return original node to notify the driver to stop replacing.
7671 SDValue resNode(AddcNode, 0);
7675 /// PerformADDCCombine - Target-specific dag combine transform from
7676 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7677 static SDValue PerformADDCCombine(SDNode *N,
7678 TargetLowering::DAGCombinerInfo &DCI,
7679 const ARMSubtarget *Subtarget) {
7681 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7685 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7686 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7687 /// called with the default operands, and if that fails, with commuted
7689 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7690 TargetLowering::DAGCombinerInfo &DCI,
7691 const ARMSubtarget *Subtarget){
7693 // Attempt to create vpaddl for this add.
7694 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7695 if (Result.getNode())
7698 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7699 if (N0.getNode()->hasOneUse()) {
7700 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7701 if (Result.getNode()) return Result;
7706 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7708 static SDValue PerformADDCombine(SDNode *N,
7709 TargetLowering::DAGCombinerInfo &DCI,
7710 const ARMSubtarget *Subtarget) {
7711 SDValue N0 = N->getOperand(0);
7712 SDValue N1 = N->getOperand(1);
7714 // First try with the default operand order.
7715 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7716 if (Result.getNode())
7719 // If that didn't work, try again with the operands commuted.
7720 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7723 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7725 static SDValue PerformSUBCombine(SDNode *N,
7726 TargetLowering::DAGCombinerInfo &DCI) {
7727 SDValue N0 = N->getOperand(0);
7728 SDValue N1 = N->getOperand(1);
7730 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7731 if (N1.getNode()->hasOneUse()) {
7732 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7733 if (Result.getNode()) return Result;
7739 /// PerformVMULCombine
7740 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7741 /// special multiplier accumulator forwarding.
7747 // However, for (A + B) * (A + B),
7754 static SDValue PerformVMULCombine(SDNode *N,
7755 TargetLowering::DAGCombinerInfo &DCI,
7756 const ARMSubtarget *Subtarget) {
7757 if (!Subtarget->hasVMLxForwarding())
7760 SelectionDAG &DAG = DCI.DAG;
7761 SDValue N0 = N->getOperand(0);
7762 SDValue N1 = N->getOperand(1);
7763 unsigned Opcode = N0.getOpcode();
7764 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7765 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7766 Opcode = N1.getOpcode();
7767 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7768 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7776 EVT VT = N->getValueType(0);
7778 SDValue N00 = N0->getOperand(0);
7779 SDValue N01 = N0->getOperand(1);
7780 return DAG.getNode(Opcode, DL, VT,
7781 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7782 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7785 static SDValue PerformMULCombine(SDNode *N,
7786 TargetLowering::DAGCombinerInfo &DCI,
7787 const ARMSubtarget *Subtarget) {
7788 SelectionDAG &DAG = DCI.DAG;
7790 if (Subtarget->isThumb1Only())
7793 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7796 EVT VT = N->getValueType(0);
7797 if (VT.is64BitVector() || VT.is128BitVector())
7798 return PerformVMULCombine(N, DCI, Subtarget);
7802 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7806 int64_t MulAmt = C->getSExtValue();
7807 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
7809 ShiftAmt = ShiftAmt & (32 - 1);
7810 SDValue V = N->getOperand(0);
7814 MulAmt >>= ShiftAmt;
7817 if (isPowerOf2_32(MulAmt - 1)) {
7818 // (mul x, 2^N + 1) => (add (shl x, N), x)
7819 Res = DAG.getNode(ISD::ADD, DL, VT,
7821 DAG.getNode(ISD::SHL, DL, VT,
7823 DAG.getConstant(Log2_32(MulAmt - 1),
7825 } else if (isPowerOf2_32(MulAmt + 1)) {
7826 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7827 Res = DAG.getNode(ISD::SUB, DL, VT,
7828 DAG.getNode(ISD::SHL, DL, VT,
7830 DAG.getConstant(Log2_32(MulAmt + 1),
7836 uint64_t MulAmtAbs = -MulAmt;
7837 if (isPowerOf2_32(MulAmtAbs + 1)) {
7838 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7839 Res = DAG.getNode(ISD::SUB, DL, VT,
7841 DAG.getNode(ISD::SHL, DL, VT,
7843 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7845 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7846 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7847 Res = DAG.getNode(ISD::ADD, DL, VT,
7849 DAG.getNode(ISD::SHL, DL, VT,
7851 DAG.getConstant(Log2_32(MulAmtAbs-1),
7853 Res = DAG.getNode(ISD::SUB, DL, VT,
7854 DAG.getConstant(0, MVT::i32),Res);
7861 Res = DAG.getNode(ISD::SHL, DL, VT,
7862 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7864 // Do not add new nodes to DAG combiner worklist.
7865 DCI.CombineTo(N, Res, false);
7869 static SDValue PerformANDCombine(SDNode *N,
7870 TargetLowering::DAGCombinerInfo &DCI,
7871 const ARMSubtarget *Subtarget) {
7873 // Attempt to use immediate-form VBIC
7874 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7876 EVT VT = N->getValueType(0);
7877 SelectionDAG &DAG = DCI.DAG;
7879 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7882 APInt SplatBits, SplatUndef;
7883 unsigned SplatBitSize;
7886 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7887 if (SplatBitSize <= 64) {
7889 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7890 SplatUndef.getZExtValue(), SplatBitSize,
7891 DAG, VbicVT, VT.is128BitVector(),
7893 if (Val.getNode()) {
7895 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7896 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7897 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7902 if (!Subtarget->isThumb1Only()) {
7903 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7904 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7905 if (Result.getNode())
7912 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7913 static SDValue PerformORCombine(SDNode *N,
7914 TargetLowering::DAGCombinerInfo &DCI,
7915 const ARMSubtarget *Subtarget) {
7916 // Attempt to use immediate-form VORR
7917 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7919 EVT VT = N->getValueType(0);
7920 SelectionDAG &DAG = DCI.DAG;
7922 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7925 APInt SplatBits, SplatUndef;
7926 unsigned SplatBitSize;
7928 if (BVN && Subtarget->hasNEON() &&
7929 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7930 if (SplatBitSize <= 64) {
7932 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7933 SplatUndef.getZExtValue(), SplatBitSize,
7934 DAG, VorrVT, VT.is128BitVector(),
7936 if (Val.getNode()) {
7938 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7939 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7940 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7945 if (!Subtarget->isThumb1Only()) {
7946 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7947 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7948 if (Result.getNode())
7952 // The code below optimizes (or (and X, Y), Z).
7953 // The AND operand needs to have a single user to make these optimizations
7955 SDValue N0 = N->getOperand(0);
7956 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7958 SDValue N1 = N->getOperand(1);
7960 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7961 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7962 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7964 unsigned SplatBitSize;
7967 APInt SplatBits0, SplatBits1;
7968 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7969 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7970 // Ensure that the second operand of both ands are constants
7971 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7972 HasAnyUndefs) && !HasAnyUndefs) {
7973 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7974 HasAnyUndefs) && !HasAnyUndefs) {
7975 // Ensure that the bit width of the constants are the same and that
7976 // the splat arguments are logical inverses as per the pattern we
7977 // are trying to simplify.
7978 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
7979 SplatBits0 == ~SplatBits1) {
7980 // Canonicalize the vector type to make instruction selection
7982 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7983 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7987 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7993 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7996 // BFI is only available on V6T2+
7997 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8001 // 1) or (and A, mask), val => ARMbfi A, val, mask
8002 // iff (val & mask) == val
8004 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8005 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8006 // && mask == ~mask2
8007 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8008 // && ~mask == mask2
8009 // (i.e., copy a bitfield value into another bitfield of the same width)
8014 SDValue N00 = N0.getOperand(0);
8016 // The value and the mask need to be constants so we can verify this is
8017 // actually a bitfield set. If the mask is 0xffff, we can do better
8018 // via a movt instruction, so don't use BFI in that case.
8019 SDValue MaskOp = N0.getOperand(1);
8020 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8023 unsigned Mask = MaskC->getZExtValue();
8027 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8028 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8030 unsigned Val = N1C->getZExtValue();
8031 if ((Val & ~Mask) != Val)
8034 if (ARM::isBitFieldInvertedMask(Mask)) {
8035 Val >>= countTrailingZeros(~Mask);
8037 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8038 DAG.getConstant(Val, MVT::i32),
8039 DAG.getConstant(Mask, MVT::i32));
8041 // Do not add new nodes to DAG combiner worklist.
8042 DCI.CombineTo(N, Res, false);
8045 } else if (N1.getOpcode() == ISD::AND) {
8046 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8047 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8050 unsigned Mask2 = N11C->getZExtValue();
8052 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8054 if (ARM::isBitFieldInvertedMask(Mask) &&
8056 // The pack halfword instruction works better for masks that fit it,
8057 // so use that when it's available.
8058 if (Subtarget->hasT2ExtractPack() &&
8059 (Mask == 0xffff || Mask == 0xffff0000))
8062 unsigned amt = countTrailingZeros(Mask2);
8063 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8064 DAG.getConstant(amt, MVT::i32));
8065 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8066 DAG.getConstant(Mask, MVT::i32));
8067 // Do not add new nodes to DAG combiner worklist.
8068 DCI.CombineTo(N, Res, false);
8070 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8072 // The pack halfword instruction works better for masks that fit it,
8073 // so use that when it's available.
8074 if (Subtarget->hasT2ExtractPack() &&
8075 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8078 unsigned lsb = countTrailingZeros(Mask);
8079 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8080 DAG.getConstant(lsb, MVT::i32));
8081 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8082 DAG.getConstant(Mask2, MVT::i32));
8083 // Do not add new nodes to DAG combiner worklist.
8084 DCI.CombineTo(N, Res, false);
8089 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8090 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8091 ARM::isBitFieldInvertedMask(~Mask)) {
8092 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8093 // where lsb(mask) == #shamt and masked bits of B are known zero.
8094 SDValue ShAmt = N00.getOperand(1);
8095 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8096 unsigned LSB = countTrailingZeros(Mask);
8100 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8101 DAG.getConstant(~Mask, MVT::i32));
8103 // Do not add new nodes to DAG combiner worklist.
8104 DCI.CombineTo(N, Res, false);
8110 static SDValue PerformXORCombine(SDNode *N,
8111 TargetLowering::DAGCombinerInfo &DCI,
8112 const ARMSubtarget *Subtarget) {
8113 EVT VT = N->getValueType(0);
8114 SelectionDAG &DAG = DCI.DAG;
8116 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8119 if (!Subtarget->isThumb1Only()) {
8120 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8121 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8122 if (Result.getNode())
8129 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8130 /// the bits being cleared by the AND are not demanded by the BFI.
8131 static SDValue PerformBFICombine(SDNode *N,
8132 TargetLowering::DAGCombinerInfo &DCI) {
8133 SDValue N1 = N->getOperand(1);
8134 if (N1.getOpcode() == ISD::AND) {
8135 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8138 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8139 unsigned LSB = countTrailingZeros(~InvMask);
8140 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8141 unsigned Mask = (1 << Width)-1;
8142 unsigned Mask2 = N11C->getZExtValue();
8143 if ((Mask & (~Mask2)) == 0)
8144 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8145 N->getOperand(0), N1.getOperand(0),
8151 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8152 /// ARMISD::VMOVRRD.
8153 static SDValue PerformVMOVRRDCombine(SDNode *N,
8154 TargetLowering::DAGCombinerInfo &DCI) {
8155 // vmovrrd(vmovdrr x, y) -> x,y
8156 SDValue InDouble = N->getOperand(0);
8157 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8158 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8160 // vmovrrd(load f64) -> (load i32), (load i32)
8161 SDNode *InNode = InDouble.getNode();
8162 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8163 InNode->getValueType(0) == MVT::f64 &&
8164 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8165 !cast<LoadSDNode>(InNode)->isVolatile()) {
8166 // TODO: Should this be done for non-FrameIndex operands?
8167 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8169 SelectionDAG &DAG = DCI.DAG;
8171 SDValue BasePtr = LD->getBasePtr();
8172 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8173 LD->getPointerInfo(), LD->isVolatile(),
8174 LD->isNonTemporal(), LD->isInvariant(),
8175 LD->getAlignment());
8177 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8178 DAG.getConstant(4, MVT::i32));
8179 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8180 LD->getPointerInfo(), LD->isVolatile(),
8181 LD->isNonTemporal(), LD->isInvariant(),
8182 std::min(4U, LD->getAlignment() / 2));
8184 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8185 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8186 DCI.RemoveFromWorklist(LD);
8194 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8195 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8196 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8197 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8198 SDValue Op0 = N->getOperand(0);
8199 SDValue Op1 = N->getOperand(1);
8200 if (Op0.getOpcode() == ISD::BITCAST)
8201 Op0 = Op0.getOperand(0);
8202 if (Op1.getOpcode() == ISD::BITCAST)
8203 Op1 = Op1.getOperand(0);
8204 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8205 Op0.getNode() == Op1.getNode() &&
8206 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8207 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8208 N->getValueType(0), Op0.getOperand(0));
8212 /// PerformSTORECombine - Target-specific dag combine xforms for
8214 static SDValue PerformSTORECombine(SDNode *N,
8215 TargetLowering::DAGCombinerInfo &DCI) {
8216 StoreSDNode *St = cast<StoreSDNode>(N);
8217 if (St->isVolatile())
8220 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8221 // pack all of the elements in one place. Next, store to memory in fewer
8223 SDValue StVal = St->getValue();
8224 EVT VT = StVal.getValueType();
8225 if (St->isTruncatingStore() && VT.isVector()) {
8226 SelectionDAG &DAG = DCI.DAG;
8227 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8228 EVT StVT = St->getMemoryVT();
8229 unsigned NumElems = VT.getVectorNumElements();
8230 assert(StVT != VT && "Cannot truncate to the same type");
8231 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8232 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8234 // From, To sizes and ElemCount must be pow of two
8235 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8237 // We are going to use the original vector elt for storing.
8238 // Accumulated smaller vector elements must be a multiple of the store size.
8239 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8241 unsigned SizeRatio = FromEltSz / ToEltSz;
8242 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8244 // Create a type on which we perform the shuffle.
8245 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8246 NumElems*SizeRatio);
8247 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8250 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8251 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8252 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8254 // Can't shuffle using an illegal type.
8255 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8257 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8258 DAG.getUNDEF(WideVec.getValueType()),
8260 // At this point all of the data is stored at the bottom of the
8261 // register. We now need to save it to mem.
8263 // Find the largest store unit
8264 MVT StoreType = MVT::i8;
8265 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8266 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8267 MVT Tp = (MVT::SimpleValueType)tp;
8268 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8271 // Didn't find a legal store type.
8272 if (!TLI.isTypeLegal(StoreType))
8275 // Bitcast the original vector into a vector of store-size units
8276 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8277 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8278 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8279 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8280 SmallVector<SDValue, 8> Chains;
8281 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8282 TLI.getPointerTy());
8283 SDValue BasePtr = St->getBasePtr();
8285 // Perform one or more big stores into memory.
8286 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8287 for (unsigned I = 0; I < E; I++) {
8288 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8289 StoreType, ShuffWide,
8290 DAG.getIntPtrConstant(I));
8291 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8292 St->getPointerInfo(), St->isVolatile(),
8293 St->isNonTemporal(), St->getAlignment());
8294 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8296 Chains.push_back(Ch);
8298 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8302 if (!ISD::isNormalStore(St))
8305 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8306 // ARM stores of arguments in the same cache line.
8307 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8308 StVal.getNode()->hasOneUse()) {
8309 SelectionDAG &DAG = DCI.DAG;
8311 SDValue BasePtr = St->getBasePtr();
8312 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8313 StVal.getNode()->getOperand(0), BasePtr,
8314 St->getPointerInfo(), St->isVolatile(),
8315 St->isNonTemporal(), St->getAlignment());
8317 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8318 DAG.getConstant(4, MVT::i32));
8319 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8320 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8321 St->isNonTemporal(),
8322 std::min(4U, St->getAlignment() / 2));
8325 if (StVal.getValueType() != MVT::i64 ||
8326 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8329 // Bitcast an i64 store extracted from a vector to f64.
8330 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8331 SelectionDAG &DAG = DCI.DAG;
8333 SDValue IntVec = StVal.getOperand(0);
8334 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8335 IntVec.getValueType().getVectorNumElements());
8336 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8337 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8338 Vec, StVal.getOperand(1));
8340 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8341 // Make the DAGCombiner fold the bitcasts.
8342 DCI.AddToWorklist(Vec.getNode());
8343 DCI.AddToWorklist(ExtElt.getNode());
8344 DCI.AddToWorklist(V.getNode());
8345 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8346 St->getPointerInfo(), St->isVolatile(),
8347 St->isNonTemporal(), St->getAlignment(),
8351 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8352 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8353 /// i64 vector to have f64 elements, since the value can then be loaded
8354 /// directly into a VFP register.
8355 static bool hasNormalLoadOperand(SDNode *N) {
8356 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8357 for (unsigned i = 0; i < NumElts; ++i) {
8358 SDNode *Elt = N->getOperand(i).getNode();
8359 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8365 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8366 /// ISD::BUILD_VECTOR.
8367 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8368 TargetLowering::DAGCombinerInfo &DCI){
8369 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8370 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8371 // into a pair of GPRs, which is fine when the value is used as a scalar,
8372 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8373 SelectionDAG &DAG = DCI.DAG;
8374 if (N->getNumOperands() == 2) {
8375 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8380 // Load i64 elements as f64 values so that type legalization does not split
8381 // them up into i32 values.
8382 EVT VT = N->getValueType(0);
8383 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8386 SmallVector<SDValue, 8> Ops;
8387 unsigned NumElts = VT.getVectorNumElements();
8388 for (unsigned i = 0; i < NumElts; ++i) {
8389 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8391 // Make the DAGCombiner fold the bitcast.
8392 DCI.AddToWorklist(V.getNode());
8394 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8395 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8396 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8399 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8401 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8402 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8403 // At that time, we may have inserted bitcasts from integer to float.
8404 // If these bitcasts have survived DAGCombine, change the lowering of this
8405 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8406 // force to use floating point types.
8408 // Make sure we can change the type of the vector.
8409 // This is possible iff:
8410 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8411 // 1.1. Vector is used only once.
8412 // 1.2. Use is a bit convert to an integer type.
8413 // 2. The size of its operands are 32-bits (64-bits are not legal).
8414 EVT VT = N->getValueType(0);
8415 EVT EltVT = VT.getVectorElementType();
8417 // Check 1.1. and 2.
8418 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8421 // By construction, the input type must be float.
8422 assert(EltVT == MVT::f32 && "Unexpected type!");
8425 SDNode *Use = *N->use_begin();
8426 if (Use->getOpcode() != ISD::BITCAST ||
8427 Use->getValueType(0).isFloatingPoint())
8430 // Check profitability.
8431 // Model is, if more than half of the relevant operands are bitcast from
8432 // i32, turn the build_vector into a sequence of insert_vector_elt.
8433 // Relevant operands are everything that is not statically
8434 // (i.e., at compile time) bitcasted.
8435 unsigned NumOfBitCastedElts = 0;
8436 unsigned NumElts = VT.getVectorNumElements();
8437 unsigned NumOfRelevantElts = NumElts;
8438 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8439 SDValue Elt = N->getOperand(Idx);
8440 if (Elt->getOpcode() == ISD::BITCAST) {
8441 // Assume only bit cast to i32 will go away.
8442 if (Elt->getOperand(0).getValueType() == MVT::i32)
8443 ++NumOfBitCastedElts;
8444 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8445 // Constants are statically casted, thus do not count them as
8446 // relevant operands.
8447 --NumOfRelevantElts;
8450 // Check if more than half of the elements require a non-free bitcast.
8451 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8454 SelectionDAG &DAG = DCI.DAG;
8455 // Create the new vector type.
8456 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8457 // Check if the type is legal.
8458 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8459 if (!TLI.isTypeLegal(VecVT))
8463 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8464 // => BITCAST INSERT_VECTOR_ELT
8465 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8467 SDValue Vec = DAG.getUNDEF(VecVT);
8469 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8470 SDValue V = N->getOperand(Idx);
8471 if (V.getOpcode() == ISD::UNDEF)
8473 if (V.getOpcode() == ISD::BITCAST &&
8474 V->getOperand(0).getValueType() == MVT::i32)
8475 // Fold obvious case.
8476 V = V.getOperand(0);
8478 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8479 // Make the DAGCombiner fold the bitcasts.
8480 DCI.AddToWorklist(V.getNode());
8482 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8483 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8485 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8486 // Make the DAGCombiner fold the bitcasts.
8487 DCI.AddToWorklist(Vec.getNode());
8491 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8492 /// ISD::INSERT_VECTOR_ELT.
8493 static SDValue PerformInsertEltCombine(SDNode *N,
8494 TargetLowering::DAGCombinerInfo &DCI) {
8495 // Bitcast an i64 load inserted into a vector to f64.
8496 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8497 EVT VT = N->getValueType(0);
8498 SDNode *Elt = N->getOperand(1).getNode();
8499 if (VT.getVectorElementType() != MVT::i64 ||
8500 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8503 SelectionDAG &DAG = DCI.DAG;
8505 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8506 VT.getVectorNumElements());
8507 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8508 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8509 // Make the DAGCombiner fold the bitcasts.
8510 DCI.AddToWorklist(Vec.getNode());
8511 DCI.AddToWorklist(V.getNode());
8512 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8513 Vec, V, N->getOperand(2));
8514 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8517 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8518 /// ISD::VECTOR_SHUFFLE.
8519 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8520 // The LLVM shufflevector instruction does not require the shuffle mask
8521 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8522 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8523 // operands do not match the mask length, they are extended by concatenating
8524 // them with undef vectors. That is probably the right thing for other
8525 // targets, but for NEON it is better to concatenate two double-register
8526 // size vector operands into a single quad-register size vector. Do that
8527 // transformation here:
8528 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8529 // shuffle(concat(v1, v2), undef)
8530 SDValue Op0 = N->getOperand(0);
8531 SDValue Op1 = N->getOperand(1);
8532 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8533 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8534 Op0.getNumOperands() != 2 ||
8535 Op1.getNumOperands() != 2)
8537 SDValue Concat0Op1 = Op0.getOperand(1);
8538 SDValue Concat1Op1 = Op1.getOperand(1);
8539 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8540 Concat1Op1.getOpcode() != ISD::UNDEF)
8542 // Skip the transformation if any of the types are illegal.
8543 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8544 EVT VT = N->getValueType(0);
8545 if (!TLI.isTypeLegal(VT) ||
8546 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8547 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8550 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8551 Op0.getOperand(0), Op1.getOperand(0));
8552 // Translate the shuffle mask.
8553 SmallVector<int, 16> NewMask;
8554 unsigned NumElts = VT.getVectorNumElements();
8555 unsigned HalfElts = NumElts/2;
8556 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8557 for (unsigned n = 0; n < NumElts; ++n) {
8558 int MaskElt = SVN->getMaskElt(n);
8560 if (MaskElt < (int)HalfElts)
8562 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8563 NewElt = HalfElts + MaskElt - NumElts;
8564 NewMask.push_back(NewElt);
8566 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8567 DAG.getUNDEF(VT), NewMask.data());
8570 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8571 /// NEON load/store intrinsics to merge base address updates.
8572 static SDValue CombineBaseUpdate(SDNode *N,
8573 TargetLowering::DAGCombinerInfo &DCI) {
8574 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8577 SelectionDAG &DAG = DCI.DAG;
8578 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8579 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8580 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8581 SDValue Addr = N->getOperand(AddrOpIdx);
8583 // Search for a use of the address operand that is an increment.
8584 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8585 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8587 if (User->getOpcode() != ISD::ADD ||
8588 UI.getUse().getResNo() != Addr.getResNo())
8591 // Check that the add is independent of the load/store. Otherwise, folding
8592 // it would create a cycle.
8593 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8596 // Find the new opcode for the updating load/store.
8598 bool isLaneOp = false;
8599 unsigned NewOpc = 0;
8600 unsigned NumVecs = 0;
8602 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8604 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8605 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8607 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8609 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8611 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8613 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8614 NumVecs = 2; isLaneOp = true; break;
8615 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8616 NumVecs = 3; isLaneOp = true; break;
8617 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8618 NumVecs = 4; isLaneOp = true; break;
8619 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8620 NumVecs = 1; isLoad = false; break;
8621 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8622 NumVecs = 2; isLoad = false; break;
8623 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8624 NumVecs = 3; isLoad = false; break;
8625 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8626 NumVecs = 4; isLoad = false; break;
8627 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8628 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8629 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8630 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8631 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8632 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8636 switch (N->getOpcode()) {
8637 default: llvm_unreachable("unexpected opcode for Neon base update");
8638 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8639 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8640 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8644 // Find the size of memory referenced by the load/store.
8647 VecTy = N->getValueType(0);
8649 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8650 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8652 NumBytes /= VecTy.getVectorNumElements();
8654 // If the increment is a constant, it must match the memory ref size.
8655 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8656 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8657 uint64_t IncVal = CInc->getZExtValue();
8658 if (IncVal != NumBytes)
8660 } else if (NumBytes >= 3 * 16) {
8661 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8662 // separate instructions that make it harder to use a non-constant update.
8666 // Create the new updating load/store node.
8668 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8670 for (n = 0; n < NumResultVecs; ++n)
8672 Tys[n++] = MVT::i32;
8673 Tys[n] = MVT::Other;
8674 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
8675 SmallVector<SDValue, 8> Ops;
8676 Ops.push_back(N->getOperand(0)); // incoming chain
8677 Ops.push_back(N->getOperand(AddrOpIdx));
8679 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8680 Ops.push_back(N->getOperand(i));
8682 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8683 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8684 Ops.data(), Ops.size(),
8685 MemInt->getMemoryVT(),
8686 MemInt->getMemOperand());
8689 std::vector<SDValue> NewResults;
8690 for (unsigned i = 0; i < NumResultVecs; ++i) {
8691 NewResults.push_back(SDValue(UpdN.getNode(), i));
8693 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8694 DCI.CombineTo(N, NewResults);
8695 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8702 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8703 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8704 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8706 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8707 SelectionDAG &DAG = DCI.DAG;
8708 EVT VT = N->getValueType(0);
8709 // vldN-dup instructions only support 64-bit vectors for N > 1.
8710 if (!VT.is64BitVector())
8713 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8714 SDNode *VLD = N->getOperand(0).getNode();
8715 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8717 unsigned NumVecs = 0;
8718 unsigned NewOpc = 0;
8719 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8720 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8722 NewOpc = ARMISD::VLD2DUP;
8723 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8725 NewOpc = ARMISD::VLD3DUP;
8726 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8728 NewOpc = ARMISD::VLD4DUP;
8733 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8734 // numbers match the load.
8735 unsigned VLDLaneNo =
8736 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8737 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8739 // Ignore uses of the chain result.
8740 if (UI.getUse().getResNo() == NumVecs)
8743 if (User->getOpcode() != ARMISD::VDUPLANE ||
8744 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8748 // Create the vldN-dup node.
8751 for (n = 0; n < NumVecs; ++n)
8753 Tys[n] = MVT::Other;
8754 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
8755 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8756 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8757 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
8758 Ops, 2, VLDMemInt->getMemoryVT(),
8759 VLDMemInt->getMemOperand());
8762 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8764 unsigned ResNo = UI.getUse().getResNo();
8765 // Ignore uses of the chain result.
8766 if (ResNo == NumVecs)
8769 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8772 // Now the vldN-lane intrinsic is dead except for its chain result.
8773 // Update uses of the chain.
8774 std::vector<SDValue> VLDDupResults;
8775 for (unsigned n = 0; n < NumVecs; ++n)
8776 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8777 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8778 DCI.CombineTo(VLD, VLDDupResults);
8783 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8784 /// ARMISD::VDUPLANE.
8785 static SDValue PerformVDUPLANECombine(SDNode *N,
8786 TargetLowering::DAGCombinerInfo &DCI) {
8787 SDValue Op = N->getOperand(0);
8789 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8790 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8791 if (CombineVLDDUP(N, DCI))
8792 return SDValue(N, 0);
8794 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8795 // redundant. Ignore bit_converts for now; element sizes are checked below.
8796 while (Op.getOpcode() == ISD::BITCAST)
8797 Op = Op.getOperand(0);
8798 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8801 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8802 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8803 // The canonical VMOV for a zero vector uses a 32-bit element size.
8804 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8806 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8808 EVT VT = N->getValueType(0);
8809 if (EltSize > VT.getVectorElementType().getSizeInBits())
8812 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
8815 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8816 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8817 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8821 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8823 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8828 APFloat APF = C->getValueAPF();
8829 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8830 != APFloat::opOK || !isExact)
8833 c0 = (I == 0) ? cN : c0;
8834 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8841 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8842 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8843 /// when the VMUL has a constant operand that is a power of 2.
8845 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8846 /// vmul.f32 d16, d17, d16
8847 /// vcvt.s32.f32 d16, d16
8849 /// vcvt.s32.f32 d16, d16, #3
8850 static SDValue PerformVCVTCombine(SDNode *N,
8851 TargetLowering::DAGCombinerInfo &DCI,
8852 const ARMSubtarget *Subtarget) {
8853 SelectionDAG &DAG = DCI.DAG;
8854 SDValue Op = N->getOperand(0);
8856 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8857 Op.getOpcode() != ISD::FMUL)
8861 SDValue N0 = Op->getOperand(0);
8862 SDValue ConstVec = Op->getOperand(1);
8863 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8865 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8866 !isConstVecPow2(ConstVec, isSigned, C))
8869 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
8870 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
8871 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
8872 // These instructions only exist converting from f32 to i32. We can handle
8873 // smaller integers by generating an extra truncate, but larger ones would
8878 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8879 Intrinsic::arm_neon_vcvtfp2fxu;
8880 unsigned NumLanes = Op.getValueType().getVectorNumElements();
8881 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8882 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
8883 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8884 DAG.getConstant(Log2_64(C), MVT::i32));
8886 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
8887 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
8892 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8893 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8894 /// when the VDIV has a constant operand that is a power of 2.
8896 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8897 /// vcvt.f32.s32 d16, d16
8898 /// vdiv.f32 d16, d17, d16
8900 /// vcvt.f32.s32 d16, d16, #3
8901 static SDValue PerformVDIVCombine(SDNode *N,
8902 TargetLowering::DAGCombinerInfo &DCI,
8903 const ARMSubtarget *Subtarget) {
8904 SelectionDAG &DAG = DCI.DAG;
8905 SDValue Op = N->getOperand(0);
8906 unsigned OpOpcode = Op.getNode()->getOpcode();
8908 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8909 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8913 SDValue ConstVec = N->getOperand(1);
8914 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8916 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8917 !isConstVecPow2(ConstVec, isSigned, C))
8920 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
8921 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
8922 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
8923 // These instructions only exist converting from i32 to f32. We can handle
8924 // smaller integers by generating an extra extend, but larger ones would
8929 SDValue ConvInput = Op.getOperand(0);
8930 unsigned NumLanes = Op.getValueType().getVectorNumElements();
8931 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
8932 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
8933 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
8936 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8937 Intrinsic::arm_neon_vcvtfxu2fp;
8938 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8940 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8941 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
8944 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8945 /// operand of a vector shift operation, where all the elements of the
8946 /// build_vector must have the same constant integer value.
8947 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8948 // Ignore bit_converts.
8949 while (Op.getOpcode() == ISD::BITCAST)
8950 Op = Op.getOperand(0);
8951 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8952 APInt SplatBits, SplatUndef;
8953 unsigned SplatBitSize;
8955 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8956 HasAnyUndefs, ElementBits) ||
8957 SplatBitSize > ElementBits)
8959 Cnt = SplatBits.getSExtValue();
8963 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8964 /// operand of a vector shift left operation. That value must be in the range:
8965 /// 0 <= Value < ElementBits for a left shift; or
8966 /// 0 <= Value <= ElementBits for a long left shift.
8967 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8968 assert(VT.isVector() && "vector shift count is not a vector type");
8969 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8970 if (! getVShiftImm(Op, ElementBits, Cnt))
8972 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8975 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8976 /// operand of a vector shift right operation. For a shift opcode, the value
8977 /// is positive, but for an intrinsic the value count must be negative. The
8978 /// absolute value must be in the range:
8979 /// 1 <= |Value| <= ElementBits for a right shift; or
8980 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8981 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8983 assert(VT.isVector() && "vector shift count is not a vector type");
8984 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8985 if (! getVShiftImm(Op, ElementBits, Cnt))
8989 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8992 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8993 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8994 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8997 // Don't do anything for most intrinsics.
9000 // Vector shifts: check for immediate versions and lower them.
9001 // Note: This is done during DAG combining instead of DAG legalizing because
9002 // the build_vectors for 64-bit vector element shift counts are generally
9003 // not legal, and it is hard to see their values after they get legalized to
9004 // loads from a constant pool.
9005 case Intrinsic::arm_neon_vshifts:
9006 case Intrinsic::arm_neon_vshiftu:
9007 case Intrinsic::arm_neon_vrshifts:
9008 case Intrinsic::arm_neon_vrshiftu:
9009 case Intrinsic::arm_neon_vrshiftn:
9010 case Intrinsic::arm_neon_vqshifts:
9011 case Intrinsic::arm_neon_vqshiftu:
9012 case Intrinsic::arm_neon_vqshiftsu:
9013 case Intrinsic::arm_neon_vqshiftns:
9014 case Intrinsic::arm_neon_vqshiftnu:
9015 case Intrinsic::arm_neon_vqshiftnsu:
9016 case Intrinsic::arm_neon_vqrshiftns:
9017 case Intrinsic::arm_neon_vqrshiftnu:
9018 case Intrinsic::arm_neon_vqrshiftnsu: {
9019 EVT VT = N->getOperand(1).getValueType();
9021 unsigned VShiftOpc = 0;
9024 case Intrinsic::arm_neon_vshifts:
9025 case Intrinsic::arm_neon_vshiftu:
9026 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9027 VShiftOpc = ARMISD::VSHL;
9030 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9031 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9032 ARMISD::VSHRs : ARMISD::VSHRu);
9037 case Intrinsic::arm_neon_vrshifts:
9038 case Intrinsic::arm_neon_vrshiftu:
9039 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9043 case Intrinsic::arm_neon_vqshifts:
9044 case Intrinsic::arm_neon_vqshiftu:
9045 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9049 case Intrinsic::arm_neon_vqshiftsu:
9050 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9052 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9054 case Intrinsic::arm_neon_vrshiftn:
9055 case Intrinsic::arm_neon_vqshiftns:
9056 case Intrinsic::arm_neon_vqshiftnu:
9057 case Intrinsic::arm_neon_vqshiftnsu:
9058 case Intrinsic::arm_neon_vqrshiftns:
9059 case Intrinsic::arm_neon_vqrshiftnu:
9060 case Intrinsic::arm_neon_vqrshiftnsu:
9061 // Narrowing shifts require an immediate right shift.
9062 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9064 llvm_unreachable("invalid shift count for narrowing vector shift "
9068 llvm_unreachable("unhandled vector shift");
9072 case Intrinsic::arm_neon_vshifts:
9073 case Intrinsic::arm_neon_vshiftu:
9074 // Opcode already set above.
9076 case Intrinsic::arm_neon_vrshifts:
9077 VShiftOpc = ARMISD::VRSHRs; break;
9078 case Intrinsic::arm_neon_vrshiftu:
9079 VShiftOpc = ARMISD::VRSHRu; break;
9080 case Intrinsic::arm_neon_vrshiftn:
9081 VShiftOpc = ARMISD::VRSHRN; break;
9082 case Intrinsic::arm_neon_vqshifts:
9083 VShiftOpc = ARMISD::VQSHLs; break;
9084 case Intrinsic::arm_neon_vqshiftu:
9085 VShiftOpc = ARMISD::VQSHLu; break;
9086 case Intrinsic::arm_neon_vqshiftsu:
9087 VShiftOpc = ARMISD::VQSHLsu; break;
9088 case Intrinsic::arm_neon_vqshiftns:
9089 VShiftOpc = ARMISD::VQSHRNs; break;
9090 case Intrinsic::arm_neon_vqshiftnu:
9091 VShiftOpc = ARMISD::VQSHRNu; break;
9092 case Intrinsic::arm_neon_vqshiftnsu:
9093 VShiftOpc = ARMISD::VQSHRNsu; break;
9094 case Intrinsic::arm_neon_vqrshiftns:
9095 VShiftOpc = ARMISD::VQRSHRNs; break;
9096 case Intrinsic::arm_neon_vqrshiftnu:
9097 VShiftOpc = ARMISD::VQRSHRNu; break;
9098 case Intrinsic::arm_neon_vqrshiftnsu:
9099 VShiftOpc = ARMISD::VQRSHRNsu; break;
9102 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9103 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9106 case Intrinsic::arm_neon_vshiftins: {
9107 EVT VT = N->getOperand(1).getValueType();
9109 unsigned VShiftOpc = 0;
9111 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9112 VShiftOpc = ARMISD::VSLI;
9113 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9114 VShiftOpc = ARMISD::VSRI;
9116 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9119 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9120 N->getOperand(1), N->getOperand(2),
9121 DAG.getConstant(Cnt, MVT::i32));
9124 case Intrinsic::arm_neon_vqrshifts:
9125 case Intrinsic::arm_neon_vqrshiftu:
9126 // No immediate versions of these to check for.
9133 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9134 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9135 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9136 /// vector element shift counts are generally not legal, and it is hard to see
9137 /// their values after they get legalized to loads from a constant pool.
9138 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9139 const ARMSubtarget *ST) {
9140 EVT VT = N->getValueType(0);
9141 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9142 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9143 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9144 SDValue N1 = N->getOperand(1);
9145 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9146 SDValue N0 = N->getOperand(0);
9147 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9148 DAG.MaskedValueIsZero(N0.getOperand(0),
9149 APInt::getHighBitsSet(32, 16)))
9150 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9154 // Nothing to be done for scalar shifts.
9155 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9156 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9159 assert(ST->hasNEON() && "unexpected vector shift");
9162 switch (N->getOpcode()) {
9163 default: llvm_unreachable("unexpected shift opcode");
9166 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9167 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9168 DAG.getConstant(Cnt, MVT::i32));
9173 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9174 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9175 ARMISD::VSHRs : ARMISD::VSHRu);
9176 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9177 DAG.getConstant(Cnt, MVT::i32));
9183 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9184 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9185 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9186 const ARMSubtarget *ST) {
9187 SDValue N0 = N->getOperand(0);
9189 // Check for sign- and zero-extensions of vector extract operations of 8-
9190 // and 16-bit vector elements. NEON supports these directly. They are
9191 // handled during DAG combining because type legalization will promote them
9192 // to 32-bit types and it is messy to recognize the operations after that.
9193 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9194 SDValue Vec = N0.getOperand(0);
9195 SDValue Lane = N0.getOperand(1);
9196 EVT VT = N->getValueType(0);
9197 EVT EltVT = N0.getValueType();
9198 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9200 if (VT == MVT::i32 &&
9201 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9202 TLI.isTypeLegal(Vec.getValueType()) &&
9203 isa<ConstantSDNode>(Lane)) {
9206 switch (N->getOpcode()) {
9207 default: llvm_unreachable("unexpected opcode");
9208 case ISD::SIGN_EXTEND:
9209 Opc = ARMISD::VGETLANEs;
9211 case ISD::ZERO_EXTEND:
9212 case ISD::ANY_EXTEND:
9213 Opc = ARMISD::VGETLANEu;
9216 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9223 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9224 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9225 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9226 const ARMSubtarget *ST) {
9227 // If the target supports NEON, try to use vmax/vmin instructions for f32
9228 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9229 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9230 // a NaN; only do the transformation when it matches that behavior.
9232 // For now only do this when using NEON for FP operations; if using VFP, it
9233 // is not obvious that the benefit outweighs the cost of switching to the
9235 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9236 N->getValueType(0) != MVT::f32)
9239 SDValue CondLHS = N->getOperand(0);
9240 SDValue CondRHS = N->getOperand(1);
9241 SDValue LHS = N->getOperand(2);
9242 SDValue RHS = N->getOperand(3);
9243 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9245 unsigned Opcode = 0;
9247 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9248 IsReversed = false; // x CC y ? x : y
9249 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9250 IsReversed = true ; // x CC y ? y : x
9264 // If LHS is NaN, an ordered comparison will be false and the result will
9265 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9266 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9267 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9268 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9270 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9271 // will return -0, so vmin can only be used for unsafe math or if one of
9272 // the operands is known to be nonzero.
9273 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9274 !DAG.getTarget().Options.UnsafeFPMath &&
9275 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9277 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9286 // If LHS is NaN, an ordered comparison will be false and the result will
9287 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9288 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9289 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9290 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9292 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9293 // will return +0, so vmax can only be used for unsafe math or if one of
9294 // the operands is known to be nonzero.
9295 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9296 !DAG.getTarget().Options.UnsafeFPMath &&
9297 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9299 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9305 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9308 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9310 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9311 SDValue Cmp = N->getOperand(4);
9312 if (Cmp.getOpcode() != ARMISD::CMPZ)
9313 // Only looking at EQ and NE cases.
9316 EVT VT = N->getValueType(0);
9318 SDValue LHS = Cmp.getOperand(0);
9319 SDValue RHS = Cmp.getOperand(1);
9320 SDValue FalseVal = N->getOperand(0);
9321 SDValue TrueVal = N->getOperand(1);
9322 SDValue ARMcc = N->getOperand(2);
9323 ARMCC::CondCodes CC =
9324 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9342 /// FIXME: Turn this into a target neutral optimization?
9344 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9345 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9346 N->getOperand(3), Cmp);
9347 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9349 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9350 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9351 N->getOperand(3), NewCmp);
9354 if (Res.getNode()) {
9355 APInt KnownZero, KnownOne;
9356 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9357 // Capture demanded bits information that would be otherwise lost.
9358 if (KnownZero == 0xfffffffe)
9359 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9360 DAG.getValueType(MVT::i1));
9361 else if (KnownZero == 0xffffff00)
9362 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9363 DAG.getValueType(MVT::i8));
9364 else if (KnownZero == 0xffff0000)
9365 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9366 DAG.getValueType(MVT::i16));
9372 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9373 DAGCombinerInfo &DCI) const {
9374 switch (N->getOpcode()) {
9376 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9377 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9378 case ISD::SUB: return PerformSUBCombine(N, DCI);
9379 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9380 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9381 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9382 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9383 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9384 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9385 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9386 case ISD::STORE: return PerformSTORECombine(N, DCI);
9387 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9388 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9389 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9390 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9391 case ISD::FP_TO_SINT:
9392 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9393 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9394 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9397 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9398 case ISD::SIGN_EXTEND:
9399 case ISD::ZERO_EXTEND:
9400 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9401 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9402 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9403 case ARMISD::VLD2DUP:
9404 case ARMISD::VLD3DUP:
9405 case ARMISD::VLD4DUP:
9406 return CombineBaseUpdate(N, DCI);
9407 case ARMISD::BUILD_VECTOR:
9408 return PerformARMBUILD_VECTORCombine(N, DCI);
9409 case ISD::INTRINSIC_VOID:
9410 case ISD::INTRINSIC_W_CHAIN:
9411 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9412 case Intrinsic::arm_neon_vld1:
9413 case Intrinsic::arm_neon_vld2:
9414 case Intrinsic::arm_neon_vld3:
9415 case Intrinsic::arm_neon_vld4:
9416 case Intrinsic::arm_neon_vld2lane:
9417 case Intrinsic::arm_neon_vld3lane:
9418 case Intrinsic::arm_neon_vld4lane:
9419 case Intrinsic::arm_neon_vst1:
9420 case Intrinsic::arm_neon_vst2:
9421 case Intrinsic::arm_neon_vst3:
9422 case Intrinsic::arm_neon_vst4:
9423 case Intrinsic::arm_neon_vst2lane:
9424 case Intrinsic::arm_neon_vst3lane:
9425 case Intrinsic::arm_neon_vst4lane:
9426 return CombineBaseUpdate(N, DCI);
9434 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9436 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9439 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
9441 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9442 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9444 switch (VT.getSimpleVT().SimpleTy) {
9450 // Unaligned access can use (for example) LRDB, LRDH, LDR
9451 if (AllowsUnaligned) {
9453 *Fast = Subtarget->hasV7Ops();
9460 // For any little-endian targets with neon, we can support unaligned ld/st
9461 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9462 // A big-endian target may also explicitly support unaligned accesses
9463 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9473 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9474 unsigned AlignCheck) {
9475 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9476 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9479 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9480 unsigned DstAlign, unsigned SrcAlign,
9481 bool IsMemset, bool ZeroMemset,
9483 MachineFunction &MF) const {
9484 const Function *F = MF.getFunction();
9486 // See if we can use NEON instructions for this...
9487 if ((!IsMemset || ZeroMemset) &&
9488 Subtarget->hasNEON() &&
9489 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9490 Attribute::NoImplicitFloat)) {
9493 (memOpAlign(SrcAlign, DstAlign, 16) ||
9494 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
9496 } else if (Size >= 8 &&
9497 (memOpAlign(SrcAlign, DstAlign, 8) ||
9498 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
9503 // Lowering to i32/i16 if the size permits.
9509 // Let the target-independent logic figure it out.
9513 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9514 if (Val.getOpcode() != ISD::LOAD)
9517 EVT VT1 = Val.getValueType();
9518 if (!VT1.isSimple() || !VT1.isInteger() ||
9519 !VT2.isSimple() || !VT2.isInteger())
9522 switch (VT1.getSimpleVT().SimpleTy) {
9527 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9534 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9535 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9538 if (!isTypeLegal(EVT::getEVT(Ty1)))
9541 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9543 // Assuming the caller doesn't have a zeroext or signext return parameter,
9544 // truncation all the way down to i1 is valid.
9549 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9554 switch (VT.getSimpleVT().SimpleTy) {
9555 default: return false;
9570 if ((V & (Scale - 1)) != 0)
9573 return V == (V & ((1LL << 5) - 1));
9576 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9577 const ARMSubtarget *Subtarget) {
9584 switch (VT.getSimpleVT().SimpleTy) {
9585 default: return false;
9590 // + imm12 or - imm8
9592 return V == (V & ((1LL << 8) - 1));
9593 return V == (V & ((1LL << 12) - 1));
9596 // Same as ARM mode. FIXME: NEON?
9597 if (!Subtarget->hasVFP2())
9602 return V == (V & ((1LL << 8) - 1));
9606 /// isLegalAddressImmediate - Return true if the integer value can be used
9607 /// as the offset of the target addressing mode for load / store of the
9609 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9610 const ARMSubtarget *Subtarget) {
9617 if (Subtarget->isThumb1Only())
9618 return isLegalT1AddressImmediate(V, VT);
9619 else if (Subtarget->isThumb2())
9620 return isLegalT2AddressImmediate(V, VT, Subtarget);
9625 switch (VT.getSimpleVT().SimpleTy) {
9626 default: return false;
9631 return V == (V & ((1LL << 12) - 1));
9634 return V == (V & ((1LL << 8) - 1));
9637 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9642 return V == (V & ((1LL << 8) - 1));
9646 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9648 int Scale = AM.Scale;
9652 switch (VT.getSimpleVT().SimpleTy) {
9653 default: return false;
9662 return Scale == 2 || Scale == 4 || Scale == 8;
9665 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9669 // Note, we allow "void" uses (basically, uses that aren't loads or
9670 // stores), because arm allows folding a scale into many arithmetic
9671 // operations. This should be made more precise and revisited later.
9673 // Allow r << imm, but the imm has to be a multiple of two.
9674 if (Scale & 1) return false;
9675 return isPowerOf2_32(Scale);
9679 /// isLegalAddressingMode - Return true if the addressing mode represented
9680 /// by AM is legal for this target, for a load/store of the specified type.
9681 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9683 EVT VT = getValueType(Ty, true);
9684 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9687 // Can never fold addr of global into load/store.
9692 case 0: // no scale reg, must be "r+i" or "r", or "i".
9695 if (Subtarget->isThumb1Only())
9699 // ARM doesn't support any R+R*scale+imm addr modes.
9706 if (Subtarget->isThumb2())
9707 return isLegalT2ScaledAddressingMode(AM, VT);
9709 int Scale = AM.Scale;
9710 switch (VT.getSimpleVT().SimpleTy) {
9711 default: return false;
9715 if (Scale < 0) Scale = -Scale;
9719 return isPowerOf2_32(Scale & ~1);
9723 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9728 // Note, we allow "void" uses (basically, uses that aren't loads or
9729 // stores), because arm allows folding a scale into many arithmetic
9730 // operations. This should be made more precise and revisited later.
9732 // Allow r << imm, but the imm has to be a multiple of two.
9733 if (Scale & 1) return false;
9734 return isPowerOf2_32(Scale);
9740 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9741 /// icmp immediate, that is the target has icmp instructions which can compare
9742 /// a register against the immediate without having to materialize the
9743 /// immediate into a register.
9744 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9745 // Thumb2 and ARM modes can use cmn for negative immediates.
9746 if (!Subtarget->isThumb())
9747 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9748 if (Subtarget->isThumb2())
9749 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9750 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9751 return Imm >= 0 && Imm <= 255;
9754 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9755 /// *or sub* immediate, that is the target has add or sub instructions which can
9756 /// add a register with the immediate without having to materialize the
9757 /// immediate into a register.
9758 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9759 // Same encoding for add/sub, just flip the sign.
9760 int64_t AbsImm = llvm::abs64(Imm);
9761 if (!Subtarget->isThumb())
9762 return ARM_AM::getSOImmVal(AbsImm) != -1;
9763 if (Subtarget->isThumb2())
9764 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9765 // Thumb1 only has 8-bit unsigned immediate.
9766 return AbsImm >= 0 && AbsImm <= 255;
9769 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9770 bool isSEXTLoad, SDValue &Base,
9771 SDValue &Offset, bool &isInc,
9772 SelectionDAG &DAG) {
9773 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9776 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9778 Base = Ptr->getOperand(0);
9779 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9780 int RHSC = (int)RHS->getZExtValue();
9781 if (RHSC < 0 && RHSC > -256) {
9782 assert(Ptr->getOpcode() == ISD::ADD);
9784 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9788 isInc = (Ptr->getOpcode() == ISD::ADD);
9789 Offset = Ptr->getOperand(1);
9791 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9793 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9794 int RHSC = (int)RHS->getZExtValue();
9795 if (RHSC < 0 && RHSC > -0x1000) {
9796 assert(Ptr->getOpcode() == ISD::ADD);
9798 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9799 Base = Ptr->getOperand(0);
9804 if (Ptr->getOpcode() == ISD::ADD) {
9806 ARM_AM::ShiftOpc ShOpcVal=
9807 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9808 if (ShOpcVal != ARM_AM::no_shift) {
9809 Base = Ptr->getOperand(1);
9810 Offset = Ptr->getOperand(0);
9812 Base = Ptr->getOperand(0);
9813 Offset = Ptr->getOperand(1);
9818 isInc = (Ptr->getOpcode() == ISD::ADD);
9819 Base = Ptr->getOperand(0);
9820 Offset = Ptr->getOperand(1);
9824 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9828 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9829 bool isSEXTLoad, SDValue &Base,
9830 SDValue &Offset, bool &isInc,
9831 SelectionDAG &DAG) {
9832 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9835 Base = Ptr->getOperand(0);
9836 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9837 int RHSC = (int)RHS->getZExtValue();
9838 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9839 assert(Ptr->getOpcode() == ISD::ADD);
9841 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9843 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9844 isInc = Ptr->getOpcode() == ISD::ADD;
9845 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9853 /// getPreIndexedAddressParts - returns true by value, base pointer and
9854 /// offset pointer and addressing mode by reference if the node's address
9855 /// can be legally represented as pre-indexed load / store address.
9857 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9859 ISD::MemIndexedMode &AM,
9860 SelectionDAG &DAG) const {
9861 if (Subtarget->isThumb1Only())
9866 bool isSEXTLoad = false;
9867 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9868 Ptr = LD->getBasePtr();
9869 VT = LD->getMemoryVT();
9870 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9871 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9872 Ptr = ST->getBasePtr();
9873 VT = ST->getMemoryVT();
9878 bool isLegal = false;
9879 if (Subtarget->isThumb2())
9880 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9881 Offset, isInc, DAG);
9883 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9884 Offset, isInc, DAG);
9888 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9892 /// getPostIndexedAddressParts - returns true by value, base pointer and
9893 /// offset pointer and addressing mode by reference if this node can be
9894 /// combined with a load / store to form a post-indexed load / store.
9895 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9898 ISD::MemIndexedMode &AM,
9899 SelectionDAG &DAG) const {
9900 if (Subtarget->isThumb1Only())
9905 bool isSEXTLoad = false;
9906 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9907 VT = LD->getMemoryVT();
9908 Ptr = LD->getBasePtr();
9909 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9910 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9911 VT = ST->getMemoryVT();
9912 Ptr = ST->getBasePtr();
9917 bool isLegal = false;
9918 if (Subtarget->isThumb2())
9919 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9922 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9928 // Swap base ptr and offset to catch more post-index load / store when
9929 // it's legal. In Thumb2 mode, offset must be an immediate.
9930 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9931 !Subtarget->isThumb2())
9932 std::swap(Base, Offset);
9934 // Post-indexed load / store update the base pointer.
9939 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9943 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9946 const SelectionDAG &DAG,
9947 unsigned Depth) const {
9948 unsigned BitWidth = KnownOne.getBitWidth();
9949 KnownZero = KnownOne = APInt(BitWidth, 0);
9950 switch (Op.getOpcode()) {
9956 // These nodes' second result is a boolean
9957 if (Op.getResNo() == 0)
9959 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
9961 case ARMISD::CMOV: {
9962 // Bits are known zero/one if known on the LHS and RHS.
9963 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9964 if (KnownZero == 0 && KnownOne == 0) return;
9966 APInt KnownZeroRHS, KnownOneRHS;
9967 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9968 KnownZero &= KnownZeroRHS;
9969 KnownOne &= KnownOneRHS;
9972 case ISD::INTRINSIC_W_CHAIN: {
9973 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
9974 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
9977 case Intrinsic::arm_ldaex:
9978 case Intrinsic::arm_ldrex: {
9979 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
9980 unsigned MemBits = VT.getScalarType().getSizeInBits();
9981 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
9989 //===----------------------------------------------------------------------===//
9990 // ARM Inline Assembly Support
9991 //===----------------------------------------------------------------------===//
9993 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9994 // Looking for "rev" which is V6+.
9995 if (!Subtarget->hasV6Ops())
9998 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9999 std::string AsmStr = IA->getAsmString();
10000 SmallVector<StringRef, 4> AsmPieces;
10001 SplitString(AsmStr, AsmPieces, ";\n");
10003 switch (AsmPieces.size()) {
10004 default: return false;
10006 AsmStr = AsmPieces[0];
10008 SplitString(AsmStr, AsmPieces, " \t,");
10011 if (AsmPieces.size() == 3 &&
10012 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10013 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10014 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10015 if (Ty && Ty->getBitWidth() == 32)
10016 return IntrinsicLowering::LowerToByteSwap(CI);
10024 /// getConstraintType - Given a constraint letter, return the type of
10025 /// constraint it is for this target.
10026 ARMTargetLowering::ConstraintType
10027 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10028 if (Constraint.size() == 1) {
10029 switch (Constraint[0]) {
10031 case 'l': return C_RegisterClass;
10032 case 'w': return C_RegisterClass;
10033 case 'h': return C_RegisterClass;
10034 case 'x': return C_RegisterClass;
10035 case 't': return C_RegisterClass;
10036 case 'j': return C_Other; // Constant for movw.
10037 // An address with a single base register. Due to the way we
10038 // currently handle addresses it is the same as an 'r' memory constraint.
10039 case 'Q': return C_Memory;
10041 } else if (Constraint.size() == 2) {
10042 switch (Constraint[0]) {
10044 // All 'U+' constraints are addresses.
10045 case 'U': return C_Memory;
10048 return TargetLowering::getConstraintType(Constraint);
10051 /// Examine constraint type and operand type and determine a weight value.
10052 /// This object must already have been set up with the operand type
10053 /// and the current alternative constraint selected.
10054 TargetLowering::ConstraintWeight
10055 ARMTargetLowering::getSingleConstraintMatchWeight(
10056 AsmOperandInfo &info, const char *constraint) const {
10057 ConstraintWeight weight = CW_Invalid;
10058 Value *CallOperandVal = info.CallOperandVal;
10059 // If we don't have a value, we can't do a match,
10060 // but allow it at the lowest weight.
10061 if (CallOperandVal == NULL)
10063 Type *type = CallOperandVal->getType();
10064 // Look at the constraint type.
10065 switch (*constraint) {
10067 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10070 if (type->isIntegerTy()) {
10071 if (Subtarget->isThumb())
10072 weight = CW_SpecificReg;
10074 weight = CW_Register;
10078 if (type->isFloatingPointTy())
10079 weight = CW_Register;
10085 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10087 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10089 if (Constraint.size() == 1) {
10090 // GCC ARM Constraint Letters
10091 switch (Constraint[0]) {
10092 case 'l': // Low regs or general regs.
10093 if (Subtarget->isThumb())
10094 return RCPair(0U, &ARM::tGPRRegClass);
10095 return RCPair(0U, &ARM::GPRRegClass);
10096 case 'h': // High regs or no regs.
10097 if (Subtarget->isThumb())
10098 return RCPair(0U, &ARM::hGPRRegClass);
10101 return RCPair(0U, &ARM::GPRRegClass);
10103 if (VT == MVT::Other)
10105 if (VT == MVT::f32)
10106 return RCPair(0U, &ARM::SPRRegClass);
10107 if (VT.getSizeInBits() == 64)
10108 return RCPair(0U, &ARM::DPRRegClass);
10109 if (VT.getSizeInBits() == 128)
10110 return RCPair(0U, &ARM::QPRRegClass);
10113 if (VT == MVT::Other)
10115 if (VT == MVT::f32)
10116 return RCPair(0U, &ARM::SPR_8RegClass);
10117 if (VT.getSizeInBits() == 64)
10118 return RCPair(0U, &ARM::DPR_8RegClass);
10119 if (VT.getSizeInBits() == 128)
10120 return RCPair(0U, &ARM::QPR_8RegClass);
10123 if (VT == MVT::f32)
10124 return RCPair(0U, &ARM::SPRRegClass);
10128 if (StringRef("{cc}").equals_lower(Constraint))
10129 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10131 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10134 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10135 /// vector. If it is invalid, don't add anything to Ops.
10136 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10137 std::string &Constraint,
10138 std::vector<SDValue>&Ops,
10139 SelectionDAG &DAG) const {
10140 SDValue Result(0, 0);
10142 // Currently only support length 1 constraints.
10143 if (Constraint.length() != 1) return;
10145 char ConstraintLetter = Constraint[0];
10146 switch (ConstraintLetter) {
10149 case 'I': case 'J': case 'K': case 'L':
10150 case 'M': case 'N': case 'O':
10151 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10155 int64_t CVal64 = C->getSExtValue();
10156 int CVal = (int) CVal64;
10157 // None of these constraints allow values larger than 32 bits. Check
10158 // that the value fits in an int.
10159 if (CVal != CVal64)
10162 switch (ConstraintLetter) {
10164 // Constant suitable for movw, must be between 0 and
10166 if (Subtarget->hasV6T2Ops())
10167 if (CVal >= 0 && CVal <= 65535)
10171 if (Subtarget->isThumb1Only()) {
10172 // This must be a constant between 0 and 255, for ADD
10174 if (CVal >= 0 && CVal <= 255)
10176 } else if (Subtarget->isThumb2()) {
10177 // A constant that can be used as an immediate value in a
10178 // data-processing instruction.
10179 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10182 // A constant that can be used as an immediate value in a
10183 // data-processing instruction.
10184 if (ARM_AM::getSOImmVal(CVal) != -1)
10190 if (Subtarget->isThumb()) { // FIXME thumb2
10191 // This must be a constant between -255 and -1, for negated ADD
10192 // immediates. This can be used in GCC with an "n" modifier that
10193 // prints the negated value, for use with SUB instructions. It is
10194 // not useful otherwise but is implemented for compatibility.
10195 if (CVal >= -255 && CVal <= -1)
10198 // This must be a constant between -4095 and 4095. It is not clear
10199 // what this constraint is intended for. Implemented for
10200 // compatibility with GCC.
10201 if (CVal >= -4095 && CVal <= 4095)
10207 if (Subtarget->isThumb1Only()) {
10208 // A 32-bit value where only one byte has a nonzero value. Exclude
10209 // zero to match GCC. This constraint is used by GCC internally for
10210 // constants that can be loaded with a move/shift combination.
10211 // It is not useful otherwise but is implemented for compatibility.
10212 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10214 } else if (Subtarget->isThumb2()) {
10215 // A constant whose bitwise inverse can be used as an immediate
10216 // value in a data-processing instruction. This can be used in GCC
10217 // with a "B" modifier that prints the inverted value, for use with
10218 // BIC and MVN instructions. It is not useful otherwise but is
10219 // implemented for compatibility.
10220 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10223 // A constant whose bitwise inverse can be used as an immediate
10224 // value in a data-processing instruction. This can be used in GCC
10225 // with a "B" modifier that prints the inverted value, for use with
10226 // BIC and MVN instructions. It is not useful otherwise but is
10227 // implemented for compatibility.
10228 if (ARM_AM::getSOImmVal(~CVal) != -1)
10234 if (Subtarget->isThumb1Only()) {
10235 // This must be a constant between -7 and 7,
10236 // for 3-operand ADD/SUB immediate instructions.
10237 if (CVal >= -7 && CVal < 7)
10239 } else if (Subtarget->isThumb2()) {
10240 // A constant whose negation can be used as an immediate value in a
10241 // data-processing instruction. This can be used in GCC with an "n"
10242 // modifier that prints the negated value, for use with SUB
10243 // instructions. It is not useful otherwise but is implemented for
10245 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10248 // A constant whose negation can be used as an immediate value in a
10249 // data-processing instruction. This can be used in GCC with an "n"
10250 // modifier that prints the negated value, for use with SUB
10251 // instructions. It is not useful otherwise but is implemented for
10253 if (ARM_AM::getSOImmVal(-CVal) != -1)
10259 if (Subtarget->isThumb()) { // FIXME thumb2
10260 // This must be a multiple of 4 between 0 and 1020, for
10261 // ADD sp + immediate.
10262 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10265 // A power of two or a constant between 0 and 32. This is used in
10266 // GCC for the shift amount on shifted register operands, but it is
10267 // useful in general for any shift amounts.
10268 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10274 if (Subtarget->isThumb()) { // FIXME thumb2
10275 // This must be a constant between 0 and 31, for shift amounts.
10276 if (CVal >= 0 && CVal <= 31)
10282 if (Subtarget->isThumb()) { // FIXME thumb2
10283 // This must be a multiple of 4 between -508 and 508, for
10284 // ADD/SUB sp = sp + immediate.
10285 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10290 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10294 if (Result.getNode()) {
10295 Ops.push_back(Result);
10298 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10301 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10302 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10303 unsigned Opcode = Op->getOpcode();
10304 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10305 "Invalid opcode for Div/Rem lowering");
10306 bool isSigned = (Opcode == ISD::SDIVREM);
10307 EVT VT = Op->getValueType(0);
10308 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10311 switch (VT.getSimpleVT().SimpleTy) {
10312 default: llvm_unreachable("Unexpected request for libcall!");
10313 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10314 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10315 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10316 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10319 SDValue InChain = DAG.getEntryNode();
10321 TargetLowering::ArgListTy Args;
10322 TargetLowering::ArgListEntry Entry;
10323 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10324 EVT ArgVT = Op->getOperand(i).getValueType();
10325 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10326 Entry.Node = Op->getOperand(i);
10328 Entry.isSExt = isSigned;
10329 Entry.isZExt = !isSigned;
10330 Args.push_back(Entry);
10333 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10336 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10340 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
10341 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
10342 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
10343 Callee, Args, DAG, dl);
10344 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10346 return CallInfo.first;
10350 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10351 // The ARM target isn't yet aware of offsets.
10355 bool ARM::isBitFieldInvertedMask(unsigned v) {
10356 if (v == 0xffffffff)
10359 // there can be 1's on either or both "outsides", all the "inside"
10360 // bits must be 0's
10361 unsigned TO = CountTrailingOnes_32(v);
10362 unsigned LO = CountLeadingOnes_32(v);
10363 v = (v >> TO) << TO;
10364 v = (v << LO) >> LO;
10368 /// isFPImmLegal - Returns true if the target can instruction select the
10369 /// specified FP immediate natively. If false, the legalizer will
10370 /// materialize the FP immediate as a load from a constant pool.
10371 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10372 if (!Subtarget->hasVFP3())
10374 if (VT == MVT::f32)
10375 return ARM_AM::getFP32Imm(Imm) != -1;
10376 if (VT == MVT::f64)
10377 return ARM_AM::getFP64Imm(Imm) != -1;
10381 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10382 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10383 /// specified in the intrinsic calls.
10384 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10386 unsigned Intrinsic) const {
10387 switch (Intrinsic) {
10388 case Intrinsic::arm_neon_vld1:
10389 case Intrinsic::arm_neon_vld2:
10390 case Intrinsic::arm_neon_vld3:
10391 case Intrinsic::arm_neon_vld4:
10392 case Intrinsic::arm_neon_vld2lane:
10393 case Intrinsic::arm_neon_vld3lane:
10394 case Intrinsic::arm_neon_vld4lane: {
10395 Info.opc = ISD::INTRINSIC_W_CHAIN;
10396 // Conservatively set memVT to the entire set of vectors loaded.
10397 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10398 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10399 Info.ptrVal = I.getArgOperand(0);
10401 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10402 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10403 Info.vol = false; // volatile loads with NEON intrinsics not supported
10404 Info.readMem = true;
10405 Info.writeMem = false;
10408 case Intrinsic::arm_neon_vst1:
10409 case Intrinsic::arm_neon_vst2:
10410 case Intrinsic::arm_neon_vst3:
10411 case Intrinsic::arm_neon_vst4:
10412 case Intrinsic::arm_neon_vst2lane:
10413 case Intrinsic::arm_neon_vst3lane:
10414 case Intrinsic::arm_neon_vst4lane: {
10415 Info.opc = ISD::INTRINSIC_VOID;
10416 // Conservatively set memVT to the entire set of vectors stored.
10417 unsigned NumElts = 0;
10418 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10419 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10420 if (!ArgTy->isVectorTy())
10422 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10424 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10425 Info.ptrVal = I.getArgOperand(0);
10427 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10428 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10429 Info.vol = false; // volatile stores with NEON intrinsics not supported
10430 Info.readMem = false;
10431 Info.writeMem = true;
10434 case Intrinsic::arm_ldaex:
10435 case Intrinsic::arm_ldrex: {
10436 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10437 Info.opc = ISD::INTRINSIC_W_CHAIN;
10438 Info.memVT = MVT::getVT(PtrTy->getElementType());
10439 Info.ptrVal = I.getArgOperand(0);
10441 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10443 Info.readMem = true;
10444 Info.writeMem = false;
10447 case Intrinsic::arm_stlex:
10448 case Intrinsic::arm_strex: {
10449 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10450 Info.opc = ISD::INTRINSIC_W_CHAIN;
10451 Info.memVT = MVT::getVT(PtrTy->getElementType());
10452 Info.ptrVal = I.getArgOperand(1);
10454 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10456 Info.readMem = false;
10457 Info.writeMem = true;
10460 case Intrinsic::arm_stlexd:
10461 case Intrinsic::arm_strexd: {
10462 Info.opc = ISD::INTRINSIC_W_CHAIN;
10463 Info.memVT = MVT::i64;
10464 Info.ptrVal = I.getArgOperand(2);
10468 Info.readMem = false;
10469 Info.writeMem = true;
10472 case Intrinsic::arm_ldaexd:
10473 case Intrinsic::arm_ldrexd: {
10474 Info.opc = ISD::INTRINSIC_W_CHAIN;
10475 Info.memVT = MVT::i64;
10476 Info.ptrVal = I.getArgOperand(0);
10480 Info.readMem = true;
10481 Info.writeMem = false;
10491 /// \brief Returns true if it is beneficial to convert a load of a constant
10492 /// to just the constant itself.
10493 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10495 assert(Ty->isIntegerTy());
10497 unsigned Bits = Ty->getPrimitiveSizeInBits();
10498 if (Bits == 0 || Bits > 32)
10503 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10504 // Loads and stores less than 64-bits are already atomic; ones above that
10505 // are doomed anyway, so defer to the default libcall and blame the OS when
10506 // things go wrong:
10507 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
10508 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 64;
10509 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
10510 return LI->getType()->getPrimitiveSizeInBits() == 64;
10512 // For the real atomic operations, we have ldrex/strex up to 64 bits.
10513 return Inst->getType()->getPrimitiveSizeInBits() <= 64;
10516 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10517 AtomicOrdering Ord) const {
10518 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10519 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10521 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10523 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
10524 // intrinsic must return {i32, i32} and we have to recombine them into a
10525 // single i64 here.
10526 if (ValTy->getPrimitiveSizeInBits() == 64) {
10527 Intrinsic::ID Int =
10528 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
10529 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
10531 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10532 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
10534 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10535 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10536 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10537 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10538 return Builder.CreateOr(
10539 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
10542 Type *Tys[] = { Addr->getType() };
10543 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
10544 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10546 return Builder.CreateTruncOrBitCast(
10547 Builder.CreateCall(Ldrex, Addr),
10548 cast<PointerType>(Addr->getType())->getElementType());
10551 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
10553 AtomicOrdering Ord) const {
10554 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10556 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10558 // Since the intrinsics must have legal type, the i64 intrinsics take two
10559 // parameters: "i32, i32". We must marshal Val into the appropriate form
10560 // before the call.
10561 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
10562 Intrinsic::ID Int =
10563 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
10564 Function *Strex = Intrinsic::getDeclaration(M, Int);
10565 Type *Int32Ty = Type::getInt32Ty(M->getContext());
10567 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
10568 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
10569 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10570 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
10573 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
10574 Type *Tys[] = { Addr->getType() };
10575 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
10577 return Builder.CreateCall2(
10578 Strex, Builder.CreateZExtOrBitCast(
10579 Val, Strex->getFunctionType()->getParamType(0)),