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::FABS, MVT::v4f32, Expand);
509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
525 // Neon does not support some operations on v1i64 and v2i64 types.
526 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
527 // Custom handling for some quad-vector types to detect VMULL.
528 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
529 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
530 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
531 // Custom handling for some vector types to avoid expensive expansions
532 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
533 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
534 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
535 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
536 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
537 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
538 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
539 // a destination type that is wider than the source, and nor does
540 // it have a FP_TO_[SU]INT instruction with a narrower destination than
542 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
543 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
544 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
545 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
547 setTargetDAGCombine(ISD::INTRINSIC_VOID);
548 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
549 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
550 setTargetDAGCombine(ISD::SHL);
551 setTargetDAGCombine(ISD::SRL);
552 setTargetDAGCombine(ISD::SRA);
553 setTargetDAGCombine(ISD::SIGN_EXTEND);
554 setTargetDAGCombine(ISD::ZERO_EXTEND);
555 setTargetDAGCombine(ISD::ANY_EXTEND);
556 setTargetDAGCombine(ISD::SELECT_CC);
557 setTargetDAGCombine(ISD::BUILD_VECTOR);
558 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
559 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
560 setTargetDAGCombine(ISD::STORE);
561 setTargetDAGCombine(ISD::FP_TO_SINT);
562 setTargetDAGCombine(ISD::FP_TO_UINT);
563 setTargetDAGCombine(ISD::FDIV);
565 // It is legal to extload from v4i8 to v4i16 or v4i32.
566 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
567 MVT::v4i16, MVT::v2i16,
569 for (unsigned i = 0; i < 6; ++i) {
570 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
571 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
572 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
576 // ARM and Thumb2 support UMLAL/SMLAL.
577 if (!Subtarget->isThumb1Only())
578 setTargetDAGCombine(ISD::ADDC);
581 computeRegisterProperties();
583 // ARM does not have f32 extending load.
584 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
586 // ARM does not have i1 sign extending load.
587 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
589 // ARM supports all 4 flavors of integer indexed load / store.
590 if (!Subtarget->isThumb1Only()) {
591 for (unsigned im = (unsigned)ISD::PRE_INC;
592 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
593 setIndexedLoadAction(im, MVT::i1, Legal);
594 setIndexedLoadAction(im, MVT::i8, Legal);
595 setIndexedLoadAction(im, MVT::i16, Legal);
596 setIndexedLoadAction(im, MVT::i32, Legal);
597 setIndexedStoreAction(im, MVT::i1, Legal);
598 setIndexedStoreAction(im, MVT::i8, Legal);
599 setIndexedStoreAction(im, MVT::i16, Legal);
600 setIndexedStoreAction(im, MVT::i32, Legal);
604 // i64 operation support.
605 setOperationAction(ISD::MUL, MVT::i64, Expand);
606 setOperationAction(ISD::MULHU, MVT::i32, Expand);
607 if (Subtarget->isThumb1Only()) {
608 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
609 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
611 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
612 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
613 setOperationAction(ISD::MULHS, MVT::i32, Expand);
615 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
616 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
617 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
618 setOperationAction(ISD::SRL, MVT::i64, Custom);
619 setOperationAction(ISD::SRA, MVT::i64, Custom);
621 if (!Subtarget->isThumb1Only()) {
622 // FIXME: We should do this for Thumb1 as well.
623 setOperationAction(ISD::ADDC, MVT::i32, Custom);
624 setOperationAction(ISD::ADDE, MVT::i32, Custom);
625 setOperationAction(ISD::SUBC, MVT::i32, Custom);
626 setOperationAction(ISD::SUBE, MVT::i32, Custom);
629 // ARM does not have ROTL.
630 setOperationAction(ISD::ROTL, MVT::i32, Expand);
631 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
632 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
633 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
634 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
636 // These just redirect to CTTZ and CTLZ on ARM.
637 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
638 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
640 // Only ARMv6 has BSWAP.
641 if (!Subtarget->hasV6Ops())
642 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
644 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
645 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
646 // These are expanded into libcalls if the cpu doesn't have HW divider.
647 setOperationAction(ISD::SDIV, MVT::i32, Expand);
648 setOperationAction(ISD::UDIV, MVT::i32, Expand);
650 setOperationAction(ISD::SREM, MVT::i32, Expand);
651 setOperationAction(ISD::UREM, MVT::i32, Expand);
652 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
653 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
655 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
656 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
657 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
658 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
659 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
661 setOperationAction(ISD::TRAP, MVT::Other, Legal);
663 // Use the default implementation.
664 setOperationAction(ISD::VASTART, MVT::Other, Custom);
665 setOperationAction(ISD::VAARG, MVT::Other, Expand);
666 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
667 setOperationAction(ISD::VAEND, MVT::Other, Expand);
668 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
669 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
671 if (!Subtarget->isTargetDarwin()) {
672 // Non-Darwin platforms may return values in these registers via the
673 // personality function.
674 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
675 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
676 setExceptionPointerRegister(ARM::R0);
677 setExceptionSelectorRegister(ARM::R1);
680 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
681 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
682 // the default expansion.
683 // FIXME: This should be checking for v6k, not just v6.
684 if (Subtarget->hasDataBarrier() ||
685 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
686 // membarrier needs custom lowering; the rest are legal and handled
688 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
689 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
690 // Custom lowering for 64-bit ops
691 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
692 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
693 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
694 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
695 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
696 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
697 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
698 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
699 setInsertFencesForAtomic(true);
701 // Set them all for expansion, which will force libcalls.
702 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
703 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
704 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
711 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
712 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
713 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
714 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
715 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
716 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
717 // Unordered/Monotonic case.
718 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
719 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
720 // Since the libcalls include locking, fold in the fences
721 setShouldFoldAtomicFences(true);
724 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
726 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
727 if (!Subtarget->hasV6Ops()) {
728 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
729 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
731 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
733 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
734 !Subtarget->isThumb1Only()) {
735 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
736 // iff target supports vfp2.
737 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
738 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
741 // We want to custom lower some of our intrinsics.
742 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
743 if (Subtarget->isTargetDarwin()) {
744 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
745 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
746 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
749 setOperationAction(ISD::SETCC, MVT::i32, Expand);
750 setOperationAction(ISD::SETCC, MVT::f32, Expand);
751 setOperationAction(ISD::SETCC, MVT::f64, Expand);
752 setOperationAction(ISD::SELECT, MVT::i32, Custom);
753 setOperationAction(ISD::SELECT, MVT::f32, Custom);
754 setOperationAction(ISD::SELECT, MVT::f64, Custom);
755 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
756 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
757 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
759 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
760 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
761 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
762 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
763 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
765 // We don't support sin/cos/fmod/copysign/pow
766 setOperationAction(ISD::FSIN, MVT::f64, Expand);
767 setOperationAction(ISD::FSIN, MVT::f32, Expand);
768 setOperationAction(ISD::FCOS, MVT::f32, Expand);
769 setOperationAction(ISD::FCOS, MVT::f64, Expand);
770 setOperationAction(ISD::FREM, MVT::f64, Expand);
771 setOperationAction(ISD::FREM, MVT::f32, Expand);
772 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
773 !Subtarget->isThumb1Only()) {
774 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
775 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
777 setOperationAction(ISD::FPOW, MVT::f64, Expand);
778 setOperationAction(ISD::FPOW, MVT::f32, Expand);
780 if (!Subtarget->hasVFP4()) {
781 setOperationAction(ISD::FMA, MVT::f64, Expand);
782 setOperationAction(ISD::FMA, MVT::f32, Expand);
785 // Various VFP goodness
786 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
787 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
788 if (Subtarget->hasVFP2()) {
789 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
790 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
791 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
792 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
794 // Special handling for half-precision FP.
795 if (!Subtarget->hasFP16()) {
796 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
797 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
801 // We have target-specific dag combine patterns for the following nodes:
802 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
803 setTargetDAGCombine(ISD::ADD);
804 setTargetDAGCombine(ISD::SUB);
805 setTargetDAGCombine(ISD::MUL);
806 setTargetDAGCombine(ISD::AND);
807 setTargetDAGCombine(ISD::OR);
808 setTargetDAGCombine(ISD::XOR);
810 if (Subtarget->hasV6Ops())
811 setTargetDAGCombine(ISD::SRL);
813 setStackPointerRegisterToSaveRestore(ARM::SP);
815 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
816 !Subtarget->hasVFP2())
817 setSchedulingPreference(Sched::RegPressure);
819 setSchedulingPreference(Sched::Hybrid);
821 //// temporary - rewrite interface to use type
822 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
823 maxStoresPerMemset = 16;
824 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
826 // On ARM arguments smaller than 4 bytes are extended, so all arguments
827 // are at least 4 bytes aligned.
828 setMinStackArgumentAlignment(4);
830 benefitFromCodePlacementOpt = true;
832 // Prefer likely predicted branches to selects on out-of-order cores.
833 predictableSelectIsExpensive = Subtarget->isLikeA9();
835 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
838 // FIXME: It might make sense to define the representative register class as the
839 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
840 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
841 // SPR's representative would be DPR_VFP2. This should work well if register
842 // pressure tracking were modified such that a register use would increment the
843 // pressure of the register class's representative and all of it's super
844 // classes' representatives transitively. We have not implemented this because
845 // of the difficulty prior to coalescing of modeling operand register classes
846 // due to the common occurrence of cross class copies and subregister insertions
848 std::pair<const TargetRegisterClass*, uint8_t>
849 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
850 const TargetRegisterClass *RRC = 0;
852 switch (VT.getSimpleVT().SimpleTy) {
854 return TargetLowering::findRepresentativeClass(VT);
855 // Use DPR as representative register class for all floating point
856 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
857 // the cost is 1 for both f32 and f64.
858 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
859 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
860 RRC = &ARM::DPRRegClass;
861 // When NEON is used for SP, only half of the register file is available
862 // because operations that define both SP and DP results will be constrained
863 // to the VFP2 class (D0-D15). We currently model this constraint prior to
864 // coalescing by double-counting the SP regs. See the FIXME above.
865 if (Subtarget->useNEONForSinglePrecisionFP())
868 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
869 case MVT::v4f32: case MVT::v2f64:
870 RRC = &ARM::DPRRegClass;
874 RRC = &ARM::DPRRegClass;
878 RRC = &ARM::DPRRegClass;
882 return std::make_pair(RRC, Cost);
885 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
888 case ARMISD::Wrapper: return "ARMISD::Wrapper";
889 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
890 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
891 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
892 case ARMISD::CALL: return "ARMISD::CALL";
893 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
894 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
895 case ARMISD::tCALL: return "ARMISD::tCALL";
896 case ARMISD::BRCOND: return "ARMISD::BRCOND";
897 case ARMISD::BR_JT: return "ARMISD::BR_JT";
898 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
899 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
900 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
901 case ARMISD::CMP: return "ARMISD::CMP";
902 case ARMISD::CMN: return "ARMISD::CMN";
903 case ARMISD::CMPZ: return "ARMISD::CMPZ";
904 case ARMISD::CMPFP: return "ARMISD::CMPFP";
905 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
906 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
907 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
909 case ARMISD::CMOV: return "ARMISD::CMOV";
911 case ARMISD::RBIT: return "ARMISD::RBIT";
913 case ARMISD::FTOSI: return "ARMISD::FTOSI";
914 case ARMISD::FTOUI: return "ARMISD::FTOUI";
915 case ARMISD::SITOF: return "ARMISD::SITOF";
916 case ARMISD::UITOF: return "ARMISD::UITOF";
918 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
919 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
920 case ARMISD::RRX: return "ARMISD::RRX";
922 case ARMISD::ADDC: return "ARMISD::ADDC";
923 case ARMISD::ADDE: return "ARMISD::ADDE";
924 case ARMISD::SUBC: return "ARMISD::SUBC";
925 case ARMISD::SUBE: return "ARMISD::SUBE";
927 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
928 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
930 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
931 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
933 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
935 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
937 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
939 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
940 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
942 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
944 case ARMISD::VCEQ: return "ARMISD::VCEQ";
945 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
946 case ARMISD::VCGE: return "ARMISD::VCGE";
947 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
948 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
949 case ARMISD::VCGEU: return "ARMISD::VCGEU";
950 case ARMISD::VCGT: return "ARMISD::VCGT";
951 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
952 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
953 case ARMISD::VCGTU: return "ARMISD::VCGTU";
954 case ARMISD::VTST: return "ARMISD::VTST";
956 case ARMISD::VSHL: return "ARMISD::VSHL";
957 case ARMISD::VSHRs: return "ARMISD::VSHRs";
958 case ARMISD::VSHRu: return "ARMISD::VSHRu";
959 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
960 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
961 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
962 case ARMISD::VSHRN: return "ARMISD::VSHRN";
963 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
964 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
965 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
966 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
967 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
968 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
969 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
970 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
971 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
972 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
973 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
974 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
975 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
976 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
977 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
978 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
979 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
980 case ARMISD::VDUP: return "ARMISD::VDUP";
981 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
982 case ARMISD::VEXT: return "ARMISD::VEXT";
983 case ARMISD::VREV64: return "ARMISD::VREV64";
984 case ARMISD::VREV32: return "ARMISD::VREV32";
985 case ARMISD::VREV16: return "ARMISD::VREV16";
986 case ARMISD::VZIP: return "ARMISD::VZIP";
987 case ARMISD::VUZP: return "ARMISD::VUZP";
988 case ARMISD::VTRN: return "ARMISD::VTRN";
989 case ARMISD::VTBL1: return "ARMISD::VTBL1";
990 case ARMISD::VTBL2: return "ARMISD::VTBL2";
991 case ARMISD::VMULLs: return "ARMISD::VMULLs";
992 case ARMISD::VMULLu: return "ARMISD::VMULLu";
993 case ARMISD::UMLAL: return "ARMISD::UMLAL";
994 case ARMISD::SMLAL: return "ARMISD::SMLAL";
995 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
996 case ARMISD::FMAX: return "ARMISD::FMAX";
997 case ARMISD::FMIN: return "ARMISD::FMIN";
998 case ARMISD::BFI: return "ARMISD::BFI";
999 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1000 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1001 case ARMISD::VBSL: return "ARMISD::VBSL";
1002 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1003 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1004 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1005 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1006 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1007 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1008 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1009 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1010 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1011 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1012 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1013 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1014 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1015 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1016 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1017 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1018 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1019 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1020 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1021 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1025 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1026 if (!VT.isVector()) return getPointerTy();
1027 return VT.changeVectorElementTypeToInteger();
1030 /// getRegClassFor - Return the register class that should be used for the
1031 /// specified value type.
1032 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1033 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1034 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1035 // load / store 4 to 8 consecutive D registers.
1036 if (Subtarget->hasNEON()) {
1037 if (VT == MVT::v4i64)
1038 return &ARM::QQPRRegClass;
1039 if (VT == MVT::v8i64)
1040 return &ARM::QQQQPRRegClass;
1042 return TargetLowering::getRegClassFor(VT);
1045 // Create a fast isel object.
1047 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1048 const TargetLibraryInfo *libInfo) const {
1049 return ARM::createFastISel(funcInfo, libInfo);
1052 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1053 /// be used for loads / stores from the global.
1054 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1055 return (Subtarget->isThumb1Only() ? 127 : 4095);
1058 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1059 unsigned NumVals = N->getNumValues();
1061 return Sched::RegPressure;
1063 for (unsigned i = 0; i != NumVals; ++i) {
1064 EVT VT = N->getValueType(i);
1065 if (VT == MVT::Glue || VT == MVT::Other)
1067 if (VT.isFloatingPoint() || VT.isVector())
1071 if (!N->isMachineOpcode())
1072 return Sched::RegPressure;
1074 // Load are scheduled for latency even if there instruction itinerary
1075 // is not available.
1076 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1077 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1079 if (MCID.getNumDefs() == 0)
1080 return Sched::RegPressure;
1081 if (!Itins->isEmpty() &&
1082 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1085 return Sched::RegPressure;
1088 //===----------------------------------------------------------------------===//
1090 //===----------------------------------------------------------------------===//
1092 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1093 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1095 default: llvm_unreachable("Unknown condition code!");
1096 case ISD::SETNE: return ARMCC::NE;
1097 case ISD::SETEQ: return ARMCC::EQ;
1098 case ISD::SETGT: return ARMCC::GT;
1099 case ISD::SETGE: return ARMCC::GE;
1100 case ISD::SETLT: return ARMCC::LT;
1101 case ISD::SETLE: return ARMCC::LE;
1102 case ISD::SETUGT: return ARMCC::HI;
1103 case ISD::SETUGE: return ARMCC::HS;
1104 case ISD::SETULT: return ARMCC::LO;
1105 case ISD::SETULE: return ARMCC::LS;
1109 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1110 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1111 ARMCC::CondCodes &CondCode2) {
1112 CondCode2 = ARMCC::AL;
1114 default: llvm_unreachable("Unknown FP condition!");
1116 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1118 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1120 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1121 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1122 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1123 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1124 case ISD::SETO: CondCode = ARMCC::VC; break;
1125 case ISD::SETUO: CondCode = ARMCC::VS; break;
1126 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1127 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1128 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1130 case ISD::SETULT: CondCode = ARMCC::LT; break;
1132 case ISD::SETULE: CondCode = ARMCC::LE; break;
1134 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1138 //===----------------------------------------------------------------------===//
1139 // Calling Convention Implementation
1140 //===----------------------------------------------------------------------===//
1142 #include "ARMGenCallingConv.inc"
1144 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1145 /// given CallingConvention value.
1146 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1148 bool isVarArg) const {
1151 llvm_unreachable("Unsupported calling convention");
1152 case CallingConv::Fast:
1153 if (Subtarget->hasVFP2() && !isVarArg) {
1154 if (!Subtarget->isAAPCS_ABI())
1155 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1156 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1157 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1160 case CallingConv::C: {
1161 // Use target triple & subtarget features to do actual dispatch.
1162 if (!Subtarget->isAAPCS_ABI())
1163 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1164 else if (Subtarget->hasVFP2() &&
1165 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1167 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1168 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1170 case CallingConv::ARM_AAPCS_VFP:
1172 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1174 case CallingConv::ARM_AAPCS:
1175 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1176 case CallingConv::ARM_APCS:
1177 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1178 case CallingConv::GHC:
1179 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1183 /// LowerCallResult - Lower the result values of a call into the
1184 /// appropriate copies out of appropriate physical registers.
1186 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1187 CallingConv::ID CallConv, bool isVarArg,
1188 const SmallVectorImpl<ISD::InputArg> &Ins,
1189 DebugLoc dl, SelectionDAG &DAG,
1190 SmallVectorImpl<SDValue> &InVals) const {
1192 // Assign locations to each value returned by this call.
1193 SmallVector<CCValAssign, 16> RVLocs;
1194 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1195 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1196 CCInfo.AnalyzeCallResult(Ins,
1197 CCAssignFnForNode(CallConv, /* Return*/ true,
1200 // Copy all of the result registers out of their specified physreg.
1201 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1202 CCValAssign VA = RVLocs[i];
1205 if (VA.needsCustom()) {
1206 // Handle f64 or half of a v2f64.
1207 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1209 Chain = Lo.getValue(1);
1210 InFlag = Lo.getValue(2);
1211 VA = RVLocs[++i]; // skip ahead to next loc
1212 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1214 Chain = Hi.getValue(1);
1215 InFlag = Hi.getValue(2);
1216 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1218 if (VA.getLocVT() == MVT::v2f64) {
1219 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1220 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1221 DAG.getConstant(0, MVT::i32));
1223 VA = RVLocs[++i]; // skip ahead to next loc
1224 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1225 Chain = Lo.getValue(1);
1226 InFlag = Lo.getValue(2);
1227 VA = RVLocs[++i]; // skip ahead to next loc
1228 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1229 Chain = Hi.getValue(1);
1230 InFlag = Hi.getValue(2);
1231 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1232 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1233 DAG.getConstant(1, MVT::i32));
1236 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1238 Chain = Val.getValue(1);
1239 InFlag = Val.getValue(2);
1242 switch (VA.getLocInfo()) {
1243 default: llvm_unreachable("Unknown loc info!");
1244 case CCValAssign::Full: break;
1245 case CCValAssign::BCvt:
1246 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1250 InVals.push_back(Val);
1256 /// LowerMemOpCallTo - Store the argument to the stack.
1258 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1259 SDValue StackPtr, SDValue Arg,
1260 DebugLoc dl, SelectionDAG &DAG,
1261 const CCValAssign &VA,
1262 ISD::ArgFlagsTy Flags) const {
1263 unsigned LocMemOffset = VA.getLocMemOffset();
1264 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1265 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1266 return DAG.getStore(Chain, dl, Arg, PtrOff,
1267 MachinePointerInfo::getStack(LocMemOffset),
1271 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1272 SDValue Chain, SDValue &Arg,
1273 RegsToPassVector &RegsToPass,
1274 CCValAssign &VA, CCValAssign &NextVA,
1276 SmallVector<SDValue, 8> &MemOpChains,
1277 ISD::ArgFlagsTy Flags) const {
1279 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1280 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1281 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1283 if (NextVA.isRegLoc())
1284 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1286 assert(NextVA.isMemLoc());
1287 if (StackPtr.getNode() == 0)
1288 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1290 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1296 /// LowerCall - Lowering a call into a callseq_start <-
1297 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1300 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1301 SmallVectorImpl<SDValue> &InVals) const {
1302 SelectionDAG &DAG = CLI.DAG;
1303 DebugLoc &dl = CLI.DL;
1304 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1305 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1306 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1307 SDValue Chain = CLI.Chain;
1308 SDValue Callee = CLI.Callee;
1309 bool &isTailCall = CLI.IsTailCall;
1310 CallingConv::ID CallConv = CLI.CallConv;
1311 bool doesNotRet = CLI.DoesNotReturn;
1312 bool isVarArg = CLI.IsVarArg;
1314 MachineFunction &MF = DAG.getMachineFunction();
1315 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1316 bool IsSibCall = false;
1317 // Disable tail calls if they're not supported.
1318 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1321 // Check if it's really possible to do a tail call.
1322 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1323 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1324 Outs, OutVals, Ins, DAG);
1325 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1326 // detected sibcalls.
1333 // Analyze operands of the call, assigning locations to each operand.
1334 SmallVector<CCValAssign, 16> ArgLocs;
1335 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1336 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1337 CCInfo.AnalyzeCallOperands(Outs,
1338 CCAssignFnForNode(CallConv, /* Return*/ false,
1341 // Get a count of how many bytes are to be pushed on the stack.
1342 unsigned NumBytes = CCInfo.getNextStackOffset();
1344 // For tail calls, memory operands are available in our caller's stack.
1348 // Adjust the stack pointer for the new arguments...
1349 // These operations are automatically eliminated by the prolog/epilog pass
1351 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1353 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1355 RegsToPassVector RegsToPass;
1356 SmallVector<SDValue, 8> MemOpChains;
1358 // Walk the register/memloc assignments, inserting copies/loads. In the case
1359 // of tail call optimization, arguments are handled later.
1360 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1362 ++i, ++realArgIdx) {
1363 CCValAssign &VA = ArgLocs[i];
1364 SDValue Arg = OutVals[realArgIdx];
1365 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1366 bool isByVal = Flags.isByVal();
1368 // Promote the value if needed.
1369 switch (VA.getLocInfo()) {
1370 default: llvm_unreachable("Unknown loc info!");
1371 case CCValAssign::Full: break;
1372 case CCValAssign::SExt:
1373 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1375 case CCValAssign::ZExt:
1376 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1378 case CCValAssign::AExt:
1379 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1381 case CCValAssign::BCvt:
1382 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1386 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1387 if (VA.needsCustom()) {
1388 if (VA.getLocVT() == MVT::v2f64) {
1389 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1390 DAG.getConstant(0, MVT::i32));
1391 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1392 DAG.getConstant(1, MVT::i32));
1394 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1395 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1397 VA = ArgLocs[++i]; // skip ahead to next loc
1398 if (VA.isRegLoc()) {
1399 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1400 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1402 assert(VA.isMemLoc());
1404 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1405 dl, DAG, VA, Flags));
1408 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1409 StackPtr, MemOpChains, Flags);
1411 } else if (VA.isRegLoc()) {
1412 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1413 } else if (isByVal) {
1414 assert(VA.isMemLoc());
1415 unsigned offset = 0;
1417 // True if this byval aggregate will be split between registers
1419 if (CCInfo.isFirstByValRegValid()) {
1420 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1422 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1423 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1424 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1425 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1426 MachinePointerInfo(),
1427 false, false, false, 0);
1428 MemOpChains.push_back(Load.getValue(1));
1429 RegsToPass.push_back(std::make_pair(j, Load));
1431 offset = ARM::R4 - CCInfo.getFirstByValReg();
1432 CCInfo.clearFirstByValReg();
1435 if (Flags.getByValSize() - 4*offset > 0) {
1436 unsigned LocMemOffset = VA.getLocMemOffset();
1437 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1438 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1440 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1441 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1442 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1444 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1446 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1447 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1448 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1449 Ops, array_lengthof(Ops)));
1451 } else if (!IsSibCall) {
1452 assert(VA.isMemLoc());
1454 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1455 dl, DAG, VA, Flags));
1459 if (!MemOpChains.empty())
1460 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1461 &MemOpChains[0], MemOpChains.size());
1463 // Build a sequence of copy-to-reg nodes chained together with token chain
1464 // and flag operands which copy the outgoing args into the appropriate regs.
1466 // Tail call byval lowering might overwrite argument registers so in case of
1467 // tail call optimization the copies to registers are lowered later.
1469 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1470 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1471 RegsToPass[i].second, InFlag);
1472 InFlag = Chain.getValue(1);
1475 // For tail calls lower the arguments to the 'real' stack slot.
1477 // Force all the incoming stack arguments to be loaded from the stack
1478 // before any new outgoing arguments are stored to the stack, because the
1479 // outgoing stack slots may alias the incoming argument stack slots, and
1480 // the alias isn't otherwise explicit. This is slightly more conservative
1481 // than necessary, because it means that each store effectively depends
1482 // on every argument instead of just those arguments it would clobber.
1484 // Do not flag preceding copytoreg stuff together with the following stuff.
1486 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1487 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1488 RegsToPass[i].second, InFlag);
1489 InFlag = Chain.getValue(1);
1494 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1495 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1496 // node so that legalize doesn't hack it.
1497 bool isDirect = false;
1498 bool isARMFunc = false;
1499 bool isLocalARMFunc = false;
1500 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1502 if (EnableARMLongCalls) {
1503 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1504 && "long-calls with non-static relocation model!");
1505 // Handle a global address or an external symbol. If it's not one of
1506 // those, the target's already in a register, so we don't need to do
1508 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1509 const GlobalValue *GV = G->getGlobal();
1510 // Create a constant pool entry for the callee address
1511 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1512 ARMConstantPoolValue *CPV =
1513 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1515 // Get the address of the callee into a register
1516 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1517 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1518 Callee = DAG.getLoad(getPointerTy(), dl,
1519 DAG.getEntryNode(), CPAddr,
1520 MachinePointerInfo::getConstantPool(),
1521 false, false, false, 0);
1522 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1523 const char *Sym = S->getSymbol();
1525 // Create a constant pool entry for the callee address
1526 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1527 ARMConstantPoolValue *CPV =
1528 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1529 ARMPCLabelIndex, 0);
1530 // Get the address of the callee into a register
1531 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1532 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1533 Callee = DAG.getLoad(getPointerTy(), dl,
1534 DAG.getEntryNode(), CPAddr,
1535 MachinePointerInfo::getConstantPool(),
1536 false, false, false, 0);
1538 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1539 const GlobalValue *GV = G->getGlobal();
1541 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1542 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1543 getTargetMachine().getRelocationModel() != Reloc::Static;
1544 isARMFunc = !Subtarget->isThumb() || isStub;
1545 // ARM call to a local ARM function is predicable.
1546 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1547 // tBX takes a register source operand.
1548 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1549 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1550 ARMConstantPoolValue *CPV =
1551 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1552 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1553 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1554 Callee = DAG.getLoad(getPointerTy(), dl,
1555 DAG.getEntryNode(), CPAddr,
1556 MachinePointerInfo::getConstantPool(),
1557 false, false, false, 0);
1558 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1559 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1560 getPointerTy(), Callee, PICLabel);
1562 // On ELF targets for PIC code, direct calls should go through the PLT
1563 unsigned OpFlags = 0;
1564 if (Subtarget->isTargetELF() &&
1565 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1566 OpFlags = ARMII::MO_PLT;
1567 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1569 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1571 bool isStub = Subtarget->isTargetDarwin() &&
1572 getTargetMachine().getRelocationModel() != Reloc::Static;
1573 isARMFunc = !Subtarget->isThumb() || isStub;
1574 // tBX takes a register source operand.
1575 const char *Sym = S->getSymbol();
1576 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1577 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1578 ARMConstantPoolValue *CPV =
1579 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1580 ARMPCLabelIndex, 4);
1581 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1582 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1583 Callee = DAG.getLoad(getPointerTy(), dl,
1584 DAG.getEntryNode(), CPAddr,
1585 MachinePointerInfo::getConstantPool(),
1586 false, false, false, 0);
1587 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1588 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1589 getPointerTy(), Callee, PICLabel);
1591 unsigned OpFlags = 0;
1592 // On ELF targets for PIC code, direct calls should go through the PLT
1593 if (Subtarget->isTargetELF() &&
1594 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1595 OpFlags = ARMII::MO_PLT;
1596 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1600 // FIXME: handle tail calls differently.
1602 bool HasMinSizeAttr = MF.getFunction()->getFnAttributes().
1603 hasAttribute(Attributes::MinSize);
1604 if (Subtarget->isThumb()) {
1605 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1606 CallOpc = ARMISD::CALL_NOLINK;
1608 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1610 if (!isDirect && !Subtarget->hasV5TOps())
1611 CallOpc = ARMISD::CALL_NOLINK;
1612 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1613 // Emit regular call when code size is the priority
1615 // "mov lr, pc; b _foo" to avoid confusing the RSP
1616 CallOpc = ARMISD::CALL_NOLINK;
1618 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1621 std::vector<SDValue> Ops;
1622 Ops.push_back(Chain);
1623 Ops.push_back(Callee);
1625 // Add argument registers to the end of the list so that they are known live
1627 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1628 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1629 RegsToPass[i].second.getValueType()));
1631 // Add a register mask operand representing the call-preserved registers.
1632 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1633 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1634 assert(Mask && "Missing call preserved mask for calling convention");
1635 Ops.push_back(DAG.getRegisterMask(Mask));
1637 if (InFlag.getNode())
1638 Ops.push_back(InFlag);
1640 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1642 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1644 // Returns a chain and a flag for retval copy to use.
1645 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1646 InFlag = Chain.getValue(1);
1648 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1649 DAG.getIntPtrConstant(0, true), InFlag);
1651 InFlag = Chain.getValue(1);
1653 // Handle result values, copying them out of physregs into vregs that we
1655 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1659 /// HandleByVal - Every parameter *after* a byval parameter is passed
1660 /// on the stack. Remember the next parameter register to allocate,
1661 /// and then confiscate the rest of the parameter registers to insure
1664 ARMTargetLowering::HandleByVal(
1665 CCState *State, unsigned &size, unsigned Align) const {
1666 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1667 assert((State->getCallOrPrologue() == Prologue ||
1668 State->getCallOrPrologue() == Call) &&
1669 "unhandled ParmContext");
1670 if ((!State->isFirstByValRegValid()) &&
1671 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1672 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1673 unsigned AlignInRegs = Align / 4;
1674 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1675 for (unsigned i = 0; i < Waste; ++i)
1676 reg = State->AllocateReg(GPRArgRegs, 4);
1679 State->setFirstByValReg(reg);
1680 // At a call site, a byval parameter that is split between
1681 // registers and memory needs its size truncated here. In a
1682 // function prologue, such byval parameters are reassembled in
1683 // memory, and are not truncated.
1684 if (State->getCallOrPrologue() == Call) {
1685 unsigned excess = 4 * (ARM::R4 - reg);
1686 assert(size >= excess && "expected larger existing stack allocation");
1691 // Confiscate any remaining parameter registers to preclude their
1692 // assignment to subsequent parameters.
1693 while (State->AllocateReg(GPRArgRegs, 4))
1697 /// MatchingStackOffset - Return true if the given stack call argument is
1698 /// already available in the same position (relatively) of the caller's
1699 /// incoming argument stack.
1701 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1702 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1703 const TargetInstrInfo *TII) {
1704 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1706 if (Arg.getOpcode() == ISD::CopyFromReg) {
1707 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1708 if (!TargetRegisterInfo::isVirtualRegister(VR))
1710 MachineInstr *Def = MRI->getVRegDef(VR);
1713 if (!Flags.isByVal()) {
1714 if (!TII->isLoadFromStackSlot(Def, FI))
1719 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1720 if (Flags.isByVal())
1721 // ByVal argument is passed in as a pointer but it's now being
1722 // dereferenced. e.g.
1723 // define @foo(%struct.X* %A) {
1724 // tail call @bar(%struct.X* byval %A)
1727 SDValue Ptr = Ld->getBasePtr();
1728 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1731 FI = FINode->getIndex();
1735 assert(FI != INT_MAX);
1736 if (!MFI->isFixedObjectIndex(FI))
1738 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1741 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1742 /// for tail call optimization. Targets which want to do tail call
1743 /// optimization should implement this function.
1745 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1746 CallingConv::ID CalleeCC,
1748 bool isCalleeStructRet,
1749 bool isCallerStructRet,
1750 const SmallVectorImpl<ISD::OutputArg> &Outs,
1751 const SmallVectorImpl<SDValue> &OutVals,
1752 const SmallVectorImpl<ISD::InputArg> &Ins,
1753 SelectionDAG& DAG) const {
1754 const Function *CallerF = DAG.getMachineFunction().getFunction();
1755 CallingConv::ID CallerCC = CallerF->getCallingConv();
1756 bool CCMatch = CallerCC == CalleeCC;
1758 // Look for obvious safe cases to perform tail call optimization that do not
1759 // require ABI changes. This is what gcc calls sibcall.
1761 // Do not sibcall optimize vararg calls unless the call site is not passing
1763 if (isVarArg && !Outs.empty())
1766 // Also avoid sibcall optimization if either caller or callee uses struct
1767 // return semantics.
1768 if (isCalleeStructRet || isCallerStructRet)
1771 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1772 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1773 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1774 // support in the assembler and linker to be used. This would need to be
1775 // fixed to fully support tail calls in Thumb1.
1777 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1778 // LR. This means if we need to reload LR, it takes an extra instructions,
1779 // which outweighs the value of the tail call; but here we don't know yet
1780 // whether LR is going to be used. Probably the right approach is to
1781 // generate the tail call here and turn it back into CALL/RET in
1782 // emitEpilogue if LR is used.
1784 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1785 // but we need to make sure there are enough registers; the only valid
1786 // registers are the 4 used for parameters. We don't currently do this
1788 if (Subtarget->isThumb1Only())
1791 // If the calling conventions do not match, then we'd better make sure the
1792 // results are returned in the same way as what the caller expects.
1794 SmallVector<CCValAssign, 16> RVLocs1;
1795 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1796 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1797 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1799 SmallVector<CCValAssign, 16> RVLocs2;
1800 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1801 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1802 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1804 if (RVLocs1.size() != RVLocs2.size())
1806 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1807 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1809 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1811 if (RVLocs1[i].isRegLoc()) {
1812 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1815 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1821 // If Caller's vararg or byval argument has been split between registers and
1822 // stack, do not perform tail call, since part of the argument is in caller's
1824 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1825 getInfo<ARMFunctionInfo>();
1826 if (AFI_Caller->getVarArgsRegSaveSize())
1829 // If the callee takes no arguments then go on to check the results of the
1831 if (!Outs.empty()) {
1832 // Check if stack adjustment is needed. For now, do not do this if any
1833 // argument is passed on the stack.
1834 SmallVector<CCValAssign, 16> ArgLocs;
1835 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1836 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1837 CCInfo.AnalyzeCallOperands(Outs,
1838 CCAssignFnForNode(CalleeCC, false, isVarArg));
1839 if (CCInfo.getNextStackOffset()) {
1840 MachineFunction &MF = DAG.getMachineFunction();
1842 // Check if the arguments are already laid out in the right way as
1843 // the caller's fixed stack objects.
1844 MachineFrameInfo *MFI = MF.getFrameInfo();
1845 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1846 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1847 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1849 ++i, ++realArgIdx) {
1850 CCValAssign &VA = ArgLocs[i];
1851 EVT RegVT = VA.getLocVT();
1852 SDValue Arg = OutVals[realArgIdx];
1853 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1854 if (VA.getLocInfo() == CCValAssign::Indirect)
1856 if (VA.needsCustom()) {
1857 // f64 and vector types are split into multiple registers or
1858 // register/stack-slot combinations. The types will not match
1859 // the registers; give up on memory f64 refs until we figure
1860 // out what to do about this.
1863 if (!ArgLocs[++i].isRegLoc())
1865 if (RegVT == MVT::v2f64) {
1866 if (!ArgLocs[++i].isRegLoc())
1868 if (!ArgLocs[++i].isRegLoc())
1871 } else if (!VA.isRegLoc()) {
1872 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1884 ARMTargetLowering::LowerReturn(SDValue Chain,
1885 CallingConv::ID CallConv, bool isVarArg,
1886 const SmallVectorImpl<ISD::OutputArg> &Outs,
1887 const SmallVectorImpl<SDValue> &OutVals,
1888 DebugLoc dl, SelectionDAG &DAG) const {
1890 // CCValAssign - represent the assignment of the return value to a location.
1891 SmallVector<CCValAssign, 16> RVLocs;
1893 // CCState - Info about the registers and stack slots.
1894 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1895 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1897 // Analyze outgoing return values.
1898 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1901 // If this is the first return lowered for this function, add
1902 // the regs to the liveout set for the function.
1903 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1904 for (unsigned i = 0; i != RVLocs.size(); ++i)
1905 if (RVLocs[i].isRegLoc())
1906 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1911 // Copy the result values into the output registers.
1912 for (unsigned i = 0, realRVLocIdx = 0;
1914 ++i, ++realRVLocIdx) {
1915 CCValAssign &VA = RVLocs[i];
1916 assert(VA.isRegLoc() && "Can only return in registers!");
1918 SDValue Arg = OutVals[realRVLocIdx];
1920 switch (VA.getLocInfo()) {
1921 default: llvm_unreachable("Unknown loc info!");
1922 case CCValAssign::Full: break;
1923 case CCValAssign::BCvt:
1924 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1928 if (VA.needsCustom()) {
1929 if (VA.getLocVT() == MVT::v2f64) {
1930 // Extract the first half and return it in two registers.
1931 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1932 DAG.getConstant(0, MVT::i32));
1933 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1934 DAG.getVTList(MVT::i32, MVT::i32), Half);
1936 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1937 Flag = Chain.getValue(1);
1938 VA = RVLocs[++i]; // skip ahead to next loc
1939 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1940 HalfGPRs.getValue(1), Flag);
1941 Flag = Chain.getValue(1);
1942 VA = RVLocs[++i]; // skip ahead to next loc
1944 // Extract the 2nd half and fall through to handle it as an f64 value.
1945 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1946 DAG.getConstant(1, MVT::i32));
1948 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1950 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1951 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1952 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1953 Flag = Chain.getValue(1);
1954 VA = RVLocs[++i]; // skip ahead to next loc
1955 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1958 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1960 // Guarantee that all emitted copies are
1961 // stuck together, avoiding something bad.
1962 Flag = Chain.getValue(1);
1967 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1969 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1974 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1975 if (N->getNumValues() != 1)
1977 if (!N->hasNUsesOfValue(1, 0))
1980 SDValue TCChain = Chain;
1981 SDNode *Copy = *N->use_begin();
1982 if (Copy->getOpcode() == ISD::CopyToReg) {
1983 // If the copy has a glue operand, we conservatively assume it isn't safe to
1984 // perform a tail call.
1985 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1987 TCChain = Copy->getOperand(0);
1988 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1989 SDNode *VMov = Copy;
1990 // f64 returned in a pair of GPRs.
1991 SmallPtrSet<SDNode*, 2> Copies;
1992 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1994 if (UI->getOpcode() != ISD::CopyToReg)
1998 if (Copies.size() > 2)
2001 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2003 SDValue UseChain = UI->getOperand(0);
2004 if (Copies.count(UseChain.getNode()))
2011 } else if (Copy->getOpcode() == ISD::BITCAST) {
2012 // f32 returned in a single GPR.
2013 if (!Copy->hasOneUse())
2015 Copy = *Copy->use_begin();
2016 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2018 Chain = Copy->getOperand(0);
2023 bool HasRet = false;
2024 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2026 if (UI->getOpcode() != ARMISD::RET_FLAG)
2038 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2039 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2042 if (!CI->isTailCall())
2045 return !Subtarget->isThumb1Only();
2048 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2049 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2050 // one of the above mentioned nodes. It has to be wrapped because otherwise
2051 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2052 // be used to form addressing mode. These wrapped nodes will be selected
2054 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2055 EVT PtrVT = Op.getValueType();
2056 // FIXME there is no actual debug info here
2057 DebugLoc dl = Op.getDebugLoc();
2058 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2060 if (CP->isMachineConstantPoolEntry())
2061 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2062 CP->getAlignment());
2064 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2065 CP->getAlignment());
2066 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2069 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2070 return MachineJumpTableInfo::EK_Inline;
2073 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2074 SelectionDAG &DAG) const {
2075 MachineFunction &MF = DAG.getMachineFunction();
2076 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2077 unsigned ARMPCLabelIndex = 0;
2078 DebugLoc DL = Op.getDebugLoc();
2079 EVT PtrVT = getPointerTy();
2080 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2081 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2083 if (RelocM == Reloc::Static) {
2084 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2086 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2087 ARMPCLabelIndex = AFI->createPICLabelUId();
2088 ARMConstantPoolValue *CPV =
2089 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2090 ARMCP::CPBlockAddress, PCAdj);
2091 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2093 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2094 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2095 MachinePointerInfo::getConstantPool(),
2096 false, false, false, 0);
2097 if (RelocM == Reloc::Static)
2099 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2100 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2103 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2105 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2106 SelectionDAG &DAG) const {
2107 DebugLoc dl = GA->getDebugLoc();
2108 EVT PtrVT = getPointerTy();
2109 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2110 MachineFunction &MF = DAG.getMachineFunction();
2111 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2112 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2113 ARMConstantPoolValue *CPV =
2114 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2115 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2116 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2117 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2118 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2119 MachinePointerInfo::getConstantPool(),
2120 false, false, false, 0);
2121 SDValue Chain = Argument.getValue(1);
2123 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2124 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2126 // call __tls_get_addr.
2129 Entry.Node = Argument;
2130 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2131 Args.push_back(Entry);
2132 // FIXME: is there useful debug info available here?
2133 TargetLowering::CallLoweringInfo CLI(Chain,
2134 (Type *) Type::getInt32Ty(*DAG.getContext()),
2135 false, false, false, false,
2136 0, CallingConv::C, /*isTailCall=*/false,
2137 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2138 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2139 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2140 return CallResult.first;
2143 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2144 // "local exec" model.
2146 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2148 TLSModel::Model model) const {
2149 const GlobalValue *GV = GA->getGlobal();
2150 DebugLoc dl = GA->getDebugLoc();
2152 SDValue Chain = DAG.getEntryNode();
2153 EVT PtrVT = getPointerTy();
2154 // Get the Thread Pointer
2155 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2157 if (model == TLSModel::InitialExec) {
2158 MachineFunction &MF = DAG.getMachineFunction();
2159 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2160 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2161 // Initial exec model.
2162 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2163 ARMConstantPoolValue *CPV =
2164 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2165 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2167 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2168 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2169 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2170 MachinePointerInfo::getConstantPool(),
2171 false, false, false, 0);
2172 Chain = Offset.getValue(1);
2174 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2175 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2177 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2178 MachinePointerInfo::getConstantPool(),
2179 false, false, false, 0);
2182 assert(model == TLSModel::LocalExec);
2183 ARMConstantPoolValue *CPV =
2184 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2185 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2186 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2187 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2188 MachinePointerInfo::getConstantPool(),
2189 false, false, false, 0);
2192 // The address of the thread local variable is the add of the thread
2193 // pointer with the offset of the variable.
2194 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2198 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2199 // TODO: implement the "local dynamic" model
2200 assert(Subtarget->isTargetELF() &&
2201 "TLS not implemented for non-ELF targets");
2202 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2204 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2207 case TLSModel::GeneralDynamic:
2208 case TLSModel::LocalDynamic:
2209 return LowerToTLSGeneralDynamicModel(GA, DAG);
2210 case TLSModel::InitialExec:
2211 case TLSModel::LocalExec:
2212 return LowerToTLSExecModels(GA, DAG, model);
2214 llvm_unreachable("bogus TLS model");
2217 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2218 SelectionDAG &DAG) const {
2219 EVT PtrVT = getPointerTy();
2220 DebugLoc dl = Op.getDebugLoc();
2221 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2222 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2223 if (RelocM == Reloc::PIC_) {
2224 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2225 ARMConstantPoolValue *CPV =
2226 ARMConstantPoolConstant::Create(GV,
2227 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2228 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2229 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2230 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2232 MachinePointerInfo::getConstantPool(),
2233 false, false, false, 0);
2234 SDValue Chain = Result.getValue(1);
2235 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2236 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2238 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2239 MachinePointerInfo::getGOT(),
2240 false, false, false, 0);
2244 // If we have T2 ops, we can materialize the address directly via movt/movw
2245 // pair. This is always cheaper.
2246 if (Subtarget->useMovt()) {
2248 // FIXME: Once remat is capable of dealing with instructions with register
2249 // operands, expand this into two nodes.
2250 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2251 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2253 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2254 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2255 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2256 MachinePointerInfo::getConstantPool(),
2257 false, false, false, 0);
2261 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2262 SelectionDAG &DAG) const {
2263 EVT PtrVT = getPointerTy();
2264 DebugLoc dl = Op.getDebugLoc();
2265 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2266 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2267 MachineFunction &MF = DAG.getMachineFunction();
2268 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2270 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2271 // update ARMFastISel::ARMMaterializeGV.
2272 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2274 // FIXME: Once remat is capable of dealing with instructions with register
2275 // operands, expand this into two nodes.
2276 if (RelocM == Reloc::Static)
2277 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2278 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2280 unsigned Wrapper = (RelocM == Reloc::PIC_)
2281 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2282 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2283 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2284 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2285 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2286 MachinePointerInfo::getGOT(),
2287 false, false, false, 0);
2291 unsigned ARMPCLabelIndex = 0;
2293 if (RelocM == Reloc::Static) {
2294 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2296 ARMPCLabelIndex = AFI->createPICLabelUId();
2297 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2298 ARMConstantPoolValue *CPV =
2299 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2301 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2303 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2305 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2306 MachinePointerInfo::getConstantPool(),
2307 false, false, false, 0);
2308 SDValue Chain = Result.getValue(1);
2310 if (RelocM == Reloc::PIC_) {
2311 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2312 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2315 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2316 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2317 false, false, false, 0);
2322 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2323 SelectionDAG &DAG) const {
2324 assert(Subtarget->isTargetELF() &&
2325 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2326 MachineFunction &MF = DAG.getMachineFunction();
2327 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2328 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2329 EVT PtrVT = getPointerTy();
2330 DebugLoc dl = Op.getDebugLoc();
2331 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2332 ARMConstantPoolValue *CPV =
2333 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2334 ARMPCLabelIndex, PCAdj);
2335 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2336 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2337 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2338 MachinePointerInfo::getConstantPool(),
2339 false, false, false, 0);
2340 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2341 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2345 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2346 DebugLoc dl = Op.getDebugLoc();
2347 SDValue Val = DAG.getConstant(0, MVT::i32);
2348 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2349 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2350 Op.getOperand(1), Val);
2354 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2355 DebugLoc dl = Op.getDebugLoc();
2356 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2357 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2361 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2362 const ARMSubtarget *Subtarget) const {
2363 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2364 DebugLoc dl = Op.getDebugLoc();
2366 default: return SDValue(); // Don't custom lower most intrinsics.
2367 case Intrinsic::arm_thread_pointer: {
2368 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2369 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2371 case Intrinsic::eh_sjlj_lsda: {
2372 MachineFunction &MF = DAG.getMachineFunction();
2373 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2374 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2375 EVT PtrVT = getPointerTy();
2376 DebugLoc dl = Op.getDebugLoc();
2377 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2379 unsigned PCAdj = (RelocM != Reloc::PIC_)
2380 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2381 ARMConstantPoolValue *CPV =
2382 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2383 ARMCP::CPLSDA, PCAdj);
2384 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2385 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2387 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2388 MachinePointerInfo::getConstantPool(),
2389 false, false, false, 0);
2391 if (RelocM == Reloc::PIC_) {
2392 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2393 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2397 case Intrinsic::arm_neon_vmulls:
2398 case Intrinsic::arm_neon_vmullu: {
2399 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2400 ? ARMISD::VMULLs : ARMISD::VMULLu;
2401 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2402 Op.getOperand(1), Op.getOperand(2));
2407 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2408 const ARMSubtarget *Subtarget) {
2409 DebugLoc dl = Op.getDebugLoc();
2410 if (!Subtarget->hasDataBarrier()) {
2411 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2412 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2414 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2415 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2416 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2417 DAG.getConstant(0, MVT::i32));
2420 SDValue Op5 = Op.getOperand(5);
2421 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2422 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2423 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2424 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2426 ARM_MB::MemBOpt DMBOpt;
2427 if (isDeviceBarrier)
2428 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2430 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2431 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2432 DAG.getConstant(DMBOpt, MVT::i32));
2436 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2437 const ARMSubtarget *Subtarget) {
2438 // FIXME: handle "fence singlethread" more efficiently.
2439 DebugLoc dl = Op.getDebugLoc();
2440 if (!Subtarget->hasDataBarrier()) {
2441 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2442 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2444 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2445 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2446 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2447 DAG.getConstant(0, MVT::i32));
2450 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2451 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2454 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2455 const ARMSubtarget *Subtarget) {
2456 // ARM pre v5TE and Thumb1 does not have preload instructions.
2457 if (!(Subtarget->isThumb2() ||
2458 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2459 // Just preserve the chain.
2460 return Op.getOperand(0);
2462 DebugLoc dl = Op.getDebugLoc();
2463 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2465 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2466 // ARMv7 with MP extension has PLDW.
2467 return Op.getOperand(0);
2469 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2470 if (Subtarget->isThumb()) {
2472 isRead = ~isRead & 1;
2473 isData = ~isData & 1;
2476 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2477 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2478 DAG.getConstant(isData, MVT::i32));
2481 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2482 MachineFunction &MF = DAG.getMachineFunction();
2483 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2485 // vastart just stores the address of the VarArgsFrameIndex slot into the
2486 // memory location argument.
2487 DebugLoc dl = Op.getDebugLoc();
2488 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2489 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2490 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2491 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2492 MachinePointerInfo(SV), false, false, 0);
2496 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2497 SDValue &Root, SelectionDAG &DAG,
2498 DebugLoc dl) const {
2499 MachineFunction &MF = DAG.getMachineFunction();
2500 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2502 const TargetRegisterClass *RC;
2503 if (AFI->isThumb1OnlyFunction())
2504 RC = &ARM::tGPRRegClass;
2506 RC = &ARM::GPRRegClass;
2508 // Transform the arguments stored in physical registers into virtual ones.
2509 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2510 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2513 if (NextVA.isMemLoc()) {
2514 MachineFrameInfo *MFI = MF.getFrameInfo();
2515 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2517 // Create load node to retrieve arguments from the stack.
2518 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2519 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2520 MachinePointerInfo::getFixedStack(FI),
2521 false, false, false, 0);
2523 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2524 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2527 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2531 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2532 unsigned &VARegSize, unsigned &VARegSaveSize)
2535 if (CCInfo.isFirstByValRegValid())
2536 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2538 unsigned int firstUnalloced;
2539 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2540 sizeof(GPRArgRegs) /
2541 sizeof(GPRArgRegs[0]));
2542 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2545 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2546 VARegSize = NumGPRs * 4;
2547 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2550 // The remaining GPRs hold either the beginning of variable-argument
2551 // data, or the beginning of an aggregate passed by value (usuall
2552 // byval). Either way, we allocate stack slots adjacent to the data
2553 // provided by our caller, and store the unallocated registers there.
2554 // If this is a variadic function, the va_list pointer will begin with
2555 // these values; otherwise, this reassembles a (byval) structure that
2556 // was split between registers and memory.
2558 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2559 DebugLoc dl, SDValue &Chain,
2560 const Value *OrigArg,
2561 unsigned OffsetFromOrigArg,
2563 bool ForceMutable) const {
2564 MachineFunction &MF = DAG.getMachineFunction();
2565 MachineFrameInfo *MFI = MF.getFrameInfo();
2566 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2567 unsigned firstRegToSaveIndex;
2568 if (CCInfo.isFirstByValRegValid())
2569 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2571 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2572 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2575 unsigned VARegSize, VARegSaveSize;
2576 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2577 if (VARegSaveSize) {
2578 // If this function is vararg, store any remaining integer argument regs
2579 // to their spots on the stack so that they may be loaded by deferencing
2580 // the result of va_next.
2581 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2582 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2583 ArgOffset + VARegSaveSize
2586 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2589 SmallVector<SDValue, 4> MemOps;
2590 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2591 const TargetRegisterClass *RC;
2592 if (AFI->isThumb1OnlyFunction())
2593 RC = &ARM::tGPRRegClass;
2595 RC = &ARM::GPRRegClass;
2597 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2598 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2600 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2601 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2603 MemOps.push_back(Store);
2604 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2605 DAG.getConstant(4, getPointerTy()));
2607 if (!MemOps.empty())
2608 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2609 &MemOps[0], MemOps.size());
2611 // This will point to the next argument passed via stack.
2612 AFI->setVarArgsFrameIndex(
2613 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2617 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2618 CallingConv::ID CallConv, bool isVarArg,
2619 const SmallVectorImpl<ISD::InputArg>
2621 DebugLoc dl, SelectionDAG &DAG,
2622 SmallVectorImpl<SDValue> &InVals)
2624 MachineFunction &MF = DAG.getMachineFunction();
2625 MachineFrameInfo *MFI = MF.getFrameInfo();
2627 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2629 // Assign locations to all of the incoming arguments.
2630 SmallVector<CCValAssign, 16> ArgLocs;
2631 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2632 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2633 CCInfo.AnalyzeFormalArguments(Ins,
2634 CCAssignFnForNode(CallConv, /* Return*/ false,
2637 SmallVector<SDValue, 16> ArgValues;
2638 int lastInsIndex = -1;
2640 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2641 unsigned CurArgIdx = 0;
2642 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2643 CCValAssign &VA = ArgLocs[i];
2644 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2645 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2646 // Arguments stored in registers.
2647 if (VA.isRegLoc()) {
2648 EVT RegVT = VA.getLocVT();
2650 if (VA.needsCustom()) {
2651 // f64 and vector types are split up into multiple registers or
2652 // combinations of registers and stack slots.
2653 if (VA.getLocVT() == MVT::v2f64) {
2654 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2656 VA = ArgLocs[++i]; // skip ahead to next loc
2658 if (VA.isMemLoc()) {
2659 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2660 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2661 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2662 MachinePointerInfo::getFixedStack(FI),
2663 false, false, false, 0);
2665 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2668 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2669 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2670 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2671 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2672 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2674 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2677 const TargetRegisterClass *RC;
2679 if (RegVT == MVT::f32)
2680 RC = &ARM::SPRRegClass;
2681 else if (RegVT == MVT::f64)
2682 RC = &ARM::DPRRegClass;
2683 else if (RegVT == MVT::v2f64)
2684 RC = &ARM::QPRRegClass;
2685 else if (RegVT == MVT::i32)
2686 RC = AFI->isThumb1OnlyFunction() ?
2687 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2688 (const TargetRegisterClass*)&ARM::GPRRegClass;
2690 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2692 // Transform the arguments in physical registers into virtual ones.
2693 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2694 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2697 // If this is an 8 or 16-bit value, it is really passed promoted
2698 // to 32 bits. Insert an assert[sz]ext to capture this, then
2699 // truncate to the right size.
2700 switch (VA.getLocInfo()) {
2701 default: llvm_unreachable("Unknown loc info!");
2702 case CCValAssign::Full: break;
2703 case CCValAssign::BCvt:
2704 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2706 case CCValAssign::SExt:
2707 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2708 DAG.getValueType(VA.getValVT()));
2709 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2711 case CCValAssign::ZExt:
2712 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2713 DAG.getValueType(VA.getValVT()));
2714 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2718 InVals.push_back(ArgValue);
2720 } else { // VA.isRegLoc()
2723 assert(VA.isMemLoc());
2724 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2726 int index = ArgLocs[i].getValNo();
2728 // Some Ins[] entries become multiple ArgLoc[] entries.
2729 // Process them only once.
2730 if (index != lastInsIndex)
2732 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2733 // FIXME: For now, all byval parameter objects are marked mutable.
2734 // This can be changed with more analysis.
2735 // In case of tail call optimization mark all arguments mutable.
2736 // Since they could be overwritten by lowering of arguments in case of
2738 if (Flags.isByVal()) {
2739 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2740 if (!AFI->getVarArgsFrameIndex()) {
2741 VarArgStyleRegisters(CCInfo, DAG,
2742 dl, Chain, CurOrigArg,
2743 Ins[VA.getValNo()].PartOffset,
2744 VA.getLocMemOffset(),
2745 true /*force mutable frames*/);
2746 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2747 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2749 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2750 VA.getLocMemOffset(), false);
2751 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2754 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2755 VA.getLocMemOffset(), true);
2757 // Create load nodes to retrieve arguments from the stack.
2758 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2759 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2760 MachinePointerInfo::getFixedStack(FI),
2761 false, false, false, 0));
2763 lastInsIndex = index;
2770 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2771 CCInfo.getNextStackOffset());
2776 /// isFloatingPointZero - Return true if this is +0.0.
2777 static bool isFloatingPointZero(SDValue Op) {
2778 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2779 return CFP->getValueAPF().isPosZero();
2780 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2781 // Maybe this has already been legalized into the constant pool?
2782 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2783 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2784 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2785 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2786 return CFP->getValueAPF().isPosZero();
2792 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2793 /// the given operands.
2795 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2796 SDValue &ARMcc, SelectionDAG &DAG,
2797 DebugLoc dl) const {
2798 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2799 unsigned C = RHSC->getZExtValue();
2800 if (!isLegalICmpImmediate(C)) {
2801 // Constant does not fit, try adjusting it by one?
2806 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2807 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2808 RHS = DAG.getConstant(C-1, MVT::i32);
2813 if (C != 0 && isLegalICmpImmediate(C-1)) {
2814 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2815 RHS = DAG.getConstant(C-1, MVT::i32);
2820 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2821 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2822 RHS = DAG.getConstant(C+1, MVT::i32);
2827 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2828 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2829 RHS = DAG.getConstant(C+1, MVT::i32);
2836 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2837 ARMISD::NodeType CompareType;
2840 CompareType = ARMISD::CMP;
2845 CompareType = ARMISD::CMPZ;
2848 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2849 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2852 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2854 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2855 DebugLoc dl) const {
2857 if (!isFloatingPointZero(RHS))
2858 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2860 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2861 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2864 /// duplicateCmp - Glue values can have only one use, so this function
2865 /// duplicates a comparison node.
2867 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2868 unsigned Opc = Cmp.getOpcode();
2869 DebugLoc DL = Cmp.getDebugLoc();
2870 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2871 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2873 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2874 Cmp = Cmp.getOperand(0);
2875 Opc = Cmp.getOpcode();
2876 if (Opc == ARMISD::CMPFP)
2877 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2879 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2880 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2882 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2885 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2886 SDValue Cond = Op.getOperand(0);
2887 SDValue SelectTrue = Op.getOperand(1);
2888 SDValue SelectFalse = Op.getOperand(2);
2889 DebugLoc dl = Op.getDebugLoc();
2893 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2894 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2896 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2897 const ConstantSDNode *CMOVTrue =
2898 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2899 const ConstantSDNode *CMOVFalse =
2900 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2902 if (CMOVTrue && CMOVFalse) {
2903 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2904 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2908 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2910 False = SelectFalse;
2911 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2916 if (True.getNode() && False.getNode()) {
2917 EVT VT = Op.getValueType();
2918 SDValue ARMcc = Cond.getOperand(2);
2919 SDValue CCR = Cond.getOperand(3);
2920 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2921 assert(True.getValueType() == VT);
2922 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2927 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2928 // undefined bits before doing a full-word comparison with zero.
2929 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2930 DAG.getConstant(1, Cond.getValueType()));
2932 return DAG.getSelectCC(dl, Cond,
2933 DAG.getConstant(0, Cond.getValueType()),
2934 SelectTrue, SelectFalse, ISD::SETNE);
2937 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2938 EVT VT = Op.getValueType();
2939 SDValue LHS = Op.getOperand(0);
2940 SDValue RHS = Op.getOperand(1);
2941 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2942 SDValue TrueVal = Op.getOperand(2);
2943 SDValue FalseVal = Op.getOperand(3);
2944 DebugLoc dl = Op.getDebugLoc();
2946 if (LHS.getValueType() == MVT::i32) {
2948 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2949 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2950 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2953 ARMCC::CondCodes CondCode, CondCode2;
2954 FPCCToARMCC(CC, CondCode, CondCode2);
2956 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2957 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2958 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2959 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2961 if (CondCode2 != ARMCC::AL) {
2962 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2963 // FIXME: Needs another CMP because flag can have but one use.
2964 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2965 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2966 Result, TrueVal, ARMcc2, CCR, Cmp2);
2971 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2972 /// to morph to an integer compare sequence.
2973 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2974 const ARMSubtarget *Subtarget) {
2975 SDNode *N = Op.getNode();
2976 if (!N->hasOneUse())
2977 // Otherwise it requires moving the value from fp to integer registers.
2979 if (!N->getNumValues())
2981 EVT VT = Op.getValueType();
2982 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2983 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2984 // vmrs are very slow, e.g. cortex-a8.
2987 if (isFloatingPointZero(Op)) {
2991 return ISD::isNormalLoad(N);
2994 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2995 if (isFloatingPointZero(Op))
2996 return DAG.getConstant(0, MVT::i32);
2998 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2999 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3000 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3001 Ld->isVolatile(), Ld->isNonTemporal(),
3002 Ld->isInvariant(), Ld->getAlignment());
3004 llvm_unreachable("Unknown VFP cmp argument!");
3007 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3008 SDValue &RetVal1, SDValue &RetVal2) {
3009 if (isFloatingPointZero(Op)) {
3010 RetVal1 = DAG.getConstant(0, MVT::i32);
3011 RetVal2 = DAG.getConstant(0, MVT::i32);
3015 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3016 SDValue Ptr = Ld->getBasePtr();
3017 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3018 Ld->getChain(), Ptr,
3019 Ld->getPointerInfo(),
3020 Ld->isVolatile(), Ld->isNonTemporal(),
3021 Ld->isInvariant(), Ld->getAlignment());
3023 EVT PtrType = Ptr.getValueType();
3024 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3025 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3026 PtrType, Ptr, DAG.getConstant(4, PtrType));
3027 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3028 Ld->getChain(), NewPtr,
3029 Ld->getPointerInfo().getWithOffset(4),
3030 Ld->isVolatile(), Ld->isNonTemporal(),
3031 Ld->isInvariant(), NewAlign);
3035 llvm_unreachable("Unknown VFP cmp argument!");
3038 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3039 /// f32 and even f64 comparisons to integer ones.
3041 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3042 SDValue Chain = Op.getOperand(0);
3043 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3044 SDValue LHS = Op.getOperand(2);
3045 SDValue RHS = Op.getOperand(3);
3046 SDValue Dest = Op.getOperand(4);
3047 DebugLoc dl = Op.getDebugLoc();
3049 bool LHSSeenZero = false;
3050 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3051 bool RHSSeenZero = false;
3052 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3053 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3054 // If unsafe fp math optimization is enabled and there are no other uses of
3055 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3056 // to an integer comparison.
3057 if (CC == ISD::SETOEQ)
3059 else if (CC == ISD::SETUNE)
3062 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3064 if (LHS.getValueType() == MVT::f32) {
3065 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3066 bitcastf32Toi32(LHS, DAG), Mask);
3067 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3068 bitcastf32Toi32(RHS, DAG), Mask);
3069 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3070 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3071 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3072 Chain, Dest, ARMcc, CCR, Cmp);
3077 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3078 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3079 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3080 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3081 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3082 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3083 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3084 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3085 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3091 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3092 SDValue Chain = Op.getOperand(0);
3093 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3094 SDValue LHS = Op.getOperand(2);
3095 SDValue RHS = Op.getOperand(3);
3096 SDValue Dest = Op.getOperand(4);
3097 DebugLoc dl = Op.getDebugLoc();
3099 if (LHS.getValueType() == MVT::i32) {
3101 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3102 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3103 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3104 Chain, Dest, ARMcc, CCR, Cmp);
3107 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3109 if (getTargetMachine().Options.UnsafeFPMath &&
3110 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3111 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3112 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3113 if (Result.getNode())
3117 ARMCC::CondCodes CondCode, CondCode2;
3118 FPCCToARMCC(CC, CondCode, CondCode2);
3120 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3121 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3122 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3123 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3124 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3125 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3126 if (CondCode2 != ARMCC::AL) {
3127 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3128 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3129 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3134 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3135 SDValue Chain = Op.getOperand(0);
3136 SDValue Table = Op.getOperand(1);
3137 SDValue Index = Op.getOperand(2);
3138 DebugLoc dl = Op.getDebugLoc();
3140 EVT PTy = getPointerTy();
3141 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3142 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3143 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3144 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3145 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3146 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3147 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3148 if (Subtarget->isThumb2()) {
3149 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3150 // which does another jump to the destination. This also makes it easier
3151 // to translate it to TBB / TBH later.
3152 // FIXME: This might not work if the function is extremely large.
3153 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3154 Addr, Op.getOperand(2), JTI, UId);
3156 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3157 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3158 MachinePointerInfo::getJumpTable(),
3159 false, false, false, 0);
3160 Chain = Addr.getValue(1);
3161 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3162 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3164 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3165 MachinePointerInfo::getJumpTable(),
3166 false, false, false, 0);
3167 Chain = Addr.getValue(1);
3168 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3172 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3173 EVT VT = Op.getValueType();
3174 DebugLoc dl = Op.getDebugLoc();
3176 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3177 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3179 return DAG.UnrollVectorOp(Op.getNode());
3182 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3183 "Invalid type for custom lowering!");
3184 if (VT != MVT::v4i16)
3185 return DAG.UnrollVectorOp(Op.getNode());
3187 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3188 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3191 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3192 EVT VT = Op.getValueType();
3194 return LowerVectorFP_TO_INT(Op, DAG);
3196 DebugLoc dl = Op.getDebugLoc();
3199 switch (Op.getOpcode()) {
3200 default: llvm_unreachable("Invalid opcode!");
3201 case ISD::FP_TO_SINT:
3202 Opc = ARMISD::FTOSI;
3204 case ISD::FP_TO_UINT:
3205 Opc = ARMISD::FTOUI;
3208 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3209 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3212 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3213 EVT VT = Op.getValueType();
3214 DebugLoc dl = Op.getDebugLoc();
3216 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3217 if (VT.getVectorElementType() == MVT::f32)
3219 return DAG.UnrollVectorOp(Op.getNode());
3222 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3223 "Invalid type for custom lowering!");
3224 if (VT != MVT::v4f32)
3225 return DAG.UnrollVectorOp(Op.getNode());
3229 switch (Op.getOpcode()) {
3230 default: llvm_unreachable("Invalid opcode!");
3231 case ISD::SINT_TO_FP:
3232 CastOpc = ISD::SIGN_EXTEND;
3233 Opc = ISD::SINT_TO_FP;
3235 case ISD::UINT_TO_FP:
3236 CastOpc = ISD::ZERO_EXTEND;
3237 Opc = ISD::UINT_TO_FP;
3241 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3242 return DAG.getNode(Opc, dl, VT, Op);
3245 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3246 EVT VT = Op.getValueType();
3248 return LowerVectorINT_TO_FP(Op, DAG);
3250 DebugLoc dl = Op.getDebugLoc();
3253 switch (Op.getOpcode()) {
3254 default: llvm_unreachable("Invalid opcode!");
3255 case ISD::SINT_TO_FP:
3256 Opc = ARMISD::SITOF;
3258 case ISD::UINT_TO_FP:
3259 Opc = ARMISD::UITOF;
3263 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3264 return DAG.getNode(Opc, dl, VT, Op);
3267 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3268 // Implement fcopysign with a fabs and a conditional fneg.
3269 SDValue Tmp0 = Op.getOperand(0);
3270 SDValue Tmp1 = Op.getOperand(1);
3271 DebugLoc dl = Op.getDebugLoc();
3272 EVT VT = Op.getValueType();
3273 EVT SrcVT = Tmp1.getValueType();
3274 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3275 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3276 bool UseNEON = !InGPR && Subtarget->hasNEON();
3279 // Use VBSL to copy the sign bit.
3280 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3281 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3282 DAG.getTargetConstant(EncodedVal, MVT::i32));
3283 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3285 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3286 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3287 DAG.getConstant(32, MVT::i32));
3288 else /*if (VT == MVT::f32)*/
3289 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3290 if (SrcVT == MVT::f32) {
3291 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3293 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3294 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3295 DAG.getConstant(32, MVT::i32));
3296 } else if (VT == MVT::f32)
3297 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3298 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3299 DAG.getConstant(32, MVT::i32));
3300 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3301 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3303 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3305 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3306 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3307 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3309 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3310 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3311 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3312 if (VT == MVT::f32) {
3313 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3314 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3315 DAG.getConstant(0, MVT::i32));
3317 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3323 // Bitcast operand 1 to i32.
3324 if (SrcVT == MVT::f64)
3325 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3326 &Tmp1, 1).getValue(1);
3327 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3329 // Or in the signbit with integer operations.
3330 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3331 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3332 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3333 if (VT == MVT::f32) {
3334 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3335 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3336 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3337 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3340 // f64: Or the high part with signbit and then combine two parts.
3341 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3343 SDValue Lo = Tmp0.getValue(0);
3344 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3345 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3346 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3349 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3350 MachineFunction &MF = DAG.getMachineFunction();
3351 MachineFrameInfo *MFI = MF.getFrameInfo();
3352 MFI->setReturnAddressIsTaken(true);
3354 EVT VT = Op.getValueType();
3355 DebugLoc dl = Op.getDebugLoc();
3356 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3358 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3359 SDValue Offset = DAG.getConstant(4, MVT::i32);
3360 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3361 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3362 MachinePointerInfo(), false, false, false, 0);
3365 // Return LR, which contains the return address. Mark it an implicit live-in.
3366 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3367 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3370 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3371 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3372 MFI->setFrameAddressIsTaken(true);
3374 EVT VT = Op.getValueType();
3375 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3376 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3377 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3378 ? ARM::R7 : ARM::R11;
3379 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3381 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3382 MachinePointerInfo(),
3383 false, false, false, 0);
3387 /// ExpandBITCAST - If the target supports VFP, this function is called to
3388 /// expand a bit convert where either the source or destination type is i64 to
3389 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3390 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3391 /// vectors), since the legalizer won't know what to do with that.
3392 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3393 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3394 DebugLoc dl = N->getDebugLoc();
3395 SDValue Op = N->getOperand(0);
3397 // This function is only supposed to be called for i64 types, either as the
3398 // source or destination of the bit convert.
3399 EVT SrcVT = Op.getValueType();
3400 EVT DstVT = N->getValueType(0);
3401 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3402 "ExpandBITCAST called for non-i64 type");
3404 // Turn i64->f64 into VMOVDRR.
3405 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3406 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3407 DAG.getConstant(0, MVT::i32));
3408 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3409 DAG.getConstant(1, MVT::i32));
3410 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3411 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3414 // Turn f64->i64 into VMOVRRD.
3415 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3416 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3417 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3418 // Merge the pieces into a single i64 value.
3419 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3425 /// getZeroVector - Returns a vector of specified type with all zero elements.
3426 /// Zero vectors are used to represent vector negation and in those cases
3427 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3428 /// not support i64 elements, so sometimes the zero vectors will need to be
3429 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3431 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3432 assert(VT.isVector() && "Expected a vector type");
3433 // The canonical modified immediate encoding of a zero vector is....0!
3434 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3435 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3436 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3437 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3440 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3441 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3442 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3443 SelectionDAG &DAG) const {
3444 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3445 EVT VT = Op.getValueType();
3446 unsigned VTBits = VT.getSizeInBits();
3447 DebugLoc dl = Op.getDebugLoc();
3448 SDValue ShOpLo = Op.getOperand(0);
3449 SDValue ShOpHi = Op.getOperand(1);
3450 SDValue ShAmt = Op.getOperand(2);
3452 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3454 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3456 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3457 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3458 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3459 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3460 DAG.getConstant(VTBits, MVT::i32));
3461 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3462 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3463 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3465 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3466 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3468 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3469 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3472 SDValue Ops[2] = { Lo, Hi };
3473 return DAG.getMergeValues(Ops, 2, dl);
3476 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3477 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3478 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3479 SelectionDAG &DAG) const {
3480 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3481 EVT VT = Op.getValueType();
3482 unsigned VTBits = VT.getSizeInBits();
3483 DebugLoc dl = Op.getDebugLoc();
3484 SDValue ShOpLo = Op.getOperand(0);
3485 SDValue ShOpHi = Op.getOperand(1);
3486 SDValue ShAmt = Op.getOperand(2);
3489 assert(Op.getOpcode() == ISD::SHL_PARTS);
3490 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3491 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3492 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3493 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3494 DAG.getConstant(VTBits, MVT::i32));
3495 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3496 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3498 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3499 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3500 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3502 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3503 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3506 SDValue Ops[2] = { Lo, Hi };
3507 return DAG.getMergeValues(Ops, 2, dl);
3510 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3511 SelectionDAG &DAG) const {
3512 // The rounding mode is in bits 23:22 of the FPSCR.
3513 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3514 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3515 // so that the shift + and get folded into a bitfield extract.
3516 DebugLoc dl = Op.getDebugLoc();
3517 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3518 DAG.getConstant(Intrinsic::arm_get_fpscr,
3520 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3521 DAG.getConstant(1U << 22, MVT::i32));
3522 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3523 DAG.getConstant(22, MVT::i32));
3524 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3525 DAG.getConstant(3, MVT::i32));
3528 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3529 const ARMSubtarget *ST) {
3530 EVT VT = N->getValueType(0);
3531 DebugLoc dl = N->getDebugLoc();
3533 if (!ST->hasV6T2Ops())
3536 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3537 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3540 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3541 const ARMSubtarget *ST) {
3542 EVT VT = N->getValueType(0);
3543 DebugLoc dl = N->getDebugLoc();
3548 // Lower vector shifts on NEON to use VSHL.
3549 assert(ST->hasNEON() && "unexpected vector shift");
3551 // Left shifts translate directly to the vshiftu intrinsic.
3552 if (N->getOpcode() == ISD::SHL)
3553 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3554 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3555 N->getOperand(0), N->getOperand(1));
3557 assert((N->getOpcode() == ISD::SRA ||
3558 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3560 // NEON uses the same intrinsics for both left and right shifts. For
3561 // right shifts, the shift amounts are negative, so negate the vector of
3563 EVT ShiftVT = N->getOperand(1).getValueType();
3564 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3565 getZeroVector(ShiftVT, DAG, dl),
3567 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3568 Intrinsic::arm_neon_vshifts :
3569 Intrinsic::arm_neon_vshiftu);
3570 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3571 DAG.getConstant(vshiftInt, MVT::i32),
3572 N->getOperand(0), NegatedCount);
3575 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3576 const ARMSubtarget *ST) {
3577 EVT VT = N->getValueType(0);
3578 DebugLoc dl = N->getDebugLoc();
3580 // We can get here for a node like i32 = ISD::SHL i32, i64
3584 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3585 "Unknown shift to lower!");
3587 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3588 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3589 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3592 // If we are in thumb mode, we don't have RRX.
3593 if (ST->isThumb1Only()) return SDValue();
3595 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3596 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3597 DAG.getConstant(0, MVT::i32));
3598 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3599 DAG.getConstant(1, MVT::i32));
3601 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3602 // captures the result into a carry flag.
3603 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3604 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3606 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3607 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3609 // Merge the pieces into a single i64 value.
3610 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3613 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3614 SDValue TmpOp0, TmpOp1;
3615 bool Invert = false;
3619 SDValue Op0 = Op.getOperand(0);
3620 SDValue Op1 = Op.getOperand(1);
3621 SDValue CC = Op.getOperand(2);
3622 EVT VT = Op.getValueType();
3623 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3624 DebugLoc dl = Op.getDebugLoc();
3626 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3627 switch (SetCCOpcode) {
3628 default: llvm_unreachable("Illegal FP comparison");
3630 case ISD::SETNE: Invert = true; // Fallthrough
3632 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3634 case ISD::SETLT: Swap = true; // Fallthrough
3636 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3638 case ISD::SETLE: Swap = true; // Fallthrough
3640 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3641 case ISD::SETUGE: Swap = true; // Fallthrough
3642 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3643 case ISD::SETUGT: Swap = true; // Fallthrough
3644 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3645 case ISD::SETUEQ: Invert = true; // Fallthrough
3647 // Expand this to (OLT | OGT).
3651 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3652 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3654 case ISD::SETUO: Invert = true; // Fallthrough
3656 // Expand this to (OLT | OGE).
3660 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3661 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3665 // Integer comparisons.
3666 switch (SetCCOpcode) {
3667 default: llvm_unreachable("Illegal integer comparison");
3668 case ISD::SETNE: Invert = true;
3669 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3670 case ISD::SETLT: Swap = true;
3671 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3672 case ISD::SETLE: Swap = true;
3673 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3674 case ISD::SETULT: Swap = true;
3675 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3676 case ISD::SETULE: Swap = true;
3677 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3680 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3681 if (Opc == ARMISD::VCEQ) {
3684 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3686 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3689 // Ignore bitconvert.
3690 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3691 AndOp = AndOp.getOperand(0);
3693 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3695 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3696 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3703 std::swap(Op0, Op1);
3705 // If one of the operands is a constant vector zero, attempt to fold the
3706 // comparison to a specialized compare-against-zero form.
3708 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3710 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3711 if (Opc == ARMISD::VCGE)
3712 Opc = ARMISD::VCLEZ;
3713 else if (Opc == ARMISD::VCGT)
3714 Opc = ARMISD::VCLTZ;
3719 if (SingleOp.getNode()) {
3722 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3724 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3726 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3728 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3730 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3732 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3735 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3739 Result = DAG.getNOT(dl, Result, VT);
3744 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3745 /// valid vector constant for a NEON instruction with a "modified immediate"
3746 /// operand (e.g., VMOV). If so, return the encoded value.
3747 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3748 unsigned SplatBitSize, SelectionDAG &DAG,
3749 EVT &VT, bool is128Bits, NEONModImmType type) {
3750 unsigned OpCmode, Imm;
3752 // SplatBitSize is set to the smallest size that splats the vector, so a
3753 // zero vector will always have SplatBitSize == 8. However, NEON modified
3754 // immediate instructions others than VMOV do not support the 8-bit encoding
3755 // of a zero vector, and the default encoding of zero is supposed to be the
3760 switch (SplatBitSize) {
3762 if (type != VMOVModImm)
3764 // Any 1-byte value is OK. Op=0, Cmode=1110.
3765 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3768 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3772 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3773 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3774 if ((SplatBits & ~0xff) == 0) {
3775 // Value = 0x00nn: Op=x, Cmode=100x.
3780 if ((SplatBits & ~0xff00) == 0) {
3781 // Value = 0xnn00: Op=x, Cmode=101x.
3783 Imm = SplatBits >> 8;
3789 // NEON's 32-bit VMOV supports splat values where:
3790 // * only one byte is nonzero, or
3791 // * the least significant byte is 0xff and the second byte is nonzero, or
3792 // * the least significant 2 bytes are 0xff and the third is nonzero.
3793 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3794 if ((SplatBits & ~0xff) == 0) {
3795 // Value = 0x000000nn: Op=x, Cmode=000x.
3800 if ((SplatBits & ~0xff00) == 0) {
3801 // Value = 0x0000nn00: Op=x, Cmode=001x.
3803 Imm = SplatBits >> 8;
3806 if ((SplatBits & ~0xff0000) == 0) {
3807 // Value = 0x00nn0000: Op=x, Cmode=010x.
3809 Imm = SplatBits >> 16;
3812 if ((SplatBits & ~0xff000000) == 0) {
3813 // Value = 0xnn000000: Op=x, Cmode=011x.
3815 Imm = SplatBits >> 24;
3819 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3820 if (type == OtherModImm) return SDValue();
3822 if ((SplatBits & ~0xffff) == 0 &&
3823 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3824 // Value = 0x0000nnff: Op=x, Cmode=1100.
3826 Imm = SplatBits >> 8;
3831 if ((SplatBits & ~0xffffff) == 0 &&
3832 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3833 // Value = 0x00nnffff: Op=x, Cmode=1101.
3835 Imm = SplatBits >> 16;
3836 SplatBits |= 0xffff;
3840 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3841 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3842 // VMOV.I32. A (very) minor optimization would be to replicate the value
3843 // and fall through here to test for a valid 64-bit splat. But, then the
3844 // caller would also need to check and handle the change in size.
3848 if (type != VMOVModImm)
3850 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3851 uint64_t BitMask = 0xff;
3853 unsigned ImmMask = 1;
3855 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3856 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3859 } else if ((SplatBits & BitMask) != 0) {
3865 // Op=1, Cmode=1110.
3868 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3873 llvm_unreachable("unexpected size for isNEONModifiedImm");
3876 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3877 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3880 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3881 const ARMSubtarget *ST) const {
3882 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3885 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3886 assert(Op.getValueType() == MVT::f32 &&
3887 "ConstantFP custom lowering should only occur for f32.");
3889 // Try splatting with a VMOV.f32...
3890 APFloat FPVal = CFP->getValueAPF();
3891 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3893 DebugLoc DL = Op.getDebugLoc();
3894 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3895 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3897 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3898 DAG.getConstant(0, MVT::i32));
3901 // If that fails, try a VMOV.i32
3903 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3904 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3906 if (NewVal != SDValue()) {
3907 DebugLoc DL = Op.getDebugLoc();
3908 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
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));
3916 // Finally, try a VMVN.i32
3917 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3919 if (NewVal != SDValue()) {
3920 DebugLoc DL = Op.getDebugLoc();
3921 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3922 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3924 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3925 DAG.getConstant(0, MVT::i32));
3931 // check if an VEXT instruction can handle the shuffle mask when the
3932 // vector sources of the shuffle are the same.
3933 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3934 unsigned NumElts = VT.getVectorNumElements();
3936 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3942 // If this is a VEXT shuffle, the immediate value is the index of the first
3943 // element. The other shuffle indices must be the successive elements after
3945 unsigned ExpectedElt = Imm;
3946 for (unsigned i = 1; i < NumElts; ++i) {
3947 // Increment the expected index. If it wraps around, just follow it
3948 // back to index zero and keep going.
3950 if (ExpectedElt == NumElts)
3953 if (M[i] < 0) continue; // ignore UNDEF indices
3954 if (ExpectedElt != static_cast<unsigned>(M[i]))
3962 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3963 bool &ReverseVEXT, unsigned &Imm) {
3964 unsigned NumElts = VT.getVectorNumElements();
3965 ReverseVEXT = false;
3967 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3973 // If this is a VEXT shuffle, the immediate value is the index of the first
3974 // element. The other shuffle indices must be the successive elements after
3976 unsigned ExpectedElt = Imm;
3977 for (unsigned i = 1; i < NumElts; ++i) {
3978 // Increment the expected index. If it wraps around, it may still be
3979 // a VEXT but the source vectors must be swapped.
3981 if (ExpectedElt == NumElts * 2) {
3986 if (M[i] < 0) continue; // ignore UNDEF indices
3987 if (ExpectedElt != static_cast<unsigned>(M[i]))
3991 // Adjust the index value if the source operands will be swapped.
3998 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3999 /// instruction with the specified blocksize. (The order of the elements
4000 /// within each block of the vector is reversed.)
4001 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4002 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4003 "Only possible block sizes for VREV are: 16, 32, 64");
4005 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4009 unsigned NumElts = VT.getVectorNumElements();
4010 unsigned BlockElts = M[0] + 1;
4011 // If the first shuffle index is UNDEF, be optimistic.
4013 BlockElts = BlockSize / EltSz;
4015 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4018 for (unsigned i = 0; i < NumElts; ++i) {
4019 if (M[i] < 0) continue; // ignore UNDEF indices
4020 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4027 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4028 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4029 // range, then 0 is placed into the resulting vector. So pretty much any mask
4030 // of 8 elements can work here.
4031 return VT == MVT::v8i8 && M.size() == 8;
4034 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4035 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4039 unsigned NumElts = VT.getVectorNumElements();
4040 WhichResult = (M[0] == 0 ? 0 : 1);
4041 for (unsigned i = 0; i < NumElts; i += 2) {
4042 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4043 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4049 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4050 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4051 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4052 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4053 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4057 unsigned NumElts = VT.getVectorNumElements();
4058 WhichResult = (M[0] == 0 ? 0 : 1);
4059 for (unsigned i = 0; i < NumElts; i += 2) {
4060 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4061 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4067 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4068 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4072 unsigned NumElts = VT.getVectorNumElements();
4073 WhichResult = (M[0] == 0 ? 0 : 1);
4074 for (unsigned i = 0; i != NumElts; ++i) {
4075 if (M[i] < 0) continue; // ignore UNDEF indices
4076 if ((unsigned) M[i] != 2 * i + WhichResult)
4080 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4081 if (VT.is64BitVector() && EltSz == 32)
4087 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4088 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4089 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4090 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4091 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4095 unsigned Half = VT.getVectorNumElements() / 2;
4096 WhichResult = (M[0] == 0 ? 0 : 1);
4097 for (unsigned j = 0; j != 2; ++j) {
4098 unsigned Idx = WhichResult;
4099 for (unsigned i = 0; i != Half; ++i) {
4100 int MIdx = M[i + j * Half];
4101 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4107 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4108 if (VT.is64BitVector() && EltSz == 32)
4114 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4115 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4119 unsigned NumElts = VT.getVectorNumElements();
4120 WhichResult = (M[0] == 0 ? 0 : 1);
4121 unsigned Idx = WhichResult * NumElts / 2;
4122 for (unsigned i = 0; i != NumElts; i += 2) {
4123 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4124 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4129 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4130 if (VT.is64BitVector() && EltSz == 32)
4136 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4137 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4138 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4139 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4140 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4144 unsigned NumElts = VT.getVectorNumElements();
4145 WhichResult = (M[0] == 0 ? 0 : 1);
4146 unsigned Idx = WhichResult * NumElts / 2;
4147 for (unsigned i = 0; i != NumElts; i += 2) {
4148 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4149 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4154 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4155 if (VT.is64BitVector() && EltSz == 32)
4161 // If N is an integer constant that can be moved into a register in one
4162 // instruction, return an SDValue of such a constant (will become a MOV
4163 // instruction). Otherwise return null.
4164 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4165 const ARMSubtarget *ST, DebugLoc dl) {
4167 if (!isa<ConstantSDNode>(N))
4169 Val = cast<ConstantSDNode>(N)->getZExtValue();
4171 if (ST->isThumb1Only()) {
4172 if (Val <= 255 || ~Val <= 255)
4173 return DAG.getConstant(Val, MVT::i32);
4175 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4176 return DAG.getConstant(Val, MVT::i32);
4181 // If this is a case we can't handle, return null and let the default
4182 // expansion code take care of it.
4183 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4184 const ARMSubtarget *ST) const {
4185 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4186 DebugLoc dl = Op.getDebugLoc();
4187 EVT VT = Op.getValueType();
4189 APInt SplatBits, SplatUndef;
4190 unsigned SplatBitSize;
4192 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4193 if (SplatBitSize <= 64) {
4194 // Check if an immediate VMOV works.
4196 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4197 SplatUndef.getZExtValue(), SplatBitSize,
4198 DAG, VmovVT, VT.is128BitVector(),
4200 if (Val.getNode()) {
4201 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4202 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4205 // Try an immediate VMVN.
4206 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4207 Val = isNEONModifiedImm(NegatedImm,
4208 SplatUndef.getZExtValue(), SplatBitSize,
4209 DAG, VmovVT, VT.is128BitVector(),
4211 if (Val.getNode()) {
4212 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4213 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4216 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4217 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4218 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4220 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4221 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4227 // Scan through the operands to see if only one value is used.
4229 // As an optimisation, even if more than one value is used it may be more
4230 // profitable to splat with one value then change some lanes.
4232 // Heuristically we decide to do this if the vector has a "dominant" value,
4233 // defined as splatted to more than half of the lanes.
4234 unsigned NumElts = VT.getVectorNumElements();
4235 bool isOnlyLowElement = true;
4236 bool usesOnlyOneValue = true;
4237 bool hasDominantValue = false;
4238 bool isConstant = true;
4240 // Map of the number of times a particular SDValue appears in the
4242 DenseMap<SDValue, unsigned> ValueCounts;
4244 for (unsigned i = 0; i < NumElts; ++i) {
4245 SDValue V = Op.getOperand(i);
4246 if (V.getOpcode() == ISD::UNDEF)
4249 isOnlyLowElement = false;
4250 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4253 ValueCounts.insert(std::make_pair(V, 0));
4254 unsigned &Count = ValueCounts[V];
4256 // Is this value dominant? (takes up more than half of the lanes)
4257 if (++Count > (NumElts / 2)) {
4258 hasDominantValue = true;
4262 if (ValueCounts.size() != 1)
4263 usesOnlyOneValue = false;
4264 if (!Value.getNode() && ValueCounts.size() > 0)
4265 Value = ValueCounts.begin()->first;
4267 if (ValueCounts.size() == 0)
4268 return DAG.getUNDEF(VT);
4270 if (isOnlyLowElement)
4271 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4273 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4275 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4276 // i32 and try again.
4277 if (hasDominantValue && EltSize <= 32) {
4281 // If we are VDUPing a value that comes directly from a vector, that will
4282 // cause an unnecessary move to and from a GPR, where instead we could
4283 // just use VDUPLANE.
4284 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4285 // We need to create a new undef vector to use for the VDUPLANE if the
4286 // size of the vector from which we get the value is different than the
4287 // size of the vector that we need to create. We will insert the element
4288 // such that the register coalescer will remove unnecessary copies.
4289 if (VT != Value->getOperand(0).getValueType()) {
4290 ConstantSDNode *constIndex;
4291 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4292 assert(constIndex && "The index is not a constant!");
4293 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4294 VT.getVectorNumElements();
4295 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4296 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4297 Value, DAG.getConstant(index, MVT::i32)),
4298 DAG.getConstant(index, MVT::i32));
4300 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4301 Value->getOperand(0), Value->getOperand(1));
4305 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4307 if (!usesOnlyOneValue) {
4308 // The dominant value was splatted as 'N', but we now have to insert
4309 // all differing elements.
4310 for (unsigned I = 0; I < NumElts; ++I) {
4311 if (Op.getOperand(I) == Value)
4313 SmallVector<SDValue, 3> Ops;
4315 Ops.push_back(Op.getOperand(I));
4316 Ops.push_back(DAG.getConstant(I, MVT::i32));
4317 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4322 if (VT.getVectorElementType().isFloatingPoint()) {
4323 SmallVector<SDValue, 8> Ops;
4324 for (unsigned i = 0; i < NumElts; ++i)
4325 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4327 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4328 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4329 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4331 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4333 if (usesOnlyOneValue) {
4334 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4335 if (isConstant && Val.getNode())
4336 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4340 // If all elements are constants and the case above didn't get hit, fall back
4341 // to the default expansion, which will generate a load from the constant
4346 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4348 SDValue shuffle = ReconstructShuffle(Op, DAG);
4349 if (shuffle != SDValue())
4353 // Vectors with 32- or 64-bit elements can be built by directly assigning
4354 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4355 // will be legalized.
4356 if (EltSize >= 32) {
4357 // Do the expansion with floating-point types, since that is what the VFP
4358 // registers are defined to use, and since i64 is not legal.
4359 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4360 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4361 SmallVector<SDValue, 8> Ops;
4362 for (unsigned i = 0; i < NumElts; ++i)
4363 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4364 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4365 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4371 // Gather data to see if the operation can be modelled as a
4372 // shuffle in combination with VEXTs.
4373 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4374 SelectionDAG &DAG) const {
4375 DebugLoc dl = Op.getDebugLoc();
4376 EVT VT = Op.getValueType();
4377 unsigned NumElts = VT.getVectorNumElements();
4379 SmallVector<SDValue, 2> SourceVecs;
4380 SmallVector<unsigned, 2> MinElts;
4381 SmallVector<unsigned, 2> MaxElts;
4383 for (unsigned i = 0; i < NumElts; ++i) {
4384 SDValue V = Op.getOperand(i);
4385 if (V.getOpcode() == ISD::UNDEF)
4387 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4388 // A shuffle can only come from building a vector from various
4389 // elements of other vectors.
4391 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4392 VT.getVectorElementType()) {
4393 // This code doesn't know how to handle shuffles where the vector
4394 // element types do not match (this happens because type legalization
4395 // promotes the return type of EXTRACT_VECTOR_ELT).
4396 // FIXME: It might be appropriate to extend this code to handle
4397 // mismatched types.
4401 // Record this extraction against the appropriate vector if possible...
4402 SDValue SourceVec = V.getOperand(0);
4403 // If the element number isn't a constant, we can't effectively
4404 // analyze what's going on.
4405 if (!isa<ConstantSDNode>(V.getOperand(1)))
4407 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4408 bool FoundSource = false;
4409 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4410 if (SourceVecs[j] == SourceVec) {
4411 if (MinElts[j] > EltNo)
4413 if (MaxElts[j] < EltNo)
4420 // Or record a new source if not...
4422 SourceVecs.push_back(SourceVec);
4423 MinElts.push_back(EltNo);
4424 MaxElts.push_back(EltNo);
4428 // Currently only do something sane when at most two source vectors
4430 if (SourceVecs.size() > 2)
4433 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4434 int VEXTOffsets[2] = {0, 0};
4436 // This loop extracts the usage patterns of the source vectors
4437 // and prepares appropriate SDValues for a shuffle if possible.
4438 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4439 if (SourceVecs[i].getValueType() == VT) {
4440 // No VEXT necessary
4441 ShuffleSrcs[i] = SourceVecs[i];
4444 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4445 // It probably isn't worth padding out a smaller vector just to
4446 // break it down again in a shuffle.
4450 // Since only 64-bit and 128-bit vectors are legal on ARM and
4451 // we've eliminated the other cases...
4452 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4453 "unexpected vector sizes in ReconstructShuffle");
4455 if (MaxElts[i] - MinElts[i] >= NumElts) {
4456 // Span too large for a VEXT to cope
4460 if (MinElts[i] >= NumElts) {
4461 // The extraction can just take the second half
4462 VEXTOffsets[i] = NumElts;
4463 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4465 DAG.getIntPtrConstant(NumElts));
4466 } else if (MaxElts[i] < NumElts) {
4467 // The extraction can just take the first half
4469 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4471 DAG.getIntPtrConstant(0));
4473 // An actual VEXT is needed
4474 VEXTOffsets[i] = MinElts[i];
4475 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4477 DAG.getIntPtrConstant(0));
4478 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4480 DAG.getIntPtrConstant(NumElts));
4481 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4482 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4486 SmallVector<int, 8> Mask;
4488 for (unsigned i = 0; i < NumElts; ++i) {
4489 SDValue Entry = Op.getOperand(i);
4490 if (Entry.getOpcode() == ISD::UNDEF) {
4495 SDValue ExtractVec = Entry.getOperand(0);
4496 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4497 .getOperand(1))->getSExtValue();
4498 if (ExtractVec == SourceVecs[0]) {
4499 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4501 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4505 // Final check before we try to produce nonsense...
4506 if (isShuffleMaskLegal(Mask, VT))
4507 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4513 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4514 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4515 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4516 /// are assumed to be legal.
4518 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4520 if (VT.getVectorNumElements() == 4 &&
4521 (VT.is128BitVector() || VT.is64BitVector())) {
4522 unsigned PFIndexes[4];
4523 for (unsigned i = 0; i != 4; ++i) {
4527 PFIndexes[i] = M[i];
4530 // Compute the index in the perfect shuffle table.
4531 unsigned PFTableIndex =
4532 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4533 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4534 unsigned Cost = (PFEntry >> 30);
4541 unsigned Imm, WhichResult;
4543 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4544 return (EltSize >= 32 ||
4545 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4546 isVREVMask(M, VT, 64) ||
4547 isVREVMask(M, VT, 32) ||
4548 isVREVMask(M, VT, 16) ||
4549 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4550 isVTBLMask(M, VT) ||
4551 isVTRNMask(M, VT, WhichResult) ||
4552 isVUZPMask(M, VT, WhichResult) ||
4553 isVZIPMask(M, VT, WhichResult) ||
4554 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4555 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4556 isVZIP_v_undef_Mask(M, VT, WhichResult));
4559 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4560 /// the specified operations to build the shuffle.
4561 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4562 SDValue RHS, SelectionDAG &DAG,
4564 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4565 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4566 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4569 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4578 OP_VUZPL, // VUZP, left result
4579 OP_VUZPR, // VUZP, right result
4580 OP_VZIPL, // VZIP, left result
4581 OP_VZIPR, // VZIP, right result
4582 OP_VTRNL, // VTRN, left result
4583 OP_VTRNR // VTRN, right result
4586 if (OpNum == OP_COPY) {
4587 if (LHSID == (1*9+2)*9+3) return LHS;
4588 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4592 SDValue OpLHS, OpRHS;
4593 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4594 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4595 EVT VT = OpLHS.getValueType();
4598 default: llvm_unreachable("Unknown shuffle opcode!");
4600 // VREV divides the vector in half and swaps within the half.
4601 if (VT.getVectorElementType() == MVT::i32 ||
4602 VT.getVectorElementType() == MVT::f32)
4603 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4604 // vrev <4 x i16> -> VREV32
4605 if (VT.getVectorElementType() == MVT::i16)
4606 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4607 // vrev <4 x i8> -> VREV16
4608 assert(VT.getVectorElementType() == MVT::i8);
4609 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4614 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4615 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4619 return DAG.getNode(ARMISD::VEXT, dl, VT,
4621 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4624 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4625 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4628 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4629 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4632 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4633 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4637 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4638 ArrayRef<int> ShuffleMask,
4639 SelectionDAG &DAG) {
4640 // Check to see if we can use the VTBL instruction.
4641 SDValue V1 = Op.getOperand(0);
4642 SDValue V2 = Op.getOperand(1);
4643 DebugLoc DL = Op.getDebugLoc();
4645 SmallVector<SDValue, 8> VTBLMask;
4646 for (ArrayRef<int>::iterator
4647 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4648 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4650 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4651 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4652 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4655 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4656 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4660 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4661 SDValue V1 = Op.getOperand(0);
4662 SDValue V2 = Op.getOperand(1);
4663 DebugLoc dl = Op.getDebugLoc();
4664 EVT VT = Op.getValueType();
4665 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4667 // Convert shuffles that are directly supported on NEON to target-specific
4668 // DAG nodes, instead of keeping them as shuffles and matching them again
4669 // during code selection. This is more efficient and avoids the possibility
4670 // of inconsistencies between legalization and selection.
4671 // FIXME: floating-point vectors should be canonicalized to integer vectors
4672 // of the same time so that they get CSEd properly.
4673 ArrayRef<int> ShuffleMask = SVN->getMask();
4675 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4676 if (EltSize <= 32) {
4677 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4678 int Lane = SVN->getSplatIndex();
4679 // If this is undef splat, generate it via "just" vdup, if possible.
4680 if (Lane == -1) Lane = 0;
4682 // Test if V1 is a SCALAR_TO_VECTOR.
4683 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4684 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4686 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4687 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4689 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4690 !isa<ConstantSDNode>(V1.getOperand(0))) {
4691 bool IsScalarToVector = true;
4692 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4693 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4694 IsScalarToVector = false;
4697 if (IsScalarToVector)
4698 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4700 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4701 DAG.getConstant(Lane, MVT::i32));
4706 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4709 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4710 DAG.getConstant(Imm, MVT::i32));
4713 if (isVREVMask(ShuffleMask, VT, 64))
4714 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4715 if (isVREVMask(ShuffleMask, VT, 32))
4716 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4717 if (isVREVMask(ShuffleMask, VT, 16))
4718 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4720 if (V2->getOpcode() == ISD::UNDEF &&
4721 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4722 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4723 DAG.getConstant(Imm, MVT::i32));
4726 // Check for Neon shuffles that modify both input vectors in place.
4727 // If both results are used, i.e., if there are two shuffles with the same
4728 // source operands and with masks corresponding to both results of one of
4729 // these operations, DAG memoization will ensure that a single node is
4730 // used for both shuffles.
4731 unsigned WhichResult;
4732 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4733 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4734 V1, V2).getValue(WhichResult);
4735 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4736 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4737 V1, V2).getValue(WhichResult);
4738 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4739 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4740 V1, V2).getValue(WhichResult);
4742 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4743 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4744 V1, V1).getValue(WhichResult);
4745 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4746 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4747 V1, V1).getValue(WhichResult);
4748 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4749 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4750 V1, V1).getValue(WhichResult);
4753 // If the shuffle is not directly supported and it has 4 elements, use
4754 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4755 unsigned NumElts = VT.getVectorNumElements();
4757 unsigned PFIndexes[4];
4758 for (unsigned i = 0; i != 4; ++i) {
4759 if (ShuffleMask[i] < 0)
4762 PFIndexes[i] = ShuffleMask[i];
4765 // Compute the index in the perfect shuffle table.
4766 unsigned PFTableIndex =
4767 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4768 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4769 unsigned Cost = (PFEntry >> 30);
4772 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4775 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4776 if (EltSize >= 32) {
4777 // Do the expansion with floating-point types, since that is what the VFP
4778 // registers are defined to use, and since i64 is not legal.
4779 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4780 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4781 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4782 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4783 SmallVector<SDValue, 8> Ops;
4784 for (unsigned i = 0; i < NumElts; ++i) {
4785 if (ShuffleMask[i] < 0)
4786 Ops.push_back(DAG.getUNDEF(EltVT));
4788 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4789 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4790 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4793 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4794 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4797 if (VT == MVT::v8i8) {
4798 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4799 if (NewOp.getNode())
4806 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4807 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4808 SDValue Lane = Op.getOperand(2);
4809 if (!isa<ConstantSDNode>(Lane))
4815 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4816 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4817 SDValue Lane = Op.getOperand(1);
4818 if (!isa<ConstantSDNode>(Lane))
4821 SDValue Vec = Op.getOperand(0);
4822 if (Op.getValueType() == MVT::i32 &&
4823 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4824 DebugLoc dl = Op.getDebugLoc();
4825 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4831 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4832 // The only time a CONCAT_VECTORS operation can have legal types is when
4833 // two 64-bit vectors are concatenated to a 128-bit vector.
4834 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4835 "unexpected CONCAT_VECTORS");
4836 DebugLoc dl = Op.getDebugLoc();
4837 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4838 SDValue Op0 = Op.getOperand(0);
4839 SDValue Op1 = Op.getOperand(1);
4840 if (Op0.getOpcode() != ISD::UNDEF)
4841 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4842 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4843 DAG.getIntPtrConstant(0));
4844 if (Op1.getOpcode() != ISD::UNDEF)
4845 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4846 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4847 DAG.getIntPtrConstant(1));
4848 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4851 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4852 /// element has been zero/sign-extended, depending on the isSigned parameter,
4853 /// from an integer type half its size.
4854 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4856 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4857 EVT VT = N->getValueType(0);
4858 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4859 SDNode *BVN = N->getOperand(0).getNode();
4860 if (BVN->getValueType(0) != MVT::v4i32 ||
4861 BVN->getOpcode() != ISD::BUILD_VECTOR)
4863 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4864 unsigned HiElt = 1 - LoElt;
4865 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4866 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4867 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4868 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4869 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4872 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4873 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4876 if (Hi0->isNullValue() && Hi1->isNullValue())
4882 if (N->getOpcode() != ISD::BUILD_VECTOR)
4885 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4886 SDNode *Elt = N->getOperand(i).getNode();
4887 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4888 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4889 unsigned HalfSize = EltSize / 2;
4891 if (!isIntN(HalfSize, C->getSExtValue()))
4894 if (!isUIntN(HalfSize, C->getZExtValue()))
4905 /// isSignExtended - Check if a node is a vector value that is sign-extended
4906 /// or a constant BUILD_VECTOR with sign-extended elements.
4907 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4908 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4910 if (isExtendedBUILD_VECTOR(N, DAG, true))
4915 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4916 /// or a constant BUILD_VECTOR with zero-extended elements.
4917 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4918 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4920 if (isExtendedBUILD_VECTOR(N, DAG, false))
4925 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4926 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4927 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4928 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4929 return N->getOperand(0);
4930 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4931 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4932 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4933 LD->isNonTemporal(), LD->isInvariant(),
4934 LD->getAlignment());
4935 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4936 // have been legalized as a BITCAST from v4i32.
4937 if (N->getOpcode() == ISD::BITCAST) {
4938 SDNode *BVN = N->getOperand(0).getNode();
4939 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4940 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4941 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4942 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4943 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4945 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4946 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4947 EVT VT = N->getValueType(0);
4948 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4949 unsigned NumElts = VT.getVectorNumElements();
4950 MVT TruncVT = MVT::getIntegerVT(EltSize);
4951 SmallVector<SDValue, 8> Ops;
4952 for (unsigned i = 0; i != NumElts; ++i) {
4953 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4954 const APInt &CInt = C->getAPIntValue();
4955 // Element types smaller than 32 bits are not legal, so use i32 elements.
4956 // The values are implicitly truncated so sext vs. zext doesn't matter.
4957 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4959 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4960 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4963 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4964 unsigned Opcode = N->getOpcode();
4965 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4966 SDNode *N0 = N->getOperand(0).getNode();
4967 SDNode *N1 = N->getOperand(1).getNode();
4968 return N0->hasOneUse() && N1->hasOneUse() &&
4969 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4974 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4975 unsigned Opcode = N->getOpcode();
4976 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4977 SDNode *N0 = N->getOperand(0).getNode();
4978 SDNode *N1 = N->getOperand(1).getNode();
4979 return N0->hasOneUse() && N1->hasOneUse() &&
4980 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4985 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4986 // Multiplications are only custom-lowered for 128-bit vectors so that
4987 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4988 EVT VT = Op.getValueType();
4989 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4990 SDNode *N0 = Op.getOperand(0).getNode();
4991 SDNode *N1 = Op.getOperand(1).getNode();
4992 unsigned NewOpc = 0;
4994 bool isN0SExt = isSignExtended(N0, DAG);
4995 bool isN1SExt = isSignExtended(N1, DAG);
4996 if (isN0SExt && isN1SExt)
4997 NewOpc = ARMISD::VMULLs;
4999 bool isN0ZExt = isZeroExtended(N0, DAG);
5000 bool isN1ZExt = isZeroExtended(N1, DAG);
5001 if (isN0ZExt && isN1ZExt)
5002 NewOpc = ARMISD::VMULLu;
5003 else if (isN1SExt || isN1ZExt) {
5004 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5005 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5006 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5007 NewOpc = ARMISD::VMULLs;
5009 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5010 NewOpc = ARMISD::VMULLu;
5012 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5014 NewOpc = ARMISD::VMULLu;
5020 if (VT == MVT::v2i64)
5021 // Fall through to expand this. It is not legal.
5024 // Other vector multiplications are legal.
5029 // Legalize to a VMULL instruction.
5030 DebugLoc DL = Op.getDebugLoc();
5032 SDValue Op1 = SkipExtension(N1, DAG);
5034 Op0 = SkipExtension(N0, DAG);
5035 assert(Op0.getValueType().is64BitVector() &&
5036 Op1.getValueType().is64BitVector() &&
5037 "unexpected types for extended operands to VMULL");
5038 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5041 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5042 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5049 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5050 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5051 EVT Op1VT = Op1.getValueType();
5052 return DAG.getNode(N0->getOpcode(), DL, VT,
5053 DAG.getNode(NewOpc, DL, VT,
5054 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5055 DAG.getNode(NewOpc, DL, VT,
5056 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5060 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5062 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5063 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5064 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5065 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5066 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5067 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5068 // Get reciprocal estimate.
5069 // float4 recip = vrecpeq_f32(yf);
5070 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5071 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5072 // Because char has a smaller range than uchar, we can actually get away
5073 // without any newton steps. This requires that we use a weird bias
5074 // of 0xb000, however (again, this has been exhaustively tested).
5075 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5076 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5077 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5078 Y = DAG.getConstant(0xb000, MVT::i32);
5079 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5080 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5081 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5082 // Convert back to short.
5083 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5084 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5089 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5091 // Convert to float.
5092 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5093 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5094 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5095 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5096 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5097 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5099 // Use reciprocal estimate and one refinement step.
5100 // float4 recip = vrecpeq_f32(yf);
5101 // recip *= vrecpsq_f32(yf, recip);
5102 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5103 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5104 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5105 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5107 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5108 // Because short has a smaller range than ushort, we can actually get away
5109 // with only a single newton step. This requires that we use a weird bias
5110 // of 89, however (again, this has been exhaustively tested).
5111 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5112 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5113 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5114 N1 = DAG.getConstant(0x89, MVT::i32);
5115 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5116 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5117 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5118 // Convert back to integer and return.
5119 // return vmovn_s32(vcvt_s32_f32(result));
5120 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5121 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5125 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5126 EVT VT = Op.getValueType();
5127 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5128 "unexpected type for custom-lowering ISD::SDIV");
5130 DebugLoc dl = Op.getDebugLoc();
5131 SDValue N0 = Op.getOperand(0);
5132 SDValue N1 = Op.getOperand(1);
5135 if (VT == MVT::v8i8) {
5136 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5137 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5139 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5140 DAG.getIntPtrConstant(4));
5141 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5142 DAG.getIntPtrConstant(4));
5143 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5144 DAG.getIntPtrConstant(0));
5145 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5146 DAG.getIntPtrConstant(0));
5148 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5149 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5151 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5152 N0 = LowerCONCAT_VECTORS(N0, DAG);
5154 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5157 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5160 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5161 EVT VT = Op.getValueType();
5162 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5163 "unexpected type for custom-lowering ISD::UDIV");
5165 DebugLoc dl = Op.getDebugLoc();
5166 SDValue N0 = Op.getOperand(0);
5167 SDValue N1 = Op.getOperand(1);
5170 if (VT == MVT::v8i8) {
5171 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5172 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5174 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5175 DAG.getIntPtrConstant(4));
5176 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5177 DAG.getIntPtrConstant(4));
5178 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5179 DAG.getIntPtrConstant(0));
5180 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5181 DAG.getIntPtrConstant(0));
5183 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5184 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5186 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5187 N0 = LowerCONCAT_VECTORS(N0, DAG);
5189 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5190 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5195 // v4i16 sdiv ... Convert to float.
5196 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5197 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5198 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5199 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5200 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5201 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5203 // Use reciprocal estimate and two refinement steps.
5204 // float4 recip = vrecpeq_f32(yf);
5205 // recip *= vrecpsq_f32(yf, recip);
5206 // recip *= vrecpsq_f32(yf, recip);
5207 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5208 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5209 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5210 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5212 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5213 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5214 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5216 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5217 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5218 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5219 // and that it will never cause us to return an answer too large).
5220 // float4 result = as_float4(as_int4(xf*recip) + 2);
5221 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5222 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5223 N1 = DAG.getConstant(2, MVT::i32);
5224 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5225 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5226 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5227 // Convert back to integer and return.
5228 // return vmovn_u32(vcvt_s32_f32(result));
5229 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5230 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5234 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5235 EVT VT = Op.getNode()->getValueType(0);
5236 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5239 bool ExtraOp = false;
5240 switch (Op.getOpcode()) {
5241 default: llvm_unreachable("Invalid code");
5242 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5243 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5244 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5245 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5249 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5251 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5252 Op.getOperand(1), Op.getOperand(2));
5255 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5256 // Monotonic load/store is legal for all targets
5257 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5260 // Aquire/Release load/store is not legal for targets without a
5261 // dmb or equivalent available.
5267 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5268 SelectionDAG &DAG, unsigned NewOp) {
5269 DebugLoc dl = Node->getDebugLoc();
5270 assert (Node->getValueType(0) == MVT::i64 &&
5271 "Only know how to expand i64 atomics");
5273 SmallVector<SDValue, 6> Ops;
5274 Ops.push_back(Node->getOperand(0)); // Chain
5275 Ops.push_back(Node->getOperand(1)); // Ptr
5277 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5278 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5279 // High part of Val1
5280 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5281 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5282 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5283 // High part of Val1
5284 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5285 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5286 // High part of Val2
5287 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5288 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5290 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5292 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5293 cast<MemSDNode>(Node)->getMemOperand());
5294 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5295 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5296 Results.push_back(Result.getValue(2));
5299 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5300 switch (Op.getOpcode()) {
5301 default: llvm_unreachable("Don't know how to custom lower this!");
5302 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5303 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5304 case ISD::GlobalAddress:
5305 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5306 LowerGlobalAddressELF(Op, DAG);
5307 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5308 case ISD::SELECT: return LowerSELECT(Op, DAG);
5309 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5310 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5311 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5312 case ISD::VASTART: return LowerVASTART(Op, DAG);
5313 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5314 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5315 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5316 case ISD::SINT_TO_FP:
5317 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5318 case ISD::FP_TO_SINT:
5319 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5320 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5321 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5322 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5323 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5324 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5325 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5326 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5328 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5331 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5332 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5333 case ISD::SRL_PARTS:
5334 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5335 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5336 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5337 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5338 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5339 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5340 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5341 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5342 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5343 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5344 case ISD::MUL: return LowerMUL(Op, DAG);
5345 case ISD::SDIV: return LowerSDIV(Op, DAG);
5346 case ISD::UDIV: return LowerUDIV(Op, DAG);
5350 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5351 case ISD::ATOMIC_LOAD:
5352 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5356 /// ReplaceNodeResults - Replace the results of node with an illegal result
5357 /// type with new values built out of custom code.
5358 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5359 SmallVectorImpl<SDValue>&Results,
5360 SelectionDAG &DAG) const {
5362 switch (N->getOpcode()) {
5364 llvm_unreachable("Don't know how to custom expand this!");
5366 Res = ExpandBITCAST(N, DAG);
5370 Res = Expand64BitShift(N, DAG, Subtarget);
5372 case ISD::ATOMIC_LOAD_ADD:
5373 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5375 case ISD::ATOMIC_LOAD_AND:
5376 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5378 case ISD::ATOMIC_LOAD_NAND:
5379 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5381 case ISD::ATOMIC_LOAD_OR:
5382 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5384 case ISD::ATOMIC_LOAD_SUB:
5385 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5387 case ISD::ATOMIC_LOAD_XOR:
5388 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5390 case ISD::ATOMIC_SWAP:
5391 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5393 case ISD::ATOMIC_CMP_SWAP:
5394 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5398 Results.push_back(Res);
5401 //===----------------------------------------------------------------------===//
5402 // ARM Scheduler Hooks
5403 //===----------------------------------------------------------------------===//
5406 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5407 MachineBasicBlock *BB,
5408 unsigned Size) const {
5409 unsigned dest = MI->getOperand(0).getReg();
5410 unsigned ptr = MI->getOperand(1).getReg();
5411 unsigned oldval = MI->getOperand(2).getReg();
5412 unsigned newval = MI->getOperand(3).getReg();
5413 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5414 DebugLoc dl = MI->getDebugLoc();
5415 bool isThumb2 = Subtarget->isThumb2();
5417 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5418 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5419 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5420 (const TargetRegisterClass*)&ARM::GPRRegClass);
5423 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5424 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5425 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5428 unsigned ldrOpc, strOpc;
5430 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5432 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5433 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5436 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5437 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5440 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5441 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5445 MachineFunction *MF = BB->getParent();
5446 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5447 MachineFunction::iterator It = BB;
5448 ++It; // insert the new blocks after the current block
5450 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5451 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5452 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5453 MF->insert(It, loop1MBB);
5454 MF->insert(It, loop2MBB);
5455 MF->insert(It, exitMBB);
5457 // Transfer the remainder of BB and its successor edges to exitMBB.
5458 exitMBB->splice(exitMBB->begin(), BB,
5459 llvm::next(MachineBasicBlock::iterator(MI)),
5461 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5465 // fallthrough --> loop1MBB
5466 BB->addSuccessor(loop1MBB);
5469 // ldrex dest, [ptr]
5473 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5474 if (ldrOpc == ARM::t2LDREX)
5476 AddDefaultPred(MIB);
5477 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5478 .addReg(dest).addReg(oldval));
5479 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5480 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5481 BB->addSuccessor(loop2MBB);
5482 BB->addSuccessor(exitMBB);
5485 // strex scratch, newval, [ptr]
5489 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5490 if (strOpc == ARM::t2STREX)
5492 AddDefaultPred(MIB);
5493 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5494 .addReg(scratch).addImm(0));
5495 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5496 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5497 BB->addSuccessor(loop1MBB);
5498 BB->addSuccessor(exitMBB);
5504 MI->eraseFromParent(); // The instruction is gone now.
5510 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5511 unsigned Size, unsigned BinOpcode) const {
5512 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5513 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5515 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5516 MachineFunction *MF = BB->getParent();
5517 MachineFunction::iterator It = BB;
5520 unsigned dest = MI->getOperand(0).getReg();
5521 unsigned ptr = MI->getOperand(1).getReg();
5522 unsigned incr = MI->getOperand(2).getReg();
5523 DebugLoc dl = MI->getDebugLoc();
5524 bool isThumb2 = Subtarget->isThumb2();
5526 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5528 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5529 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5532 unsigned ldrOpc, strOpc;
5534 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5536 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5537 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5540 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5541 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5544 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5545 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5549 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5550 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5551 MF->insert(It, loopMBB);
5552 MF->insert(It, exitMBB);
5554 // Transfer the remainder of BB and its successor edges to exitMBB.
5555 exitMBB->splice(exitMBB->begin(), BB,
5556 llvm::next(MachineBasicBlock::iterator(MI)),
5558 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5560 const TargetRegisterClass *TRC = isThumb2 ?
5561 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5562 (const TargetRegisterClass*)&ARM::GPRRegClass;
5563 unsigned scratch = MRI.createVirtualRegister(TRC);
5564 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5568 // fallthrough --> loopMBB
5569 BB->addSuccessor(loopMBB);
5573 // <binop> scratch2, dest, incr
5574 // strex scratch, scratch2, ptr
5577 // fallthrough --> exitMBB
5579 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5580 if (ldrOpc == ARM::t2LDREX)
5582 AddDefaultPred(MIB);
5584 // operand order needs to go the other way for NAND
5585 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5586 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5587 addReg(incr).addReg(dest)).addReg(0);
5589 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5590 addReg(dest).addReg(incr)).addReg(0);
5593 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5594 if (strOpc == ARM::t2STREX)
5596 AddDefaultPred(MIB);
5597 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5598 .addReg(scratch).addImm(0));
5599 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5600 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5602 BB->addSuccessor(loopMBB);
5603 BB->addSuccessor(exitMBB);
5609 MI->eraseFromParent(); // The instruction is gone now.
5615 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5616 MachineBasicBlock *BB,
5619 ARMCC::CondCodes Cond) const {
5620 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5622 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5623 MachineFunction *MF = BB->getParent();
5624 MachineFunction::iterator It = BB;
5627 unsigned dest = MI->getOperand(0).getReg();
5628 unsigned ptr = MI->getOperand(1).getReg();
5629 unsigned incr = MI->getOperand(2).getReg();
5630 unsigned oldval = dest;
5631 DebugLoc dl = MI->getDebugLoc();
5632 bool isThumb2 = Subtarget->isThumb2();
5634 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5636 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5637 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5640 unsigned ldrOpc, strOpc, extendOpc;
5642 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5644 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5645 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5646 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5649 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5650 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5651 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5654 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5655 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5660 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5661 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5662 MF->insert(It, loopMBB);
5663 MF->insert(It, exitMBB);
5665 // Transfer the remainder of BB and its successor edges to exitMBB.
5666 exitMBB->splice(exitMBB->begin(), BB,
5667 llvm::next(MachineBasicBlock::iterator(MI)),
5669 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5671 const TargetRegisterClass *TRC = isThumb2 ?
5672 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5673 (const TargetRegisterClass*)&ARM::GPRRegClass;
5674 unsigned scratch = MRI.createVirtualRegister(TRC);
5675 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5679 // fallthrough --> loopMBB
5680 BB->addSuccessor(loopMBB);
5684 // (sign extend dest, if required)
5686 // cmov.cond scratch2, incr, dest
5687 // strex scratch, scratch2, ptr
5690 // fallthrough --> exitMBB
5692 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5693 if (ldrOpc == ARM::t2LDREX)
5695 AddDefaultPred(MIB);
5697 // Sign extend the value, if necessary.
5698 if (signExtend && extendOpc) {
5699 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5700 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5705 // Build compare and cmov instructions.
5706 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5707 .addReg(oldval).addReg(incr));
5708 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5709 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5711 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5712 if (strOpc == ARM::t2STREX)
5714 AddDefaultPred(MIB);
5715 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5716 .addReg(scratch).addImm(0));
5717 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5718 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5720 BB->addSuccessor(loopMBB);
5721 BB->addSuccessor(exitMBB);
5727 MI->eraseFromParent(); // The instruction is gone now.
5733 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5734 unsigned Op1, unsigned Op2,
5735 bool NeedsCarry, bool IsCmpxchg) const {
5736 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5737 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5739 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5740 MachineFunction *MF = BB->getParent();
5741 MachineFunction::iterator It = BB;
5744 unsigned destlo = MI->getOperand(0).getReg();
5745 unsigned desthi = MI->getOperand(1).getReg();
5746 unsigned ptr = MI->getOperand(2).getReg();
5747 unsigned vallo = MI->getOperand(3).getReg();
5748 unsigned valhi = MI->getOperand(4).getReg();
5749 DebugLoc dl = MI->getDebugLoc();
5750 bool isThumb2 = Subtarget->isThumb2();
5752 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5754 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5755 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5756 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5759 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5760 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5762 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5763 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5765 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5766 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5768 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5769 MF->insert(It, loopMBB);
5771 MF->insert(It, contBB);
5772 MF->insert(It, cont2BB);
5774 MF->insert(It, exitMBB);
5776 // Transfer the remainder of BB and its successor edges to exitMBB.
5777 exitMBB->splice(exitMBB->begin(), BB,
5778 llvm::next(MachineBasicBlock::iterator(MI)),
5780 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5782 const TargetRegisterClass *TRC = isThumb2 ?
5783 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5784 (const TargetRegisterClass*)&ARM::GPRRegClass;
5785 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5789 // fallthrough --> loopMBB
5790 BB->addSuccessor(loopMBB);
5793 // ldrexd r2, r3, ptr
5794 // <binopa> r0, r2, incr
5795 // <binopb> r1, r3, incr
5796 // strexd storesuccess, r0, r1, ptr
5797 // cmp storesuccess, #0
5799 // fallthrough --> exitMBB
5801 // Note that the registers are explicitly specified because there is not any
5802 // way to force the register allocator to allocate a register pair.
5804 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5805 // need to properly enforce the restriction that the two output registers
5806 // for ldrexd must be different.
5809 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5810 .addReg(ARM::R2, RegState::Define)
5811 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5812 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5813 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5814 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5818 for (unsigned i = 0; i < 2; i++) {
5819 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5821 .addReg(i == 0 ? destlo : desthi)
5822 .addReg(i == 0 ? vallo : valhi));
5823 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5824 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5825 BB->addSuccessor(exitMBB);
5826 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5827 BB = (i == 0 ? contBB : cont2BB);
5830 // Copy to physregs for strexd
5831 unsigned setlo = MI->getOperand(5).getReg();
5832 unsigned sethi = MI->getOperand(6).getReg();
5833 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5834 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5836 // Perform binary operation
5837 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5838 .addReg(destlo).addReg(vallo))
5839 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5840 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5841 .addReg(desthi).addReg(valhi)).addReg(0);
5843 // Copy to physregs for strexd
5844 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5845 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5849 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5850 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5852 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5853 .addReg(storesuccess).addImm(0));
5854 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5855 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5857 BB->addSuccessor(loopMBB);
5858 BB->addSuccessor(exitMBB);
5864 MI->eraseFromParent(); // The instruction is gone now.
5869 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5870 /// registers the function context.
5871 void ARMTargetLowering::
5872 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5873 MachineBasicBlock *DispatchBB, int FI) const {
5874 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5875 DebugLoc dl = MI->getDebugLoc();
5876 MachineFunction *MF = MBB->getParent();
5877 MachineRegisterInfo *MRI = &MF->getRegInfo();
5878 MachineConstantPool *MCP = MF->getConstantPool();
5879 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5880 const Function *F = MF->getFunction();
5882 bool isThumb = Subtarget->isThumb();
5883 bool isThumb2 = Subtarget->isThumb2();
5885 unsigned PCLabelId = AFI->createPICLabelUId();
5886 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5887 ARMConstantPoolValue *CPV =
5888 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5889 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5891 const TargetRegisterClass *TRC = isThumb ?
5892 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5893 (const TargetRegisterClass*)&ARM::GPRRegClass;
5895 // Grab constant pool and fixed stack memory operands.
5896 MachineMemOperand *CPMMO =
5897 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5898 MachineMemOperand::MOLoad, 4, 4);
5900 MachineMemOperand *FIMMOSt =
5901 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5902 MachineMemOperand::MOStore, 4, 4);
5904 // Load the address of the dispatch MBB into the jump buffer.
5906 // Incoming value: jbuf
5907 // ldr.n r5, LCPI1_1
5910 // str r5, [$jbuf, #+4] ; &jbuf[1]
5911 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5912 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5913 .addConstantPoolIndex(CPI)
5914 .addMemOperand(CPMMO));
5915 // Set the low bit because of thumb mode.
5916 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5918 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5919 .addReg(NewVReg1, RegState::Kill)
5921 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5922 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5923 .addReg(NewVReg2, RegState::Kill)
5925 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5926 .addReg(NewVReg3, RegState::Kill)
5928 .addImm(36) // &jbuf[1] :: pc
5929 .addMemOperand(FIMMOSt));
5930 } else if (isThumb) {
5931 // Incoming value: jbuf
5932 // ldr.n r1, LCPI1_4
5936 // add r2, $jbuf, #+4 ; &jbuf[1]
5938 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5939 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5940 .addConstantPoolIndex(CPI)
5941 .addMemOperand(CPMMO));
5942 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5943 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5944 .addReg(NewVReg1, RegState::Kill)
5946 // Set the low bit because of thumb mode.
5947 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5948 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5949 .addReg(ARM::CPSR, RegState::Define)
5951 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5952 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5953 .addReg(ARM::CPSR, RegState::Define)
5954 .addReg(NewVReg2, RegState::Kill)
5955 .addReg(NewVReg3, RegState::Kill));
5956 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5957 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5959 .addImm(36)); // &jbuf[1] :: pc
5960 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5961 .addReg(NewVReg4, RegState::Kill)
5962 .addReg(NewVReg5, RegState::Kill)
5964 .addMemOperand(FIMMOSt));
5966 // Incoming value: jbuf
5969 // str r1, [$jbuf, #+4] ; &jbuf[1]
5970 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5971 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5972 .addConstantPoolIndex(CPI)
5974 .addMemOperand(CPMMO));
5975 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5976 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5977 .addReg(NewVReg1, RegState::Kill)
5978 .addImm(PCLabelId));
5979 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5980 .addReg(NewVReg2, RegState::Kill)
5982 .addImm(36) // &jbuf[1] :: pc
5983 .addMemOperand(FIMMOSt));
5987 MachineBasicBlock *ARMTargetLowering::
5988 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5989 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5990 DebugLoc dl = MI->getDebugLoc();
5991 MachineFunction *MF = MBB->getParent();
5992 MachineRegisterInfo *MRI = &MF->getRegInfo();
5993 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5994 MachineFrameInfo *MFI = MF->getFrameInfo();
5995 int FI = MFI->getFunctionContextIndex();
5997 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5998 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5999 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6001 // Get a mapping of the call site numbers to all of the landing pads they're
6003 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6004 unsigned MaxCSNum = 0;
6005 MachineModuleInfo &MMI = MF->getMMI();
6006 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6008 if (!BB->isLandingPad()) continue;
6010 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6012 for (MachineBasicBlock::iterator
6013 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6014 if (!II->isEHLabel()) continue;
6016 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6017 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6019 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6020 for (SmallVectorImpl<unsigned>::iterator
6021 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6022 CSI != CSE; ++CSI) {
6023 CallSiteNumToLPad[*CSI].push_back(BB);
6024 MaxCSNum = std::max(MaxCSNum, *CSI);
6030 // Get an ordered list of the machine basic blocks for the jump table.
6031 std::vector<MachineBasicBlock*> LPadList;
6032 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6033 LPadList.reserve(CallSiteNumToLPad.size());
6034 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6035 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6036 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6037 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6038 LPadList.push_back(*II);
6039 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6043 assert(!LPadList.empty() &&
6044 "No landing pad destinations for the dispatch jump table!");
6046 // Create the jump table and associated information.
6047 MachineJumpTableInfo *JTI =
6048 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6049 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6050 unsigned UId = AFI->createJumpTableUId();
6052 // Create the MBBs for the dispatch code.
6054 // Shove the dispatch's address into the return slot in the function context.
6055 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6056 DispatchBB->setIsLandingPad();
6058 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6059 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6060 DispatchBB->addSuccessor(TrapBB);
6062 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6063 DispatchBB->addSuccessor(DispContBB);
6066 MF->insert(MF->end(), DispatchBB);
6067 MF->insert(MF->end(), DispContBB);
6068 MF->insert(MF->end(), TrapBB);
6070 // Insert code into the entry block that creates and registers the function
6072 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6074 MachineMemOperand *FIMMOLd =
6075 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6076 MachineMemOperand::MOLoad |
6077 MachineMemOperand::MOVolatile, 4, 4);
6079 MachineInstrBuilder MIB;
6080 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6082 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6083 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6085 // Add a register mask with no preserved registers. This results in all
6086 // registers being marked as clobbered.
6087 MIB.addRegMask(RI.getNoPreservedMask());
6089 unsigned NumLPads = LPadList.size();
6090 if (Subtarget->isThumb2()) {
6091 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6092 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6095 .addMemOperand(FIMMOLd));
6097 if (NumLPads < 256) {
6098 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6100 .addImm(LPadList.size()));
6102 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6103 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6104 .addImm(NumLPads & 0xFFFF));
6106 unsigned VReg2 = VReg1;
6107 if ((NumLPads & 0xFFFF0000) != 0) {
6108 VReg2 = MRI->createVirtualRegister(TRC);
6109 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6111 .addImm(NumLPads >> 16));
6114 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6119 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6124 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6125 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6126 .addJumpTableIndex(MJTI)
6129 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6132 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6133 .addReg(NewVReg3, RegState::Kill)
6135 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6137 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6138 .addReg(NewVReg4, RegState::Kill)
6140 .addJumpTableIndex(MJTI)
6142 } else if (Subtarget->isThumb()) {
6143 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6144 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6147 .addMemOperand(FIMMOLd));
6149 if (NumLPads < 256) {
6150 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6154 MachineConstantPool *ConstantPool = MF->getConstantPool();
6155 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6156 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6158 // MachineConstantPool wants an explicit alignment.
6159 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6161 Align = getDataLayout()->getTypeAllocSize(C->getType());
6162 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6164 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6165 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6166 .addReg(VReg1, RegState::Define)
6167 .addConstantPoolIndex(Idx));
6168 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6173 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6178 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6179 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6180 .addReg(ARM::CPSR, RegState::Define)
6184 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6185 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6186 .addJumpTableIndex(MJTI)
6189 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6190 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6191 .addReg(ARM::CPSR, RegState::Define)
6192 .addReg(NewVReg2, RegState::Kill)
6195 MachineMemOperand *JTMMOLd =
6196 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6197 MachineMemOperand::MOLoad, 4, 4);
6199 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6200 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6201 .addReg(NewVReg4, RegState::Kill)
6203 .addMemOperand(JTMMOLd));
6205 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6206 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6207 .addReg(ARM::CPSR, RegState::Define)
6208 .addReg(NewVReg5, RegState::Kill)
6211 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6212 .addReg(NewVReg6, RegState::Kill)
6213 .addJumpTableIndex(MJTI)
6216 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6217 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6220 .addMemOperand(FIMMOLd));
6222 if (NumLPads < 256) {
6223 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6226 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6227 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6228 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6229 .addImm(NumLPads & 0xFFFF));
6231 unsigned VReg2 = VReg1;
6232 if ((NumLPads & 0xFFFF0000) != 0) {
6233 VReg2 = MRI->createVirtualRegister(TRC);
6234 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6236 .addImm(NumLPads >> 16));
6239 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6243 MachineConstantPool *ConstantPool = MF->getConstantPool();
6244 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6245 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6247 // MachineConstantPool wants an explicit alignment.
6248 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6250 Align = getDataLayout()->getTypeAllocSize(C->getType());
6251 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6253 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6254 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6255 .addReg(VReg1, RegState::Define)
6256 .addConstantPoolIndex(Idx)
6258 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6260 .addReg(VReg1, RegState::Kill));
6263 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6268 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6270 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6272 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6273 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6274 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6275 .addJumpTableIndex(MJTI)
6278 MachineMemOperand *JTMMOLd =
6279 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6280 MachineMemOperand::MOLoad, 4, 4);
6281 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6283 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6284 .addReg(NewVReg3, RegState::Kill)
6287 .addMemOperand(JTMMOLd));
6289 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6290 .addReg(NewVReg5, RegState::Kill)
6292 .addJumpTableIndex(MJTI)
6296 // Add the jump table entries as successors to the MBB.
6297 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6298 for (std::vector<MachineBasicBlock*>::iterator
6299 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6300 MachineBasicBlock *CurMBB = *I;
6301 if (SeenMBBs.insert(CurMBB))
6302 DispContBB->addSuccessor(CurMBB);
6305 // N.B. the order the invoke BBs are processed in doesn't matter here.
6306 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6307 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6308 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6309 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6310 MachineBasicBlock *BB = *I;
6312 // Remove the landing pad successor from the invoke block and replace it
6313 // with the new dispatch block.
6314 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6316 while (!Successors.empty()) {
6317 MachineBasicBlock *SMBB = Successors.pop_back_val();
6318 if (SMBB->isLandingPad()) {
6319 BB->removeSuccessor(SMBB);
6320 MBBLPads.push_back(SMBB);
6324 BB->addSuccessor(DispatchBB);
6326 // Find the invoke call and mark all of the callee-saved registers as
6327 // 'implicit defined' so that they're spilled. This prevents code from
6328 // moving instructions to before the EH block, where they will never be
6330 for (MachineBasicBlock::reverse_iterator
6331 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6332 if (!II->isCall()) continue;
6334 DenseMap<unsigned, bool> DefRegs;
6335 for (MachineInstr::mop_iterator
6336 OI = II->operands_begin(), OE = II->operands_end();
6338 if (!OI->isReg()) continue;
6339 DefRegs[OI->getReg()] = true;
6342 MachineInstrBuilder MIB(&*II);
6344 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6345 unsigned Reg = SavedRegs[i];
6346 if (Subtarget->isThumb2() &&
6347 !ARM::tGPRRegClass.contains(Reg) &&
6348 !ARM::hGPRRegClass.contains(Reg))
6350 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6352 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6355 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6362 // Mark all former landing pads as non-landing pads. The dispatch is the only
6364 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6365 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6366 (*I)->setIsLandingPad(false);
6368 // The instruction is gone now.
6369 MI->eraseFromParent();
6375 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6376 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6377 E = MBB->succ_end(); I != E; ++I)
6380 llvm_unreachable("Expecting a BB with two successors!");
6383 MachineBasicBlock *ARMTargetLowering::
6384 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6385 // This pseudo instruction has 3 operands: dst, src, size
6386 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6387 // Otherwise, we will generate unrolled scalar copies.
6388 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6389 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6390 MachineFunction::iterator It = BB;
6393 unsigned dest = MI->getOperand(0).getReg();
6394 unsigned src = MI->getOperand(1).getReg();
6395 unsigned SizeVal = MI->getOperand(2).getImm();
6396 unsigned Align = MI->getOperand(3).getImm();
6397 DebugLoc dl = MI->getDebugLoc();
6399 bool isThumb2 = Subtarget->isThumb2();
6400 MachineFunction *MF = BB->getParent();
6401 MachineRegisterInfo &MRI = MF->getRegInfo();
6402 unsigned ldrOpc, strOpc, UnitSize = 0;
6404 const TargetRegisterClass *TRC = isThumb2 ?
6405 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6406 (const TargetRegisterClass*)&ARM::GPRRegClass;
6407 const TargetRegisterClass *TRC_Vec = 0;
6410 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6411 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6413 } else if (Align & 2) {
6414 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6415 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6418 // Check whether we can use NEON instructions.
6419 if (!MF->getFunction()->getFnAttributes().
6420 hasAttribute(Attributes::NoImplicitFloat) &&
6421 Subtarget->hasNEON()) {
6422 if ((Align % 16 == 0) && SizeVal >= 16) {
6423 ldrOpc = ARM::VLD1q32wb_fixed;
6424 strOpc = ARM::VST1q32wb_fixed;
6426 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6428 else if ((Align % 8 == 0) && SizeVal >= 8) {
6429 ldrOpc = ARM::VLD1d32wb_fixed;
6430 strOpc = ARM::VST1d32wb_fixed;
6432 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6435 // Can't use NEON instructions.
6436 if (UnitSize == 0) {
6437 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6438 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6443 unsigned BytesLeft = SizeVal % UnitSize;
6444 unsigned LoopSize = SizeVal - BytesLeft;
6446 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6447 // Use LDR and STR to copy.
6448 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6449 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6450 unsigned srcIn = src;
6451 unsigned destIn = dest;
6452 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6453 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6454 unsigned srcOut = MRI.createVirtualRegister(TRC);
6455 unsigned destOut = MRI.createVirtualRegister(TRC);
6456 if (UnitSize >= 8) {
6457 AddDefaultPred(BuildMI(*BB, MI, dl,
6458 TII->get(ldrOpc), scratch)
6459 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6461 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6462 .addReg(destIn).addImm(0).addReg(scratch));
6463 } else if (isThumb2) {
6464 AddDefaultPred(BuildMI(*BB, MI, dl,
6465 TII->get(ldrOpc), scratch)
6466 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6468 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6469 .addReg(scratch).addReg(destIn)
6472 AddDefaultPred(BuildMI(*BB, MI, dl,
6473 TII->get(ldrOpc), scratch)
6474 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6477 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6478 .addReg(scratch).addReg(destIn)
6479 .addReg(0).addImm(UnitSize));
6485 // Handle the leftover bytes with LDRB and STRB.
6486 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6487 // [destOut] = STRB_POST(scratch, destIn, 1)
6488 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6489 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6490 for (unsigned i = 0; i < BytesLeft; i++) {
6491 unsigned scratch = MRI.createVirtualRegister(TRC);
6492 unsigned srcOut = MRI.createVirtualRegister(TRC);
6493 unsigned destOut = MRI.createVirtualRegister(TRC);
6495 AddDefaultPred(BuildMI(*BB, MI, dl,
6496 TII->get(ldrOpc),scratch)
6497 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6499 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6500 .addReg(scratch).addReg(destIn)
6501 .addReg(0).addImm(1));
6503 AddDefaultPred(BuildMI(*BB, MI, dl,
6504 TII->get(ldrOpc),scratch)
6505 .addReg(srcOut, RegState::Define).addReg(srcIn)
6506 .addReg(0).addImm(1));
6508 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6509 .addReg(scratch).addReg(destIn)
6510 .addReg(0).addImm(1));
6515 MI->eraseFromParent(); // The instruction is gone now.
6519 // Expand the pseudo op to a loop.
6522 // movw varEnd, # --> with thumb2
6524 // ldrcp varEnd, idx --> without thumb2
6525 // fallthrough --> loopMBB
6527 // PHI varPhi, varEnd, varLoop
6528 // PHI srcPhi, src, srcLoop
6529 // PHI destPhi, dst, destLoop
6530 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6531 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6532 // subs varLoop, varPhi, #UnitSize
6534 // fallthrough --> exitMBB
6536 // epilogue to handle left-over bytes
6537 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6538 // [destOut] = STRB_POST(scratch, destLoop, 1)
6539 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6540 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6541 MF->insert(It, loopMBB);
6542 MF->insert(It, exitMBB);
6544 // Transfer the remainder of BB and its successor edges to exitMBB.
6545 exitMBB->splice(exitMBB->begin(), BB,
6546 llvm::next(MachineBasicBlock::iterator(MI)),
6548 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6550 // Load an immediate to varEnd.
6551 unsigned varEnd = MRI.createVirtualRegister(TRC);
6553 unsigned VReg1 = varEnd;
6554 if ((LoopSize & 0xFFFF0000) != 0)
6555 VReg1 = MRI.createVirtualRegister(TRC);
6556 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6557 .addImm(LoopSize & 0xFFFF));
6559 if ((LoopSize & 0xFFFF0000) != 0)
6560 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6562 .addImm(LoopSize >> 16));
6564 MachineConstantPool *ConstantPool = MF->getConstantPool();
6565 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6566 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6568 // MachineConstantPool wants an explicit alignment.
6569 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6571 Align = getDataLayout()->getTypeAllocSize(C->getType());
6572 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6574 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6575 .addReg(varEnd, RegState::Define)
6576 .addConstantPoolIndex(Idx)
6579 BB->addSuccessor(loopMBB);
6581 // Generate the loop body:
6582 // varPhi = PHI(varLoop, varEnd)
6583 // srcPhi = PHI(srcLoop, src)
6584 // destPhi = PHI(destLoop, dst)
6585 MachineBasicBlock *entryBB = BB;
6587 unsigned varLoop = MRI.createVirtualRegister(TRC);
6588 unsigned varPhi = MRI.createVirtualRegister(TRC);
6589 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6590 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6591 unsigned destLoop = MRI.createVirtualRegister(TRC);
6592 unsigned destPhi = MRI.createVirtualRegister(TRC);
6594 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6595 .addReg(varLoop).addMBB(loopMBB)
6596 .addReg(varEnd).addMBB(entryBB);
6597 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6598 .addReg(srcLoop).addMBB(loopMBB)
6599 .addReg(src).addMBB(entryBB);
6600 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6601 .addReg(destLoop).addMBB(loopMBB)
6602 .addReg(dest).addMBB(entryBB);
6604 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6605 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6606 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6607 if (UnitSize >= 8) {
6608 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6609 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6611 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6612 .addReg(destPhi).addImm(0).addReg(scratch));
6613 } else if (isThumb2) {
6614 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6615 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6617 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6618 .addReg(scratch).addReg(destPhi)
6621 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6622 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6625 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6626 .addReg(scratch).addReg(destPhi)
6627 .addReg(0).addImm(UnitSize));
6630 // Decrement loop variable by UnitSize.
6631 MachineInstrBuilder MIB = BuildMI(BB, dl,
6632 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6633 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6634 MIB->getOperand(5).setReg(ARM::CPSR);
6635 MIB->getOperand(5).setIsDef(true);
6637 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6638 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6640 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6641 BB->addSuccessor(loopMBB);
6642 BB->addSuccessor(exitMBB);
6644 // Add epilogue to handle BytesLeft.
6646 MachineInstr *StartOfExit = exitMBB->begin();
6647 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6648 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6650 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6651 // [destOut] = STRB_POST(scratch, destLoop, 1)
6652 unsigned srcIn = srcLoop;
6653 unsigned destIn = destLoop;
6654 for (unsigned i = 0; i < BytesLeft; i++) {
6655 unsigned scratch = MRI.createVirtualRegister(TRC);
6656 unsigned srcOut = MRI.createVirtualRegister(TRC);
6657 unsigned destOut = MRI.createVirtualRegister(TRC);
6659 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6660 TII->get(ldrOpc),scratch)
6661 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6663 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6664 .addReg(scratch).addReg(destIn)
6667 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6668 TII->get(ldrOpc),scratch)
6669 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6671 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6672 .addReg(scratch).addReg(destIn)
6673 .addReg(0).addImm(1));
6679 MI->eraseFromParent(); // The instruction is gone now.
6684 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6685 MachineBasicBlock *BB) const {
6686 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6687 DebugLoc dl = MI->getDebugLoc();
6688 bool isThumb2 = Subtarget->isThumb2();
6689 switch (MI->getOpcode()) {
6692 llvm_unreachable("Unexpected instr type to insert");
6694 // The Thumb2 pre-indexed stores have the same MI operands, they just
6695 // define them differently in the .td files from the isel patterns, so
6696 // they need pseudos.
6697 case ARM::t2STR_preidx:
6698 MI->setDesc(TII->get(ARM::t2STR_PRE));
6700 case ARM::t2STRB_preidx:
6701 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6703 case ARM::t2STRH_preidx:
6704 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6707 case ARM::STRi_preidx:
6708 case ARM::STRBi_preidx: {
6709 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6710 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6711 // Decode the offset.
6712 unsigned Offset = MI->getOperand(4).getImm();
6713 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6714 Offset = ARM_AM::getAM2Offset(Offset);
6718 MachineMemOperand *MMO = *MI->memoperands_begin();
6719 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6720 .addOperand(MI->getOperand(0)) // Rn_wb
6721 .addOperand(MI->getOperand(1)) // Rt
6722 .addOperand(MI->getOperand(2)) // Rn
6723 .addImm(Offset) // offset (skip GPR==zero_reg)
6724 .addOperand(MI->getOperand(5)) // pred
6725 .addOperand(MI->getOperand(6))
6726 .addMemOperand(MMO);
6727 MI->eraseFromParent();
6730 case ARM::STRr_preidx:
6731 case ARM::STRBr_preidx:
6732 case ARM::STRH_preidx: {
6734 switch (MI->getOpcode()) {
6735 default: llvm_unreachable("unexpected opcode!");
6736 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6737 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6738 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6740 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6741 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6742 MIB.addOperand(MI->getOperand(i));
6743 MI->eraseFromParent();
6746 case ARM::ATOMIC_LOAD_ADD_I8:
6747 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6748 case ARM::ATOMIC_LOAD_ADD_I16:
6749 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6750 case ARM::ATOMIC_LOAD_ADD_I32:
6751 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6753 case ARM::ATOMIC_LOAD_AND_I8:
6754 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6755 case ARM::ATOMIC_LOAD_AND_I16:
6756 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6757 case ARM::ATOMIC_LOAD_AND_I32:
6758 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6760 case ARM::ATOMIC_LOAD_OR_I8:
6761 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6762 case ARM::ATOMIC_LOAD_OR_I16:
6763 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6764 case ARM::ATOMIC_LOAD_OR_I32:
6765 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6767 case ARM::ATOMIC_LOAD_XOR_I8:
6768 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6769 case ARM::ATOMIC_LOAD_XOR_I16:
6770 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6771 case ARM::ATOMIC_LOAD_XOR_I32:
6772 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6774 case ARM::ATOMIC_LOAD_NAND_I8:
6775 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6776 case ARM::ATOMIC_LOAD_NAND_I16:
6777 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6778 case ARM::ATOMIC_LOAD_NAND_I32:
6779 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6781 case ARM::ATOMIC_LOAD_SUB_I8:
6782 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6783 case ARM::ATOMIC_LOAD_SUB_I16:
6784 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6785 case ARM::ATOMIC_LOAD_SUB_I32:
6786 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6788 case ARM::ATOMIC_LOAD_MIN_I8:
6789 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6790 case ARM::ATOMIC_LOAD_MIN_I16:
6791 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6792 case ARM::ATOMIC_LOAD_MIN_I32:
6793 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6795 case ARM::ATOMIC_LOAD_MAX_I8:
6796 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6797 case ARM::ATOMIC_LOAD_MAX_I16:
6798 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6799 case ARM::ATOMIC_LOAD_MAX_I32:
6800 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6802 case ARM::ATOMIC_LOAD_UMIN_I8:
6803 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6804 case ARM::ATOMIC_LOAD_UMIN_I16:
6805 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6806 case ARM::ATOMIC_LOAD_UMIN_I32:
6807 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6809 case ARM::ATOMIC_LOAD_UMAX_I8:
6810 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6811 case ARM::ATOMIC_LOAD_UMAX_I16:
6812 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6813 case ARM::ATOMIC_LOAD_UMAX_I32:
6814 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6816 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6817 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6818 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6820 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6821 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6822 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6825 case ARM::ATOMADD6432:
6826 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6827 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6828 /*NeedsCarry*/ true);
6829 case ARM::ATOMSUB6432:
6830 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6831 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6832 /*NeedsCarry*/ true);
6833 case ARM::ATOMOR6432:
6834 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6835 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6836 case ARM::ATOMXOR6432:
6837 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6838 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6839 case ARM::ATOMAND6432:
6840 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6841 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6842 case ARM::ATOMSWAP6432:
6843 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6844 case ARM::ATOMCMPXCHG6432:
6845 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6846 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6847 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6849 case ARM::tMOVCCr_pseudo: {
6850 // To "insert" a SELECT_CC instruction, we actually have to insert the
6851 // diamond control-flow pattern. The incoming instruction knows the
6852 // destination vreg to set, the condition code register to branch on, the
6853 // true/false values to select between, and a branch opcode to use.
6854 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6855 MachineFunction::iterator It = BB;
6861 // cmpTY ccX, r1, r2
6863 // fallthrough --> copy0MBB
6864 MachineBasicBlock *thisMBB = BB;
6865 MachineFunction *F = BB->getParent();
6866 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6867 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6868 F->insert(It, copy0MBB);
6869 F->insert(It, sinkMBB);
6871 // Transfer the remainder of BB and its successor edges to sinkMBB.
6872 sinkMBB->splice(sinkMBB->begin(), BB,
6873 llvm::next(MachineBasicBlock::iterator(MI)),
6875 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6877 BB->addSuccessor(copy0MBB);
6878 BB->addSuccessor(sinkMBB);
6880 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6881 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6884 // %FalseValue = ...
6885 // # fallthrough to sinkMBB
6888 // Update machine-CFG edges
6889 BB->addSuccessor(sinkMBB);
6892 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6895 BuildMI(*BB, BB->begin(), dl,
6896 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6897 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6898 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6900 MI->eraseFromParent(); // The pseudo instruction is gone now.
6905 case ARM::BCCZi64: {
6906 // If there is an unconditional branch to the other successor, remove it.
6907 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6909 // Compare both parts that make up the double comparison separately for
6911 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6913 unsigned LHS1 = MI->getOperand(1).getReg();
6914 unsigned LHS2 = MI->getOperand(2).getReg();
6916 AddDefaultPred(BuildMI(BB, dl,
6917 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6918 .addReg(LHS1).addImm(0));
6919 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6920 .addReg(LHS2).addImm(0)
6921 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6923 unsigned RHS1 = MI->getOperand(3).getReg();
6924 unsigned RHS2 = MI->getOperand(4).getReg();
6925 AddDefaultPred(BuildMI(BB, dl,
6926 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6927 .addReg(LHS1).addReg(RHS1));
6928 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6929 .addReg(LHS2).addReg(RHS2)
6930 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6933 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6934 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6935 if (MI->getOperand(0).getImm() == ARMCC::NE)
6936 std::swap(destMBB, exitMBB);
6938 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6939 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6941 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6943 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6945 MI->eraseFromParent(); // The pseudo instruction is gone now.
6949 case ARM::Int_eh_sjlj_setjmp:
6950 case ARM::Int_eh_sjlj_setjmp_nofp:
6951 case ARM::tInt_eh_sjlj_setjmp:
6952 case ARM::t2Int_eh_sjlj_setjmp:
6953 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6954 EmitSjLjDispatchBlock(MI, BB);
6959 // To insert an ABS instruction, we have to insert the
6960 // diamond control-flow pattern. The incoming instruction knows the
6961 // source vreg to test against 0, the destination vreg to set,
6962 // the condition code register to branch on, the
6963 // true/false values to select between, and a branch opcode to use.
6968 // BCC (branch to SinkBB if V0 >= 0)
6969 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6970 // SinkBB: V1 = PHI(V2, V3)
6971 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6972 MachineFunction::iterator BBI = BB;
6974 MachineFunction *Fn = BB->getParent();
6975 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6976 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6977 Fn->insert(BBI, RSBBB);
6978 Fn->insert(BBI, SinkBB);
6980 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6981 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6982 bool isThumb2 = Subtarget->isThumb2();
6983 MachineRegisterInfo &MRI = Fn->getRegInfo();
6984 // In Thumb mode S must not be specified if source register is the SP or
6985 // PC and if destination register is the SP, so restrict register class
6986 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6987 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6988 (const TargetRegisterClass*)&ARM::GPRRegClass);
6990 // Transfer the remainder of BB and its successor edges to sinkMBB.
6991 SinkBB->splice(SinkBB->begin(), BB,
6992 llvm::next(MachineBasicBlock::iterator(MI)),
6994 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6996 BB->addSuccessor(RSBBB);
6997 BB->addSuccessor(SinkBB);
6999 // fall through to SinkMBB
7000 RSBBB->addSuccessor(SinkBB);
7002 // insert a cmp at the end of BB
7003 AddDefaultPred(BuildMI(BB, dl,
7004 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7005 .addReg(ABSSrcReg).addImm(0));
7007 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7009 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7010 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7012 // insert rsbri in RSBBB
7013 // Note: BCC and rsbri will be converted into predicated rsbmi
7014 // by if-conversion pass
7015 BuildMI(*RSBBB, RSBBB->begin(), dl,
7016 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7017 .addReg(ABSSrcReg, RegState::Kill)
7018 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7020 // insert PHI in SinkBB,
7021 // reuse ABSDstReg to not change uses of ABS instruction
7022 BuildMI(*SinkBB, SinkBB->begin(), dl,
7023 TII->get(ARM::PHI), ABSDstReg)
7024 .addReg(NewRsbDstReg).addMBB(RSBBB)
7025 .addReg(ABSSrcReg).addMBB(BB);
7027 // remove ABS instruction
7028 MI->eraseFromParent();
7030 // return last added BB
7033 case ARM::COPY_STRUCT_BYVAL_I32:
7035 return EmitStructByval(MI, BB);
7039 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7040 SDNode *Node) const {
7041 if (!MI->hasPostISelHook()) {
7042 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7043 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7047 const MCInstrDesc *MCID = &MI->getDesc();
7048 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7049 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7050 // operand is still set to noreg. If needed, set the optional operand's
7051 // register to CPSR, and remove the redundant implicit def.
7053 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7055 // Rename pseudo opcodes.
7056 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7058 const ARMBaseInstrInfo *TII =
7059 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7060 MCID = &TII->get(NewOpc);
7062 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7063 "converted opcode should be the same except for cc_out");
7067 // Add the optional cc_out operand
7068 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7070 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7072 // Any ARM instruction that sets the 's' bit should specify an optional
7073 // "cc_out" operand in the last operand position.
7074 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7075 assert(!NewOpc && "Optional cc_out operand required");
7078 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7079 // since we already have an optional CPSR def.
7080 bool definesCPSR = false;
7081 bool deadCPSR = false;
7082 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7084 const MachineOperand &MO = MI->getOperand(i);
7085 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7089 MI->RemoveOperand(i);
7094 assert(!NewOpc && "Optional cc_out operand required");
7097 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7099 assert(!MI->getOperand(ccOutIdx).getReg() &&
7100 "expect uninitialized optional cc_out operand");
7104 // If this instruction was defined with an optional CPSR def and its dag node
7105 // had a live implicit CPSR def, then activate the optional CPSR def.
7106 MachineOperand &MO = MI->getOperand(ccOutIdx);
7107 MO.setReg(ARM::CPSR);
7111 //===----------------------------------------------------------------------===//
7112 // ARM Optimization Hooks
7113 //===----------------------------------------------------------------------===//
7115 // Helper function that checks if N is a null or all ones constant.
7116 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7117 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7120 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7123 // Return true if N is conditionally 0 or all ones.
7124 // Detects these expressions where cc is an i1 value:
7126 // (select cc 0, y) [AllOnes=0]
7127 // (select cc y, 0) [AllOnes=0]
7128 // (zext cc) [AllOnes=0]
7129 // (sext cc) [AllOnes=0/1]
7130 // (select cc -1, y) [AllOnes=1]
7131 // (select cc y, -1) [AllOnes=1]
7133 // Invert is set when N is the null/all ones constant when CC is false.
7134 // OtherOp is set to the alternative value of N.
7135 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7136 SDValue &CC, bool &Invert,
7138 SelectionDAG &DAG) {
7139 switch (N->getOpcode()) {
7140 default: return false;
7142 CC = N->getOperand(0);
7143 SDValue N1 = N->getOperand(1);
7144 SDValue N2 = N->getOperand(2);
7145 if (isZeroOrAllOnes(N1, AllOnes)) {
7150 if (isZeroOrAllOnes(N2, AllOnes)) {
7157 case ISD::ZERO_EXTEND:
7158 // (zext cc) can never be the all ones value.
7162 case ISD::SIGN_EXTEND: {
7163 EVT VT = N->getValueType(0);
7164 CC = N->getOperand(0);
7165 if (CC.getValueType() != MVT::i1)
7169 // When looking for an AllOnes constant, N is an sext, and the 'other'
7171 OtherOp = DAG.getConstant(0, VT);
7172 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7173 // When looking for a 0 constant, N can be zext or sext.
7174 OtherOp = DAG.getConstant(1, VT);
7176 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7182 // Combine a constant select operand into its use:
7184 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7185 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7186 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7187 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7188 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7190 // The transform is rejected if the select doesn't have a constant operand that
7191 // is null, or all ones when AllOnes is set.
7193 // Also recognize sext/zext from i1:
7195 // (add (zext cc), x) -> (select cc (add x, 1), x)
7196 // (add (sext cc), x) -> (select cc (add x, -1), x)
7198 // These transformations eventually create predicated instructions.
7200 // @param N The node to transform.
7201 // @param Slct The N operand that is a select.
7202 // @param OtherOp The other N operand (x above).
7203 // @param DCI Context.
7204 // @param AllOnes Require the select constant to be all ones instead of null.
7205 // @returns The new node, or SDValue() on failure.
7207 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7208 TargetLowering::DAGCombinerInfo &DCI,
7209 bool AllOnes = false) {
7210 SelectionDAG &DAG = DCI.DAG;
7211 EVT VT = N->getValueType(0);
7212 SDValue NonConstantVal;
7215 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7216 NonConstantVal, DAG))
7219 // Slct is now know to be the desired identity constant when CC is true.
7220 SDValue TrueVal = OtherOp;
7221 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7222 OtherOp, NonConstantVal);
7223 // Unless SwapSelectOps says CC should be false.
7225 std::swap(TrueVal, FalseVal);
7227 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7228 CCOp, TrueVal, FalseVal);
7231 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7233 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7234 TargetLowering::DAGCombinerInfo &DCI) {
7235 SDValue N0 = N->getOperand(0);
7236 SDValue N1 = N->getOperand(1);
7237 if (N0.getNode()->hasOneUse()) {
7238 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7239 if (Result.getNode())
7242 if (N1.getNode()->hasOneUse()) {
7243 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7244 if (Result.getNode())
7250 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7251 // (only after legalization).
7252 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7253 TargetLowering::DAGCombinerInfo &DCI,
7254 const ARMSubtarget *Subtarget) {
7256 // Only perform optimization if after legalize, and if NEON is available. We
7257 // also expected both operands to be BUILD_VECTORs.
7258 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7259 || N0.getOpcode() != ISD::BUILD_VECTOR
7260 || N1.getOpcode() != ISD::BUILD_VECTOR)
7263 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7264 EVT VT = N->getValueType(0);
7265 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7268 // Check that the vector operands are of the right form.
7269 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7270 // operands, where N is the size of the formed vector.
7271 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7272 // index such that we have a pair wise add pattern.
7274 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7275 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7277 SDValue Vec = N0->getOperand(0)->getOperand(0);
7278 SDNode *V = Vec.getNode();
7279 unsigned nextIndex = 0;
7281 // For each operands to the ADD which are BUILD_VECTORs,
7282 // check to see if each of their operands are an EXTRACT_VECTOR with
7283 // the same vector and appropriate index.
7284 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7285 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7286 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7288 SDValue ExtVec0 = N0->getOperand(i);
7289 SDValue ExtVec1 = N1->getOperand(i);
7291 // First operand is the vector, verify its the same.
7292 if (V != ExtVec0->getOperand(0).getNode() ||
7293 V != ExtVec1->getOperand(0).getNode())
7296 // Second is the constant, verify its correct.
7297 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7298 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7300 // For the constant, we want to see all the even or all the odd.
7301 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7302 || C1->getZExtValue() != nextIndex+1)
7311 // Create VPADDL node.
7312 SelectionDAG &DAG = DCI.DAG;
7313 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7315 // Build operand list.
7316 SmallVector<SDValue, 8> Ops;
7317 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7318 TLI.getPointerTy()));
7320 // Input is the vector.
7323 // Get widened type and narrowed type.
7325 unsigned numElem = VT.getVectorNumElements();
7326 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7327 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7328 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7329 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7331 llvm_unreachable("Invalid vector element type for padd optimization.");
7334 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7335 widenType, &Ops[0], Ops.size());
7336 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7339 static SDValue findMUL_LOHI(SDValue V) {
7340 if (V->getOpcode() == ISD::UMUL_LOHI ||
7341 V->getOpcode() == ISD::SMUL_LOHI)
7346 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7347 TargetLowering::DAGCombinerInfo &DCI,
7348 const ARMSubtarget *Subtarget) {
7350 if (Subtarget->isThumb1Only()) return SDValue();
7352 // Only perform the checks after legalize when the pattern is available.
7353 if (DCI.isBeforeLegalize()) return SDValue();
7355 // Look for multiply add opportunities.
7356 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7357 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7358 // a glue link from the first add to the second add.
7359 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7360 // a S/UMLAL instruction.
7363 // \ / \ [no multiline comment]
7369 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7370 SDValue AddcOp0 = AddcNode->getOperand(0);
7371 SDValue AddcOp1 = AddcNode->getOperand(1);
7373 // Check if the two operands are from the same mul_lohi node.
7374 if (AddcOp0.getNode() == AddcOp1.getNode())
7377 assert(AddcNode->getNumValues() == 2 &&
7378 AddcNode->getValueType(0) == MVT::i32 &&
7379 AddcNode->getValueType(1) == MVT::Glue &&
7380 "Expect ADDC with two result values: i32, glue");
7382 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7383 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7384 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7385 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7386 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7389 // Look for the glued ADDE.
7390 SDNode* AddeNode = AddcNode->getGluedUser();
7391 if (AddeNode == NULL)
7394 // Make sure it is really an ADDE.
7395 if (AddeNode->getOpcode() != ISD::ADDE)
7398 assert(AddeNode->getNumOperands() == 3 &&
7399 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7400 "ADDE node has the wrong inputs");
7402 // Check for the triangle shape.
7403 SDValue AddeOp0 = AddeNode->getOperand(0);
7404 SDValue AddeOp1 = AddeNode->getOperand(1);
7406 // Make sure that the ADDE operands are not coming from the same node.
7407 if (AddeOp0.getNode() == AddeOp1.getNode())
7410 // Find the MUL_LOHI node walking up ADDE's operands.
7411 bool IsLeftOperandMUL = false;
7412 SDValue MULOp = findMUL_LOHI(AddeOp0);
7413 if (MULOp == SDValue())
7414 MULOp = findMUL_LOHI(AddeOp1);
7416 IsLeftOperandMUL = true;
7417 if (MULOp == SDValue())
7420 // Figure out the right opcode.
7421 unsigned Opc = MULOp->getOpcode();
7422 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7424 // Figure out the high and low input values to the MLAL node.
7425 SDValue* HiMul = &MULOp;
7426 SDValue* HiAdd = NULL;
7427 SDValue* LoMul = NULL;
7428 SDValue* LowAdd = NULL;
7430 if (IsLeftOperandMUL)
7436 if (AddcOp0->getOpcode() == Opc) {
7440 if (AddcOp1->getOpcode() == Opc) {
7448 if (LoMul->getNode() != HiMul->getNode())
7451 // Create the merged node.
7452 SelectionDAG &DAG = DCI.DAG;
7454 // Build operand list.
7455 SmallVector<SDValue, 8> Ops;
7456 Ops.push_back(LoMul->getOperand(0));
7457 Ops.push_back(LoMul->getOperand(1));
7458 Ops.push_back(*LowAdd);
7459 Ops.push_back(*HiAdd);
7461 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7462 DAG.getVTList(MVT::i32, MVT::i32),
7463 &Ops[0], Ops.size());
7465 // Replace the ADDs' nodes uses by the MLA node's values.
7466 SDValue HiMLALResult(MLALNode.getNode(), 1);
7467 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7469 SDValue LoMLALResult(MLALNode.getNode(), 0);
7470 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7472 // Return original node to notify the driver to stop replacing.
7473 SDValue resNode(AddcNode, 0);
7477 /// PerformADDCCombine - Target-specific dag combine transform from
7478 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7479 static SDValue PerformADDCCombine(SDNode *N,
7480 TargetLowering::DAGCombinerInfo &DCI,
7481 const ARMSubtarget *Subtarget) {
7483 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7487 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7488 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7489 /// called with the default operands, and if that fails, with commuted
7491 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7492 TargetLowering::DAGCombinerInfo &DCI,
7493 const ARMSubtarget *Subtarget){
7495 // Attempt to create vpaddl for this add.
7496 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7497 if (Result.getNode())
7500 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7501 if (N0.getNode()->hasOneUse()) {
7502 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7503 if (Result.getNode()) return Result;
7508 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7510 static SDValue PerformADDCombine(SDNode *N,
7511 TargetLowering::DAGCombinerInfo &DCI,
7512 const ARMSubtarget *Subtarget) {
7513 SDValue N0 = N->getOperand(0);
7514 SDValue N1 = N->getOperand(1);
7516 // First try with the default operand order.
7517 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7518 if (Result.getNode())
7521 // If that didn't work, try again with the operands commuted.
7522 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7525 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7527 static SDValue PerformSUBCombine(SDNode *N,
7528 TargetLowering::DAGCombinerInfo &DCI) {
7529 SDValue N0 = N->getOperand(0);
7530 SDValue N1 = N->getOperand(1);
7532 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7533 if (N1.getNode()->hasOneUse()) {
7534 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7535 if (Result.getNode()) return Result;
7541 /// PerformVMULCombine
7542 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7543 /// special multiplier accumulator forwarding.
7549 static SDValue PerformVMULCombine(SDNode *N,
7550 TargetLowering::DAGCombinerInfo &DCI,
7551 const ARMSubtarget *Subtarget) {
7552 if (!Subtarget->hasVMLxForwarding())
7555 SelectionDAG &DAG = DCI.DAG;
7556 SDValue N0 = N->getOperand(0);
7557 SDValue N1 = N->getOperand(1);
7558 unsigned Opcode = N0.getOpcode();
7559 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7560 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7561 Opcode = N1.getOpcode();
7562 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7563 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7568 EVT VT = N->getValueType(0);
7569 DebugLoc DL = N->getDebugLoc();
7570 SDValue N00 = N0->getOperand(0);
7571 SDValue N01 = N0->getOperand(1);
7572 return DAG.getNode(Opcode, DL, VT,
7573 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7574 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7577 static SDValue PerformMULCombine(SDNode *N,
7578 TargetLowering::DAGCombinerInfo &DCI,
7579 const ARMSubtarget *Subtarget) {
7580 SelectionDAG &DAG = DCI.DAG;
7582 if (Subtarget->isThumb1Only())
7585 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7588 EVT VT = N->getValueType(0);
7589 if (VT.is64BitVector() || VT.is128BitVector())
7590 return PerformVMULCombine(N, DCI, Subtarget);
7594 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7598 int64_t MulAmt = C->getSExtValue();
7599 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7601 ShiftAmt = ShiftAmt & (32 - 1);
7602 SDValue V = N->getOperand(0);
7603 DebugLoc DL = N->getDebugLoc();
7606 MulAmt >>= ShiftAmt;
7609 if (isPowerOf2_32(MulAmt - 1)) {
7610 // (mul x, 2^N + 1) => (add (shl x, N), x)
7611 Res = DAG.getNode(ISD::ADD, DL, VT,
7613 DAG.getNode(ISD::SHL, DL, VT,
7615 DAG.getConstant(Log2_32(MulAmt - 1),
7617 } else if (isPowerOf2_32(MulAmt + 1)) {
7618 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7619 Res = DAG.getNode(ISD::SUB, DL, VT,
7620 DAG.getNode(ISD::SHL, DL, VT,
7622 DAG.getConstant(Log2_32(MulAmt + 1),
7628 uint64_t MulAmtAbs = -MulAmt;
7629 if (isPowerOf2_32(MulAmtAbs + 1)) {
7630 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7631 Res = DAG.getNode(ISD::SUB, DL, VT,
7633 DAG.getNode(ISD::SHL, DL, VT,
7635 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7637 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7638 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7639 Res = DAG.getNode(ISD::ADD, DL, VT,
7641 DAG.getNode(ISD::SHL, DL, VT,
7643 DAG.getConstant(Log2_32(MulAmtAbs-1),
7645 Res = DAG.getNode(ISD::SUB, DL, VT,
7646 DAG.getConstant(0, MVT::i32),Res);
7653 Res = DAG.getNode(ISD::SHL, DL, VT,
7654 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7656 // Do not add new nodes to DAG combiner worklist.
7657 DCI.CombineTo(N, Res, false);
7661 static SDValue PerformANDCombine(SDNode *N,
7662 TargetLowering::DAGCombinerInfo &DCI,
7663 const ARMSubtarget *Subtarget) {
7665 // Attempt to use immediate-form VBIC
7666 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7667 DebugLoc dl = N->getDebugLoc();
7668 EVT VT = N->getValueType(0);
7669 SelectionDAG &DAG = DCI.DAG;
7671 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7674 APInt SplatBits, SplatUndef;
7675 unsigned SplatBitSize;
7678 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7679 if (SplatBitSize <= 64) {
7681 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7682 SplatUndef.getZExtValue(), SplatBitSize,
7683 DAG, VbicVT, VT.is128BitVector(),
7685 if (Val.getNode()) {
7687 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7688 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7689 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7694 if (!Subtarget->isThumb1Only()) {
7695 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7696 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7697 if (Result.getNode())
7704 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7705 static SDValue PerformORCombine(SDNode *N,
7706 TargetLowering::DAGCombinerInfo &DCI,
7707 const ARMSubtarget *Subtarget) {
7708 // Attempt to use immediate-form VORR
7709 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7710 DebugLoc dl = N->getDebugLoc();
7711 EVT VT = N->getValueType(0);
7712 SelectionDAG &DAG = DCI.DAG;
7714 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7717 APInt SplatBits, SplatUndef;
7718 unsigned SplatBitSize;
7720 if (BVN && Subtarget->hasNEON() &&
7721 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7722 if (SplatBitSize <= 64) {
7724 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7725 SplatUndef.getZExtValue(), SplatBitSize,
7726 DAG, VorrVT, VT.is128BitVector(),
7728 if (Val.getNode()) {
7730 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7731 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7732 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7737 if (!Subtarget->isThumb1Only()) {
7738 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7739 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7740 if (Result.getNode())
7744 // The code below optimizes (or (and X, Y), Z).
7745 // The AND operand needs to have a single user to make these optimizations
7747 SDValue N0 = N->getOperand(0);
7748 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7750 SDValue N1 = N->getOperand(1);
7752 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7753 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7754 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7756 unsigned SplatBitSize;
7759 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7761 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7762 HasAnyUndefs) && !HasAnyUndefs) {
7763 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7765 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7766 HasAnyUndefs) && !HasAnyUndefs &&
7767 SplatBits0 == ~SplatBits1) {
7768 // Canonicalize the vector type to make instruction selection simpler.
7769 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7770 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7771 N0->getOperand(1), N0->getOperand(0),
7773 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7778 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7781 // BFI is only available on V6T2+
7782 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7785 DebugLoc DL = N->getDebugLoc();
7786 // 1) or (and A, mask), val => ARMbfi A, val, mask
7787 // iff (val & mask) == val
7789 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7790 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7791 // && mask == ~mask2
7792 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7793 // && ~mask == mask2
7794 // (i.e., copy a bitfield value into another bitfield of the same width)
7799 SDValue N00 = N0.getOperand(0);
7801 // The value and the mask need to be constants so we can verify this is
7802 // actually a bitfield set. If the mask is 0xffff, we can do better
7803 // via a movt instruction, so don't use BFI in that case.
7804 SDValue MaskOp = N0.getOperand(1);
7805 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7808 unsigned Mask = MaskC->getZExtValue();
7812 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7813 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7815 unsigned Val = N1C->getZExtValue();
7816 if ((Val & ~Mask) != Val)
7819 if (ARM::isBitFieldInvertedMask(Mask)) {
7820 Val >>= CountTrailingZeros_32(~Mask);
7822 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7823 DAG.getConstant(Val, MVT::i32),
7824 DAG.getConstant(Mask, MVT::i32));
7826 // Do not add new nodes to DAG combiner worklist.
7827 DCI.CombineTo(N, Res, false);
7830 } else if (N1.getOpcode() == ISD::AND) {
7831 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7832 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7835 unsigned Mask2 = N11C->getZExtValue();
7837 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7839 if (ARM::isBitFieldInvertedMask(Mask) &&
7841 // The pack halfword instruction works better for masks that fit it,
7842 // so use that when it's available.
7843 if (Subtarget->hasT2ExtractPack() &&
7844 (Mask == 0xffff || Mask == 0xffff0000))
7847 unsigned amt = CountTrailingZeros_32(Mask2);
7848 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7849 DAG.getConstant(amt, MVT::i32));
7850 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7851 DAG.getConstant(Mask, MVT::i32));
7852 // Do not add new nodes to DAG combiner worklist.
7853 DCI.CombineTo(N, Res, false);
7855 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7857 // The pack halfword instruction works better for masks that fit it,
7858 // so use that when it's available.
7859 if (Subtarget->hasT2ExtractPack() &&
7860 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7863 unsigned lsb = CountTrailingZeros_32(Mask);
7864 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7865 DAG.getConstant(lsb, MVT::i32));
7866 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7867 DAG.getConstant(Mask2, MVT::i32));
7868 // Do not add new nodes to DAG combiner worklist.
7869 DCI.CombineTo(N, Res, false);
7874 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7875 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7876 ARM::isBitFieldInvertedMask(~Mask)) {
7877 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7878 // where lsb(mask) == #shamt and masked bits of B are known zero.
7879 SDValue ShAmt = N00.getOperand(1);
7880 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7881 unsigned LSB = CountTrailingZeros_32(Mask);
7885 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7886 DAG.getConstant(~Mask, MVT::i32));
7888 // Do not add new nodes to DAG combiner worklist.
7889 DCI.CombineTo(N, Res, false);
7895 static SDValue PerformXORCombine(SDNode *N,
7896 TargetLowering::DAGCombinerInfo &DCI,
7897 const ARMSubtarget *Subtarget) {
7898 EVT VT = N->getValueType(0);
7899 SelectionDAG &DAG = DCI.DAG;
7901 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7904 if (!Subtarget->isThumb1Only()) {
7905 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7906 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7907 if (Result.getNode())
7914 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7915 /// the bits being cleared by the AND are not demanded by the BFI.
7916 static SDValue PerformBFICombine(SDNode *N,
7917 TargetLowering::DAGCombinerInfo &DCI) {
7918 SDValue N1 = N->getOperand(1);
7919 if (N1.getOpcode() == ISD::AND) {
7920 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7923 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7924 unsigned LSB = CountTrailingZeros_32(~InvMask);
7925 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7926 unsigned Mask = (1 << Width)-1;
7927 unsigned Mask2 = N11C->getZExtValue();
7928 if ((Mask & (~Mask2)) == 0)
7929 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7930 N->getOperand(0), N1.getOperand(0),
7936 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7937 /// ARMISD::VMOVRRD.
7938 static SDValue PerformVMOVRRDCombine(SDNode *N,
7939 TargetLowering::DAGCombinerInfo &DCI) {
7940 // vmovrrd(vmovdrr x, y) -> x,y
7941 SDValue InDouble = N->getOperand(0);
7942 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7943 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7945 // vmovrrd(load f64) -> (load i32), (load i32)
7946 SDNode *InNode = InDouble.getNode();
7947 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7948 InNode->getValueType(0) == MVT::f64 &&
7949 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7950 !cast<LoadSDNode>(InNode)->isVolatile()) {
7951 // TODO: Should this be done for non-FrameIndex operands?
7952 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7954 SelectionDAG &DAG = DCI.DAG;
7955 DebugLoc DL = LD->getDebugLoc();
7956 SDValue BasePtr = LD->getBasePtr();
7957 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7958 LD->getPointerInfo(), LD->isVolatile(),
7959 LD->isNonTemporal(), LD->isInvariant(),
7960 LD->getAlignment());
7962 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7963 DAG.getConstant(4, MVT::i32));
7964 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7965 LD->getPointerInfo(), LD->isVolatile(),
7966 LD->isNonTemporal(), LD->isInvariant(),
7967 std::min(4U, LD->getAlignment() / 2));
7969 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7970 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7971 DCI.RemoveFromWorklist(LD);
7979 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7980 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7981 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7982 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7983 SDValue Op0 = N->getOperand(0);
7984 SDValue Op1 = N->getOperand(1);
7985 if (Op0.getOpcode() == ISD::BITCAST)
7986 Op0 = Op0.getOperand(0);
7987 if (Op1.getOpcode() == ISD::BITCAST)
7988 Op1 = Op1.getOperand(0);
7989 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7990 Op0.getNode() == Op1.getNode() &&
7991 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7992 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7993 N->getValueType(0), Op0.getOperand(0));
7997 /// PerformSTORECombine - Target-specific dag combine xforms for
7999 static SDValue PerformSTORECombine(SDNode *N,
8000 TargetLowering::DAGCombinerInfo &DCI) {
8001 StoreSDNode *St = cast<StoreSDNode>(N);
8002 if (St->isVolatile())
8005 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8006 // pack all of the elements in one place. Next, store to memory in fewer
8008 SDValue StVal = St->getValue();
8009 EVT VT = StVal.getValueType();
8010 if (St->isTruncatingStore() && VT.isVector()) {
8011 SelectionDAG &DAG = DCI.DAG;
8012 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8013 EVT StVT = St->getMemoryVT();
8014 unsigned NumElems = VT.getVectorNumElements();
8015 assert(StVT != VT && "Cannot truncate to the same type");
8016 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8017 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8019 // From, To sizes and ElemCount must be pow of two
8020 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8022 // We are going to use the original vector elt for storing.
8023 // Accumulated smaller vector elements must be a multiple of the store size.
8024 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8026 unsigned SizeRatio = FromEltSz / ToEltSz;
8027 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8029 // Create a type on which we perform the shuffle.
8030 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8031 NumElems*SizeRatio);
8032 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8034 DebugLoc DL = St->getDebugLoc();
8035 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8036 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8037 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8039 // Can't shuffle using an illegal type.
8040 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8042 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8043 DAG.getUNDEF(WideVec.getValueType()),
8045 // At this point all of the data is stored at the bottom of the
8046 // register. We now need to save it to mem.
8048 // Find the largest store unit
8049 MVT StoreType = MVT::i8;
8050 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8051 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8052 MVT Tp = (MVT::SimpleValueType)tp;
8053 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8056 // Didn't find a legal store type.
8057 if (!TLI.isTypeLegal(StoreType))
8060 // Bitcast the original vector into a vector of store-size units
8061 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8062 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8063 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8064 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8065 SmallVector<SDValue, 8> Chains;
8066 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8067 TLI.getPointerTy());
8068 SDValue BasePtr = St->getBasePtr();
8070 // Perform one or more big stores into memory.
8071 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8072 for (unsigned I = 0; I < E; I++) {
8073 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8074 StoreType, ShuffWide,
8075 DAG.getIntPtrConstant(I));
8076 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8077 St->getPointerInfo(), St->isVolatile(),
8078 St->isNonTemporal(), St->getAlignment());
8079 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8081 Chains.push_back(Ch);
8083 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8087 if (!ISD::isNormalStore(St))
8090 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8091 // ARM stores of arguments in the same cache line.
8092 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8093 StVal.getNode()->hasOneUse()) {
8094 SelectionDAG &DAG = DCI.DAG;
8095 DebugLoc DL = St->getDebugLoc();
8096 SDValue BasePtr = St->getBasePtr();
8097 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8098 StVal.getNode()->getOperand(0), BasePtr,
8099 St->getPointerInfo(), St->isVolatile(),
8100 St->isNonTemporal(), St->getAlignment());
8102 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8103 DAG.getConstant(4, MVT::i32));
8104 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8105 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8106 St->isNonTemporal(),
8107 std::min(4U, St->getAlignment() / 2));
8110 if (StVal.getValueType() != MVT::i64 ||
8111 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8114 // Bitcast an i64 store extracted from a vector to f64.
8115 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8116 SelectionDAG &DAG = DCI.DAG;
8117 DebugLoc dl = StVal.getDebugLoc();
8118 SDValue IntVec = StVal.getOperand(0);
8119 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8120 IntVec.getValueType().getVectorNumElements());
8121 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8122 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8123 Vec, StVal.getOperand(1));
8124 dl = N->getDebugLoc();
8125 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8126 // Make the DAGCombiner fold the bitcasts.
8127 DCI.AddToWorklist(Vec.getNode());
8128 DCI.AddToWorklist(ExtElt.getNode());
8129 DCI.AddToWorklist(V.getNode());
8130 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8131 St->getPointerInfo(), St->isVolatile(),
8132 St->isNonTemporal(), St->getAlignment(),
8136 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8137 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8138 /// i64 vector to have f64 elements, since the value can then be loaded
8139 /// directly into a VFP register.
8140 static bool hasNormalLoadOperand(SDNode *N) {
8141 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8142 for (unsigned i = 0; i < NumElts; ++i) {
8143 SDNode *Elt = N->getOperand(i).getNode();
8144 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8150 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8151 /// ISD::BUILD_VECTOR.
8152 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8153 TargetLowering::DAGCombinerInfo &DCI){
8154 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8155 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8156 // into a pair of GPRs, which is fine when the value is used as a scalar,
8157 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8158 SelectionDAG &DAG = DCI.DAG;
8159 if (N->getNumOperands() == 2) {
8160 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8165 // Load i64 elements as f64 values so that type legalization does not split
8166 // them up into i32 values.
8167 EVT VT = N->getValueType(0);
8168 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8170 DebugLoc dl = N->getDebugLoc();
8171 SmallVector<SDValue, 8> Ops;
8172 unsigned NumElts = VT.getVectorNumElements();
8173 for (unsigned i = 0; i < NumElts; ++i) {
8174 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8176 // Make the DAGCombiner fold the bitcast.
8177 DCI.AddToWorklist(V.getNode());
8179 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8180 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8181 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8184 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8185 /// ISD::INSERT_VECTOR_ELT.
8186 static SDValue PerformInsertEltCombine(SDNode *N,
8187 TargetLowering::DAGCombinerInfo &DCI) {
8188 // Bitcast an i64 load inserted into a vector to f64.
8189 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8190 EVT VT = N->getValueType(0);
8191 SDNode *Elt = N->getOperand(1).getNode();
8192 if (VT.getVectorElementType() != MVT::i64 ||
8193 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8196 SelectionDAG &DAG = DCI.DAG;
8197 DebugLoc dl = N->getDebugLoc();
8198 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8199 VT.getVectorNumElements());
8200 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8201 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8202 // Make the DAGCombiner fold the bitcasts.
8203 DCI.AddToWorklist(Vec.getNode());
8204 DCI.AddToWorklist(V.getNode());
8205 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8206 Vec, V, N->getOperand(2));
8207 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8210 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8211 /// ISD::VECTOR_SHUFFLE.
8212 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8213 // The LLVM shufflevector instruction does not require the shuffle mask
8214 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8215 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8216 // operands do not match the mask length, they are extended by concatenating
8217 // them with undef vectors. That is probably the right thing for other
8218 // targets, but for NEON it is better to concatenate two double-register
8219 // size vector operands into a single quad-register size vector. Do that
8220 // transformation here:
8221 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8222 // shuffle(concat(v1, v2), undef)
8223 SDValue Op0 = N->getOperand(0);
8224 SDValue Op1 = N->getOperand(1);
8225 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8226 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8227 Op0.getNumOperands() != 2 ||
8228 Op1.getNumOperands() != 2)
8230 SDValue Concat0Op1 = Op0.getOperand(1);
8231 SDValue Concat1Op1 = Op1.getOperand(1);
8232 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8233 Concat1Op1.getOpcode() != ISD::UNDEF)
8235 // Skip the transformation if any of the types are illegal.
8236 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8237 EVT VT = N->getValueType(0);
8238 if (!TLI.isTypeLegal(VT) ||
8239 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8240 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8243 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8244 Op0.getOperand(0), Op1.getOperand(0));
8245 // Translate the shuffle mask.
8246 SmallVector<int, 16> NewMask;
8247 unsigned NumElts = VT.getVectorNumElements();
8248 unsigned HalfElts = NumElts/2;
8249 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8250 for (unsigned n = 0; n < NumElts; ++n) {
8251 int MaskElt = SVN->getMaskElt(n);
8253 if (MaskElt < (int)HalfElts)
8255 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8256 NewElt = HalfElts + MaskElt - NumElts;
8257 NewMask.push_back(NewElt);
8259 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8260 DAG.getUNDEF(VT), NewMask.data());
8263 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8264 /// NEON load/store intrinsics to merge base address updates.
8265 static SDValue CombineBaseUpdate(SDNode *N,
8266 TargetLowering::DAGCombinerInfo &DCI) {
8267 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8270 SelectionDAG &DAG = DCI.DAG;
8271 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8272 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8273 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8274 SDValue Addr = N->getOperand(AddrOpIdx);
8276 // Search for a use of the address operand that is an increment.
8277 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8278 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8280 if (User->getOpcode() != ISD::ADD ||
8281 UI.getUse().getResNo() != Addr.getResNo())
8284 // Check that the add is independent of the load/store. Otherwise, folding
8285 // it would create a cycle.
8286 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8289 // Find the new opcode for the updating load/store.
8291 bool isLaneOp = false;
8292 unsigned NewOpc = 0;
8293 unsigned NumVecs = 0;
8295 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8297 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8298 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8300 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8302 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8304 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8306 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8307 NumVecs = 2; isLaneOp = true; break;
8308 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8309 NumVecs = 3; isLaneOp = true; break;
8310 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8311 NumVecs = 4; isLaneOp = true; break;
8312 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8313 NumVecs = 1; isLoad = false; break;
8314 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8315 NumVecs = 2; isLoad = false; break;
8316 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8317 NumVecs = 3; isLoad = false; break;
8318 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8319 NumVecs = 4; isLoad = false; break;
8320 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8321 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8322 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8323 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8324 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8325 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8329 switch (N->getOpcode()) {
8330 default: llvm_unreachable("unexpected opcode for Neon base update");
8331 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8332 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8333 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8337 // Find the size of memory referenced by the load/store.
8340 VecTy = N->getValueType(0);
8342 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8343 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8345 NumBytes /= VecTy.getVectorNumElements();
8347 // If the increment is a constant, it must match the memory ref size.
8348 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8349 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8350 uint64_t IncVal = CInc->getZExtValue();
8351 if (IncVal != NumBytes)
8353 } else if (NumBytes >= 3 * 16) {
8354 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8355 // separate instructions that make it harder to use a non-constant update.
8359 // Create the new updating load/store node.
8361 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8363 for (n = 0; n < NumResultVecs; ++n)
8365 Tys[n++] = MVT::i32;
8366 Tys[n] = MVT::Other;
8367 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8368 SmallVector<SDValue, 8> Ops;
8369 Ops.push_back(N->getOperand(0)); // incoming chain
8370 Ops.push_back(N->getOperand(AddrOpIdx));
8372 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8373 Ops.push_back(N->getOperand(i));
8375 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8376 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8377 Ops.data(), Ops.size(),
8378 MemInt->getMemoryVT(),
8379 MemInt->getMemOperand());
8382 std::vector<SDValue> NewResults;
8383 for (unsigned i = 0; i < NumResultVecs; ++i) {
8384 NewResults.push_back(SDValue(UpdN.getNode(), i));
8386 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8387 DCI.CombineTo(N, NewResults);
8388 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8395 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8396 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8397 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8399 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8400 SelectionDAG &DAG = DCI.DAG;
8401 EVT VT = N->getValueType(0);
8402 // vldN-dup instructions only support 64-bit vectors for N > 1.
8403 if (!VT.is64BitVector())
8406 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8407 SDNode *VLD = N->getOperand(0).getNode();
8408 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8410 unsigned NumVecs = 0;
8411 unsigned NewOpc = 0;
8412 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8413 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8415 NewOpc = ARMISD::VLD2DUP;
8416 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8418 NewOpc = ARMISD::VLD3DUP;
8419 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8421 NewOpc = ARMISD::VLD4DUP;
8426 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8427 // numbers match the load.
8428 unsigned VLDLaneNo =
8429 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8430 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8432 // Ignore uses of the chain result.
8433 if (UI.getUse().getResNo() == NumVecs)
8436 if (User->getOpcode() != ARMISD::VDUPLANE ||
8437 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8441 // Create the vldN-dup node.
8444 for (n = 0; n < NumVecs; ++n)
8446 Tys[n] = MVT::Other;
8447 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8448 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8449 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8450 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8451 Ops, 2, VLDMemInt->getMemoryVT(),
8452 VLDMemInt->getMemOperand());
8455 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8457 unsigned ResNo = UI.getUse().getResNo();
8458 // Ignore uses of the chain result.
8459 if (ResNo == NumVecs)
8462 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8465 // Now the vldN-lane intrinsic is dead except for its chain result.
8466 // Update uses of the chain.
8467 std::vector<SDValue> VLDDupResults;
8468 for (unsigned n = 0; n < NumVecs; ++n)
8469 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8470 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8471 DCI.CombineTo(VLD, VLDDupResults);
8476 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8477 /// ARMISD::VDUPLANE.
8478 static SDValue PerformVDUPLANECombine(SDNode *N,
8479 TargetLowering::DAGCombinerInfo &DCI) {
8480 SDValue Op = N->getOperand(0);
8482 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8483 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8484 if (CombineVLDDUP(N, DCI))
8485 return SDValue(N, 0);
8487 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8488 // redundant. Ignore bit_converts for now; element sizes are checked below.
8489 while (Op.getOpcode() == ISD::BITCAST)
8490 Op = Op.getOperand(0);
8491 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8494 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8495 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8496 // The canonical VMOV for a zero vector uses a 32-bit element size.
8497 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8499 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8501 EVT VT = N->getValueType(0);
8502 if (EltSize > VT.getVectorElementType().getSizeInBits())
8505 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8508 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8509 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8510 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8514 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8516 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8521 APFloat APF = C->getValueAPF();
8522 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8523 != APFloat::opOK || !isExact)
8526 c0 = (I == 0) ? cN : c0;
8527 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8534 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8535 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8536 /// when the VMUL has a constant operand that is a power of 2.
8538 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8539 /// vmul.f32 d16, d17, d16
8540 /// vcvt.s32.f32 d16, d16
8542 /// vcvt.s32.f32 d16, d16, #3
8543 static SDValue PerformVCVTCombine(SDNode *N,
8544 TargetLowering::DAGCombinerInfo &DCI,
8545 const ARMSubtarget *Subtarget) {
8546 SelectionDAG &DAG = DCI.DAG;
8547 SDValue Op = N->getOperand(0);
8549 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8550 Op.getOpcode() != ISD::FMUL)
8554 SDValue N0 = Op->getOperand(0);
8555 SDValue ConstVec = Op->getOperand(1);
8556 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8558 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8559 !isConstVecPow2(ConstVec, isSigned, C))
8562 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8563 Intrinsic::arm_neon_vcvtfp2fxu;
8564 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8566 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8567 DAG.getConstant(Log2_64(C), MVT::i32));
8570 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8571 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8572 /// when the VDIV has a constant operand that is a power of 2.
8574 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8575 /// vcvt.f32.s32 d16, d16
8576 /// vdiv.f32 d16, d17, d16
8578 /// vcvt.f32.s32 d16, d16, #3
8579 static SDValue PerformVDIVCombine(SDNode *N,
8580 TargetLowering::DAGCombinerInfo &DCI,
8581 const ARMSubtarget *Subtarget) {
8582 SelectionDAG &DAG = DCI.DAG;
8583 SDValue Op = N->getOperand(0);
8584 unsigned OpOpcode = Op.getNode()->getOpcode();
8586 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8587 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8591 SDValue ConstVec = N->getOperand(1);
8592 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8594 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8595 !isConstVecPow2(ConstVec, isSigned, C))
8598 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8599 Intrinsic::arm_neon_vcvtfxu2fp;
8600 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8602 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8603 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8606 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8607 /// operand of a vector shift operation, where all the elements of the
8608 /// build_vector must have the same constant integer value.
8609 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8610 // Ignore bit_converts.
8611 while (Op.getOpcode() == ISD::BITCAST)
8612 Op = Op.getOperand(0);
8613 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8614 APInt SplatBits, SplatUndef;
8615 unsigned SplatBitSize;
8617 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8618 HasAnyUndefs, ElementBits) ||
8619 SplatBitSize > ElementBits)
8621 Cnt = SplatBits.getSExtValue();
8625 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8626 /// operand of a vector shift left operation. That value must be in the range:
8627 /// 0 <= Value < ElementBits for a left shift; or
8628 /// 0 <= Value <= ElementBits for a long left shift.
8629 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8630 assert(VT.isVector() && "vector shift count is not a vector type");
8631 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8632 if (! getVShiftImm(Op, ElementBits, Cnt))
8634 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8637 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8638 /// operand of a vector shift right operation. For a shift opcode, the value
8639 /// is positive, but for an intrinsic the value count must be negative. The
8640 /// absolute value must be in the range:
8641 /// 1 <= |Value| <= ElementBits for a right shift; or
8642 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8643 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8645 assert(VT.isVector() && "vector shift count is not a vector type");
8646 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8647 if (! getVShiftImm(Op, ElementBits, Cnt))
8651 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8654 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8655 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8656 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8659 // Don't do anything for most intrinsics.
8662 // Vector shifts: check for immediate versions and lower them.
8663 // Note: This is done during DAG combining instead of DAG legalizing because
8664 // the build_vectors for 64-bit vector element shift counts are generally
8665 // not legal, and it is hard to see their values after they get legalized to
8666 // loads from a constant pool.
8667 case Intrinsic::arm_neon_vshifts:
8668 case Intrinsic::arm_neon_vshiftu:
8669 case Intrinsic::arm_neon_vshiftls:
8670 case Intrinsic::arm_neon_vshiftlu:
8671 case Intrinsic::arm_neon_vshiftn:
8672 case Intrinsic::arm_neon_vrshifts:
8673 case Intrinsic::arm_neon_vrshiftu:
8674 case Intrinsic::arm_neon_vrshiftn:
8675 case Intrinsic::arm_neon_vqshifts:
8676 case Intrinsic::arm_neon_vqshiftu:
8677 case Intrinsic::arm_neon_vqshiftsu:
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 EVT VT = N->getOperand(1).getValueType();
8686 unsigned VShiftOpc = 0;
8689 case Intrinsic::arm_neon_vshifts:
8690 case Intrinsic::arm_neon_vshiftu:
8691 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8692 VShiftOpc = ARMISD::VSHL;
8695 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8696 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8697 ARMISD::VSHRs : ARMISD::VSHRu);
8702 case Intrinsic::arm_neon_vshiftls:
8703 case Intrinsic::arm_neon_vshiftlu:
8704 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8706 llvm_unreachable("invalid shift count for vshll intrinsic");
8708 case Intrinsic::arm_neon_vrshifts:
8709 case Intrinsic::arm_neon_vrshiftu:
8710 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8714 case Intrinsic::arm_neon_vqshifts:
8715 case Intrinsic::arm_neon_vqshiftu:
8716 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8720 case Intrinsic::arm_neon_vqshiftsu:
8721 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8723 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8725 case Intrinsic::arm_neon_vshiftn:
8726 case Intrinsic::arm_neon_vrshiftn:
8727 case Intrinsic::arm_neon_vqshiftns:
8728 case Intrinsic::arm_neon_vqshiftnu:
8729 case Intrinsic::arm_neon_vqshiftnsu:
8730 case Intrinsic::arm_neon_vqrshiftns:
8731 case Intrinsic::arm_neon_vqrshiftnu:
8732 case Intrinsic::arm_neon_vqrshiftnsu:
8733 // Narrowing shifts require an immediate right shift.
8734 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8736 llvm_unreachable("invalid shift count for narrowing vector shift "
8740 llvm_unreachable("unhandled vector shift");
8744 case Intrinsic::arm_neon_vshifts:
8745 case Intrinsic::arm_neon_vshiftu:
8746 // Opcode already set above.
8748 case Intrinsic::arm_neon_vshiftls:
8749 case Intrinsic::arm_neon_vshiftlu:
8750 if (Cnt == VT.getVectorElementType().getSizeInBits())
8751 VShiftOpc = ARMISD::VSHLLi;
8753 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8754 ARMISD::VSHLLs : ARMISD::VSHLLu);
8756 case Intrinsic::arm_neon_vshiftn:
8757 VShiftOpc = ARMISD::VSHRN; break;
8758 case Intrinsic::arm_neon_vrshifts:
8759 VShiftOpc = ARMISD::VRSHRs; break;
8760 case Intrinsic::arm_neon_vrshiftu:
8761 VShiftOpc = ARMISD::VRSHRu; break;
8762 case Intrinsic::arm_neon_vrshiftn:
8763 VShiftOpc = ARMISD::VRSHRN; break;
8764 case Intrinsic::arm_neon_vqshifts:
8765 VShiftOpc = ARMISD::VQSHLs; break;
8766 case Intrinsic::arm_neon_vqshiftu:
8767 VShiftOpc = ARMISD::VQSHLu; break;
8768 case Intrinsic::arm_neon_vqshiftsu:
8769 VShiftOpc = ARMISD::VQSHLsu; break;
8770 case Intrinsic::arm_neon_vqshiftns:
8771 VShiftOpc = ARMISD::VQSHRNs; break;
8772 case Intrinsic::arm_neon_vqshiftnu:
8773 VShiftOpc = ARMISD::VQSHRNu; break;
8774 case Intrinsic::arm_neon_vqshiftnsu:
8775 VShiftOpc = ARMISD::VQSHRNsu; break;
8776 case Intrinsic::arm_neon_vqrshiftns:
8777 VShiftOpc = ARMISD::VQRSHRNs; break;
8778 case Intrinsic::arm_neon_vqrshiftnu:
8779 VShiftOpc = ARMISD::VQRSHRNu; break;
8780 case Intrinsic::arm_neon_vqrshiftnsu:
8781 VShiftOpc = ARMISD::VQRSHRNsu; break;
8784 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8785 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8788 case Intrinsic::arm_neon_vshiftins: {
8789 EVT VT = N->getOperand(1).getValueType();
8791 unsigned VShiftOpc = 0;
8793 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8794 VShiftOpc = ARMISD::VSLI;
8795 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8796 VShiftOpc = ARMISD::VSRI;
8798 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8801 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8802 N->getOperand(1), N->getOperand(2),
8803 DAG.getConstant(Cnt, MVT::i32));
8806 case Intrinsic::arm_neon_vqrshifts:
8807 case Intrinsic::arm_neon_vqrshiftu:
8808 // No immediate versions of these to check for.
8815 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8816 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8817 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8818 /// vector element shift counts are generally not legal, and it is hard to see
8819 /// their values after they get legalized to loads from a constant pool.
8820 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8821 const ARMSubtarget *ST) {
8822 EVT VT = N->getValueType(0);
8823 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8824 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8825 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8826 SDValue N1 = N->getOperand(1);
8827 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8828 SDValue N0 = N->getOperand(0);
8829 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8830 DAG.MaskedValueIsZero(N0.getOperand(0),
8831 APInt::getHighBitsSet(32, 16)))
8832 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8836 // Nothing to be done for scalar shifts.
8837 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8838 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8841 assert(ST->hasNEON() && "unexpected vector shift");
8844 switch (N->getOpcode()) {
8845 default: llvm_unreachable("unexpected shift opcode");
8848 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8849 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8850 DAG.getConstant(Cnt, MVT::i32));
8855 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8856 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8857 ARMISD::VSHRs : ARMISD::VSHRu);
8858 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8859 DAG.getConstant(Cnt, MVT::i32));
8865 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8866 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8867 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8868 const ARMSubtarget *ST) {
8869 SDValue N0 = N->getOperand(0);
8871 // Check for sign- and zero-extensions of vector extract operations of 8-
8872 // and 16-bit vector elements. NEON supports these directly. They are
8873 // handled during DAG combining because type legalization will promote them
8874 // to 32-bit types and it is messy to recognize the operations after that.
8875 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8876 SDValue Vec = N0.getOperand(0);
8877 SDValue Lane = N0.getOperand(1);
8878 EVT VT = N->getValueType(0);
8879 EVT EltVT = N0.getValueType();
8880 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8882 if (VT == MVT::i32 &&
8883 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8884 TLI.isTypeLegal(Vec.getValueType()) &&
8885 isa<ConstantSDNode>(Lane)) {
8888 switch (N->getOpcode()) {
8889 default: llvm_unreachable("unexpected opcode");
8890 case ISD::SIGN_EXTEND:
8891 Opc = ARMISD::VGETLANEs;
8893 case ISD::ZERO_EXTEND:
8894 case ISD::ANY_EXTEND:
8895 Opc = ARMISD::VGETLANEu;
8898 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8905 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8906 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8907 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8908 const ARMSubtarget *ST) {
8909 // If the target supports NEON, try to use vmax/vmin instructions for f32
8910 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8911 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8912 // a NaN; only do the transformation when it matches that behavior.
8914 // For now only do this when using NEON for FP operations; if using VFP, it
8915 // is not obvious that the benefit outweighs the cost of switching to the
8917 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8918 N->getValueType(0) != MVT::f32)
8921 SDValue CondLHS = N->getOperand(0);
8922 SDValue CondRHS = N->getOperand(1);
8923 SDValue LHS = N->getOperand(2);
8924 SDValue RHS = N->getOperand(3);
8925 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8927 unsigned Opcode = 0;
8929 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8930 IsReversed = false; // x CC y ? x : y
8931 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8932 IsReversed = true ; // x CC y ? y : x
8946 // If LHS is NaN, an ordered comparison will be false and the result will
8947 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8948 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8949 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8950 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8952 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8953 // will return -0, so vmin can only be used for unsafe math or if one of
8954 // the operands is known to be nonzero.
8955 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8956 !DAG.getTarget().Options.UnsafeFPMath &&
8957 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8959 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8968 // If LHS is NaN, an ordered comparison will be false and the result will
8969 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8970 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8971 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8972 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8974 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8975 // will return +0, so vmax can only be used for unsafe math or if one of
8976 // the operands is known to be nonzero.
8977 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8978 !DAG.getTarget().Options.UnsafeFPMath &&
8979 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8981 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8987 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8990 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8992 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8993 SDValue Cmp = N->getOperand(4);
8994 if (Cmp.getOpcode() != ARMISD::CMPZ)
8995 // Only looking at EQ and NE cases.
8998 EVT VT = N->getValueType(0);
8999 DebugLoc dl = N->getDebugLoc();
9000 SDValue LHS = Cmp.getOperand(0);
9001 SDValue RHS = Cmp.getOperand(1);
9002 SDValue FalseVal = N->getOperand(0);
9003 SDValue TrueVal = N->getOperand(1);
9004 SDValue ARMcc = N->getOperand(2);
9005 ARMCC::CondCodes CC =
9006 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9024 /// FIXME: Turn this into a target neutral optimization?
9026 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9027 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9028 N->getOperand(3), Cmp);
9029 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9031 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9032 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9033 N->getOperand(3), NewCmp);
9036 if (Res.getNode()) {
9037 APInt KnownZero, KnownOne;
9038 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9039 // Capture demanded bits information that would be otherwise lost.
9040 if (KnownZero == 0xfffffffe)
9041 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9042 DAG.getValueType(MVT::i1));
9043 else if (KnownZero == 0xffffff00)
9044 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9045 DAG.getValueType(MVT::i8));
9046 else if (KnownZero == 0xffff0000)
9047 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9048 DAG.getValueType(MVT::i16));
9054 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9055 DAGCombinerInfo &DCI) const {
9056 switch (N->getOpcode()) {
9058 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9059 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9060 case ISD::SUB: return PerformSUBCombine(N, DCI);
9061 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9062 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9063 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9064 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9065 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9066 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9067 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9068 case ISD::STORE: return PerformSTORECombine(N, DCI);
9069 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9070 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9071 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9072 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9073 case ISD::FP_TO_SINT:
9074 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9075 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9076 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9079 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9080 case ISD::SIGN_EXTEND:
9081 case ISD::ZERO_EXTEND:
9082 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9083 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9084 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9085 case ARMISD::VLD2DUP:
9086 case ARMISD::VLD3DUP:
9087 case ARMISD::VLD4DUP:
9088 return CombineBaseUpdate(N, DCI);
9089 case ISD::INTRINSIC_VOID:
9090 case ISD::INTRINSIC_W_CHAIN:
9091 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9092 case Intrinsic::arm_neon_vld1:
9093 case Intrinsic::arm_neon_vld2:
9094 case Intrinsic::arm_neon_vld3:
9095 case Intrinsic::arm_neon_vld4:
9096 case Intrinsic::arm_neon_vld2lane:
9097 case Intrinsic::arm_neon_vld3lane:
9098 case Intrinsic::arm_neon_vld4lane:
9099 case Intrinsic::arm_neon_vst1:
9100 case Intrinsic::arm_neon_vst2:
9101 case Intrinsic::arm_neon_vst3:
9102 case Intrinsic::arm_neon_vst4:
9103 case Intrinsic::arm_neon_vst2lane:
9104 case Intrinsic::arm_neon_vst3lane:
9105 case Intrinsic::arm_neon_vst4lane:
9106 return CombineBaseUpdate(N, DCI);
9114 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9116 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9119 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9120 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9121 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9123 switch (VT.getSimpleVT().SimpleTy) {
9129 // Unaligned access can use (for example) LRDB, LRDH, LDR
9130 return AllowsUnaligned;
9133 // For any little-endian targets with neon, we can support unaligned ld/st
9134 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9135 // A big-endian target may also explictly support unaligned accesses
9136 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9140 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9141 unsigned AlignCheck) {
9142 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9143 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9146 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9147 unsigned DstAlign, unsigned SrcAlign,
9150 MachineFunction &MF) const {
9151 const Function *F = MF.getFunction();
9153 // See if we can use NEON instructions for this...
9155 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9156 Subtarget->hasNEON()) {
9157 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9159 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9164 // Lowering to i32/i16 if the size permits.
9167 } else if (Size >= 2) {
9171 // Let the target-independent logic figure it out.
9175 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9180 switch (VT.getSimpleVT().SimpleTy) {
9181 default: return false;
9196 if ((V & (Scale - 1)) != 0)
9199 return V == (V & ((1LL << 5) - 1));
9202 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9203 const ARMSubtarget *Subtarget) {
9210 switch (VT.getSimpleVT().SimpleTy) {
9211 default: return false;
9216 // + imm12 or - imm8
9218 return V == (V & ((1LL << 8) - 1));
9219 return V == (V & ((1LL << 12) - 1));
9222 // Same as ARM mode. FIXME: NEON?
9223 if (!Subtarget->hasVFP2())
9228 return V == (V & ((1LL << 8) - 1));
9232 /// isLegalAddressImmediate - Return true if the integer value can be used
9233 /// as the offset of the target addressing mode for load / store of the
9235 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9236 const ARMSubtarget *Subtarget) {
9243 if (Subtarget->isThumb1Only())
9244 return isLegalT1AddressImmediate(V, VT);
9245 else if (Subtarget->isThumb2())
9246 return isLegalT2AddressImmediate(V, VT, Subtarget);
9251 switch (VT.getSimpleVT().SimpleTy) {
9252 default: return false;
9257 return V == (V & ((1LL << 12) - 1));
9260 return V == (V & ((1LL << 8) - 1));
9263 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9268 return V == (V & ((1LL << 8) - 1));
9272 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9274 int Scale = AM.Scale;
9278 switch (VT.getSimpleVT().SimpleTy) {
9279 default: return false;
9288 return Scale == 2 || Scale == 4 || Scale == 8;
9291 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9295 // Note, we allow "void" uses (basically, uses that aren't loads or
9296 // stores), because arm allows folding a scale into many arithmetic
9297 // operations. This should be made more precise and revisited later.
9299 // Allow r << imm, but the imm has to be a multiple of two.
9300 if (Scale & 1) return false;
9301 return isPowerOf2_32(Scale);
9305 /// isLegalAddressingMode - Return true if the addressing mode represented
9306 /// by AM is legal for this target, for a load/store of the specified type.
9307 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9309 EVT VT = getValueType(Ty, true);
9310 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9313 // Can never fold addr of global into load/store.
9318 case 0: // no scale reg, must be "r+i" or "r", or "i".
9321 if (Subtarget->isThumb1Only())
9325 // ARM doesn't support any R+R*scale+imm addr modes.
9332 if (Subtarget->isThumb2())
9333 return isLegalT2ScaledAddressingMode(AM, VT);
9335 int Scale = AM.Scale;
9336 switch (VT.getSimpleVT().SimpleTy) {
9337 default: return false;
9341 if (Scale < 0) Scale = -Scale;
9345 return isPowerOf2_32(Scale & ~1);
9349 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9354 // Note, we allow "void" uses (basically, uses that aren't loads or
9355 // stores), because arm allows folding a scale into many arithmetic
9356 // operations. This should be made more precise and revisited later.
9358 // Allow r << imm, but the imm has to be a multiple of two.
9359 if (Scale & 1) return false;
9360 return isPowerOf2_32(Scale);
9366 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9367 /// icmp immediate, that is the target has icmp instructions which can compare
9368 /// a register against the immediate without having to materialize the
9369 /// immediate into a register.
9370 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9371 // Thumb2 and ARM modes can use cmn for negative immediates.
9372 if (!Subtarget->isThumb())
9373 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9374 if (Subtarget->isThumb2())
9375 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9376 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9377 return Imm >= 0 && Imm <= 255;
9380 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9381 /// *or sub* immediate, that is the target has add or sub instructions which can
9382 /// add a register with the immediate without having to materialize the
9383 /// immediate into a register.
9384 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9385 // Same encoding for add/sub, just flip the sign.
9386 int64_t AbsImm = llvm::abs64(Imm);
9387 if (!Subtarget->isThumb())
9388 return ARM_AM::getSOImmVal(AbsImm) != -1;
9389 if (Subtarget->isThumb2())
9390 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9391 // Thumb1 only has 8-bit unsigned immediate.
9392 return AbsImm >= 0 && AbsImm <= 255;
9395 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9396 bool isSEXTLoad, SDValue &Base,
9397 SDValue &Offset, bool &isInc,
9398 SelectionDAG &DAG) {
9399 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9402 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9404 Base = Ptr->getOperand(0);
9405 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9406 int RHSC = (int)RHS->getZExtValue();
9407 if (RHSC < 0 && RHSC > -256) {
9408 assert(Ptr->getOpcode() == ISD::ADD);
9410 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9414 isInc = (Ptr->getOpcode() == ISD::ADD);
9415 Offset = Ptr->getOperand(1);
9417 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9419 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9420 int RHSC = (int)RHS->getZExtValue();
9421 if (RHSC < 0 && RHSC > -0x1000) {
9422 assert(Ptr->getOpcode() == ISD::ADD);
9424 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9425 Base = Ptr->getOperand(0);
9430 if (Ptr->getOpcode() == ISD::ADD) {
9432 ARM_AM::ShiftOpc ShOpcVal=
9433 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9434 if (ShOpcVal != ARM_AM::no_shift) {
9435 Base = Ptr->getOperand(1);
9436 Offset = Ptr->getOperand(0);
9438 Base = Ptr->getOperand(0);
9439 Offset = Ptr->getOperand(1);
9444 isInc = (Ptr->getOpcode() == ISD::ADD);
9445 Base = Ptr->getOperand(0);
9446 Offset = Ptr->getOperand(1);
9450 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9454 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9455 bool isSEXTLoad, SDValue &Base,
9456 SDValue &Offset, bool &isInc,
9457 SelectionDAG &DAG) {
9458 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9461 Base = Ptr->getOperand(0);
9462 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9463 int RHSC = (int)RHS->getZExtValue();
9464 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9465 assert(Ptr->getOpcode() == ISD::ADD);
9467 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9469 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9470 isInc = Ptr->getOpcode() == ISD::ADD;
9471 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9479 /// getPreIndexedAddressParts - returns true by value, base pointer and
9480 /// offset pointer and addressing mode by reference if the node's address
9481 /// can be legally represented as pre-indexed load / store address.
9483 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9485 ISD::MemIndexedMode &AM,
9486 SelectionDAG &DAG) const {
9487 if (Subtarget->isThumb1Only())
9492 bool isSEXTLoad = false;
9493 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9494 Ptr = LD->getBasePtr();
9495 VT = LD->getMemoryVT();
9496 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9497 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9498 Ptr = ST->getBasePtr();
9499 VT = ST->getMemoryVT();
9504 bool isLegal = false;
9505 if (Subtarget->isThumb2())
9506 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9507 Offset, isInc, DAG);
9509 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9510 Offset, isInc, DAG);
9514 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9518 /// getPostIndexedAddressParts - returns true by value, base pointer and
9519 /// offset pointer and addressing mode by reference if this node can be
9520 /// combined with a load / store to form a post-indexed load / store.
9521 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9524 ISD::MemIndexedMode &AM,
9525 SelectionDAG &DAG) const {
9526 if (Subtarget->isThumb1Only())
9531 bool isSEXTLoad = false;
9532 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9533 VT = LD->getMemoryVT();
9534 Ptr = LD->getBasePtr();
9535 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9536 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9537 VT = ST->getMemoryVT();
9538 Ptr = ST->getBasePtr();
9543 bool isLegal = false;
9544 if (Subtarget->isThumb2())
9545 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9548 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9554 // Swap base ptr and offset to catch more post-index load / store when
9555 // it's legal. In Thumb2 mode, offset must be an immediate.
9556 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9557 !Subtarget->isThumb2())
9558 std::swap(Base, Offset);
9560 // Post-indexed load / store update the base pointer.
9565 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9569 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9572 const SelectionDAG &DAG,
9573 unsigned Depth) const {
9574 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9575 switch (Op.getOpcode()) {
9577 case ARMISD::CMOV: {
9578 // Bits are known zero/one if known on the LHS and RHS.
9579 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9580 if (KnownZero == 0 && KnownOne == 0) return;
9582 APInt KnownZeroRHS, KnownOneRHS;
9583 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9584 KnownZero &= KnownZeroRHS;
9585 KnownOne &= KnownOneRHS;
9591 //===----------------------------------------------------------------------===//
9592 // ARM Inline Assembly Support
9593 //===----------------------------------------------------------------------===//
9595 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9596 // Looking for "rev" which is V6+.
9597 if (!Subtarget->hasV6Ops())
9600 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9601 std::string AsmStr = IA->getAsmString();
9602 SmallVector<StringRef, 4> AsmPieces;
9603 SplitString(AsmStr, AsmPieces, ";\n");
9605 switch (AsmPieces.size()) {
9606 default: return false;
9608 AsmStr = AsmPieces[0];
9610 SplitString(AsmStr, AsmPieces, " \t,");
9613 if (AsmPieces.size() == 3 &&
9614 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9615 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9616 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9617 if (Ty && Ty->getBitWidth() == 32)
9618 return IntrinsicLowering::LowerToByteSwap(CI);
9626 /// getConstraintType - Given a constraint letter, return the type of
9627 /// constraint it is for this target.
9628 ARMTargetLowering::ConstraintType
9629 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9630 if (Constraint.size() == 1) {
9631 switch (Constraint[0]) {
9633 case 'l': return C_RegisterClass;
9634 case 'w': return C_RegisterClass;
9635 case 'h': return C_RegisterClass;
9636 case 'x': return C_RegisterClass;
9637 case 't': return C_RegisterClass;
9638 case 'j': return C_Other; // Constant for movw.
9639 // An address with a single base register. Due to the way we
9640 // currently handle addresses it is the same as an 'r' memory constraint.
9641 case 'Q': return C_Memory;
9643 } else if (Constraint.size() == 2) {
9644 switch (Constraint[0]) {
9646 // All 'U+' constraints are addresses.
9647 case 'U': return C_Memory;
9650 return TargetLowering::getConstraintType(Constraint);
9653 /// Examine constraint type and operand type and determine a weight value.
9654 /// This object must already have been set up with the operand type
9655 /// and the current alternative constraint selected.
9656 TargetLowering::ConstraintWeight
9657 ARMTargetLowering::getSingleConstraintMatchWeight(
9658 AsmOperandInfo &info, const char *constraint) const {
9659 ConstraintWeight weight = CW_Invalid;
9660 Value *CallOperandVal = info.CallOperandVal;
9661 // If we don't have a value, we can't do a match,
9662 // but allow it at the lowest weight.
9663 if (CallOperandVal == NULL)
9665 Type *type = CallOperandVal->getType();
9666 // Look at the constraint type.
9667 switch (*constraint) {
9669 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9672 if (type->isIntegerTy()) {
9673 if (Subtarget->isThumb())
9674 weight = CW_SpecificReg;
9676 weight = CW_Register;
9680 if (type->isFloatingPointTy())
9681 weight = CW_Register;
9687 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9689 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9691 if (Constraint.size() == 1) {
9692 // GCC ARM Constraint Letters
9693 switch (Constraint[0]) {
9694 case 'l': // Low regs or general regs.
9695 if (Subtarget->isThumb())
9696 return RCPair(0U, &ARM::tGPRRegClass);
9697 return RCPair(0U, &ARM::GPRRegClass);
9698 case 'h': // High regs or no regs.
9699 if (Subtarget->isThumb())
9700 return RCPair(0U, &ARM::hGPRRegClass);
9703 return RCPair(0U, &ARM::GPRRegClass);
9706 return RCPair(0U, &ARM::SPRRegClass);
9707 if (VT.getSizeInBits() == 64)
9708 return RCPair(0U, &ARM::DPRRegClass);
9709 if (VT.getSizeInBits() == 128)
9710 return RCPair(0U, &ARM::QPRRegClass);
9714 return RCPair(0U, &ARM::SPR_8RegClass);
9715 if (VT.getSizeInBits() == 64)
9716 return RCPair(0U, &ARM::DPR_8RegClass);
9717 if (VT.getSizeInBits() == 128)
9718 return RCPair(0U, &ARM::QPR_8RegClass);
9722 return RCPair(0U, &ARM::SPRRegClass);
9726 if (StringRef("{cc}").equals_lower(Constraint))
9727 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9729 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9732 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9733 /// vector. If it is invalid, don't add anything to Ops.
9734 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9735 std::string &Constraint,
9736 std::vector<SDValue>&Ops,
9737 SelectionDAG &DAG) const {
9738 SDValue Result(0, 0);
9740 // Currently only support length 1 constraints.
9741 if (Constraint.length() != 1) return;
9743 char ConstraintLetter = Constraint[0];
9744 switch (ConstraintLetter) {
9747 case 'I': case 'J': case 'K': case 'L':
9748 case 'M': case 'N': case 'O':
9749 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9753 int64_t CVal64 = C->getSExtValue();
9754 int CVal = (int) CVal64;
9755 // None of these constraints allow values larger than 32 bits. Check
9756 // that the value fits in an int.
9760 switch (ConstraintLetter) {
9762 // Constant suitable for movw, must be between 0 and
9764 if (Subtarget->hasV6T2Ops())
9765 if (CVal >= 0 && CVal <= 65535)
9769 if (Subtarget->isThumb1Only()) {
9770 // This must be a constant between 0 and 255, for ADD
9772 if (CVal >= 0 && CVal <= 255)
9774 } else if (Subtarget->isThumb2()) {
9775 // A constant that can be used as an immediate value in a
9776 // data-processing instruction.
9777 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9780 // A constant that can be used as an immediate value in a
9781 // data-processing instruction.
9782 if (ARM_AM::getSOImmVal(CVal) != -1)
9788 if (Subtarget->isThumb()) { // FIXME thumb2
9789 // This must be a constant between -255 and -1, for negated ADD
9790 // immediates. This can be used in GCC with an "n" modifier that
9791 // prints the negated value, for use with SUB instructions. It is
9792 // not useful otherwise but is implemented for compatibility.
9793 if (CVal >= -255 && CVal <= -1)
9796 // This must be a constant between -4095 and 4095. It is not clear
9797 // what this constraint is intended for. Implemented for
9798 // compatibility with GCC.
9799 if (CVal >= -4095 && CVal <= 4095)
9805 if (Subtarget->isThumb1Only()) {
9806 // A 32-bit value where only one byte has a nonzero value. Exclude
9807 // zero to match GCC. This constraint is used by GCC internally for
9808 // constants that can be loaded with a move/shift combination.
9809 // It is not useful otherwise but is implemented for compatibility.
9810 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9812 } else if (Subtarget->isThumb2()) {
9813 // A constant whose bitwise inverse can be used as an immediate
9814 // value in a data-processing instruction. This can be used in GCC
9815 // with a "B" modifier that prints the inverted value, for use with
9816 // BIC and MVN instructions. It is not useful otherwise but is
9817 // implemented for compatibility.
9818 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9821 // A constant whose bitwise inverse can be used as an immediate
9822 // value in a data-processing instruction. This can be used in GCC
9823 // with a "B" modifier that prints the inverted value, for use with
9824 // BIC and MVN instructions. It is not useful otherwise but is
9825 // implemented for compatibility.
9826 if (ARM_AM::getSOImmVal(~CVal) != -1)
9832 if (Subtarget->isThumb1Only()) {
9833 // This must be a constant between -7 and 7,
9834 // for 3-operand ADD/SUB immediate instructions.
9835 if (CVal >= -7 && CVal < 7)
9837 } else if (Subtarget->isThumb2()) {
9838 // A constant whose negation can be used as an immediate value in a
9839 // data-processing instruction. This can be used in GCC with an "n"
9840 // modifier that prints the negated value, for use with SUB
9841 // instructions. It is not useful otherwise but is implemented for
9843 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9846 // A constant whose negation can be used as an immediate value in a
9847 // data-processing instruction. This can be used in GCC with an "n"
9848 // modifier that prints the negated value, for use with SUB
9849 // instructions. It is not useful otherwise but is implemented for
9851 if (ARM_AM::getSOImmVal(-CVal) != -1)
9857 if (Subtarget->isThumb()) { // FIXME thumb2
9858 // This must be a multiple of 4 between 0 and 1020, for
9859 // ADD sp + immediate.
9860 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9863 // A power of two or a constant between 0 and 32. This is used in
9864 // GCC for the shift amount on shifted register operands, but it is
9865 // useful in general for any shift amounts.
9866 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9872 if (Subtarget->isThumb()) { // FIXME thumb2
9873 // This must be a constant between 0 and 31, for shift amounts.
9874 if (CVal >= 0 && CVal <= 31)
9880 if (Subtarget->isThumb()) { // FIXME thumb2
9881 // This must be a multiple of 4 between -508 and 508, for
9882 // ADD/SUB sp = sp + immediate.
9883 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9888 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9892 if (Result.getNode()) {
9893 Ops.push_back(Result);
9896 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9900 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9901 // The ARM target isn't yet aware of offsets.
9905 bool ARM::isBitFieldInvertedMask(unsigned v) {
9906 if (v == 0xffffffff)
9908 // there can be 1's on either or both "outsides", all the "inside"
9910 unsigned int lsb = 0, msb = 31;
9911 while (v & (1 << msb)) --msb;
9912 while (v & (1 << lsb)) ++lsb;
9913 for (unsigned int i = lsb; i <= msb; ++i) {
9920 /// isFPImmLegal - Returns true if the target can instruction select the
9921 /// specified FP immediate natively. If false, the legalizer will
9922 /// materialize the FP immediate as a load from a constant pool.
9923 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9924 if (!Subtarget->hasVFP3())
9927 return ARM_AM::getFP32Imm(Imm) != -1;
9929 return ARM_AM::getFP64Imm(Imm) != -1;
9933 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9934 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9935 /// specified in the intrinsic calls.
9936 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9938 unsigned Intrinsic) const {
9939 switch (Intrinsic) {
9940 case Intrinsic::arm_neon_vld1:
9941 case Intrinsic::arm_neon_vld2:
9942 case Intrinsic::arm_neon_vld3:
9943 case Intrinsic::arm_neon_vld4:
9944 case Intrinsic::arm_neon_vld2lane:
9945 case Intrinsic::arm_neon_vld3lane:
9946 case Intrinsic::arm_neon_vld4lane: {
9947 Info.opc = ISD::INTRINSIC_W_CHAIN;
9948 // Conservatively set memVT to the entire set of vectors loaded.
9949 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9950 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9951 Info.ptrVal = I.getArgOperand(0);
9953 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9954 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9955 Info.vol = false; // volatile loads with NEON intrinsics not supported
9956 Info.readMem = true;
9957 Info.writeMem = false;
9960 case Intrinsic::arm_neon_vst1:
9961 case Intrinsic::arm_neon_vst2:
9962 case Intrinsic::arm_neon_vst3:
9963 case Intrinsic::arm_neon_vst4:
9964 case Intrinsic::arm_neon_vst2lane:
9965 case Intrinsic::arm_neon_vst3lane:
9966 case Intrinsic::arm_neon_vst4lane: {
9967 Info.opc = ISD::INTRINSIC_VOID;
9968 // Conservatively set memVT to the entire set of vectors stored.
9969 unsigned NumElts = 0;
9970 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9971 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9972 if (!ArgTy->isVectorTy())
9974 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9976 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9977 Info.ptrVal = I.getArgOperand(0);
9979 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9980 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9981 Info.vol = false; // volatile stores with NEON intrinsics not supported
9982 Info.readMem = false;
9983 Info.writeMem = true;
9986 case Intrinsic::arm_strexd: {
9987 Info.opc = ISD::INTRINSIC_W_CHAIN;
9988 Info.memVT = MVT::i64;
9989 Info.ptrVal = I.getArgOperand(2);
9993 Info.readMem = false;
9994 Info.writeMem = true;
9997 case Intrinsic::arm_ldrexd: {
9998 Info.opc = ISD::INTRINSIC_W_CHAIN;
9999 Info.memVT = MVT::i64;
10000 Info.ptrVal = I.getArgOperand(0);
10004 Info.readMem = true;
10005 Info.writeMem = false;