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/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/IntrinsicLowering.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.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);
508 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
509 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
510 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
518 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
520 // Neon does not support some operations on v1i64 and v2i64 types.
521 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
522 // Custom handling for some quad-vector types to detect VMULL.
523 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
524 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
525 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
526 // Custom handling for some vector types to avoid expensive expansions
527 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
528 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
529 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
532 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
533 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
534 // a destination type that is wider than the source, and nor does
535 // it have a FP_TO_[SU]INT instruction with a narrower destination than
537 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
538 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
539 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
540 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
542 setTargetDAGCombine(ISD::INTRINSIC_VOID);
543 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
544 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
545 setTargetDAGCombine(ISD::SHL);
546 setTargetDAGCombine(ISD::SRL);
547 setTargetDAGCombine(ISD::SRA);
548 setTargetDAGCombine(ISD::SIGN_EXTEND);
549 setTargetDAGCombine(ISD::ZERO_EXTEND);
550 setTargetDAGCombine(ISD::ANY_EXTEND);
551 setTargetDAGCombine(ISD::SELECT_CC);
552 setTargetDAGCombine(ISD::BUILD_VECTOR);
553 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
554 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
555 setTargetDAGCombine(ISD::STORE);
556 setTargetDAGCombine(ISD::FP_TO_SINT);
557 setTargetDAGCombine(ISD::FP_TO_UINT);
558 setTargetDAGCombine(ISD::FDIV);
560 // It is legal to extload from v4i8 to v4i16 or v4i32.
561 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
562 MVT::v4i16, MVT::v2i16,
564 for (unsigned i = 0; i < 6; ++i) {
565 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
566 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
567 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
571 // ARM and Thumb2 support UMLAL/SMLAL.
572 if (!Subtarget->isThumb1Only())
573 setTargetDAGCombine(ISD::ADDC);
576 computeRegisterProperties();
578 // ARM does not have f32 extending load.
579 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
581 // ARM does not have i1 sign extending load.
582 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
584 // ARM supports all 4 flavors of integer indexed load / store.
585 if (!Subtarget->isThumb1Only()) {
586 for (unsigned im = (unsigned)ISD::PRE_INC;
587 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
588 setIndexedLoadAction(im, MVT::i1, Legal);
589 setIndexedLoadAction(im, MVT::i8, Legal);
590 setIndexedLoadAction(im, MVT::i16, Legal);
591 setIndexedLoadAction(im, MVT::i32, Legal);
592 setIndexedStoreAction(im, MVT::i1, Legal);
593 setIndexedStoreAction(im, MVT::i8, Legal);
594 setIndexedStoreAction(im, MVT::i16, Legal);
595 setIndexedStoreAction(im, MVT::i32, Legal);
599 // i64 operation support.
600 setOperationAction(ISD::MUL, MVT::i64, Expand);
601 setOperationAction(ISD::MULHU, MVT::i32, Expand);
602 if (Subtarget->isThumb1Only()) {
603 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
604 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
606 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
607 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
608 setOperationAction(ISD::MULHS, MVT::i32, Expand);
610 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
612 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
613 setOperationAction(ISD::SRL, MVT::i64, Custom);
614 setOperationAction(ISD::SRA, MVT::i64, Custom);
616 if (!Subtarget->isThumb1Only()) {
617 // FIXME: We should do this for Thumb1 as well.
618 setOperationAction(ISD::ADDC, MVT::i32, Custom);
619 setOperationAction(ISD::ADDE, MVT::i32, Custom);
620 setOperationAction(ISD::SUBC, MVT::i32, Custom);
621 setOperationAction(ISD::SUBE, MVT::i32, Custom);
624 // ARM does not have ROTL.
625 setOperationAction(ISD::ROTL, MVT::i32, Expand);
626 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
627 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
628 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
629 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
631 // These just redirect to CTTZ and CTLZ on ARM.
632 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
633 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
635 // Only ARMv6 has BSWAP.
636 if (!Subtarget->hasV6Ops())
637 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
639 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
640 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
641 // These are expanded into libcalls if the cpu doesn't have HW divider.
642 setOperationAction(ISD::SDIV, MVT::i32, Expand);
643 setOperationAction(ISD::UDIV, MVT::i32, Expand);
645 setOperationAction(ISD::SREM, MVT::i32, Expand);
646 setOperationAction(ISD::UREM, MVT::i32, Expand);
647 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
648 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
650 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
651 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
652 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
653 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
654 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
656 setOperationAction(ISD::TRAP, MVT::Other, Legal);
658 // Use the default implementation.
659 setOperationAction(ISD::VASTART, MVT::Other, Custom);
660 setOperationAction(ISD::VAARG, MVT::Other, Expand);
661 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
662 setOperationAction(ISD::VAEND, MVT::Other, Expand);
663 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
664 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
666 if (!Subtarget->isTargetDarwin()) {
667 // Non-Darwin platforms may return values in these registers via the
668 // personality function.
669 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
670 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
671 setExceptionPointerRegister(ARM::R0);
672 setExceptionSelectorRegister(ARM::R1);
675 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
676 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
677 // the default expansion.
678 // FIXME: This should be checking for v6k, not just v6.
679 if (Subtarget->hasDataBarrier() ||
680 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
681 // membarrier needs custom lowering; the rest are legal and handled
683 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
684 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
685 // Custom lowering for 64-bit ops
686 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
691 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
692 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
693 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
694 setInsertFencesForAtomic(true);
696 // Set them all for expansion, which will force libcalls.
697 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
698 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
699 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
711 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
712 // Unordered/Monotonic case.
713 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
714 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
715 // Since the libcalls include locking, fold in the fences
716 setShouldFoldAtomicFences(true);
719 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
721 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
722 if (!Subtarget->hasV6Ops()) {
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
726 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
728 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
729 !Subtarget->isThumb1Only()) {
730 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
731 // iff target supports vfp2.
732 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
733 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
736 // We want to custom lower some of our intrinsics.
737 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
738 if (Subtarget->isTargetDarwin()) {
739 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
740 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
741 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
744 setOperationAction(ISD::SETCC, MVT::i32, Expand);
745 setOperationAction(ISD::SETCC, MVT::f32, Expand);
746 setOperationAction(ISD::SETCC, MVT::f64, Expand);
747 setOperationAction(ISD::SELECT, MVT::i32, Custom);
748 setOperationAction(ISD::SELECT, MVT::f32, Custom);
749 setOperationAction(ISD::SELECT, MVT::f64, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
751 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
752 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
754 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
755 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
756 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
757 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
758 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
760 // We don't support sin/cos/fmod/copysign/pow
761 setOperationAction(ISD::FSIN, MVT::f64, Expand);
762 setOperationAction(ISD::FSIN, MVT::f32, Expand);
763 setOperationAction(ISD::FCOS, MVT::f32, Expand);
764 setOperationAction(ISD::FCOS, MVT::f64, Expand);
765 setOperationAction(ISD::FREM, MVT::f64, Expand);
766 setOperationAction(ISD::FREM, MVT::f32, Expand);
767 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
768 !Subtarget->isThumb1Only()) {
769 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
770 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
772 setOperationAction(ISD::FPOW, MVT::f64, Expand);
773 setOperationAction(ISD::FPOW, MVT::f32, Expand);
775 if (!Subtarget->hasVFP4()) {
776 setOperationAction(ISD::FMA, MVT::f64, Expand);
777 setOperationAction(ISD::FMA, MVT::f32, Expand);
780 // Various VFP goodness
781 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
782 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
783 if (Subtarget->hasVFP2()) {
784 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
785 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
786 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
787 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
789 // Special handling for half-precision FP.
790 if (!Subtarget->hasFP16()) {
791 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
792 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
796 // We have target-specific dag combine patterns for the following nodes:
797 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
798 setTargetDAGCombine(ISD::ADD);
799 setTargetDAGCombine(ISD::SUB);
800 setTargetDAGCombine(ISD::MUL);
801 setTargetDAGCombine(ISD::AND);
802 setTargetDAGCombine(ISD::OR);
803 setTargetDAGCombine(ISD::XOR);
805 if (Subtarget->hasV6Ops())
806 setTargetDAGCombine(ISD::SRL);
808 setStackPointerRegisterToSaveRestore(ARM::SP);
810 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
811 !Subtarget->hasVFP2())
812 setSchedulingPreference(Sched::RegPressure);
814 setSchedulingPreference(Sched::Hybrid);
816 //// temporary - rewrite interface to use type
817 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
818 maxStoresPerMemset = 16;
819 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
821 // On ARM arguments smaller than 4 bytes are extended, so all arguments
822 // are at least 4 bytes aligned.
823 setMinStackArgumentAlignment(4);
825 benefitFromCodePlacementOpt = true;
827 // Prefer likely predicted branches to selects on out-of-order cores.
828 predictableSelectIsExpensive = Subtarget->isLikeA9();
830 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
833 // FIXME: It might make sense to define the representative register class as the
834 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
835 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
836 // SPR's representative would be DPR_VFP2. This should work well if register
837 // pressure tracking were modified such that a register use would increment the
838 // pressure of the register class's representative and all of it's super
839 // classes' representatives transitively. We have not implemented this because
840 // of the difficulty prior to coalescing of modeling operand register classes
841 // due to the common occurrence of cross class copies and subregister insertions
843 std::pair<const TargetRegisterClass*, uint8_t>
844 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
845 const TargetRegisterClass *RRC = 0;
847 switch (VT.getSimpleVT().SimpleTy) {
849 return TargetLowering::findRepresentativeClass(VT);
850 // Use DPR as representative register class for all floating point
851 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
852 // the cost is 1 for both f32 and f64.
853 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
854 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
855 RRC = &ARM::DPRRegClass;
856 // When NEON is used for SP, only half of the register file is available
857 // because operations that define both SP and DP results will be constrained
858 // to the VFP2 class (D0-D15). We currently model this constraint prior to
859 // coalescing by double-counting the SP regs. See the FIXME above.
860 if (Subtarget->useNEONForSinglePrecisionFP())
863 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
864 case MVT::v4f32: case MVT::v2f64:
865 RRC = &ARM::DPRRegClass;
869 RRC = &ARM::DPRRegClass;
873 RRC = &ARM::DPRRegClass;
877 return std::make_pair(RRC, Cost);
880 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
883 case ARMISD::Wrapper: return "ARMISD::Wrapper";
884 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
885 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
886 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
887 case ARMISD::CALL: return "ARMISD::CALL";
888 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
889 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
890 case ARMISD::tCALL: return "ARMISD::tCALL";
891 case ARMISD::BRCOND: return "ARMISD::BRCOND";
892 case ARMISD::BR_JT: return "ARMISD::BR_JT";
893 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
894 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
895 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
896 case ARMISD::CMP: return "ARMISD::CMP";
897 case ARMISD::CMN: return "ARMISD::CMN";
898 case ARMISD::CMPZ: return "ARMISD::CMPZ";
899 case ARMISD::CMPFP: return "ARMISD::CMPFP";
900 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
901 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
902 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
904 case ARMISD::CMOV: return "ARMISD::CMOV";
906 case ARMISD::RBIT: return "ARMISD::RBIT";
908 case ARMISD::FTOSI: return "ARMISD::FTOSI";
909 case ARMISD::FTOUI: return "ARMISD::FTOUI";
910 case ARMISD::SITOF: return "ARMISD::SITOF";
911 case ARMISD::UITOF: return "ARMISD::UITOF";
913 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
914 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
915 case ARMISD::RRX: return "ARMISD::RRX";
917 case ARMISD::ADDC: return "ARMISD::ADDC";
918 case ARMISD::ADDE: return "ARMISD::ADDE";
919 case ARMISD::SUBC: return "ARMISD::SUBC";
920 case ARMISD::SUBE: return "ARMISD::SUBE";
922 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
923 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
925 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
926 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
928 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
930 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
932 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
934 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
935 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
937 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
939 case ARMISD::VCEQ: return "ARMISD::VCEQ";
940 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
941 case ARMISD::VCGE: return "ARMISD::VCGE";
942 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
943 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
944 case ARMISD::VCGEU: return "ARMISD::VCGEU";
945 case ARMISD::VCGT: return "ARMISD::VCGT";
946 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
947 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
948 case ARMISD::VCGTU: return "ARMISD::VCGTU";
949 case ARMISD::VTST: return "ARMISD::VTST";
951 case ARMISD::VSHL: return "ARMISD::VSHL";
952 case ARMISD::VSHRs: return "ARMISD::VSHRs";
953 case ARMISD::VSHRu: return "ARMISD::VSHRu";
954 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
955 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
956 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
957 case ARMISD::VSHRN: return "ARMISD::VSHRN";
958 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
959 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
960 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
961 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
962 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
963 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
964 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
965 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
966 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
967 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
968 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
969 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
970 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
971 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
972 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
973 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
974 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
975 case ARMISD::VDUP: return "ARMISD::VDUP";
976 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
977 case ARMISD::VEXT: return "ARMISD::VEXT";
978 case ARMISD::VREV64: return "ARMISD::VREV64";
979 case ARMISD::VREV32: return "ARMISD::VREV32";
980 case ARMISD::VREV16: return "ARMISD::VREV16";
981 case ARMISD::VZIP: return "ARMISD::VZIP";
982 case ARMISD::VUZP: return "ARMISD::VUZP";
983 case ARMISD::VTRN: return "ARMISD::VTRN";
984 case ARMISD::VTBL1: return "ARMISD::VTBL1";
985 case ARMISD::VTBL2: return "ARMISD::VTBL2";
986 case ARMISD::VMULLs: return "ARMISD::VMULLs";
987 case ARMISD::VMULLu: return "ARMISD::VMULLu";
988 case ARMISD::UMLAL: return "ARMISD::UMLAL";
989 case ARMISD::SMLAL: return "ARMISD::SMLAL";
990 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
991 case ARMISD::FMAX: return "ARMISD::FMAX";
992 case ARMISD::FMIN: return "ARMISD::FMIN";
993 case ARMISD::BFI: return "ARMISD::BFI";
994 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
995 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
996 case ARMISD::VBSL: return "ARMISD::VBSL";
997 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
998 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
999 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1000 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1001 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1002 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1003 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1004 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1005 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1006 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1007 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1008 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1009 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1010 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1011 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1012 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1013 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1014 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1015 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1016 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1020 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1021 if (!VT.isVector()) return getPointerTy();
1022 return VT.changeVectorElementTypeToInteger();
1025 /// getRegClassFor - Return the register class that should be used for the
1026 /// specified value type.
1027 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1028 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1029 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1030 // load / store 4 to 8 consecutive D registers.
1031 if (Subtarget->hasNEON()) {
1032 if (VT == MVT::v4i64)
1033 return &ARM::QQPRRegClass;
1034 if (VT == MVT::v8i64)
1035 return &ARM::QQQQPRRegClass;
1037 return TargetLowering::getRegClassFor(VT);
1040 // Create a fast isel object.
1042 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1043 const TargetLibraryInfo *libInfo) const {
1044 return ARM::createFastISel(funcInfo, libInfo);
1047 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1048 /// be used for loads / stores from the global.
1049 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1050 return (Subtarget->isThumb1Only() ? 127 : 4095);
1053 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1054 unsigned NumVals = N->getNumValues();
1056 return Sched::RegPressure;
1058 for (unsigned i = 0; i != NumVals; ++i) {
1059 EVT VT = N->getValueType(i);
1060 if (VT == MVT::Glue || VT == MVT::Other)
1062 if (VT.isFloatingPoint() || VT.isVector())
1066 if (!N->isMachineOpcode())
1067 return Sched::RegPressure;
1069 // Load are scheduled for latency even if there instruction itinerary
1070 // is not available.
1071 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1072 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1074 if (MCID.getNumDefs() == 0)
1075 return Sched::RegPressure;
1076 if (!Itins->isEmpty() &&
1077 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1080 return Sched::RegPressure;
1083 //===----------------------------------------------------------------------===//
1085 //===----------------------------------------------------------------------===//
1087 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1088 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1090 default: llvm_unreachable("Unknown condition code!");
1091 case ISD::SETNE: return ARMCC::NE;
1092 case ISD::SETEQ: return ARMCC::EQ;
1093 case ISD::SETGT: return ARMCC::GT;
1094 case ISD::SETGE: return ARMCC::GE;
1095 case ISD::SETLT: return ARMCC::LT;
1096 case ISD::SETLE: return ARMCC::LE;
1097 case ISD::SETUGT: return ARMCC::HI;
1098 case ISD::SETUGE: return ARMCC::HS;
1099 case ISD::SETULT: return ARMCC::LO;
1100 case ISD::SETULE: return ARMCC::LS;
1104 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1105 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1106 ARMCC::CondCodes &CondCode2) {
1107 CondCode2 = ARMCC::AL;
1109 default: llvm_unreachable("Unknown FP condition!");
1111 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1113 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1115 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1116 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1117 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1118 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1119 case ISD::SETO: CondCode = ARMCC::VC; break;
1120 case ISD::SETUO: CondCode = ARMCC::VS; break;
1121 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1122 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1123 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1125 case ISD::SETULT: CondCode = ARMCC::LT; break;
1127 case ISD::SETULE: CondCode = ARMCC::LE; break;
1129 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1133 //===----------------------------------------------------------------------===//
1134 // Calling Convention Implementation
1135 //===----------------------------------------------------------------------===//
1137 #include "ARMGenCallingConv.inc"
1139 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1140 /// given CallingConvention value.
1141 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1143 bool isVarArg) const {
1146 llvm_unreachable("Unsupported calling convention");
1147 case CallingConv::Fast:
1148 if (Subtarget->hasVFP2() && !isVarArg) {
1149 if (!Subtarget->isAAPCS_ABI())
1150 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1151 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1152 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1155 case CallingConv::C: {
1156 // Use target triple & subtarget features to do actual dispatch.
1157 if (!Subtarget->isAAPCS_ABI())
1158 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1159 else if (Subtarget->hasVFP2() &&
1160 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1162 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_AAPCS_VFP:
1167 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1169 case CallingConv::ARM_AAPCS:
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1171 case CallingConv::ARM_APCS:
1172 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1173 case CallingConv::GHC:
1174 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1178 /// LowerCallResult - Lower the result values of a call into the
1179 /// appropriate copies out of appropriate physical registers.
1181 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1182 CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<ISD::InputArg> &Ins,
1184 DebugLoc dl, SelectionDAG &DAG,
1185 SmallVectorImpl<SDValue> &InVals) const {
1187 // Assign locations to each value returned by this call.
1188 SmallVector<CCValAssign, 16> RVLocs;
1189 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1190 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1191 CCInfo.AnalyzeCallResult(Ins,
1192 CCAssignFnForNode(CallConv, /* Return*/ true,
1195 // Copy all of the result registers out of their specified physreg.
1196 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1197 CCValAssign VA = RVLocs[i];
1200 if (VA.needsCustom()) {
1201 // Handle f64 or half of a v2f64.
1202 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1204 Chain = Lo.getValue(1);
1205 InFlag = Lo.getValue(2);
1206 VA = RVLocs[++i]; // skip ahead to next loc
1207 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1209 Chain = Hi.getValue(1);
1210 InFlag = Hi.getValue(2);
1211 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1213 if (VA.getLocVT() == MVT::v2f64) {
1214 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1215 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1216 DAG.getConstant(0, MVT::i32));
1218 VA = RVLocs[++i]; // skip ahead to next loc
1219 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1220 Chain = Lo.getValue(1);
1221 InFlag = Lo.getValue(2);
1222 VA = RVLocs[++i]; // skip ahead to next loc
1223 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1224 Chain = Hi.getValue(1);
1225 InFlag = Hi.getValue(2);
1226 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1227 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1228 DAG.getConstant(1, MVT::i32));
1231 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1233 Chain = Val.getValue(1);
1234 InFlag = Val.getValue(2);
1237 switch (VA.getLocInfo()) {
1238 default: llvm_unreachable("Unknown loc info!");
1239 case CCValAssign::Full: break;
1240 case CCValAssign::BCvt:
1241 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1245 InVals.push_back(Val);
1251 /// LowerMemOpCallTo - Store the argument to the stack.
1253 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1254 SDValue StackPtr, SDValue Arg,
1255 DebugLoc dl, SelectionDAG &DAG,
1256 const CCValAssign &VA,
1257 ISD::ArgFlagsTy Flags) const {
1258 unsigned LocMemOffset = VA.getLocMemOffset();
1259 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1260 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1261 return DAG.getStore(Chain, dl, Arg, PtrOff,
1262 MachinePointerInfo::getStack(LocMemOffset),
1266 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1267 SDValue Chain, SDValue &Arg,
1268 RegsToPassVector &RegsToPass,
1269 CCValAssign &VA, CCValAssign &NextVA,
1271 SmallVector<SDValue, 8> &MemOpChains,
1272 ISD::ArgFlagsTy Flags) const {
1274 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1275 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1276 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1278 if (NextVA.isRegLoc())
1279 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1281 assert(NextVA.isMemLoc());
1282 if (StackPtr.getNode() == 0)
1283 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1285 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1291 /// LowerCall - Lowering a call into a callseq_start <-
1292 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1295 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1296 SmallVectorImpl<SDValue> &InVals) const {
1297 SelectionDAG &DAG = CLI.DAG;
1298 DebugLoc &dl = CLI.DL;
1299 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1300 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1301 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1302 SDValue Chain = CLI.Chain;
1303 SDValue Callee = CLI.Callee;
1304 bool &isTailCall = CLI.IsTailCall;
1305 CallingConv::ID CallConv = CLI.CallConv;
1306 bool doesNotRet = CLI.DoesNotReturn;
1307 bool isVarArg = CLI.IsVarArg;
1309 MachineFunction &MF = DAG.getMachineFunction();
1310 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1311 bool IsSibCall = false;
1312 // Disable tail calls if they're not supported.
1313 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1316 // Check if it's really possible to do a tail call.
1317 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1318 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1319 Outs, OutVals, Ins, DAG);
1320 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1321 // detected sibcalls.
1328 // Analyze operands of the call, assigning locations to each operand.
1329 SmallVector<CCValAssign, 16> ArgLocs;
1330 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1331 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1332 CCInfo.AnalyzeCallOperands(Outs,
1333 CCAssignFnForNode(CallConv, /* Return*/ false,
1336 // Get a count of how many bytes are to be pushed on the stack.
1337 unsigned NumBytes = CCInfo.getNextStackOffset();
1339 // For tail calls, memory operands are available in our caller's stack.
1343 // Adjust the stack pointer for the new arguments...
1344 // These operations are automatically eliminated by the prolog/epilog pass
1346 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1348 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1350 RegsToPassVector RegsToPass;
1351 SmallVector<SDValue, 8> MemOpChains;
1353 // Walk the register/memloc assignments, inserting copies/loads. In the case
1354 // of tail call optimization, arguments are handled later.
1355 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1357 ++i, ++realArgIdx) {
1358 CCValAssign &VA = ArgLocs[i];
1359 SDValue Arg = OutVals[realArgIdx];
1360 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1361 bool isByVal = Flags.isByVal();
1363 // Promote the value if needed.
1364 switch (VA.getLocInfo()) {
1365 default: llvm_unreachable("Unknown loc info!");
1366 case CCValAssign::Full: break;
1367 case CCValAssign::SExt:
1368 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1370 case CCValAssign::ZExt:
1371 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1373 case CCValAssign::AExt:
1374 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1376 case CCValAssign::BCvt:
1377 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1381 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1382 if (VA.needsCustom()) {
1383 if (VA.getLocVT() == MVT::v2f64) {
1384 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1385 DAG.getConstant(0, MVT::i32));
1386 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1387 DAG.getConstant(1, MVT::i32));
1389 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1390 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1392 VA = ArgLocs[++i]; // skip ahead to next loc
1393 if (VA.isRegLoc()) {
1394 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1395 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1397 assert(VA.isMemLoc());
1399 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1400 dl, DAG, VA, Flags));
1403 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1404 StackPtr, MemOpChains, Flags);
1406 } else if (VA.isRegLoc()) {
1407 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1408 } else if (isByVal) {
1409 assert(VA.isMemLoc());
1410 unsigned offset = 0;
1412 // True if this byval aggregate will be split between registers
1414 if (CCInfo.isFirstByValRegValid()) {
1415 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1417 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1418 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1419 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1420 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1421 MachinePointerInfo(),
1422 false, false, false, 0);
1423 MemOpChains.push_back(Load.getValue(1));
1424 RegsToPass.push_back(std::make_pair(j, Load));
1426 offset = ARM::R4 - CCInfo.getFirstByValReg();
1427 CCInfo.clearFirstByValReg();
1430 if (Flags.getByValSize() - 4*offset > 0) {
1431 unsigned LocMemOffset = VA.getLocMemOffset();
1432 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1433 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1435 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1436 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1437 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1439 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1441 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1442 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1443 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1444 Ops, array_lengthof(Ops)));
1446 } else if (!IsSibCall) {
1447 assert(VA.isMemLoc());
1449 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1450 dl, DAG, VA, Flags));
1454 if (!MemOpChains.empty())
1455 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1456 &MemOpChains[0], MemOpChains.size());
1458 // Build a sequence of copy-to-reg nodes chained together with token chain
1459 // and flag operands which copy the outgoing args into the appropriate regs.
1461 // Tail call byval lowering might overwrite argument registers so in case of
1462 // tail call optimization the copies to registers are lowered later.
1464 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1465 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1466 RegsToPass[i].second, InFlag);
1467 InFlag = Chain.getValue(1);
1470 // For tail calls lower the arguments to the 'real' stack slot.
1472 // Force all the incoming stack arguments to be loaded from the stack
1473 // before any new outgoing arguments are stored to the stack, because the
1474 // outgoing stack slots may alias the incoming argument stack slots, and
1475 // the alias isn't otherwise explicit. This is slightly more conservative
1476 // than necessary, because it means that each store effectively depends
1477 // on every argument instead of just those arguments it would clobber.
1479 // Do not flag preceding copytoreg stuff together with the following stuff.
1481 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1482 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1483 RegsToPass[i].second, InFlag);
1484 InFlag = Chain.getValue(1);
1489 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1490 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1491 // node so that legalize doesn't hack it.
1492 bool isDirect = false;
1493 bool isARMFunc = false;
1494 bool isLocalARMFunc = false;
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1497 if (EnableARMLongCalls) {
1498 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1499 && "long-calls with non-static relocation model!");
1500 // Handle a global address or an external symbol. If it's not one of
1501 // those, the target's already in a register, so we don't need to do
1503 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1504 const GlobalValue *GV = G->getGlobal();
1505 // Create a constant pool entry for the callee address
1506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507 ARMConstantPoolValue *CPV =
1508 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1510 // Get the address of the callee into a register
1511 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513 Callee = DAG.getLoad(getPointerTy(), dl,
1514 DAG.getEntryNode(), CPAddr,
1515 MachinePointerInfo::getConstantPool(),
1516 false, false, false, 0);
1517 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1518 const char *Sym = S->getSymbol();
1520 // Create a constant pool entry for the callee address
1521 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1522 ARMConstantPoolValue *CPV =
1523 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1524 ARMPCLabelIndex, 0);
1525 // Get the address of the callee into a register
1526 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1527 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1528 Callee = DAG.getLoad(getPointerTy(), dl,
1529 DAG.getEntryNode(), CPAddr,
1530 MachinePointerInfo::getConstantPool(),
1531 false, false, false, 0);
1533 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1534 const GlobalValue *GV = G->getGlobal();
1536 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1537 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1538 getTargetMachine().getRelocationModel() != Reloc::Static;
1539 isARMFunc = !Subtarget->isThumb() || isStub;
1540 // ARM call to a local ARM function is predicable.
1541 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1542 // tBX takes a register source operand.
1543 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1544 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1545 ARMConstantPoolValue *CPV =
1546 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1547 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1548 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1549 Callee = DAG.getLoad(getPointerTy(), dl,
1550 DAG.getEntryNode(), CPAddr,
1551 MachinePointerInfo::getConstantPool(),
1552 false, false, false, 0);
1553 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1554 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1555 getPointerTy(), Callee, PICLabel);
1557 // On ELF targets for PIC code, direct calls should go through the PLT
1558 unsigned OpFlags = 0;
1559 if (Subtarget->isTargetELF() &&
1560 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1561 OpFlags = ARMII::MO_PLT;
1562 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1564 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1566 bool isStub = Subtarget->isTargetDarwin() &&
1567 getTargetMachine().getRelocationModel() != Reloc::Static;
1568 isARMFunc = !Subtarget->isThumb() || isStub;
1569 // tBX takes a register source operand.
1570 const char *Sym = S->getSymbol();
1571 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1572 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1573 ARMConstantPoolValue *CPV =
1574 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1575 ARMPCLabelIndex, 4);
1576 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1577 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1578 Callee = DAG.getLoad(getPointerTy(), dl,
1579 DAG.getEntryNode(), CPAddr,
1580 MachinePointerInfo::getConstantPool(),
1581 false, false, false, 0);
1582 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1583 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1584 getPointerTy(), Callee, PICLabel);
1586 unsigned OpFlags = 0;
1587 // On ELF targets for PIC code, direct calls should go through the PLT
1588 if (Subtarget->isTargetELF() &&
1589 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1590 OpFlags = ARMII::MO_PLT;
1591 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1595 // FIXME: handle tail calls differently.
1597 if (Subtarget->isThumb()) {
1598 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1599 CallOpc = ARMISD::CALL_NOLINK;
1600 else if (doesNotRet && isDirect && !isARMFunc &&
1601 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1602 // "mov lr, pc; b _foo" to avoid confusing the RSP
1603 CallOpc = ARMISD::CALL_NOLINK;
1605 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1607 if (!isDirect && !Subtarget->hasV5TOps()) {
1608 CallOpc = ARMISD::CALL_NOLINK;
1609 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1610 // "mov lr, pc; b _foo" to avoid confusing the RSP
1611 CallOpc = ARMISD::CALL_NOLINK;
1613 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1616 std::vector<SDValue> Ops;
1617 Ops.push_back(Chain);
1618 Ops.push_back(Callee);
1620 // Add argument registers to the end of the list so that they are known live
1622 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1623 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1624 RegsToPass[i].second.getValueType()));
1626 // Add a register mask operand representing the call-preserved registers.
1627 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1628 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1629 assert(Mask && "Missing call preserved mask for calling convention");
1630 Ops.push_back(DAG.getRegisterMask(Mask));
1632 if (InFlag.getNode())
1633 Ops.push_back(InFlag);
1635 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1637 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1639 // Returns a chain and a flag for retval copy to use.
1640 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1641 InFlag = Chain.getValue(1);
1643 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1644 DAG.getIntPtrConstant(0, true), InFlag);
1646 InFlag = Chain.getValue(1);
1648 // Handle result values, copying them out of physregs into vregs that we
1650 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1654 /// HandleByVal - Every parameter *after* a byval parameter is passed
1655 /// on the stack. Remember the next parameter register to allocate,
1656 /// and then confiscate the rest of the parameter registers to insure
1659 ARMTargetLowering::HandleByVal(
1660 CCState *State, unsigned &size, unsigned Align) const {
1661 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1662 assert((State->getCallOrPrologue() == Prologue ||
1663 State->getCallOrPrologue() == Call) &&
1664 "unhandled ParmContext");
1665 if ((!State->isFirstByValRegValid()) &&
1666 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1667 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1668 unsigned AlignInRegs = Align / 4;
1669 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1670 for (unsigned i = 0; i < Waste; ++i)
1671 reg = State->AllocateReg(GPRArgRegs, 4);
1674 State->setFirstByValReg(reg);
1675 // At a call site, a byval parameter that is split between
1676 // registers and memory needs its size truncated here. In a
1677 // function prologue, such byval parameters are reassembled in
1678 // memory, and are not truncated.
1679 if (State->getCallOrPrologue() == Call) {
1680 unsigned excess = 4 * (ARM::R4 - reg);
1681 assert(size >= excess && "expected larger existing stack allocation");
1686 // Confiscate any remaining parameter registers to preclude their
1687 // assignment to subsequent parameters.
1688 while (State->AllocateReg(GPRArgRegs, 4))
1692 /// MatchingStackOffset - Return true if the given stack call argument is
1693 /// already available in the same position (relatively) of the caller's
1694 /// incoming argument stack.
1696 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1697 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1698 const TargetInstrInfo *TII) {
1699 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1701 if (Arg.getOpcode() == ISD::CopyFromReg) {
1702 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1703 if (!TargetRegisterInfo::isVirtualRegister(VR))
1705 MachineInstr *Def = MRI->getVRegDef(VR);
1708 if (!Flags.isByVal()) {
1709 if (!TII->isLoadFromStackSlot(Def, FI))
1714 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1715 if (Flags.isByVal())
1716 // ByVal argument is passed in as a pointer but it's now being
1717 // dereferenced. e.g.
1718 // define @foo(%struct.X* %A) {
1719 // tail call @bar(%struct.X* byval %A)
1722 SDValue Ptr = Ld->getBasePtr();
1723 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1726 FI = FINode->getIndex();
1730 assert(FI != INT_MAX);
1731 if (!MFI->isFixedObjectIndex(FI))
1733 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1736 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1737 /// for tail call optimization. Targets which want to do tail call
1738 /// optimization should implement this function.
1740 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1741 CallingConv::ID CalleeCC,
1743 bool isCalleeStructRet,
1744 bool isCallerStructRet,
1745 const SmallVectorImpl<ISD::OutputArg> &Outs,
1746 const SmallVectorImpl<SDValue> &OutVals,
1747 const SmallVectorImpl<ISD::InputArg> &Ins,
1748 SelectionDAG& DAG) const {
1749 const Function *CallerF = DAG.getMachineFunction().getFunction();
1750 CallingConv::ID CallerCC = CallerF->getCallingConv();
1751 bool CCMatch = CallerCC == CalleeCC;
1753 // Look for obvious safe cases to perform tail call optimization that do not
1754 // require ABI changes. This is what gcc calls sibcall.
1756 // Do not sibcall optimize vararg calls unless the call site is not passing
1758 if (isVarArg && !Outs.empty())
1761 // Also avoid sibcall optimization if either caller or callee uses struct
1762 // return semantics.
1763 if (isCalleeStructRet || isCallerStructRet)
1766 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1767 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1768 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1769 // support in the assembler and linker to be used. This would need to be
1770 // fixed to fully support tail calls in Thumb1.
1772 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1773 // LR. This means if we need to reload LR, it takes an extra instructions,
1774 // which outweighs the value of the tail call; but here we don't know yet
1775 // whether LR is going to be used. Probably the right approach is to
1776 // generate the tail call here and turn it back into CALL/RET in
1777 // emitEpilogue if LR is used.
1779 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1780 // but we need to make sure there are enough registers; the only valid
1781 // registers are the 4 used for parameters. We don't currently do this
1783 if (Subtarget->isThumb1Only())
1786 // If the calling conventions do not match, then we'd better make sure the
1787 // results are returned in the same way as what the caller expects.
1789 SmallVector<CCValAssign, 16> RVLocs1;
1790 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1791 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1792 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1794 SmallVector<CCValAssign, 16> RVLocs2;
1795 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1796 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1797 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1799 if (RVLocs1.size() != RVLocs2.size())
1801 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1802 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1804 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1806 if (RVLocs1[i].isRegLoc()) {
1807 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1810 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1816 // If Caller's vararg or byval argument has been split between registers and
1817 // stack, do not perform tail call, since part of the argument is in caller's
1819 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1820 getInfo<ARMFunctionInfo>();
1821 if (AFI_Caller->getVarArgsRegSaveSize())
1824 // If the callee takes no arguments then go on to check the results of the
1826 if (!Outs.empty()) {
1827 // Check if stack adjustment is needed. For now, do not do this if any
1828 // argument is passed on the stack.
1829 SmallVector<CCValAssign, 16> ArgLocs;
1830 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1831 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1832 CCInfo.AnalyzeCallOperands(Outs,
1833 CCAssignFnForNode(CalleeCC, false, isVarArg));
1834 if (CCInfo.getNextStackOffset()) {
1835 MachineFunction &MF = DAG.getMachineFunction();
1837 // Check if the arguments are already laid out in the right way as
1838 // the caller's fixed stack objects.
1839 MachineFrameInfo *MFI = MF.getFrameInfo();
1840 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1841 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1842 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1844 ++i, ++realArgIdx) {
1845 CCValAssign &VA = ArgLocs[i];
1846 EVT RegVT = VA.getLocVT();
1847 SDValue Arg = OutVals[realArgIdx];
1848 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1849 if (VA.getLocInfo() == CCValAssign::Indirect)
1851 if (VA.needsCustom()) {
1852 // f64 and vector types are split into multiple registers or
1853 // register/stack-slot combinations. The types will not match
1854 // the registers; give up on memory f64 refs until we figure
1855 // out what to do about this.
1858 if (!ArgLocs[++i].isRegLoc())
1860 if (RegVT == MVT::v2f64) {
1861 if (!ArgLocs[++i].isRegLoc())
1863 if (!ArgLocs[++i].isRegLoc())
1866 } else if (!VA.isRegLoc()) {
1867 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1879 ARMTargetLowering::LowerReturn(SDValue Chain,
1880 CallingConv::ID CallConv, bool isVarArg,
1881 const SmallVectorImpl<ISD::OutputArg> &Outs,
1882 const SmallVectorImpl<SDValue> &OutVals,
1883 DebugLoc dl, SelectionDAG &DAG) const {
1885 // CCValAssign - represent the assignment of the return value to a location.
1886 SmallVector<CCValAssign, 16> RVLocs;
1888 // CCState - Info about the registers and stack slots.
1889 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1890 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1892 // Analyze outgoing return values.
1893 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1896 // If this is the first return lowered for this function, add
1897 // the regs to the liveout set for the function.
1898 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1899 for (unsigned i = 0; i != RVLocs.size(); ++i)
1900 if (RVLocs[i].isRegLoc())
1901 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1906 // Copy the result values into the output registers.
1907 for (unsigned i = 0, realRVLocIdx = 0;
1909 ++i, ++realRVLocIdx) {
1910 CCValAssign &VA = RVLocs[i];
1911 assert(VA.isRegLoc() && "Can only return in registers!");
1913 SDValue Arg = OutVals[realRVLocIdx];
1915 switch (VA.getLocInfo()) {
1916 default: llvm_unreachable("Unknown loc info!");
1917 case CCValAssign::Full: break;
1918 case CCValAssign::BCvt:
1919 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1923 if (VA.needsCustom()) {
1924 if (VA.getLocVT() == MVT::v2f64) {
1925 // Extract the first half and return it in two registers.
1926 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1927 DAG.getConstant(0, MVT::i32));
1928 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1929 DAG.getVTList(MVT::i32, MVT::i32), Half);
1931 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1932 Flag = Chain.getValue(1);
1933 VA = RVLocs[++i]; // skip ahead to next loc
1934 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1935 HalfGPRs.getValue(1), Flag);
1936 Flag = Chain.getValue(1);
1937 VA = RVLocs[++i]; // skip ahead to next loc
1939 // Extract the 2nd half and fall through to handle it as an f64 value.
1940 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1941 DAG.getConstant(1, MVT::i32));
1943 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1945 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1946 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1947 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1948 Flag = Chain.getValue(1);
1949 VA = RVLocs[++i]; // skip ahead to next loc
1950 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1953 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1955 // Guarantee that all emitted copies are
1956 // stuck together, avoiding something bad.
1957 Flag = Chain.getValue(1);
1962 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1964 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1969 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1970 if (N->getNumValues() != 1)
1972 if (!N->hasNUsesOfValue(1, 0))
1975 SDValue TCChain = Chain;
1976 SDNode *Copy = *N->use_begin();
1977 if (Copy->getOpcode() == ISD::CopyToReg) {
1978 // If the copy has a glue operand, we conservatively assume it isn't safe to
1979 // perform a tail call.
1980 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1982 TCChain = Copy->getOperand(0);
1983 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1984 SDNode *VMov = Copy;
1985 // f64 returned in a pair of GPRs.
1986 SmallPtrSet<SDNode*, 2> Copies;
1987 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1989 if (UI->getOpcode() != ISD::CopyToReg)
1993 if (Copies.size() > 2)
1996 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1998 SDValue UseChain = UI->getOperand(0);
1999 if (Copies.count(UseChain.getNode()))
2006 } else if (Copy->getOpcode() == ISD::BITCAST) {
2007 // f32 returned in a single GPR.
2008 if (!Copy->hasOneUse())
2010 Copy = *Copy->use_begin();
2011 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2013 Chain = Copy->getOperand(0);
2018 bool HasRet = false;
2019 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2021 if (UI->getOpcode() != ARMISD::RET_FLAG)
2033 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2034 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2037 if (!CI->isTailCall())
2040 return !Subtarget->isThumb1Only();
2043 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2044 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2045 // one of the above mentioned nodes. It has to be wrapped because otherwise
2046 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2047 // be used to form addressing mode. These wrapped nodes will be selected
2049 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2050 EVT PtrVT = Op.getValueType();
2051 // FIXME there is no actual debug info here
2052 DebugLoc dl = Op.getDebugLoc();
2053 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2055 if (CP->isMachineConstantPoolEntry())
2056 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2057 CP->getAlignment());
2059 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2060 CP->getAlignment());
2061 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2064 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2065 return MachineJumpTableInfo::EK_Inline;
2068 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2069 SelectionDAG &DAG) const {
2070 MachineFunction &MF = DAG.getMachineFunction();
2071 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2072 unsigned ARMPCLabelIndex = 0;
2073 DebugLoc DL = Op.getDebugLoc();
2074 EVT PtrVT = getPointerTy();
2075 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2076 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2078 if (RelocM == Reloc::Static) {
2079 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2081 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2082 ARMPCLabelIndex = AFI->createPICLabelUId();
2083 ARMConstantPoolValue *CPV =
2084 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2085 ARMCP::CPBlockAddress, PCAdj);
2086 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2088 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2089 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2090 MachinePointerInfo::getConstantPool(),
2091 false, false, false, 0);
2092 if (RelocM == Reloc::Static)
2094 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2095 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2098 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2100 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2101 SelectionDAG &DAG) const {
2102 DebugLoc dl = GA->getDebugLoc();
2103 EVT PtrVT = getPointerTy();
2104 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2105 MachineFunction &MF = DAG.getMachineFunction();
2106 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2107 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2108 ARMConstantPoolValue *CPV =
2109 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2110 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2111 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2112 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2113 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2114 MachinePointerInfo::getConstantPool(),
2115 false, false, false, 0);
2116 SDValue Chain = Argument.getValue(1);
2118 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2119 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2121 // call __tls_get_addr.
2124 Entry.Node = Argument;
2125 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2126 Args.push_back(Entry);
2127 // FIXME: is there useful debug info available here?
2128 TargetLowering::CallLoweringInfo CLI(Chain,
2129 (Type *) Type::getInt32Ty(*DAG.getContext()),
2130 false, false, false, false,
2131 0, CallingConv::C, /*isTailCall=*/false,
2132 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2133 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2134 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2135 return CallResult.first;
2138 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2139 // "local exec" model.
2141 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2143 TLSModel::Model model) const {
2144 const GlobalValue *GV = GA->getGlobal();
2145 DebugLoc dl = GA->getDebugLoc();
2147 SDValue Chain = DAG.getEntryNode();
2148 EVT PtrVT = getPointerTy();
2149 // Get the Thread Pointer
2150 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2152 if (model == TLSModel::InitialExec) {
2153 MachineFunction &MF = DAG.getMachineFunction();
2154 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2155 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2156 // Initial exec model.
2157 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2158 ARMConstantPoolValue *CPV =
2159 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2160 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2162 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2163 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2164 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2165 MachinePointerInfo::getConstantPool(),
2166 false, false, false, 0);
2167 Chain = Offset.getValue(1);
2169 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2170 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2172 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2173 MachinePointerInfo::getConstantPool(),
2174 false, false, false, 0);
2177 assert(model == TLSModel::LocalExec);
2178 ARMConstantPoolValue *CPV =
2179 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2180 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2181 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2182 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2183 MachinePointerInfo::getConstantPool(),
2184 false, false, false, 0);
2187 // The address of the thread local variable is the add of the thread
2188 // pointer with the offset of the variable.
2189 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2193 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2194 // TODO: implement the "local dynamic" model
2195 assert(Subtarget->isTargetELF() &&
2196 "TLS not implemented for non-ELF targets");
2197 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2199 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2202 case TLSModel::GeneralDynamic:
2203 case TLSModel::LocalDynamic:
2204 return LowerToTLSGeneralDynamicModel(GA, DAG);
2205 case TLSModel::InitialExec:
2206 case TLSModel::LocalExec:
2207 return LowerToTLSExecModels(GA, DAG, model);
2209 llvm_unreachable("bogus TLS model");
2212 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2213 SelectionDAG &DAG) const {
2214 EVT PtrVT = getPointerTy();
2215 DebugLoc dl = Op.getDebugLoc();
2216 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2217 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2218 if (RelocM == Reloc::PIC_) {
2219 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2220 ARMConstantPoolValue *CPV =
2221 ARMConstantPoolConstant::Create(GV,
2222 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2223 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2224 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2225 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2227 MachinePointerInfo::getConstantPool(),
2228 false, false, false, 0);
2229 SDValue Chain = Result.getValue(1);
2230 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2231 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2233 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2234 MachinePointerInfo::getGOT(),
2235 false, false, false, 0);
2239 // If we have T2 ops, we can materialize the address directly via movt/movw
2240 // pair. This is always cheaper.
2241 if (Subtarget->useMovt()) {
2243 // FIXME: Once remat is capable of dealing with instructions with register
2244 // operands, expand this into two nodes.
2245 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2246 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2248 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2249 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2250 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2251 MachinePointerInfo::getConstantPool(),
2252 false, false, false, 0);
2256 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2257 SelectionDAG &DAG) const {
2258 EVT PtrVT = getPointerTy();
2259 DebugLoc dl = Op.getDebugLoc();
2260 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2261 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2262 MachineFunction &MF = DAG.getMachineFunction();
2263 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2265 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2266 // update ARMFastISel::ARMMaterializeGV.
2267 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2269 // FIXME: Once remat is capable of dealing with instructions with register
2270 // operands, expand this into two nodes.
2271 if (RelocM == Reloc::Static)
2272 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2273 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2275 unsigned Wrapper = (RelocM == Reloc::PIC_)
2276 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2277 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2278 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2279 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2280 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2281 MachinePointerInfo::getGOT(),
2282 false, false, false, 0);
2286 unsigned ARMPCLabelIndex = 0;
2288 if (RelocM == Reloc::Static) {
2289 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2291 ARMPCLabelIndex = AFI->createPICLabelUId();
2292 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2293 ARMConstantPoolValue *CPV =
2294 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2296 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2298 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2300 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2301 MachinePointerInfo::getConstantPool(),
2302 false, false, false, 0);
2303 SDValue Chain = Result.getValue(1);
2305 if (RelocM == Reloc::PIC_) {
2306 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2307 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2310 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2311 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2312 false, false, false, 0);
2317 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2318 SelectionDAG &DAG) const {
2319 assert(Subtarget->isTargetELF() &&
2320 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2321 MachineFunction &MF = DAG.getMachineFunction();
2322 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2323 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2324 EVT PtrVT = getPointerTy();
2325 DebugLoc dl = Op.getDebugLoc();
2326 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2327 ARMConstantPoolValue *CPV =
2328 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2329 ARMPCLabelIndex, PCAdj);
2330 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2331 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2332 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2333 MachinePointerInfo::getConstantPool(),
2334 false, false, false, 0);
2335 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2336 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2340 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2341 DebugLoc dl = Op.getDebugLoc();
2342 SDValue Val = DAG.getConstant(0, MVT::i32);
2343 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2344 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2345 Op.getOperand(1), Val);
2349 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2350 DebugLoc dl = Op.getDebugLoc();
2351 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2352 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2356 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2357 const ARMSubtarget *Subtarget) const {
2358 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2359 DebugLoc dl = Op.getDebugLoc();
2361 default: return SDValue(); // Don't custom lower most intrinsics.
2362 case Intrinsic::arm_thread_pointer: {
2363 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2364 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2366 case Intrinsic::eh_sjlj_lsda: {
2367 MachineFunction &MF = DAG.getMachineFunction();
2368 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2369 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2370 EVT PtrVT = getPointerTy();
2371 DebugLoc dl = Op.getDebugLoc();
2372 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2374 unsigned PCAdj = (RelocM != Reloc::PIC_)
2375 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2376 ARMConstantPoolValue *CPV =
2377 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2378 ARMCP::CPLSDA, PCAdj);
2379 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2380 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2382 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2383 MachinePointerInfo::getConstantPool(),
2384 false, false, false, 0);
2386 if (RelocM == Reloc::PIC_) {
2387 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2388 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2392 case Intrinsic::arm_neon_vmulls:
2393 case Intrinsic::arm_neon_vmullu: {
2394 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2395 ? ARMISD::VMULLs : ARMISD::VMULLu;
2396 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2397 Op.getOperand(1), Op.getOperand(2));
2402 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2403 const ARMSubtarget *Subtarget) {
2404 DebugLoc dl = Op.getDebugLoc();
2405 if (!Subtarget->hasDataBarrier()) {
2406 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2407 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2409 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2410 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2411 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2412 DAG.getConstant(0, MVT::i32));
2415 SDValue Op5 = Op.getOperand(5);
2416 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2417 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2418 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2419 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2421 ARM_MB::MemBOpt DMBOpt;
2422 if (isDeviceBarrier)
2423 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2425 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2426 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2427 DAG.getConstant(DMBOpt, MVT::i32));
2431 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2432 const ARMSubtarget *Subtarget) {
2433 // FIXME: handle "fence singlethread" more efficiently.
2434 DebugLoc dl = Op.getDebugLoc();
2435 if (!Subtarget->hasDataBarrier()) {
2436 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2437 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2439 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2440 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2441 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2442 DAG.getConstant(0, MVT::i32));
2445 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2446 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2449 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2450 const ARMSubtarget *Subtarget) {
2451 // ARM pre v5TE and Thumb1 does not have preload instructions.
2452 if (!(Subtarget->isThumb2() ||
2453 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2454 // Just preserve the chain.
2455 return Op.getOperand(0);
2457 DebugLoc dl = Op.getDebugLoc();
2458 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2460 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2461 // ARMv7 with MP extension has PLDW.
2462 return Op.getOperand(0);
2464 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2465 if (Subtarget->isThumb()) {
2467 isRead = ~isRead & 1;
2468 isData = ~isData & 1;
2471 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2472 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2473 DAG.getConstant(isData, MVT::i32));
2476 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2477 MachineFunction &MF = DAG.getMachineFunction();
2478 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2480 // vastart just stores the address of the VarArgsFrameIndex slot into the
2481 // memory location argument.
2482 DebugLoc dl = Op.getDebugLoc();
2483 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2484 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2485 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2486 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2487 MachinePointerInfo(SV), false, false, 0);
2491 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2492 SDValue &Root, SelectionDAG &DAG,
2493 DebugLoc dl) const {
2494 MachineFunction &MF = DAG.getMachineFunction();
2495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2497 const TargetRegisterClass *RC;
2498 if (AFI->isThumb1OnlyFunction())
2499 RC = &ARM::tGPRRegClass;
2501 RC = &ARM::GPRRegClass;
2503 // Transform the arguments stored in physical registers into virtual ones.
2504 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2505 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2508 if (NextVA.isMemLoc()) {
2509 MachineFrameInfo *MFI = MF.getFrameInfo();
2510 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2512 // Create load node to retrieve arguments from the stack.
2513 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2514 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2515 MachinePointerInfo::getFixedStack(FI),
2516 false, false, false, 0);
2518 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2519 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2522 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2526 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2527 unsigned &VARegSize, unsigned &VARegSaveSize)
2530 if (CCInfo.isFirstByValRegValid())
2531 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2533 unsigned int firstUnalloced;
2534 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2535 sizeof(GPRArgRegs) /
2536 sizeof(GPRArgRegs[0]));
2537 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2540 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2541 VARegSize = NumGPRs * 4;
2542 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2545 // The remaining GPRs hold either the beginning of variable-argument
2546 // data, or the beginning of an aggregate passed by value (usuall
2547 // byval). Either way, we allocate stack slots adjacent to the data
2548 // provided by our caller, and store the unallocated registers there.
2549 // If this is a variadic function, the va_list pointer will begin with
2550 // these values; otherwise, this reassembles a (byval) structure that
2551 // was split between registers and memory.
2553 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2554 DebugLoc dl, SDValue &Chain,
2555 const Value *OrigArg,
2556 unsigned OffsetFromOrigArg,
2557 unsigned ArgOffset) const {
2558 MachineFunction &MF = DAG.getMachineFunction();
2559 MachineFrameInfo *MFI = MF.getFrameInfo();
2560 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2561 unsigned firstRegToSaveIndex;
2562 if (CCInfo.isFirstByValRegValid())
2563 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2565 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2566 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2569 unsigned VARegSize, VARegSaveSize;
2570 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2571 if (VARegSaveSize) {
2572 // If this function is vararg, store any remaining integer argument regs
2573 // to their spots on the stack so that they may be loaded by deferencing
2574 // the result of va_next.
2575 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2576 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2577 ArgOffset + VARegSaveSize
2580 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2583 SmallVector<SDValue, 4> MemOps;
2584 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2585 const TargetRegisterClass *RC;
2586 if (AFI->isThumb1OnlyFunction())
2587 RC = &ARM::tGPRRegClass;
2589 RC = &ARM::GPRRegClass;
2591 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2592 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2594 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2595 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2597 MemOps.push_back(Store);
2598 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2599 DAG.getConstant(4, getPointerTy()));
2601 if (!MemOps.empty())
2602 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2603 &MemOps[0], MemOps.size());
2605 // This will point to the next argument passed via stack.
2606 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2610 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2611 CallingConv::ID CallConv, bool isVarArg,
2612 const SmallVectorImpl<ISD::InputArg>
2614 DebugLoc dl, SelectionDAG &DAG,
2615 SmallVectorImpl<SDValue> &InVals)
2617 MachineFunction &MF = DAG.getMachineFunction();
2618 MachineFrameInfo *MFI = MF.getFrameInfo();
2620 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2622 // Assign locations to all of the incoming arguments.
2623 SmallVector<CCValAssign, 16> ArgLocs;
2624 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2625 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2626 CCInfo.AnalyzeFormalArguments(Ins,
2627 CCAssignFnForNode(CallConv, /* Return*/ false,
2630 SmallVector<SDValue, 16> ArgValues;
2631 int lastInsIndex = -1;
2633 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2634 unsigned CurArgIdx = 0;
2635 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2636 CCValAssign &VA = ArgLocs[i];
2637 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2638 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2639 // Arguments stored in registers.
2640 if (VA.isRegLoc()) {
2641 EVT RegVT = VA.getLocVT();
2643 if (VA.needsCustom()) {
2644 // f64 and vector types are split up into multiple registers or
2645 // combinations of registers and stack slots.
2646 if (VA.getLocVT() == MVT::v2f64) {
2647 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2649 VA = ArgLocs[++i]; // skip ahead to next loc
2651 if (VA.isMemLoc()) {
2652 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2653 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2654 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2655 MachinePointerInfo::getFixedStack(FI),
2656 false, false, false, 0);
2658 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2661 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2662 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2663 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2664 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2665 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2667 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2670 const TargetRegisterClass *RC;
2672 if (RegVT == MVT::f32)
2673 RC = &ARM::SPRRegClass;
2674 else if (RegVT == MVT::f64)
2675 RC = &ARM::DPRRegClass;
2676 else if (RegVT == MVT::v2f64)
2677 RC = &ARM::QPRRegClass;
2678 else if (RegVT == MVT::i32)
2679 RC = AFI->isThumb1OnlyFunction() ?
2680 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2681 (const TargetRegisterClass*)&ARM::GPRRegClass;
2683 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2685 // Transform the arguments in physical registers into virtual ones.
2686 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2687 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2690 // If this is an 8 or 16-bit value, it is really passed promoted
2691 // to 32 bits. Insert an assert[sz]ext to capture this, then
2692 // truncate to the right size.
2693 switch (VA.getLocInfo()) {
2694 default: llvm_unreachable("Unknown loc info!");
2695 case CCValAssign::Full: break;
2696 case CCValAssign::BCvt:
2697 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2699 case CCValAssign::SExt:
2700 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2701 DAG.getValueType(VA.getValVT()));
2702 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2704 case CCValAssign::ZExt:
2705 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2706 DAG.getValueType(VA.getValVT()));
2707 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2711 InVals.push_back(ArgValue);
2713 } else { // VA.isRegLoc()
2716 assert(VA.isMemLoc());
2717 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2719 int index = ArgLocs[i].getValNo();
2721 // Some Ins[] entries become multiple ArgLoc[] entries.
2722 // Process them only once.
2723 if (index != lastInsIndex)
2725 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2726 // FIXME: For now, all byval parameter objects are marked mutable.
2727 // This can be changed with more analysis.
2728 // In case of tail call optimization mark all arguments mutable.
2729 // Since they could be overwritten by lowering of arguments in case of
2731 if (Flags.isByVal()) {
2732 unsigned VARegSize, VARegSaveSize;
2733 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2734 VarArgStyleRegisters(CCInfo, DAG,
2735 dl, Chain, CurOrigArg, Ins[VA.getValNo()].PartOffset, 0);
2736 unsigned Bytes = Flags.getByValSize() - VARegSize;
2737 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2738 int FI = MFI->CreateFixedObject(Bytes,
2739 VA.getLocMemOffset(), false);
2740 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2742 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2743 VA.getLocMemOffset(), true);
2745 // Create load nodes to retrieve arguments from the stack.
2746 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2747 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2748 MachinePointerInfo::getFixedStack(FI),
2749 false, false, false, 0));
2751 lastInsIndex = index;
2758 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2759 CCInfo.getNextStackOffset());
2764 /// isFloatingPointZero - Return true if this is +0.0.
2765 static bool isFloatingPointZero(SDValue Op) {
2766 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2767 return CFP->getValueAPF().isPosZero();
2768 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2769 // Maybe this has already been legalized into the constant pool?
2770 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2771 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2772 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2773 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2774 return CFP->getValueAPF().isPosZero();
2780 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2781 /// the given operands.
2783 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2784 SDValue &ARMcc, SelectionDAG &DAG,
2785 DebugLoc dl) const {
2786 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2787 unsigned C = RHSC->getZExtValue();
2788 if (!isLegalICmpImmediate(C)) {
2789 // Constant does not fit, try adjusting it by one?
2794 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2795 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2796 RHS = DAG.getConstant(C-1, MVT::i32);
2801 if (C != 0 && isLegalICmpImmediate(C-1)) {
2802 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2803 RHS = DAG.getConstant(C-1, MVT::i32);
2808 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2809 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2810 RHS = DAG.getConstant(C+1, MVT::i32);
2815 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2816 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2817 RHS = DAG.getConstant(C+1, MVT::i32);
2824 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2825 ARMISD::NodeType CompareType;
2828 CompareType = ARMISD::CMP;
2833 CompareType = ARMISD::CMPZ;
2836 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2837 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2840 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2842 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2843 DebugLoc dl) const {
2845 if (!isFloatingPointZero(RHS))
2846 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2848 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2849 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2852 /// duplicateCmp - Glue values can have only one use, so this function
2853 /// duplicates a comparison node.
2855 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2856 unsigned Opc = Cmp.getOpcode();
2857 DebugLoc DL = Cmp.getDebugLoc();
2858 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2859 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2861 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2862 Cmp = Cmp.getOperand(0);
2863 Opc = Cmp.getOpcode();
2864 if (Opc == ARMISD::CMPFP)
2865 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2867 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2868 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2870 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2873 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2874 SDValue Cond = Op.getOperand(0);
2875 SDValue SelectTrue = Op.getOperand(1);
2876 SDValue SelectFalse = Op.getOperand(2);
2877 DebugLoc dl = Op.getDebugLoc();
2881 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2882 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2884 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2885 const ConstantSDNode *CMOVTrue =
2886 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2887 const ConstantSDNode *CMOVFalse =
2888 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2890 if (CMOVTrue && CMOVFalse) {
2891 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2892 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2896 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2898 False = SelectFalse;
2899 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2904 if (True.getNode() && False.getNode()) {
2905 EVT VT = Op.getValueType();
2906 SDValue ARMcc = Cond.getOperand(2);
2907 SDValue CCR = Cond.getOperand(3);
2908 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2909 assert(True.getValueType() == VT);
2910 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2915 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2916 // undefined bits before doing a full-word comparison with zero.
2917 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2918 DAG.getConstant(1, Cond.getValueType()));
2920 return DAG.getSelectCC(dl, Cond,
2921 DAG.getConstant(0, Cond.getValueType()),
2922 SelectTrue, SelectFalse, ISD::SETNE);
2925 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2926 EVT VT = Op.getValueType();
2927 SDValue LHS = Op.getOperand(0);
2928 SDValue RHS = Op.getOperand(1);
2929 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2930 SDValue TrueVal = Op.getOperand(2);
2931 SDValue FalseVal = Op.getOperand(3);
2932 DebugLoc dl = Op.getDebugLoc();
2934 if (LHS.getValueType() == MVT::i32) {
2936 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2937 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2938 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2941 ARMCC::CondCodes CondCode, CondCode2;
2942 FPCCToARMCC(CC, CondCode, CondCode2);
2944 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2945 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2946 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2947 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2949 if (CondCode2 != ARMCC::AL) {
2950 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2951 // FIXME: Needs another CMP because flag can have but one use.
2952 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2953 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2954 Result, TrueVal, ARMcc2, CCR, Cmp2);
2959 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2960 /// to morph to an integer compare sequence.
2961 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2962 const ARMSubtarget *Subtarget) {
2963 SDNode *N = Op.getNode();
2964 if (!N->hasOneUse())
2965 // Otherwise it requires moving the value from fp to integer registers.
2967 if (!N->getNumValues())
2969 EVT VT = Op.getValueType();
2970 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2971 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2972 // vmrs are very slow, e.g. cortex-a8.
2975 if (isFloatingPointZero(Op)) {
2979 return ISD::isNormalLoad(N);
2982 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2983 if (isFloatingPointZero(Op))
2984 return DAG.getConstant(0, MVT::i32);
2986 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2987 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2988 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2989 Ld->isVolatile(), Ld->isNonTemporal(),
2990 Ld->isInvariant(), Ld->getAlignment());
2992 llvm_unreachable("Unknown VFP cmp argument!");
2995 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2996 SDValue &RetVal1, SDValue &RetVal2) {
2997 if (isFloatingPointZero(Op)) {
2998 RetVal1 = DAG.getConstant(0, MVT::i32);
2999 RetVal2 = DAG.getConstant(0, MVT::i32);
3003 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3004 SDValue Ptr = Ld->getBasePtr();
3005 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3006 Ld->getChain(), Ptr,
3007 Ld->getPointerInfo(),
3008 Ld->isVolatile(), Ld->isNonTemporal(),
3009 Ld->isInvariant(), Ld->getAlignment());
3011 EVT PtrType = Ptr.getValueType();
3012 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3013 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3014 PtrType, Ptr, DAG.getConstant(4, PtrType));
3015 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3016 Ld->getChain(), NewPtr,
3017 Ld->getPointerInfo().getWithOffset(4),
3018 Ld->isVolatile(), Ld->isNonTemporal(),
3019 Ld->isInvariant(), NewAlign);
3023 llvm_unreachable("Unknown VFP cmp argument!");
3026 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3027 /// f32 and even f64 comparisons to integer ones.
3029 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3030 SDValue Chain = Op.getOperand(0);
3031 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3032 SDValue LHS = Op.getOperand(2);
3033 SDValue RHS = Op.getOperand(3);
3034 SDValue Dest = Op.getOperand(4);
3035 DebugLoc dl = Op.getDebugLoc();
3037 bool LHSSeenZero = false;
3038 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3039 bool RHSSeenZero = false;
3040 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3041 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3042 // If unsafe fp math optimization is enabled and there are no other uses of
3043 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3044 // to an integer comparison.
3045 if (CC == ISD::SETOEQ)
3047 else if (CC == ISD::SETUNE)
3050 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3052 if (LHS.getValueType() == MVT::f32) {
3053 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3054 bitcastf32Toi32(LHS, DAG), Mask);
3055 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3056 bitcastf32Toi32(RHS, DAG), Mask);
3057 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3058 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3059 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3060 Chain, Dest, ARMcc, CCR, Cmp);
3065 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3066 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3067 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3068 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3069 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3070 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3071 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3072 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3073 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3079 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3080 SDValue Chain = Op.getOperand(0);
3081 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3082 SDValue LHS = Op.getOperand(2);
3083 SDValue RHS = Op.getOperand(3);
3084 SDValue Dest = Op.getOperand(4);
3085 DebugLoc dl = Op.getDebugLoc();
3087 if (LHS.getValueType() == MVT::i32) {
3089 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3090 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3091 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3092 Chain, Dest, ARMcc, CCR, Cmp);
3095 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3097 if (getTargetMachine().Options.UnsafeFPMath &&
3098 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3099 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3100 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3101 if (Result.getNode())
3105 ARMCC::CondCodes CondCode, CondCode2;
3106 FPCCToARMCC(CC, CondCode, CondCode2);
3108 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3109 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3110 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3111 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3112 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3113 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3114 if (CondCode2 != ARMCC::AL) {
3115 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3116 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3117 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3122 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3123 SDValue Chain = Op.getOperand(0);
3124 SDValue Table = Op.getOperand(1);
3125 SDValue Index = Op.getOperand(2);
3126 DebugLoc dl = Op.getDebugLoc();
3128 EVT PTy = getPointerTy();
3129 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3130 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3131 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3132 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3133 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3134 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3135 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3136 if (Subtarget->isThumb2()) {
3137 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3138 // which does another jump to the destination. This also makes it easier
3139 // to translate it to TBB / TBH later.
3140 // FIXME: This might not work if the function is extremely large.
3141 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3142 Addr, Op.getOperand(2), JTI, UId);
3144 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3145 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3146 MachinePointerInfo::getJumpTable(),
3147 false, false, false, 0);
3148 Chain = Addr.getValue(1);
3149 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3150 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3152 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3153 MachinePointerInfo::getJumpTable(),
3154 false, false, false, 0);
3155 Chain = Addr.getValue(1);
3156 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3160 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3161 EVT VT = Op.getValueType();
3162 DebugLoc dl = Op.getDebugLoc();
3164 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3165 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3167 return DAG.UnrollVectorOp(Op.getNode());
3170 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3171 "Invalid type for custom lowering!");
3172 if (VT != MVT::v4i16)
3173 return DAG.UnrollVectorOp(Op.getNode());
3175 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3176 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3179 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3180 EVT VT = Op.getValueType();
3182 return LowerVectorFP_TO_INT(Op, DAG);
3184 DebugLoc dl = Op.getDebugLoc();
3187 switch (Op.getOpcode()) {
3188 default: llvm_unreachable("Invalid opcode!");
3189 case ISD::FP_TO_SINT:
3190 Opc = ARMISD::FTOSI;
3192 case ISD::FP_TO_UINT:
3193 Opc = ARMISD::FTOUI;
3196 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3197 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3200 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3201 EVT VT = Op.getValueType();
3202 DebugLoc dl = Op.getDebugLoc();
3204 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3205 if (VT.getVectorElementType() == MVT::f32)
3207 return DAG.UnrollVectorOp(Op.getNode());
3210 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3211 "Invalid type for custom lowering!");
3212 if (VT != MVT::v4f32)
3213 return DAG.UnrollVectorOp(Op.getNode());
3217 switch (Op.getOpcode()) {
3218 default: llvm_unreachable("Invalid opcode!");
3219 case ISD::SINT_TO_FP:
3220 CastOpc = ISD::SIGN_EXTEND;
3221 Opc = ISD::SINT_TO_FP;
3223 case ISD::UINT_TO_FP:
3224 CastOpc = ISD::ZERO_EXTEND;
3225 Opc = ISD::UINT_TO_FP;
3229 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3230 return DAG.getNode(Opc, dl, VT, Op);
3233 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3234 EVT VT = Op.getValueType();
3236 return LowerVectorINT_TO_FP(Op, DAG);
3238 DebugLoc dl = Op.getDebugLoc();
3241 switch (Op.getOpcode()) {
3242 default: llvm_unreachable("Invalid opcode!");
3243 case ISD::SINT_TO_FP:
3244 Opc = ARMISD::SITOF;
3246 case ISD::UINT_TO_FP:
3247 Opc = ARMISD::UITOF;
3251 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3252 return DAG.getNode(Opc, dl, VT, Op);
3255 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3256 // Implement fcopysign with a fabs and a conditional fneg.
3257 SDValue Tmp0 = Op.getOperand(0);
3258 SDValue Tmp1 = Op.getOperand(1);
3259 DebugLoc dl = Op.getDebugLoc();
3260 EVT VT = Op.getValueType();
3261 EVT SrcVT = Tmp1.getValueType();
3262 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3263 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3264 bool UseNEON = !InGPR && Subtarget->hasNEON();
3267 // Use VBSL to copy the sign bit.
3268 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3269 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3270 DAG.getTargetConstant(EncodedVal, MVT::i32));
3271 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3273 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3274 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3275 DAG.getConstant(32, MVT::i32));
3276 else /*if (VT == MVT::f32)*/
3277 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3278 if (SrcVT == MVT::f32) {
3279 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3281 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3282 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3283 DAG.getConstant(32, MVT::i32));
3284 } else if (VT == MVT::f32)
3285 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3286 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3287 DAG.getConstant(32, MVT::i32));
3288 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3289 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3291 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3293 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3294 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3295 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3297 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3298 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3299 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3300 if (VT == MVT::f32) {
3301 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3302 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3303 DAG.getConstant(0, MVT::i32));
3305 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3311 // Bitcast operand 1 to i32.
3312 if (SrcVT == MVT::f64)
3313 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3314 &Tmp1, 1).getValue(1);
3315 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3317 // Or in the signbit with integer operations.
3318 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3319 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3320 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3321 if (VT == MVT::f32) {
3322 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3323 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3324 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3325 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3328 // f64: Or the high part with signbit and then combine two parts.
3329 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3331 SDValue Lo = Tmp0.getValue(0);
3332 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3333 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3334 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3337 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3338 MachineFunction &MF = DAG.getMachineFunction();
3339 MachineFrameInfo *MFI = MF.getFrameInfo();
3340 MFI->setReturnAddressIsTaken(true);
3342 EVT VT = Op.getValueType();
3343 DebugLoc dl = Op.getDebugLoc();
3344 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3346 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3347 SDValue Offset = DAG.getConstant(4, MVT::i32);
3348 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3349 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3350 MachinePointerInfo(), false, false, false, 0);
3353 // Return LR, which contains the return address. Mark it an implicit live-in.
3354 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3355 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3358 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3359 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3360 MFI->setFrameAddressIsTaken(true);
3362 EVT VT = Op.getValueType();
3363 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3364 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3365 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3366 ? ARM::R7 : ARM::R11;
3367 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3369 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3370 MachinePointerInfo(),
3371 false, false, false, 0);
3375 /// ExpandBITCAST - If the target supports VFP, this function is called to
3376 /// expand a bit convert where either the source or destination type is i64 to
3377 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3378 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3379 /// vectors), since the legalizer won't know what to do with that.
3380 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3381 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3382 DebugLoc dl = N->getDebugLoc();
3383 SDValue Op = N->getOperand(0);
3385 // This function is only supposed to be called for i64 types, either as the
3386 // source or destination of the bit convert.
3387 EVT SrcVT = Op.getValueType();
3388 EVT DstVT = N->getValueType(0);
3389 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3390 "ExpandBITCAST called for non-i64 type");
3392 // Turn i64->f64 into VMOVDRR.
3393 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3394 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3395 DAG.getConstant(0, MVT::i32));
3396 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3397 DAG.getConstant(1, MVT::i32));
3398 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3399 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3402 // Turn f64->i64 into VMOVRRD.
3403 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3404 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3405 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3406 // Merge the pieces into a single i64 value.
3407 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3413 /// getZeroVector - Returns a vector of specified type with all zero elements.
3414 /// Zero vectors are used to represent vector negation and in those cases
3415 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3416 /// not support i64 elements, so sometimes the zero vectors will need to be
3417 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3419 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3420 assert(VT.isVector() && "Expected a vector type");
3421 // The canonical modified immediate encoding of a zero vector is....0!
3422 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3423 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3424 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3425 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3428 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3429 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3430 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3431 SelectionDAG &DAG) const {
3432 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3433 EVT VT = Op.getValueType();
3434 unsigned VTBits = VT.getSizeInBits();
3435 DebugLoc dl = Op.getDebugLoc();
3436 SDValue ShOpLo = Op.getOperand(0);
3437 SDValue ShOpHi = Op.getOperand(1);
3438 SDValue ShAmt = Op.getOperand(2);
3440 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3442 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3444 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3445 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3446 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3447 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3448 DAG.getConstant(VTBits, MVT::i32));
3449 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3450 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3451 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3453 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3454 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3456 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3457 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3460 SDValue Ops[2] = { Lo, Hi };
3461 return DAG.getMergeValues(Ops, 2, dl);
3464 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3465 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3466 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3467 SelectionDAG &DAG) const {
3468 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3469 EVT VT = Op.getValueType();
3470 unsigned VTBits = VT.getSizeInBits();
3471 DebugLoc dl = Op.getDebugLoc();
3472 SDValue ShOpLo = Op.getOperand(0);
3473 SDValue ShOpHi = Op.getOperand(1);
3474 SDValue ShAmt = Op.getOperand(2);
3477 assert(Op.getOpcode() == ISD::SHL_PARTS);
3478 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3479 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3480 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3481 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3482 DAG.getConstant(VTBits, MVT::i32));
3483 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3484 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3486 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3487 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3488 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3490 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3491 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3494 SDValue Ops[2] = { Lo, Hi };
3495 return DAG.getMergeValues(Ops, 2, dl);
3498 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3499 SelectionDAG &DAG) const {
3500 // The rounding mode is in bits 23:22 of the FPSCR.
3501 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3502 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3503 // so that the shift + and get folded into a bitfield extract.
3504 DebugLoc dl = Op.getDebugLoc();
3505 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3506 DAG.getConstant(Intrinsic::arm_get_fpscr,
3508 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3509 DAG.getConstant(1U << 22, MVT::i32));
3510 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3511 DAG.getConstant(22, MVT::i32));
3512 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3513 DAG.getConstant(3, MVT::i32));
3516 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3517 const ARMSubtarget *ST) {
3518 EVT VT = N->getValueType(0);
3519 DebugLoc dl = N->getDebugLoc();
3521 if (!ST->hasV6T2Ops())
3524 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3525 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3528 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3529 const ARMSubtarget *ST) {
3530 EVT VT = N->getValueType(0);
3531 DebugLoc dl = N->getDebugLoc();
3536 // Lower vector shifts on NEON to use VSHL.
3537 assert(ST->hasNEON() && "unexpected vector shift");
3539 // Left shifts translate directly to the vshiftu intrinsic.
3540 if (N->getOpcode() == ISD::SHL)
3541 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3542 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3543 N->getOperand(0), N->getOperand(1));
3545 assert((N->getOpcode() == ISD::SRA ||
3546 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3548 // NEON uses the same intrinsics for both left and right shifts. For
3549 // right shifts, the shift amounts are negative, so negate the vector of
3551 EVT ShiftVT = N->getOperand(1).getValueType();
3552 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3553 getZeroVector(ShiftVT, DAG, dl),
3555 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3556 Intrinsic::arm_neon_vshifts :
3557 Intrinsic::arm_neon_vshiftu);
3558 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3559 DAG.getConstant(vshiftInt, MVT::i32),
3560 N->getOperand(0), NegatedCount);
3563 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3564 const ARMSubtarget *ST) {
3565 EVT VT = N->getValueType(0);
3566 DebugLoc dl = N->getDebugLoc();
3568 // We can get here for a node like i32 = ISD::SHL i32, i64
3572 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3573 "Unknown shift to lower!");
3575 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3576 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3577 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3580 // If we are in thumb mode, we don't have RRX.
3581 if (ST->isThumb1Only()) return SDValue();
3583 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3584 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3585 DAG.getConstant(0, MVT::i32));
3586 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3587 DAG.getConstant(1, MVT::i32));
3589 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3590 // captures the result into a carry flag.
3591 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3592 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3594 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3595 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3597 // Merge the pieces into a single i64 value.
3598 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3601 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3602 SDValue TmpOp0, TmpOp1;
3603 bool Invert = false;
3607 SDValue Op0 = Op.getOperand(0);
3608 SDValue Op1 = Op.getOperand(1);
3609 SDValue CC = Op.getOperand(2);
3610 EVT VT = Op.getValueType();
3611 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3612 DebugLoc dl = Op.getDebugLoc();
3614 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3615 switch (SetCCOpcode) {
3616 default: llvm_unreachable("Illegal FP comparison");
3618 case ISD::SETNE: Invert = true; // Fallthrough
3620 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3622 case ISD::SETLT: Swap = true; // Fallthrough
3624 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3626 case ISD::SETLE: Swap = true; // Fallthrough
3628 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3629 case ISD::SETUGE: Swap = true; // Fallthrough
3630 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3631 case ISD::SETUGT: Swap = true; // Fallthrough
3632 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3633 case ISD::SETUEQ: Invert = true; // Fallthrough
3635 // Expand this to (OLT | OGT).
3639 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3640 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3642 case ISD::SETUO: Invert = true; // Fallthrough
3644 // Expand this to (OLT | OGE).
3648 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3649 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3653 // Integer comparisons.
3654 switch (SetCCOpcode) {
3655 default: llvm_unreachable("Illegal integer comparison");
3656 case ISD::SETNE: Invert = true;
3657 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3658 case ISD::SETLT: Swap = true;
3659 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3660 case ISD::SETLE: Swap = true;
3661 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3662 case ISD::SETULT: Swap = true;
3663 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3664 case ISD::SETULE: Swap = true;
3665 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3668 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3669 if (Opc == ARMISD::VCEQ) {
3672 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3674 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3677 // Ignore bitconvert.
3678 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3679 AndOp = AndOp.getOperand(0);
3681 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3683 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3684 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3691 std::swap(Op0, Op1);
3693 // If one of the operands is a constant vector zero, attempt to fold the
3694 // comparison to a specialized compare-against-zero form.
3696 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3698 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3699 if (Opc == ARMISD::VCGE)
3700 Opc = ARMISD::VCLEZ;
3701 else if (Opc == ARMISD::VCGT)
3702 Opc = ARMISD::VCLTZ;
3707 if (SingleOp.getNode()) {
3710 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3712 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3714 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3716 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3718 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3720 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3723 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3727 Result = DAG.getNOT(dl, Result, VT);
3732 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3733 /// valid vector constant for a NEON instruction with a "modified immediate"
3734 /// operand (e.g., VMOV). If so, return the encoded value.
3735 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3736 unsigned SplatBitSize, SelectionDAG &DAG,
3737 EVT &VT, bool is128Bits, NEONModImmType type) {
3738 unsigned OpCmode, Imm;
3740 // SplatBitSize is set to the smallest size that splats the vector, so a
3741 // zero vector will always have SplatBitSize == 8. However, NEON modified
3742 // immediate instructions others than VMOV do not support the 8-bit encoding
3743 // of a zero vector, and the default encoding of zero is supposed to be the
3748 switch (SplatBitSize) {
3750 if (type != VMOVModImm)
3752 // Any 1-byte value is OK. Op=0, Cmode=1110.
3753 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3756 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3760 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3761 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3762 if ((SplatBits & ~0xff) == 0) {
3763 // Value = 0x00nn: Op=x, Cmode=100x.
3768 if ((SplatBits & ~0xff00) == 0) {
3769 // Value = 0xnn00: Op=x, Cmode=101x.
3771 Imm = SplatBits >> 8;
3777 // NEON's 32-bit VMOV supports splat values where:
3778 // * only one byte is nonzero, or
3779 // * the least significant byte is 0xff and the second byte is nonzero, or
3780 // * the least significant 2 bytes are 0xff and the third is nonzero.
3781 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3782 if ((SplatBits & ~0xff) == 0) {
3783 // Value = 0x000000nn: Op=x, Cmode=000x.
3788 if ((SplatBits & ~0xff00) == 0) {
3789 // Value = 0x0000nn00: Op=x, Cmode=001x.
3791 Imm = SplatBits >> 8;
3794 if ((SplatBits & ~0xff0000) == 0) {
3795 // Value = 0x00nn0000: Op=x, Cmode=010x.
3797 Imm = SplatBits >> 16;
3800 if ((SplatBits & ~0xff000000) == 0) {
3801 // Value = 0xnn000000: Op=x, Cmode=011x.
3803 Imm = SplatBits >> 24;
3807 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3808 if (type == OtherModImm) return SDValue();
3810 if ((SplatBits & ~0xffff) == 0 &&
3811 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3812 // Value = 0x0000nnff: Op=x, Cmode=1100.
3814 Imm = SplatBits >> 8;
3819 if ((SplatBits & ~0xffffff) == 0 &&
3820 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3821 // Value = 0x00nnffff: Op=x, Cmode=1101.
3823 Imm = SplatBits >> 16;
3824 SplatBits |= 0xffff;
3828 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3829 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3830 // VMOV.I32. A (very) minor optimization would be to replicate the value
3831 // and fall through here to test for a valid 64-bit splat. But, then the
3832 // caller would also need to check and handle the change in size.
3836 if (type != VMOVModImm)
3838 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3839 uint64_t BitMask = 0xff;
3841 unsigned ImmMask = 1;
3843 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3844 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3847 } else if ((SplatBits & BitMask) != 0) {
3853 // Op=1, Cmode=1110.
3856 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3861 llvm_unreachable("unexpected size for isNEONModifiedImm");
3864 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3865 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3868 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3869 const ARMSubtarget *ST) const {
3870 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3873 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3874 assert(Op.getValueType() == MVT::f32 &&
3875 "ConstantFP custom lowering should only occur for f32.");
3877 // Try splatting with a VMOV.f32...
3878 APFloat FPVal = CFP->getValueAPF();
3879 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3881 DebugLoc DL = Op.getDebugLoc();
3882 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3883 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3885 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3886 DAG.getConstant(0, MVT::i32));
3889 // If that fails, try a VMOV.i32
3891 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3892 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3894 if (NewVal != SDValue()) {
3895 DebugLoc DL = Op.getDebugLoc();
3896 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3898 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3900 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3901 DAG.getConstant(0, MVT::i32));
3904 // Finally, try a VMVN.i32
3905 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3907 if (NewVal != SDValue()) {
3908 DebugLoc DL = Op.getDebugLoc();
3909 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3910 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3912 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3913 DAG.getConstant(0, MVT::i32));
3920 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3921 bool &ReverseVEXT, unsigned &Imm) {
3922 unsigned NumElts = VT.getVectorNumElements();
3923 ReverseVEXT = false;
3925 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3931 // If this is a VEXT shuffle, the immediate value is the index of the first
3932 // element. The other shuffle indices must be the successive elements after
3934 unsigned ExpectedElt = Imm;
3935 for (unsigned i = 1; i < NumElts; ++i) {
3936 // Increment the expected index. If it wraps around, it may still be
3937 // a VEXT but the source vectors must be swapped.
3939 if (ExpectedElt == NumElts * 2) {
3944 if (M[i] < 0) continue; // ignore UNDEF indices
3945 if (ExpectedElt != static_cast<unsigned>(M[i]))
3949 // Adjust the index value if the source operands will be swapped.
3956 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3957 /// instruction with the specified blocksize. (The order of the elements
3958 /// within each block of the vector is reversed.)
3959 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3960 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3961 "Only possible block sizes for VREV are: 16, 32, 64");
3963 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3967 unsigned NumElts = VT.getVectorNumElements();
3968 unsigned BlockElts = M[0] + 1;
3969 // If the first shuffle index is UNDEF, be optimistic.
3971 BlockElts = BlockSize / EltSz;
3973 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3976 for (unsigned i = 0; i < NumElts; ++i) {
3977 if (M[i] < 0) continue; // ignore UNDEF indices
3978 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3985 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3986 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3987 // range, then 0 is placed into the resulting vector. So pretty much any mask
3988 // of 8 elements can work here.
3989 return VT == MVT::v8i8 && M.size() == 8;
3992 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3993 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3997 unsigned NumElts = VT.getVectorNumElements();
3998 WhichResult = (M[0] == 0 ? 0 : 1);
3999 for (unsigned i = 0; i < NumElts; i += 2) {
4000 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4001 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4007 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4008 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4009 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4010 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4011 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4015 unsigned NumElts = VT.getVectorNumElements();
4016 WhichResult = (M[0] == 0 ? 0 : 1);
4017 for (unsigned i = 0; i < NumElts; i += 2) {
4018 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4019 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4025 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4026 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4030 unsigned NumElts = VT.getVectorNumElements();
4031 WhichResult = (M[0] == 0 ? 0 : 1);
4032 for (unsigned i = 0; i != NumElts; ++i) {
4033 if (M[i] < 0) continue; // ignore UNDEF indices
4034 if ((unsigned) M[i] != 2 * i + WhichResult)
4038 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4039 if (VT.is64BitVector() && EltSz == 32)
4045 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4046 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4047 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4048 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4049 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4053 unsigned Half = VT.getVectorNumElements() / 2;
4054 WhichResult = (M[0] == 0 ? 0 : 1);
4055 for (unsigned j = 0; j != 2; ++j) {
4056 unsigned Idx = WhichResult;
4057 for (unsigned i = 0; i != Half; ++i) {
4058 int MIdx = M[i + j * Half];
4059 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4065 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4066 if (VT.is64BitVector() && EltSz == 32)
4072 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4073 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4077 unsigned NumElts = VT.getVectorNumElements();
4078 WhichResult = (M[0] == 0 ? 0 : 1);
4079 unsigned Idx = WhichResult * NumElts / 2;
4080 for (unsigned i = 0; i != NumElts; i += 2) {
4081 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4082 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4087 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4088 if (VT.is64BitVector() && EltSz == 32)
4094 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4095 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4096 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4097 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4098 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4102 unsigned NumElts = VT.getVectorNumElements();
4103 WhichResult = (M[0] == 0 ? 0 : 1);
4104 unsigned Idx = WhichResult * NumElts / 2;
4105 for (unsigned i = 0; i != NumElts; i += 2) {
4106 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4107 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4112 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4113 if (VT.is64BitVector() && EltSz == 32)
4119 // If N is an integer constant that can be moved into a register in one
4120 // instruction, return an SDValue of such a constant (will become a MOV
4121 // instruction). Otherwise return null.
4122 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4123 const ARMSubtarget *ST, DebugLoc dl) {
4125 if (!isa<ConstantSDNode>(N))
4127 Val = cast<ConstantSDNode>(N)->getZExtValue();
4129 if (ST->isThumb1Only()) {
4130 if (Val <= 255 || ~Val <= 255)
4131 return DAG.getConstant(Val, MVT::i32);
4133 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4134 return DAG.getConstant(Val, MVT::i32);
4139 // If this is a case we can't handle, return null and let the default
4140 // expansion code take care of it.
4141 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4142 const ARMSubtarget *ST) const {
4143 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4144 DebugLoc dl = Op.getDebugLoc();
4145 EVT VT = Op.getValueType();
4147 APInt SplatBits, SplatUndef;
4148 unsigned SplatBitSize;
4150 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4151 if (SplatBitSize <= 64) {
4152 // Check if an immediate VMOV works.
4154 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4155 SplatUndef.getZExtValue(), SplatBitSize,
4156 DAG, VmovVT, VT.is128BitVector(),
4158 if (Val.getNode()) {
4159 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4160 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4163 // Try an immediate VMVN.
4164 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4165 Val = isNEONModifiedImm(NegatedImm,
4166 SplatUndef.getZExtValue(), SplatBitSize,
4167 DAG, VmovVT, VT.is128BitVector(),
4169 if (Val.getNode()) {
4170 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4171 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4174 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4175 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4176 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4178 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4179 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4185 // Scan through the operands to see if only one value is used.
4187 // As an optimisation, even if more than one value is used it may be more
4188 // profitable to splat with one value then change some lanes.
4190 // Heuristically we decide to do this if the vector has a "dominant" value,
4191 // defined as splatted to more than half of the lanes.
4192 unsigned NumElts = VT.getVectorNumElements();
4193 bool isOnlyLowElement = true;
4194 bool usesOnlyOneValue = true;
4195 bool hasDominantValue = false;
4196 bool isConstant = true;
4198 // Map of the number of times a particular SDValue appears in the
4200 DenseMap<SDValue, unsigned> ValueCounts;
4202 for (unsigned i = 0; i < NumElts; ++i) {
4203 SDValue V = Op.getOperand(i);
4204 if (V.getOpcode() == ISD::UNDEF)
4207 isOnlyLowElement = false;
4208 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4211 ValueCounts.insert(std::make_pair(V, 0));
4212 unsigned &Count = ValueCounts[V];
4214 // Is this value dominant? (takes up more than half of the lanes)
4215 if (++Count > (NumElts / 2)) {
4216 hasDominantValue = true;
4220 if (ValueCounts.size() != 1)
4221 usesOnlyOneValue = false;
4222 if (!Value.getNode() && ValueCounts.size() > 0)
4223 Value = ValueCounts.begin()->first;
4225 if (ValueCounts.size() == 0)
4226 return DAG.getUNDEF(VT);
4228 if (isOnlyLowElement)
4229 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4231 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4233 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4234 // i32 and try again.
4235 if (hasDominantValue && EltSize <= 32) {
4239 // If we are VDUPing a value that comes directly from a vector, that will
4240 // cause an unnecessary move to and from a GPR, where instead we could
4241 // just use VDUPLANE.
4242 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4243 // We need to create a new undef vector to use for the VDUPLANE if the
4244 // size of the vector from which we get the value is different than the
4245 // size of the vector that we need to create. We will insert the element
4246 // such that the register coalescer will remove unnecessary copies.
4247 if (VT != Value->getOperand(0).getValueType()) {
4248 ConstantSDNode *constIndex;
4249 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4250 assert(constIndex && "The index is not a constant!");
4251 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4252 VT.getVectorNumElements();
4253 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4254 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4255 Value, DAG.getConstant(index, MVT::i32)),
4256 DAG.getConstant(index, MVT::i32));
4258 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4259 Value->getOperand(0), Value->getOperand(1));
4263 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4265 if (!usesOnlyOneValue) {
4266 // The dominant value was splatted as 'N', but we now have to insert
4267 // all differing elements.
4268 for (unsigned I = 0; I < NumElts; ++I) {
4269 if (Op.getOperand(I) == Value)
4271 SmallVector<SDValue, 3> Ops;
4273 Ops.push_back(Op.getOperand(I));
4274 Ops.push_back(DAG.getConstant(I, MVT::i32));
4275 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4280 if (VT.getVectorElementType().isFloatingPoint()) {
4281 SmallVector<SDValue, 8> Ops;
4282 for (unsigned i = 0; i < NumElts; ++i)
4283 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4285 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4286 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4287 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4289 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4291 if (usesOnlyOneValue) {
4292 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4293 if (isConstant && Val.getNode())
4294 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4298 // If all elements are constants and the case above didn't get hit, fall back
4299 // to the default expansion, which will generate a load from the constant
4304 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4306 SDValue shuffle = ReconstructShuffle(Op, DAG);
4307 if (shuffle != SDValue())
4311 // Vectors with 32- or 64-bit elements can be built by directly assigning
4312 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4313 // will be legalized.
4314 if (EltSize >= 32) {
4315 // Do the expansion with floating-point types, since that is what the VFP
4316 // registers are defined to use, and since i64 is not legal.
4317 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4318 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4319 SmallVector<SDValue, 8> Ops;
4320 for (unsigned i = 0; i < NumElts; ++i)
4321 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4322 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4323 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4329 // Gather data to see if the operation can be modelled as a
4330 // shuffle in combination with VEXTs.
4331 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4332 SelectionDAG &DAG) const {
4333 DebugLoc dl = Op.getDebugLoc();
4334 EVT VT = Op.getValueType();
4335 unsigned NumElts = VT.getVectorNumElements();
4337 SmallVector<SDValue, 2> SourceVecs;
4338 SmallVector<unsigned, 2> MinElts;
4339 SmallVector<unsigned, 2> MaxElts;
4341 for (unsigned i = 0; i < NumElts; ++i) {
4342 SDValue V = Op.getOperand(i);
4343 if (V.getOpcode() == ISD::UNDEF)
4345 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4346 // A shuffle can only come from building a vector from various
4347 // elements of other vectors.
4349 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4350 VT.getVectorElementType()) {
4351 // This code doesn't know how to handle shuffles where the vector
4352 // element types do not match (this happens because type legalization
4353 // promotes the return type of EXTRACT_VECTOR_ELT).
4354 // FIXME: It might be appropriate to extend this code to handle
4355 // mismatched types.
4359 // Record this extraction against the appropriate vector if possible...
4360 SDValue SourceVec = V.getOperand(0);
4361 // If the element number isn't a constant, we can't effectively
4362 // analyze what's going on.
4363 if (!isa<ConstantSDNode>(V.getOperand(1)))
4365 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4366 bool FoundSource = false;
4367 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4368 if (SourceVecs[j] == SourceVec) {
4369 if (MinElts[j] > EltNo)
4371 if (MaxElts[j] < EltNo)
4378 // Or record a new source if not...
4380 SourceVecs.push_back(SourceVec);
4381 MinElts.push_back(EltNo);
4382 MaxElts.push_back(EltNo);
4386 // Currently only do something sane when at most two source vectors
4388 if (SourceVecs.size() > 2)
4391 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4392 int VEXTOffsets[2] = {0, 0};
4394 // This loop extracts the usage patterns of the source vectors
4395 // and prepares appropriate SDValues for a shuffle if possible.
4396 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4397 if (SourceVecs[i].getValueType() == VT) {
4398 // No VEXT necessary
4399 ShuffleSrcs[i] = SourceVecs[i];
4402 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4403 // It probably isn't worth padding out a smaller vector just to
4404 // break it down again in a shuffle.
4408 // Since only 64-bit and 128-bit vectors are legal on ARM and
4409 // we've eliminated the other cases...
4410 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4411 "unexpected vector sizes in ReconstructShuffle");
4413 if (MaxElts[i] - MinElts[i] >= NumElts) {
4414 // Span too large for a VEXT to cope
4418 if (MinElts[i] >= NumElts) {
4419 // The extraction can just take the second half
4420 VEXTOffsets[i] = NumElts;
4421 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4423 DAG.getIntPtrConstant(NumElts));
4424 } else if (MaxElts[i] < NumElts) {
4425 // The extraction can just take the first half
4427 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4429 DAG.getIntPtrConstant(0));
4431 // An actual VEXT is needed
4432 VEXTOffsets[i] = MinElts[i];
4433 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4435 DAG.getIntPtrConstant(0));
4436 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4438 DAG.getIntPtrConstant(NumElts));
4439 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4440 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4444 SmallVector<int, 8> Mask;
4446 for (unsigned i = 0; i < NumElts; ++i) {
4447 SDValue Entry = Op.getOperand(i);
4448 if (Entry.getOpcode() == ISD::UNDEF) {
4453 SDValue ExtractVec = Entry.getOperand(0);
4454 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4455 .getOperand(1))->getSExtValue();
4456 if (ExtractVec == SourceVecs[0]) {
4457 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4459 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4463 // Final check before we try to produce nonsense...
4464 if (isShuffleMaskLegal(Mask, VT))
4465 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4471 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4472 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4473 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4474 /// are assumed to be legal.
4476 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4478 if (VT.getVectorNumElements() == 4 &&
4479 (VT.is128BitVector() || VT.is64BitVector())) {
4480 unsigned PFIndexes[4];
4481 for (unsigned i = 0; i != 4; ++i) {
4485 PFIndexes[i] = M[i];
4488 // Compute the index in the perfect shuffle table.
4489 unsigned PFTableIndex =
4490 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4491 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4492 unsigned Cost = (PFEntry >> 30);
4499 unsigned Imm, WhichResult;
4501 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4502 return (EltSize >= 32 ||
4503 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4504 isVREVMask(M, VT, 64) ||
4505 isVREVMask(M, VT, 32) ||
4506 isVREVMask(M, VT, 16) ||
4507 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4508 isVTBLMask(M, VT) ||
4509 isVTRNMask(M, VT, WhichResult) ||
4510 isVUZPMask(M, VT, WhichResult) ||
4511 isVZIPMask(M, VT, WhichResult) ||
4512 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4513 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4514 isVZIP_v_undef_Mask(M, VT, WhichResult));
4517 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4518 /// the specified operations to build the shuffle.
4519 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4520 SDValue RHS, SelectionDAG &DAG,
4522 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4523 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4524 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4527 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4536 OP_VUZPL, // VUZP, left result
4537 OP_VUZPR, // VUZP, right result
4538 OP_VZIPL, // VZIP, left result
4539 OP_VZIPR, // VZIP, right result
4540 OP_VTRNL, // VTRN, left result
4541 OP_VTRNR // VTRN, right result
4544 if (OpNum == OP_COPY) {
4545 if (LHSID == (1*9+2)*9+3) return LHS;
4546 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4550 SDValue OpLHS, OpRHS;
4551 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4552 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4553 EVT VT = OpLHS.getValueType();
4556 default: llvm_unreachable("Unknown shuffle opcode!");
4558 // VREV divides the vector in half and swaps within the half.
4559 if (VT.getVectorElementType() == MVT::i32 ||
4560 VT.getVectorElementType() == MVT::f32)
4561 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4562 // vrev <4 x i16> -> VREV32
4563 if (VT.getVectorElementType() == MVT::i16)
4564 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4565 // vrev <4 x i8> -> VREV16
4566 assert(VT.getVectorElementType() == MVT::i8);
4567 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4572 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4573 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4577 return DAG.getNode(ARMISD::VEXT, dl, VT,
4579 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4582 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4583 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4586 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4587 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4590 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4591 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4595 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4596 ArrayRef<int> ShuffleMask,
4597 SelectionDAG &DAG) {
4598 // Check to see if we can use the VTBL instruction.
4599 SDValue V1 = Op.getOperand(0);
4600 SDValue V2 = Op.getOperand(1);
4601 DebugLoc DL = Op.getDebugLoc();
4603 SmallVector<SDValue, 8> VTBLMask;
4604 for (ArrayRef<int>::iterator
4605 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4606 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4608 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4609 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4610 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4613 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4614 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4618 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4619 SDValue V1 = Op.getOperand(0);
4620 SDValue V2 = Op.getOperand(1);
4621 DebugLoc dl = Op.getDebugLoc();
4622 EVT VT = Op.getValueType();
4623 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4625 // Convert shuffles that are directly supported on NEON to target-specific
4626 // DAG nodes, instead of keeping them as shuffles and matching them again
4627 // during code selection. This is more efficient and avoids the possibility
4628 // of inconsistencies between legalization and selection.
4629 // FIXME: floating-point vectors should be canonicalized to integer vectors
4630 // of the same time so that they get CSEd properly.
4631 ArrayRef<int> ShuffleMask = SVN->getMask();
4633 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4634 if (EltSize <= 32) {
4635 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4636 int Lane = SVN->getSplatIndex();
4637 // If this is undef splat, generate it via "just" vdup, if possible.
4638 if (Lane == -1) Lane = 0;
4640 // Test if V1 is a SCALAR_TO_VECTOR.
4641 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4642 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4644 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4645 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4647 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4648 !isa<ConstantSDNode>(V1.getOperand(0))) {
4649 bool IsScalarToVector = true;
4650 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4651 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4652 IsScalarToVector = false;
4655 if (IsScalarToVector)
4656 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4658 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4659 DAG.getConstant(Lane, MVT::i32));
4664 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4667 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4668 DAG.getConstant(Imm, MVT::i32));
4671 if (isVREVMask(ShuffleMask, VT, 64))
4672 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4673 if (isVREVMask(ShuffleMask, VT, 32))
4674 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4675 if (isVREVMask(ShuffleMask, VT, 16))
4676 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4678 // Check for Neon shuffles that modify both input vectors in place.
4679 // If both results are used, i.e., if there are two shuffles with the same
4680 // source operands and with masks corresponding to both results of one of
4681 // these operations, DAG memoization will ensure that a single node is
4682 // used for both shuffles.
4683 unsigned WhichResult;
4684 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4685 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4686 V1, V2).getValue(WhichResult);
4687 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4688 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4689 V1, V2).getValue(WhichResult);
4690 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4691 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4692 V1, V2).getValue(WhichResult);
4694 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4695 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4696 V1, V1).getValue(WhichResult);
4697 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4698 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4699 V1, V1).getValue(WhichResult);
4700 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4701 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4702 V1, V1).getValue(WhichResult);
4705 // If the shuffle is not directly supported and it has 4 elements, use
4706 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4707 unsigned NumElts = VT.getVectorNumElements();
4709 unsigned PFIndexes[4];
4710 for (unsigned i = 0; i != 4; ++i) {
4711 if (ShuffleMask[i] < 0)
4714 PFIndexes[i] = ShuffleMask[i];
4717 // Compute the index in the perfect shuffle table.
4718 unsigned PFTableIndex =
4719 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4720 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4721 unsigned Cost = (PFEntry >> 30);
4724 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4727 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4728 if (EltSize >= 32) {
4729 // Do the expansion with floating-point types, since that is what the VFP
4730 // registers are defined to use, and since i64 is not legal.
4731 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4732 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4733 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4734 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4735 SmallVector<SDValue, 8> Ops;
4736 for (unsigned i = 0; i < NumElts; ++i) {
4737 if (ShuffleMask[i] < 0)
4738 Ops.push_back(DAG.getUNDEF(EltVT));
4740 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4741 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4742 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4745 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4746 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4749 if (VT == MVT::v8i8) {
4750 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4751 if (NewOp.getNode())
4758 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4759 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4760 SDValue Lane = Op.getOperand(2);
4761 if (!isa<ConstantSDNode>(Lane))
4767 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4768 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4769 SDValue Lane = Op.getOperand(1);
4770 if (!isa<ConstantSDNode>(Lane))
4773 SDValue Vec = Op.getOperand(0);
4774 if (Op.getValueType() == MVT::i32 &&
4775 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4776 DebugLoc dl = Op.getDebugLoc();
4777 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4783 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4784 // The only time a CONCAT_VECTORS operation can have legal types is when
4785 // two 64-bit vectors are concatenated to a 128-bit vector.
4786 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4787 "unexpected CONCAT_VECTORS");
4788 DebugLoc dl = Op.getDebugLoc();
4789 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4790 SDValue Op0 = Op.getOperand(0);
4791 SDValue Op1 = Op.getOperand(1);
4792 if (Op0.getOpcode() != ISD::UNDEF)
4793 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4794 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4795 DAG.getIntPtrConstant(0));
4796 if (Op1.getOpcode() != ISD::UNDEF)
4797 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4798 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4799 DAG.getIntPtrConstant(1));
4800 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4803 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4804 /// element has been zero/sign-extended, depending on the isSigned parameter,
4805 /// from an integer type half its size.
4806 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4808 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4809 EVT VT = N->getValueType(0);
4810 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4811 SDNode *BVN = N->getOperand(0).getNode();
4812 if (BVN->getValueType(0) != MVT::v4i32 ||
4813 BVN->getOpcode() != ISD::BUILD_VECTOR)
4815 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4816 unsigned HiElt = 1 - LoElt;
4817 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4818 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4819 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4820 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4821 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4824 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4825 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4828 if (Hi0->isNullValue() && Hi1->isNullValue())
4834 if (N->getOpcode() != ISD::BUILD_VECTOR)
4837 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4838 SDNode *Elt = N->getOperand(i).getNode();
4839 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4840 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4841 unsigned HalfSize = EltSize / 2;
4843 if (!isIntN(HalfSize, C->getSExtValue()))
4846 if (!isUIntN(HalfSize, C->getZExtValue()))
4857 /// isSignExtended - Check if a node is a vector value that is sign-extended
4858 /// or a constant BUILD_VECTOR with sign-extended elements.
4859 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4860 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4862 if (isExtendedBUILD_VECTOR(N, DAG, true))
4867 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4868 /// or a constant BUILD_VECTOR with zero-extended elements.
4869 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4870 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4872 if (isExtendedBUILD_VECTOR(N, DAG, false))
4877 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4878 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4879 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4880 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4881 return N->getOperand(0);
4882 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4883 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4884 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4885 LD->isNonTemporal(), LD->isInvariant(),
4886 LD->getAlignment());
4887 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4888 // have been legalized as a BITCAST from v4i32.
4889 if (N->getOpcode() == ISD::BITCAST) {
4890 SDNode *BVN = N->getOperand(0).getNode();
4891 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4892 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4893 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4894 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4895 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4897 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4898 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4899 EVT VT = N->getValueType(0);
4900 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4901 unsigned NumElts = VT.getVectorNumElements();
4902 MVT TruncVT = MVT::getIntegerVT(EltSize);
4903 SmallVector<SDValue, 8> Ops;
4904 for (unsigned i = 0; i != NumElts; ++i) {
4905 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4906 const APInt &CInt = C->getAPIntValue();
4907 // Element types smaller than 32 bits are not legal, so use i32 elements.
4908 // The values are implicitly truncated so sext vs. zext doesn't matter.
4909 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4911 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4912 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4915 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4916 unsigned Opcode = N->getOpcode();
4917 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4918 SDNode *N0 = N->getOperand(0).getNode();
4919 SDNode *N1 = N->getOperand(1).getNode();
4920 return N0->hasOneUse() && N1->hasOneUse() &&
4921 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4926 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4927 unsigned Opcode = N->getOpcode();
4928 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4929 SDNode *N0 = N->getOperand(0).getNode();
4930 SDNode *N1 = N->getOperand(1).getNode();
4931 return N0->hasOneUse() && N1->hasOneUse() &&
4932 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4937 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4938 // Multiplications are only custom-lowered for 128-bit vectors so that
4939 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4940 EVT VT = Op.getValueType();
4941 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4942 SDNode *N0 = Op.getOperand(0).getNode();
4943 SDNode *N1 = Op.getOperand(1).getNode();
4944 unsigned NewOpc = 0;
4946 bool isN0SExt = isSignExtended(N0, DAG);
4947 bool isN1SExt = isSignExtended(N1, DAG);
4948 if (isN0SExt && isN1SExt)
4949 NewOpc = ARMISD::VMULLs;
4951 bool isN0ZExt = isZeroExtended(N0, DAG);
4952 bool isN1ZExt = isZeroExtended(N1, DAG);
4953 if (isN0ZExt && isN1ZExt)
4954 NewOpc = ARMISD::VMULLu;
4955 else if (isN1SExt || isN1ZExt) {
4956 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4957 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4958 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4959 NewOpc = ARMISD::VMULLs;
4961 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4962 NewOpc = ARMISD::VMULLu;
4964 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4966 NewOpc = ARMISD::VMULLu;
4972 if (VT == MVT::v2i64)
4973 // Fall through to expand this. It is not legal.
4976 // Other vector multiplications are legal.
4981 // Legalize to a VMULL instruction.
4982 DebugLoc DL = Op.getDebugLoc();
4984 SDValue Op1 = SkipExtension(N1, DAG);
4986 Op0 = SkipExtension(N0, DAG);
4987 assert(Op0.getValueType().is64BitVector() &&
4988 Op1.getValueType().is64BitVector() &&
4989 "unexpected types for extended operands to VMULL");
4990 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4993 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4994 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5001 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5002 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5003 EVT Op1VT = Op1.getValueType();
5004 return DAG.getNode(N0->getOpcode(), DL, VT,
5005 DAG.getNode(NewOpc, DL, VT,
5006 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5007 DAG.getNode(NewOpc, DL, VT,
5008 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5012 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5014 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5015 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5016 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5017 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5018 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5019 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5020 // Get reciprocal estimate.
5021 // float4 recip = vrecpeq_f32(yf);
5022 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5023 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5024 // Because char has a smaller range than uchar, we can actually get away
5025 // without any newton steps. This requires that we use a weird bias
5026 // of 0xb000, however (again, this has been exhaustively tested).
5027 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5028 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5029 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5030 Y = DAG.getConstant(0xb000, MVT::i32);
5031 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5032 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5033 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5034 // Convert back to short.
5035 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5036 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5041 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5043 // Convert to float.
5044 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5045 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5046 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5047 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5048 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5049 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5051 // Use reciprocal estimate and one refinement step.
5052 // float4 recip = vrecpeq_f32(yf);
5053 // recip *= vrecpsq_f32(yf, recip);
5054 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5055 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5056 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5057 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5059 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5060 // Because short has a smaller range than ushort, we can actually get away
5061 // with only a single newton step. This requires that we use a weird bias
5062 // of 89, however (again, this has been exhaustively tested).
5063 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5064 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5065 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5066 N1 = DAG.getConstant(0x89, MVT::i32);
5067 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5068 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5069 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5070 // Convert back to integer and return.
5071 // return vmovn_s32(vcvt_s32_f32(result));
5072 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5073 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5077 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5078 EVT VT = Op.getValueType();
5079 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5080 "unexpected type for custom-lowering ISD::SDIV");
5082 DebugLoc dl = Op.getDebugLoc();
5083 SDValue N0 = Op.getOperand(0);
5084 SDValue N1 = Op.getOperand(1);
5087 if (VT == MVT::v8i8) {
5088 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5089 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5091 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5092 DAG.getIntPtrConstant(4));
5093 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5094 DAG.getIntPtrConstant(4));
5095 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5096 DAG.getIntPtrConstant(0));
5097 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5098 DAG.getIntPtrConstant(0));
5100 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5101 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5103 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5104 N0 = LowerCONCAT_VECTORS(N0, DAG);
5106 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5109 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5112 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5113 EVT VT = Op.getValueType();
5114 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5115 "unexpected type for custom-lowering ISD::UDIV");
5117 DebugLoc dl = Op.getDebugLoc();
5118 SDValue N0 = Op.getOperand(0);
5119 SDValue N1 = Op.getOperand(1);
5122 if (VT == MVT::v8i8) {
5123 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5124 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5126 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5127 DAG.getIntPtrConstant(4));
5128 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5129 DAG.getIntPtrConstant(4));
5130 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5131 DAG.getIntPtrConstant(0));
5132 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5133 DAG.getIntPtrConstant(0));
5135 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5136 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5138 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5139 N0 = LowerCONCAT_VECTORS(N0, DAG);
5141 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5142 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5147 // v4i16 sdiv ... Convert to float.
5148 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5149 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5150 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5151 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5152 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5153 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5155 // Use reciprocal estimate and two refinement steps.
5156 // float4 recip = vrecpeq_f32(yf);
5157 // recip *= vrecpsq_f32(yf, recip);
5158 // recip *= vrecpsq_f32(yf, recip);
5159 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5160 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5161 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5162 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5164 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5165 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5166 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5168 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5169 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5170 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5171 // and that it will never cause us to return an answer too large).
5172 // float4 result = as_float4(as_int4(xf*recip) + 2);
5173 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5174 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5175 N1 = DAG.getConstant(2, MVT::i32);
5176 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5177 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5178 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5179 // Convert back to integer and return.
5180 // return vmovn_u32(vcvt_s32_f32(result));
5181 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5182 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5186 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5187 EVT VT = Op.getNode()->getValueType(0);
5188 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5191 bool ExtraOp = false;
5192 switch (Op.getOpcode()) {
5193 default: llvm_unreachable("Invalid code");
5194 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5195 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5196 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5197 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5201 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5203 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5204 Op.getOperand(1), Op.getOperand(2));
5207 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5208 // Monotonic load/store is legal for all targets
5209 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5212 // Aquire/Release load/store is not legal for targets without a
5213 // dmb or equivalent available.
5219 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5220 SelectionDAG &DAG, unsigned NewOp) {
5221 DebugLoc dl = Node->getDebugLoc();
5222 assert (Node->getValueType(0) == MVT::i64 &&
5223 "Only know how to expand i64 atomics");
5225 SmallVector<SDValue, 6> Ops;
5226 Ops.push_back(Node->getOperand(0)); // Chain
5227 Ops.push_back(Node->getOperand(1)); // Ptr
5229 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5230 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5231 // High part of Val1
5232 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5233 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5234 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5235 // High part of Val1
5236 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5237 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5238 // High part of Val2
5239 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5240 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5242 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5244 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5245 cast<MemSDNode>(Node)->getMemOperand());
5246 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5247 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5248 Results.push_back(Result.getValue(2));
5251 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5252 switch (Op.getOpcode()) {
5253 default: llvm_unreachable("Don't know how to custom lower this!");
5254 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5255 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5256 case ISD::GlobalAddress:
5257 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5258 LowerGlobalAddressELF(Op, DAG);
5259 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5260 case ISD::SELECT: return LowerSELECT(Op, DAG);
5261 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5262 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5263 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5264 case ISD::VASTART: return LowerVASTART(Op, DAG);
5265 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5266 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5267 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5268 case ISD::SINT_TO_FP:
5269 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5270 case ISD::FP_TO_SINT:
5271 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5272 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5273 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5274 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5275 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5276 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5277 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5278 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5280 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5283 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5284 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5285 case ISD::SRL_PARTS:
5286 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5287 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5288 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5289 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5290 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5291 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5292 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5293 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5294 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5295 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5296 case ISD::MUL: return LowerMUL(Op, DAG);
5297 case ISD::SDIV: return LowerSDIV(Op, DAG);
5298 case ISD::UDIV: return LowerUDIV(Op, DAG);
5302 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5303 case ISD::ATOMIC_LOAD:
5304 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5308 /// ReplaceNodeResults - Replace the results of node with an illegal result
5309 /// type with new values built out of custom code.
5310 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5311 SmallVectorImpl<SDValue>&Results,
5312 SelectionDAG &DAG) const {
5314 switch (N->getOpcode()) {
5316 llvm_unreachable("Don't know how to custom expand this!");
5318 Res = ExpandBITCAST(N, DAG);
5322 Res = Expand64BitShift(N, DAG, Subtarget);
5324 case ISD::ATOMIC_LOAD_ADD:
5325 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5327 case ISD::ATOMIC_LOAD_AND:
5328 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5330 case ISD::ATOMIC_LOAD_NAND:
5331 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5333 case ISD::ATOMIC_LOAD_OR:
5334 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5336 case ISD::ATOMIC_LOAD_SUB:
5337 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5339 case ISD::ATOMIC_LOAD_XOR:
5340 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5342 case ISD::ATOMIC_SWAP:
5343 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5345 case ISD::ATOMIC_CMP_SWAP:
5346 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5350 Results.push_back(Res);
5353 //===----------------------------------------------------------------------===//
5354 // ARM Scheduler Hooks
5355 //===----------------------------------------------------------------------===//
5358 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5359 MachineBasicBlock *BB,
5360 unsigned Size) const {
5361 unsigned dest = MI->getOperand(0).getReg();
5362 unsigned ptr = MI->getOperand(1).getReg();
5363 unsigned oldval = MI->getOperand(2).getReg();
5364 unsigned newval = MI->getOperand(3).getReg();
5365 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5366 DebugLoc dl = MI->getDebugLoc();
5367 bool isThumb2 = Subtarget->isThumb2();
5369 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5370 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5371 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5372 (const TargetRegisterClass*)&ARM::GPRRegClass);
5375 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5376 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5377 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5380 unsigned ldrOpc, strOpc;
5382 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5384 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5385 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5388 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5389 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5392 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5393 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5397 MachineFunction *MF = BB->getParent();
5398 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5399 MachineFunction::iterator It = BB;
5400 ++It; // insert the new blocks after the current block
5402 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5403 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5404 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5405 MF->insert(It, loop1MBB);
5406 MF->insert(It, loop2MBB);
5407 MF->insert(It, exitMBB);
5409 // Transfer the remainder of BB and its successor edges to exitMBB.
5410 exitMBB->splice(exitMBB->begin(), BB,
5411 llvm::next(MachineBasicBlock::iterator(MI)),
5413 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5417 // fallthrough --> loop1MBB
5418 BB->addSuccessor(loop1MBB);
5421 // ldrex dest, [ptr]
5425 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5426 if (ldrOpc == ARM::t2LDREX)
5428 AddDefaultPred(MIB);
5429 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5430 .addReg(dest).addReg(oldval));
5431 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5432 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5433 BB->addSuccessor(loop2MBB);
5434 BB->addSuccessor(exitMBB);
5437 // strex scratch, newval, [ptr]
5441 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5442 if (strOpc == ARM::t2STREX)
5444 AddDefaultPred(MIB);
5445 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5446 .addReg(scratch).addImm(0));
5447 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5448 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5449 BB->addSuccessor(loop1MBB);
5450 BB->addSuccessor(exitMBB);
5456 MI->eraseFromParent(); // The instruction is gone now.
5462 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5463 unsigned Size, unsigned BinOpcode) const {
5464 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5465 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5467 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5468 MachineFunction *MF = BB->getParent();
5469 MachineFunction::iterator It = BB;
5472 unsigned dest = MI->getOperand(0).getReg();
5473 unsigned ptr = MI->getOperand(1).getReg();
5474 unsigned incr = MI->getOperand(2).getReg();
5475 DebugLoc dl = MI->getDebugLoc();
5476 bool isThumb2 = Subtarget->isThumb2();
5478 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5480 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5481 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5484 unsigned ldrOpc, strOpc;
5486 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5488 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5489 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5492 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5493 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5496 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5497 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5501 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5502 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5503 MF->insert(It, loopMBB);
5504 MF->insert(It, exitMBB);
5506 // Transfer the remainder of BB and its successor edges to exitMBB.
5507 exitMBB->splice(exitMBB->begin(), BB,
5508 llvm::next(MachineBasicBlock::iterator(MI)),
5510 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5512 const TargetRegisterClass *TRC = isThumb2 ?
5513 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5514 (const TargetRegisterClass*)&ARM::GPRRegClass;
5515 unsigned scratch = MRI.createVirtualRegister(TRC);
5516 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5520 // fallthrough --> loopMBB
5521 BB->addSuccessor(loopMBB);
5525 // <binop> scratch2, dest, incr
5526 // strex scratch, scratch2, ptr
5529 // fallthrough --> exitMBB
5531 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5532 if (ldrOpc == ARM::t2LDREX)
5534 AddDefaultPred(MIB);
5536 // operand order needs to go the other way for NAND
5537 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5538 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5539 addReg(incr).addReg(dest)).addReg(0);
5541 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5542 addReg(dest).addReg(incr)).addReg(0);
5545 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5546 if (strOpc == ARM::t2STREX)
5548 AddDefaultPred(MIB);
5549 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5550 .addReg(scratch).addImm(0));
5551 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5552 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5554 BB->addSuccessor(loopMBB);
5555 BB->addSuccessor(exitMBB);
5561 MI->eraseFromParent(); // The instruction is gone now.
5567 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5568 MachineBasicBlock *BB,
5571 ARMCC::CondCodes Cond) const {
5572 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5574 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5575 MachineFunction *MF = BB->getParent();
5576 MachineFunction::iterator It = BB;
5579 unsigned dest = MI->getOperand(0).getReg();
5580 unsigned ptr = MI->getOperand(1).getReg();
5581 unsigned incr = MI->getOperand(2).getReg();
5582 unsigned oldval = dest;
5583 DebugLoc dl = MI->getDebugLoc();
5584 bool isThumb2 = Subtarget->isThumb2();
5586 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5588 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5589 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5592 unsigned ldrOpc, strOpc, extendOpc;
5594 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5596 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5597 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5598 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5601 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5602 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5603 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5606 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5607 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5612 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5613 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5614 MF->insert(It, loopMBB);
5615 MF->insert(It, exitMBB);
5617 // Transfer the remainder of BB and its successor edges to exitMBB.
5618 exitMBB->splice(exitMBB->begin(), BB,
5619 llvm::next(MachineBasicBlock::iterator(MI)),
5621 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5623 const TargetRegisterClass *TRC = isThumb2 ?
5624 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5625 (const TargetRegisterClass*)&ARM::GPRRegClass;
5626 unsigned scratch = MRI.createVirtualRegister(TRC);
5627 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5631 // fallthrough --> loopMBB
5632 BB->addSuccessor(loopMBB);
5636 // (sign extend dest, if required)
5638 // cmov.cond scratch2, incr, dest
5639 // strex scratch, scratch2, ptr
5642 // fallthrough --> exitMBB
5644 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5645 if (ldrOpc == ARM::t2LDREX)
5647 AddDefaultPred(MIB);
5649 // Sign extend the value, if necessary.
5650 if (signExtend && extendOpc) {
5651 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5652 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5657 // Build compare and cmov instructions.
5658 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5659 .addReg(oldval).addReg(incr));
5660 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5661 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5663 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5664 if (strOpc == ARM::t2STREX)
5666 AddDefaultPred(MIB);
5667 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5668 .addReg(scratch).addImm(0));
5669 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5670 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5672 BB->addSuccessor(loopMBB);
5673 BB->addSuccessor(exitMBB);
5679 MI->eraseFromParent(); // The instruction is gone now.
5685 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5686 unsigned Op1, unsigned Op2,
5687 bool NeedsCarry, bool IsCmpxchg) const {
5688 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5689 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5691 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5692 MachineFunction *MF = BB->getParent();
5693 MachineFunction::iterator It = BB;
5696 unsigned destlo = MI->getOperand(0).getReg();
5697 unsigned desthi = MI->getOperand(1).getReg();
5698 unsigned ptr = MI->getOperand(2).getReg();
5699 unsigned vallo = MI->getOperand(3).getReg();
5700 unsigned valhi = MI->getOperand(4).getReg();
5701 DebugLoc dl = MI->getDebugLoc();
5702 bool isThumb2 = Subtarget->isThumb2();
5704 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5706 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5707 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5708 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5711 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5712 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5714 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5715 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5717 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5718 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5720 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5721 MF->insert(It, loopMBB);
5723 MF->insert(It, contBB);
5724 MF->insert(It, cont2BB);
5726 MF->insert(It, exitMBB);
5728 // Transfer the remainder of BB and its successor edges to exitMBB.
5729 exitMBB->splice(exitMBB->begin(), BB,
5730 llvm::next(MachineBasicBlock::iterator(MI)),
5732 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5734 const TargetRegisterClass *TRC = isThumb2 ?
5735 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5736 (const TargetRegisterClass*)&ARM::GPRRegClass;
5737 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5741 // fallthrough --> loopMBB
5742 BB->addSuccessor(loopMBB);
5745 // ldrexd r2, r3, ptr
5746 // <binopa> r0, r2, incr
5747 // <binopb> r1, r3, incr
5748 // strexd storesuccess, r0, r1, ptr
5749 // cmp storesuccess, #0
5751 // fallthrough --> exitMBB
5753 // Note that the registers are explicitly specified because there is not any
5754 // way to force the register allocator to allocate a register pair.
5756 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5757 // need to properly enforce the restriction that the two output registers
5758 // for ldrexd must be different.
5761 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5762 .addReg(ARM::R2, RegState::Define)
5763 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5764 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5765 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5766 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5770 for (unsigned i = 0; i < 2; i++) {
5771 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5773 .addReg(i == 0 ? destlo : desthi)
5774 .addReg(i == 0 ? vallo : valhi));
5775 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5776 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5777 BB->addSuccessor(exitMBB);
5778 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5779 BB = (i == 0 ? contBB : cont2BB);
5782 // Copy to physregs for strexd
5783 unsigned setlo = MI->getOperand(5).getReg();
5784 unsigned sethi = MI->getOperand(6).getReg();
5785 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5786 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5788 // Perform binary operation
5789 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5790 .addReg(destlo).addReg(vallo))
5791 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5792 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5793 .addReg(desthi).addReg(valhi)).addReg(0);
5795 // Copy to physregs for strexd
5796 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5797 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5801 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5802 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5804 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5805 .addReg(storesuccess).addImm(0));
5806 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5807 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5809 BB->addSuccessor(loopMBB);
5810 BB->addSuccessor(exitMBB);
5816 MI->eraseFromParent(); // The instruction is gone now.
5821 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5822 /// registers the function context.
5823 void ARMTargetLowering::
5824 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5825 MachineBasicBlock *DispatchBB, int FI) const {
5826 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5827 DebugLoc dl = MI->getDebugLoc();
5828 MachineFunction *MF = MBB->getParent();
5829 MachineRegisterInfo *MRI = &MF->getRegInfo();
5830 MachineConstantPool *MCP = MF->getConstantPool();
5831 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5832 const Function *F = MF->getFunction();
5834 bool isThumb = Subtarget->isThumb();
5835 bool isThumb2 = Subtarget->isThumb2();
5837 unsigned PCLabelId = AFI->createPICLabelUId();
5838 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5839 ARMConstantPoolValue *CPV =
5840 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5841 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5843 const TargetRegisterClass *TRC = isThumb ?
5844 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5845 (const TargetRegisterClass*)&ARM::GPRRegClass;
5847 // Grab constant pool and fixed stack memory operands.
5848 MachineMemOperand *CPMMO =
5849 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5850 MachineMemOperand::MOLoad, 4, 4);
5852 MachineMemOperand *FIMMOSt =
5853 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5854 MachineMemOperand::MOStore, 4, 4);
5856 // Load the address of the dispatch MBB into the jump buffer.
5858 // Incoming value: jbuf
5859 // ldr.n r5, LCPI1_1
5862 // str r5, [$jbuf, #+4] ; &jbuf[1]
5863 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5864 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5865 .addConstantPoolIndex(CPI)
5866 .addMemOperand(CPMMO));
5867 // Set the low bit because of thumb mode.
5868 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5870 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5871 .addReg(NewVReg1, RegState::Kill)
5873 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5874 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5875 .addReg(NewVReg2, RegState::Kill)
5877 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5878 .addReg(NewVReg3, RegState::Kill)
5880 .addImm(36) // &jbuf[1] :: pc
5881 .addMemOperand(FIMMOSt));
5882 } else if (isThumb) {
5883 // Incoming value: jbuf
5884 // ldr.n r1, LCPI1_4
5888 // add r2, $jbuf, #+4 ; &jbuf[1]
5890 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5891 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5892 .addConstantPoolIndex(CPI)
5893 .addMemOperand(CPMMO));
5894 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5895 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5896 .addReg(NewVReg1, RegState::Kill)
5898 // Set the low bit because of thumb mode.
5899 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5900 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5901 .addReg(ARM::CPSR, RegState::Define)
5903 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5904 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5905 .addReg(ARM::CPSR, RegState::Define)
5906 .addReg(NewVReg2, RegState::Kill)
5907 .addReg(NewVReg3, RegState::Kill));
5908 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5909 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5911 .addImm(36)); // &jbuf[1] :: pc
5912 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5913 .addReg(NewVReg4, RegState::Kill)
5914 .addReg(NewVReg5, RegState::Kill)
5916 .addMemOperand(FIMMOSt));
5918 // Incoming value: jbuf
5921 // str r1, [$jbuf, #+4] ; &jbuf[1]
5922 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5923 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5924 .addConstantPoolIndex(CPI)
5926 .addMemOperand(CPMMO));
5927 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5928 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5929 .addReg(NewVReg1, RegState::Kill)
5930 .addImm(PCLabelId));
5931 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5932 .addReg(NewVReg2, RegState::Kill)
5934 .addImm(36) // &jbuf[1] :: pc
5935 .addMemOperand(FIMMOSt));
5939 MachineBasicBlock *ARMTargetLowering::
5940 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5941 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5942 DebugLoc dl = MI->getDebugLoc();
5943 MachineFunction *MF = MBB->getParent();
5944 MachineRegisterInfo *MRI = &MF->getRegInfo();
5945 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5946 MachineFrameInfo *MFI = MF->getFrameInfo();
5947 int FI = MFI->getFunctionContextIndex();
5949 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5950 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5951 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
5953 // Get a mapping of the call site numbers to all of the landing pads they're
5955 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5956 unsigned MaxCSNum = 0;
5957 MachineModuleInfo &MMI = MF->getMMI();
5958 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5960 if (!BB->isLandingPad()) continue;
5962 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5964 for (MachineBasicBlock::iterator
5965 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5966 if (!II->isEHLabel()) continue;
5968 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5969 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5971 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5972 for (SmallVectorImpl<unsigned>::iterator
5973 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5974 CSI != CSE; ++CSI) {
5975 CallSiteNumToLPad[*CSI].push_back(BB);
5976 MaxCSNum = std::max(MaxCSNum, *CSI);
5982 // Get an ordered list of the machine basic blocks for the jump table.
5983 std::vector<MachineBasicBlock*> LPadList;
5984 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5985 LPadList.reserve(CallSiteNumToLPad.size());
5986 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5987 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5988 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5989 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5990 LPadList.push_back(*II);
5991 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5995 assert(!LPadList.empty() &&
5996 "No landing pad destinations for the dispatch jump table!");
5998 // Create the jump table and associated information.
5999 MachineJumpTableInfo *JTI =
6000 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6001 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6002 unsigned UId = AFI->createJumpTableUId();
6004 // Create the MBBs for the dispatch code.
6006 // Shove the dispatch's address into the return slot in the function context.
6007 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6008 DispatchBB->setIsLandingPad();
6010 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6011 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6012 DispatchBB->addSuccessor(TrapBB);
6014 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6015 DispatchBB->addSuccessor(DispContBB);
6018 MF->insert(MF->end(), DispatchBB);
6019 MF->insert(MF->end(), DispContBB);
6020 MF->insert(MF->end(), TrapBB);
6022 // Insert code into the entry block that creates and registers the function
6024 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6026 MachineMemOperand *FIMMOLd =
6027 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6028 MachineMemOperand::MOLoad |
6029 MachineMemOperand::MOVolatile, 4, 4);
6031 if (AFI->isThumb1OnlyFunction())
6032 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
6033 else if (!Subtarget->hasVFP2())
6034 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
6036 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6038 unsigned NumLPads = LPadList.size();
6039 if (Subtarget->isThumb2()) {
6040 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6041 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6044 .addMemOperand(FIMMOLd));
6046 if (NumLPads < 256) {
6047 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6049 .addImm(LPadList.size()));
6051 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6052 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6053 .addImm(NumLPads & 0xFFFF));
6055 unsigned VReg2 = VReg1;
6056 if ((NumLPads & 0xFFFF0000) != 0) {
6057 VReg2 = MRI->createVirtualRegister(TRC);
6058 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6060 .addImm(NumLPads >> 16));
6063 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6068 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6073 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6074 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6075 .addJumpTableIndex(MJTI)
6078 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6081 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6082 .addReg(NewVReg3, RegState::Kill)
6084 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6086 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6087 .addReg(NewVReg4, RegState::Kill)
6089 .addJumpTableIndex(MJTI)
6091 } else if (Subtarget->isThumb()) {
6092 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6093 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6096 .addMemOperand(FIMMOLd));
6098 if (NumLPads < 256) {
6099 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6103 MachineConstantPool *ConstantPool = MF->getConstantPool();
6104 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6105 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6107 // MachineConstantPool wants an explicit alignment.
6108 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6110 Align = getDataLayout()->getTypeAllocSize(C->getType());
6111 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6113 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6114 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6115 .addReg(VReg1, RegState::Define)
6116 .addConstantPoolIndex(Idx));
6117 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6122 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6127 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6128 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6129 .addReg(ARM::CPSR, RegState::Define)
6133 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6134 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6135 .addJumpTableIndex(MJTI)
6138 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6139 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6140 .addReg(ARM::CPSR, RegState::Define)
6141 .addReg(NewVReg2, RegState::Kill)
6144 MachineMemOperand *JTMMOLd =
6145 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6146 MachineMemOperand::MOLoad, 4, 4);
6148 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6149 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6150 .addReg(NewVReg4, RegState::Kill)
6152 .addMemOperand(JTMMOLd));
6154 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6155 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6156 .addReg(ARM::CPSR, RegState::Define)
6157 .addReg(NewVReg5, RegState::Kill)
6160 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6161 .addReg(NewVReg6, RegState::Kill)
6162 .addJumpTableIndex(MJTI)
6165 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6166 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6169 .addMemOperand(FIMMOLd));
6171 if (NumLPads < 256) {
6172 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6175 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6176 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6177 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6178 .addImm(NumLPads & 0xFFFF));
6180 unsigned VReg2 = VReg1;
6181 if ((NumLPads & 0xFFFF0000) != 0) {
6182 VReg2 = MRI->createVirtualRegister(TRC);
6183 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6185 .addImm(NumLPads >> 16));
6188 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6192 MachineConstantPool *ConstantPool = MF->getConstantPool();
6193 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6194 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6196 // MachineConstantPool wants an explicit alignment.
6197 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6199 Align = getDataLayout()->getTypeAllocSize(C->getType());
6200 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6202 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6203 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6204 .addReg(VReg1, RegState::Define)
6205 .addConstantPoolIndex(Idx)
6207 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6209 .addReg(VReg1, RegState::Kill));
6212 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6217 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6219 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6221 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6222 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6223 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6224 .addJumpTableIndex(MJTI)
6227 MachineMemOperand *JTMMOLd =
6228 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6229 MachineMemOperand::MOLoad, 4, 4);
6230 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6232 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6233 .addReg(NewVReg3, RegState::Kill)
6236 .addMemOperand(JTMMOLd));
6238 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6239 .addReg(NewVReg5, RegState::Kill)
6241 .addJumpTableIndex(MJTI)
6245 // Add the jump table entries as successors to the MBB.
6246 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6247 for (std::vector<MachineBasicBlock*>::iterator
6248 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6249 MachineBasicBlock *CurMBB = *I;
6250 if (SeenMBBs.insert(CurMBB))
6251 DispContBB->addSuccessor(CurMBB);
6254 // N.B. the order the invoke BBs are processed in doesn't matter here.
6255 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6256 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6257 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6258 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6259 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6260 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6261 MachineBasicBlock *BB = *I;
6263 // Remove the landing pad successor from the invoke block and replace it
6264 // with the new dispatch block.
6265 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6267 while (!Successors.empty()) {
6268 MachineBasicBlock *SMBB = Successors.pop_back_val();
6269 if (SMBB->isLandingPad()) {
6270 BB->removeSuccessor(SMBB);
6271 MBBLPads.push_back(SMBB);
6275 BB->addSuccessor(DispatchBB);
6277 // Find the invoke call and mark all of the callee-saved registers as
6278 // 'implicit defined' so that they're spilled. This prevents code from
6279 // moving instructions to before the EH block, where they will never be
6281 for (MachineBasicBlock::reverse_iterator
6282 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6283 if (!II->isCall()) continue;
6285 DenseMap<unsigned, bool> DefRegs;
6286 for (MachineInstr::mop_iterator
6287 OI = II->operands_begin(), OE = II->operands_end();
6289 if (!OI->isReg()) continue;
6290 DefRegs[OI->getReg()] = true;
6293 MachineInstrBuilder MIB(&*II);
6295 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6296 unsigned Reg = SavedRegs[i];
6297 if (Subtarget->isThumb2() &&
6298 !ARM::tGPRRegClass.contains(Reg) &&
6299 !ARM::hGPRRegClass.contains(Reg))
6301 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6303 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6306 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6313 // Mark all former landing pads as non-landing pads. The dispatch is the only
6315 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6316 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6317 (*I)->setIsLandingPad(false);
6319 // The instruction is gone now.
6320 MI->eraseFromParent();
6326 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6327 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6328 E = MBB->succ_end(); I != E; ++I)
6331 llvm_unreachable("Expecting a BB with two successors!");
6334 MachineBasicBlock *ARMTargetLowering::
6335 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6336 // This pseudo instruction has 3 operands: dst, src, size
6337 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6338 // Otherwise, we will generate unrolled scalar copies.
6339 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6340 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6341 MachineFunction::iterator It = BB;
6344 unsigned dest = MI->getOperand(0).getReg();
6345 unsigned src = MI->getOperand(1).getReg();
6346 unsigned SizeVal = MI->getOperand(2).getImm();
6347 unsigned Align = MI->getOperand(3).getImm();
6348 DebugLoc dl = MI->getDebugLoc();
6350 bool isThumb2 = Subtarget->isThumb2();
6351 MachineFunction *MF = BB->getParent();
6352 MachineRegisterInfo &MRI = MF->getRegInfo();
6353 unsigned ldrOpc, strOpc, UnitSize = 0;
6355 const TargetRegisterClass *TRC = isThumb2 ?
6356 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6357 (const TargetRegisterClass*)&ARM::GPRRegClass;
6358 const TargetRegisterClass *TRC_Vec = 0;
6361 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6362 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6364 } else if (Align & 2) {
6365 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6366 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6369 // Check whether we can use NEON instructions.
6370 if (!MF->getFunction()->getFnAttributes().
6371 hasAttribute(Attributes::NoImplicitFloat) &&
6372 Subtarget->hasNEON()) {
6373 if ((Align % 16 == 0) && SizeVal >= 16) {
6374 ldrOpc = ARM::VLD1q32wb_fixed;
6375 strOpc = ARM::VST1q32wb_fixed;
6377 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6379 else if ((Align % 8 == 0) && SizeVal >= 8) {
6380 ldrOpc = ARM::VLD1d32wb_fixed;
6381 strOpc = ARM::VST1d32wb_fixed;
6383 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6386 // Can't use NEON instructions.
6387 if (UnitSize == 0) {
6388 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6389 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6394 unsigned BytesLeft = SizeVal % UnitSize;
6395 unsigned LoopSize = SizeVal - BytesLeft;
6397 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6398 // Use LDR and STR to copy.
6399 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6400 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6401 unsigned srcIn = src;
6402 unsigned destIn = dest;
6403 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6404 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6405 unsigned srcOut = MRI.createVirtualRegister(TRC);
6406 unsigned destOut = MRI.createVirtualRegister(TRC);
6407 if (UnitSize >= 8) {
6408 AddDefaultPred(BuildMI(*BB, MI, dl,
6409 TII->get(ldrOpc), scratch)
6410 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6412 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6413 .addReg(destIn).addImm(0).addReg(scratch));
6414 } else if (isThumb2) {
6415 AddDefaultPred(BuildMI(*BB, MI, dl,
6416 TII->get(ldrOpc), scratch)
6417 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6419 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6420 .addReg(scratch).addReg(destIn)
6423 AddDefaultPred(BuildMI(*BB, MI, dl,
6424 TII->get(ldrOpc), scratch)
6425 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6428 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6429 .addReg(scratch).addReg(destIn)
6430 .addReg(0).addImm(UnitSize));
6436 // Handle the leftover bytes with LDRB and STRB.
6437 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6438 // [destOut] = STRB_POST(scratch, destIn, 1)
6439 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6440 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6441 for (unsigned i = 0; i < BytesLeft; i++) {
6442 unsigned scratch = MRI.createVirtualRegister(TRC);
6443 unsigned srcOut = MRI.createVirtualRegister(TRC);
6444 unsigned destOut = MRI.createVirtualRegister(TRC);
6446 AddDefaultPred(BuildMI(*BB, MI, dl,
6447 TII->get(ldrOpc),scratch)
6448 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6450 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6451 .addReg(scratch).addReg(destIn)
6452 .addReg(0).addImm(1));
6454 AddDefaultPred(BuildMI(*BB, MI, dl,
6455 TII->get(ldrOpc),scratch)
6456 .addReg(srcOut, RegState::Define).addReg(srcIn)
6457 .addReg(0).addImm(1));
6459 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6460 .addReg(scratch).addReg(destIn)
6461 .addReg(0).addImm(1));
6466 MI->eraseFromParent(); // The instruction is gone now.
6470 // Expand the pseudo op to a loop.
6473 // movw varEnd, # --> with thumb2
6475 // ldrcp varEnd, idx --> without thumb2
6476 // fallthrough --> loopMBB
6478 // PHI varPhi, varEnd, varLoop
6479 // PHI srcPhi, src, srcLoop
6480 // PHI destPhi, dst, destLoop
6481 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6482 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6483 // subs varLoop, varPhi, #UnitSize
6485 // fallthrough --> exitMBB
6487 // epilogue to handle left-over bytes
6488 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6489 // [destOut] = STRB_POST(scratch, destLoop, 1)
6490 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6491 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6492 MF->insert(It, loopMBB);
6493 MF->insert(It, exitMBB);
6495 // Transfer the remainder of BB and its successor edges to exitMBB.
6496 exitMBB->splice(exitMBB->begin(), BB,
6497 llvm::next(MachineBasicBlock::iterator(MI)),
6499 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6501 // Load an immediate to varEnd.
6502 unsigned varEnd = MRI.createVirtualRegister(TRC);
6504 unsigned VReg1 = varEnd;
6505 if ((LoopSize & 0xFFFF0000) != 0)
6506 VReg1 = MRI.createVirtualRegister(TRC);
6507 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6508 .addImm(LoopSize & 0xFFFF));
6510 if ((LoopSize & 0xFFFF0000) != 0)
6511 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6513 .addImm(LoopSize >> 16));
6515 MachineConstantPool *ConstantPool = MF->getConstantPool();
6516 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6517 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6519 // MachineConstantPool wants an explicit alignment.
6520 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6522 Align = getDataLayout()->getTypeAllocSize(C->getType());
6523 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6525 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6526 .addReg(varEnd, RegState::Define)
6527 .addConstantPoolIndex(Idx)
6530 BB->addSuccessor(loopMBB);
6532 // Generate the loop body:
6533 // varPhi = PHI(varLoop, varEnd)
6534 // srcPhi = PHI(srcLoop, src)
6535 // destPhi = PHI(destLoop, dst)
6536 MachineBasicBlock *entryBB = BB;
6538 unsigned varLoop = MRI.createVirtualRegister(TRC);
6539 unsigned varPhi = MRI.createVirtualRegister(TRC);
6540 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6541 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6542 unsigned destLoop = MRI.createVirtualRegister(TRC);
6543 unsigned destPhi = MRI.createVirtualRegister(TRC);
6545 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6546 .addReg(varLoop).addMBB(loopMBB)
6547 .addReg(varEnd).addMBB(entryBB);
6548 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6549 .addReg(srcLoop).addMBB(loopMBB)
6550 .addReg(src).addMBB(entryBB);
6551 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6552 .addReg(destLoop).addMBB(loopMBB)
6553 .addReg(dest).addMBB(entryBB);
6555 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6556 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6557 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6558 if (UnitSize >= 8) {
6559 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6560 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6562 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6563 .addReg(destPhi).addImm(0).addReg(scratch));
6564 } else if (isThumb2) {
6565 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6566 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6568 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6569 .addReg(scratch).addReg(destPhi)
6572 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6573 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6576 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6577 .addReg(scratch).addReg(destPhi)
6578 .addReg(0).addImm(UnitSize));
6581 // Decrement loop variable by UnitSize.
6582 MachineInstrBuilder MIB = BuildMI(BB, dl,
6583 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6584 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6585 MIB->getOperand(5).setReg(ARM::CPSR);
6586 MIB->getOperand(5).setIsDef(true);
6588 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6589 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6591 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6592 BB->addSuccessor(loopMBB);
6593 BB->addSuccessor(exitMBB);
6595 // Add epilogue to handle BytesLeft.
6597 MachineInstr *StartOfExit = exitMBB->begin();
6598 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6599 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6601 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6602 // [destOut] = STRB_POST(scratch, destLoop, 1)
6603 unsigned srcIn = srcLoop;
6604 unsigned destIn = destLoop;
6605 for (unsigned i = 0; i < BytesLeft; i++) {
6606 unsigned scratch = MRI.createVirtualRegister(TRC);
6607 unsigned srcOut = MRI.createVirtualRegister(TRC);
6608 unsigned destOut = MRI.createVirtualRegister(TRC);
6610 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6611 TII->get(ldrOpc),scratch)
6612 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6614 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6615 .addReg(scratch).addReg(destIn)
6618 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6619 TII->get(ldrOpc),scratch)
6620 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6622 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6623 .addReg(scratch).addReg(destIn)
6624 .addReg(0).addImm(1));
6630 MI->eraseFromParent(); // The instruction is gone now.
6635 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6636 MachineBasicBlock *BB) const {
6637 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6638 DebugLoc dl = MI->getDebugLoc();
6639 bool isThumb2 = Subtarget->isThumb2();
6640 switch (MI->getOpcode()) {
6643 llvm_unreachable("Unexpected instr type to insert");
6645 // The Thumb2 pre-indexed stores have the same MI operands, they just
6646 // define them differently in the .td files from the isel patterns, so
6647 // they need pseudos.
6648 case ARM::t2STR_preidx:
6649 MI->setDesc(TII->get(ARM::t2STR_PRE));
6651 case ARM::t2STRB_preidx:
6652 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6654 case ARM::t2STRH_preidx:
6655 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6658 case ARM::STRi_preidx:
6659 case ARM::STRBi_preidx: {
6660 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6661 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6662 // Decode the offset.
6663 unsigned Offset = MI->getOperand(4).getImm();
6664 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6665 Offset = ARM_AM::getAM2Offset(Offset);
6669 MachineMemOperand *MMO = *MI->memoperands_begin();
6670 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6671 .addOperand(MI->getOperand(0)) // Rn_wb
6672 .addOperand(MI->getOperand(1)) // Rt
6673 .addOperand(MI->getOperand(2)) // Rn
6674 .addImm(Offset) // offset (skip GPR==zero_reg)
6675 .addOperand(MI->getOperand(5)) // pred
6676 .addOperand(MI->getOperand(6))
6677 .addMemOperand(MMO);
6678 MI->eraseFromParent();
6681 case ARM::STRr_preidx:
6682 case ARM::STRBr_preidx:
6683 case ARM::STRH_preidx: {
6685 switch (MI->getOpcode()) {
6686 default: llvm_unreachable("unexpected opcode!");
6687 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6688 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6689 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6691 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6692 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6693 MIB.addOperand(MI->getOperand(i));
6694 MI->eraseFromParent();
6697 case ARM::ATOMIC_LOAD_ADD_I8:
6698 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6699 case ARM::ATOMIC_LOAD_ADD_I16:
6700 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6701 case ARM::ATOMIC_LOAD_ADD_I32:
6702 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6704 case ARM::ATOMIC_LOAD_AND_I8:
6705 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6706 case ARM::ATOMIC_LOAD_AND_I16:
6707 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6708 case ARM::ATOMIC_LOAD_AND_I32:
6709 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6711 case ARM::ATOMIC_LOAD_OR_I8:
6712 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6713 case ARM::ATOMIC_LOAD_OR_I16:
6714 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6715 case ARM::ATOMIC_LOAD_OR_I32:
6716 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6718 case ARM::ATOMIC_LOAD_XOR_I8:
6719 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6720 case ARM::ATOMIC_LOAD_XOR_I16:
6721 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6722 case ARM::ATOMIC_LOAD_XOR_I32:
6723 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6725 case ARM::ATOMIC_LOAD_NAND_I8:
6726 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6727 case ARM::ATOMIC_LOAD_NAND_I16:
6728 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6729 case ARM::ATOMIC_LOAD_NAND_I32:
6730 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6732 case ARM::ATOMIC_LOAD_SUB_I8:
6733 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6734 case ARM::ATOMIC_LOAD_SUB_I16:
6735 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6736 case ARM::ATOMIC_LOAD_SUB_I32:
6737 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6739 case ARM::ATOMIC_LOAD_MIN_I8:
6740 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6741 case ARM::ATOMIC_LOAD_MIN_I16:
6742 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6743 case ARM::ATOMIC_LOAD_MIN_I32:
6744 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6746 case ARM::ATOMIC_LOAD_MAX_I8:
6747 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6748 case ARM::ATOMIC_LOAD_MAX_I16:
6749 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6750 case ARM::ATOMIC_LOAD_MAX_I32:
6751 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6753 case ARM::ATOMIC_LOAD_UMIN_I8:
6754 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6755 case ARM::ATOMIC_LOAD_UMIN_I16:
6756 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6757 case ARM::ATOMIC_LOAD_UMIN_I32:
6758 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6760 case ARM::ATOMIC_LOAD_UMAX_I8:
6761 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6762 case ARM::ATOMIC_LOAD_UMAX_I16:
6763 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6764 case ARM::ATOMIC_LOAD_UMAX_I32:
6765 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6767 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6768 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6769 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6771 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6772 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6773 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6776 case ARM::ATOMADD6432:
6777 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6778 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6779 /*NeedsCarry*/ true);
6780 case ARM::ATOMSUB6432:
6781 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6782 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6783 /*NeedsCarry*/ true);
6784 case ARM::ATOMOR6432:
6785 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6786 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6787 case ARM::ATOMXOR6432:
6788 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6789 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6790 case ARM::ATOMAND6432:
6791 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6792 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6793 case ARM::ATOMSWAP6432:
6794 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6795 case ARM::ATOMCMPXCHG6432:
6796 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6797 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6798 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6800 case ARM::tMOVCCr_pseudo: {
6801 // To "insert" a SELECT_CC instruction, we actually have to insert the
6802 // diamond control-flow pattern. The incoming instruction knows the
6803 // destination vreg to set, the condition code register to branch on, the
6804 // true/false values to select between, and a branch opcode to use.
6805 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6806 MachineFunction::iterator It = BB;
6812 // cmpTY ccX, r1, r2
6814 // fallthrough --> copy0MBB
6815 MachineBasicBlock *thisMBB = BB;
6816 MachineFunction *F = BB->getParent();
6817 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6818 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6819 F->insert(It, copy0MBB);
6820 F->insert(It, sinkMBB);
6822 // Transfer the remainder of BB and its successor edges to sinkMBB.
6823 sinkMBB->splice(sinkMBB->begin(), BB,
6824 llvm::next(MachineBasicBlock::iterator(MI)),
6826 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6828 BB->addSuccessor(copy0MBB);
6829 BB->addSuccessor(sinkMBB);
6831 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6832 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6835 // %FalseValue = ...
6836 // # fallthrough to sinkMBB
6839 // Update machine-CFG edges
6840 BB->addSuccessor(sinkMBB);
6843 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6846 BuildMI(*BB, BB->begin(), dl,
6847 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6848 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6849 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6851 MI->eraseFromParent(); // The pseudo instruction is gone now.
6856 case ARM::BCCZi64: {
6857 // If there is an unconditional branch to the other successor, remove it.
6858 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6860 // Compare both parts that make up the double comparison separately for
6862 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6864 unsigned LHS1 = MI->getOperand(1).getReg();
6865 unsigned LHS2 = MI->getOperand(2).getReg();
6867 AddDefaultPred(BuildMI(BB, dl,
6868 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6869 .addReg(LHS1).addImm(0));
6870 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6871 .addReg(LHS2).addImm(0)
6872 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6874 unsigned RHS1 = MI->getOperand(3).getReg();
6875 unsigned RHS2 = MI->getOperand(4).getReg();
6876 AddDefaultPred(BuildMI(BB, dl,
6877 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6878 .addReg(LHS1).addReg(RHS1));
6879 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6880 .addReg(LHS2).addReg(RHS2)
6881 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6884 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6885 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6886 if (MI->getOperand(0).getImm() == ARMCC::NE)
6887 std::swap(destMBB, exitMBB);
6889 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6890 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6892 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6894 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6896 MI->eraseFromParent(); // The pseudo instruction is gone now.
6900 case ARM::Int_eh_sjlj_setjmp:
6901 case ARM::Int_eh_sjlj_setjmp_nofp:
6902 case ARM::tInt_eh_sjlj_setjmp:
6903 case ARM::t2Int_eh_sjlj_setjmp:
6904 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6905 EmitSjLjDispatchBlock(MI, BB);
6910 // To insert an ABS instruction, we have to insert the
6911 // diamond control-flow pattern. The incoming instruction knows the
6912 // source vreg to test against 0, the destination vreg to set,
6913 // the condition code register to branch on, the
6914 // true/false values to select between, and a branch opcode to use.
6919 // BCC (branch to SinkBB if V0 >= 0)
6920 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6921 // SinkBB: V1 = PHI(V2, V3)
6922 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6923 MachineFunction::iterator BBI = BB;
6925 MachineFunction *Fn = BB->getParent();
6926 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6927 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6928 Fn->insert(BBI, RSBBB);
6929 Fn->insert(BBI, SinkBB);
6931 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6932 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6933 bool isThumb2 = Subtarget->isThumb2();
6934 MachineRegisterInfo &MRI = Fn->getRegInfo();
6935 // In Thumb mode S must not be specified if source register is the SP or
6936 // PC and if destination register is the SP, so restrict register class
6937 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6938 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6939 (const TargetRegisterClass*)&ARM::GPRRegClass);
6941 // Transfer the remainder of BB and its successor edges to sinkMBB.
6942 SinkBB->splice(SinkBB->begin(), BB,
6943 llvm::next(MachineBasicBlock::iterator(MI)),
6945 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6947 BB->addSuccessor(RSBBB);
6948 BB->addSuccessor(SinkBB);
6950 // fall through to SinkMBB
6951 RSBBB->addSuccessor(SinkBB);
6953 // insert a cmp at the end of BB
6954 AddDefaultPred(BuildMI(BB, dl,
6955 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6956 .addReg(ABSSrcReg).addImm(0));
6958 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6960 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6961 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6963 // insert rsbri in RSBBB
6964 // Note: BCC and rsbri will be converted into predicated rsbmi
6965 // by if-conversion pass
6966 BuildMI(*RSBBB, RSBBB->begin(), dl,
6967 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6968 .addReg(ABSSrcReg, RegState::Kill)
6969 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6971 // insert PHI in SinkBB,
6972 // reuse ABSDstReg to not change uses of ABS instruction
6973 BuildMI(*SinkBB, SinkBB->begin(), dl,
6974 TII->get(ARM::PHI), ABSDstReg)
6975 .addReg(NewRsbDstReg).addMBB(RSBBB)
6976 .addReg(ABSSrcReg).addMBB(BB);
6978 // remove ABS instruction
6979 MI->eraseFromParent();
6981 // return last added BB
6984 case ARM::COPY_STRUCT_BYVAL_I32:
6986 return EmitStructByval(MI, BB);
6990 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6991 SDNode *Node) const {
6992 if (!MI->hasPostISelHook()) {
6993 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6994 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6998 const MCInstrDesc *MCID = &MI->getDesc();
6999 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7000 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7001 // operand is still set to noreg. If needed, set the optional operand's
7002 // register to CPSR, and remove the redundant implicit def.
7004 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7006 // Rename pseudo opcodes.
7007 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7009 const ARMBaseInstrInfo *TII =
7010 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7011 MCID = &TII->get(NewOpc);
7013 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7014 "converted opcode should be the same except for cc_out");
7018 // Add the optional cc_out operand
7019 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7021 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7023 // Any ARM instruction that sets the 's' bit should specify an optional
7024 // "cc_out" operand in the last operand position.
7025 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7026 assert(!NewOpc && "Optional cc_out operand required");
7029 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7030 // since we already have an optional CPSR def.
7031 bool definesCPSR = false;
7032 bool deadCPSR = false;
7033 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7035 const MachineOperand &MO = MI->getOperand(i);
7036 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7040 MI->RemoveOperand(i);
7045 assert(!NewOpc && "Optional cc_out operand required");
7048 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7050 assert(!MI->getOperand(ccOutIdx).getReg() &&
7051 "expect uninitialized optional cc_out operand");
7055 // If this instruction was defined with an optional CPSR def and its dag node
7056 // had a live implicit CPSR def, then activate the optional CPSR def.
7057 MachineOperand &MO = MI->getOperand(ccOutIdx);
7058 MO.setReg(ARM::CPSR);
7062 //===----------------------------------------------------------------------===//
7063 // ARM Optimization Hooks
7064 //===----------------------------------------------------------------------===//
7066 // Helper function that checks if N is a null or all ones constant.
7067 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7068 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7071 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7074 // Return true if N is conditionally 0 or all ones.
7075 // Detects these expressions where cc is an i1 value:
7077 // (select cc 0, y) [AllOnes=0]
7078 // (select cc y, 0) [AllOnes=0]
7079 // (zext cc) [AllOnes=0]
7080 // (sext cc) [AllOnes=0/1]
7081 // (select cc -1, y) [AllOnes=1]
7082 // (select cc y, -1) [AllOnes=1]
7084 // Invert is set when N is the null/all ones constant when CC is false.
7085 // OtherOp is set to the alternative value of N.
7086 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7087 SDValue &CC, bool &Invert,
7089 SelectionDAG &DAG) {
7090 switch (N->getOpcode()) {
7091 default: return false;
7093 CC = N->getOperand(0);
7094 SDValue N1 = N->getOperand(1);
7095 SDValue N2 = N->getOperand(2);
7096 if (isZeroOrAllOnes(N1, AllOnes)) {
7101 if (isZeroOrAllOnes(N2, AllOnes)) {
7108 case ISD::ZERO_EXTEND:
7109 // (zext cc) can never be the all ones value.
7113 case ISD::SIGN_EXTEND: {
7114 EVT VT = N->getValueType(0);
7115 CC = N->getOperand(0);
7116 if (CC.getValueType() != MVT::i1)
7120 // When looking for an AllOnes constant, N is an sext, and the 'other'
7122 OtherOp = DAG.getConstant(0, VT);
7123 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7124 // When looking for a 0 constant, N can be zext or sext.
7125 OtherOp = DAG.getConstant(1, VT);
7127 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7133 // Combine a constant select operand into its use:
7135 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7136 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7137 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7138 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7139 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7141 // The transform is rejected if the select doesn't have a constant operand that
7142 // is null, or all ones when AllOnes is set.
7144 // Also recognize sext/zext from i1:
7146 // (add (zext cc), x) -> (select cc (add x, 1), x)
7147 // (add (sext cc), x) -> (select cc (add x, -1), x)
7149 // These transformations eventually create predicated instructions.
7151 // @param N The node to transform.
7152 // @param Slct The N operand that is a select.
7153 // @param OtherOp The other N operand (x above).
7154 // @param DCI Context.
7155 // @param AllOnes Require the select constant to be all ones instead of null.
7156 // @returns The new node, or SDValue() on failure.
7158 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7159 TargetLowering::DAGCombinerInfo &DCI,
7160 bool AllOnes = false) {
7161 SelectionDAG &DAG = DCI.DAG;
7162 EVT VT = N->getValueType(0);
7163 SDValue NonConstantVal;
7166 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7167 NonConstantVal, DAG))
7170 // Slct is now know to be the desired identity constant when CC is true.
7171 SDValue TrueVal = OtherOp;
7172 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7173 OtherOp, NonConstantVal);
7174 // Unless SwapSelectOps says CC should be false.
7176 std::swap(TrueVal, FalseVal);
7178 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7179 CCOp, TrueVal, FalseVal);
7182 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7184 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7185 TargetLowering::DAGCombinerInfo &DCI) {
7186 SDValue N0 = N->getOperand(0);
7187 SDValue N1 = N->getOperand(1);
7188 if (N0.getNode()->hasOneUse()) {
7189 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7190 if (Result.getNode())
7193 if (N1.getNode()->hasOneUse()) {
7194 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7195 if (Result.getNode())
7201 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7202 // (only after legalization).
7203 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7204 TargetLowering::DAGCombinerInfo &DCI,
7205 const ARMSubtarget *Subtarget) {
7207 // Only perform optimization if after legalize, and if NEON is available. We
7208 // also expected both operands to be BUILD_VECTORs.
7209 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7210 || N0.getOpcode() != ISD::BUILD_VECTOR
7211 || N1.getOpcode() != ISD::BUILD_VECTOR)
7214 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7215 EVT VT = N->getValueType(0);
7216 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7219 // Check that the vector operands are of the right form.
7220 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7221 // operands, where N is the size of the formed vector.
7222 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7223 // index such that we have a pair wise add pattern.
7225 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7226 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7228 SDValue Vec = N0->getOperand(0)->getOperand(0);
7229 SDNode *V = Vec.getNode();
7230 unsigned nextIndex = 0;
7232 // For each operands to the ADD which are BUILD_VECTORs,
7233 // check to see if each of their operands are an EXTRACT_VECTOR with
7234 // the same vector and appropriate index.
7235 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7236 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7237 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7239 SDValue ExtVec0 = N0->getOperand(i);
7240 SDValue ExtVec1 = N1->getOperand(i);
7242 // First operand is the vector, verify its the same.
7243 if (V != ExtVec0->getOperand(0).getNode() ||
7244 V != ExtVec1->getOperand(0).getNode())
7247 // Second is the constant, verify its correct.
7248 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7249 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7251 // For the constant, we want to see all the even or all the odd.
7252 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7253 || C1->getZExtValue() != nextIndex+1)
7262 // Create VPADDL node.
7263 SelectionDAG &DAG = DCI.DAG;
7264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7266 // Build operand list.
7267 SmallVector<SDValue, 8> Ops;
7268 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7269 TLI.getPointerTy()));
7271 // Input is the vector.
7274 // Get widened type and narrowed type.
7276 unsigned numElem = VT.getVectorNumElements();
7277 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7278 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7279 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7280 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7282 llvm_unreachable("Invalid vector element type for padd optimization.");
7285 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7286 widenType, &Ops[0], Ops.size());
7287 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7290 static SDValue findMUL_LOHI(SDValue V) {
7291 if (V->getOpcode() == ISD::UMUL_LOHI ||
7292 V->getOpcode() == ISD::SMUL_LOHI)
7297 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7298 TargetLowering::DAGCombinerInfo &DCI,
7299 const ARMSubtarget *Subtarget) {
7301 if (Subtarget->isThumb1Only()) return SDValue();
7303 // Only perform the checks after legalize when the pattern is available.
7304 if (DCI.isBeforeLegalize()) return SDValue();
7306 // Look for multiply add opportunities.
7307 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7308 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7309 // a glue link from the first add to the second add.
7310 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7311 // a S/UMLAL instruction.
7314 // \ / \ [no multiline comment]
7320 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7321 SDValue AddcOp0 = AddcNode->getOperand(0);
7322 SDValue AddcOp1 = AddcNode->getOperand(1);
7324 // Check if the two operands are from the same mul_lohi node.
7325 if (AddcOp0.getNode() == AddcOp1.getNode())
7328 assert(AddcNode->getNumValues() == 2 &&
7329 AddcNode->getValueType(0) == MVT::i32 &&
7330 AddcNode->getValueType(1) == MVT::Glue &&
7331 "Expect ADDC with two result values: i32, glue");
7333 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7334 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7335 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7336 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7337 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7340 // Look for the glued ADDE.
7341 SDNode* AddeNode = AddcNode->getGluedUser();
7342 if (AddeNode == NULL)
7345 // Make sure it is really an ADDE.
7346 if (AddeNode->getOpcode() != ISD::ADDE)
7349 assert(AddeNode->getNumOperands() == 3 &&
7350 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7351 "ADDE node has the wrong inputs");
7353 // Check for the triangle shape.
7354 SDValue AddeOp0 = AddeNode->getOperand(0);
7355 SDValue AddeOp1 = AddeNode->getOperand(1);
7357 // Make sure that the ADDE operands are not coming from the same node.
7358 if (AddeOp0.getNode() == AddeOp1.getNode())
7361 // Find the MUL_LOHI node walking up ADDE's operands.
7362 bool IsLeftOperandMUL = false;
7363 SDValue MULOp = findMUL_LOHI(AddeOp0);
7364 if (MULOp == SDValue())
7365 MULOp = findMUL_LOHI(AddeOp1);
7367 IsLeftOperandMUL = true;
7368 if (MULOp == SDValue())
7371 // Figure out the right opcode.
7372 unsigned Opc = MULOp->getOpcode();
7373 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7375 // Figure out the high and low input values to the MLAL node.
7376 SDValue* HiMul = &MULOp;
7377 SDValue* HiAdd = NULL;
7378 SDValue* LoMul = NULL;
7379 SDValue* LowAdd = NULL;
7381 if (IsLeftOperandMUL)
7387 if (AddcOp0->getOpcode() == Opc) {
7391 if (AddcOp1->getOpcode() == Opc) {
7399 if (LoMul->getNode() != HiMul->getNode())
7402 // Create the merged node.
7403 SelectionDAG &DAG = DCI.DAG;
7405 // Build operand list.
7406 SmallVector<SDValue, 8> Ops;
7407 Ops.push_back(LoMul->getOperand(0));
7408 Ops.push_back(LoMul->getOperand(1));
7409 Ops.push_back(*LowAdd);
7410 Ops.push_back(*HiAdd);
7412 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7413 DAG.getVTList(MVT::i32, MVT::i32),
7414 &Ops[0], Ops.size());
7416 // Replace the ADDs' nodes uses by the MLA node's values.
7417 SDValue HiMLALResult(MLALNode.getNode(), 1);
7418 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7420 SDValue LoMLALResult(MLALNode.getNode(), 0);
7421 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7423 // Return original node to notify the driver to stop replacing.
7424 SDValue resNode(AddcNode, 0);
7428 /// PerformADDCCombine - Target-specific dag combine transform from
7429 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7430 static SDValue PerformADDCCombine(SDNode *N,
7431 TargetLowering::DAGCombinerInfo &DCI,
7432 const ARMSubtarget *Subtarget) {
7434 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7438 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7439 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7440 /// called with the default operands, and if that fails, with commuted
7442 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7443 TargetLowering::DAGCombinerInfo &DCI,
7444 const ARMSubtarget *Subtarget){
7446 // Attempt to create vpaddl for this add.
7447 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7448 if (Result.getNode())
7451 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7452 if (N0.getNode()->hasOneUse()) {
7453 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7454 if (Result.getNode()) return Result;
7459 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7461 static SDValue PerformADDCombine(SDNode *N,
7462 TargetLowering::DAGCombinerInfo &DCI,
7463 const ARMSubtarget *Subtarget) {
7464 SDValue N0 = N->getOperand(0);
7465 SDValue N1 = N->getOperand(1);
7467 // First try with the default operand order.
7468 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7469 if (Result.getNode())
7472 // If that didn't work, try again with the operands commuted.
7473 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7476 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7478 static SDValue PerformSUBCombine(SDNode *N,
7479 TargetLowering::DAGCombinerInfo &DCI) {
7480 SDValue N0 = N->getOperand(0);
7481 SDValue N1 = N->getOperand(1);
7483 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7484 if (N1.getNode()->hasOneUse()) {
7485 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7486 if (Result.getNode()) return Result;
7492 /// PerformVMULCombine
7493 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7494 /// special multiplier accumulator forwarding.
7500 static SDValue PerformVMULCombine(SDNode *N,
7501 TargetLowering::DAGCombinerInfo &DCI,
7502 const ARMSubtarget *Subtarget) {
7503 if (!Subtarget->hasVMLxForwarding())
7506 SelectionDAG &DAG = DCI.DAG;
7507 SDValue N0 = N->getOperand(0);
7508 SDValue N1 = N->getOperand(1);
7509 unsigned Opcode = N0.getOpcode();
7510 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7511 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7512 Opcode = N1.getOpcode();
7513 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7514 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7519 EVT VT = N->getValueType(0);
7520 DebugLoc DL = N->getDebugLoc();
7521 SDValue N00 = N0->getOperand(0);
7522 SDValue N01 = N0->getOperand(1);
7523 return DAG.getNode(Opcode, DL, VT,
7524 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7525 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7528 static SDValue PerformMULCombine(SDNode *N,
7529 TargetLowering::DAGCombinerInfo &DCI,
7530 const ARMSubtarget *Subtarget) {
7531 SelectionDAG &DAG = DCI.DAG;
7533 if (Subtarget->isThumb1Only())
7536 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7539 EVT VT = N->getValueType(0);
7540 if (VT.is64BitVector() || VT.is128BitVector())
7541 return PerformVMULCombine(N, DCI, Subtarget);
7545 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7549 int64_t MulAmt = C->getSExtValue();
7550 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7552 ShiftAmt = ShiftAmt & (32 - 1);
7553 SDValue V = N->getOperand(0);
7554 DebugLoc DL = N->getDebugLoc();
7557 MulAmt >>= ShiftAmt;
7560 if (isPowerOf2_32(MulAmt - 1)) {
7561 // (mul x, 2^N + 1) => (add (shl x, N), x)
7562 Res = DAG.getNode(ISD::ADD, DL, VT,
7564 DAG.getNode(ISD::SHL, DL, VT,
7566 DAG.getConstant(Log2_32(MulAmt - 1),
7568 } else if (isPowerOf2_32(MulAmt + 1)) {
7569 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7570 Res = DAG.getNode(ISD::SUB, DL, VT,
7571 DAG.getNode(ISD::SHL, DL, VT,
7573 DAG.getConstant(Log2_32(MulAmt + 1),
7579 uint64_t MulAmtAbs = -MulAmt;
7580 if (isPowerOf2_32(MulAmtAbs + 1)) {
7581 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7582 Res = DAG.getNode(ISD::SUB, DL, VT,
7584 DAG.getNode(ISD::SHL, DL, VT,
7586 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7588 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7589 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7590 Res = DAG.getNode(ISD::ADD, DL, VT,
7592 DAG.getNode(ISD::SHL, DL, VT,
7594 DAG.getConstant(Log2_32(MulAmtAbs-1),
7596 Res = DAG.getNode(ISD::SUB, DL, VT,
7597 DAG.getConstant(0, MVT::i32),Res);
7604 Res = DAG.getNode(ISD::SHL, DL, VT,
7605 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7607 // Do not add new nodes to DAG combiner worklist.
7608 DCI.CombineTo(N, Res, false);
7612 static SDValue PerformANDCombine(SDNode *N,
7613 TargetLowering::DAGCombinerInfo &DCI,
7614 const ARMSubtarget *Subtarget) {
7616 // Attempt to use immediate-form VBIC
7617 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7618 DebugLoc dl = N->getDebugLoc();
7619 EVT VT = N->getValueType(0);
7620 SelectionDAG &DAG = DCI.DAG;
7622 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7625 APInt SplatBits, SplatUndef;
7626 unsigned SplatBitSize;
7629 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7630 if (SplatBitSize <= 64) {
7632 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7633 SplatUndef.getZExtValue(), SplatBitSize,
7634 DAG, VbicVT, VT.is128BitVector(),
7636 if (Val.getNode()) {
7638 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7639 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7640 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7645 if (!Subtarget->isThumb1Only()) {
7646 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7647 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7648 if (Result.getNode())
7655 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7656 static SDValue PerformORCombine(SDNode *N,
7657 TargetLowering::DAGCombinerInfo &DCI,
7658 const ARMSubtarget *Subtarget) {
7659 // Attempt to use immediate-form VORR
7660 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7661 DebugLoc dl = N->getDebugLoc();
7662 EVT VT = N->getValueType(0);
7663 SelectionDAG &DAG = DCI.DAG;
7665 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7668 APInt SplatBits, SplatUndef;
7669 unsigned SplatBitSize;
7671 if (BVN && Subtarget->hasNEON() &&
7672 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7673 if (SplatBitSize <= 64) {
7675 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7676 SplatUndef.getZExtValue(), SplatBitSize,
7677 DAG, VorrVT, VT.is128BitVector(),
7679 if (Val.getNode()) {
7681 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7682 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7683 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7688 if (!Subtarget->isThumb1Only()) {
7689 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7690 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7691 if (Result.getNode())
7695 // The code below optimizes (or (and X, Y), Z).
7696 // The AND operand needs to have a single user to make these optimizations
7698 SDValue N0 = N->getOperand(0);
7699 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7701 SDValue N1 = N->getOperand(1);
7703 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7704 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7705 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7707 unsigned SplatBitSize;
7710 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7712 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7713 HasAnyUndefs) && !HasAnyUndefs) {
7714 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7716 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7717 HasAnyUndefs) && !HasAnyUndefs &&
7718 SplatBits0 == ~SplatBits1) {
7719 // Canonicalize the vector type to make instruction selection simpler.
7720 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7721 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7722 N0->getOperand(1), N0->getOperand(0),
7724 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7729 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7732 // BFI is only available on V6T2+
7733 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7736 DebugLoc DL = N->getDebugLoc();
7737 // 1) or (and A, mask), val => ARMbfi A, val, mask
7738 // iff (val & mask) == val
7740 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7741 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7742 // && mask == ~mask2
7743 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7744 // && ~mask == mask2
7745 // (i.e., copy a bitfield value into another bitfield of the same width)
7750 SDValue N00 = N0.getOperand(0);
7752 // The value and the mask need to be constants so we can verify this is
7753 // actually a bitfield set. If the mask is 0xffff, we can do better
7754 // via a movt instruction, so don't use BFI in that case.
7755 SDValue MaskOp = N0.getOperand(1);
7756 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7759 unsigned Mask = MaskC->getZExtValue();
7763 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7764 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7766 unsigned Val = N1C->getZExtValue();
7767 if ((Val & ~Mask) != Val)
7770 if (ARM::isBitFieldInvertedMask(Mask)) {
7771 Val >>= CountTrailingZeros_32(~Mask);
7773 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7774 DAG.getConstant(Val, MVT::i32),
7775 DAG.getConstant(Mask, MVT::i32));
7777 // Do not add new nodes to DAG combiner worklist.
7778 DCI.CombineTo(N, Res, false);
7781 } else if (N1.getOpcode() == ISD::AND) {
7782 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7783 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7786 unsigned Mask2 = N11C->getZExtValue();
7788 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7790 if (ARM::isBitFieldInvertedMask(Mask) &&
7792 // The pack halfword instruction works better for masks that fit it,
7793 // so use that when it's available.
7794 if (Subtarget->hasT2ExtractPack() &&
7795 (Mask == 0xffff || Mask == 0xffff0000))
7798 unsigned amt = CountTrailingZeros_32(Mask2);
7799 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7800 DAG.getConstant(amt, MVT::i32));
7801 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7802 DAG.getConstant(Mask, MVT::i32));
7803 // Do not add new nodes to DAG combiner worklist.
7804 DCI.CombineTo(N, Res, false);
7806 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7808 // The pack halfword instruction works better for masks that fit it,
7809 // so use that when it's available.
7810 if (Subtarget->hasT2ExtractPack() &&
7811 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7814 unsigned lsb = CountTrailingZeros_32(Mask);
7815 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7816 DAG.getConstant(lsb, MVT::i32));
7817 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7818 DAG.getConstant(Mask2, MVT::i32));
7819 // Do not add new nodes to DAG combiner worklist.
7820 DCI.CombineTo(N, Res, false);
7825 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7826 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7827 ARM::isBitFieldInvertedMask(~Mask)) {
7828 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7829 // where lsb(mask) == #shamt and masked bits of B are known zero.
7830 SDValue ShAmt = N00.getOperand(1);
7831 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7832 unsigned LSB = CountTrailingZeros_32(Mask);
7836 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7837 DAG.getConstant(~Mask, MVT::i32));
7839 // Do not add new nodes to DAG combiner worklist.
7840 DCI.CombineTo(N, Res, false);
7846 static SDValue PerformXORCombine(SDNode *N,
7847 TargetLowering::DAGCombinerInfo &DCI,
7848 const ARMSubtarget *Subtarget) {
7849 EVT VT = N->getValueType(0);
7850 SelectionDAG &DAG = DCI.DAG;
7852 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7855 if (!Subtarget->isThumb1Only()) {
7856 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7857 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7858 if (Result.getNode())
7865 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7866 /// the bits being cleared by the AND are not demanded by the BFI.
7867 static SDValue PerformBFICombine(SDNode *N,
7868 TargetLowering::DAGCombinerInfo &DCI) {
7869 SDValue N1 = N->getOperand(1);
7870 if (N1.getOpcode() == ISD::AND) {
7871 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7874 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7875 unsigned LSB = CountTrailingZeros_32(~InvMask);
7876 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7877 unsigned Mask = (1 << Width)-1;
7878 unsigned Mask2 = N11C->getZExtValue();
7879 if ((Mask & (~Mask2)) == 0)
7880 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7881 N->getOperand(0), N1.getOperand(0),
7887 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7888 /// ARMISD::VMOVRRD.
7889 static SDValue PerformVMOVRRDCombine(SDNode *N,
7890 TargetLowering::DAGCombinerInfo &DCI) {
7891 // vmovrrd(vmovdrr x, y) -> x,y
7892 SDValue InDouble = N->getOperand(0);
7893 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7894 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7896 // vmovrrd(load f64) -> (load i32), (load i32)
7897 SDNode *InNode = InDouble.getNode();
7898 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7899 InNode->getValueType(0) == MVT::f64 &&
7900 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7901 !cast<LoadSDNode>(InNode)->isVolatile()) {
7902 // TODO: Should this be done for non-FrameIndex operands?
7903 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7905 SelectionDAG &DAG = DCI.DAG;
7906 DebugLoc DL = LD->getDebugLoc();
7907 SDValue BasePtr = LD->getBasePtr();
7908 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7909 LD->getPointerInfo(), LD->isVolatile(),
7910 LD->isNonTemporal(), LD->isInvariant(),
7911 LD->getAlignment());
7913 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7914 DAG.getConstant(4, MVT::i32));
7915 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7916 LD->getPointerInfo(), LD->isVolatile(),
7917 LD->isNonTemporal(), LD->isInvariant(),
7918 std::min(4U, LD->getAlignment() / 2));
7920 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7921 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7922 DCI.RemoveFromWorklist(LD);
7930 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7931 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7932 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7933 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7934 SDValue Op0 = N->getOperand(0);
7935 SDValue Op1 = N->getOperand(1);
7936 if (Op0.getOpcode() == ISD::BITCAST)
7937 Op0 = Op0.getOperand(0);
7938 if (Op1.getOpcode() == ISD::BITCAST)
7939 Op1 = Op1.getOperand(0);
7940 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7941 Op0.getNode() == Op1.getNode() &&
7942 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7943 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7944 N->getValueType(0), Op0.getOperand(0));
7948 /// PerformSTORECombine - Target-specific dag combine xforms for
7950 static SDValue PerformSTORECombine(SDNode *N,
7951 TargetLowering::DAGCombinerInfo &DCI) {
7952 StoreSDNode *St = cast<StoreSDNode>(N);
7953 if (St->isVolatile())
7956 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7957 // pack all of the elements in one place. Next, store to memory in fewer
7959 SDValue StVal = St->getValue();
7960 EVT VT = StVal.getValueType();
7961 if (St->isTruncatingStore() && VT.isVector()) {
7962 SelectionDAG &DAG = DCI.DAG;
7963 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7964 EVT StVT = St->getMemoryVT();
7965 unsigned NumElems = VT.getVectorNumElements();
7966 assert(StVT != VT && "Cannot truncate to the same type");
7967 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7968 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7970 // From, To sizes and ElemCount must be pow of two
7971 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7973 // We are going to use the original vector elt for storing.
7974 // Accumulated smaller vector elements must be a multiple of the store size.
7975 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7977 unsigned SizeRatio = FromEltSz / ToEltSz;
7978 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7980 // Create a type on which we perform the shuffle.
7981 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7982 NumElems*SizeRatio);
7983 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7985 DebugLoc DL = St->getDebugLoc();
7986 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
7987 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
7988 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
7990 // Can't shuffle using an illegal type.
7991 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
7993 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
7994 DAG.getUNDEF(WideVec.getValueType()),
7996 // At this point all of the data is stored at the bottom of the
7997 // register. We now need to save it to mem.
7999 // Find the largest store unit
8000 MVT StoreType = MVT::i8;
8001 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8002 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8003 MVT Tp = (MVT::SimpleValueType)tp;
8004 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8007 // Didn't find a legal store type.
8008 if (!TLI.isTypeLegal(StoreType))
8011 // Bitcast the original vector into a vector of store-size units
8012 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8013 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8014 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8015 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8016 SmallVector<SDValue, 8> Chains;
8017 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8018 TLI.getPointerTy());
8019 SDValue BasePtr = St->getBasePtr();
8021 // Perform one or more big stores into memory.
8022 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8023 for (unsigned I = 0; I < E; I++) {
8024 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8025 StoreType, ShuffWide,
8026 DAG.getIntPtrConstant(I));
8027 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8028 St->getPointerInfo(), St->isVolatile(),
8029 St->isNonTemporal(), St->getAlignment());
8030 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8032 Chains.push_back(Ch);
8034 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8038 if (!ISD::isNormalStore(St))
8041 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8042 // ARM stores of arguments in the same cache line.
8043 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8044 StVal.getNode()->hasOneUse()) {
8045 SelectionDAG &DAG = DCI.DAG;
8046 DebugLoc DL = St->getDebugLoc();
8047 SDValue BasePtr = St->getBasePtr();
8048 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8049 StVal.getNode()->getOperand(0), BasePtr,
8050 St->getPointerInfo(), St->isVolatile(),
8051 St->isNonTemporal(), St->getAlignment());
8053 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8054 DAG.getConstant(4, MVT::i32));
8055 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8056 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8057 St->isNonTemporal(),
8058 std::min(4U, St->getAlignment() / 2));
8061 if (StVal.getValueType() != MVT::i64 ||
8062 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8065 // Bitcast an i64 store extracted from a vector to f64.
8066 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8067 SelectionDAG &DAG = DCI.DAG;
8068 DebugLoc dl = StVal.getDebugLoc();
8069 SDValue IntVec = StVal.getOperand(0);
8070 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8071 IntVec.getValueType().getVectorNumElements());
8072 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8073 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8074 Vec, StVal.getOperand(1));
8075 dl = N->getDebugLoc();
8076 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8077 // Make the DAGCombiner fold the bitcasts.
8078 DCI.AddToWorklist(Vec.getNode());
8079 DCI.AddToWorklist(ExtElt.getNode());
8080 DCI.AddToWorklist(V.getNode());
8081 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8082 St->getPointerInfo(), St->isVolatile(),
8083 St->isNonTemporal(), St->getAlignment(),
8087 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8088 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8089 /// i64 vector to have f64 elements, since the value can then be loaded
8090 /// directly into a VFP register.
8091 static bool hasNormalLoadOperand(SDNode *N) {
8092 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8093 for (unsigned i = 0; i < NumElts; ++i) {
8094 SDNode *Elt = N->getOperand(i).getNode();
8095 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8101 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8102 /// ISD::BUILD_VECTOR.
8103 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8104 TargetLowering::DAGCombinerInfo &DCI){
8105 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8106 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8107 // into a pair of GPRs, which is fine when the value is used as a scalar,
8108 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8109 SelectionDAG &DAG = DCI.DAG;
8110 if (N->getNumOperands() == 2) {
8111 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8116 // Load i64 elements as f64 values so that type legalization does not split
8117 // them up into i32 values.
8118 EVT VT = N->getValueType(0);
8119 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8121 DebugLoc dl = N->getDebugLoc();
8122 SmallVector<SDValue, 8> Ops;
8123 unsigned NumElts = VT.getVectorNumElements();
8124 for (unsigned i = 0; i < NumElts; ++i) {
8125 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8127 // Make the DAGCombiner fold the bitcast.
8128 DCI.AddToWorklist(V.getNode());
8130 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8131 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8132 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8135 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8136 /// ISD::INSERT_VECTOR_ELT.
8137 static SDValue PerformInsertEltCombine(SDNode *N,
8138 TargetLowering::DAGCombinerInfo &DCI) {
8139 // Bitcast an i64 load inserted into a vector to f64.
8140 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8141 EVT VT = N->getValueType(0);
8142 SDNode *Elt = N->getOperand(1).getNode();
8143 if (VT.getVectorElementType() != MVT::i64 ||
8144 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8147 SelectionDAG &DAG = DCI.DAG;
8148 DebugLoc dl = N->getDebugLoc();
8149 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8150 VT.getVectorNumElements());
8151 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8152 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8153 // Make the DAGCombiner fold the bitcasts.
8154 DCI.AddToWorklist(Vec.getNode());
8155 DCI.AddToWorklist(V.getNode());
8156 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8157 Vec, V, N->getOperand(2));
8158 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8161 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8162 /// ISD::VECTOR_SHUFFLE.
8163 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8164 // The LLVM shufflevector instruction does not require the shuffle mask
8165 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8166 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8167 // operands do not match the mask length, they are extended by concatenating
8168 // them with undef vectors. That is probably the right thing for other
8169 // targets, but for NEON it is better to concatenate two double-register
8170 // size vector operands into a single quad-register size vector. Do that
8171 // transformation here:
8172 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8173 // shuffle(concat(v1, v2), undef)
8174 SDValue Op0 = N->getOperand(0);
8175 SDValue Op1 = N->getOperand(1);
8176 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8177 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8178 Op0.getNumOperands() != 2 ||
8179 Op1.getNumOperands() != 2)
8181 SDValue Concat0Op1 = Op0.getOperand(1);
8182 SDValue Concat1Op1 = Op1.getOperand(1);
8183 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8184 Concat1Op1.getOpcode() != ISD::UNDEF)
8186 // Skip the transformation if any of the types are illegal.
8187 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8188 EVT VT = N->getValueType(0);
8189 if (!TLI.isTypeLegal(VT) ||
8190 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8191 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8194 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8195 Op0.getOperand(0), Op1.getOperand(0));
8196 // Translate the shuffle mask.
8197 SmallVector<int, 16> NewMask;
8198 unsigned NumElts = VT.getVectorNumElements();
8199 unsigned HalfElts = NumElts/2;
8200 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8201 for (unsigned n = 0; n < NumElts; ++n) {
8202 int MaskElt = SVN->getMaskElt(n);
8204 if (MaskElt < (int)HalfElts)
8206 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8207 NewElt = HalfElts + MaskElt - NumElts;
8208 NewMask.push_back(NewElt);
8210 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8211 DAG.getUNDEF(VT), NewMask.data());
8214 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8215 /// NEON load/store intrinsics to merge base address updates.
8216 static SDValue CombineBaseUpdate(SDNode *N,
8217 TargetLowering::DAGCombinerInfo &DCI) {
8218 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8221 SelectionDAG &DAG = DCI.DAG;
8222 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8223 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8224 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8225 SDValue Addr = N->getOperand(AddrOpIdx);
8227 // Search for a use of the address operand that is an increment.
8228 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8229 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8231 if (User->getOpcode() != ISD::ADD ||
8232 UI.getUse().getResNo() != Addr.getResNo())
8235 // Check that the add is independent of the load/store. Otherwise, folding
8236 // it would create a cycle.
8237 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8240 // Find the new opcode for the updating load/store.
8242 bool isLaneOp = false;
8243 unsigned NewOpc = 0;
8244 unsigned NumVecs = 0;
8246 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8248 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8249 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8251 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8253 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8255 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8257 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8258 NumVecs = 2; isLaneOp = true; break;
8259 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8260 NumVecs = 3; isLaneOp = true; break;
8261 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8262 NumVecs = 4; isLaneOp = true; break;
8263 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8264 NumVecs = 1; isLoad = false; break;
8265 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8266 NumVecs = 2; isLoad = false; break;
8267 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8268 NumVecs = 3; isLoad = false; break;
8269 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8270 NumVecs = 4; isLoad = false; break;
8271 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8272 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8273 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8274 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8275 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8276 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8280 switch (N->getOpcode()) {
8281 default: llvm_unreachable("unexpected opcode for Neon base update");
8282 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8283 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8284 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8288 // Find the size of memory referenced by the load/store.
8291 VecTy = N->getValueType(0);
8293 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8294 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8296 NumBytes /= VecTy.getVectorNumElements();
8298 // If the increment is a constant, it must match the memory ref size.
8299 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8300 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8301 uint64_t IncVal = CInc->getZExtValue();
8302 if (IncVal != NumBytes)
8304 } else if (NumBytes >= 3 * 16) {
8305 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8306 // separate instructions that make it harder to use a non-constant update.
8310 // Create the new updating load/store node.
8312 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8314 for (n = 0; n < NumResultVecs; ++n)
8316 Tys[n++] = MVT::i32;
8317 Tys[n] = MVT::Other;
8318 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8319 SmallVector<SDValue, 8> Ops;
8320 Ops.push_back(N->getOperand(0)); // incoming chain
8321 Ops.push_back(N->getOperand(AddrOpIdx));
8323 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8324 Ops.push_back(N->getOperand(i));
8326 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8327 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8328 Ops.data(), Ops.size(),
8329 MemInt->getMemoryVT(),
8330 MemInt->getMemOperand());
8333 std::vector<SDValue> NewResults;
8334 for (unsigned i = 0; i < NumResultVecs; ++i) {
8335 NewResults.push_back(SDValue(UpdN.getNode(), i));
8337 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8338 DCI.CombineTo(N, NewResults);
8339 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8346 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8347 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8348 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8350 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8351 SelectionDAG &DAG = DCI.DAG;
8352 EVT VT = N->getValueType(0);
8353 // vldN-dup instructions only support 64-bit vectors for N > 1.
8354 if (!VT.is64BitVector())
8357 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8358 SDNode *VLD = N->getOperand(0).getNode();
8359 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8361 unsigned NumVecs = 0;
8362 unsigned NewOpc = 0;
8363 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8364 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8366 NewOpc = ARMISD::VLD2DUP;
8367 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8369 NewOpc = ARMISD::VLD3DUP;
8370 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8372 NewOpc = ARMISD::VLD4DUP;
8377 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8378 // numbers match the load.
8379 unsigned VLDLaneNo =
8380 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8381 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8383 // Ignore uses of the chain result.
8384 if (UI.getUse().getResNo() == NumVecs)
8387 if (User->getOpcode() != ARMISD::VDUPLANE ||
8388 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8392 // Create the vldN-dup node.
8395 for (n = 0; n < NumVecs; ++n)
8397 Tys[n] = MVT::Other;
8398 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8399 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8400 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8401 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8402 Ops, 2, VLDMemInt->getMemoryVT(),
8403 VLDMemInt->getMemOperand());
8406 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8408 unsigned ResNo = UI.getUse().getResNo();
8409 // Ignore uses of the chain result.
8410 if (ResNo == NumVecs)
8413 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8416 // Now the vldN-lane intrinsic is dead except for its chain result.
8417 // Update uses of the chain.
8418 std::vector<SDValue> VLDDupResults;
8419 for (unsigned n = 0; n < NumVecs; ++n)
8420 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8421 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8422 DCI.CombineTo(VLD, VLDDupResults);
8427 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8428 /// ARMISD::VDUPLANE.
8429 static SDValue PerformVDUPLANECombine(SDNode *N,
8430 TargetLowering::DAGCombinerInfo &DCI) {
8431 SDValue Op = N->getOperand(0);
8433 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8434 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8435 if (CombineVLDDUP(N, DCI))
8436 return SDValue(N, 0);
8438 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8439 // redundant. Ignore bit_converts for now; element sizes are checked below.
8440 while (Op.getOpcode() == ISD::BITCAST)
8441 Op = Op.getOperand(0);
8442 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8445 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8446 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8447 // The canonical VMOV for a zero vector uses a 32-bit element size.
8448 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8450 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8452 EVT VT = N->getValueType(0);
8453 if (EltSize > VT.getVectorElementType().getSizeInBits())
8456 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8459 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8460 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8461 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8465 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8467 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8472 APFloat APF = C->getValueAPF();
8473 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8474 != APFloat::opOK || !isExact)
8477 c0 = (I == 0) ? cN : c0;
8478 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8485 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8486 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8487 /// when the VMUL has a constant operand that is a power of 2.
8489 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8490 /// vmul.f32 d16, d17, d16
8491 /// vcvt.s32.f32 d16, d16
8493 /// vcvt.s32.f32 d16, d16, #3
8494 static SDValue PerformVCVTCombine(SDNode *N,
8495 TargetLowering::DAGCombinerInfo &DCI,
8496 const ARMSubtarget *Subtarget) {
8497 SelectionDAG &DAG = DCI.DAG;
8498 SDValue Op = N->getOperand(0);
8500 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8501 Op.getOpcode() != ISD::FMUL)
8505 SDValue N0 = Op->getOperand(0);
8506 SDValue ConstVec = Op->getOperand(1);
8507 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8509 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8510 !isConstVecPow2(ConstVec, isSigned, C))
8513 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8514 Intrinsic::arm_neon_vcvtfp2fxu;
8515 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8517 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8518 DAG.getConstant(Log2_64(C), MVT::i32));
8521 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8522 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8523 /// when the VDIV has a constant operand that is a power of 2.
8525 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8526 /// vcvt.f32.s32 d16, d16
8527 /// vdiv.f32 d16, d17, d16
8529 /// vcvt.f32.s32 d16, d16, #3
8530 static SDValue PerformVDIVCombine(SDNode *N,
8531 TargetLowering::DAGCombinerInfo &DCI,
8532 const ARMSubtarget *Subtarget) {
8533 SelectionDAG &DAG = DCI.DAG;
8534 SDValue Op = N->getOperand(0);
8535 unsigned OpOpcode = Op.getNode()->getOpcode();
8537 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8538 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8542 SDValue ConstVec = N->getOperand(1);
8543 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8545 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8546 !isConstVecPow2(ConstVec, isSigned, C))
8549 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8550 Intrinsic::arm_neon_vcvtfxu2fp;
8551 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8553 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8554 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8557 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8558 /// operand of a vector shift operation, where all the elements of the
8559 /// build_vector must have the same constant integer value.
8560 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8561 // Ignore bit_converts.
8562 while (Op.getOpcode() == ISD::BITCAST)
8563 Op = Op.getOperand(0);
8564 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8565 APInt SplatBits, SplatUndef;
8566 unsigned SplatBitSize;
8568 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8569 HasAnyUndefs, ElementBits) ||
8570 SplatBitSize > ElementBits)
8572 Cnt = SplatBits.getSExtValue();
8576 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8577 /// operand of a vector shift left operation. That value must be in the range:
8578 /// 0 <= Value < ElementBits for a left shift; or
8579 /// 0 <= Value <= ElementBits for a long left shift.
8580 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8581 assert(VT.isVector() && "vector shift count is not a vector type");
8582 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8583 if (! getVShiftImm(Op, ElementBits, Cnt))
8585 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8588 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8589 /// operand of a vector shift right operation. For a shift opcode, the value
8590 /// is positive, but for an intrinsic the value count must be negative. The
8591 /// absolute value must be in the range:
8592 /// 1 <= |Value| <= ElementBits for a right shift; or
8593 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8594 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8596 assert(VT.isVector() && "vector shift count is not a vector type");
8597 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8598 if (! getVShiftImm(Op, ElementBits, Cnt))
8602 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8605 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8606 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8607 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8610 // Don't do anything for most intrinsics.
8613 // Vector shifts: check for immediate versions and lower them.
8614 // Note: This is done during DAG combining instead of DAG legalizing because
8615 // the build_vectors for 64-bit vector element shift counts are generally
8616 // not legal, and it is hard to see their values after they get legalized to
8617 // loads from a constant pool.
8618 case Intrinsic::arm_neon_vshifts:
8619 case Intrinsic::arm_neon_vshiftu:
8620 case Intrinsic::arm_neon_vshiftls:
8621 case Intrinsic::arm_neon_vshiftlu:
8622 case Intrinsic::arm_neon_vshiftn:
8623 case Intrinsic::arm_neon_vrshifts:
8624 case Intrinsic::arm_neon_vrshiftu:
8625 case Intrinsic::arm_neon_vrshiftn:
8626 case Intrinsic::arm_neon_vqshifts:
8627 case Intrinsic::arm_neon_vqshiftu:
8628 case Intrinsic::arm_neon_vqshiftsu:
8629 case Intrinsic::arm_neon_vqshiftns:
8630 case Intrinsic::arm_neon_vqshiftnu:
8631 case Intrinsic::arm_neon_vqshiftnsu:
8632 case Intrinsic::arm_neon_vqrshiftns:
8633 case Intrinsic::arm_neon_vqrshiftnu:
8634 case Intrinsic::arm_neon_vqrshiftnsu: {
8635 EVT VT = N->getOperand(1).getValueType();
8637 unsigned VShiftOpc = 0;
8640 case Intrinsic::arm_neon_vshifts:
8641 case Intrinsic::arm_neon_vshiftu:
8642 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8643 VShiftOpc = ARMISD::VSHL;
8646 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8647 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8648 ARMISD::VSHRs : ARMISD::VSHRu);
8653 case Intrinsic::arm_neon_vshiftls:
8654 case Intrinsic::arm_neon_vshiftlu:
8655 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8657 llvm_unreachable("invalid shift count for vshll intrinsic");
8659 case Intrinsic::arm_neon_vrshifts:
8660 case Intrinsic::arm_neon_vrshiftu:
8661 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8665 case Intrinsic::arm_neon_vqshifts:
8666 case Intrinsic::arm_neon_vqshiftu:
8667 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8671 case Intrinsic::arm_neon_vqshiftsu:
8672 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8674 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8676 case Intrinsic::arm_neon_vshiftn:
8677 case Intrinsic::arm_neon_vrshiftn:
8678 case Intrinsic::arm_neon_vqshiftns:
8679 case Intrinsic::arm_neon_vqshiftnu:
8680 case Intrinsic::arm_neon_vqshiftnsu:
8681 case Intrinsic::arm_neon_vqrshiftns:
8682 case Intrinsic::arm_neon_vqrshiftnu:
8683 case Intrinsic::arm_neon_vqrshiftnsu:
8684 // Narrowing shifts require an immediate right shift.
8685 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8687 llvm_unreachable("invalid shift count for narrowing vector shift "
8691 llvm_unreachable("unhandled vector shift");
8695 case Intrinsic::arm_neon_vshifts:
8696 case Intrinsic::arm_neon_vshiftu:
8697 // Opcode already set above.
8699 case Intrinsic::arm_neon_vshiftls:
8700 case Intrinsic::arm_neon_vshiftlu:
8701 if (Cnt == VT.getVectorElementType().getSizeInBits())
8702 VShiftOpc = ARMISD::VSHLLi;
8704 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8705 ARMISD::VSHLLs : ARMISD::VSHLLu);
8707 case Intrinsic::arm_neon_vshiftn:
8708 VShiftOpc = ARMISD::VSHRN; break;
8709 case Intrinsic::arm_neon_vrshifts:
8710 VShiftOpc = ARMISD::VRSHRs; break;
8711 case Intrinsic::arm_neon_vrshiftu:
8712 VShiftOpc = ARMISD::VRSHRu; break;
8713 case Intrinsic::arm_neon_vrshiftn:
8714 VShiftOpc = ARMISD::VRSHRN; break;
8715 case Intrinsic::arm_neon_vqshifts:
8716 VShiftOpc = ARMISD::VQSHLs; break;
8717 case Intrinsic::arm_neon_vqshiftu:
8718 VShiftOpc = ARMISD::VQSHLu; break;
8719 case Intrinsic::arm_neon_vqshiftsu:
8720 VShiftOpc = ARMISD::VQSHLsu; break;
8721 case Intrinsic::arm_neon_vqshiftns:
8722 VShiftOpc = ARMISD::VQSHRNs; break;
8723 case Intrinsic::arm_neon_vqshiftnu:
8724 VShiftOpc = ARMISD::VQSHRNu; break;
8725 case Intrinsic::arm_neon_vqshiftnsu:
8726 VShiftOpc = ARMISD::VQSHRNsu; break;
8727 case Intrinsic::arm_neon_vqrshiftns:
8728 VShiftOpc = ARMISD::VQRSHRNs; break;
8729 case Intrinsic::arm_neon_vqrshiftnu:
8730 VShiftOpc = ARMISD::VQRSHRNu; break;
8731 case Intrinsic::arm_neon_vqrshiftnsu:
8732 VShiftOpc = ARMISD::VQRSHRNsu; break;
8735 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8736 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8739 case Intrinsic::arm_neon_vshiftins: {
8740 EVT VT = N->getOperand(1).getValueType();
8742 unsigned VShiftOpc = 0;
8744 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8745 VShiftOpc = ARMISD::VSLI;
8746 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8747 VShiftOpc = ARMISD::VSRI;
8749 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8752 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8753 N->getOperand(1), N->getOperand(2),
8754 DAG.getConstant(Cnt, MVT::i32));
8757 case Intrinsic::arm_neon_vqrshifts:
8758 case Intrinsic::arm_neon_vqrshiftu:
8759 // No immediate versions of these to check for.
8766 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8767 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8768 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8769 /// vector element shift counts are generally not legal, and it is hard to see
8770 /// their values after they get legalized to loads from a constant pool.
8771 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8772 const ARMSubtarget *ST) {
8773 EVT VT = N->getValueType(0);
8774 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8775 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8776 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8777 SDValue N1 = N->getOperand(1);
8778 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8779 SDValue N0 = N->getOperand(0);
8780 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8781 DAG.MaskedValueIsZero(N0.getOperand(0),
8782 APInt::getHighBitsSet(32, 16)))
8783 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8787 // Nothing to be done for scalar shifts.
8788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8789 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8792 assert(ST->hasNEON() && "unexpected vector shift");
8795 switch (N->getOpcode()) {
8796 default: llvm_unreachable("unexpected shift opcode");
8799 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8800 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8801 DAG.getConstant(Cnt, MVT::i32));
8806 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8807 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8808 ARMISD::VSHRs : ARMISD::VSHRu);
8809 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8810 DAG.getConstant(Cnt, MVT::i32));
8816 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8817 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8818 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8819 const ARMSubtarget *ST) {
8820 SDValue N0 = N->getOperand(0);
8822 // Check for sign- and zero-extensions of vector extract operations of 8-
8823 // and 16-bit vector elements. NEON supports these directly. They are
8824 // handled during DAG combining because type legalization will promote them
8825 // to 32-bit types and it is messy to recognize the operations after that.
8826 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8827 SDValue Vec = N0.getOperand(0);
8828 SDValue Lane = N0.getOperand(1);
8829 EVT VT = N->getValueType(0);
8830 EVT EltVT = N0.getValueType();
8831 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8833 if (VT == MVT::i32 &&
8834 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8835 TLI.isTypeLegal(Vec.getValueType()) &&
8836 isa<ConstantSDNode>(Lane)) {
8839 switch (N->getOpcode()) {
8840 default: llvm_unreachable("unexpected opcode");
8841 case ISD::SIGN_EXTEND:
8842 Opc = ARMISD::VGETLANEs;
8844 case ISD::ZERO_EXTEND:
8845 case ISD::ANY_EXTEND:
8846 Opc = ARMISD::VGETLANEu;
8849 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8856 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8857 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8858 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8859 const ARMSubtarget *ST) {
8860 // If the target supports NEON, try to use vmax/vmin instructions for f32
8861 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8862 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8863 // a NaN; only do the transformation when it matches that behavior.
8865 // For now only do this when using NEON for FP operations; if using VFP, it
8866 // is not obvious that the benefit outweighs the cost of switching to the
8868 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8869 N->getValueType(0) != MVT::f32)
8872 SDValue CondLHS = N->getOperand(0);
8873 SDValue CondRHS = N->getOperand(1);
8874 SDValue LHS = N->getOperand(2);
8875 SDValue RHS = N->getOperand(3);
8876 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8878 unsigned Opcode = 0;
8880 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8881 IsReversed = false; // x CC y ? x : y
8882 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8883 IsReversed = true ; // x CC y ? y : x
8897 // If LHS is NaN, an ordered comparison will be false and the result will
8898 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8899 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8900 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8901 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8903 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8904 // will return -0, so vmin can only be used for unsafe math or if one of
8905 // the operands is known to be nonzero.
8906 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8907 !DAG.getTarget().Options.UnsafeFPMath &&
8908 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8910 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8919 // If LHS is NaN, an ordered comparison will be false and the result will
8920 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8921 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8922 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8923 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8925 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8926 // will return +0, so vmax can only be used for unsafe math or if one of
8927 // the operands is known to be nonzero.
8928 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8929 !DAG.getTarget().Options.UnsafeFPMath &&
8930 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8932 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8938 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8941 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8943 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8944 SDValue Cmp = N->getOperand(4);
8945 if (Cmp.getOpcode() != ARMISD::CMPZ)
8946 // Only looking at EQ and NE cases.
8949 EVT VT = N->getValueType(0);
8950 DebugLoc dl = N->getDebugLoc();
8951 SDValue LHS = Cmp.getOperand(0);
8952 SDValue RHS = Cmp.getOperand(1);
8953 SDValue FalseVal = N->getOperand(0);
8954 SDValue TrueVal = N->getOperand(1);
8955 SDValue ARMcc = N->getOperand(2);
8956 ARMCC::CondCodes CC =
8957 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8975 /// FIXME: Turn this into a target neutral optimization?
8977 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8978 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8979 N->getOperand(3), Cmp);
8980 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8982 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8983 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8984 N->getOperand(3), NewCmp);
8987 if (Res.getNode()) {
8988 APInt KnownZero, KnownOne;
8989 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8990 // Capture demanded bits information that would be otherwise lost.
8991 if (KnownZero == 0xfffffffe)
8992 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8993 DAG.getValueType(MVT::i1));
8994 else if (KnownZero == 0xffffff00)
8995 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8996 DAG.getValueType(MVT::i8));
8997 else if (KnownZero == 0xffff0000)
8998 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8999 DAG.getValueType(MVT::i16));
9005 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9006 DAGCombinerInfo &DCI) const {
9007 switch (N->getOpcode()) {
9009 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9010 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9011 case ISD::SUB: return PerformSUBCombine(N, DCI);
9012 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9013 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9014 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9015 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9016 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9017 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9018 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9019 case ISD::STORE: return PerformSTORECombine(N, DCI);
9020 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9021 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9022 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9023 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9024 case ISD::FP_TO_SINT:
9025 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9026 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9027 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9030 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9031 case ISD::SIGN_EXTEND:
9032 case ISD::ZERO_EXTEND:
9033 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9034 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9035 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9036 case ARMISD::VLD2DUP:
9037 case ARMISD::VLD3DUP:
9038 case ARMISD::VLD4DUP:
9039 return CombineBaseUpdate(N, DCI);
9040 case ISD::INTRINSIC_VOID:
9041 case ISD::INTRINSIC_W_CHAIN:
9042 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9043 case Intrinsic::arm_neon_vld1:
9044 case Intrinsic::arm_neon_vld2:
9045 case Intrinsic::arm_neon_vld3:
9046 case Intrinsic::arm_neon_vld4:
9047 case Intrinsic::arm_neon_vld2lane:
9048 case Intrinsic::arm_neon_vld3lane:
9049 case Intrinsic::arm_neon_vld4lane:
9050 case Intrinsic::arm_neon_vst1:
9051 case Intrinsic::arm_neon_vst2:
9052 case Intrinsic::arm_neon_vst3:
9053 case Intrinsic::arm_neon_vst4:
9054 case Intrinsic::arm_neon_vst2lane:
9055 case Intrinsic::arm_neon_vst3lane:
9056 case Intrinsic::arm_neon_vst4lane:
9057 return CombineBaseUpdate(N, DCI);
9065 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9067 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9070 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9071 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9072 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9074 switch (VT.getSimpleVT().SimpleTy) {
9080 // Unaligned access can use (for example) LRDB, LRDH, LDR
9081 return AllowsUnaligned;
9084 // For any little-endian targets with neon, we can support unaligned ld/st
9085 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9086 // A big-endian target may also explictly support unaligned accesses
9087 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9091 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9092 unsigned AlignCheck) {
9093 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9094 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9097 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9098 unsigned DstAlign, unsigned SrcAlign,
9101 MachineFunction &MF) const {
9102 const Function *F = MF.getFunction();
9104 // See if we can use NEON instructions for this...
9106 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9107 Subtarget->hasNEON()) {
9108 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9110 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9115 // Lowering to i32/i16 if the size permits.
9118 } else if (Size >= 2) {
9122 // Let the target-independent logic figure it out.
9126 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9131 switch (VT.getSimpleVT().SimpleTy) {
9132 default: return false;
9147 if ((V & (Scale - 1)) != 0)
9150 return V == (V & ((1LL << 5) - 1));
9153 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9154 const ARMSubtarget *Subtarget) {
9161 switch (VT.getSimpleVT().SimpleTy) {
9162 default: return false;
9167 // + imm12 or - imm8
9169 return V == (V & ((1LL << 8) - 1));
9170 return V == (V & ((1LL << 12) - 1));
9173 // Same as ARM mode. FIXME: NEON?
9174 if (!Subtarget->hasVFP2())
9179 return V == (V & ((1LL << 8) - 1));
9183 /// isLegalAddressImmediate - Return true if the integer value can be used
9184 /// as the offset of the target addressing mode for load / store of the
9186 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9187 const ARMSubtarget *Subtarget) {
9194 if (Subtarget->isThumb1Only())
9195 return isLegalT1AddressImmediate(V, VT);
9196 else if (Subtarget->isThumb2())
9197 return isLegalT2AddressImmediate(V, VT, Subtarget);
9202 switch (VT.getSimpleVT().SimpleTy) {
9203 default: return false;
9208 return V == (V & ((1LL << 12) - 1));
9211 return V == (V & ((1LL << 8) - 1));
9214 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9219 return V == (V & ((1LL << 8) - 1));
9223 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9225 int Scale = AM.Scale;
9229 switch (VT.getSimpleVT().SimpleTy) {
9230 default: return false;
9239 return Scale == 2 || Scale == 4 || Scale == 8;
9242 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9246 // Note, we allow "void" uses (basically, uses that aren't loads or
9247 // stores), because arm allows folding a scale into many arithmetic
9248 // operations. This should be made more precise and revisited later.
9250 // Allow r << imm, but the imm has to be a multiple of two.
9251 if (Scale & 1) return false;
9252 return isPowerOf2_32(Scale);
9256 /// isLegalAddressingMode - Return true if the addressing mode represented
9257 /// by AM is legal for this target, for a load/store of the specified type.
9258 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9260 EVT VT = getValueType(Ty, true);
9261 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9264 // Can never fold addr of global into load/store.
9269 case 0: // no scale reg, must be "r+i" or "r", or "i".
9272 if (Subtarget->isThumb1Only())
9276 // ARM doesn't support any R+R*scale+imm addr modes.
9283 if (Subtarget->isThumb2())
9284 return isLegalT2ScaledAddressingMode(AM, VT);
9286 int Scale = AM.Scale;
9287 switch (VT.getSimpleVT().SimpleTy) {
9288 default: return false;
9292 if (Scale < 0) Scale = -Scale;
9296 return isPowerOf2_32(Scale & ~1);
9300 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9305 // Note, we allow "void" uses (basically, uses that aren't loads or
9306 // stores), because arm allows folding a scale into many arithmetic
9307 // operations. This should be made more precise and revisited later.
9309 // Allow r << imm, but the imm has to be a multiple of two.
9310 if (Scale & 1) return false;
9311 return isPowerOf2_32(Scale);
9317 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9318 /// icmp immediate, that is the target has icmp instructions which can compare
9319 /// a register against the immediate without having to materialize the
9320 /// immediate into a register.
9321 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9322 // Thumb2 and ARM modes can use cmn for negative immediates.
9323 if (!Subtarget->isThumb())
9324 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9325 if (Subtarget->isThumb2())
9326 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9327 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9328 return Imm >= 0 && Imm <= 255;
9331 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9332 /// *or sub* immediate, that is the target has add or sub instructions which can
9333 /// add a register with the immediate without having to materialize the
9334 /// immediate into a register.
9335 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9336 // Same encoding for add/sub, just flip the sign.
9337 int64_t AbsImm = llvm::abs64(Imm);
9338 if (!Subtarget->isThumb())
9339 return ARM_AM::getSOImmVal(AbsImm) != -1;
9340 if (Subtarget->isThumb2())
9341 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9342 // Thumb1 only has 8-bit unsigned immediate.
9343 return AbsImm >= 0 && AbsImm <= 255;
9346 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9347 bool isSEXTLoad, SDValue &Base,
9348 SDValue &Offset, bool &isInc,
9349 SelectionDAG &DAG) {
9350 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9353 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9355 Base = Ptr->getOperand(0);
9356 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9357 int RHSC = (int)RHS->getZExtValue();
9358 if (RHSC < 0 && RHSC > -256) {
9359 assert(Ptr->getOpcode() == ISD::ADD);
9361 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9365 isInc = (Ptr->getOpcode() == ISD::ADD);
9366 Offset = Ptr->getOperand(1);
9368 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9370 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9371 int RHSC = (int)RHS->getZExtValue();
9372 if (RHSC < 0 && RHSC > -0x1000) {
9373 assert(Ptr->getOpcode() == ISD::ADD);
9375 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9376 Base = Ptr->getOperand(0);
9381 if (Ptr->getOpcode() == ISD::ADD) {
9383 ARM_AM::ShiftOpc ShOpcVal=
9384 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9385 if (ShOpcVal != ARM_AM::no_shift) {
9386 Base = Ptr->getOperand(1);
9387 Offset = Ptr->getOperand(0);
9389 Base = Ptr->getOperand(0);
9390 Offset = Ptr->getOperand(1);
9395 isInc = (Ptr->getOpcode() == ISD::ADD);
9396 Base = Ptr->getOperand(0);
9397 Offset = Ptr->getOperand(1);
9401 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9405 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9406 bool isSEXTLoad, SDValue &Base,
9407 SDValue &Offset, bool &isInc,
9408 SelectionDAG &DAG) {
9409 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9412 Base = Ptr->getOperand(0);
9413 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9414 int RHSC = (int)RHS->getZExtValue();
9415 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9416 assert(Ptr->getOpcode() == ISD::ADD);
9418 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9420 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9421 isInc = Ptr->getOpcode() == ISD::ADD;
9422 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9430 /// getPreIndexedAddressParts - returns true by value, base pointer and
9431 /// offset pointer and addressing mode by reference if the node's address
9432 /// can be legally represented as pre-indexed load / store address.
9434 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9436 ISD::MemIndexedMode &AM,
9437 SelectionDAG &DAG) const {
9438 if (Subtarget->isThumb1Only())
9443 bool isSEXTLoad = false;
9444 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9445 Ptr = LD->getBasePtr();
9446 VT = LD->getMemoryVT();
9447 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9448 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9449 Ptr = ST->getBasePtr();
9450 VT = ST->getMemoryVT();
9455 bool isLegal = false;
9456 if (Subtarget->isThumb2())
9457 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9458 Offset, isInc, DAG);
9460 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9461 Offset, isInc, DAG);
9465 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9469 /// getPostIndexedAddressParts - returns true by value, base pointer and
9470 /// offset pointer and addressing mode by reference if this node can be
9471 /// combined with a load / store to form a post-indexed load / store.
9472 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9475 ISD::MemIndexedMode &AM,
9476 SelectionDAG &DAG) const {
9477 if (Subtarget->isThumb1Only())
9482 bool isSEXTLoad = false;
9483 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9484 VT = LD->getMemoryVT();
9485 Ptr = LD->getBasePtr();
9486 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9487 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9488 VT = ST->getMemoryVT();
9489 Ptr = ST->getBasePtr();
9494 bool isLegal = false;
9495 if (Subtarget->isThumb2())
9496 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9499 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9505 // Swap base ptr and offset to catch more post-index load / store when
9506 // it's legal. In Thumb2 mode, offset must be an immediate.
9507 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9508 !Subtarget->isThumb2())
9509 std::swap(Base, Offset);
9511 // Post-indexed load / store update the base pointer.
9516 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9520 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9523 const SelectionDAG &DAG,
9524 unsigned Depth) const {
9525 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9526 switch (Op.getOpcode()) {
9528 case ARMISD::CMOV: {
9529 // Bits are known zero/one if known on the LHS and RHS.
9530 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9531 if (KnownZero == 0 && KnownOne == 0) return;
9533 APInt KnownZeroRHS, KnownOneRHS;
9534 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9535 KnownZero &= KnownZeroRHS;
9536 KnownOne &= KnownOneRHS;
9542 //===----------------------------------------------------------------------===//
9543 // ARM Inline Assembly Support
9544 //===----------------------------------------------------------------------===//
9546 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9547 // Looking for "rev" which is V6+.
9548 if (!Subtarget->hasV6Ops())
9551 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9552 std::string AsmStr = IA->getAsmString();
9553 SmallVector<StringRef, 4> AsmPieces;
9554 SplitString(AsmStr, AsmPieces, ";\n");
9556 switch (AsmPieces.size()) {
9557 default: return false;
9559 AsmStr = AsmPieces[0];
9561 SplitString(AsmStr, AsmPieces, " \t,");
9564 if (AsmPieces.size() == 3 &&
9565 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9566 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9567 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9568 if (Ty && Ty->getBitWidth() == 32)
9569 return IntrinsicLowering::LowerToByteSwap(CI);
9577 /// getConstraintType - Given a constraint letter, return the type of
9578 /// constraint it is for this target.
9579 ARMTargetLowering::ConstraintType
9580 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9581 if (Constraint.size() == 1) {
9582 switch (Constraint[0]) {
9584 case 'l': return C_RegisterClass;
9585 case 'w': return C_RegisterClass;
9586 case 'h': return C_RegisterClass;
9587 case 'x': return C_RegisterClass;
9588 case 't': return C_RegisterClass;
9589 case 'j': return C_Other; // Constant for movw.
9590 // An address with a single base register. Due to the way we
9591 // currently handle addresses it is the same as an 'r' memory constraint.
9592 case 'Q': return C_Memory;
9594 } else if (Constraint.size() == 2) {
9595 switch (Constraint[0]) {
9597 // All 'U+' constraints are addresses.
9598 case 'U': return C_Memory;
9601 return TargetLowering::getConstraintType(Constraint);
9604 /// Examine constraint type and operand type and determine a weight value.
9605 /// This object must already have been set up with the operand type
9606 /// and the current alternative constraint selected.
9607 TargetLowering::ConstraintWeight
9608 ARMTargetLowering::getSingleConstraintMatchWeight(
9609 AsmOperandInfo &info, const char *constraint) const {
9610 ConstraintWeight weight = CW_Invalid;
9611 Value *CallOperandVal = info.CallOperandVal;
9612 // If we don't have a value, we can't do a match,
9613 // but allow it at the lowest weight.
9614 if (CallOperandVal == NULL)
9616 Type *type = CallOperandVal->getType();
9617 // Look at the constraint type.
9618 switch (*constraint) {
9620 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9623 if (type->isIntegerTy()) {
9624 if (Subtarget->isThumb())
9625 weight = CW_SpecificReg;
9627 weight = CW_Register;
9631 if (type->isFloatingPointTy())
9632 weight = CW_Register;
9638 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9640 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9642 if (Constraint.size() == 1) {
9643 // GCC ARM Constraint Letters
9644 switch (Constraint[0]) {
9645 case 'l': // Low regs or general regs.
9646 if (Subtarget->isThumb())
9647 return RCPair(0U, &ARM::tGPRRegClass);
9648 return RCPair(0U, &ARM::GPRRegClass);
9649 case 'h': // High regs or no regs.
9650 if (Subtarget->isThumb())
9651 return RCPair(0U, &ARM::hGPRRegClass);
9654 return RCPair(0U, &ARM::GPRRegClass);
9657 return RCPair(0U, &ARM::SPRRegClass);
9658 if (VT.getSizeInBits() == 64)
9659 return RCPair(0U, &ARM::DPRRegClass);
9660 if (VT.getSizeInBits() == 128)
9661 return RCPair(0U, &ARM::QPRRegClass);
9665 return RCPair(0U, &ARM::SPR_8RegClass);
9666 if (VT.getSizeInBits() == 64)
9667 return RCPair(0U, &ARM::DPR_8RegClass);
9668 if (VT.getSizeInBits() == 128)
9669 return RCPair(0U, &ARM::QPR_8RegClass);
9673 return RCPair(0U, &ARM::SPRRegClass);
9677 if (StringRef("{cc}").equals_lower(Constraint))
9678 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9680 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9683 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9684 /// vector. If it is invalid, don't add anything to Ops.
9685 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9686 std::string &Constraint,
9687 std::vector<SDValue>&Ops,
9688 SelectionDAG &DAG) const {
9689 SDValue Result(0, 0);
9691 // Currently only support length 1 constraints.
9692 if (Constraint.length() != 1) return;
9694 char ConstraintLetter = Constraint[0];
9695 switch (ConstraintLetter) {
9698 case 'I': case 'J': case 'K': case 'L':
9699 case 'M': case 'N': case 'O':
9700 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9704 int64_t CVal64 = C->getSExtValue();
9705 int CVal = (int) CVal64;
9706 // None of these constraints allow values larger than 32 bits. Check
9707 // that the value fits in an int.
9711 switch (ConstraintLetter) {
9713 // Constant suitable for movw, must be between 0 and
9715 if (Subtarget->hasV6T2Ops())
9716 if (CVal >= 0 && CVal <= 65535)
9720 if (Subtarget->isThumb1Only()) {
9721 // This must be a constant between 0 and 255, for ADD
9723 if (CVal >= 0 && CVal <= 255)
9725 } else if (Subtarget->isThumb2()) {
9726 // A constant that can be used as an immediate value in a
9727 // data-processing instruction.
9728 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9731 // A constant that can be used as an immediate value in a
9732 // data-processing instruction.
9733 if (ARM_AM::getSOImmVal(CVal) != -1)
9739 if (Subtarget->isThumb()) { // FIXME thumb2
9740 // This must be a constant between -255 and -1, for negated ADD
9741 // immediates. This can be used in GCC with an "n" modifier that
9742 // prints the negated value, for use with SUB instructions. It is
9743 // not useful otherwise but is implemented for compatibility.
9744 if (CVal >= -255 && CVal <= -1)
9747 // This must be a constant between -4095 and 4095. It is not clear
9748 // what this constraint is intended for. Implemented for
9749 // compatibility with GCC.
9750 if (CVal >= -4095 && CVal <= 4095)
9756 if (Subtarget->isThumb1Only()) {
9757 // A 32-bit value where only one byte has a nonzero value. Exclude
9758 // zero to match GCC. This constraint is used by GCC internally for
9759 // constants that can be loaded with a move/shift combination.
9760 // It is not useful otherwise but is implemented for compatibility.
9761 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9763 } else if (Subtarget->isThumb2()) {
9764 // A constant whose bitwise inverse can be used as an immediate
9765 // value in a data-processing instruction. This can be used in GCC
9766 // with a "B" modifier that prints the inverted value, for use with
9767 // BIC and MVN instructions. It is not useful otherwise but is
9768 // implemented for compatibility.
9769 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9772 // A constant whose bitwise inverse can be used as an immediate
9773 // value in a data-processing instruction. This can be used in GCC
9774 // with a "B" modifier that prints the inverted value, for use with
9775 // BIC and MVN instructions. It is not useful otherwise but is
9776 // implemented for compatibility.
9777 if (ARM_AM::getSOImmVal(~CVal) != -1)
9783 if (Subtarget->isThumb1Only()) {
9784 // This must be a constant between -7 and 7,
9785 // for 3-operand ADD/SUB immediate instructions.
9786 if (CVal >= -7 && CVal < 7)
9788 } else if (Subtarget->isThumb2()) {
9789 // A constant whose negation can be used as an immediate value in a
9790 // data-processing instruction. This can be used in GCC with an "n"
9791 // modifier that prints the negated value, for use with SUB
9792 // instructions. It is not useful otherwise but is implemented for
9794 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9797 // A constant whose negation can be used as an immediate value in a
9798 // data-processing instruction. This can be used in GCC with an "n"
9799 // modifier that prints the negated value, for use with SUB
9800 // instructions. It is not useful otherwise but is implemented for
9802 if (ARM_AM::getSOImmVal(-CVal) != -1)
9808 if (Subtarget->isThumb()) { // FIXME thumb2
9809 // This must be a multiple of 4 between 0 and 1020, for
9810 // ADD sp + immediate.
9811 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9814 // A power of two or a constant between 0 and 32. This is used in
9815 // GCC for the shift amount on shifted register operands, but it is
9816 // useful in general for any shift amounts.
9817 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9823 if (Subtarget->isThumb()) { // FIXME thumb2
9824 // This must be a constant between 0 and 31, for shift amounts.
9825 if (CVal >= 0 && CVal <= 31)
9831 if (Subtarget->isThumb()) { // FIXME thumb2
9832 // This must be a multiple of 4 between -508 and 508, for
9833 // ADD/SUB sp = sp + immediate.
9834 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9839 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9843 if (Result.getNode()) {
9844 Ops.push_back(Result);
9847 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9851 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9852 // The ARM target isn't yet aware of offsets.
9856 bool ARM::isBitFieldInvertedMask(unsigned v) {
9857 if (v == 0xffffffff)
9859 // there can be 1's on either or both "outsides", all the "inside"
9861 unsigned int lsb = 0, msb = 31;
9862 while (v & (1 << msb)) --msb;
9863 while (v & (1 << lsb)) ++lsb;
9864 for (unsigned int i = lsb; i <= msb; ++i) {
9871 /// isFPImmLegal - Returns true if the target can instruction select the
9872 /// specified FP immediate natively. If false, the legalizer will
9873 /// materialize the FP immediate as a load from a constant pool.
9874 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9875 if (!Subtarget->hasVFP3())
9878 return ARM_AM::getFP32Imm(Imm) != -1;
9880 return ARM_AM::getFP64Imm(Imm) != -1;
9884 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9885 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9886 /// specified in the intrinsic calls.
9887 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9889 unsigned Intrinsic) const {
9890 switch (Intrinsic) {
9891 case Intrinsic::arm_neon_vld1:
9892 case Intrinsic::arm_neon_vld2:
9893 case Intrinsic::arm_neon_vld3:
9894 case Intrinsic::arm_neon_vld4:
9895 case Intrinsic::arm_neon_vld2lane:
9896 case Intrinsic::arm_neon_vld3lane:
9897 case Intrinsic::arm_neon_vld4lane: {
9898 Info.opc = ISD::INTRINSIC_W_CHAIN;
9899 // Conservatively set memVT to the entire set of vectors loaded.
9900 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9901 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9902 Info.ptrVal = I.getArgOperand(0);
9904 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9905 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9906 Info.vol = false; // volatile loads with NEON intrinsics not supported
9907 Info.readMem = true;
9908 Info.writeMem = false;
9911 case Intrinsic::arm_neon_vst1:
9912 case Intrinsic::arm_neon_vst2:
9913 case Intrinsic::arm_neon_vst3:
9914 case Intrinsic::arm_neon_vst4:
9915 case Intrinsic::arm_neon_vst2lane:
9916 case Intrinsic::arm_neon_vst3lane:
9917 case Intrinsic::arm_neon_vst4lane: {
9918 Info.opc = ISD::INTRINSIC_VOID;
9919 // Conservatively set memVT to the entire set of vectors stored.
9920 unsigned NumElts = 0;
9921 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9922 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9923 if (!ArgTy->isVectorTy())
9925 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9927 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9928 Info.ptrVal = I.getArgOperand(0);
9930 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9931 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9932 Info.vol = false; // volatile stores with NEON intrinsics not supported
9933 Info.readMem = false;
9934 Info.writeMem = true;
9937 case Intrinsic::arm_strexd: {
9938 Info.opc = ISD::INTRINSIC_W_CHAIN;
9939 Info.memVT = MVT::i64;
9940 Info.ptrVal = I.getArgOperand(2);
9944 Info.readMem = false;
9945 Info.writeMem = true;
9948 case Intrinsic::arm_ldrexd: {
9949 Info.opc = ISD::INTRINSIC_W_CHAIN;
9950 Info.memVT = MVT::i64;
9951 Info.ptrVal = I.getArgOperand(0);
9955 Info.readMem = true;
9956 Info.writeMem = false;