1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "MCTargetDesc/ARMAddressingModes.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Target/TargetOptions.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::QPRRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
164 return new TargetLoweringObjectFileMachO();
166 return new ARMElfTargetObjectFile();
169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<ARMSubtarget>();
172 RegInfo = TM.getRegisterInfo();
173 Itins = TM.getInstrItineraryData();
175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
177 if (Subtarget->isTargetDarwin()) {
178 // Uses VFP for Thumb libfuncs if available.
179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
180 // Single-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
186 // Double-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
192 // Single-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
211 // Double-precision comparisons.
212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
230 // Floating-point to integer conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
238 // Conversions between floating types.
239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
242 // Integer to floating-point conversions.
243 // i64 conversions are done via library routines even when generating VFP
244 // instructions, so use the same ones.
245 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
246 // e.g., __floatunsidf vs. __floatunssidfvfp.
247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, 0);
256 setLibcallName(RTLIB::SRL_I128, 0);
257 setLibcallName(RTLIB::SRA_I128, 0);
259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point arithmetic helper functions
299 // RTABI chapter 4.1.2, Table 4
300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
309 // Single-precision floating-point comparison helper functions
310 // RTABI chapter 4.1.2, Table 5
311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
336 // Floating-point to integer conversions.
337 // RTABI chapter 4.1.2, Table 6
338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
355 // Conversions between floating types.
356 // RTABI chapter 4.1.2, Table 7
357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
362 // Integer to floating-point conversions.
363 // RTABI chapter 4.1.2, Table 8
364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 // Long long helper functions
382 // RTABI chapter 4.2, Table 9
383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
394 // Integer division functions
395 // RTABI chapter 4.3.1
396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
414 // RTABI chapter 4.3.4
415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
423 // Use divmod compiler-rt calls for iOS 5.0 and later.
424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430 if (Subtarget->isThumb1Only())
431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
433 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
435 !Subtarget->isThumb1Only()) {
436 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
437 if (!Subtarget->isFPOnlySP())
438 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
447 setTruncStoreAction((MVT::SimpleValueType)VT,
448 (MVT::SimpleValueType)InnerVT, Expand);
449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
456 if (Subtarget->hasNEON()) {
457 addDRTypeForNEON(MVT::v2f32);
458 addDRTypeForNEON(MVT::v8i8);
459 addDRTypeForNEON(MVT::v4i16);
460 addDRTypeForNEON(MVT::v2i32);
461 addDRTypeForNEON(MVT::v1i64);
463 addQRTypeForNEON(MVT::v4f32);
464 addQRTypeForNEON(MVT::v2f64);
465 addQRTypeForNEON(MVT::v16i8);
466 addQRTypeForNEON(MVT::v8i16);
467 addQRTypeForNEON(MVT::v4i32);
468 addQRTypeForNEON(MVT::v2i64);
470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
471 // neither Neon nor VFP support any arithmetic operations on it.
472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
473 // supported for v4f32.
474 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
477 // FIXME: Code duplication: FDIV and FREM are expanded always, see
478 // ARMTargetLowering::addTypeForNEON method for details.
479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
480 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
481 // FIXME: Create unittest.
482 // In another words, find a way when "copysign" appears in DAG with vector
484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
485 // FIXME: Code duplication: SETCC has custom operation action, see
486 // ARMTargetLowering::addTypeForNEON method for details.
487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
488 // FIXME: Create unittest for FNEG and for FABS.
489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
490 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
507 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
525 // Mark v2f32 intrinsics.
526 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
527 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
528 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
529 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
530 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
531 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
532 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
533 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
534 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
535 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
536 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
537 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
538 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
539 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
540 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
542 // Neon does not support some operations on v1i64 and v2i64 types.
543 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
544 // Custom handling for some quad-vector types to detect VMULL.
545 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
546 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
547 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
548 // Custom handling for some vector types to avoid expensive expansions
549 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
550 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
551 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
552 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
553 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
554 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
555 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
556 // a destination type that is wider than the source, and nor does
557 // it have a FP_TO_[SU]INT instruction with a narrower destination than
559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
561 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
564 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
565 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
567 // Custom expand long extensions to vectors.
568 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
569 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
570 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
571 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
572 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
573 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
574 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
575 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
577 // NEON does not have single instruction CTPOP for vectors with element
578 // types wider than 8-bits. However, custom lowering can leverage the
579 // v8i8/v16i8 vcnt instruction.
580 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
581 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
582 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
583 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
585 // NEON only has FMA instructions as of VFP4.
586 if (!Subtarget->hasVFP4()) {
587 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
588 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
591 setTargetDAGCombine(ISD::INTRINSIC_VOID);
592 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
593 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
594 setTargetDAGCombine(ISD::SHL);
595 setTargetDAGCombine(ISD::SRL);
596 setTargetDAGCombine(ISD::SRA);
597 setTargetDAGCombine(ISD::SIGN_EXTEND);
598 setTargetDAGCombine(ISD::ZERO_EXTEND);
599 setTargetDAGCombine(ISD::ANY_EXTEND);
600 setTargetDAGCombine(ISD::SELECT_CC);
601 setTargetDAGCombine(ISD::BUILD_VECTOR);
602 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
603 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
604 setTargetDAGCombine(ISD::STORE);
605 setTargetDAGCombine(ISD::FP_TO_SINT);
606 setTargetDAGCombine(ISD::FP_TO_UINT);
607 setTargetDAGCombine(ISD::FDIV);
609 // It is legal to extload from v4i8 to v4i16 or v4i32.
610 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
611 MVT::v4i16, MVT::v2i16,
613 for (unsigned i = 0; i < 6; ++i) {
614 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
615 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
616 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
620 // ARM and Thumb2 support UMLAL/SMLAL.
621 if (!Subtarget->isThumb1Only())
622 setTargetDAGCombine(ISD::ADDC);
625 computeRegisterProperties();
627 // ARM does not have f32 extending load.
628 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
630 // ARM does not have i1 sign extending load.
631 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
633 // ARM supports all 4 flavors of integer indexed load / store.
634 if (!Subtarget->isThumb1Only()) {
635 for (unsigned im = (unsigned)ISD::PRE_INC;
636 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
637 setIndexedLoadAction(im, MVT::i1, Legal);
638 setIndexedLoadAction(im, MVT::i8, Legal);
639 setIndexedLoadAction(im, MVT::i16, Legal);
640 setIndexedLoadAction(im, MVT::i32, Legal);
641 setIndexedStoreAction(im, MVT::i1, Legal);
642 setIndexedStoreAction(im, MVT::i8, Legal);
643 setIndexedStoreAction(im, MVT::i16, Legal);
644 setIndexedStoreAction(im, MVT::i32, Legal);
648 // i64 operation support.
649 setOperationAction(ISD::MUL, MVT::i64, Expand);
650 setOperationAction(ISD::MULHU, MVT::i32, Expand);
651 if (Subtarget->isThumb1Only()) {
652 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
653 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
655 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
656 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
657 setOperationAction(ISD::MULHS, MVT::i32, Expand);
659 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
660 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
661 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
662 setOperationAction(ISD::SRL, MVT::i64, Custom);
663 setOperationAction(ISD::SRA, MVT::i64, Custom);
665 if (!Subtarget->isThumb1Only()) {
666 // FIXME: We should do this for Thumb1 as well.
667 setOperationAction(ISD::ADDC, MVT::i32, Custom);
668 setOperationAction(ISD::ADDE, MVT::i32, Custom);
669 setOperationAction(ISD::SUBC, MVT::i32, Custom);
670 setOperationAction(ISD::SUBE, MVT::i32, Custom);
673 // ARM does not have ROTL.
674 setOperationAction(ISD::ROTL, MVT::i32, Expand);
675 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
676 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
677 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
678 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
680 // These just redirect to CTTZ and CTLZ on ARM.
681 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
682 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
684 // Only ARMv6 has BSWAP.
685 if (!Subtarget->hasV6Ops())
686 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
688 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
689 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
690 // These are expanded into libcalls if the cpu doesn't have HW divider.
691 setOperationAction(ISD::SDIV, MVT::i32, Expand);
692 setOperationAction(ISD::UDIV, MVT::i32, Expand);
694 setOperationAction(ISD::SREM, MVT::i32, Expand);
695 setOperationAction(ISD::UREM, MVT::i32, Expand);
696 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
697 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
699 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
700 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
701 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
702 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
703 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
705 setOperationAction(ISD::TRAP, MVT::Other, Legal);
707 // Use the default implementation.
708 setOperationAction(ISD::VASTART, MVT::Other, Custom);
709 setOperationAction(ISD::VAARG, MVT::Other, Expand);
710 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
711 setOperationAction(ISD::VAEND, MVT::Other, Expand);
712 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
713 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
715 if (!Subtarget->isTargetDarwin()) {
716 // Non-Darwin platforms may return values in these registers via the
717 // personality function.
718 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
719 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
720 setExceptionPointerRegister(ARM::R0);
721 setExceptionSelectorRegister(ARM::R1);
724 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
725 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
726 // the default expansion.
727 // FIXME: This should be checking for v6k, not just v6.
728 if (Subtarget->hasDataBarrier() ||
729 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
730 // membarrier needs custom lowering; the rest are legal and handled
732 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
733 // Custom lowering for 64-bit ops
734 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
735 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
736 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
737 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
738 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
739 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
740 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
741 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
742 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
743 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
744 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
745 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
746 setInsertFencesForAtomic(true);
748 // Set them all for expansion, which will force libcalls.
749 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
750 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
751 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
752 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
753 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
754 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
755 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
756 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
757 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
758 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
759 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
760 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
761 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
762 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
763 // Unordered/Monotonic case.
764 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
765 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
768 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
770 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
771 if (!Subtarget->hasV6Ops()) {
772 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
773 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
775 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
777 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
778 !Subtarget->isThumb1Only()) {
779 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
780 // iff target supports vfp2.
781 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
782 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
785 // We want to custom lower some of our intrinsics.
786 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
787 if (Subtarget->isTargetDarwin()) {
788 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
789 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
790 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
793 setOperationAction(ISD::SETCC, MVT::i32, Expand);
794 setOperationAction(ISD::SETCC, MVT::f32, Expand);
795 setOperationAction(ISD::SETCC, MVT::f64, Expand);
796 setOperationAction(ISD::SELECT, MVT::i32, Custom);
797 setOperationAction(ISD::SELECT, MVT::f32, Custom);
798 setOperationAction(ISD::SELECT, MVT::f64, Custom);
799 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
800 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
801 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
803 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
804 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
805 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
806 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
807 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
809 // We don't support sin/cos/fmod/copysign/pow
810 setOperationAction(ISD::FSIN, MVT::f64, Expand);
811 setOperationAction(ISD::FSIN, MVT::f32, Expand);
812 setOperationAction(ISD::FCOS, MVT::f32, Expand);
813 setOperationAction(ISD::FCOS, MVT::f64, Expand);
814 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
815 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
816 setOperationAction(ISD::FREM, MVT::f64, Expand);
817 setOperationAction(ISD::FREM, MVT::f32, Expand);
818 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
819 !Subtarget->isThumb1Only()) {
820 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
821 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
823 setOperationAction(ISD::FPOW, MVT::f64, Expand);
824 setOperationAction(ISD::FPOW, MVT::f32, Expand);
826 if (!Subtarget->hasVFP4()) {
827 setOperationAction(ISD::FMA, MVT::f64, Expand);
828 setOperationAction(ISD::FMA, MVT::f32, Expand);
831 // Various VFP goodness
832 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
833 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
834 if (Subtarget->hasVFP2()) {
835 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
836 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
837 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
838 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
840 // Special handling for half-precision FP.
841 if (!Subtarget->hasFP16()) {
842 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
843 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
847 // We have target-specific dag combine patterns for the following nodes:
848 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
849 setTargetDAGCombine(ISD::ADD);
850 setTargetDAGCombine(ISD::SUB);
851 setTargetDAGCombine(ISD::MUL);
852 setTargetDAGCombine(ISD::AND);
853 setTargetDAGCombine(ISD::OR);
854 setTargetDAGCombine(ISD::XOR);
856 if (Subtarget->hasV6Ops())
857 setTargetDAGCombine(ISD::SRL);
859 setStackPointerRegisterToSaveRestore(ARM::SP);
861 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
862 !Subtarget->hasVFP2())
863 setSchedulingPreference(Sched::RegPressure);
865 setSchedulingPreference(Sched::Hybrid);
867 //// temporary - rewrite interface to use type
868 MaxStoresPerMemset = 8;
869 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
870 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
871 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
872 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
873 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
875 // On ARM arguments smaller than 4 bytes are extended, so all arguments
876 // are at least 4 bytes aligned.
877 setMinStackArgumentAlignment(4);
879 // Prefer likely predicted branches to selects on out-of-order cores.
880 PredictableSelectIsExpensive = Subtarget->isLikeA9();
882 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
885 // FIXME: It might make sense to define the representative register class as the
886 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
887 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
888 // SPR's representative would be DPR_VFP2. This should work well if register
889 // pressure tracking were modified such that a register use would increment the
890 // pressure of the register class's representative and all of it's super
891 // classes' representatives transitively. We have not implemented this because
892 // of the difficulty prior to coalescing of modeling operand register classes
893 // due to the common occurrence of cross class copies and subregister insertions
895 std::pair<const TargetRegisterClass*, uint8_t>
896 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
897 const TargetRegisterClass *RRC = 0;
899 switch (VT.SimpleTy) {
901 return TargetLowering::findRepresentativeClass(VT);
902 // Use DPR as representative register class for all floating point
903 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
904 // the cost is 1 for both f32 and f64.
905 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
906 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
907 RRC = &ARM::DPRRegClass;
908 // When NEON is used for SP, only half of the register file is available
909 // because operations that define both SP and DP results will be constrained
910 // to the VFP2 class (D0-D15). We currently model this constraint prior to
911 // coalescing by double-counting the SP regs. See the FIXME above.
912 if (Subtarget->useNEONForSinglePrecisionFP())
915 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
916 case MVT::v4f32: case MVT::v2f64:
917 RRC = &ARM::DPRRegClass;
921 RRC = &ARM::DPRRegClass;
925 RRC = &ARM::DPRRegClass;
929 return std::make_pair(RRC, Cost);
932 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
935 case ARMISD::Wrapper: return "ARMISD::Wrapper";
936 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
937 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
938 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
939 case ARMISD::CALL: return "ARMISD::CALL";
940 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
941 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
942 case ARMISD::tCALL: return "ARMISD::tCALL";
943 case ARMISD::BRCOND: return "ARMISD::BRCOND";
944 case ARMISD::BR_JT: return "ARMISD::BR_JT";
945 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
946 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
947 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
948 case ARMISD::CMP: return "ARMISD::CMP";
949 case ARMISD::CMN: return "ARMISD::CMN";
950 case ARMISD::CMPZ: return "ARMISD::CMPZ";
951 case ARMISD::CMPFP: return "ARMISD::CMPFP";
952 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
953 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
954 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
956 case ARMISD::CMOV: return "ARMISD::CMOV";
958 case ARMISD::RBIT: return "ARMISD::RBIT";
960 case ARMISD::FTOSI: return "ARMISD::FTOSI";
961 case ARMISD::FTOUI: return "ARMISD::FTOUI";
962 case ARMISD::SITOF: return "ARMISD::SITOF";
963 case ARMISD::UITOF: return "ARMISD::UITOF";
965 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
966 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
967 case ARMISD::RRX: return "ARMISD::RRX";
969 case ARMISD::ADDC: return "ARMISD::ADDC";
970 case ARMISD::ADDE: return "ARMISD::ADDE";
971 case ARMISD::SUBC: return "ARMISD::SUBC";
972 case ARMISD::SUBE: return "ARMISD::SUBE";
974 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
975 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
977 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
978 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
980 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
982 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
984 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
986 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
987 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
989 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
991 case ARMISD::VCEQ: return "ARMISD::VCEQ";
992 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
993 case ARMISD::VCGE: return "ARMISD::VCGE";
994 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
995 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
996 case ARMISD::VCGEU: return "ARMISD::VCGEU";
997 case ARMISD::VCGT: return "ARMISD::VCGT";
998 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
999 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1000 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1001 case ARMISD::VTST: return "ARMISD::VTST";
1003 case ARMISD::VSHL: return "ARMISD::VSHL";
1004 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1005 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1006 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1007 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1008 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1009 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1010 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1011 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1012 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1013 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1014 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1015 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1016 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1017 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1018 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1019 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1020 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1021 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1022 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1023 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1024 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1025 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1026 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1027 case ARMISD::VDUP: return "ARMISD::VDUP";
1028 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1029 case ARMISD::VEXT: return "ARMISD::VEXT";
1030 case ARMISD::VREV64: return "ARMISD::VREV64";
1031 case ARMISD::VREV32: return "ARMISD::VREV32";
1032 case ARMISD::VREV16: return "ARMISD::VREV16";
1033 case ARMISD::VZIP: return "ARMISD::VZIP";
1034 case ARMISD::VUZP: return "ARMISD::VUZP";
1035 case ARMISD::VTRN: return "ARMISD::VTRN";
1036 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1037 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1038 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1039 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1040 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1041 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1042 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1043 case ARMISD::FMAX: return "ARMISD::FMAX";
1044 case ARMISD::FMIN: return "ARMISD::FMIN";
1045 case ARMISD::BFI: return "ARMISD::BFI";
1046 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1047 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1048 case ARMISD::VBSL: return "ARMISD::VBSL";
1049 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1050 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1051 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1052 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1053 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1054 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1055 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1056 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1057 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1058 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1059 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1060 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1061 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1062 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1063 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1064 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1065 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1066 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1067 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1068 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1072 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1073 if (!VT.isVector()) return getPointerTy();
1074 return VT.changeVectorElementTypeToInteger();
1077 /// getRegClassFor - Return the register class that should be used for the
1078 /// specified value type.
1079 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1080 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1081 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1082 // load / store 4 to 8 consecutive D registers.
1083 if (Subtarget->hasNEON()) {
1084 if (VT == MVT::v4i64)
1085 return &ARM::QQPRRegClass;
1086 if (VT == MVT::v8i64)
1087 return &ARM::QQQQPRRegClass;
1089 return TargetLowering::getRegClassFor(VT);
1092 // Create a fast isel object.
1094 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1095 const TargetLibraryInfo *libInfo) const {
1096 return ARM::createFastISel(funcInfo, libInfo);
1099 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1100 /// be used for loads / stores from the global.
1101 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1102 return (Subtarget->isThumb1Only() ? 127 : 4095);
1105 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1106 unsigned NumVals = N->getNumValues();
1108 return Sched::RegPressure;
1110 for (unsigned i = 0; i != NumVals; ++i) {
1111 EVT VT = N->getValueType(i);
1112 if (VT == MVT::Glue || VT == MVT::Other)
1114 if (VT.isFloatingPoint() || VT.isVector())
1118 if (!N->isMachineOpcode())
1119 return Sched::RegPressure;
1121 // Load are scheduled for latency even if there instruction itinerary
1122 // is not available.
1123 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1124 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1126 if (MCID.getNumDefs() == 0)
1127 return Sched::RegPressure;
1128 if (!Itins->isEmpty() &&
1129 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1132 return Sched::RegPressure;
1135 //===----------------------------------------------------------------------===//
1137 //===----------------------------------------------------------------------===//
1139 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1140 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1142 default: llvm_unreachable("Unknown condition code!");
1143 case ISD::SETNE: return ARMCC::NE;
1144 case ISD::SETEQ: return ARMCC::EQ;
1145 case ISD::SETGT: return ARMCC::GT;
1146 case ISD::SETGE: return ARMCC::GE;
1147 case ISD::SETLT: return ARMCC::LT;
1148 case ISD::SETLE: return ARMCC::LE;
1149 case ISD::SETUGT: return ARMCC::HI;
1150 case ISD::SETUGE: return ARMCC::HS;
1151 case ISD::SETULT: return ARMCC::LO;
1152 case ISD::SETULE: return ARMCC::LS;
1156 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1157 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1158 ARMCC::CondCodes &CondCode2) {
1159 CondCode2 = ARMCC::AL;
1161 default: llvm_unreachable("Unknown FP condition!");
1163 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1165 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1167 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1168 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1169 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1170 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1171 case ISD::SETO: CondCode = ARMCC::VC; break;
1172 case ISD::SETUO: CondCode = ARMCC::VS; break;
1173 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1174 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1175 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1177 case ISD::SETULT: CondCode = ARMCC::LT; break;
1179 case ISD::SETULE: CondCode = ARMCC::LE; break;
1181 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1185 //===----------------------------------------------------------------------===//
1186 // Calling Convention Implementation
1187 //===----------------------------------------------------------------------===//
1189 #include "ARMGenCallingConv.inc"
1191 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1192 /// given CallingConvention value.
1193 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1195 bool isVarArg) const {
1198 llvm_unreachable("Unsupported calling convention");
1199 case CallingConv::Fast:
1200 if (Subtarget->hasVFP2() && !isVarArg) {
1201 if (!Subtarget->isAAPCS_ABI())
1202 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1203 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1204 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1207 case CallingConv::C: {
1208 // Use target triple & subtarget features to do actual dispatch.
1209 if (!Subtarget->isAAPCS_ABI())
1210 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1211 else if (Subtarget->hasVFP2() &&
1212 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1214 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1215 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1217 case CallingConv::ARM_AAPCS_VFP:
1219 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1221 case CallingConv::ARM_AAPCS:
1222 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1223 case CallingConv::ARM_APCS:
1224 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1225 case CallingConv::GHC:
1226 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1230 /// LowerCallResult - Lower the result values of a call into the
1231 /// appropriate copies out of appropriate physical registers.
1233 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1234 CallingConv::ID CallConv, bool isVarArg,
1235 const SmallVectorImpl<ISD::InputArg> &Ins,
1236 DebugLoc dl, SelectionDAG &DAG,
1237 SmallVectorImpl<SDValue> &InVals,
1238 bool isThisReturn, SDValue ThisVal) const {
1240 // Assign locations to each value returned by this call.
1241 SmallVector<CCValAssign, 16> RVLocs;
1242 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1243 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1244 CCInfo.AnalyzeCallResult(Ins,
1245 CCAssignFnForNode(CallConv, /* Return*/ true,
1248 // Copy all of the result registers out of their specified physreg.
1249 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1250 CCValAssign VA = RVLocs[i];
1252 // Pass 'this' value directly from the argument to return value, to avoid
1253 // reg unit interference
1254 if (i == 0 && isThisReturn) {
1255 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32);
1256 InVals.push_back(ThisVal);
1261 if (VA.needsCustom()) {
1262 // Handle f64 or half of a v2f64.
1263 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1265 Chain = Lo.getValue(1);
1266 InFlag = Lo.getValue(2);
1267 VA = RVLocs[++i]; // skip ahead to next loc
1268 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1270 Chain = Hi.getValue(1);
1271 InFlag = Hi.getValue(2);
1272 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1274 if (VA.getLocVT() == MVT::v2f64) {
1275 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1276 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1277 DAG.getConstant(0, MVT::i32));
1279 VA = RVLocs[++i]; // skip ahead to next loc
1280 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1281 Chain = Lo.getValue(1);
1282 InFlag = Lo.getValue(2);
1283 VA = RVLocs[++i]; // skip ahead to next loc
1284 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1285 Chain = Hi.getValue(1);
1286 InFlag = Hi.getValue(2);
1287 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1288 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1289 DAG.getConstant(1, MVT::i32));
1292 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1294 Chain = Val.getValue(1);
1295 InFlag = Val.getValue(2);
1298 switch (VA.getLocInfo()) {
1299 default: llvm_unreachable("Unknown loc info!");
1300 case CCValAssign::Full: break;
1301 case CCValAssign::BCvt:
1302 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1306 InVals.push_back(Val);
1312 /// LowerMemOpCallTo - Store the argument to the stack.
1314 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1315 SDValue StackPtr, SDValue Arg,
1316 DebugLoc dl, SelectionDAG &DAG,
1317 const CCValAssign &VA,
1318 ISD::ArgFlagsTy Flags) const {
1319 unsigned LocMemOffset = VA.getLocMemOffset();
1320 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1321 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1322 return DAG.getStore(Chain, dl, Arg, PtrOff,
1323 MachinePointerInfo::getStack(LocMemOffset),
1327 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1328 SDValue Chain, SDValue &Arg,
1329 RegsToPassVector &RegsToPass,
1330 CCValAssign &VA, CCValAssign &NextVA,
1332 SmallVector<SDValue, 8> &MemOpChains,
1333 ISD::ArgFlagsTy Flags) const {
1335 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1336 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1337 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1339 if (NextVA.isRegLoc())
1340 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1342 assert(NextVA.isMemLoc());
1343 if (StackPtr.getNode() == 0)
1344 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1346 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1352 /// LowerCall - Lowering a call into a callseq_start <-
1353 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1356 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1357 SmallVectorImpl<SDValue> &InVals) const {
1358 SelectionDAG &DAG = CLI.DAG;
1359 DebugLoc &dl = CLI.DL;
1360 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1361 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1362 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1363 SDValue Chain = CLI.Chain;
1364 SDValue Callee = CLI.Callee;
1365 bool &isTailCall = CLI.IsTailCall;
1366 CallingConv::ID CallConv = CLI.CallConv;
1367 bool doesNotRet = CLI.DoesNotReturn;
1368 bool isVarArg = CLI.IsVarArg;
1370 MachineFunction &MF = DAG.getMachineFunction();
1371 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1372 bool IsThisReturn = false;
1373 bool IsSibCall = false;
1374 // Disable tail calls if they're not supported.
1375 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1378 // Check if it's really possible to do a tail call.
1379 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1380 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1381 Outs, OutVals, Ins, DAG);
1382 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1383 // detected sibcalls.
1390 // Analyze operands of the call, assigning locations to each operand.
1391 SmallVector<CCValAssign, 16> ArgLocs;
1392 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1393 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1394 CCInfo.AnalyzeCallOperands(Outs,
1395 CCAssignFnForNode(CallConv, /* Return*/ false,
1398 // Get a count of how many bytes are to be pushed on the stack.
1399 unsigned NumBytes = CCInfo.getNextStackOffset();
1401 // For tail calls, memory operands are available in our caller's stack.
1405 // Adjust the stack pointer for the new arguments...
1406 // These operations are automatically eliminated by the prolog/epilog pass
1408 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1410 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1412 RegsToPassVector RegsToPass;
1413 SmallVector<SDValue, 8> MemOpChains;
1415 // Walk the register/memloc assignments, inserting copies/loads. In the case
1416 // of tail call optimization, arguments are handled later.
1417 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1419 ++i, ++realArgIdx) {
1420 CCValAssign &VA = ArgLocs[i];
1421 SDValue Arg = OutVals[realArgIdx];
1422 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1423 bool isByVal = Flags.isByVal();
1425 // Promote the value if needed.
1426 switch (VA.getLocInfo()) {
1427 default: llvm_unreachable("Unknown loc info!");
1428 case CCValAssign::Full: break;
1429 case CCValAssign::SExt:
1430 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1432 case CCValAssign::ZExt:
1433 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1435 case CCValAssign::AExt:
1436 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1438 case CCValAssign::BCvt:
1439 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1443 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1444 if (VA.needsCustom()) {
1445 if (VA.getLocVT() == MVT::v2f64) {
1446 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1447 DAG.getConstant(0, MVT::i32));
1448 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1449 DAG.getConstant(1, MVT::i32));
1451 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1452 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1454 VA = ArgLocs[++i]; // skip ahead to next loc
1455 if (VA.isRegLoc()) {
1456 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1457 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1459 assert(VA.isMemLoc());
1461 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1462 dl, DAG, VA, Flags));
1465 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1466 StackPtr, MemOpChains, Flags);
1468 } else if (VA.isRegLoc()) {
1469 if (realArgIdx == 0 && Flags.isReturned() && VA.getLocVT() == MVT::i32) {
1470 assert(!Ins.empty() && Ins[0].VT == Outs[0].VT &&
1471 "unexpected use of 'returned'");
1472 IsThisReturn = true;
1474 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1475 } else if (isByVal) {
1476 assert(VA.isMemLoc());
1477 unsigned offset = 0;
1479 // True if this byval aggregate will be split between registers
1481 if (CCInfo.isFirstByValRegValid()) {
1482 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1484 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1485 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1486 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1487 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1488 MachinePointerInfo(),
1489 false, false, false, 0);
1490 MemOpChains.push_back(Load.getValue(1));
1491 RegsToPass.push_back(std::make_pair(j, Load));
1493 offset = ARM::R4 - CCInfo.getFirstByValReg();
1494 CCInfo.clearFirstByValReg();
1497 if (Flags.getByValSize() - 4*offset > 0) {
1498 unsigned LocMemOffset = VA.getLocMemOffset();
1499 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1500 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1502 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1503 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1504 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1506 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1508 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1509 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1510 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1511 Ops, array_lengthof(Ops)));
1513 } else if (!IsSibCall) {
1514 assert(VA.isMemLoc());
1516 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1517 dl, DAG, VA, Flags));
1521 if (!MemOpChains.empty())
1522 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1523 &MemOpChains[0], MemOpChains.size());
1525 // Build a sequence of copy-to-reg nodes chained together with token chain
1526 // and flag operands which copy the outgoing args into the appropriate regs.
1528 // Tail call byval lowering might overwrite argument registers so in case of
1529 // tail call optimization the copies to registers are lowered later.
1531 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1532 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1533 RegsToPass[i].second, InFlag);
1534 InFlag = Chain.getValue(1);
1537 // For tail calls lower the arguments to the 'real' stack slot.
1539 // Force all the incoming stack arguments to be loaded from the stack
1540 // before any new outgoing arguments are stored to the stack, because the
1541 // outgoing stack slots may alias the incoming argument stack slots, and
1542 // the alias isn't otherwise explicit. This is slightly more conservative
1543 // than necessary, because it means that each store effectively depends
1544 // on every argument instead of just those arguments it would clobber.
1546 // Do not flag preceding copytoreg stuff together with the following stuff.
1548 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1549 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1550 RegsToPass[i].second, InFlag);
1551 InFlag = Chain.getValue(1);
1556 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1557 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1558 // node so that legalize doesn't hack it.
1559 bool isDirect = false;
1560 bool isARMFunc = false;
1561 bool isLocalARMFunc = false;
1562 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1564 if (EnableARMLongCalls) {
1565 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1566 && "long-calls with non-static relocation model!");
1567 // Handle a global address or an external symbol. If it's not one of
1568 // those, the target's already in a register, so we don't need to do
1570 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1571 const GlobalValue *GV = G->getGlobal();
1572 // Create a constant pool entry for the callee address
1573 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1574 ARMConstantPoolValue *CPV =
1575 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1577 // Get the address of the callee into a register
1578 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1579 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1580 Callee = DAG.getLoad(getPointerTy(), dl,
1581 DAG.getEntryNode(), CPAddr,
1582 MachinePointerInfo::getConstantPool(),
1583 false, false, false, 0);
1584 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1585 const char *Sym = S->getSymbol();
1587 // Create a constant pool entry for the callee address
1588 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1589 ARMConstantPoolValue *CPV =
1590 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1591 ARMPCLabelIndex, 0);
1592 // Get the address of the callee into a register
1593 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1594 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1595 Callee = DAG.getLoad(getPointerTy(), dl,
1596 DAG.getEntryNode(), CPAddr,
1597 MachinePointerInfo::getConstantPool(),
1598 false, false, false, 0);
1600 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1601 const GlobalValue *GV = G->getGlobal();
1603 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1604 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1605 getTargetMachine().getRelocationModel() != Reloc::Static;
1606 isARMFunc = !Subtarget->isThumb() || isStub;
1607 // ARM call to a local ARM function is predicable.
1608 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1609 // tBX takes a register source operand.
1610 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1611 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1612 ARMConstantPoolValue *CPV =
1613 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1614 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1615 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1616 Callee = DAG.getLoad(getPointerTy(), dl,
1617 DAG.getEntryNode(), CPAddr,
1618 MachinePointerInfo::getConstantPool(),
1619 false, false, false, 0);
1620 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1621 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1622 getPointerTy(), Callee, PICLabel);
1624 // On ELF targets for PIC code, direct calls should go through the PLT
1625 unsigned OpFlags = 0;
1626 if (Subtarget->isTargetELF() &&
1627 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1628 OpFlags = ARMII::MO_PLT;
1629 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1631 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1633 bool isStub = Subtarget->isTargetDarwin() &&
1634 getTargetMachine().getRelocationModel() != Reloc::Static;
1635 isARMFunc = !Subtarget->isThumb() || isStub;
1636 // tBX takes a register source operand.
1637 const char *Sym = S->getSymbol();
1638 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1639 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1640 ARMConstantPoolValue *CPV =
1641 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1642 ARMPCLabelIndex, 4);
1643 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1644 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1645 Callee = DAG.getLoad(getPointerTy(), dl,
1646 DAG.getEntryNode(), CPAddr,
1647 MachinePointerInfo::getConstantPool(),
1648 false, false, false, 0);
1649 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1650 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1651 getPointerTy(), Callee, PICLabel);
1653 unsigned OpFlags = 0;
1654 // On ELF targets for PIC code, direct calls should go through the PLT
1655 if (Subtarget->isTargetELF() &&
1656 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1657 OpFlags = ARMII::MO_PLT;
1658 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1662 // FIXME: handle tail calls differently.
1664 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1665 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1666 if (Subtarget->isThumb()) {
1667 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1668 CallOpc = ARMISD::CALL_NOLINK;
1670 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1672 if (!isDirect && !Subtarget->hasV5TOps())
1673 CallOpc = ARMISD::CALL_NOLINK;
1674 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1675 // Emit regular call when code size is the priority
1677 // "mov lr, pc; b _foo" to avoid confusing the RSP
1678 CallOpc = ARMISD::CALL_NOLINK;
1680 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1683 std::vector<SDValue> Ops;
1684 Ops.push_back(Chain);
1685 Ops.push_back(Callee);
1687 // Add argument registers to the end of the list so that they are known live
1689 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1690 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1691 RegsToPass[i].second.getValueType()));
1693 // Add a register mask operand representing the call-preserved registers.
1694 const uint32_t *Mask;
1695 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1696 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1698 // For 'this' returns, use the R0-preserving mask
1699 Mask = ARI->getThisReturnPreservedMask(CallConv);
1701 Mask = ARI->getCallPreservedMask(CallConv);
1703 assert(Mask && "Missing call preserved mask for calling convention");
1704 Ops.push_back(DAG.getRegisterMask(Mask));
1706 if (InFlag.getNode())
1707 Ops.push_back(InFlag);
1709 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1711 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1713 // Returns a chain and a flag for retval copy to use.
1714 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1715 InFlag = Chain.getValue(1);
1717 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1718 DAG.getIntPtrConstant(0, true), InFlag);
1720 InFlag = Chain.getValue(1);
1722 // Handle result values, copying them out of physregs into vregs that we
1724 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1725 InVals, IsThisReturn,
1726 IsThisReturn ? OutVals[0] : SDValue());
1729 /// HandleByVal - Every parameter *after* a byval parameter is passed
1730 /// on the stack. Remember the next parameter register to allocate,
1731 /// and then confiscate the rest of the parameter registers to insure
1734 ARMTargetLowering::HandleByVal(
1735 CCState *State, unsigned &size, unsigned Align) const {
1736 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1737 assert((State->getCallOrPrologue() == Prologue ||
1738 State->getCallOrPrologue() == Call) &&
1739 "unhandled ParmContext");
1740 if ((!State->isFirstByValRegValid()) &&
1741 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1742 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1743 unsigned AlignInRegs = Align / 4;
1744 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1745 for (unsigned i = 0; i < Waste; ++i)
1746 reg = State->AllocateReg(GPRArgRegs, 4);
1749 State->setFirstByValReg(reg);
1750 // At a call site, a byval parameter that is split between
1751 // registers and memory needs its size truncated here. In a
1752 // function prologue, such byval parameters are reassembled in
1753 // memory, and are not truncated.
1754 if (State->getCallOrPrologue() == Call) {
1755 unsigned excess = 4 * (ARM::R4 - reg);
1756 assert(size >= excess && "expected larger existing stack allocation");
1761 // Confiscate any remaining parameter registers to preclude their
1762 // assignment to subsequent parameters.
1763 while (State->AllocateReg(GPRArgRegs, 4))
1767 /// MatchingStackOffset - Return true if the given stack call argument is
1768 /// already available in the same position (relatively) of the caller's
1769 /// incoming argument stack.
1771 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1772 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1773 const TargetInstrInfo *TII) {
1774 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1776 if (Arg.getOpcode() == ISD::CopyFromReg) {
1777 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1778 if (!TargetRegisterInfo::isVirtualRegister(VR))
1780 MachineInstr *Def = MRI->getVRegDef(VR);
1783 if (!Flags.isByVal()) {
1784 if (!TII->isLoadFromStackSlot(Def, FI))
1789 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1790 if (Flags.isByVal())
1791 // ByVal argument is passed in as a pointer but it's now being
1792 // dereferenced. e.g.
1793 // define @foo(%struct.X* %A) {
1794 // tail call @bar(%struct.X* byval %A)
1797 SDValue Ptr = Ld->getBasePtr();
1798 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1801 FI = FINode->getIndex();
1805 assert(FI != INT_MAX);
1806 if (!MFI->isFixedObjectIndex(FI))
1808 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1811 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1812 /// for tail call optimization. Targets which want to do tail call
1813 /// optimization should implement this function.
1815 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1816 CallingConv::ID CalleeCC,
1818 bool isCalleeStructRet,
1819 bool isCallerStructRet,
1820 const SmallVectorImpl<ISD::OutputArg> &Outs,
1821 const SmallVectorImpl<SDValue> &OutVals,
1822 const SmallVectorImpl<ISD::InputArg> &Ins,
1823 SelectionDAG& DAG) const {
1824 const Function *CallerF = DAG.getMachineFunction().getFunction();
1825 CallingConv::ID CallerCC = CallerF->getCallingConv();
1826 bool CCMatch = CallerCC == CalleeCC;
1828 // Look for obvious safe cases to perform tail call optimization that do not
1829 // require ABI changes. This is what gcc calls sibcall.
1831 // Do not sibcall optimize vararg calls unless the call site is not passing
1833 if (isVarArg && !Outs.empty())
1836 // Also avoid sibcall optimization if either caller or callee uses struct
1837 // return semantics.
1838 if (isCalleeStructRet || isCallerStructRet)
1841 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1842 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1843 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1844 // support in the assembler and linker to be used. This would need to be
1845 // fixed to fully support tail calls in Thumb1.
1847 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1848 // LR. This means if we need to reload LR, it takes an extra instructions,
1849 // which outweighs the value of the tail call; but here we don't know yet
1850 // whether LR is going to be used. Probably the right approach is to
1851 // generate the tail call here and turn it back into CALL/RET in
1852 // emitEpilogue if LR is used.
1854 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1855 // but we need to make sure there are enough registers; the only valid
1856 // registers are the 4 used for parameters. We don't currently do this
1858 if (Subtarget->isThumb1Only())
1861 // If the calling conventions do not match, then we'd better make sure the
1862 // results are returned in the same way as what the caller expects.
1864 SmallVector<CCValAssign, 16> RVLocs1;
1865 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1866 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1867 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1869 SmallVector<CCValAssign, 16> RVLocs2;
1870 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1871 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1872 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1874 if (RVLocs1.size() != RVLocs2.size())
1876 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1877 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1879 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1881 if (RVLocs1[i].isRegLoc()) {
1882 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1885 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1891 // If Caller's vararg or byval argument has been split between registers and
1892 // stack, do not perform tail call, since part of the argument is in caller's
1894 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1895 getInfo<ARMFunctionInfo>();
1896 if (AFI_Caller->getVarArgsRegSaveSize())
1899 // If the callee takes no arguments then go on to check the results of the
1901 if (!Outs.empty()) {
1902 // Check if stack adjustment is needed. For now, do not do this if any
1903 // argument is passed on the stack.
1904 SmallVector<CCValAssign, 16> ArgLocs;
1905 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1906 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1907 CCInfo.AnalyzeCallOperands(Outs,
1908 CCAssignFnForNode(CalleeCC, false, isVarArg));
1909 if (CCInfo.getNextStackOffset()) {
1910 MachineFunction &MF = DAG.getMachineFunction();
1912 // Check if the arguments are already laid out in the right way as
1913 // the caller's fixed stack objects.
1914 MachineFrameInfo *MFI = MF.getFrameInfo();
1915 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1916 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1917 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1919 ++i, ++realArgIdx) {
1920 CCValAssign &VA = ArgLocs[i];
1921 EVT RegVT = VA.getLocVT();
1922 SDValue Arg = OutVals[realArgIdx];
1923 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1924 if (VA.getLocInfo() == CCValAssign::Indirect)
1926 if (VA.needsCustom()) {
1927 // f64 and vector types are split into multiple registers or
1928 // register/stack-slot combinations. The types will not match
1929 // the registers; give up on memory f64 refs until we figure
1930 // out what to do about this.
1933 if (!ArgLocs[++i].isRegLoc())
1935 if (RegVT == MVT::v2f64) {
1936 if (!ArgLocs[++i].isRegLoc())
1938 if (!ArgLocs[++i].isRegLoc())
1941 } else if (!VA.isRegLoc()) {
1942 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1954 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1955 MachineFunction &MF, bool isVarArg,
1956 const SmallVectorImpl<ISD::OutputArg> &Outs,
1957 LLVMContext &Context) const {
1958 SmallVector<CCValAssign, 16> RVLocs;
1959 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
1960 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
1965 ARMTargetLowering::LowerReturn(SDValue Chain,
1966 CallingConv::ID CallConv, bool isVarArg,
1967 const SmallVectorImpl<ISD::OutputArg> &Outs,
1968 const SmallVectorImpl<SDValue> &OutVals,
1969 DebugLoc dl, SelectionDAG &DAG) const {
1971 // CCValAssign - represent the assignment of the return value to a location.
1972 SmallVector<CCValAssign, 16> RVLocs;
1974 // CCState - Info about the registers and stack slots.
1975 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1976 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1978 // Analyze outgoing return values.
1979 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1983 SmallVector<SDValue, 4> RetOps;
1984 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1986 // Copy the result values into the output registers.
1987 for (unsigned i = 0, realRVLocIdx = 0;
1989 ++i, ++realRVLocIdx) {
1990 CCValAssign &VA = RVLocs[i];
1991 assert(VA.isRegLoc() && "Can only return in registers!");
1993 SDValue Arg = OutVals[realRVLocIdx];
1995 switch (VA.getLocInfo()) {
1996 default: llvm_unreachable("Unknown loc info!");
1997 case CCValAssign::Full: break;
1998 case CCValAssign::BCvt:
1999 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2003 if (VA.needsCustom()) {
2004 if (VA.getLocVT() == MVT::v2f64) {
2005 // Extract the first half and return it in two registers.
2006 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2007 DAG.getConstant(0, MVT::i32));
2008 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2009 DAG.getVTList(MVT::i32, MVT::i32), Half);
2011 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2012 Flag = Chain.getValue(1);
2013 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2014 VA = RVLocs[++i]; // skip ahead to next loc
2015 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2016 HalfGPRs.getValue(1), Flag);
2017 Flag = Chain.getValue(1);
2018 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2019 VA = RVLocs[++i]; // skip ahead to next loc
2021 // Extract the 2nd half and fall through to handle it as an f64 value.
2022 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2023 DAG.getConstant(1, MVT::i32));
2025 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2027 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2028 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2029 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2030 Flag = Chain.getValue(1);
2031 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2032 VA = RVLocs[++i]; // skip ahead to next loc
2033 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2036 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2038 // Guarantee that all emitted copies are
2039 // stuck together, avoiding something bad.
2040 Flag = Chain.getValue(1);
2041 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2044 // Update chain and glue.
2047 RetOps.push_back(Flag);
2049 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2050 RetOps.data(), RetOps.size());
2053 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2054 if (N->getNumValues() != 1)
2056 if (!N->hasNUsesOfValue(1, 0))
2059 SDValue TCChain = Chain;
2060 SDNode *Copy = *N->use_begin();
2061 if (Copy->getOpcode() == ISD::CopyToReg) {
2062 // If the copy has a glue operand, we conservatively assume it isn't safe to
2063 // perform a tail call.
2064 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2066 TCChain = Copy->getOperand(0);
2067 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2068 SDNode *VMov = Copy;
2069 // f64 returned in a pair of GPRs.
2070 SmallPtrSet<SDNode*, 2> Copies;
2071 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2073 if (UI->getOpcode() != ISD::CopyToReg)
2077 if (Copies.size() > 2)
2080 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2082 SDValue UseChain = UI->getOperand(0);
2083 if (Copies.count(UseChain.getNode()))
2090 } else if (Copy->getOpcode() == ISD::BITCAST) {
2091 // f32 returned in a single GPR.
2092 if (!Copy->hasOneUse())
2094 Copy = *Copy->use_begin();
2095 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2097 Chain = Copy->getOperand(0);
2102 bool HasRet = false;
2103 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2105 if (UI->getOpcode() != ARMISD::RET_FLAG)
2117 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2118 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2121 if (!CI->isTailCall())
2124 return !Subtarget->isThumb1Only();
2127 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2128 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2129 // one of the above mentioned nodes. It has to be wrapped because otherwise
2130 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2131 // be used to form addressing mode. These wrapped nodes will be selected
2133 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2134 EVT PtrVT = Op.getValueType();
2135 // FIXME there is no actual debug info here
2136 DebugLoc dl = Op.getDebugLoc();
2137 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2139 if (CP->isMachineConstantPoolEntry())
2140 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2141 CP->getAlignment());
2143 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2144 CP->getAlignment());
2145 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2148 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2149 return MachineJumpTableInfo::EK_Inline;
2152 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2153 SelectionDAG &DAG) const {
2154 MachineFunction &MF = DAG.getMachineFunction();
2155 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2156 unsigned ARMPCLabelIndex = 0;
2157 DebugLoc DL = Op.getDebugLoc();
2158 EVT PtrVT = getPointerTy();
2159 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2160 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2162 if (RelocM == Reloc::Static) {
2163 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2165 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2166 ARMPCLabelIndex = AFI->createPICLabelUId();
2167 ARMConstantPoolValue *CPV =
2168 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2169 ARMCP::CPBlockAddress, PCAdj);
2170 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2172 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2173 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2174 MachinePointerInfo::getConstantPool(),
2175 false, false, false, 0);
2176 if (RelocM == Reloc::Static)
2178 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2179 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2182 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2184 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2185 SelectionDAG &DAG) const {
2186 DebugLoc dl = GA->getDebugLoc();
2187 EVT PtrVT = getPointerTy();
2188 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2189 MachineFunction &MF = DAG.getMachineFunction();
2190 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2191 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2192 ARMConstantPoolValue *CPV =
2193 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2194 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2195 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2196 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2197 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2198 MachinePointerInfo::getConstantPool(),
2199 false, false, false, 0);
2200 SDValue Chain = Argument.getValue(1);
2202 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2203 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2205 // call __tls_get_addr.
2208 Entry.Node = Argument;
2209 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2210 Args.push_back(Entry);
2211 // FIXME: is there useful debug info available here?
2212 TargetLowering::CallLoweringInfo CLI(Chain,
2213 (Type *) Type::getInt32Ty(*DAG.getContext()),
2214 false, false, false, false,
2215 0, CallingConv::C, /*isTailCall=*/false,
2216 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2217 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2218 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2219 return CallResult.first;
2222 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2223 // "local exec" model.
2225 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2227 TLSModel::Model model) const {
2228 const GlobalValue *GV = GA->getGlobal();
2229 DebugLoc dl = GA->getDebugLoc();
2231 SDValue Chain = DAG.getEntryNode();
2232 EVT PtrVT = getPointerTy();
2233 // Get the Thread Pointer
2234 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2236 if (model == TLSModel::InitialExec) {
2237 MachineFunction &MF = DAG.getMachineFunction();
2238 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2239 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2240 // Initial exec model.
2241 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2242 ARMConstantPoolValue *CPV =
2243 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2244 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2246 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2247 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2248 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2249 MachinePointerInfo::getConstantPool(),
2250 false, false, false, 0);
2251 Chain = Offset.getValue(1);
2253 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2254 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2256 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2257 MachinePointerInfo::getConstantPool(),
2258 false, false, false, 0);
2261 assert(model == TLSModel::LocalExec);
2262 ARMConstantPoolValue *CPV =
2263 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2264 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2265 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2266 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2267 MachinePointerInfo::getConstantPool(),
2268 false, false, false, 0);
2271 // The address of the thread local variable is the add of the thread
2272 // pointer with the offset of the variable.
2273 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2277 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2278 // TODO: implement the "local dynamic" model
2279 assert(Subtarget->isTargetELF() &&
2280 "TLS not implemented for non-ELF targets");
2281 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2283 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2286 case TLSModel::GeneralDynamic:
2287 case TLSModel::LocalDynamic:
2288 return LowerToTLSGeneralDynamicModel(GA, DAG);
2289 case TLSModel::InitialExec:
2290 case TLSModel::LocalExec:
2291 return LowerToTLSExecModels(GA, DAG, model);
2293 llvm_unreachable("bogus TLS model");
2296 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2297 SelectionDAG &DAG) const {
2298 EVT PtrVT = getPointerTy();
2299 DebugLoc dl = Op.getDebugLoc();
2300 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2301 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2302 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2303 ARMConstantPoolValue *CPV =
2304 ARMConstantPoolConstant::Create(GV,
2305 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2306 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2307 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2308 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2310 MachinePointerInfo::getConstantPool(),
2311 false, false, false, 0);
2312 SDValue Chain = Result.getValue(1);
2313 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2314 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2316 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2317 MachinePointerInfo::getGOT(),
2318 false, false, false, 0);
2322 // If we have T2 ops, we can materialize the address directly via movt/movw
2323 // pair. This is always cheaper.
2324 if (Subtarget->useMovt()) {
2326 // FIXME: Once remat is capable of dealing with instructions with register
2327 // operands, expand this into two nodes.
2328 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2329 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2331 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2332 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2333 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2334 MachinePointerInfo::getConstantPool(),
2335 false, false, false, 0);
2339 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2340 SelectionDAG &DAG) const {
2341 EVT PtrVT = getPointerTy();
2342 DebugLoc dl = Op.getDebugLoc();
2343 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2344 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2346 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2347 // update ARMFastISel::ARMMaterializeGV.
2348 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2350 // FIXME: Once remat is capable of dealing with instructions with register
2351 // operands, expand this into two nodes.
2352 if (RelocM == Reloc::Static)
2353 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2354 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2356 unsigned Wrapper = (RelocM == Reloc::PIC_)
2357 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2358 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2359 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2360 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2361 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2362 MachinePointerInfo::getGOT(),
2363 false, false, false, 0);
2367 unsigned ARMPCLabelIndex = 0;
2369 if (RelocM == Reloc::Static) {
2370 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2372 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2373 ARMPCLabelIndex = AFI->createPICLabelUId();
2374 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2375 ARMConstantPoolValue *CPV =
2376 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2378 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2380 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2382 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2383 MachinePointerInfo::getConstantPool(),
2384 false, false, false, 0);
2385 SDValue Chain = Result.getValue(1);
2387 if (RelocM == Reloc::PIC_) {
2388 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2389 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2392 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2393 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2394 false, false, false, 0);
2399 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2400 SelectionDAG &DAG) const {
2401 assert(Subtarget->isTargetELF() &&
2402 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2403 MachineFunction &MF = DAG.getMachineFunction();
2404 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2405 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2406 EVT PtrVT = getPointerTy();
2407 DebugLoc dl = Op.getDebugLoc();
2408 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2409 ARMConstantPoolValue *CPV =
2410 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2411 ARMPCLabelIndex, PCAdj);
2412 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2413 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2414 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2415 MachinePointerInfo::getConstantPool(),
2416 false, false, false, 0);
2417 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2418 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2422 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2423 DebugLoc dl = Op.getDebugLoc();
2424 SDValue Val = DAG.getConstant(0, MVT::i32);
2425 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2426 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2427 Op.getOperand(1), Val);
2431 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2432 DebugLoc dl = Op.getDebugLoc();
2433 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2434 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2438 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2439 const ARMSubtarget *Subtarget) const {
2440 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2441 DebugLoc dl = Op.getDebugLoc();
2443 default: return SDValue(); // Don't custom lower most intrinsics.
2444 case Intrinsic::arm_thread_pointer: {
2445 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2446 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2448 case Intrinsic::eh_sjlj_lsda: {
2449 MachineFunction &MF = DAG.getMachineFunction();
2450 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2451 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2452 EVT PtrVT = getPointerTy();
2453 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2455 unsigned PCAdj = (RelocM != Reloc::PIC_)
2456 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2457 ARMConstantPoolValue *CPV =
2458 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2459 ARMCP::CPLSDA, PCAdj);
2460 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2461 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2463 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2464 MachinePointerInfo::getConstantPool(),
2465 false, false, false, 0);
2467 if (RelocM == Reloc::PIC_) {
2468 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2469 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2473 case Intrinsic::arm_neon_vmulls:
2474 case Intrinsic::arm_neon_vmullu: {
2475 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2476 ? ARMISD::VMULLs : ARMISD::VMULLu;
2477 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2478 Op.getOperand(1), Op.getOperand(2));
2483 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2484 const ARMSubtarget *Subtarget) {
2485 // FIXME: handle "fence singlethread" more efficiently.
2486 DebugLoc dl = Op.getDebugLoc();
2487 if (!Subtarget->hasDataBarrier()) {
2488 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2489 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2491 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2492 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2493 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2494 DAG.getConstant(0, MVT::i32));
2497 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2498 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2501 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2502 const ARMSubtarget *Subtarget) {
2503 // ARM pre v5TE and Thumb1 does not have preload instructions.
2504 if (!(Subtarget->isThumb2() ||
2505 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2506 // Just preserve the chain.
2507 return Op.getOperand(0);
2509 DebugLoc dl = Op.getDebugLoc();
2510 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2512 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2513 // ARMv7 with MP extension has PLDW.
2514 return Op.getOperand(0);
2516 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2517 if (Subtarget->isThumb()) {
2519 isRead = ~isRead & 1;
2520 isData = ~isData & 1;
2523 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2524 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2525 DAG.getConstant(isData, MVT::i32));
2528 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2529 MachineFunction &MF = DAG.getMachineFunction();
2530 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2532 // vastart just stores the address of the VarArgsFrameIndex slot into the
2533 // memory location argument.
2534 DebugLoc dl = Op.getDebugLoc();
2535 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2536 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2537 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2538 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2539 MachinePointerInfo(SV), false, false, 0);
2543 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2544 SDValue &Root, SelectionDAG &DAG,
2545 DebugLoc dl) const {
2546 MachineFunction &MF = DAG.getMachineFunction();
2547 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2549 const TargetRegisterClass *RC;
2550 if (AFI->isThumb1OnlyFunction())
2551 RC = &ARM::tGPRRegClass;
2553 RC = &ARM::GPRRegClass;
2555 // Transform the arguments stored in physical registers into virtual ones.
2556 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2557 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2560 if (NextVA.isMemLoc()) {
2561 MachineFrameInfo *MFI = MF.getFrameInfo();
2562 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2564 // Create load node to retrieve arguments from the stack.
2565 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2566 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2567 MachinePointerInfo::getFixedStack(FI),
2568 false, false, false, 0);
2570 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2571 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2574 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2578 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2579 unsigned &VARegSize, unsigned &VARegSaveSize)
2582 if (CCInfo.isFirstByValRegValid())
2583 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2585 unsigned int firstUnalloced;
2586 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2587 sizeof(GPRArgRegs) /
2588 sizeof(GPRArgRegs[0]));
2589 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2592 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2593 VARegSize = NumGPRs * 4;
2594 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2597 // The remaining GPRs hold either the beginning of variable-argument
2598 // data, or the beginning of an aggregate passed by value (usually
2599 // byval). Either way, we allocate stack slots adjacent to the data
2600 // provided by our caller, and store the unallocated registers there.
2601 // If this is a variadic function, the va_list pointer will begin with
2602 // these values; otherwise, this reassembles a (byval) structure that
2603 // was split between registers and memory.
2605 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2606 DebugLoc dl, SDValue &Chain,
2607 const Value *OrigArg,
2608 unsigned OffsetFromOrigArg,
2610 bool ForceMutable) const {
2611 MachineFunction &MF = DAG.getMachineFunction();
2612 MachineFrameInfo *MFI = MF.getFrameInfo();
2613 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2614 unsigned firstRegToSaveIndex;
2615 if (CCInfo.isFirstByValRegValid())
2616 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2618 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2619 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2622 unsigned VARegSize, VARegSaveSize;
2623 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2624 if (VARegSaveSize) {
2625 // If this function is vararg, store any remaining integer argument regs
2626 // to their spots on the stack so that they may be loaded by deferencing
2627 // the result of va_next.
2628 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2629 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2630 ArgOffset + VARegSaveSize
2633 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2636 SmallVector<SDValue, 4> MemOps;
2637 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2638 const TargetRegisterClass *RC;
2639 if (AFI->isThumb1OnlyFunction())
2640 RC = &ARM::tGPRRegClass;
2642 RC = &ARM::GPRRegClass;
2644 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2645 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2647 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2648 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2650 MemOps.push_back(Store);
2651 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2652 DAG.getConstant(4, getPointerTy()));
2654 if (!MemOps.empty())
2655 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2656 &MemOps[0], MemOps.size());
2658 // This will point to the next argument passed via stack.
2659 AFI->setVarArgsFrameIndex(
2660 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2664 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2665 CallingConv::ID CallConv, bool isVarArg,
2666 const SmallVectorImpl<ISD::InputArg>
2668 DebugLoc dl, SelectionDAG &DAG,
2669 SmallVectorImpl<SDValue> &InVals)
2671 MachineFunction &MF = DAG.getMachineFunction();
2672 MachineFrameInfo *MFI = MF.getFrameInfo();
2674 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2676 // Assign locations to all of the incoming arguments.
2677 SmallVector<CCValAssign, 16> ArgLocs;
2678 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2679 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2680 CCInfo.AnalyzeFormalArguments(Ins,
2681 CCAssignFnForNode(CallConv, /* Return*/ false,
2684 SmallVector<SDValue, 16> ArgValues;
2685 int lastInsIndex = -1;
2687 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2688 unsigned CurArgIdx = 0;
2689 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2690 CCValAssign &VA = ArgLocs[i];
2691 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2692 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2693 // Arguments stored in registers.
2694 if (VA.isRegLoc()) {
2695 EVT RegVT = VA.getLocVT();
2697 if (VA.needsCustom()) {
2698 // f64 and vector types are split up into multiple registers or
2699 // combinations of registers and stack slots.
2700 if (VA.getLocVT() == MVT::v2f64) {
2701 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2703 VA = ArgLocs[++i]; // skip ahead to next loc
2705 if (VA.isMemLoc()) {
2706 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2707 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2708 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2709 MachinePointerInfo::getFixedStack(FI),
2710 false, false, false, 0);
2712 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2715 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2716 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2717 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2718 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2719 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2721 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2724 const TargetRegisterClass *RC;
2726 if (RegVT == MVT::f32)
2727 RC = &ARM::SPRRegClass;
2728 else if (RegVT == MVT::f64)
2729 RC = &ARM::DPRRegClass;
2730 else if (RegVT == MVT::v2f64)
2731 RC = &ARM::QPRRegClass;
2732 else if (RegVT == MVT::i32)
2733 RC = AFI->isThumb1OnlyFunction() ?
2734 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2735 (const TargetRegisterClass*)&ARM::GPRRegClass;
2737 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2739 // Transform the arguments in physical registers into virtual ones.
2740 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2741 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2744 // If this is an 8 or 16-bit value, it is really passed promoted
2745 // to 32 bits. Insert an assert[sz]ext to capture this, then
2746 // truncate to the right size.
2747 switch (VA.getLocInfo()) {
2748 default: llvm_unreachable("Unknown loc info!");
2749 case CCValAssign::Full: break;
2750 case CCValAssign::BCvt:
2751 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2753 case CCValAssign::SExt:
2754 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2755 DAG.getValueType(VA.getValVT()));
2756 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2758 case CCValAssign::ZExt:
2759 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2760 DAG.getValueType(VA.getValVT()));
2761 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2765 InVals.push_back(ArgValue);
2767 } else { // VA.isRegLoc()
2770 assert(VA.isMemLoc());
2771 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2773 int index = ArgLocs[i].getValNo();
2775 // Some Ins[] entries become multiple ArgLoc[] entries.
2776 // Process them only once.
2777 if (index != lastInsIndex)
2779 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2780 // FIXME: For now, all byval parameter objects are marked mutable.
2781 // This can be changed with more analysis.
2782 // In case of tail call optimization mark all arguments mutable.
2783 // Since they could be overwritten by lowering of arguments in case of
2785 if (Flags.isByVal()) {
2786 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2787 if (!AFI->getVarArgsFrameIndex()) {
2788 VarArgStyleRegisters(CCInfo, DAG,
2789 dl, Chain, CurOrigArg,
2790 Ins[VA.getValNo()].PartOffset,
2791 VA.getLocMemOffset(),
2792 true /*force mutable frames*/);
2793 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2794 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2796 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2797 VA.getLocMemOffset(), false);
2798 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2801 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2802 VA.getLocMemOffset(), true);
2804 // Create load nodes to retrieve arguments from the stack.
2805 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2806 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2807 MachinePointerInfo::getFixedStack(FI),
2808 false, false, false, 0));
2810 lastInsIndex = index;
2817 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2818 CCInfo.getNextStackOffset());
2823 /// isFloatingPointZero - Return true if this is +0.0.
2824 static bool isFloatingPointZero(SDValue Op) {
2825 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2826 return CFP->getValueAPF().isPosZero();
2827 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2828 // Maybe this has already been legalized into the constant pool?
2829 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2830 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2831 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2832 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2833 return CFP->getValueAPF().isPosZero();
2839 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2840 /// the given operands.
2842 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2843 SDValue &ARMcc, SelectionDAG &DAG,
2844 DebugLoc dl) const {
2845 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2846 unsigned C = RHSC->getZExtValue();
2847 if (!isLegalICmpImmediate(C)) {
2848 // Constant does not fit, try adjusting it by one?
2853 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2854 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2855 RHS = DAG.getConstant(C-1, MVT::i32);
2860 if (C != 0 && isLegalICmpImmediate(C-1)) {
2861 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2862 RHS = DAG.getConstant(C-1, MVT::i32);
2867 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2868 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2869 RHS = DAG.getConstant(C+1, MVT::i32);
2874 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2875 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2876 RHS = DAG.getConstant(C+1, MVT::i32);
2883 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2884 ARMISD::NodeType CompareType;
2887 CompareType = ARMISD::CMP;
2892 CompareType = ARMISD::CMPZ;
2895 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2896 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2899 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2901 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2902 DebugLoc dl) const {
2904 if (!isFloatingPointZero(RHS))
2905 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2907 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2908 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2911 /// duplicateCmp - Glue values can have only one use, so this function
2912 /// duplicates a comparison node.
2914 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2915 unsigned Opc = Cmp.getOpcode();
2916 DebugLoc DL = Cmp.getDebugLoc();
2917 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2918 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2920 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2921 Cmp = Cmp.getOperand(0);
2922 Opc = Cmp.getOpcode();
2923 if (Opc == ARMISD::CMPFP)
2924 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2926 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2927 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2929 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2932 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2933 SDValue Cond = Op.getOperand(0);
2934 SDValue SelectTrue = Op.getOperand(1);
2935 SDValue SelectFalse = Op.getOperand(2);
2936 DebugLoc dl = Op.getDebugLoc();
2940 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2941 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2943 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2944 const ConstantSDNode *CMOVTrue =
2945 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2946 const ConstantSDNode *CMOVFalse =
2947 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2949 if (CMOVTrue && CMOVFalse) {
2950 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2951 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2955 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2957 False = SelectFalse;
2958 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2963 if (True.getNode() && False.getNode()) {
2964 EVT VT = Op.getValueType();
2965 SDValue ARMcc = Cond.getOperand(2);
2966 SDValue CCR = Cond.getOperand(3);
2967 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2968 assert(True.getValueType() == VT);
2969 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2974 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2975 // undefined bits before doing a full-word comparison with zero.
2976 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2977 DAG.getConstant(1, Cond.getValueType()));
2979 return DAG.getSelectCC(dl, Cond,
2980 DAG.getConstant(0, Cond.getValueType()),
2981 SelectTrue, SelectFalse, ISD::SETNE);
2984 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2985 EVT VT = Op.getValueType();
2986 SDValue LHS = Op.getOperand(0);
2987 SDValue RHS = Op.getOperand(1);
2988 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2989 SDValue TrueVal = Op.getOperand(2);
2990 SDValue FalseVal = Op.getOperand(3);
2991 DebugLoc dl = Op.getDebugLoc();
2993 if (LHS.getValueType() == MVT::i32) {
2995 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2996 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2997 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
3000 ARMCC::CondCodes CondCode, CondCode2;
3001 FPCCToARMCC(CC, CondCode, CondCode2);
3003 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3004 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3005 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3006 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3008 if (CondCode2 != ARMCC::AL) {
3009 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3010 // FIXME: Needs another CMP because flag can have but one use.
3011 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3012 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3013 Result, TrueVal, ARMcc2, CCR, Cmp2);
3018 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3019 /// to morph to an integer compare sequence.
3020 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3021 const ARMSubtarget *Subtarget) {
3022 SDNode *N = Op.getNode();
3023 if (!N->hasOneUse())
3024 // Otherwise it requires moving the value from fp to integer registers.
3026 if (!N->getNumValues())
3028 EVT VT = Op.getValueType();
3029 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3030 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3031 // vmrs are very slow, e.g. cortex-a8.
3034 if (isFloatingPointZero(Op)) {
3038 return ISD::isNormalLoad(N);
3041 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3042 if (isFloatingPointZero(Op))
3043 return DAG.getConstant(0, MVT::i32);
3045 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3046 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3047 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3048 Ld->isVolatile(), Ld->isNonTemporal(),
3049 Ld->isInvariant(), Ld->getAlignment());
3051 llvm_unreachable("Unknown VFP cmp argument!");
3054 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3055 SDValue &RetVal1, SDValue &RetVal2) {
3056 if (isFloatingPointZero(Op)) {
3057 RetVal1 = DAG.getConstant(0, MVT::i32);
3058 RetVal2 = DAG.getConstant(0, MVT::i32);
3062 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3063 SDValue Ptr = Ld->getBasePtr();
3064 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3065 Ld->getChain(), Ptr,
3066 Ld->getPointerInfo(),
3067 Ld->isVolatile(), Ld->isNonTemporal(),
3068 Ld->isInvariant(), Ld->getAlignment());
3070 EVT PtrType = Ptr.getValueType();
3071 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3072 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3073 PtrType, Ptr, DAG.getConstant(4, PtrType));
3074 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3075 Ld->getChain(), NewPtr,
3076 Ld->getPointerInfo().getWithOffset(4),
3077 Ld->isVolatile(), Ld->isNonTemporal(),
3078 Ld->isInvariant(), NewAlign);
3082 llvm_unreachable("Unknown VFP cmp argument!");
3085 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3086 /// f32 and even f64 comparisons to integer ones.
3088 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3089 SDValue Chain = Op.getOperand(0);
3090 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3091 SDValue LHS = Op.getOperand(2);
3092 SDValue RHS = Op.getOperand(3);
3093 SDValue Dest = Op.getOperand(4);
3094 DebugLoc dl = Op.getDebugLoc();
3096 bool LHSSeenZero = false;
3097 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3098 bool RHSSeenZero = false;
3099 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3100 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3101 // If unsafe fp math optimization is enabled and there are no other uses of
3102 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3103 // to an integer comparison.
3104 if (CC == ISD::SETOEQ)
3106 else if (CC == ISD::SETUNE)
3109 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3111 if (LHS.getValueType() == MVT::f32) {
3112 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3113 bitcastf32Toi32(LHS, DAG), Mask);
3114 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3115 bitcastf32Toi32(RHS, DAG), Mask);
3116 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3117 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3118 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3119 Chain, Dest, ARMcc, CCR, Cmp);
3124 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3125 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3126 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3127 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3128 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3129 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3130 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3131 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3132 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3138 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3139 SDValue Chain = Op.getOperand(0);
3140 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3141 SDValue LHS = Op.getOperand(2);
3142 SDValue RHS = Op.getOperand(3);
3143 SDValue Dest = Op.getOperand(4);
3144 DebugLoc dl = Op.getDebugLoc();
3146 if (LHS.getValueType() == MVT::i32) {
3148 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3149 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3150 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3151 Chain, Dest, ARMcc, CCR, Cmp);
3154 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3156 if (getTargetMachine().Options.UnsafeFPMath &&
3157 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3158 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3159 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3160 if (Result.getNode())
3164 ARMCC::CondCodes CondCode, CondCode2;
3165 FPCCToARMCC(CC, CondCode, CondCode2);
3167 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3168 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3169 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3170 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3171 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3172 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3173 if (CondCode2 != ARMCC::AL) {
3174 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3175 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3176 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3181 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3182 SDValue Chain = Op.getOperand(0);
3183 SDValue Table = Op.getOperand(1);
3184 SDValue Index = Op.getOperand(2);
3185 DebugLoc dl = Op.getDebugLoc();
3187 EVT PTy = getPointerTy();
3188 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3189 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3190 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3191 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3192 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3193 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3194 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3195 if (Subtarget->isThumb2()) {
3196 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3197 // which does another jump to the destination. This also makes it easier
3198 // to translate it to TBB / TBH later.
3199 // FIXME: This might not work if the function is extremely large.
3200 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3201 Addr, Op.getOperand(2), JTI, UId);
3203 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3204 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3205 MachinePointerInfo::getJumpTable(),
3206 false, false, false, 0);
3207 Chain = Addr.getValue(1);
3208 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3209 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3211 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3212 MachinePointerInfo::getJumpTable(),
3213 false, false, false, 0);
3214 Chain = Addr.getValue(1);
3215 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3219 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3220 EVT VT = Op.getValueType();
3221 DebugLoc dl = Op.getDebugLoc();
3223 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3224 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3226 return DAG.UnrollVectorOp(Op.getNode());
3229 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3230 "Invalid type for custom lowering!");
3231 if (VT != MVT::v4i16)
3232 return DAG.UnrollVectorOp(Op.getNode());
3234 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3235 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3238 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3239 EVT VT = Op.getValueType();
3241 return LowerVectorFP_TO_INT(Op, DAG);
3243 DebugLoc dl = Op.getDebugLoc();
3246 switch (Op.getOpcode()) {
3247 default: llvm_unreachable("Invalid opcode!");
3248 case ISD::FP_TO_SINT:
3249 Opc = ARMISD::FTOSI;
3251 case ISD::FP_TO_UINT:
3252 Opc = ARMISD::FTOUI;
3255 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3256 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3259 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3260 EVT VT = Op.getValueType();
3261 DebugLoc dl = Op.getDebugLoc();
3263 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3264 if (VT.getVectorElementType() == MVT::f32)
3266 return DAG.UnrollVectorOp(Op.getNode());
3269 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3270 "Invalid type for custom lowering!");
3271 if (VT != MVT::v4f32)
3272 return DAG.UnrollVectorOp(Op.getNode());
3276 switch (Op.getOpcode()) {
3277 default: llvm_unreachable("Invalid opcode!");
3278 case ISD::SINT_TO_FP:
3279 CastOpc = ISD::SIGN_EXTEND;
3280 Opc = ISD::SINT_TO_FP;
3282 case ISD::UINT_TO_FP:
3283 CastOpc = ISD::ZERO_EXTEND;
3284 Opc = ISD::UINT_TO_FP;
3288 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3289 return DAG.getNode(Opc, dl, VT, Op);
3292 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3293 EVT VT = Op.getValueType();
3295 return LowerVectorINT_TO_FP(Op, DAG);
3297 DebugLoc dl = Op.getDebugLoc();
3300 switch (Op.getOpcode()) {
3301 default: llvm_unreachable("Invalid opcode!");
3302 case ISD::SINT_TO_FP:
3303 Opc = ARMISD::SITOF;
3305 case ISD::UINT_TO_FP:
3306 Opc = ARMISD::UITOF;
3310 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3311 return DAG.getNode(Opc, dl, VT, Op);
3314 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3315 // Implement fcopysign with a fabs and a conditional fneg.
3316 SDValue Tmp0 = Op.getOperand(0);
3317 SDValue Tmp1 = Op.getOperand(1);
3318 DebugLoc dl = Op.getDebugLoc();
3319 EVT VT = Op.getValueType();
3320 EVT SrcVT = Tmp1.getValueType();
3321 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3322 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3323 bool UseNEON = !InGPR && Subtarget->hasNEON();
3326 // Use VBSL to copy the sign bit.
3327 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3328 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3329 DAG.getTargetConstant(EncodedVal, MVT::i32));
3330 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3332 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3333 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3334 DAG.getConstant(32, MVT::i32));
3335 else /*if (VT == MVT::f32)*/
3336 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3337 if (SrcVT == MVT::f32) {
3338 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3340 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3341 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3342 DAG.getConstant(32, MVT::i32));
3343 } else if (VT == MVT::f32)
3344 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3345 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3346 DAG.getConstant(32, MVT::i32));
3347 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3348 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3350 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3352 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3353 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3354 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3356 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3357 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3358 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3359 if (VT == MVT::f32) {
3360 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3361 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3362 DAG.getConstant(0, MVT::i32));
3364 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3370 // Bitcast operand 1 to i32.
3371 if (SrcVT == MVT::f64)
3372 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3373 &Tmp1, 1).getValue(1);
3374 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3376 // Or in the signbit with integer operations.
3377 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3378 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3379 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3380 if (VT == MVT::f32) {
3381 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3382 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3383 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3384 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3387 // f64: Or the high part with signbit and then combine two parts.
3388 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3390 SDValue Lo = Tmp0.getValue(0);
3391 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3392 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3393 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3396 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3397 MachineFunction &MF = DAG.getMachineFunction();
3398 MachineFrameInfo *MFI = MF.getFrameInfo();
3399 MFI->setReturnAddressIsTaken(true);
3401 EVT VT = Op.getValueType();
3402 DebugLoc dl = Op.getDebugLoc();
3403 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3405 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3406 SDValue Offset = DAG.getConstant(4, MVT::i32);
3407 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3408 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3409 MachinePointerInfo(), false, false, false, 0);
3412 // Return LR, which contains the return address. Mark it an implicit live-in.
3413 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3414 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3417 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3418 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3419 MFI->setFrameAddressIsTaken(true);
3421 EVT VT = Op.getValueType();
3422 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3423 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3424 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3425 ? ARM::R7 : ARM::R11;
3426 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3428 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3429 MachinePointerInfo(),
3430 false, false, false, 0);
3434 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
3435 /// and size(DestVec) > 128-bits.
3436 /// This is achieved by doing the one extension from the SrcVec, splitting the
3437 /// result, extending these parts, and then concatenating these into the
3439 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
3440 SDValue Op = N->getOperand(0);
3441 EVT SrcVT = Op.getValueType();
3442 EVT DestVT = N->getValueType(0);
3444 assert(DestVT.getSizeInBits() > 128 &&
3445 "Custom sext/zext expansion needs >128-bit vector.");
3446 // If this is a normal length extension, use the default expansion.
3447 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
3448 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
3451 DebugLoc dl = N->getDebugLoc();
3452 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
3453 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
3454 unsigned NumElts = SrcVT.getVectorNumElements();
3455 LLVMContext &Ctx = *DAG.getContext();
3456 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;
3458 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3460 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3462 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
3465 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
3466 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3467 DAG.getIntPtrConstant(0));
3468 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3469 DAG.getIntPtrConstant(NumElts/2));
3470 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
3471 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
3472 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
3475 /// ExpandBITCAST - If the target supports VFP, this function is called to
3476 /// expand a bit convert where either the source or destination type is i64 to
3477 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3478 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3479 /// vectors), since the legalizer won't know what to do with that.
3480 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3481 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3482 DebugLoc dl = N->getDebugLoc();
3483 SDValue Op = N->getOperand(0);
3485 // This function is only supposed to be called for i64 types, either as the
3486 // source or destination of the bit convert.
3487 EVT SrcVT = Op.getValueType();
3488 EVT DstVT = N->getValueType(0);
3489 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3490 "ExpandBITCAST called for non-i64 type");
3492 // Turn i64->f64 into VMOVDRR.
3493 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3494 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3495 DAG.getConstant(0, MVT::i32));
3496 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3497 DAG.getConstant(1, MVT::i32));
3498 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3499 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3502 // Turn f64->i64 into VMOVRRD.
3503 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3504 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3505 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3506 // Merge the pieces into a single i64 value.
3507 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3513 /// getZeroVector - Returns a vector of specified type with all zero elements.
3514 /// Zero vectors are used to represent vector negation and in those cases
3515 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3516 /// not support i64 elements, so sometimes the zero vectors will need to be
3517 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3519 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3520 assert(VT.isVector() && "Expected a vector type");
3521 // The canonical modified immediate encoding of a zero vector is....0!
3522 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3523 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3524 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3525 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3528 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3529 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3530 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3531 SelectionDAG &DAG) const {
3532 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3533 EVT VT = Op.getValueType();
3534 unsigned VTBits = VT.getSizeInBits();
3535 DebugLoc dl = Op.getDebugLoc();
3536 SDValue ShOpLo = Op.getOperand(0);
3537 SDValue ShOpHi = Op.getOperand(1);
3538 SDValue ShAmt = Op.getOperand(2);
3540 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3542 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3544 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3545 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3546 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3547 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3548 DAG.getConstant(VTBits, MVT::i32));
3549 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3550 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3551 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3553 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3554 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3556 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3557 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3560 SDValue Ops[2] = { Lo, Hi };
3561 return DAG.getMergeValues(Ops, 2, dl);
3564 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3565 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3566 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3567 SelectionDAG &DAG) const {
3568 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3569 EVT VT = Op.getValueType();
3570 unsigned VTBits = VT.getSizeInBits();
3571 DebugLoc dl = Op.getDebugLoc();
3572 SDValue ShOpLo = Op.getOperand(0);
3573 SDValue ShOpHi = Op.getOperand(1);
3574 SDValue ShAmt = Op.getOperand(2);
3577 assert(Op.getOpcode() == ISD::SHL_PARTS);
3578 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3579 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3580 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3581 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3582 DAG.getConstant(VTBits, MVT::i32));
3583 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3584 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3586 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3587 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3588 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3590 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3591 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3594 SDValue Ops[2] = { Lo, Hi };
3595 return DAG.getMergeValues(Ops, 2, dl);
3598 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3599 SelectionDAG &DAG) const {
3600 // The rounding mode is in bits 23:22 of the FPSCR.
3601 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3602 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3603 // so that the shift + and get folded into a bitfield extract.
3604 DebugLoc dl = Op.getDebugLoc();
3605 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3606 DAG.getConstant(Intrinsic::arm_get_fpscr,
3608 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3609 DAG.getConstant(1U << 22, MVT::i32));
3610 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3611 DAG.getConstant(22, MVT::i32));
3612 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3613 DAG.getConstant(3, MVT::i32));
3616 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3617 const ARMSubtarget *ST) {
3618 EVT VT = N->getValueType(0);
3619 DebugLoc dl = N->getDebugLoc();
3621 if (!ST->hasV6T2Ops())
3624 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3625 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3628 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3629 /// for each 16-bit element from operand, repeated. The basic idea is to
3630 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3632 /// Trace for v4i16:
3633 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3634 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3635 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3636 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3637 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3638 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3639 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3640 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3641 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3642 EVT VT = N->getValueType(0);
3643 DebugLoc DL = N->getDebugLoc();
3645 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3646 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3647 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3648 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3649 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3650 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3653 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3654 /// bit-count for each 16-bit element from the operand. We need slightly
3655 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3656 /// 64/128-bit registers.
3658 /// Trace for v4i16:
3659 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3660 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3661 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3662 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3663 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3664 EVT VT = N->getValueType(0);
3665 DebugLoc DL = N->getDebugLoc();
3667 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3668 if (VT.is64BitVector()) {
3669 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3670 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3671 DAG.getIntPtrConstant(0));
3673 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3674 BitCounts, DAG.getIntPtrConstant(0));
3675 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3679 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3680 /// bit-count for each 32-bit element from the operand. The idea here is
3681 /// to split the vector into 16-bit elements, leverage the 16-bit count
3682 /// routine, and then combine the results.
3684 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3685 /// input = [v0 v1 ] (vi: 32-bit elements)
3686 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3687 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3688 /// vrev: N0 = [k1 k0 k3 k2 ]
3690 /// N1 =+[k1 k0 k3 k2 ]
3692 /// N2 =+[k1 k3 k0 k2 ]
3694 /// Extended =+[k1 k3 k0 k2 ]
3696 /// Extracted=+[k1 k3 ]
3698 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
3699 EVT VT = N->getValueType(0);
3700 DebugLoc DL = N->getDebugLoc();
3702 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
3704 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
3705 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
3706 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
3707 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
3708 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
3710 if (VT.is64BitVector()) {
3711 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
3712 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
3713 DAG.getIntPtrConstant(0));
3715 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
3716 DAG.getIntPtrConstant(0));
3717 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
3721 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
3722 const ARMSubtarget *ST) {
3723 EVT VT = N->getValueType(0);
3725 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
3726 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
3727 VT == MVT::v4i16 || VT == MVT::v8i16) &&
3728 "Unexpected type for custom ctpop lowering");
3730 if (VT.getVectorElementType() == MVT::i32)
3731 return lowerCTPOP32BitElements(N, DAG);
3733 return lowerCTPOP16BitElements(N, DAG);
3736 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3737 const ARMSubtarget *ST) {
3738 EVT VT = N->getValueType(0);
3739 DebugLoc dl = N->getDebugLoc();
3744 // Lower vector shifts on NEON to use VSHL.
3745 assert(ST->hasNEON() && "unexpected vector shift");
3747 // Left shifts translate directly to the vshiftu intrinsic.
3748 if (N->getOpcode() == ISD::SHL)
3749 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3750 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3751 N->getOperand(0), N->getOperand(1));
3753 assert((N->getOpcode() == ISD::SRA ||
3754 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3756 // NEON uses the same intrinsics for both left and right shifts. For
3757 // right shifts, the shift amounts are negative, so negate the vector of
3759 EVT ShiftVT = N->getOperand(1).getValueType();
3760 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3761 getZeroVector(ShiftVT, DAG, dl),
3763 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3764 Intrinsic::arm_neon_vshifts :
3765 Intrinsic::arm_neon_vshiftu);
3766 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3767 DAG.getConstant(vshiftInt, MVT::i32),
3768 N->getOperand(0), NegatedCount);
3771 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3772 const ARMSubtarget *ST) {
3773 EVT VT = N->getValueType(0);
3774 DebugLoc dl = N->getDebugLoc();
3776 // We can get here for a node like i32 = ISD::SHL i32, i64
3780 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3781 "Unknown shift to lower!");
3783 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3784 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3785 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3788 // If we are in thumb mode, we don't have RRX.
3789 if (ST->isThumb1Only()) return SDValue();
3791 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3792 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3793 DAG.getConstant(0, MVT::i32));
3794 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3795 DAG.getConstant(1, MVT::i32));
3797 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3798 // captures the result into a carry flag.
3799 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3800 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3802 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3803 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3805 // Merge the pieces into a single i64 value.
3806 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3809 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3810 SDValue TmpOp0, TmpOp1;
3811 bool Invert = false;
3815 SDValue Op0 = Op.getOperand(0);
3816 SDValue Op1 = Op.getOperand(1);
3817 SDValue CC = Op.getOperand(2);
3818 EVT VT = Op.getValueType();
3819 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3820 DebugLoc dl = Op.getDebugLoc();
3822 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3823 switch (SetCCOpcode) {
3824 default: llvm_unreachable("Illegal FP comparison");
3826 case ISD::SETNE: Invert = true; // Fallthrough
3828 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3830 case ISD::SETLT: Swap = true; // Fallthrough
3832 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3834 case ISD::SETLE: Swap = true; // Fallthrough
3836 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3837 case ISD::SETUGE: Swap = true; // Fallthrough
3838 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3839 case ISD::SETUGT: Swap = true; // Fallthrough
3840 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3841 case ISD::SETUEQ: Invert = true; // Fallthrough
3843 // Expand this to (OLT | OGT).
3847 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3848 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3850 case ISD::SETUO: Invert = true; // Fallthrough
3852 // Expand this to (OLT | OGE).
3856 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3857 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3861 // Integer comparisons.
3862 switch (SetCCOpcode) {
3863 default: llvm_unreachable("Illegal integer comparison");
3864 case ISD::SETNE: Invert = true;
3865 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3866 case ISD::SETLT: Swap = true;
3867 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3868 case ISD::SETLE: Swap = true;
3869 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3870 case ISD::SETULT: Swap = true;
3871 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3872 case ISD::SETULE: Swap = true;
3873 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3876 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3877 if (Opc == ARMISD::VCEQ) {
3880 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3882 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3885 // Ignore bitconvert.
3886 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3887 AndOp = AndOp.getOperand(0);
3889 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3891 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3892 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3899 std::swap(Op0, Op1);
3901 // If one of the operands is a constant vector zero, attempt to fold the
3902 // comparison to a specialized compare-against-zero form.
3904 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3906 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3907 if (Opc == ARMISD::VCGE)
3908 Opc = ARMISD::VCLEZ;
3909 else if (Opc == ARMISD::VCGT)
3910 Opc = ARMISD::VCLTZ;
3915 if (SingleOp.getNode()) {
3918 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3920 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3922 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3924 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3926 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3928 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3931 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3935 Result = DAG.getNOT(dl, Result, VT);
3940 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3941 /// valid vector constant for a NEON instruction with a "modified immediate"
3942 /// operand (e.g., VMOV). If so, return the encoded value.
3943 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3944 unsigned SplatBitSize, SelectionDAG &DAG,
3945 EVT &VT, bool is128Bits, NEONModImmType type) {
3946 unsigned OpCmode, Imm;
3948 // SplatBitSize is set to the smallest size that splats the vector, so a
3949 // zero vector will always have SplatBitSize == 8. However, NEON modified
3950 // immediate instructions others than VMOV do not support the 8-bit encoding
3951 // of a zero vector, and the default encoding of zero is supposed to be the
3956 switch (SplatBitSize) {
3958 if (type != VMOVModImm)
3960 // Any 1-byte value is OK. Op=0, Cmode=1110.
3961 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3964 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3968 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3969 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3970 if ((SplatBits & ~0xff) == 0) {
3971 // Value = 0x00nn: Op=x, Cmode=100x.
3976 if ((SplatBits & ~0xff00) == 0) {
3977 // Value = 0xnn00: Op=x, Cmode=101x.
3979 Imm = SplatBits >> 8;
3985 // NEON's 32-bit VMOV supports splat values where:
3986 // * only one byte is nonzero, or
3987 // * the least significant byte is 0xff and the second byte is nonzero, or
3988 // * the least significant 2 bytes are 0xff and the third is nonzero.
3989 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3990 if ((SplatBits & ~0xff) == 0) {
3991 // Value = 0x000000nn: Op=x, Cmode=000x.
3996 if ((SplatBits & ~0xff00) == 0) {
3997 // Value = 0x0000nn00: Op=x, Cmode=001x.
3999 Imm = SplatBits >> 8;
4002 if ((SplatBits & ~0xff0000) == 0) {
4003 // Value = 0x00nn0000: Op=x, Cmode=010x.
4005 Imm = SplatBits >> 16;
4008 if ((SplatBits & ~0xff000000) == 0) {
4009 // Value = 0xnn000000: Op=x, Cmode=011x.
4011 Imm = SplatBits >> 24;
4015 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4016 if (type == OtherModImm) return SDValue();
4018 if ((SplatBits & ~0xffff) == 0 &&
4019 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4020 // Value = 0x0000nnff: Op=x, Cmode=1100.
4022 Imm = SplatBits >> 8;
4027 if ((SplatBits & ~0xffffff) == 0 &&
4028 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4029 // Value = 0x00nnffff: Op=x, Cmode=1101.
4031 Imm = SplatBits >> 16;
4032 SplatBits |= 0xffff;
4036 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4037 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4038 // VMOV.I32. A (very) minor optimization would be to replicate the value
4039 // and fall through here to test for a valid 64-bit splat. But, then the
4040 // caller would also need to check and handle the change in size.
4044 if (type != VMOVModImm)
4046 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4047 uint64_t BitMask = 0xff;
4049 unsigned ImmMask = 1;
4051 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4052 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4055 } else if ((SplatBits & BitMask) != 0) {
4061 // Op=1, Cmode=1110.
4064 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4069 llvm_unreachable("unexpected size for isNEONModifiedImm");
4072 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4073 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4076 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4077 const ARMSubtarget *ST) const {
4078 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
4081 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4082 assert(Op.getValueType() == MVT::f32 &&
4083 "ConstantFP custom lowering should only occur for f32.");
4085 // Try splatting with a VMOV.f32...
4086 APFloat FPVal = CFP->getValueAPF();
4087 int ImmVal = ARM_AM::getFP32Imm(FPVal);
4089 DebugLoc DL = Op.getDebugLoc();
4090 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4091 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4093 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4094 DAG.getConstant(0, MVT::i32));
4097 // If that fails, try a VMOV.i32
4099 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
4100 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
4102 if (NewVal != SDValue()) {
4103 DebugLoc DL = Op.getDebugLoc();
4104 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4106 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4108 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4109 DAG.getConstant(0, MVT::i32));
4112 // Finally, try a VMVN.i32
4113 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
4115 if (NewVal != SDValue()) {
4116 DebugLoc DL = Op.getDebugLoc();
4117 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4118 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4120 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4121 DAG.getConstant(0, MVT::i32));
4127 // check if an VEXT instruction can handle the shuffle mask when the
4128 // vector sources of the shuffle are the same.
4129 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4130 unsigned NumElts = VT.getVectorNumElements();
4132 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4138 // If this is a VEXT shuffle, the immediate value is the index of the first
4139 // element. The other shuffle indices must be the successive elements after
4141 unsigned ExpectedElt = Imm;
4142 for (unsigned i = 1; i < NumElts; ++i) {
4143 // Increment the expected index. If it wraps around, just follow it
4144 // back to index zero and keep going.
4146 if (ExpectedElt == NumElts)
4149 if (M[i] < 0) continue; // ignore UNDEF indices
4150 if (ExpectedElt != static_cast<unsigned>(M[i]))
4158 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4159 bool &ReverseVEXT, unsigned &Imm) {
4160 unsigned NumElts = VT.getVectorNumElements();
4161 ReverseVEXT = false;
4163 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4169 // If this is a VEXT shuffle, the immediate value is the index of the first
4170 // element. The other shuffle indices must be the successive elements after
4172 unsigned ExpectedElt = Imm;
4173 for (unsigned i = 1; i < NumElts; ++i) {
4174 // Increment the expected index. If it wraps around, it may still be
4175 // a VEXT but the source vectors must be swapped.
4177 if (ExpectedElt == NumElts * 2) {
4182 if (M[i] < 0) continue; // ignore UNDEF indices
4183 if (ExpectedElt != static_cast<unsigned>(M[i]))
4187 // Adjust the index value if the source operands will be swapped.
4194 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4195 /// instruction with the specified blocksize. (The order of the elements
4196 /// within each block of the vector is reversed.)
4197 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4198 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4199 "Only possible block sizes for VREV are: 16, 32, 64");
4201 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4205 unsigned NumElts = VT.getVectorNumElements();
4206 unsigned BlockElts = M[0] + 1;
4207 // If the first shuffle index is UNDEF, be optimistic.
4209 BlockElts = BlockSize / EltSz;
4211 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4214 for (unsigned i = 0; i < NumElts; ++i) {
4215 if (M[i] < 0) continue; // ignore UNDEF indices
4216 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4223 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4224 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4225 // range, then 0 is placed into the resulting vector. So pretty much any mask
4226 // of 8 elements can work here.
4227 return VT == MVT::v8i8 && M.size() == 8;
4230 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4231 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4235 unsigned NumElts = VT.getVectorNumElements();
4236 WhichResult = (M[0] == 0 ? 0 : 1);
4237 for (unsigned i = 0; i < NumElts; i += 2) {
4238 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4239 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4245 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4246 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4247 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4248 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4249 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4253 unsigned NumElts = VT.getVectorNumElements();
4254 WhichResult = (M[0] == 0 ? 0 : 1);
4255 for (unsigned i = 0; i < NumElts; i += 2) {
4256 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4257 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4263 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4264 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4268 unsigned NumElts = VT.getVectorNumElements();
4269 WhichResult = (M[0] == 0 ? 0 : 1);
4270 for (unsigned i = 0; i != NumElts; ++i) {
4271 if (M[i] < 0) continue; // ignore UNDEF indices
4272 if ((unsigned) M[i] != 2 * i + WhichResult)
4276 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4277 if (VT.is64BitVector() && EltSz == 32)
4283 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4284 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4285 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4286 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4287 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4291 unsigned Half = VT.getVectorNumElements() / 2;
4292 WhichResult = (M[0] == 0 ? 0 : 1);
4293 for (unsigned j = 0; j != 2; ++j) {
4294 unsigned Idx = WhichResult;
4295 for (unsigned i = 0; i != Half; ++i) {
4296 int MIdx = M[i + j * Half];
4297 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4303 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4304 if (VT.is64BitVector() && EltSz == 32)
4310 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4311 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4315 unsigned NumElts = VT.getVectorNumElements();
4316 WhichResult = (M[0] == 0 ? 0 : 1);
4317 unsigned Idx = WhichResult * NumElts / 2;
4318 for (unsigned i = 0; i != NumElts; i += 2) {
4319 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4320 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4325 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4326 if (VT.is64BitVector() && EltSz == 32)
4332 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4333 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4334 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4335 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4336 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4340 unsigned NumElts = VT.getVectorNumElements();
4341 WhichResult = (M[0] == 0 ? 0 : 1);
4342 unsigned Idx = WhichResult * NumElts / 2;
4343 for (unsigned i = 0; i != NumElts; i += 2) {
4344 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4345 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4350 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4351 if (VT.is64BitVector() && EltSz == 32)
4357 /// \return true if this is a reverse operation on an vector.
4358 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4359 unsigned NumElts = VT.getVectorNumElements();
4360 // Make sure the mask has the right size.
4361 if (NumElts != M.size())
4364 // Look for <15, ..., 3, -1, 1, 0>.
4365 for (unsigned i = 0; i != NumElts; ++i)
4366 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4372 // If N is an integer constant that can be moved into a register in one
4373 // instruction, return an SDValue of such a constant (will become a MOV
4374 // instruction). Otherwise return null.
4375 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4376 const ARMSubtarget *ST, DebugLoc dl) {
4378 if (!isa<ConstantSDNode>(N))
4380 Val = cast<ConstantSDNode>(N)->getZExtValue();
4382 if (ST->isThumb1Only()) {
4383 if (Val <= 255 || ~Val <= 255)
4384 return DAG.getConstant(Val, MVT::i32);
4386 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4387 return DAG.getConstant(Val, MVT::i32);
4392 // If this is a case we can't handle, return null and let the default
4393 // expansion code take care of it.
4394 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4395 const ARMSubtarget *ST) const {
4396 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4397 DebugLoc dl = Op.getDebugLoc();
4398 EVT VT = Op.getValueType();
4400 APInt SplatBits, SplatUndef;
4401 unsigned SplatBitSize;
4403 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4404 if (SplatBitSize <= 64) {
4405 // Check if an immediate VMOV works.
4407 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4408 SplatUndef.getZExtValue(), SplatBitSize,
4409 DAG, VmovVT, VT.is128BitVector(),
4411 if (Val.getNode()) {
4412 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4413 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4416 // Try an immediate VMVN.
4417 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4418 Val = isNEONModifiedImm(NegatedImm,
4419 SplatUndef.getZExtValue(), SplatBitSize,
4420 DAG, VmovVT, VT.is128BitVector(),
4422 if (Val.getNode()) {
4423 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4424 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4427 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4428 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4429 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4431 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4432 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4438 // Scan through the operands to see if only one value is used.
4440 // As an optimisation, even if more than one value is used it may be more
4441 // profitable to splat with one value then change some lanes.
4443 // Heuristically we decide to do this if the vector has a "dominant" value,
4444 // defined as splatted to more than half of the lanes.
4445 unsigned NumElts = VT.getVectorNumElements();
4446 bool isOnlyLowElement = true;
4447 bool usesOnlyOneValue = true;
4448 bool hasDominantValue = false;
4449 bool isConstant = true;
4451 // Map of the number of times a particular SDValue appears in the
4453 DenseMap<SDValue, unsigned> ValueCounts;
4455 for (unsigned i = 0; i < NumElts; ++i) {
4456 SDValue V = Op.getOperand(i);
4457 if (V.getOpcode() == ISD::UNDEF)
4460 isOnlyLowElement = false;
4461 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4464 ValueCounts.insert(std::make_pair(V, 0));
4465 unsigned &Count = ValueCounts[V];
4467 // Is this value dominant? (takes up more than half of the lanes)
4468 if (++Count > (NumElts / 2)) {
4469 hasDominantValue = true;
4473 if (ValueCounts.size() != 1)
4474 usesOnlyOneValue = false;
4475 if (!Value.getNode() && ValueCounts.size() > 0)
4476 Value = ValueCounts.begin()->first;
4478 if (ValueCounts.size() == 0)
4479 return DAG.getUNDEF(VT);
4481 if (isOnlyLowElement)
4482 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4484 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4486 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4487 // i32 and try again.
4488 if (hasDominantValue && EltSize <= 32) {
4492 // If we are VDUPing a value that comes directly from a vector, that will
4493 // cause an unnecessary move to and from a GPR, where instead we could
4494 // just use VDUPLANE. We can only do this if the lane being extracted
4495 // is at a constant index, as the VDUP from lane instructions only have
4496 // constant-index forms.
4497 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4498 isa<ConstantSDNode>(Value->getOperand(1))) {
4499 // We need to create a new undef vector to use for the VDUPLANE if the
4500 // size of the vector from which we get the value is different than the
4501 // size of the vector that we need to create. We will insert the element
4502 // such that the register coalescer will remove unnecessary copies.
4503 if (VT != Value->getOperand(0).getValueType()) {
4504 ConstantSDNode *constIndex;
4505 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4506 assert(constIndex && "The index is not a constant!");
4507 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4508 VT.getVectorNumElements();
4509 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4510 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4511 Value, DAG.getConstant(index, MVT::i32)),
4512 DAG.getConstant(index, MVT::i32));
4514 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4515 Value->getOperand(0), Value->getOperand(1));
4517 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4519 if (!usesOnlyOneValue) {
4520 // The dominant value was splatted as 'N', but we now have to insert
4521 // all differing elements.
4522 for (unsigned I = 0; I < NumElts; ++I) {
4523 if (Op.getOperand(I) == Value)
4525 SmallVector<SDValue, 3> Ops;
4527 Ops.push_back(Op.getOperand(I));
4528 Ops.push_back(DAG.getConstant(I, MVT::i32));
4529 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4534 if (VT.getVectorElementType().isFloatingPoint()) {
4535 SmallVector<SDValue, 8> Ops;
4536 for (unsigned i = 0; i < NumElts; ++i)
4537 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4539 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4540 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4541 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4543 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4545 if (usesOnlyOneValue) {
4546 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4547 if (isConstant && Val.getNode())
4548 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4552 // If all elements are constants and the case above didn't get hit, fall back
4553 // to the default expansion, which will generate a load from the constant
4558 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4560 SDValue shuffle = ReconstructShuffle(Op, DAG);
4561 if (shuffle != SDValue())
4565 // Vectors with 32- or 64-bit elements can be built by directly assigning
4566 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4567 // will be legalized.
4568 if (EltSize >= 32) {
4569 // Do the expansion with floating-point types, since that is what the VFP
4570 // registers are defined to use, and since i64 is not legal.
4571 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4572 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4573 SmallVector<SDValue, 8> Ops;
4574 for (unsigned i = 0; i < NumElts; ++i)
4575 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4576 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4577 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4583 // Gather data to see if the operation can be modelled as a
4584 // shuffle in combination with VEXTs.
4585 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4586 SelectionDAG &DAG) const {
4587 DebugLoc dl = Op.getDebugLoc();
4588 EVT VT = Op.getValueType();
4589 unsigned NumElts = VT.getVectorNumElements();
4591 SmallVector<SDValue, 2> SourceVecs;
4592 SmallVector<unsigned, 2> MinElts;
4593 SmallVector<unsigned, 2> MaxElts;
4595 for (unsigned i = 0; i < NumElts; ++i) {
4596 SDValue V = Op.getOperand(i);
4597 if (V.getOpcode() == ISD::UNDEF)
4599 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4600 // A shuffle can only come from building a vector from various
4601 // elements of other vectors.
4603 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4604 VT.getVectorElementType()) {
4605 // This code doesn't know how to handle shuffles where the vector
4606 // element types do not match (this happens because type legalization
4607 // promotes the return type of EXTRACT_VECTOR_ELT).
4608 // FIXME: It might be appropriate to extend this code to handle
4609 // mismatched types.
4613 // Record this extraction against the appropriate vector if possible...
4614 SDValue SourceVec = V.getOperand(0);
4615 // If the element number isn't a constant, we can't effectively
4616 // analyze what's going on.
4617 if (!isa<ConstantSDNode>(V.getOperand(1)))
4619 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4620 bool FoundSource = false;
4621 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4622 if (SourceVecs[j] == SourceVec) {
4623 if (MinElts[j] > EltNo)
4625 if (MaxElts[j] < EltNo)
4632 // Or record a new source if not...
4634 SourceVecs.push_back(SourceVec);
4635 MinElts.push_back(EltNo);
4636 MaxElts.push_back(EltNo);
4640 // Currently only do something sane when at most two source vectors
4642 if (SourceVecs.size() > 2)
4645 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4646 int VEXTOffsets[2] = {0, 0};
4648 // This loop extracts the usage patterns of the source vectors
4649 // and prepares appropriate SDValues for a shuffle if possible.
4650 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4651 if (SourceVecs[i].getValueType() == VT) {
4652 // No VEXT necessary
4653 ShuffleSrcs[i] = SourceVecs[i];
4656 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4657 // It probably isn't worth padding out a smaller vector just to
4658 // break it down again in a shuffle.
4662 // Since only 64-bit and 128-bit vectors are legal on ARM and
4663 // we've eliminated the other cases...
4664 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4665 "unexpected vector sizes in ReconstructShuffle");
4667 if (MaxElts[i] - MinElts[i] >= NumElts) {
4668 // Span too large for a VEXT to cope
4672 if (MinElts[i] >= NumElts) {
4673 // The extraction can just take the second half
4674 VEXTOffsets[i] = NumElts;
4675 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4677 DAG.getIntPtrConstant(NumElts));
4678 } else if (MaxElts[i] < NumElts) {
4679 // The extraction can just take the first half
4681 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4683 DAG.getIntPtrConstant(0));
4685 // An actual VEXT is needed
4686 VEXTOffsets[i] = MinElts[i];
4687 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4689 DAG.getIntPtrConstant(0));
4690 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4692 DAG.getIntPtrConstant(NumElts));
4693 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4694 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4698 SmallVector<int, 8> Mask;
4700 for (unsigned i = 0; i < NumElts; ++i) {
4701 SDValue Entry = Op.getOperand(i);
4702 if (Entry.getOpcode() == ISD::UNDEF) {
4707 SDValue ExtractVec = Entry.getOperand(0);
4708 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4709 .getOperand(1))->getSExtValue();
4710 if (ExtractVec == SourceVecs[0]) {
4711 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4713 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4717 // Final check before we try to produce nonsense...
4718 if (isShuffleMaskLegal(Mask, VT))
4719 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4725 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4726 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4727 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4728 /// are assumed to be legal.
4730 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4732 if (VT.getVectorNumElements() == 4 &&
4733 (VT.is128BitVector() || VT.is64BitVector())) {
4734 unsigned PFIndexes[4];
4735 for (unsigned i = 0; i != 4; ++i) {
4739 PFIndexes[i] = M[i];
4742 // Compute the index in the perfect shuffle table.
4743 unsigned PFTableIndex =
4744 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4745 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4746 unsigned Cost = (PFEntry >> 30);
4753 unsigned Imm, WhichResult;
4755 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4756 return (EltSize >= 32 ||
4757 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4758 isVREVMask(M, VT, 64) ||
4759 isVREVMask(M, VT, 32) ||
4760 isVREVMask(M, VT, 16) ||
4761 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4762 isVTBLMask(M, VT) ||
4763 isVTRNMask(M, VT, WhichResult) ||
4764 isVUZPMask(M, VT, WhichResult) ||
4765 isVZIPMask(M, VT, WhichResult) ||
4766 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4767 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4768 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
4769 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
4772 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4773 /// the specified operations to build the shuffle.
4774 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4775 SDValue RHS, SelectionDAG &DAG,
4777 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4778 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4779 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4782 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4791 OP_VUZPL, // VUZP, left result
4792 OP_VUZPR, // VUZP, right result
4793 OP_VZIPL, // VZIP, left result
4794 OP_VZIPR, // VZIP, right result
4795 OP_VTRNL, // VTRN, left result
4796 OP_VTRNR // VTRN, right result
4799 if (OpNum == OP_COPY) {
4800 if (LHSID == (1*9+2)*9+3) return LHS;
4801 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4805 SDValue OpLHS, OpRHS;
4806 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4807 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4808 EVT VT = OpLHS.getValueType();
4811 default: llvm_unreachable("Unknown shuffle opcode!");
4813 // VREV divides the vector in half and swaps within the half.
4814 if (VT.getVectorElementType() == MVT::i32 ||
4815 VT.getVectorElementType() == MVT::f32)
4816 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4817 // vrev <4 x i16> -> VREV32
4818 if (VT.getVectorElementType() == MVT::i16)
4819 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4820 // vrev <4 x i8> -> VREV16
4821 assert(VT.getVectorElementType() == MVT::i8);
4822 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4827 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4828 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4832 return DAG.getNode(ARMISD::VEXT, dl, VT,
4834 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4837 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4838 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4841 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4842 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4845 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4846 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4850 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4851 ArrayRef<int> ShuffleMask,
4852 SelectionDAG &DAG) {
4853 // Check to see if we can use the VTBL instruction.
4854 SDValue V1 = Op.getOperand(0);
4855 SDValue V2 = Op.getOperand(1);
4856 DebugLoc DL = Op.getDebugLoc();
4858 SmallVector<SDValue, 8> VTBLMask;
4859 for (ArrayRef<int>::iterator
4860 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4861 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4863 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4864 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4865 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4868 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4869 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4873 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
4874 SelectionDAG &DAG) {
4875 DebugLoc DL = Op.getDebugLoc();
4876 SDValue OpLHS = Op.getOperand(0);
4877 EVT VT = OpLHS.getValueType();
4879 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
4880 "Expect an v8i16/v16i8 type");
4881 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
4882 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
4883 // extract the first 8 bytes into the top double word and the last 8 bytes
4884 // into the bottom double word. The v8i16 case is similar.
4885 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
4886 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
4887 DAG.getConstant(ExtractNum, MVT::i32));
4890 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4891 SDValue V1 = Op.getOperand(0);
4892 SDValue V2 = Op.getOperand(1);
4893 DebugLoc dl = Op.getDebugLoc();
4894 EVT VT = Op.getValueType();
4895 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4897 // Convert shuffles that are directly supported on NEON to target-specific
4898 // DAG nodes, instead of keeping them as shuffles and matching them again
4899 // during code selection. This is more efficient and avoids the possibility
4900 // of inconsistencies between legalization and selection.
4901 // FIXME: floating-point vectors should be canonicalized to integer vectors
4902 // of the same time so that they get CSEd properly.
4903 ArrayRef<int> ShuffleMask = SVN->getMask();
4905 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4906 if (EltSize <= 32) {
4907 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4908 int Lane = SVN->getSplatIndex();
4909 // If this is undef splat, generate it via "just" vdup, if possible.
4910 if (Lane == -1) Lane = 0;
4912 // Test if V1 is a SCALAR_TO_VECTOR.
4913 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4914 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4916 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4917 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4919 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4920 !isa<ConstantSDNode>(V1.getOperand(0))) {
4921 bool IsScalarToVector = true;
4922 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4923 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4924 IsScalarToVector = false;
4927 if (IsScalarToVector)
4928 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4930 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4931 DAG.getConstant(Lane, MVT::i32));
4936 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4939 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4940 DAG.getConstant(Imm, MVT::i32));
4943 if (isVREVMask(ShuffleMask, VT, 64))
4944 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4945 if (isVREVMask(ShuffleMask, VT, 32))
4946 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4947 if (isVREVMask(ShuffleMask, VT, 16))
4948 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4950 if (V2->getOpcode() == ISD::UNDEF &&
4951 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4952 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4953 DAG.getConstant(Imm, MVT::i32));
4956 // Check for Neon shuffles that modify both input vectors in place.
4957 // If both results are used, i.e., if there are two shuffles with the same
4958 // source operands and with masks corresponding to both results of one of
4959 // these operations, DAG memoization will ensure that a single node is
4960 // used for both shuffles.
4961 unsigned WhichResult;
4962 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4963 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4964 V1, V2).getValue(WhichResult);
4965 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4966 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4967 V1, V2).getValue(WhichResult);
4968 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4969 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4970 V1, V2).getValue(WhichResult);
4972 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4973 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4974 V1, V1).getValue(WhichResult);
4975 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4976 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4977 V1, V1).getValue(WhichResult);
4978 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4979 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4980 V1, V1).getValue(WhichResult);
4983 // If the shuffle is not directly supported and it has 4 elements, use
4984 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4985 unsigned NumElts = VT.getVectorNumElements();
4987 unsigned PFIndexes[4];
4988 for (unsigned i = 0; i != 4; ++i) {
4989 if (ShuffleMask[i] < 0)
4992 PFIndexes[i] = ShuffleMask[i];
4995 // Compute the index in the perfect shuffle table.
4996 unsigned PFTableIndex =
4997 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4998 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4999 unsigned Cost = (PFEntry >> 30);
5002 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5005 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5006 if (EltSize >= 32) {
5007 // Do the expansion with floating-point types, since that is what the VFP
5008 // registers are defined to use, and since i64 is not legal.
5009 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5010 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5011 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5012 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5013 SmallVector<SDValue, 8> Ops;
5014 for (unsigned i = 0; i < NumElts; ++i) {
5015 if (ShuffleMask[i] < 0)
5016 Ops.push_back(DAG.getUNDEF(EltVT));
5018 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5019 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5020 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5023 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5024 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5027 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5028 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5030 if (VT == MVT::v8i8) {
5031 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5032 if (NewOp.getNode())
5039 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5040 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5041 SDValue Lane = Op.getOperand(2);
5042 if (!isa<ConstantSDNode>(Lane))
5048 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5049 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5050 SDValue Lane = Op.getOperand(1);
5051 if (!isa<ConstantSDNode>(Lane))
5054 SDValue Vec = Op.getOperand(0);
5055 if (Op.getValueType() == MVT::i32 &&
5056 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5057 DebugLoc dl = Op.getDebugLoc();
5058 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5064 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5065 // The only time a CONCAT_VECTORS operation can have legal types is when
5066 // two 64-bit vectors are concatenated to a 128-bit vector.
5067 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5068 "unexpected CONCAT_VECTORS");
5069 DebugLoc dl = Op.getDebugLoc();
5070 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5071 SDValue Op0 = Op.getOperand(0);
5072 SDValue Op1 = Op.getOperand(1);
5073 if (Op0.getOpcode() != ISD::UNDEF)
5074 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5075 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5076 DAG.getIntPtrConstant(0));
5077 if (Op1.getOpcode() != ISD::UNDEF)
5078 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5079 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5080 DAG.getIntPtrConstant(1));
5081 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5084 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5085 /// element has been zero/sign-extended, depending on the isSigned parameter,
5086 /// from an integer type half its size.
5087 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5089 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5090 EVT VT = N->getValueType(0);
5091 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5092 SDNode *BVN = N->getOperand(0).getNode();
5093 if (BVN->getValueType(0) != MVT::v4i32 ||
5094 BVN->getOpcode() != ISD::BUILD_VECTOR)
5096 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5097 unsigned HiElt = 1 - LoElt;
5098 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5099 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5100 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5101 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5102 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5105 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5106 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5109 if (Hi0->isNullValue() && Hi1->isNullValue())
5115 if (N->getOpcode() != ISD::BUILD_VECTOR)
5118 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5119 SDNode *Elt = N->getOperand(i).getNode();
5120 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5121 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5122 unsigned HalfSize = EltSize / 2;
5124 if (!isIntN(HalfSize, C->getSExtValue()))
5127 if (!isUIntN(HalfSize, C->getZExtValue()))
5138 /// isSignExtended - Check if a node is a vector value that is sign-extended
5139 /// or a constant BUILD_VECTOR with sign-extended elements.
5140 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5141 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5143 if (isExtendedBUILD_VECTOR(N, DAG, true))
5148 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5149 /// or a constant BUILD_VECTOR with zero-extended elements.
5150 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5151 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5153 if (isExtendedBUILD_VECTOR(N, DAG, false))
5158 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5159 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5160 /// We insert the required extension here to get the vector to fill a D register.
5161 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5164 unsigned ExtOpcode) {
5165 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5166 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5167 // 64-bits we need to insert a new extension so that it will be 64-bits.
5168 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5169 if (OrigTy.getSizeInBits() >= 64)
5172 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5173 MVT::SimpleValueType OrigSimpleTy = OrigTy.getSimpleVT().SimpleTy;
5175 switch (OrigSimpleTy) {
5176 default: llvm_unreachable("Unexpected Orig Vector Type");
5185 return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N);
5188 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5189 /// does not do any sign/zero extension. If the original vector is less
5190 /// than 64 bits, an appropriate extension will be added after the load to
5191 /// reach a total size of 64 bits. We have to add the extension separately
5192 /// because ARM does not have a sign/zero extending load for vectors.
5193 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5194 SDValue NonExtendingLoad =
5195 DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(),
5196 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5197 LD->isNonTemporal(), LD->isInvariant(),
5198 LD->getAlignment());
5200 switch (LD->getExtensionType()) {
5201 default: llvm_unreachable("Unexpected LoadExtType");
5203 case ISD::SEXTLOAD: ExtOp = ISD::SIGN_EXTEND; break;
5204 case ISD::ZEXTLOAD: ExtOp = ISD::ZERO_EXTEND; break;
5206 MVT::SimpleValueType MemType = LD->getMemoryVT().getSimpleVT().SimpleTy;
5207 MVT::SimpleValueType ExtType = LD->getValueType(0).getSimpleVT().SimpleTy;
5208 return AddRequiredExtensionForVMULL(NonExtendingLoad, DAG,
5209 MemType, ExtType, ExtOp);
5212 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5213 /// extending load, or BUILD_VECTOR with extended elements, return the
5214 /// unextended value. The unextended vector should be 64 bits so that it can
5215 /// be used as an operand to a VMULL instruction. If the original vector size
5216 /// before extension is less than 64 bits we add a an extension to resize
5217 /// the vector to 64 bits.
5218 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5219 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5220 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5221 N->getOperand(0)->getValueType(0),
5225 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5226 return SkipLoadExtensionForVMULL(LD, DAG);
5228 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5229 // have been legalized as a BITCAST from v4i32.
5230 if (N->getOpcode() == ISD::BITCAST) {
5231 SDNode *BVN = N->getOperand(0).getNode();
5232 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5233 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5234 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5235 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
5236 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5238 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5239 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5240 EVT VT = N->getValueType(0);
5241 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5242 unsigned NumElts = VT.getVectorNumElements();
5243 MVT TruncVT = MVT::getIntegerVT(EltSize);
5244 SmallVector<SDValue, 8> Ops;
5245 for (unsigned i = 0; i != NumElts; ++i) {
5246 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5247 const APInt &CInt = C->getAPIntValue();
5248 // Element types smaller than 32 bits are not legal, so use i32 elements.
5249 // The values are implicitly truncated so sext vs. zext doesn't matter.
5250 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5252 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5253 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5256 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5257 unsigned Opcode = N->getOpcode();
5258 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5259 SDNode *N0 = N->getOperand(0).getNode();
5260 SDNode *N1 = N->getOperand(1).getNode();
5261 return N0->hasOneUse() && N1->hasOneUse() &&
5262 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5267 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5268 unsigned Opcode = N->getOpcode();
5269 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5270 SDNode *N0 = N->getOperand(0).getNode();
5271 SDNode *N1 = N->getOperand(1).getNode();
5272 return N0->hasOneUse() && N1->hasOneUse() &&
5273 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5278 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5279 // Multiplications are only custom-lowered for 128-bit vectors so that
5280 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5281 EVT VT = Op.getValueType();
5282 assert(VT.is128BitVector() && VT.isInteger() &&
5283 "unexpected type for custom-lowering ISD::MUL");
5284 SDNode *N0 = Op.getOperand(0).getNode();
5285 SDNode *N1 = Op.getOperand(1).getNode();
5286 unsigned NewOpc = 0;
5288 bool isN0SExt = isSignExtended(N0, DAG);
5289 bool isN1SExt = isSignExtended(N1, DAG);
5290 if (isN0SExt && isN1SExt)
5291 NewOpc = ARMISD::VMULLs;
5293 bool isN0ZExt = isZeroExtended(N0, DAG);
5294 bool isN1ZExt = isZeroExtended(N1, DAG);
5295 if (isN0ZExt && isN1ZExt)
5296 NewOpc = ARMISD::VMULLu;
5297 else if (isN1SExt || isN1ZExt) {
5298 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5299 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5300 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5301 NewOpc = ARMISD::VMULLs;
5303 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5304 NewOpc = ARMISD::VMULLu;
5306 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5308 NewOpc = ARMISD::VMULLu;
5314 if (VT == MVT::v2i64)
5315 // Fall through to expand this. It is not legal.
5318 // Other vector multiplications are legal.
5323 // Legalize to a VMULL instruction.
5324 DebugLoc DL = Op.getDebugLoc();
5326 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5328 Op0 = SkipExtensionForVMULL(N0, DAG);
5329 assert(Op0.getValueType().is64BitVector() &&
5330 Op1.getValueType().is64BitVector() &&
5331 "unexpected types for extended operands to VMULL");
5332 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5335 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5336 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5343 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5344 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5345 EVT Op1VT = Op1.getValueType();
5346 return DAG.getNode(N0->getOpcode(), DL, VT,
5347 DAG.getNode(NewOpc, DL, VT,
5348 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5349 DAG.getNode(NewOpc, DL, VT,
5350 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5354 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5356 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5357 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5358 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5359 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5360 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5361 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5362 // Get reciprocal estimate.
5363 // float4 recip = vrecpeq_f32(yf);
5364 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5365 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5366 // Because char has a smaller range than uchar, we can actually get away
5367 // without any newton steps. This requires that we use a weird bias
5368 // of 0xb000, however (again, this has been exhaustively tested).
5369 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5370 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5371 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5372 Y = DAG.getConstant(0xb000, MVT::i32);
5373 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5374 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5375 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5376 // Convert back to short.
5377 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5378 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5383 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5385 // Convert to float.
5386 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5387 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5388 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5389 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5390 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5391 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5393 // Use reciprocal estimate and one refinement step.
5394 // float4 recip = vrecpeq_f32(yf);
5395 // recip *= vrecpsq_f32(yf, recip);
5396 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5397 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5398 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5399 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5401 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5402 // Because short has a smaller range than ushort, we can actually get away
5403 // with only a single newton step. This requires that we use a weird bias
5404 // of 89, however (again, this has been exhaustively tested).
5405 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5406 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5407 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5408 N1 = DAG.getConstant(0x89, MVT::i32);
5409 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5410 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5411 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5412 // Convert back to integer and return.
5413 // return vmovn_s32(vcvt_s32_f32(result));
5414 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5415 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5419 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5420 EVT VT = Op.getValueType();
5421 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5422 "unexpected type for custom-lowering ISD::SDIV");
5424 DebugLoc dl = Op.getDebugLoc();
5425 SDValue N0 = Op.getOperand(0);
5426 SDValue N1 = Op.getOperand(1);
5429 if (VT == MVT::v8i8) {
5430 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5431 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5433 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5434 DAG.getIntPtrConstant(4));
5435 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5436 DAG.getIntPtrConstant(4));
5437 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5438 DAG.getIntPtrConstant(0));
5439 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5440 DAG.getIntPtrConstant(0));
5442 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5443 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5445 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5446 N0 = LowerCONCAT_VECTORS(N0, DAG);
5448 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5451 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5454 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5455 EVT VT = Op.getValueType();
5456 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5457 "unexpected type for custom-lowering ISD::UDIV");
5459 DebugLoc dl = Op.getDebugLoc();
5460 SDValue N0 = Op.getOperand(0);
5461 SDValue N1 = Op.getOperand(1);
5464 if (VT == MVT::v8i8) {
5465 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5466 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5468 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5469 DAG.getIntPtrConstant(4));
5470 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5471 DAG.getIntPtrConstant(4));
5472 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5473 DAG.getIntPtrConstant(0));
5474 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5475 DAG.getIntPtrConstant(0));
5477 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5478 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5480 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5481 N0 = LowerCONCAT_VECTORS(N0, DAG);
5483 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5484 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5489 // v4i16 sdiv ... Convert to float.
5490 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5491 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5492 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5493 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5494 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5495 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5497 // Use reciprocal estimate and two refinement steps.
5498 // float4 recip = vrecpeq_f32(yf);
5499 // recip *= vrecpsq_f32(yf, recip);
5500 // recip *= vrecpsq_f32(yf, recip);
5501 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5502 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5503 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5504 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5506 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5507 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5508 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5510 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5511 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5512 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5513 // and that it will never cause us to return an answer too large).
5514 // float4 result = as_float4(as_int4(xf*recip) + 2);
5515 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5516 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5517 N1 = DAG.getConstant(2, MVT::i32);
5518 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5519 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5520 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5521 // Convert back to integer and return.
5522 // return vmovn_u32(vcvt_s32_f32(result));
5523 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5524 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5528 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5529 EVT VT = Op.getNode()->getValueType(0);
5530 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5533 bool ExtraOp = false;
5534 switch (Op.getOpcode()) {
5535 default: llvm_unreachable("Invalid code");
5536 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5537 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5538 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5539 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5543 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5545 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5546 Op.getOperand(1), Op.getOperand(2));
5549 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5550 // Monotonic load/store is legal for all targets
5551 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5554 // Aquire/Release load/store is not legal for targets without a
5555 // dmb or equivalent available.
5561 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5562 SelectionDAG &DAG, unsigned NewOp) {
5563 DebugLoc dl = Node->getDebugLoc();
5564 assert (Node->getValueType(0) == MVT::i64 &&
5565 "Only know how to expand i64 atomics");
5567 SmallVector<SDValue, 6> Ops;
5568 Ops.push_back(Node->getOperand(0)); // Chain
5569 Ops.push_back(Node->getOperand(1)); // Ptr
5571 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5572 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5573 // High part of Val1
5574 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5575 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5576 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5577 // High part of Val1
5578 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5579 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5580 // High part of Val2
5581 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5582 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5584 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5586 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5587 cast<MemSDNode>(Node)->getMemOperand());
5588 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5589 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5590 Results.push_back(Result.getValue(2));
5593 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5594 switch (Op.getOpcode()) {
5595 default: llvm_unreachable("Don't know how to custom lower this!");
5596 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5597 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5598 case ISD::GlobalAddress:
5599 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5600 LowerGlobalAddressELF(Op, DAG);
5601 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5602 case ISD::SELECT: return LowerSELECT(Op, DAG);
5603 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5604 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5605 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5606 case ISD::VASTART: return LowerVASTART(Op, DAG);
5607 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5608 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5609 case ISD::SINT_TO_FP:
5610 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5611 case ISD::FP_TO_SINT:
5612 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5613 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5614 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5615 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5616 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5617 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5618 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5619 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5621 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5624 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5625 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5626 case ISD::SRL_PARTS:
5627 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5628 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5629 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
5630 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5631 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5632 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5633 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5634 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5635 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5636 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5637 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5638 case ISD::MUL: return LowerMUL(Op, DAG);
5639 case ISD::SDIV: return LowerSDIV(Op, DAG);
5640 case ISD::UDIV: return LowerUDIV(Op, DAG);
5644 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5645 case ISD::ATOMIC_LOAD:
5646 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5650 /// ReplaceNodeResults - Replace the results of node with an illegal result
5651 /// type with new values built out of custom code.
5652 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5653 SmallVectorImpl<SDValue>&Results,
5654 SelectionDAG &DAG) const {
5656 switch (N->getOpcode()) {
5658 llvm_unreachable("Don't know how to custom expand this!");
5660 Res = ExpandBITCAST(N, DAG);
5662 case ISD::SIGN_EXTEND:
5663 case ISD::ZERO_EXTEND:
5664 Res = ExpandVectorExtension(N, DAG);
5668 Res = Expand64BitShift(N, DAG, Subtarget);
5670 case ISD::ATOMIC_LOAD_ADD:
5671 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5673 case ISD::ATOMIC_LOAD_AND:
5674 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5676 case ISD::ATOMIC_LOAD_NAND:
5677 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5679 case ISD::ATOMIC_LOAD_OR:
5680 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5682 case ISD::ATOMIC_LOAD_SUB:
5683 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5685 case ISD::ATOMIC_LOAD_XOR:
5686 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5688 case ISD::ATOMIC_SWAP:
5689 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5691 case ISD::ATOMIC_CMP_SWAP:
5692 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5694 case ISD::ATOMIC_LOAD_MIN:
5695 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
5697 case ISD::ATOMIC_LOAD_UMIN:
5698 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
5700 case ISD::ATOMIC_LOAD_MAX:
5701 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
5703 case ISD::ATOMIC_LOAD_UMAX:
5704 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
5708 Results.push_back(Res);
5711 //===----------------------------------------------------------------------===//
5712 // ARM Scheduler Hooks
5713 //===----------------------------------------------------------------------===//
5716 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5717 MachineBasicBlock *BB,
5718 unsigned Size) const {
5719 unsigned dest = MI->getOperand(0).getReg();
5720 unsigned ptr = MI->getOperand(1).getReg();
5721 unsigned oldval = MI->getOperand(2).getReg();
5722 unsigned newval = MI->getOperand(3).getReg();
5723 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5724 DebugLoc dl = MI->getDebugLoc();
5725 bool isThumb2 = Subtarget->isThumb2();
5727 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5728 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5729 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5730 (const TargetRegisterClass*)&ARM::GPRRegClass);
5733 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5734 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5735 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5738 unsigned ldrOpc, strOpc;
5740 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5742 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5743 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5746 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5747 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5750 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5751 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5755 MachineFunction *MF = BB->getParent();
5756 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5757 MachineFunction::iterator It = BB;
5758 ++It; // insert the new blocks after the current block
5760 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5761 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5762 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5763 MF->insert(It, loop1MBB);
5764 MF->insert(It, loop2MBB);
5765 MF->insert(It, exitMBB);
5767 // Transfer the remainder of BB and its successor edges to exitMBB.
5768 exitMBB->splice(exitMBB->begin(), BB,
5769 llvm::next(MachineBasicBlock::iterator(MI)),
5771 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5775 // fallthrough --> loop1MBB
5776 BB->addSuccessor(loop1MBB);
5779 // ldrex dest, [ptr]
5783 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5784 if (ldrOpc == ARM::t2LDREX)
5786 AddDefaultPred(MIB);
5787 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5788 .addReg(dest).addReg(oldval));
5789 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5790 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5791 BB->addSuccessor(loop2MBB);
5792 BB->addSuccessor(exitMBB);
5795 // strex scratch, newval, [ptr]
5799 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5800 if (strOpc == ARM::t2STREX)
5802 AddDefaultPred(MIB);
5803 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5804 .addReg(scratch).addImm(0));
5805 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5806 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5807 BB->addSuccessor(loop1MBB);
5808 BB->addSuccessor(exitMBB);
5814 MI->eraseFromParent(); // The instruction is gone now.
5820 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5821 unsigned Size, unsigned BinOpcode) const {
5822 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5823 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5825 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5826 MachineFunction *MF = BB->getParent();
5827 MachineFunction::iterator It = BB;
5830 unsigned dest = MI->getOperand(0).getReg();
5831 unsigned ptr = MI->getOperand(1).getReg();
5832 unsigned incr = MI->getOperand(2).getReg();
5833 DebugLoc dl = MI->getDebugLoc();
5834 bool isThumb2 = Subtarget->isThumb2();
5836 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5838 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5839 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5842 unsigned ldrOpc, strOpc;
5844 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5846 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5847 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5850 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5851 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5854 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5855 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5859 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5860 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5861 MF->insert(It, loopMBB);
5862 MF->insert(It, exitMBB);
5864 // Transfer the remainder of BB and its successor edges to exitMBB.
5865 exitMBB->splice(exitMBB->begin(), BB,
5866 llvm::next(MachineBasicBlock::iterator(MI)),
5868 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5870 const TargetRegisterClass *TRC = isThumb2 ?
5871 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5872 (const TargetRegisterClass*)&ARM::GPRRegClass;
5873 unsigned scratch = MRI.createVirtualRegister(TRC);
5874 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5878 // fallthrough --> loopMBB
5879 BB->addSuccessor(loopMBB);
5883 // <binop> scratch2, dest, incr
5884 // strex scratch, scratch2, ptr
5887 // fallthrough --> exitMBB
5889 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5890 if (ldrOpc == ARM::t2LDREX)
5892 AddDefaultPred(MIB);
5894 // operand order needs to go the other way for NAND
5895 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5896 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5897 addReg(incr).addReg(dest)).addReg(0);
5899 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5900 addReg(dest).addReg(incr)).addReg(0);
5903 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5904 if (strOpc == ARM::t2STREX)
5906 AddDefaultPred(MIB);
5907 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5908 .addReg(scratch).addImm(0));
5909 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5910 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5912 BB->addSuccessor(loopMBB);
5913 BB->addSuccessor(exitMBB);
5919 MI->eraseFromParent(); // The instruction is gone now.
5925 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5926 MachineBasicBlock *BB,
5929 ARMCC::CondCodes Cond) const {
5930 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5932 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5933 MachineFunction *MF = BB->getParent();
5934 MachineFunction::iterator It = BB;
5937 unsigned dest = MI->getOperand(0).getReg();
5938 unsigned ptr = MI->getOperand(1).getReg();
5939 unsigned incr = MI->getOperand(2).getReg();
5940 unsigned oldval = dest;
5941 DebugLoc dl = MI->getDebugLoc();
5942 bool isThumb2 = Subtarget->isThumb2();
5944 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5946 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5947 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5950 unsigned ldrOpc, strOpc, extendOpc;
5952 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5954 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5955 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5956 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5959 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5960 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5961 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5964 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5965 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5970 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5971 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5972 MF->insert(It, loopMBB);
5973 MF->insert(It, exitMBB);
5975 // Transfer the remainder of BB and its successor edges to exitMBB.
5976 exitMBB->splice(exitMBB->begin(), BB,
5977 llvm::next(MachineBasicBlock::iterator(MI)),
5979 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5981 const TargetRegisterClass *TRC = isThumb2 ?
5982 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5983 (const TargetRegisterClass*)&ARM::GPRRegClass;
5984 unsigned scratch = MRI.createVirtualRegister(TRC);
5985 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5989 // fallthrough --> loopMBB
5990 BB->addSuccessor(loopMBB);
5994 // (sign extend dest, if required)
5996 // cmov.cond scratch2, incr, dest
5997 // strex scratch, scratch2, ptr
6000 // fallthrough --> exitMBB
6002 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6003 if (ldrOpc == ARM::t2LDREX)
6005 AddDefaultPred(MIB);
6007 // Sign extend the value, if necessary.
6008 if (signExtend && extendOpc) {
6009 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
6010 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6015 // Build compare and cmov instructions.
6016 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6017 .addReg(oldval).addReg(incr));
6018 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6019 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6021 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6022 if (strOpc == ARM::t2STREX)
6024 AddDefaultPred(MIB);
6025 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6026 .addReg(scratch).addImm(0));
6027 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6028 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6030 BB->addSuccessor(loopMBB);
6031 BB->addSuccessor(exitMBB);
6037 MI->eraseFromParent(); // The instruction is gone now.
6043 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6044 unsigned Op1, unsigned Op2,
6045 bool NeedsCarry, bool IsCmpxchg,
6046 bool IsMinMax, ARMCC::CondCodes CC) const {
6047 // This also handles ATOMIC_SWAP, indicated by Op1==0.
6048 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6050 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6051 MachineFunction *MF = BB->getParent();
6052 MachineFunction::iterator It = BB;
6055 unsigned destlo = MI->getOperand(0).getReg();
6056 unsigned desthi = MI->getOperand(1).getReg();
6057 unsigned ptr = MI->getOperand(2).getReg();
6058 unsigned vallo = MI->getOperand(3).getReg();
6059 unsigned valhi = MI->getOperand(4).getReg();
6060 DebugLoc dl = MI->getDebugLoc();
6061 bool isThumb2 = Subtarget->isThumb2();
6063 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6065 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6066 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6067 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6070 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6071 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6072 if (IsCmpxchg || IsMinMax)
6073 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6075 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6076 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6078 MF->insert(It, loopMBB);
6079 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6080 if (IsCmpxchg) MF->insert(It, cont2BB);
6081 MF->insert(It, exitMBB);
6083 // Transfer the remainder of BB and its successor edges to exitMBB.
6084 exitMBB->splice(exitMBB->begin(), BB,
6085 llvm::next(MachineBasicBlock::iterator(MI)),
6087 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6089 const TargetRegisterClass *TRC = isThumb2 ?
6090 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6091 (const TargetRegisterClass*)&ARM::GPRRegClass;
6092 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6096 // fallthrough --> loopMBB
6097 BB->addSuccessor(loopMBB);
6100 // ldrexd r2, r3, ptr
6101 // <binopa> r0, r2, incr
6102 // <binopb> r1, r3, incr
6103 // strexd storesuccess, r0, r1, ptr
6104 // cmp storesuccess, #0
6106 // fallthrough --> exitMBB
6111 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD))
6112 .addReg(destlo, RegState::Define)
6113 .addReg(desthi, RegState::Define)
6116 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6117 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD))
6118 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6119 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6120 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6121 .addReg(GPRPair0, 0, ARM::gsub_0);
6122 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6123 .addReg(GPRPair0, 0, ARM::gsub_1);
6126 unsigned StoreLo, StoreHi;
6129 for (unsigned i = 0; i < 2; i++) {
6130 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6132 .addReg(i == 0 ? destlo : desthi)
6133 .addReg(i == 0 ? vallo : valhi));
6134 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6135 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6136 BB->addSuccessor(exitMBB);
6137 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6138 BB = (i == 0 ? contBB : cont2BB);
6141 // Copy to physregs for strexd
6142 StoreLo = MI->getOperand(5).getReg();
6143 StoreHi = MI->getOperand(6).getReg();
6145 // Perform binary operation
6146 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6147 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6148 .addReg(destlo).addReg(vallo))
6149 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6150 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6151 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6152 .addReg(desthi).addReg(valhi))
6153 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6158 // Copy to physregs for strexd
6163 // Compare and branch to exit block.
6164 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6165 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6166 BB->addSuccessor(exitMBB);
6167 BB->addSuccessor(contBB);
6175 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess)
6176 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6178 // Marshal a pair...
6179 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6180 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6181 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6182 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6183 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6186 .addImm(ARM::gsub_0);
6187 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6190 .addImm(ARM::gsub_1);
6193 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess)
6194 .addReg(StorePair).addReg(ptr));
6197 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6198 .addReg(storesuccess).addImm(0));
6199 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6200 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6202 BB->addSuccessor(loopMBB);
6203 BB->addSuccessor(exitMBB);
6209 MI->eraseFromParent(); // The instruction is gone now.
6214 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6215 /// registers the function context.
6216 void ARMTargetLowering::
6217 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6218 MachineBasicBlock *DispatchBB, int FI) const {
6219 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6220 DebugLoc dl = MI->getDebugLoc();
6221 MachineFunction *MF = MBB->getParent();
6222 MachineRegisterInfo *MRI = &MF->getRegInfo();
6223 MachineConstantPool *MCP = MF->getConstantPool();
6224 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6225 const Function *F = MF->getFunction();
6227 bool isThumb = Subtarget->isThumb();
6228 bool isThumb2 = Subtarget->isThumb2();
6230 unsigned PCLabelId = AFI->createPICLabelUId();
6231 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6232 ARMConstantPoolValue *CPV =
6233 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6234 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6236 const TargetRegisterClass *TRC = isThumb ?
6237 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6238 (const TargetRegisterClass*)&ARM::GPRRegClass;
6240 // Grab constant pool and fixed stack memory operands.
6241 MachineMemOperand *CPMMO =
6242 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6243 MachineMemOperand::MOLoad, 4, 4);
6245 MachineMemOperand *FIMMOSt =
6246 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6247 MachineMemOperand::MOStore, 4, 4);
6249 // Load the address of the dispatch MBB into the jump buffer.
6251 // Incoming value: jbuf
6252 // ldr.n r5, LCPI1_1
6255 // str r5, [$jbuf, #+4] ; &jbuf[1]
6256 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6257 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6258 .addConstantPoolIndex(CPI)
6259 .addMemOperand(CPMMO));
6260 // Set the low bit because of thumb mode.
6261 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6263 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6264 .addReg(NewVReg1, RegState::Kill)
6266 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6267 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6268 .addReg(NewVReg2, RegState::Kill)
6270 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6271 .addReg(NewVReg3, RegState::Kill)
6273 .addImm(36) // &jbuf[1] :: pc
6274 .addMemOperand(FIMMOSt));
6275 } else if (isThumb) {
6276 // Incoming value: jbuf
6277 // ldr.n r1, LCPI1_4
6281 // add r2, $jbuf, #+4 ; &jbuf[1]
6283 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6284 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6285 .addConstantPoolIndex(CPI)
6286 .addMemOperand(CPMMO));
6287 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6288 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6289 .addReg(NewVReg1, RegState::Kill)
6291 // Set the low bit because of thumb mode.
6292 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6293 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6294 .addReg(ARM::CPSR, RegState::Define)
6296 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6297 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6298 .addReg(ARM::CPSR, RegState::Define)
6299 .addReg(NewVReg2, RegState::Kill)
6300 .addReg(NewVReg3, RegState::Kill));
6301 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6302 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6304 .addImm(36)); // &jbuf[1] :: pc
6305 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6306 .addReg(NewVReg4, RegState::Kill)
6307 .addReg(NewVReg5, RegState::Kill)
6309 .addMemOperand(FIMMOSt));
6311 // Incoming value: jbuf
6314 // str r1, [$jbuf, #+4] ; &jbuf[1]
6315 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6316 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6317 .addConstantPoolIndex(CPI)
6319 .addMemOperand(CPMMO));
6320 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6321 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6322 .addReg(NewVReg1, RegState::Kill)
6323 .addImm(PCLabelId));
6324 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6325 .addReg(NewVReg2, RegState::Kill)
6327 .addImm(36) // &jbuf[1] :: pc
6328 .addMemOperand(FIMMOSt));
6332 MachineBasicBlock *ARMTargetLowering::
6333 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6334 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6335 DebugLoc dl = MI->getDebugLoc();
6336 MachineFunction *MF = MBB->getParent();
6337 MachineRegisterInfo *MRI = &MF->getRegInfo();
6338 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6339 MachineFrameInfo *MFI = MF->getFrameInfo();
6340 int FI = MFI->getFunctionContextIndex();
6342 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6343 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6344 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6346 // Get a mapping of the call site numbers to all of the landing pads they're
6348 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6349 unsigned MaxCSNum = 0;
6350 MachineModuleInfo &MMI = MF->getMMI();
6351 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6353 if (!BB->isLandingPad()) continue;
6355 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6357 for (MachineBasicBlock::iterator
6358 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6359 if (!II->isEHLabel()) continue;
6361 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6362 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6364 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6365 for (SmallVectorImpl<unsigned>::iterator
6366 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6367 CSI != CSE; ++CSI) {
6368 CallSiteNumToLPad[*CSI].push_back(BB);
6369 MaxCSNum = std::max(MaxCSNum, *CSI);
6375 // Get an ordered list of the machine basic blocks for the jump table.
6376 std::vector<MachineBasicBlock*> LPadList;
6377 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6378 LPadList.reserve(CallSiteNumToLPad.size());
6379 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6380 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6381 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6382 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6383 LPadList.push_back(*II);
6384 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6388 assert(!LPadList.empty() &&
6389 "No landing pad destinations for the dispatch jump table!");
6391 // Create the jump table and associated information.
6392 MachineJumpTableInfo *JTI =
6393 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6394 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6395 unsigned UId = AFI->createJumpTableUId();
6396 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6398 // Create the MBBs for the dispatch code.
6400 // Shove the dispatch's address into the return slot in the function context.
6401 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6402 DispatchBB->setIsLandingPad();
6404 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6405 unsigned trap_opcode;
6406 if (Subtarget->isThumb())
6407 trap_opcode = ARM::tTRAP;
6409 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6411 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6412 DispatchBB->addSuccessor(TrapBB);
6414 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6415 DispatchBB->addSuccessor(DispContBB);
6418 MF->insert(MF->end(), DispatchBB);
6419 MF->insert(MF->end(), DispContBB);
6420 MF->insert(MF->end(), TrapBB);
6422 // Insert code into the entry block that creates and registers the function
6424 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6426 MachineMemOperand *FIMMOLd =
6427 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6428 MachineMemOperand::MOLoad |
6429 MachineMemOperand::MOVolatile, 4, 4);
6431 MachineInstrBuilder MIB;
6432 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6434 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6435 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6437 // Add a register mask with no preserved registers. This results in all
6438 // registers being marked as clobbered.
6439 MIB.addRegMask(RI.getNoPreservedMask());
6441 unsigned NumLPads = LPadList.size();
6442 if (Subtarget->isThumb2()) {
6443 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6444 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6447 .addMemOperand(FIMMOLd));
6449 if (NumLPads < 256) {
6450 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6452 .addImm(LPadList.size()));
6454 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6455 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6456 .addImm(NumLPads & 0xFFFF));
6458 unsigned VReg2 = VReg1;
6459 if ((NumLPads & 0xFFFF0000) != 0) {
6460 VReg2 = MRI->createVirtualRegister(TRC);
6461 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6463 .addImm(NumLPads >> 16));
6466 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6471 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6476 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6477 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6478 .addJumpTableIndex(MJTI)
6481 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6484 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6485 .addReg(NewVReg3, RegState::Kill)
6487 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6489 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6490 .addReg(NewVReg4, RegState::Kill)
6492 .addJumpTableIndex(MJTI)
6494 } else if (Subtarget->isThumb()) {
6495 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6496 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6499 .addMemOperand(FIMMOLd));
6501 if (NumLPads < 256) {
6502 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6506 MachineConstantPool *ConstantPool = MF->getConstantPool();
6507 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6508 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6510 // MachineConstantPool wants an explicit alignment.
6511 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6513 Align = getDataLayout()->getTypeAllocSize(C->getType());
6514 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6516 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6517 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6518 .addReg(VReg1, RegState::Define)
6519 .addConstantPoolIndex(Idx));
6520 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6525 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6530 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6531 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6532 .addReg(ARM::CPSR, RegState::Define)
6536 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6537 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6538 .addJumpTableIndex(MJTI)
6541 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6542 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6543 .addReg(ARM::CPSR, RegState::Define)
6544 .addReg(NewVReg2, RegState::Kill)
6547 MachineMemOperand *JTMMOLd =
6548 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6549 MachineMemOperand::MOLoad, 4, 4);
6551 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6552 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6553 .addReg(NewVReg4, RegState::Kill)
6555 .addMemOperand(JTMMOLd));
6557 unsigned NewVReg6 = NewVReg5;
6558 if (RelocM == Reloc::PIC_) {
6559 NewVReg6 = MRI->createVirtualRegister(TRC);
6560 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6561 .addReg(ARM::CPSR, RegState::Define)
6562 .addReg(NewVReg5, RegState::Kill)
6566 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6567 .addReg(NewVReg6, RegState::Kill)
6568 .addJumpTableIndex(MJTI)
6571 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6572 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6575 .addMemOperand(FIMMOLd));
6577 if (NumLPads < 256) {
6578 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6581 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6582 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6583 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6584 .addImm(NumLPads & 0xFFFF));
6586 unsigned VReg2 = VReg1;
6587 if ((NumLPads & 0xFFFF0000) != 0) {
6588 VReg2 = MRI->createVirtualRegister(TRC);
6589 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6591 .addImm(NumLPads >> 16));
6594 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6598 MachineConstantPool *ConstantPool = MF->getConstantPool();
6599 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6600 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6602 // MachineConstantPool wants an explicit alignment.
6603 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6605 Align = getDataLayout()->getTypeAllocSize(C->getType());
6606 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6608 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6609 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6610 .addReg(VReg1, RegState::Define)
6611 .addConstantPoolIndex(Idx)
6613 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6615 .addReg(VReg1, RegState::Kill));
6618 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6623 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6625 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6627 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6628 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6629 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6630 .addJumpTableIndex(MJTI)
6633 MachineMemOperand *JTMMOLd =
6634 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6635 MachineMemOperand::MOLoad, 4, 4);
6636 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6638 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6639 .addReg(NewVReg3, RegState::Kill)
6642 .addMemOperand(JTMMOLd));
6644 if (RelocM == Reloc::PIC_) {
6645 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6646 .addReg(NewVReg5, RegState::Kill)
6648 .addJumpTableIndex(MJTI)
6651 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6652 .addReg(NewVReg5, RegState::Kill)
6653 .addJumpTableIndex(MJTI)
6658 // Add the jump table entries as successors to the MBB.
6659 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6660 for (std::vector<MachineBasicBlock*>::iterator
6661 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6662 MachineBasicBlock *CurMBB = *I;
6663 if (SeenMBBs.insert(CurMBB))
6664 DispContBB->addSuccessor(CurMBB);
6667 // N.B. the order the invoke BBs are processed in doesn't matter here.
6668 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6669 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6670 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6671 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6672 MachineBasicBlock *BB = *I;
6674 // Remove the landing pad successor from the invoke block and replace it
6675 // with the new dispatch block.
6676 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6678 while (!Successors.empty()) {
6679 MachineBasicBlock *SMBB = Successors.pop_back_val();
6680 if (SMBB->isLandingPad()) {
6681 BB->removeSuccessor(SMBB);
6682 MBBLPads.push_back(SMBB);
6686 BB->addSuccessor(DispatchBB);
6688 // Find the invoke call and mark all of the callee-saved registers as
6689 // 'implicit defined' so that they're spilled. This prevents code from
6690 // moving instructions to before the EH block, where they will never be
6692 for (MachineBasicBlock::reverse_iterator
6693 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6694 if (!II->isCall()) continue;
6696 DenseMap<unsigned, bool> DefRegs;
6697 for (MachineInstr::mop_iterator
6698 OI = II->operands_begin(), OE = II->operands_end();
6700 if (!OI->isReg()) continue;
6701 DefRegs[OI->getReg()] = true;
6704 MachineInstrBuilder MIB(*MF, &*II);
6706 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6707 unsigned Reg = SavedRegs[i];
6708 if (Subtarget->isThumb2() &&
6709 !ARM::tGPRRegClass.contains(Reg) &&
6710 !ARM::hGPRRegClass.contains(Reg))
6712 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6714 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6717 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6724 // Mark all former landing pads as non-landing pads. The dispatch is the only
6726 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6727 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6728 (*I)->setIsLandingPad(false);
6730 // The instruction is gone now.
6731 MI->eraseFromParent();
6737 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6738 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6739 E = MBB->succ_end(); I != E; ++I)
6742 llvm_unreachable("Expecting a BB with two successors!");
6745 MachineBasicBlock *ARMTargetLowering::
6746 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6747 // This pseudo instruction has 3 operands: dst, src, size
6748 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6749 // Otherwise, we will generate unrolled scalar copies.
6750 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6751 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6752 MachineFunction::iterator It = BB;
6755 unsigned dest = MI->getOperand(0).getReg();
6756 unsigned src = MI->getOperand(1).getReg();
6757 unsigned SizeVal = MI->getOperand(2).getImm();
6758 unsigned Align = MI->getOperand(3).getImm();
6759 DebugLoc dl = MI->getDebugLoc();
6761 bool isThumb2 = Subtarget->isThumb2();
6762 MachineFunction *MF = BB->getParent();
6763 MachineRegisterInfo &MRI = MF->getRegInfo();
6764 unsigned ldrOpc, strOpc, UnitSize = 0;
6766 const TargetRegisterClass *TRC = isThumb2 ?
6767 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6768 (const TargetRegisterClass*)&ARM::GPRRegClass;
6769 const TargetRegisterClass *TRC_Vec = 0;
6772 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6773 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6775 } else if (Align & 2) {
6776 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6777 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6780 // Check whether we can use NEON instructions.
6781 if (!MF->getFunction()->getAttributes().
6782 hasAttribute(AttributeSet::FunctionIndex,
6783 Attribute::NoImplicitFloat) &&
6784 Subtarget->hasNEON()) {
6785 if ((Align % 16 == 0) && SizeVal >= 16) {
6786 ldrOpc = ARM::VLD1q32wb_fixed;
6787 strOpc = ARM::VST1q32wb_fixed;
6789 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6791 else if ((Align % 8 == 0) && SizeVal >= 8) {
6792 ldrOpc = ARM::VLD1d32wb_fixed;
6793 strOpc = ARM::VST1d32wb_fixed;
6795 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6798 // Can't use NEON instructions.
6799 if (UnitSize == 0) {
6800 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6801 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6806 unsigned BytesLeft = SizeVal % UnitSize;
6807 unsigned LoopSize = SizeVal - BytesLeft;
6809 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6810 // Use LDR and STR to copy.
6811 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6812 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6813 unsigned srcIn = src;
6814 unsigned destIn = dest;
6815 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6816 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6817 unsigned srcOut = MRI.createVirtualRegister(TRC);
6818 unsigned destOut = MRI.createVirtualRegister(TRC);
6819 if (UnitSize >= 8) {
6820 AddDefaultPred(BuildMI(*BB, MI, dl,
6821 TII->get(ldrOpc), scratch)
6822 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6824 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6825 .addReg(destIn).addImm(0).addReg(scratch));
6826 } else if (isThumb2) {
6827 AddDefaultPred(BuildMI(*BB, MI, dl,
6828 TII->get(ldrOpc), scratch)
6829 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6831 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6832 .addReg(scratch).addReg(destIn)
6835 AddDefaultPred(BuildMI(*BB, MI, dl,
6836 TII->get(ldrOpc), scratch)
6837 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6840 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6841 .addReg(scratch).addReg(destIn)
6842 .addReg(0).addImm(UnitSize));
6848 // Handle the leftover bytes with LDRB and STRB.
6849 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6850 // [destOut] = STRB_POST(scratch, destIn, 1)
6851 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6852 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6853 for (unsigned i = 0; i < BytesLeft; i++) {
6854 unsigned scratch = MRI.createVirtualRegister(TRC);
6855 unsigned srcOut = MRI.createVirtualRegister(TRC);
6856 unsigned destOut = MRI.createVirtualRegister(TRC);
6858 AddDefaultPred(BuildMI(*BB, MI, dl,
6859 TII->get(ldrOpc),scratch)
6860 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6862 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6863 .addReg(scratch).addReg(destIn)
6864 .addReg(0).addImm(1));
6866 AddDefaultPred(BuildMI(*BB, MI, dl,
6867 TII->get(ldrOpc),scratch)
6868 .addReg(srcOut, RegState::Define).addReg(srcIn)
6869 .addReg(0).addImm(1));
6871 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6872 .addReg(scratch).addReg(destIn)
6873 .addReg(0).addImm(1));
6878 MI->eraseFromParent(); // The instruction is gone now.
6882 // Expand the pseudo op to a loop.
6885 // movw varEnd, # --> with thumb2
6887 // ldrcp varEnd, idx --> without thumb2
6888 // fallthrough --> loopMBB
6890 // PHI varPhi, varEnd, varLoop
6891 // PHI srcPhi, src, srcLoop
6892 // PHI destPhi, dst, destLoop
6893 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6894 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6895 // subs varLoop, varPhi, #UnitSize
6897 // fallthrough --> exitMBB
6899 // epilogue to handle left-over bytes
6900 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6901 // [destOut] = STRB_POST(scratch, destLoop, 1)
6902 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6903 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6904 MF->insert(It, loopMBB);
6905 MF->insert(It, exitMBB);
6907 // Transfer the remainder of BB and its successor edges to exitMBB.
6908 exitMBB->splice(exitMBB->begin(), BB,
6909 llvm::next(MachineBasicBlock::iterator(MI)),
6911 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6913 // Load an immediate to varEnd.
6914 unsigned varEnd = MRI.createVirtualRegister(TRC);
6916 unsigned VReg1 = varEnd;
6917 if ((LoopSize & 0xFFFF0000) != 0)
6918 VReg1 = MRI.createVirtualRegister(TRC);
6919 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6920 .addImm(LoopSize & 0xFFFF));
6922 if ((LoopSize & 0xFFFF0000) != 0)
6923 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6925 .addImm(LoopSize >> 16));
6927 MachineConstantPool *ConstantPool = MF->getConstantPool();
6928 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6929 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6931 // MachineConstantPool wants an explicit alignment.
6932 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6934 Align = getDataLayout()->getTypeAllocSize(C->getType());
6935 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6937 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6938 .addReg(varEnd, RegState::Define)
6939 .addConstantPoolIndex(Idx)
6942 BB->addSuccessor(loopMBB);
6944 // Generate the loop body:
6945 // varPhi = PHI(varLoop, varEnd)
6946 // srcPhi = PHI(srcLoop, src)
6947 // destPhi = PHI(destLoop, dst)
6948 MachineBasicBlock *entryBB = BB;
6950 unsigned varLoop = MRI.createVirtualRegister(TRC);
6951 unsigned varPhi = MRI.createVirtualRegister(TRC);
6952 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6953 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6954 unsigned destLoop = MRI.createVirtualRegister(TRC);
6955 unsigned destPhi = MRI.createVirtualRegister(TRC);
6957 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6958 .addReg(varLoop).addMBB(loopMBB)
6959 .addReg(varEnd).addMBB(entryBB);
6960 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6961 .addReg(srcLoop).addMBB(loopMBB)
6962 .addReg(src).addMBB(entryBB);
6963 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6964 .addReg(destLoop).addMBB(loopMBB)
6965 .addReg(dest).addMBB(entryBB);
6967 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6968 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6969 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6970 if (UnitSize >= 8) {
6971 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6972 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6974 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6975 .addReg(destPhi).addImm(0).addReg(scratch));
6976 } else if (isThumb2) {
6977 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6978 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6980 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6981 .addReg(scratch).addReg(destPhi)
6984 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6985 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6988 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6989 .addReg(scratch).addReg(destPhi)
6990 .addReg(0).addImm(UnitSize));
6993 // Decrement loop variable by UnitSize.
6994 MachineInstrBuilder MIB = BuildMI(BB, dl,
6995 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6996 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6997 MIB->getOperand(5).setReg(ARM::CPSR);
6998 MIB->getOperand(5).setIsDef(true);
7000 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7001 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7003 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7004 BB->addSuccessor(loopMBB);
7005 BB->addSuccessor(exitMBB);
7007 // Add epilogue to handle BytesLeft.
7009 MachineInstr *StartOfExit = exitMBB->begin();
7010 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7011 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7013 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7014 // [destOut] = STRB_POST(scratch, destLoop, 1)
7015 unsigned srcIn = srcLoop;
7016 unsigned destIn = destLoop;
7017 for (unsigned i = 0; i < BytesLeft; i++) {
7018 unsigned scratch = MRI.createVirtualRegister(TRC);
7019 unsigned srcOut = MRI.createVirtualRegister(TRC);
7020 unsigned destOut = MRI.createVirtualRegister(TRC);
7022 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7023 TII->get(ldrOpc),scratch)
7024 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7026 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7027 .addReg(scratch).addReg(destIn)
7030 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7031 TII->get(ldrOpc),scratch)
7032 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
7034 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7035 .addReg(scratch).addReg(destIn)
7036 .addReg(0).addImm(1));
7042 MI->eraseFromParent(); // The instruction is gone now.
7047 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7048 MachineBasicBlock *BB) const {
7049 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7050 DebugLoc dl = MI->getDebugLoc();
7051 bool isThumb2 = Subtarget->isThumb2();
7052 switch (MI->getOpcode()) {
7055 llvm_unreachable("Unexpected instr type to insert");
7057 // The Thumb2 pre-indexed stores have the same MI operands, they just
7058 // define them differently in the .td files from the isel patterns, so
7059 // they need pseudos.
7060 case ARM::t2STR_preidx:
7061 MI->setDesc(TII->get(ARM::t2STR_PRE));
7063 case ARM::t2STRB_preidx:
7064 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7066 case ARM::t2STRH_preidx:
7067 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7070 case ARM::STRi_preidx:
7071 case ARM::STRBi_preidx: {
7072 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7073 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7074 // Decode the offset.
7075 unsigned Offset = MI->getOperand(4).getImm();
7076 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7077 Offset = ARM_AM::getAM2Offset(Offset);
7081 MachineMemOperand *MMO = *MI->memoperands_begin();
7082 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7083 .addOperand(MI->getOperand(0)) // Rn_wb
7084 .addOperand(MI->getOperand(1)) // Rt
7085 .addOperand(MI->getOperand(2)) // Rn
7086 .addImm(Offset) // offset (skip GPR==zero_reg)
7087 .addOperand(MI->getOperand(5)) // pred
7088 .addOperand(MI->getOperand(6))
7089 .addMemOperand(MMO);
7090 MI->eraseFromParent();
7093 case ARM::STRr_preidx:
7094 case ARM::STRBr_preidx:
7095 case ARM::STRH_preidx: {
7097 switch (MI->getOpcode()) {
7098 default: llvm_unreachable("unexpected opcode!");
7099 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7100 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7101 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7103 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7104 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7105 MIB.addOperand(MI->getOperand(i));
7106 MI->eraseFromParent();
7109 case ARM::ATOMIC_LOAD_ADD_I8:
7110 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7111 case ARM::ATOMIC_LOAD_ADD_I16:
7112 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7113 case ARM::ATOMIC_LOAD_ADD_I32:
7114 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7116 case ARM::ATOMIC_LOAD_AND_I8:
7117 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7118 case ARM::ATOMIC_LOAD_AND_I16:
7119 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7120 case ARM::ATOMIC_LOAD_AND_I32:
7121 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7123 case ARM::ATOMIC_LOAD_OR_I8:
7124 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7125 case ARM::ATOMIC_LOAD_OR_I16:
7126 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7127 case ARM::ATOMIC_LOAD_OR_I32:
7128 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7130 case ARM::ATOMIC_LOAD_XOR_I8:
7131 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7132 case ARM::ATOMIC_LOAD_XOR_I16:
7133 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7134 case ARM::ATOMIC_LOAD_XOR_I32:
7135 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7137 case ARM::ATOMIC_LOAD_NAND_I8:
7138 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7139 case ARM::ATOMIC_LOAD_NAND_I16:
7140 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7141 case ARM::ATOMIC_LOAD_NAND_I32:
7142 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7144 case ARM::ATOMIC_LOAD_SUB_I8:
7145 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7146 case ARM::ATOMIC_LOAD_SUB_I16:
7147 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7148 case ARM::ATOMIC_LOAD_SUB_I32:
7149 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7151 case ARM::ATOMIC_LOAD_MIN_I8:
7152 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7153 case ARM::ATOMIC_LOAD_MIN_I16:
7154 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7155 case ARM::ATOMIC_LOAD_MIN_I32:
7156 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7158 case ARM::ATOMIC_LOAD_MAX_I8:
7159 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7160 case ARM::ATOMIC_LOAD_MAX_I16:
7161 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7162 case ARM::ATOMIC_LOAD_MAX_I32:
7163 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7165 case ARM::ATOMIC_LOAD_UMIN_I8:
7166 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7167 case ARM::ATOMIC_LOAD_UMIN_I16:
7168 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7169 case ARM::ATOMIC_LOAD_UMIN_I32:
7170 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7172 case ARM::ATOMIC_LOAD_UMAX_I8:
7173 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7174 case ARM::ATOMIC_LOAD_UMAX_I16:
7175 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7176 case ARM::ATOMIC_LOAD_UMAX_I32:
7177 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7179 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7180 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7181 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7183 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7184 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7185 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7188 case ARM::ATOMADD6432:
7189 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7190 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7191 /*NeedsCarry*/ true);
7192 case ARM::ATOMSUB6432:
7193 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7194 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7195 /*NeedsCarry*/ true);
7196 case ARM::ATOMOR6432:
7197 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7198 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7199 case ARM::ATOMXOR6432:
7200 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7201 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7202 case ARM::ATOMAND6432:
7203 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7204 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7205 case ARM::ATOMSWAP6432:
7206 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7207 case ARM::ATOMCMPXCHG6432:
7208 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7209 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7210 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7211 case ARM::ATOMMIN6432:
7212 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7213 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7214 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7215 /*IsMinMax*/ true, ARMCC::LT);
7216 case ARM::ATOMMAX6432:
7217 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7218 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7219 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7220 /*IsMinMax*/ true, ARMCC::GE);
7221 case ARM::ATOMUMIN6432:
7222 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7223 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7224 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7225 /*IsMinMax*/ true, ARMCC::LO);
7226 case ARM::ATOMUMAX6432:
7227 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7228 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7229 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7230 /*IsMinMax*/ true, ARMCC::HS);
7232 case ARM::tMOVCCr_pseudo: {
7233 // To "insert" a SELECT_CC instruction, we actually have to insert the
7234 // diamond control-flow pattern. The incoming instruction knows the
7235 // destination vreg to set, the condition code register to branch on, the
7236 // true/false values to select between, and a branch opcode to use.
7237 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7238 MachineFunction::iterator It = BB;
7244 // cmpTY ccX, r1, r2
7246 // fallthrough --> copy0MBB
7247 MachineBasicBlock *thisMBB = BB;
7248 MachineFunction *F = BB->getParent();
7249 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7250 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7251 F->insert(It, copy0MBB);
7252 F->insert(It, sinkMBB);
7254 // Transfer the remainder of BB and its successor edges to sinkMBB.
7255 sinkMBB->splice(sinkMBB->begin(), BB,
7256 llvm::next(MachineBasicBlock::iterator(MI)),
7258 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7260 BB->addSuccessor(copy0MBB);
7261 BB->addSuccessor(sinkMBB);
7263 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7264 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7267 // %FalseValue = ...
7268 // # fallthrough to sinkMBB
7271 // Update machine-CFG edges
7272 BB->addSuccessor(sinkMBB);
7275 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7278 BuildMI(*BB, BB->begin(), dl,
7279 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7280 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7281 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7283 MI->eraseFromParent(); // The pseudo instruction is gone now.
7288 case ARM::BCCZi64: {
7289 // If there is an unconditional branch to the other successor, remove it.
7290 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7292 // Compare both parts that make up the double comparison separately for
7294 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7296 unsigned LHS1 = MI->getOperand(1).getReg();
7297 unsigned LHS2 = MI->getOperand(2).getReg();
7299 AddDefaultPred(BuildMI(BB, dl,
7300 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7301 .addReg(LHS1).addImm(0));
7302 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7303 .addReg(LHS2).addImm(0)
7304 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7306 unsigned RHS1 = MI->getOperand(3).getReg();
7307 unsigned RHS2 = MI->getOperand(4).getReg();
7308 AddDefaultPred(BuildMI(BB, dl,
7309 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7310 .addReg(LHS1).addReg(RHS1));
7311 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7312 .addReg(LHS2).addReg(RHS2)
7313 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7316 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7317 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7318 if (MI->getOperand(0).getImm() == ARMCC::NE)
7319 std::swap(destMBB, exitMBB);
7321 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7322 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7324 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7326 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7328 MI->eraseFromParent(); // The pseudo instruction is gone now.
7332 case ARM::Int_eh_sjlj_setjmp:
7333 case ARM::Int_eh_sjlj_setjmp_nofp:
7334 case ARM::tInt_eh_sjlj_setjmp:
7335 case ARM::t2Int_eh_sjlj_setjmp:
7336 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7337 EmitSjLjDispatchBlock(MI, BB);
7342 // To insert an ABS instruction, we have to insert the
7343 // diamond control-flow pattern. The incoming instruction knows the
7344 // source vreg to test against 0, the destination vreg to set,
7345 // the condition code register to branch on, the
7346 // true/false values to select between, and a branch opcode to use.
7351 // BCC (branch to SinkBB if V0 >= 0)
7352 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7353 // SinkBB: V1 = PHI(V2, V3)
7354 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7355 MachineFunction::iterator BBI = BB;
7357 MachineFunction *Fn = BB->getParent();
7358 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7359 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7360 Fn->insert(BBI, RSBBB);
7361 Fn->insert(BBI, SinkBB);
7363 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7364 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7365 bool isThumb2 = Subtarget->isThumb2();
7366 MachineRegisterInfo &MRI = Fn->getRegInfo();
7367 // In Thumb mode S must not be specified if source register is the SP or
7368 // PC and if destination register is the SP, so restrict register class
7369 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7370 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7371 (const TargetRegisterClass*)&ARM::GPRRegClass);
7373 // Transfer the remainder of BB and its successor edges to sinkMBB.
7374 SinkBB->splice(SinkBB->begin(), BB,
7375 llvm::next(MachineBasicBlock::iterator(MI)),
7377 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7379 BB->addSuccessor(RSBBB);
7380 BB->addSuccessor(SinkBB);
7382 // fall through to SinkMBB
7383 RSBBB->addSuccessor(SinkBB);
7385 // insert a cmp at the end of BB
7386 AddDefaultPred(BuildMI(BB, dl,
7387 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7388 .addReg(ABSSrcReg).addImm(0));
7390 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7392 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7393 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7395 // insert rsbri in RSBBB
7396 // Note: BCC and rsbri will be converted into predicated rsbmi
7397 // by if-conversion pass
7398 BuildMI(*RSBBB, RSBBB->begin(), dl,
7399 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7400 .addReg(ABSSrcReg, RegState::Kill)
7401 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7403 // insert PHI in SinkBB,
7404 // reuse ABSDstReg to not change uses of ABS instruction
7405 BuildMI(*SinkBB, SinkBB->begin(), dl,
7406 TII->get(ARM::PHI), ABSDstReg)
7407 .addReg(NewRsbDstReg).addMBB(RSBBB)
7408 .addReg(ABSSrcReg).addMBB(BB);
7410 // remove ABS instruction
7411 MI->eraseFromParent();
7413 // return last added BB
7416 case ARM::COPY_STRUCT_BYVAL_I32:
7418 return EmitStructByval(MI, BB);
7422 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7423 SDNode *Node) const {
7424 if (!MI->hasPostISelHook()) {
7425 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7426 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7430 const MCInstrDesc *MCID = &MI->getDesc();
7431 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7432 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7433 // operand is still set to noreg. If needed, set the optional operand's
7434 // register to CPSR, and remove the redundant implicit def.
7436 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7438 // Rename pseudo opcodes.
7439 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7441 const ARMBaseInstrInfo *TII =
7442 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7443 MCID = &TII->get(NewOpc);
7445 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7446 "converted opcode should be the same except for cc_out");
7450 // Add the optional cc_out operand
7451 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7453 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7455 // Any ARM instruction that sets the 's' bit should specify an optional
7456 // "cc_out" operand in the last operand position.
7457 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7458 assert(!NewOpc && "Optional cc_out operand required");
7461 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7462 // since we already have an optional CPSR def.
7463 bool definesCPSR = false;
7464 bool deadCPSR = false;
7465 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7467 const MachineOperand &MO = MI->getOperand(i);
7468 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7472 MI->RemoveOperand(i);
7477 assert(!NewOpc && "Optional cc_out operand required");
7480 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7482 assert(!MI->getOperand(ccOutIdx).getReg() &&
7483 "expect uninitialized optional cc_out operand");
7487 // If this instruction was defined with an optional CPSR def and its dag node
7488 // had a live implicit CPSR def, then activate the optional CPSR def.
7489 MachineOperand &MO = MI->getOperand(ccOutIdx);
7490 MO.setReg(ARM::CPSR);
7494 //===----------------------------------------------------------------------===//
7495 // ARM Optimization Hooks
7496 //===----------------------------------------------------------------------===//
7498 // Helper function that checks if N is a null or all ones constant.
7499 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7500 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7503 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7506 // Return true if N is conditionally 0 or all ones.
7507 // Detects these expressions where cc is an i1 value:
7509 // (select cc 0, y) [AllOnes=0]
7510 // (select cc y, 0) [AllOnes=0]
7511 // (zext cc) [AllOnes=0]
7512 // (sext cc) [AllOnes=0/1]
7513 // (select cc -1, y) [AllOnes=1]
7514 // (select cc y, -1) [AllOnes=1]
7516 // Invert is set when N is the null/all ones constant when CC is false.
7517 // OtherOp is set to the alternative value of N.
7518 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7519 SDValue &CC, bool &Invert,
7521 SelectionDAG &DAG) {
7522 switch (N->getOpcode()) {
7523 default: return false;
7525 CC = N->getOperand(0);
7526 SDValue N1 = N->getOperand(1);
7527 SDValue N2 = N->getOperand(2);
7528 if (isZeroOrAllOnes(N1, AllOnes)) {
7533 if (isZeroOrAllOnes(N2, AllOnes)) {
7540 case ISD::ZERO_EXTEND:
7541 // (zext cc) can never be the all ones value.
7545 case ISD::SIGN_EXTEND: {
7546 EVT VT = N->getValueType(0);
7547 CC = N->getOperand(0);
7548 if (CC.getValueType() != MVT::i1)
7552 // When looking for an AllOnes constant, N is an sext, and the 'other'
7554 OtherOp = DAG.getConstant(0, VT);
7555 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7556 // When looking for a 0 constant, N can be zext or sext.
7557 OtherOp = DAG.getConstant(1, VT);
7559 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7565 // Combine a constant select operand into its use:
7567 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7568 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7569 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7570 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7571 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7573 // The transform is rejected if the select doesn't have a constant operand that
7574 // is null, or all ones when AllOnes is set.
7576 // Also recognize sext/zext from i1:
7578 // (add (zext cc), x) -> (select cc (add x, 1), x)
7579 // (add (sext cc), x) -> (select cc (add x, -1), x)
7581 // These transformations eventually create predicated instructions.
7583 // @param N The node to transform.
7584 // @param Slct The N operand that is a select.
7585 // @param OtherOp The other N operand (x above).
7586 // @param DCI Context.
7587 // @param AllOnes Require the select constant to be all ones instead of null.
7588 // @returns The new node, or SDValue() on failure.
7590 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7591 TargetLowering::DAGCombinerInfo &DCI,
7592 bool AllOnes = false) {
7593 SelectionDAG &DAG = DCI.DAG;
7594 EVT VT = N->getValueType(0);
7595 SDValue NonConstantVal;
7598 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7599 NonConstantVal, DAG))
7602 // Slct is now know to be the desired identity constant when CC is true.
7603 SDValue TrueVal = OtherOp;
7604 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7605 OtherOp, NonConstantVal);
7606 // Unless SwapSelectOps says CC should be false.
7608 std::swap(TrueVal, FalseVal);
7610 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7611 CCOp, TrueVal, FalseVal);
7614 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7616 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7617 TargetLowering::DAGCombinerInfo &DCI) {
7618 SDValue N0 = N->getOperand(0);
7619 SDValue N1 = N->getOperand(1);
7620 if (N0.getNode()->hasOneUse()) {
7621 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7622 if (Result.getNode())
7625 if (N1.getNode()->hasOneUse()) {
7626 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7627 if (Result.getNode())
7633 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7634 // (only after legalization).
7635 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7636 TargetLowering::DAGCombinerInfo &DCI,
7637 const ARMSubtarget *Subtarget) {
7639 // Only perform optimization if after legalize, and if NEON is available. We
7640 // also expected both operands to be BUILD_VECTORs.
7641 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7642 || N0.getOpcode() != ISD::BUILD_VECTOR
7643 || N1.getOpcode() != ISD::BUILD_VECTOR)
7646 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7647 EVT VT = N->getValueType(0);
7648 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7651 // Check that the vector operands are of the right form.
7652 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7653 // operands, where N is the size of the formed vector.
7654 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7655 // index such that we have a pair wise add pattern.
7657 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7658 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7660 SDValue Vec = N0->getOperand(0)->getOperand(0);
7661 SDNode *V = Vec.getNode();
7662 unsigned nextIndex = 0;
7664 // For each operands to the ADD which are BUILD_VECTORs,
7665 // check to see if each of their operands are an EXTRACT_VECTOR with
7666 // the same vector and appropriate index.
7667 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7668 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7669 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7671 SDValue ExtVec0 = N0->getOperand(i);
7672 SDValue ExtVec1 = N1->getOperand(i);
7674 // First operand is the vector, verify its the same.
7675 if (V != ExtVec0->getOperand(0).getNode() ||
7676 V != ExtVec1->getOperand(0).getNode())
7679 // Second is the constant, verify its correct.
7680 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7681 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7683 // For the constant, we want to see all the even or all the odd.
7684 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7685 || C1->getZExtValue() != nextIndex+1)
7694 // Create VPADDL node.
7695 SelectionDAG &DAG = DCI.DAG;
7696 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7698 // Build operand list.
7699 SmallVector<SDValue, 8> Ops;
7700 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7701 TLI.getPointerTy()));
7703 // Input is the vector.
7706 // Get widened type and narrowed type.
7708 unsigned numElem = VT.getVectorNumElements();
7709 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7710 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7711 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7712 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7714 llvm_unreachable("Invalid vector element type for padd optimization.");
7717 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7718 widenType, &Ops[0], Ops.size());
7719 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7722 static SDValue findMUL_LOHI(SDValue V) {
7723 if (V->getOpcode() == ISD::UMUL_LOHI ||
7724 V->getOpcode() == ISD::SMUL_LOHI)
7729 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7730 TargetLowering::DAGCombinerInfo &DCI,
7731 const ARMSubtarget *Subtarget) {
7733 if (Subtarget->isThumb1Only()) return SDValue();
7735 // Only perform the checks after legalize when the pattern is available.
7736 if (DCI.isBeforeLegalize()) return SDValue();
7738 // Look for multiply add opportunities.
7739 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7740 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7741 // a glue link from the first add to the second add.
7742 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7743 // a S/UMLAL instruction.
7746 // \ / \ [no multiline comment]
7752 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7753 SDValue AddcOp0 = AddcNode->getOperand(0);
7754 SDValue AddcOp1 = AddcNode->getOperand(1);
7756 // Check if the two operands are from the same mul_lohi node.
7757 if (AddcOp0.getNode() == AddcOp1.getNode())
7760 assert(AddcNode->getNumValues() == 2 &&
7761 AddcNode->getValueType(0) == MVT::i32 &&
7762 AddcNode->getValueType(1) == MVT::Glue &&
7763 "Expect ADDC with two result values: i32, glue");
7765 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7766 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7767 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7768 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7769 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7772 // Look for the glued ADDE.
7773 SDNode* AddeNode = AddcNode->getGluedUser();
7774 if (AddeNode == NULL)
7777 // Make sure it is really an ADDE.
7778 if (AddeNode->getOpcode() != ISD::ADDE)
7781 assert(AddeNode->getNumOperands() == 3 &&
7782 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7783 "ADDE node has the wrong inputs");
7785 // Check for the triangle shape.
7786 SDValue AddeOp0 = AddeNode->getOperand(0);
7787 SDValue AddeOp1 = AddeNode->getOperand(1);
7789 // Make sure that the ADDE operands are not coming from the same node.
7790 if (AddeOp0.getNode() == AddeOp1.getNode())
7793 // Find the MUL_LOHI node walking up ADDE's operands.
7794 bool IsLeftOperandMUL = false;
7795 SDValue MULOp = findMUL_LOHI(AddeOp0);
7796 if (MULOp == SDValue())
7797 MULOp = findMUL_LOHI(AddeOp1);
7799 IsLeftOperandMUL = true;
7800 if (MULOp == SDValue())
7803 // Figure out the right opcode.
7804 unsigned Opc = MULOp->getOpcode();
7805 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7807 // Figure out the high and low input values to the MLAL node.
7808 SDValue* HiMul = &MULOp;
7809 SDValue* HiAdd = NULL;
7810 SDValue* LoMul = NULL;
7811 SDValue* LowAdd = NULL;
7813 if (IsLeftOperandMUL)
7819 if (AddcOp0->getOpcode() == Opc) {
7823 if (AddcOp1->getOpcode() == Opc) {
7831 if (LoMul->getNode() != HiMul->getNode())
7834 // Create the merged node.
7835 SelectionDAG &DAG = DCI.DAG;
7837 // Build operand list.
7838 SmallVector<SDValue, 8> Ops;
7839 Ops.push_back(LoMul->getOperand(0));
7840 Ops.push_back(LoMul->getOperand(1));
7841 Ops.push_back(*LowAdd);
7842 Ops.push_back(*HiAdd);
7844 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7845 DAG.getVTList(MVT::i32, MVT::i32),
7846 &Ops[0], Ops.size());
7848 // Replace the ADDs' nodes uses by the MLA node's values.
7849 SDValue HiMLALResult(MLALNode.getNode(), 1);
7850 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7852 SDValue LoMLALResult(MLALNode.getNode(), 0);
7853 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7855 // Return original node to notify the driver to stop replacing.
7856 SDValue resNode(AddcNode, 0);
7860 /// PerformADDCCombine - Target-specific dag combine transform from
7861 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7862 static SDValue PerformADDCCombine(SDNode *N,
7863 TargetLowering::DAGCombinerInfo &DCI,
7864 const ARMSubtarget *Subtarget) {
7866 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7870 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7871 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7872 /// called with the default operands, and if that fails, with commuted
7874 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7875 TargetLowering::DAGCombinerInfo &DCI,
7876 const ARMSubtarget *Subtarget){
7878 // Attempt to create vpaddl for this add.
7879 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7880 if (Result.getNode())
7883 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7884 if (N0.getNode()->hasOneUse()) {
7885 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7886 if (Result.getNode()) return Result;
7891 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7893 static SDValue PerformADDCombine(SDNode *N,
7894 TargetLowering::DAGCombinerInfo &DCI,
7895 const ARMSubtarget *Subtarget) {
7896 SDValue N0 = N->getOperand(0);
7897 SDValue N1 = N->getOperand(1);
7899 // First try with the default operand order.
7900 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7901 if (Result.getNode())
7904 // If that didn't work, try again with the operands commuted.
7905 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7908 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7910 static SDValue PerformSUBCombine(SDNode *N,
7911 TargetLowering::DAGCombinerInfo &DCI) {
7912 SDValue N0 = N->getOperand(0);
7913 SDValue N1 = N->getOperand(1);
7915 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7916 if (N1.getNode()->hasOneUse()) {
7917 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7918 if (Result.getNode()) return Result;
7924 /// PerformVMULCombine
7925 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7926 /// special multiplier accumulator forwarding.
7932 static SDValue PerformVMULCombine(SDNode *N,
7933 TargetLowering::DAGCombinerInfo &DCI,
7934 const ARMSubtarget *Subtarget) {
7935 if (!Subtarget->hasVMLxForwarding())
7938 SelectionDAG &DAG = DCI.DAG;
7939 SDValue N0 = N->getOperand(0);
7940 SDValue N1 = N->getOperand(1);
7941 unsigned Opcode = N0.getOpcode();
7942 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7943 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7944 Opcode = N1.getOpcode();
7945 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7946 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7951 EVT VT = N->getValueType(0);
7952 DebugLoc DL = N->getDebugLoc();
7953 SDValue N00 = N0->getOperand(0);
7954 SDValue N01 = N0->getOperand(1);
7955 return DAG.getNode(Opcode, DL, VT,
7956 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7957 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7960 static SDValue PerformMULCombine(SDNode *N,
7961 TargetLowering::DAGCombinerInfo &DCI,
7962 const ARMSubtarget *Subtarget) {
7963 SelectionDAG &DAG = DCI.DAG;
7965 if (Subtarget->isThumb1Only())
7968 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7971 EVT VT = N->getValueType(0);
7972 if (VT.is64BitVector() || VT.is128BitVector())
7973 return PerformVMULCombine(N, DCI, Subtarget);
7977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7981 int64_t MulAmt = C->getSExtValue();
7982 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7984 ShiftAmt = ShiftAmt & (32 - 1);
7985 SDValue V = N->getOperand(0);
7986 DebugLoc DL = N->getDebugLoc();
7989 MulAmt >>= ShiftAmt;
7992 if (isPowerOf2_32(MulAmt - 1)) {
7993 // (mul x, 2^N + 1) => (add (shl x, N), x)
7994 Res = DAG.getNode(ISD::ADD, DL, VT,
7996 DAG.getNode(ISD::SHL, DL, VT,
7998 DAG.getConstant(Log2_32(MulAmt - 1),
8000 } else if (isPowerOf2_32(MulAmt + 1)) {
8001 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8002 Res = DAG.getNode(ISD::SUB, DL, VT,
8003 DAG.getNode(ISD::SHL, DL, VT,
8005 DAG.getConstant(Log2_32(MulAmt + 1),
8011 uint64_t MulAmtAbs = -MulAmt;
8012 if (isPowerOf2_32(MulAmtAbs + 1)) {
8013 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8014 Res = DAG.getNode(ISD::SUB, DL, VT,
8016 DAG.getNode(ISD::SHL, DL, VT,
8018 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8020 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8021 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8022 Res = DAG.getNode(ISD::ADD, DL, VT,
8024 DAG.getNode(ISD::SHL, DL, VT,
8026 DAG.getConstant(Log2_32(MulAmtAbs-1),
8028 Res = DAG.getNode(ISD::SUB, DL, VT,
8029 DAG.getConstant(0, MVT::i32),Res);
8036 Res = DAG.getNode(ISD::SHL, DL, VT,
8037 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8039 // Do not add new nodes to DAG combiner worklist.
8040 DCI.CombineTo(N, Res, false);
8044 static SDValue PerformANDCombine(SDNode *N,
8045 TargetLowering::DAGCombinerInfo &DCI,
8046 const ARMSubtarget *Subtarget) {
8048 // Attempt to use immediate-form VBIC
8049 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8050 DebugLoc dl = N->getDebugLoc();
8051 EVT VT = N->getValueType(0);
8052 SelectionDAG &DAG = DCI.DAG;
8054 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8057 APInt SplatBits, SplatUndef;
8058 unsigned SplatBitSize;
8061 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8062 if (SplatBitSize <= 64) {
8064 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8065 SplatUndef.getZExtValue(), SplatBitSize,
8066 DAG, VbicVT, VT.is128BitVector(),
8068 if (Val.getNode()) {
8070 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8071 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8072 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8077 if (!Subtarget->isThumb1Only()) {
8078 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8079 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8080 if (Result.getNode())
8087 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8088 static SDValue PerformORCombine(SDNode *N,
8089 TargetLowering::DAGCombinerInfo &DCI,
8090 const ARMSubtarget *Subtarget) {
8091 // Attempt to use immediate-form VORR
8092 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8093 DebugLoc dl = N->getDebugLoc();
8094 EVT VT = N->getValueType(0);
8095 SelectionDAG &DAG = DCI.DAG;
8097 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8100 APInt SplatBits, SplatUndef;
8101 unsigned SplatBitSize;
8103 if (BVN && Subtarget->hasNEON() &&
8104 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8105 if (SplatBitSize <= 64) {
8107 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8108 SplatUndef.getZExtValue(), SplatBitSize,
8109 DAG, VorrVT, VT.is128BitVector(),
8111 if (Val.getNode()) {
8113 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8114 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8115 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8120 if (!Subtarget->isThumb1Only()) {
8121 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8122 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8123 if (Result.getNode())
8127 // The code below optimizes (or (and X, Y), Z).
8128 // The AND operand needs to have a single user to make these optimizations
8130 SDValue N0 = N->getOperand(0);
8131 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8133 SDValue N1 = N->getOperand(1);
8135 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8136 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8137 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8139 unsigned SplatBitSize;
8142 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8144 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8145 HasAnyUndefs) && !HasAnyUndefs) {
8146 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8148 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8149 HasAnyUndefs) && !HasAnyUndefs &&
8150 SplatBits0 == ~SplatBits1) {
8151 // Canonicalize the vector type to make instruction selection simpler.
8152 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8153 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8154 N0->getOperand(1), N0->getOperand(0),
8156 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8161 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8164 // BFI is only available on V6T2+
8165 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8168 DebugLoc DL = N->getDebugLoc();
8169 // 1) or (and A, mask), val => ARMbfi A, val, mask
8170 // iff (val & mask) == val
8172 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8173 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8174 // && mask == ~mask2
8175 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8176 // && ~mask == mask2
8177 // (i.e., copy a bitfield value into another bitfield of the same width)
8182 SDValue N00 = N0.getOperand(0);
8184 // The value and the mask need to be constants so we can verify this is
8185 // actually a bitfield set. If the mask is 0xffff, we can do better
8186 // via a movt instruction, so don't use BFI in that case.
8187 SDValue MaskOp = N0.getOperand(1);
8188 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8191 unsigned Mask = MaskC->getZExtValue();
8195 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8196 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8198 unsigned Val = N1C->getZExtValue();
8199 if ((Val & ~Mask) != Val)
8202 if (ARM::isBitFieldInvertedMask(Mask)) {
8203 Val >>= CountTrailingZeros_32(~Mask);
8205 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8206 DAG.getConstant(Val, MVT::i32),
8207 DAG.getConstant(Mask, MVT::i32));
8209 // Do not add new nodes to DAG combiner worklist.
8210 DCI.CombineTo(N, Res, false);
8213 } else if (N1.getOpcode() == ISD::AND) {
8214 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8215 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8218 unsigned Mask2 = N11C->getZExtValue();
8220 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8222 if (ARM::isBitFieldInvertedMask(Mask) &&
8224 // The pack halfword instruction works better for masks that fit it,
8225 // so use that when it's available.
8226 if (Subtarget->hasT2ExtractPack() &&
8227 (Mask == 0xffff || Mask == 0xffff0000))
8230 unsigned amt = CountTrailingZeros_32(Mask2);
8231 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8232 DAG.getConstant(amt, MVT::i32));
8233 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8234 DAG.getConstant(Mask, MVT::i32));
8235 // Do not add new nodes to DAG combiner worklist.
8236 DCI.CombineTo(N, Res, false);
8238 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8240 // The pack halfword instruction works better for masks that fit it,
8241 // so use that when it's available.
8242 if (Subtarget->hasT2ExtractPack() &&
8243 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8246 unsigned lsb = CountTrailingZeros_32(Mask);
8247 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8248 DAG.getConstant(lsb, MVT::i32));
8249 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8250 DAG.getConstant(Mask2, MVT::i32));
8251 // Do not add new nodes to DAG combiner worklist.
8252 DCI.CombineTo(N, Res, false);
8257 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8258 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8259 ARM::isBitFieldInvertedMask(~Mask)) {
8260 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8261 // where lsb(mask) == #shamt and masked bits of B are known zero.
8262 SDValue ShAmt = N00.getOperand(1);
8263 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8264 unsigned LSB = CountTrailingZeros_32(Mask);
8268 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8269 DAG.getConstant(~Mask, MVT::i32));
8271 // Do not add new nodes to DAG combiner worklist.
8272 DCI.CombineTo(N, Res, false);
8278 static SDValue PerformXORCombine(SDNode *N,
8279 TargetLowering::DAGCombinerInfo &DCI,
8280 const ARMSubtarget *Subtarget) {
8281 EVT VT = N->getValueType(0);
8282 SelectionDAG &DAG = DCI.DAG;
8284 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8287 if (!Subtarget->isThumb1Only()) {
8288 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8289 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8290 if (Result.getNode())
8297 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8298 /// the bits being cleared by the AND are not demanded by the BFI.
8299 static SDValue PerformBFICombine(SDNode *N,
8300 TargetLowering::DAGCombinerInfo &DCI) {
8301 SDValue N1 = N->getOperand(1);
8302 if (N1.getOpcode() == ISD::AND) {
8303 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8306 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8307 unsigned LSB = CountTrailingZeros_32(~InvMask);
8308 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
8309 unsigned Mask = (1 << Width)-1;
8310 unsigned Mask2 = N11C->getZExtValue();
8311 if ((Mask & (~Mask2)) == 0)
8312 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
8313 N->getOperand(0), N1.getOperand(0),
8319 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8320 /// ARMISD::VMOVRRD.
8321 static SDValue PerformVMOVRRDCombine(SDNode *N,
8322 TargetLowering::DAGCombinerInfo &DCI) {
8323 // vmovrrd(vmovdrr x, y) -> x,y
8324 SDValue InDouble = N->getOperand(0);
8325 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8326 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8328 // vmovrrd(load f64) -> (load i32), (load i32)
8329 SDNode *InNode = InDouble.getNode();
8330 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8331 InNode->getValueType(0) == MVT::f64 &&
8332 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8333 !cast<LoadSDNode>(InNode)->isVolatile()) {
8334 // TODO: Should this be done for non-FrameIndex operands?
8335 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8337 SelectionDAG &DAG = DCI.DAG;
8338 DebugLoc DL = LD->getDebugLoc();
8339 SDValue BasePtr = LD->getBasePtr();
8340 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8341 LD->getPointerInfo(), LD->isVolatile(),
8342 LD->isNonTemporal(), LD->isInvariant(),
8343 LD->getAlignment());
8345 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8346 DAG.getConstant(4, MVT::i32));
8347 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8348 LD->getPointerInfo(), LD->isVolatile(),
8349 LD->isNonTemporal(), LD->isInvariant(),
8350 std::min(4U, LD->getAlignment() / 2));
8352 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8353 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8354 DCI.RemoveFromWorklist(LD);
8362 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8363 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8364 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8365 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8366 SDValue Op0 = N->getOperand(0);
8367 SDValue Op1 = N->getOperand(1);
8368 if (Op0.getOpcode() == ISD::BITCAST)
8369 Op0 = Op0.getOperand(0);
8370 if (Op1.getOpcode() == ISD::BITCAST)
8371 Op1 = Op1.getOperand(0);
8372 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8373 Op0.getNode() == Op1.getNode() &&
8374 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8375 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
8376 N->getValueType(0), Op0.getOperand(0));
8380 /// PerformSTORECombine - Target-specific dag combine xforms for
8382 static SDValue PerformSTORECombine(SDNode *N,
8383 TargetLowering::DAGCombinerInfo &DCI) {
8384 StoreSDNode *St = cast<StoreSDNode>(N);
8385 if (St->isVolatile())
8388 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8389 // pack all of the elements in one place. Next, store to memory in fewer
8391 SDValue StVal = St->getValue();
8392 EVT VT = StVal.getValueType();
8393 if (St->isTruncatingStore() && VT.isVector()) {
8394 SelectionDAG &DAG = DCI.DAG;
8395 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8396 EVT StVT = St->getMemoryVT();
8397 unsigned NumElems = VT.getVectorNumElements();
8398 assert(StVT != VT && "Cannot truncate to the same type");
8399 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8400 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8402 // From, To sizes and ElemCount must be pow of two
8403 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8405 // We are going to use the original vector elt for storing.
8406 // Accumulated smaller vector elements must be a multiple of the store size.
8407 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8409 unsigned SizeRatio = FromEltSz / ToEltSz;
8410 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8412 // Create a type on which we perform the shuffle.
8413 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8414 NumElems*SizeRatio);
8415 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8417 DebugLoc DL = St->getDebugLoc();
8418 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8419 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8420 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8422 // Can't shuffle using an illegal type.
8423 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8425 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8426 DAG.getUNDEF(WideVec.getValueType()),
8428 // At this point all of the data is stored at the bottom of the
8429 // register. We now need to save it to mem.
8431 // Find the largest store unit
8432 MVT StoreType = MVT::i8;
8433 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8434 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8435 MVT Tp = (MVT::SimpleValueType)tp;
8436 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8439 // Didn't find a legal store type.
8440 if (!TLI.isTypeLegal(StoreType))
8443 // Bitcast the original vector into a vector of store-size units
8444 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8445 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8446 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8447 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8448 SmallVector<SDValue, 8> Chains;
8449 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8450 TLI.getPointerTy());
8451 SDValue BasePtr = St->getBasePtr();
8453 // Perform one or more big stores into memory.
8454 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8455 for (unsigned I = 0; I < E; I++) {
8456 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8457 StoreType, ShuffWide,
8458 DAG.getIntPtrConstant(I));
8459 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8460 St->getPointerInfo(), St->isVolatile(),
8461 St->isNonTemporal(), St->getAlignment());
8462 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8464 Chains.push_back(Ch);
8466 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8470 if (!ISD::isNormalStore(St))
8473 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8474 // ARM stores of arguments in the same cache line.
8475 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8476 StVal.getNode()->hasOneUse()) {
8477 SelectionDAG &DAG = DCI.DAG;
8478 DebugLoc DL = St->getDebugLoc();
8479 SDValue BasePtr = St->getBasePtr();
8480 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8481 StVal.getNode()->getOperand(0), BasePtr,
8482 St->getPointerInfo(), St->isVolatile(),
8483 St->isNonTemporal(), St->getAlignment());
8485 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8486 DAG.getConstant(4, MVT::i32));
8487 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8488 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8489 St->isNonTemporal(),
8490 std::min(4U, St->getAlignment() / 2));
8493 if (StVal.getValueType() != MVT::i64 ||
8494 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8497 // Bitcast an i64 store extracted from a vector to f64.
8498 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8499 SelectionDAG &DAG = DCI.DAG;
8500 DebugLoc dl = StVal.getDebugLoc();
8501 SDValue IntVec = StVal.getOperand(0);
8502 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8503 IntVec.getValueType().getVectorNumElements());
8504 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8505 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8506 Vec, StVal.getOperand(1));
8507 dl = N->getDebugLoc();
8508 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8509 // Make the DAGCombiner fold the bitcasts.
8510 DCI.AddToWorklist(Vec.getNode());
8511 DCI.AddToWorklist(ExtElt.getNode());
8512 DCI.AddToWorklist(V.getNode());
8513 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8514 St->getPointerInfo(), St->isVolatile(),
8515 St->isNonTemporal(), St->getAlignment(),
8519 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8520 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8521 /// i64 vector to have f64 elements, since the value can then be loaded
8522 /// directly into a VFP register.
8523 static bool hasNormalLoadOperand(SDNode *N) {
8524 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8525 for (unsigned i = 0; i < NumElts; ++i) {
8526 SDNode *Elt = N->getOperand(i).getNode();
8527 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8533 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8534 /// ISD::BUILD_VECTOR.
8535 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8536 TargetLowering::DAGCombinerInfo &DCI){
8537 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8538 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8539 // into a pair of GPRs, which is fine when the value is used as a scalar,
8540 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8541 SelectionDAG &DAG = DCI.DAG;
8542 if (N->getNumOperands() == 2) {
8543 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8548 // Load i64 elements as f64 values so that type legalization does not split
8549 // them up into i32 values.
8550 EVT VT = N->getValueType(0);
8551 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8553 DebugLoc dl = N->getDebugLoc();
8554 SmallVector<SDValue, 8> Ops;
8555 unsigned NumElts = VT.getVectorNumElements();
8556 for (unsigned i = 0; i < NumElts; ++i) {
8557 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8559 // Make the DAGCombiner fold the bitcast.
8560 DCI.AddToWorklist(V.getNode());
8562 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8563 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8564 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8567 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8568 /// ISD::INSERT_VECTOR_ELT.
8569 static SDValue PerformInsertEltCombine(SDNode *N,
8570 TargetLowering::DAGCombinerInfo &DCI) {
8571 // Bitcast an i64 load inserted into a vector to f64.
8572 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8573 EVT VT = N->getValueType(0);
8574 SDNode *Elt = N->getOperand(1).getNode();
8575 if (VT.getVectorElementType() != MVT::i64 ||
8576 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8579 SelectionDAG &DAG = DCI.DAG;
8580 DebugLoc dl = N->getDebugLoc();
8581 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8582 VT.getVectorNumElements());
8583 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8584 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8585 // Make the DAGCombiner fold the bitcasts.
8586 DCI.AddToWorklist(Vec.getNode());
8587 DCI.AddToWorklist(V.getNode());
8588 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8589 Vec, V, N->getOperand(2));
8590 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8593 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8594 /// ISD::VECTOR_SHUFFLE.
8595 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8596 // The LLVM shufflevector instruction does not require the shuffle mask
8597 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8598 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8599 // operands do not match the mask length, they are extended by concatenating
8600 // them with undef vectors. That is probably the right thing for other
8601 // targets, but for NEON it is better to concatenate two double-register
8602 // size vector operands into a single quad-register size vector. Do that
8603 // transformation here:
8604 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8605 // shuffle(concat(v1, v2), undef)
8606 SDValue Op0 = N->getOperand(0);
8607 SDValue Op1 = N->getOperand(1);
8608 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8609 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8610 Op0.getNumOperands() != 2 ||
8611 Op1.getNumOperands() != 2)
8613 SDValue Concat0Op1 = Op0.getOperand(1);
8614 SDValue Concat1Op1 = Op1.getOperand(1);
8615 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8616 Concat1Op1.getOpcode() != ISD::UNDEF)
8618 // Skip the transformation if any of the types are illegal.
8619 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8620 EVT VT = N->getValueType(0);
8621 if (!TLI.isTypeLegal(VT) ||
8622 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8623 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8626 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8627 Op0.getOperand(0), Op1.getOperand(0));
8628 // Translate the shuffle mask.
8629 SmallVector<int, 16> NewMask;
8630 unsigned NumElts = VT.getVectorNumElements();
8631 unsigned HalfElts = NumElts/2;
8632 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8633 for (unsigned n = 0; n < NumElts; ++n) {
8634 int MaskElt = SVN->getMaskElt(n);
8636 if (MaskElt < (int)HalfElts)
8638 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8639 NewElt = HalfElts + MaskElt - NumElts;
8640 NewMask.push_back(NewElt);
8642 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8643 DAG.getUNDEF(VT), NewMask.data());
8646 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8647 /// NEON load/store intrinsics to merge base address updates.
8648 static SDValue CombineBaseUpdate(SDNode *N,
8649 TargetLowering::DAGCombinerInfo &DCI) {
8650 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8653 SelectionDAG &DAG = DCI.DAG;
8654 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8655 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8656 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8657 SDValue Addr = N->getOperand(AddrOpIdx);
8659 // Search for a use of the address operand that is an increment.
8660 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8661 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8663 if (User->getOpcode() != ISD::ADD ||
8664 UI.getUse().getResNo() != Addr.getResNo())
8667 // Check that the add is independent of the load/store. Otherwise, folding
8668 // it would create a cycle.
8669 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8672 // Find the new opcode for the updating load/store.
8674 bool isLaneOp = false;
8675 unsigned NewOpc = 0;
8676 unsigned NumVecs = 0;
8678 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8680 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8681 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8683 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8685 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8687 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8689 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8690 NumVecs = 2; isLaneOp = true; break;
8691 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8692 NumVecs = 3; isLaneOp = true; break;
8693 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8694 NumVecs = 4; isLaneOp = true; break;
8695 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8696 NumVecs = 1; isLoad = false; break;
8697 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8698 NumVecs = 2; isLoad = false; break;
8699 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8700 NumVecs = 3; isLoad = false; break;
8701 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8702 NumVecs = 4; isLoad = false; break;
8703 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8704 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8705 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8706 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8707 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8708 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8712 switch (N->getOpcode()) {
8713 default: llvm_unreachable("unexpected opcode for Neon base update");
8714 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8715 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8716 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8720 // Find the size of memory referenced by the load/store.
8723 VecTy = N->getValueType(0);
8725 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8726 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8728 NumBytes /= VecTy.getVectorNumElements();
8730 // If the increment is a constant, it must match the memory ref size.
8731 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8732 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8733 uint64_t IncVal = CInc->getZExtValue();
8734 if (IncVal != NumBytes)
8736 } else if (NumBytes >= 3 * 16) {
8737 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8738 // separate instructions that make it harder to use a non-constant update.
8742 // Create the new updating load/store node.
8744 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8746 for (n = 0; n < NumResultVecs; ++n)
8748 Tys[n++] = MVT::i32;
8749 Tys[n] = MVT::Other;
8750 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8751 SmallVector<SDValue, 8> Ops;
8752 Ops.push_back(N->getOperand(0)); // incoming chain
8753 Ops.push_back(N->getOperand(AddrOpIdx));
8755 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8756 Ops.push_back(N->getOperand(i));
8758 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8759 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8760 Ops.data(), Ops.size(),
8761 MemInt->getMemoryVT(),
8762 MemInt->getMemOperand());
8765 std::vector<SDValue> NewResults;
8766 for (unsigned i = 0; i < NumResultVecs; ++i) {
8767 NewResults.push_back(SDValue(UpdN.getNode(), i));
8769 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8770 DCI.CombineTo(N, NewResults);
8771 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8778 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8779 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8780 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8782 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8783 SelectionDAG &DAG = DCI.DAG;
8784 EVT VT = N->getValueType(0);
8785 // vldN-dup instructions only support 64-bit vectors for N > 1.
8786 if (!VT.is64BitVector())
8789 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8790 SDNode *VLD = N->getOperand(0).getNode();
8791 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8793 unsigned NumVecs = 0;
8794 unsigned NewOpc = 0;
8795 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8796 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8798 NewOpc = ARMISD::VLD2DUP;
8799 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8801 NewOpc = ARMISD::VLD3DUP;
8802 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8804 NewOpc = ARMISD::VLD4DUP;
8809 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8810 // numbers match the load.
8811 unsigned VLDLaneNo =
8812 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8813 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8815 // Ignore uses of the chain result.
8816 if (UI.getUse().getResNo() == NumVecs)
8819 if (User->getOpcode() != ARMISD::VDUPLANE ||
8820 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8824 // Create the vldN-dup node.
8827 for (n = 0; n < NumVecs; ++n)
8829 Tys[n] = MVT::Other;
8830 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8831 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8832 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8833 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8834 Ops, 2, VLDMemInt->getMemoryVT(),
8835 VLDMemInt->getMemOperand());
8838 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8840 unsigned ResNo = UI.getUse().getResNo();
8841 // Ignore uses of the chain result.
8842 if (ResNo == NumVecs)
8845 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8848 // Now the vldN-lane intrinsic is dead except for its chain result.
8849 // Update uses of the chain.
8850 std::vector<SDValue> VLDDupResults;
8851 for (unsigned n = 0; n < NumVecs; ++n)
8852 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8853 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8854 DCI.CombineTo(VLD, VLDDupResults);
8859 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8860 /// ARMISD::VDUPLANE.
8861 static SDValue PerformVDUPLANECombine(SDNode *N,
8862 TargetLowering::DAGCombinerInfo &DCI) {
8863 SDValue Op = N->getOperand(0);
8865 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8866 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8867 if (CombineVLDDUP(N, DCI))
8868 return SDValue(N, 0);
8870 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8871 // redundant. Ignore bit_converts for now; element sizes are checked below.
8872 while (Op.getOpcode() == ISD::BITCAST)
8873 Op = Op.getOperand(0);
8874 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8877 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8878 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8879 // The canonical VMOV for a zero vector uses a 32-bit element size.
8880 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8882 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8884 EVT VT = N->getValueType(0);
8885 if (EltSize > VT.getVectorElementType().getSizeInBits())
8888 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8891 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8892 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8893 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8897 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8899 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8904 APFloat APF = C->getValueAPF();
8905 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8906 != APFloat::opOK || !isExact)
8909 c0 = (I == 0) ? cN : c0;
8910 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8917 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8918 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8919 /// when the VMUL has a constant operand that is a power of 2.
8921 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8922 /// vmul.f32 d16, d17, d16
8923 /// vcvt.s32.f32 d16, d16
8925 /// vcvt.s32.f32 d16, d16, #3
8926 static SDValue PerformVCVTCombine(SDNode *N,
8927 TargetLowering::DAGCombinerInfo &DCI,
8928 const ARMSubtarget *Subtarget) {
8929 SelectionDAG &DAG = DCI.DAG;
8930 SDValue Op = N->getOperand(0);
8932 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8933 Op.getOpcode() != ISD::FMUL)
8937 SDValue N0 = Op->getOperand(0);
8938 SDValue ConstVec = Op->getOperand(1);
8939 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8941 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8942 !isConstVecPow2(ConstVec, isSigned, C))
8945 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8946 Intrinsic::arm_neon_vcvtfp2fxu;
8947 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8949 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8950 DAG.getConstant(Log2_64(C), MVT::i32));
8953 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8954 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8955 /// when the VDIV has a constant operand that is a power of 2.
8957 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8958 /// vcvt.f32.s32 d16, d16
8959 /// vdiv.f32 d16, d17, d16
8961 /// vcvt.f32.s32 d16, d16, #3
8962 static SDValue PerformVDIVCombine(SDNode *N,
8963 TargetLowering::DAGCombinerInfo &DCI,
8964 const ARMSubtarget *Subtarget) {
8965 SelectionDAG &DAG = DCI.DAG;
8966 SDValue Op = N->getOperand(0);
8967 unsigned OpOpcode = Op.getNode()->getOpcode();
8969 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8970 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8974 SDValue ConstVec = N->getOperand(1);
8975 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8977 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8978 !isConstVecPow2(ConstVec, isSigned, C))
8981 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8982 Intrinsic::arm_neon_vcvtfxu2fp;
8983 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8985 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8986 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8989 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8990 /// operand of a vector shift operation, where all the elements of the
8991 /// build_vector must have the same constant integer value.
8992 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8993 // Ignore bit_converts.
8994 while (Op.getOpcode() == ISD::BITCAST)
8995 Op = Op.getOperand(0);
8996 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8997 APInt SplatBits, SplatUndef;
8998 unsigned SplatBitSize;
9000 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9001 HasAnyUndefs, ElementBits) ||
9002 SplatBitSize > ElementBits)
9004 Cnt = SplatBits.getSExtValue();
9008 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9009 /// operand of a vector shift left operation. That value must be in the range:
9010 /// 0 <= Value < ElementBits for a left shift; or
9011 /// 0 <= Value <= ElementBits for a long left shift.
9012 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9013 assert(VT.isVector() && "vector shift count is not a vector type");
9014 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9015 if (! getVShiftImm(Op, ElementBits, Cnt))
9017 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9020 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9021 /// operand of a vector shift right operation. For a shift opcode, the value
9022 /// is positive, but for an intrinsic the value count must be negative. The
9023 /// absolute value must be in the range:
9024 /// 1 <= |Value| <= ElementBits for a right shift; or
9025 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9026 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9028 assert(VT.isVector() && "vector shift count is not a vector type");
9029 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9030 if (! getVShiftImm(Op, ElementBits, Cnt))
9034 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9037 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9038 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9039 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9042 // Don't do anything for most intrinsics.
9045 // Vector shifts: check for immediate versions and lower them.
9046 // Note: This is done during DAG combining instead of DAG legalizing because
9047 // the build_vectors for 64-bit vector element shift counts are generally
9048 // not legal, and it is hard to see their values after they get legalized to
9049 // loads from a constant pool.
9050 case Intrinsic::arm_neon_vshifts:
9051 case Intrinsic::arm_neon_vshiftu:
9052 case Intrinsic::arm_neon_vshiftls:
9053 case Intrinsic::arm_neon_vshiftlu:
9054 case Intrinsic::arm_neon_vshiftn:
9055 case Intrinsic::arm_neon_vrshifts:
9056 case Intrinsic::arm_neon_vrshiftu:
9057 case Intrinsic::arm_neon_vrshiftn:
9058 case Intrinsic::arm_neon_vqshifts:
9059 case Intrinsic::arm_neon_vqshiftu:
9060 case Intrinsic::arm_neon_vqshiftsu:
9061 case Intrinsic::arm_neon_vqshiftns:
9062 case Intrinsic::arm_neon_vqshiftnu:
9063 case Intrinsic::arm_neon_vqshiftnsu:
9064 case Intrinsic::arm_neon_vqrshiftns:
9065 case Intrinsic::arm_neon_vqrshiftnu:
9066 case Intrinsic::arm_neon_vqrshiftnsu: {
9067 EVT VT = N->getOperand(1).getValueType();
9069 unsigned VShiftOpc = 0;
9072 case Intrinsic::arm_neon_vshifts:
9073 case Intrinsic::arm_neon_vshiftu:
9074 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9075 VShiftOpc = ARMISD::VSHL;
9078 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9079 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9080 ARMISD::VSHRs : ARMISD::VSHRu);
9085 case Intrinsic::arm_neon_vshiftls:
9086 case Intrinsic::arm_neon_vshiftlu:
9087 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9089 llvm_unreachable("invalid shift count for vshll intrinsic");
9091 case Intrinsic::arm_neon_vrshifts:
9092 case Intrinsic::arm_neon_vrshiftu:
9093 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9097 case Intrinsic::arm_neon_vqshifts:
9098 case Intrinsic::arm_neon_vqshiftu:
9099 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9103 case Intrinsic::arm_neon_vqshiftsu:
9104 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9106 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9108 case Intrinsic::arm_neon_vshiftn:
9109 case Intrinsic::arm_neon_vrshiftn:
9110 case Intrinsic::arm_neon_vqshiftns:
9111 case Intrinsic::arm_neon_vqshiftnu:
9112 case Intrinsic::arm_neon_vqshiftnsu:
9113 case Intrinsic::arm_neon_vqrshiftns:
9114 case Intrinsic::arm_neon_vqrshiftnu:
9115 case Intrinsic::arm_neon_vqrshiftnsu:
9116 // Narrowing shifts require an immediate right shift.
9117 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9119 llvm_unreachable("invalid shift count for narrowing vector shift "
9123 llvm_unreachable("unhandled vector shift");
9127 case Intrinsic::arm_neon_vshifts:
9128 case Intrinsic::arm_neon_vshiftu:
9129 // Opcode already set above.
9131 case Intrinsic::arm_neon_vshiftls:
9132 case Intrinsic::arm_neon_vshiftlu:
9133 if (Cnt == VT.getVectorElementType().getSizeInBits())
9134 VShiftOpc = ARMISD::VSHLLi;
9136 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9137 ARMISD::VSHLLs : ARMISD::VSHLLu);
9139 case Intrinsic::arm_neon_vshiftn:
9140 VShiftOpc = ARMISD::VSHRN; break;
9141 case Intrinsic::arm_neon_vrshifts:
9142 VShiftOpc = ARMISD::VRSHRs; break;
9143 case Intrinsic::arm_neon_vrshiftu:
9144 VShiftOpc = ARMISD::VRSHRu; break;
9145 case Intrinsic::arm_neon_vrshiftn:
9146 VShiftOpc = ARMISD::VRSHRN; break;
9147 case Intrinsic::arm_neon_vqshifts:
9148 VShiftOpc = ARMISD::VQSHLs; break;
9149 case Intrinsic::arm_neon_vqshiftu:
9150 VShiftOpc = ARMISD::VQSHLu; break;
9151 case Intrinsic::arm_neon_vqshiftsu:
9152 VShiftOpc = ARMISD::VQSHLsu; break;
9153 case Intrinsic::arm_neon_vqshiftns:
9154 VShiftOpc = ARMISD::VQSHRNs; break;
9155 case Intrinsic::arm_neon_vqshiftnu:
9156 VShiftOpc = ARMISD::VQSHRNu; break;
9157 case Intrinsic::arm_neon_vqshiftnsu:
9158 VShiftOpc = ARMISD::VQSHRNsu; break;
9159 case Intrinsic::arm_neon_vqrshiftns:
9160 VShiftOpc = ARMISD::VQRSHRNs; break;
9161 case Intrinsic::arm_neon_vqrshiftnu:
9162 VShiftOpc = ARMISD::VQRSHRNu; break;
9163 case Intrinsic::arm_neon_vqrshiftnsu:
9164 VShiftOpc = ARMISD::VQRSHRNsu; break;
9167 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9168 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9171 case Intrinsic::arm_neon_vshiftins: {
9172 EVT VT = N->getOperand(1).getValueType();
9174 unsigned VShiftOpc = 0;
9176 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9177 VShiftOpc = ARMISD::VSLI;
9178 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9179 VShiftOpc = ARMISD::VSRI;
9181 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9184 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9185 N->getOperand(1), N->getOperand(2),
9186 DAG.getConstant(Cnt, MVT::i32));
9189 case Intrinsic::arm_neon_vqrshifts:
9190 case Intrinsic::arm_neon_vqrshiftu:
9191 // No immediate versions of these to check for.
9198 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9199 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9200 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9201 /// vector element shift counts are generally not legal, and it is hard to see
9202 /// their values after they get legalized to loads from a constant pool.
9203 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9204 const ARMSubtarget *ST) {
9205 EVT VT = N->getValueType(0);
9206 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9207 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9208 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9209 SDValue N1 = N->getOperand(1);
9210 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9211 SDValue N0 = N->getOperand(0);
9212 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9213 DAG.MaskedValueIsZero(N0.getOperand(0),
9214 APInt::getHighBitsSet(32, 16)))
9215 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
9219 // Nothing to be done for scalar shifts.
9220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9221 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9224 assert(ST->hasNEON() && "unexpected vector shift");
9227 switch (N->getOpcode()) {
9228 default: llvm_unreachable("unexpected shift opcode");
9231 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9232 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
9233 DAG.getConstant(Cnt, MVT::i32));
9238 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9239 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9240 ARMISD::VSHRs : ARMISD::VSHRu);
9241 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
9242 DAG.getConstant(Cnt, MVT::i32));
9248 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9249 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9250 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9251 const ARMSubtarget *ST) {
9252 SDValue N0 = N->getOperand(0);
9254 // Check for sign- and zero-extensions of vector extract operations of 8-
9255 // and 16-bit vector elements. NEON supports these directly. They are
9256 // handled during DAG combining because type legalization will promote them
9257 // to 32-bit types and it is messy to recognize the operations after that.
9258 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9259 SDValue Vec = N0.getOperand(0);
9260 SDValue Lane = N0.getOperand(1);
9261 EVT VT = N->getValueType(0);
9262 EVT EltVT = N0.getValueType();
9263 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9265 if (VT == MVT::i32 &&
9266 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9267 TLI.isTypeLegal(Vec.getValueType()) &&
9268 isa<ConstantSDNode>(Lane)) {
9271 switch (N->getOpcode()) {
9272 default: llvm_unreachable("unexpected opcode");
9273 case ISD::SIGN_EXTEND:
9274 Opc = ARMISD::VGETLANEs;
9276 case ISD::ZERO_EXTEND:
9277 case ISD::ANY_EXTEND:
9278 Opc = ARMISD::VGETLANEu;
9281 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
9288 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9289 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9290 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9291 const ARMSubtarget *ST) {
9292 // If the target supports NEON, try to use vmax/vmin instructions for f32
9293 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9294 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9295 // a NaN; only do the transformation when it matches that behavior.
9297 // For now only do this when using NEON for FP operations; if using VFP, it
9298 // is not obvious that the benefit outweighs the cost of switching to the
9300 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9301 N->getValueType(0) != MVT::f32)
9304 SDValue CondLHS = N->getOperand(0);
9305 SDValue CondRHS = N->getOperand(1);
9306 SDValue LHS = N->getOperand(2);
9307 SDValue RHS = N->getOperand(3);
9308 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9310 unsigned Opcode = 0;
9312 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9313 IsReversed = false; // x CC y ? x : y
9314 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9315 IsReversed = true ; // x CC y ? y : x
9329 // If LHS is NaN, an ordered comparison will be false and the result will
9330 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9331 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9332 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9333 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9335 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9336 // will return -0, so vmin can only be used for unsafe math or if one of
9337 // the operands is known to be nonzero.
9338 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9339 !DAG.getTarget().Options.UnsafeFPMath &&
9340 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9342 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9351 // If LHS is NaN, an ordered comparison will be false and the result will
9352 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9353 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9354 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9355 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9357 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9358 // will return +0, so vmax can only be used for unsafe math or if one of
9359 // the operands is known to be nonzero.
9360 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9361 !DAG.getTarget().Options.UnsafeFPMath &&
9362 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9364 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9370 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
9373 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9375 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9376 SDValue Cmp = N->getOperand(4);
9377 if (Cmp.getOpcode() != ARMISD::CMPZ)
9378 // Only looking at EQ and NE cases.
9381 EVT VT = N->getValueType(0);
9382 DebugLoc dl = N->getDebugLoc();
9383 SDValue LHS = Cmp.getOperand(0);
9384 SDValue RHS = Cmp.getOperand(1);
9385 SDValue FalseVal = N->getOperand(0);
9386 SDValue TrueVal = N->getOperand(1);
9387 SDValue ARMcc = N->getOperand(2);
9388 ARMCC::CondCodes CC =
9389 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9407 /// FIXME: Turn this into a target neutral optimization?
9409 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9410 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9411 N->getOperand(3), Cmp);
9412 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9414 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9415 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9416 N->getOperand(3), NewCmp);
9419 if (Res.getNode()) {
9420 APInt KnownZero, KnownOne;
9421 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9422 // Capture demanded bits information that would be otherwise lost.
9423 if (KnownZero == 0xfffffffe)
9424 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9425 DAG.getValueType(MVT::i1));
9426 else if (KnownZero == 0xffffff00)
9427 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9428 DAG.getValueType(MVT::i8));
9429 else if (KnownZero == 0xffff0000)
9430 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9431 DAG.getValueType(MVT::i16));
9437 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9438 DAGCombinerInfo &DCI) const {
9439 switch (N->getOpcode()) {
9441 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9442 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9443 case ISD::SUB: return PerformSUBCombine(N, DCI);
9444 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9445 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9446 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9447 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9448 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9449 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9450 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9451 case ISD::STORE: return PerformSTORECombine(N, DCI);
9452 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9453 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9454 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9455 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9456 case ISD::FP_TO_SINT:
9457 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9458 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9459 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9462 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9463 case ISD::SIGN_EXTEND:
9464 case ISD::ZERO_EXTEND:
9465 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9466 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9467 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9468 case ARMISD::VLD2DUP:
9469 case ARMISD::VLD3DUP:
9470 case ARMISD::VLD4DUP:
9471 return CombineBaseUpdate(N, DCI);
9472 case ISD::INTRINSIC_VOID:
9473 case ISD::INTRINSIC_W_CHAIN:
9474 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9475 case Intrinsic::arm_neon_vld1:
9476 case Intrinsic::arm_neon_vld2:
9477 case Intrinsic::arm_neon_vld3:
9478 case Intrinsic::arm_neon_vld4:
9479 case Intrinsic::arm_neon_vld2lane:
9480 case Intrinsic::arm_neon_vld3lane:
9481 case Intrinsic::arm_neon_vld4lane:
9482 case Intrinsic::arm_neon_vst1:
9483 case Intrinsic::arm_neon_vst2:
9484 case Intrinsic::arm_neon_vst3:
9485 case Intrinsic::arm_neon_vst4:
9486 case Intrinsic::arm_neon_vst2lane:
9487 case Intrinsic::arm_neon_vst3lane:
9488 case Intrinsic::arm_neon_vst4lane:
9489 return CombineBaseUpdate(N, DCI);
9497 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9499 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9502 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
9503 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9504 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9506 switch (VT.getSimpleVT().SimpleTy) {
9512 // Unaligned access can use (for example) LRDB, LRDH, LDR
9513 if (AllowsUnaligned) {
9515 *Fast = Subtarget->hasV7Ops();
9522 // For any little-endian targets with neon, we can support unaligned ld/st
9523 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9524 // A big-endian target may also explictly support unaligned accesses
9525 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9535 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9536 unsigned AlignCheck) {
9537 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9538 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9541 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9542 unsigned DstAlign, unsigned SrcAlign,
9543 bool IsMemset, bool ZeroMemset,
9545 MachineFunction &MF) const {
9546 const Function *F = MF.getFunction();
9548 // See if we can use NEON instructions for this...
9549 if ((!IsMemset || ZeroMemset) &&
9550 Subtarget->hasNEON() &&
9551 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9552 Attribute::NoImplicitFloat)) {
9555 (memOpAlign(SrcAlign, DstAlign, 16) ||
9556 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
9558 } else if (Size >= 8 &&
9559 (memOpAlign(SrcAlign, DstAlign, 8) ||
9560 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
9565 // Lowering to i32/i16 if the size permits.
9571 // Let the target-independent logic figure it out.
9575 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9576 if (Val.getOpcode() != ISD::LOAD)
9579 EVT VT1 = Val.getValueType();
9580 if (!VT1.isSimple() || !VT1.isInteger() ||
9581 !VT2.isSimple() || !VT2.isInteger())
9584 switch (VT1.getSimpleVT().SimpleTy) {
9589 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9596 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9601 switch (VT.getSimpleVT().SimpleTy) {
9602 default: return false;
9617 if ((V & (Scale - 1)) != 0)
9620 return V == (V & ((1LL << 5) - 1));
9623 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9624 const ARMSubtarget *Subtarget) {
9631 switch (VT.getSimpleVT().SimpleTy) {
9632 default: return false;
9637 // + imm12 or - imm8
9639 return V == (V & ((1LL << 8) - 1));
9640 return V == (V & ((1LL << 12) - 1));
9643 // Same as ARM mode. FIXME: NEON?
9644 if (!Subtarget->hasVFP2())
9649 return V == (V & ((1LL << 8) - 1));
9653 /// isLegalAddressImmediate - Return true if the integer value can be used
9654 /// as the offset of the target addressing mode for load / store of the
9656 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9657 const ARMSubtarget *Subtarget) {
9664 if (Subtarget->isThumb1Only())
9665 return isLegalT1AddressImmediate(V, VT);
9666 else if (Subtarget->isThumb2())
9667 return isLegalT2AddressImmediate(V, VT, Subtarget);
9672 switch (VT.getSimpleVT().SimpleTy) {
9673 default: return false;
9678 return V == (V & ((1LL << 12) - 1));
9681 return V == (V & ((1LL << 8) - 1));
9684 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9689 return V == (V & ((1LL << 8) - 1));
9693 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9695 int Scale = AM.Scale;
9699 switch (VT.getSimpleVT().SimpleTy) {
9700 default: return false;
9709 return Scale == 2 || Scale == 4 || Scale == 8;
9712 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9716 // Note, we allow "void" uses (basically, uses that aren't loads or
9717 // stores), because arm allows folding a scale into many arithmetic
9718 // operations. This should be made more precise and revisited later.
9720 // Allow r << imm, but the imm has to be a multiple of two.
9721 if (Scale & 1) return false;
9722 return isPowerOf2_32(Scale);
9726 /// isLegalAddressingMode - Return true if the addressing mode represented
9727 /// by AM is legal for this target, for a load/store of the specified type.
9728 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9730 EVT VT = getValueType(Ty, true);
9731 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9734 // Can never fold addr of global into load/store.
9739 case 0: // no scale reg, must be "r+i" or "r", or "i".
9742 if (Subtarget->isThumb1Only())
9746 // ARM doesn't support any R+R*scale+imm addr modes.
9753 if (Subtarget->isThumb2())
9754 return isLegalT2ScaledAddressingMode(AM, VT);
9756 int Scale = AM.Scale;
9757 switch (VT.getSimpleVT().SimpleTy) {
9758 default: return false;
9762 if (Scale < 0) Scale = -Scale;
9766 return isPowerOf2_32(Scale & ~1);
9770 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9775 // Note, we allow "void" uses (basically, uses that aren't loads or
9776 // stores), because arm allows folding a scale into many arithmetic
9777 // operations. This should be made more precise and revisited later.
9779 // Allow r << imm, but the imm has to be a multiple of two.
9780 if (Scale & 1) return false;
9781 return isPowerOf2_32(Scale);
9787 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9788 /// icmp immediate, that is the target has icmp instructions which can compare
9789 /// a register against the immediate without having to materialize the
9790 /// immediate into a register.
9791 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9792 // Thumb2 and ARM modes can use cmn for negative immediates.
9793 if (!Subtarget->isThumb())
9794 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9795 if (Subtarget->isThumb2())
9796 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9797 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9798 return Imm >= 0 && Imm <= 255;
9801 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9802 /// *or sub* immediate, that is the target has add or sub instructions which can
9803 /// add a register with the immediate without having to materialize the
9804 /// immediate into a register.
9805 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9806 // Same encoding for add/sub, just flip the sign.
9807 int64_t AbsImm = llvm::abs64(Imm);
9808 if (!Subtarget->isThumb())
9809 return ARM_AM::getSOImmVal(AbsImm) != -1;
9810 if (Subtarget->isThumb2())
9811 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9812 // Thumb1 only has 8-bit unsigned immediate.
9813 return AbsImm >= 0 && AbsImm <= 255;
9816 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9817 bool isSEXTLoad, SDValue &Base,
9818 SDValue &Offset, bool &isInc,
9819 SelectionDAG &DAG) {
9820 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9823 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9825 Base = Ptr->getOperand(0);
9826 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9827 int RHSC = (int)RHS->getZExtValue();
9828 if (RHSC < 0 && RHSC > -256) {
9829 assert(Ptr->getOpcode() == ISD::ADD);
9831 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9835 isInc = (Ptr->getOpcode() == ISD::ADD);
9836 Offset = Ptr->getOperand(1);
9838 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9840 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9841 int RHSC = (int)RHS->getZExtValue();
9842 if (RHSC < 0 && RHSC > -0x1000) {
9843 assert(Ptr->getOpcode() == ISD::ADD);
9845 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9846 Base = Ptr->getOperand(0);
9851 if (Ptr->getOpcode() == ISD::ADD) {
9853 ARM_AM::ShiftOpc ShOpcVal=
9854 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9855 if (ShOpcVal != ARM_AM::no_shift) {
9856 Base = Ptr->getOperand(1);
9857 Offset = Ptr->getOperand(0);
9859 Base = Ptr->getOperand(0);
9860 Offset = Ptr->getOperand(1);
9865 isInc = (Ptr->getOpcode() == ISD::ADD);
9866 Base = Ptr->getOperand(0);
9867 Offset = Ptr->getOperand(1);
9871 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9875 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9876 bool isSEXTLoad, SDValue &Base,
9877 SDValue &Offset, bool &isInc,
9878 SelectionDAG &DAG) {
9879 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9882 Base = Ptr->getOperand(0);
9883 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9884 int RHSC = (int)RHS->getZExtValue();
9885 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9886 assert(Ptr->getOpcode() == ISD::ADD);
9888 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9890 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9891 isInc = Ptr->getOpcode() == ISD::ADD;
9892 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9900 /// getPreIndexedAddressParts - returns true by value, base pointer and
9901 /// offset pointer and addressing mode by reference if the node's address
9902 /// can be legally represented as pre-indexed load / store address.
9904 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9906 ISD::MemIndexedMode &AM,
9907 SelectionDAG &DAG) const {
9908 if (Subtarget->isThumb1Only())
9913 bool isSEXTLoad = false;
9914 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9915 Ptr = LD->getBasePtr();
9916 VT = LD->getMemoryVT();
9917 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9918 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9919 Ptr = ST->getBasePtr();
9920 VT = ST->getMemoryVT();
9925 bool isLegal = false;
9926 if (Subtarget->isThumb2())
9927 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9928 Offset, isInc, DAG);
9930 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9931 Offset, isInc, DAG);
9935 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9939 /// getPostIndexedAddressParts - returns true by value, base pointer and
9940 /// offset pointer and addressing mode by reference if this node can be
9941 /// combined with a load / store to form a post-indexed load / store.
9942 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9945 ISD::MemIndexedMode &AM,
9946 SelectionDAG &DAG) const {
9947 if (Subtarget->isThumb1Only())
9952 bool isSEXTLoad = false;
9953 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9954 VT = LD->getMemoryVT();
9955 Ptr = LD->getBasePtr();
9956 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9957 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9958 VT = ST->getMemoryVT();
9959 Ptr = ST->getBasePtr();
9964 bool isLegal = false;
9965 if (Subtarget->isThumb2())
9966 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9969 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9975 // Swap base ptr and offset to catch more post-index load / store when
9976 // it's legal. In Thumb2 mode, offset must be an immediate.
9977 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9978 !Subtarget->isThumb2())
9979 std::swap(Base, Offset);
9981 // Post-indexed load / store update the base pointer.
9986 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9990 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9993 const SelectionDAG &DAG,
9994 unsigned Depth) const {
9995 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9996 switch (Op.getOpcode()) {
9998 case ARMISD::CMOV: {
9999 // Bits are known zero/one if known on the LHS and RHS.
10000 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10001 if (KnownZero == 0 && KnownOne == 0) return;
10003 APInt KnownZeroRHS, KnownOneRHS;
10004 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10005 KnownZero &= KnownZeroRHS;
10006 KnownOne &= KnownOneRHS;
10012 //===----------------------------------------------------------------------===//
10013 // ARM Inline Assembly Support
10014 //===----------------------------------------------------------------------===//
10016 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10017 // Looking for "rev" which is V6+.
10018 if (!Subtarget->hasV6Ops())
10021 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10022 std::string AsmStr = IA->getAsmString();
10023 SmallVector<StringRef, 4> AsmPieces;
10024 SplitString(AsmStr, AsmPieces, ";\n");
10026 switch (AsmPieces.size()) {
10027 default: return false;
10029 AsmStr = AsmPieces[0];
10031 SplitString(AsmStr, AsmPieces, " \t,");
10034 if (AsmPieces.size() == 3 &&
10035 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10036 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10037 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10038 if (Ty && Ty->getBitWidth() == 32)
10039 return IntrinsicLowering::LowerToByteSwap(CI);
10047 /// getConstraintType - Given a constraint letter, return the type of
10048 /// constraint it is for this target.
10049 ARMTargetLowering::ConstraintType
10050 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10051 if (Constraint.size() == 1) {
10052 switch (Constraint[0]) {
10054 case 'l': return C_RegisterClass;
10055 case 'w': return C_RegisterClass;
10056 case 'h': return C_RegisterClass;
10057 case 'x': return C_RegisterClass;
10058 case 't': return C_RegisterClass;
10059 case 'j': return C_Other; // Constant for movw.
10060 // An address with a single base register. Due to the way we
10061 // currently handle addresses it is the same as an 'r' memory constraint.
10062 case 'Q': return C_Memory;
10064 } else if (Constraint.size() == 2) {
10065 switch (Constraint[0]) {
10067 // All 'U+' constraints are addresses.
10068 case 'U': return C_Memory;
10071 return TargetLowering::getConstraintType(Constraint);
10074 /// Examine constraint type and operand type and determine a weight value.
10075 /// This object must already have been set up with the operand type
10076 /// and the current alternative constraint selected.
10077 TargetLowering::ConstraintWeight
10078 ARMTargetLowering::getSingleConstraintMatchWeight(
10079 AsmOperandInfo &info, const char *constraint) const {
10080 ConstraintWeight weight = CW_Invalid;
10081 Value *CallOperandVal = info.CallOperandVal;
10082 // If we don't have a value, we can't do a match,
10083 // but allow it at the lowest weight.
10084 if (CallOperandVal == NULL)
10086 Type *type = CallOperandVal->getType();
10087 // Look at the constraint type.
10088 switch (*constraint) {
10090 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10093 if (type->isIntegerTy()) {
10094 if (Subtarget->isThumb())
10095 weight = CW_SpecificReg;
10097 weight = CW_Register;
10101 if (type->isFloatingPointTy())
10102 weight = CW_Register;
10108 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10110 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10112 if (Constraint.size() == 1) {
10113 // GCC ARM Constraint Letters
10114 switch (Constraint[0]) {
10115 case 'l': // Low regs or general regs.
10116 if (Subtarget->isThumb())
10117 return RCPair(0U, &ARM::tGPRRegClass);
10118 return RCPair(0U, &ARM::GPRRegClass);
10119 case 'h': // High regs or no regs.
10120 if (Subtarget->isThumb())
10121 return RCPair(0U, &ARM::hGPRRegClass);
10124 return RCPair(0U, &ARM::GPRRegClass);
10126 if (VT == MVT::f32)
10127 return RCPair(0U, &ARM::SPRRegClass);
10128 if (VT.getSizeInBits() == 64)
10129 return RCPair(0U, &ARM::DPRRegClass);
10130 if (VT.getSizeInBits() == 128)
10131 return RCPair(0U, &ARM::QPRRegClass);
10134 if (VT == MVT::f32)
10135 return RCPair(0U, &ARM::SPR_8RegClass);
10136 if (VT.getSizeInBits() == 64)
10137 return RCPair(0U, &ARM::DPR_8RegClass);
10138 if (VT.getSizeInBits() == 128)
10139 return RCPair(0U, &ARM::QPR_8RegClass);
10142 if (VT == MVT::f32)
10143 return RCPair(0U, &ARM::SPRRegClass);
10147 if (StringRef("{cc}").equals_lower(Constraint))
10148 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10150 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10153 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10154 /// vector. If it is invalid, don't add anything to Ops.
10155 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10156 std::string &Constraint,
10157 std::vector<SDValue>&Ops,
10158 SelectionDAG &DAG) const {
10159 SDValue Result(0, 0);
10161 // Currently only support length 1 constraints.
10162 if (Constraint.length() != 1) return;
10164 char ConstraintLetter = Constraint[0];
10165 switch (ConstraintLetter) {
10168 case 'I': case 'J': case 'K': case 'L':
10169 case 'M': case 'N': case 'O':
10170 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10174 int64_t CVal64 = C->getSExtValue();
10175 int CVal = (int) CVal64;
10176 // None of these constraints allow values larger than 32 bits. Check
10177 // that the value fits in an int.
10178 if (CVal != CVal64)
10181 switch (ConstraintLetter) {
10183 // Constant suitable for movw, must be between 0 and
10185 if (Subtarget->hasV6T2Ops())
10186 if (CVal >= 0 && CVal <= 65535)
10190 if (Subtarget->isThumb1Only()) {
10191 // This must be a constant between 0 and 255, for ADD
10193 if (CVal >= 0 && CVal <= 255)
10195 } else if (Subtarget->isThumb2()) {
10196 // A constant that can be used as an immediate value in a
10197 // data-processing instruction.
10198 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10201 // A constant that can be used as an immediate value in a
10202 // data-processing instruction.
10203 if (ARM_AM::getSOImmVal(CVal) != -1)
10209 if (Subtarget->isThumb()) { // FIXME thumb2
10210 // This must be a constant between -255 and -1, for negated ADD
10211 // immediates. This can be used in GCC with an "n" modifier that
10212 // prints the negated value, for use with SUB instructions. It is
10213 // not useful otherwise but is implemented for compatibility.
10214 if (CVal >= -255 && CVal <= -1)
10217 // This must be a constant between -4095 and 4095. It is not clear
10218 // what this constraint is intended for. Implemented for
10219 // compatibility with GCC.
10220 if (CVal >= -4095 && CVal <= 4095)
10226 if (Subtarget->isThumb1Only()) {
10227 // A 32-bit value where only one byte has a nonzero value. Exclude
10228 // zero to match GCC. This constraint is used by GCC internally for
10229 // constants that can be loaded with a move/shift combination.
10230 // It is not useful otherwise but is implemented for compatibility.
10231 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10233 } else if (Subtarget->isThumb2()) {
10234 // A constant whose bitwise inverse can be used as an immediate
10235 // value in a data-processing instruction. This can be used in GCC
10236 // with a "B" modifier that prints the inverted value, for use with
10237 // BIC and MVN instructions. It is not useful otherwise but is
10238 // implemented for compatibility.
10239 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10242 // A constant whose bitwise inverse can be used as an immediate
10243 // value in a data-processing instruction. This can be used in GCC
10244 // with a "B" modifier that prints the inverted value, for use with
10245 // BIC and MVN instructions. It is not useful otherwise but is
10246 // implemented for compatibility.
10247 if (ARM_AM::getSOImmVal(~CVal) != -1)
10253 if (Subtarget->isThumb1Only()) {
10254 // This must be a constant between -7 and 7,
10255 // for 3-operand ADD/SUB immediate instructions.
10256 if (CVal >= -7 && CVal < 7)
10258 } else if (Subtarget->isThumb2()) {
10259 // A constant whose negation can be used as an immediate value in a
10260 // data-processing instruction. This can be used in GCC with an "n"
10261 // modifier that prints the negated value, for use with SUB
10262 // instructions. It is not useful otherwise but is implemented for
10264 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10267 // A constant whose negation can be used as an immediate value in a
10268 // data-processing instruction. This can be used in GCC with an "n"
10269 // modifier that prints the negated value, for use with SUB
10270 // instructions. It is not useful otherwise but is implemented for
10272 if (ARM_AM::getSOImmVal(-CVal) != -1)
10278 if (Subtarget->isThumb()) { // FIXME thumb2
10279 // This must be a multiple of 4 between 0 and 1020, for
10280 // ADD sp + immediate.
10281 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10284 // A power of two or a constant between 0 and 32. This is used in
10285 // GCC for the shift amount on shifted register operands, but it is
10286 // useful in general for any shift amounts.
10287 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10293 if (Subtarget->isThumb()) { // FIXME thumb2
10294 // This must be a constant between 0 and 31, for shift amounts.
10295 if (CVal >= 0 && CVal <= 31)
10301 if (Subtarget->isThumb()) { // FIXME thumb2
10302 // This must be a multiple of 4 between -508 and 508, for
10303 // ADD/SUB sp = sp + immediate.
10304 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10309 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10313 if (Result.getNode()) {
10314 Ops.push_back(Result);
10317 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10321 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10322 // The ARM target isn't yet aware of offsets.
10326 bool ARM::isBitFieldInvertedMask(unsigned v) {
10327 if (v == 0xffffffff)
10329 // there can be 1's on either or both "outsides", all the "inside"
10330 // bits must be 0's
10331 unsigned int lsb = 0, msb = 31;
10332 while (v & (1 << msb)) --msb;
10333 while (v & (1 << lsb)) ++lsb;
10334 for (unsigned int i = lsb; i <= msb; ++i) {
10341 /// isFPImmLegal - Returns true if the target can instruction select the
10342 /// specified FP immediate natively. If false, the legalizer will
10343 /// materialize the FP immediate as a load from a constant pool.
10344 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10345 if (!Subtarget->hasVFP3())
10347 if (VT == MVT::f32)
10348 return ARM_AM::getFP32Imm(Imm) != -1;
10349 if (VT == MVT::f64)
10350 return ARM_AM::getFP64Imm(Imm) != -1;
10354 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10355 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10356 /// specified in the intrinsic calls.
10357 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10359 unsigned Intrinsic) const {
10360 switch (Intrinsic) {
10361 case Intrinsic::arm_neon_vld1:
10362 case Intrinsic::arm_neon_vld2:
10363 case Intrinsic::arm_neon_vld3:
10364 case Intrinsic::arm_neon_vld4:
10365 case Intrinsic::arm_neon_vld2lane:
10366 case Intrinsic::arm_neon_vld3lane:
10367 case Intrinsic::arm_neon_vld4lane: {
10368 Info.opc = ISD::INTRINSIC_W_CHAIN;
10369 // Conservatively set memVT to the entire set of vectors loaded.
10370 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10371 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10372 Info.ptrVal = I.getArgOperand(0);
10374 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10375 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10376 Info.vol = false; // volatile loads with NEON intrinsics not supported
10377 Info.readMem = true;
10378 Info.writeMem = false;
10381 case Intrinsic::arm_neon_vst1:
10382 case Intrinsic::arm_neon_vst2:
10383 case Intrinsic::arm_neon_vst3:
10384 case Intrinsic::arm_neon_vst4:
10385 case Intrinsic::arm_neon_vst2lane:
10386 case Intrinsic::arm_neon_vst3lane:
10387 case Intrinsic::arm_neon_vst4lane: {
10388 Info.opc = ISD::INTRINSIC_VOID;
10389 // Conservatively set memVT to the entire set of vectors stored.
10390 unsigned NumElts = 0;
10391 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10392 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10393 if (!ArgTy->isVectorTy())
10395 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10397 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10398 Info.ptrVal = I.getArgOperand(0);
10400 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10401 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10402 Info.vol = false; // volatile stores with NEON intrinsics not supported
10403 Info.readMem = false;
10404 Info.writeMem = true;
10407 case Intrinsic::arm_strexd: {
10408 Info.opc = ISD::INTRINSIC_W_CHAIN;
10409 Info.memVT = MVT::i64;
10410 Info.ptrVal = I.getArgOperand(2);
10414 Info.readMem = false;
10415 Info.writeMem = true;
10418 case Intrinsic::arm_ldrexd: {
10419 Info.opc = ISD::INTRINSIC_W_CHAIN;
10420 Info.memVT = MVT::i64;
10421 Info.ptrVal = I.getArgOperand(0);
10425 Info.readMem = true;
10426 Info.writeMem = false;