1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "MCTargetDesc/ARMAddressingModes.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Target/TargetOptions.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::QPRRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
164 return new TargetLoweringObjectFileMachO();
166 return new ARMElfTargetObjectFile();
169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<ARMSubtarget>();
172 RegInfo = TM.getRegisterInfo();
173 Itins = TM.getInstrItineraryData();
175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
177 if (Subtarget->isTargetDarwin()) {
178 // Uses VFP for Thumb libfuncs if available.
179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
180 // Single-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
186 // Double-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
192 // Single-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
211 // Double-precision comparisons.
212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
230 // Floating-point to integer conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
238 // Conversions between floating types.
239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
242 // Integer to floating-point conversions.
243 // i64 conversions are done via library routines even when generating VFP
244 // instructions, so use the same ones.
245 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
246 // e.g., __floatunsidf vs. __floatunssidfvfp.
247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, 0);
256 setLibcallName(RTLIB::SRL_I128, 0);
257 setLibcallName(RTLIB::SRA_I128, 0);
259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point arithmetic helper functions
299 // RTABI chapter 4.1.2, Table 4
300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
309 // Single-precision floating-point comparison helper functions
310 // RTABI chapter 4.1.2, Table 5
311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
336 // Floating-point to integer conversions.
337 // RTABI chapter 4.1.2, Table 6
338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
355 // Conversions between floating types.
356 // RTABI chapter 4.1.2, Table 7
357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
362 // Integer to floating-point conversions.
363 // RTABI chapter 4.1.2, Table 8
364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 // Long long helper functions
382 // RTABI chapter 4.2, Table 9
383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
394 // Integer division functions
395 // RTABI chapter 4.3.1
396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
414 // RTABI chapter 4.3.4
415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
423 // Use divmod compiler-rt calls for iOS 5.0 and later.
424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430 if (Subtarget->isThumb1Only())
431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
433 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
435 !Subtarget->isThumb1Only()) {
436 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
437 if (!Subtarget->isFPOnlySP())
438 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
447 setTruncStoreAction((MVT::SimpleValueType)VT,
448 (MVT::SimpleValueType)InnerVT, Expand);
449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
456 if (Subtarget->hasNEON()) {
457 addDRTypeForNEON(MVT::v2f32);
458 addDRTypeForNEON(MVT::v8i8);
459 addDRTypeForNEON(MVT::v4i16);
460 addDRTypeForNEON(MVT::v2i32);
461 addDRTypeForNEON(MVT::v1i64);
463 addQRTypeForNEON(MVT::v4f32);
464 addQRTypeForNEON(MVT::v2f64);
465 addQRTypeForNEON(MVT::v16i8);
466 addQRTypeForNEON(MVT::v8i16);
467 addQRTypeForNEON(MVT::v4i32);
468 addQRTypeForNEON(MVT::v2i64);
470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
471 // neither Neon nor VFP support any arithmetic operations on it.
472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
473 // supported for v4f32.
474 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
477 // FIXME: Code duplication: FDIV and FREM are expanded always, see
478 // ARMTargetLowering::addTypeForNEON method for details.
479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
480 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
481 // FIXME: Create unittest.
482 // In another words, find a way when "copysign" appears in DAG with vector
484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
485 // FIXME: Code duplication: SETCC has custom operation action, see
486 // ARMTargetLowering::addTypeForNEON method for details.
487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
488 // FIXME: Create unittest for FNEG and for FABS.
489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
490 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
507 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
525 // Mark v2f32 intrinsics.
526 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
527 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
528 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
529 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
530 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
531 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
532 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
533 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
534 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
535 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
536 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
537 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
538 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
539 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
540 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
542 // Neon does not support some operations on v1i64 and v2i64 types.
543 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
544 // Custom handling for some quad-vector types to detect VMULL.
545 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
546 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
547 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
548 // Custom handling for some vector types to avoid expensive expansions
549 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
550 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
551 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
552 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
553 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
554 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
555 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
556 // a destination type that is wider than the source, and nor does
557 // it have a FP_TO_[SU]INT instruction with a narrower destination than
559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
561 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
564 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
565 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
567 // Custom expand long extensions to vectors.
568 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
569 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
570 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
571 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
572 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
573 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
574 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
575 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
577 // NEON does not have single instruction CTPOP for vectors with element
578 // types wider than 8-bits. However, custom lowering can leverage the
579 // v8i8/v16i8 vcnt instruction.
580 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
581 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
582 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
583 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
585 // NEON only has FMA instructions as of VFP4.
586 if (!Subtarget->hasVFP4()) {
587 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
588 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
591 setTargetDAGCombine(ISD::INTRINSIC_VOID);
592 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
593 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
594 setTargetDAGCombine(ISD::SHL);
595 setTargetDAGCombine(ISD::SRL);
596 setTargetDAGCombine(ISD::SRA);
597 setTargetDAGCombine(ISD::SIGN_EXTEND);
598 setTargetDAGCombine(ISD::ZERO_EXTEND);
599 setTargetDAGCombine(ISD::ANY_EXTEND);
600 setTargetDAGCombine(ISD::SELECT_CC);
601 setTargetDAGCombine(ISD::BUILD_VECTOR);
602 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
603 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
604 setTargetDAGCombine(ISD::STORE);
605 setTargetDAGCombine(ISD::FP_TO_SINT);
606 setTargetDAGCombine(ISD::FP_TO_UINT);
607 setTargetDAGCombine(ISD::FDIV);
609 // It is legal to extload from v4i8 to v4i16 or v4i32.
610 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
611 MVT::v4i16, MVT::v2i16,
613 for (unsigned i = 0; i < 6; ++i) {
614 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
615 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
616 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
620 // ARM and Thumb2 support UMLAL/SMLAL.
621 if (!Subtarget->isThumb1Only())
622 setTargetDAGCombine(ISD::ADDC);
625 computeRegisterProperties();
627 // ARM does not have f32 extending load.
628 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
630 // ARM does not have i1 sign extending load.
631 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
633 // ARM supports all 4 flavors of integer indexed load / store.
634 if (!Subtarget->isThumb1Only()) {
635 for (unsigned im = (unsigned)ISD::PRE_INC;
636 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
637 setIndexedLoadAction(im, MVT::i1, Legal);
638 setIndexedLoadAction(im, MVT::i8, Legal);
639 setIndexedLoadAction(im, MVT::i16, Legal);
640 setIndexedLoadAction(im, MVT::i32, Legal);
641 setIndexedStoreAction(im, MVT::i1, Legal);
642 setIndexedStoreAction(im, MVT::i8, Legal);
643 setIndexedStoreAction(im, MVT::i16, Legal);
644 setIndexedStoreAction(im, MVT::i32, Legal);
648 // i64 operation support.
649 setOperationAction(ISD::MUL, MVT::i64, Expand);
650 setOperationAction(ISD::MULHU, MVT::i32, Expand);
651 if (Subtarget->isThumb1Only()) {
652 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
653 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
655 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
656 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
657 setOperationAction(ISD::MULHS, MVT::i32, Expand);
659 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
660 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
661 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
662 setOperationAction(ISD::SRL, MVT::i64, Custom);
663 setOperationAction(ISD::SRA, MVT::i64, Custom);
665 if (!Subtarget->isThumb1Only()) {
666 // FIXME: We should do this for Thumb1 as well.
667 setOperationAction(ISD::ADDC, MVT::i32, Custom);
668 setOperationAction(ISD::ADDE, MVT::i32, Custom);
669 setOperationAction(ISD::SUBC, MVT::i32, Custom);
670 setOperationAction(ISD::SUBE, MVT::i32, Custom);
673 // ARM does not have ROTL.
674 setOperationAction(ISD::ROTL, MVT::i32, Expand);
675 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
676 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
677 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
678 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
680 // These just redirect to CTTZ and CTLZ on ARM.
681 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
682 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
684 // Only ARMv6 has BSWAP.
685 if (!Subtarget->hasV6Ops())
686 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
688 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
689 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
690 // These are expanded into libcalls if the cpu doesn't have HW divider.
691 setOperationAction(ISD::SDIV, MVT::i32, Expand);
692 setOperationAction(ISD::UDIV, MVT::i32, Expand);
694 setOperationAction(ISD::SREM, MVT::i32, Expand);
695 setOperationAction(ISD::UREM, MVT::i32, Expand);
696 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
697 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
699 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
700 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
701 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
702 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
703 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
705 setOperationAction(ISD::TRAP, MVT::Other, Legal);
707 // Use the default implementation.
708 setOperationAction(ISD::VASTART, MVT::Other, Custom);
709 setOperationAction(ISD::VAARG, MVT::Other, Expand);
710 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
711 setOperationAction(ISD::VAEND, MVT::Other, Expand);
712 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
713 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
715 if (!Subtarget->isTargetDarwin()) {
716 // Non-Darwin platforms may return values in these registers via the
717 // personality function.
718 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
719 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
720 setExceptionPointerRegister(ARM::R0);
721 setExceptionSelectorRegister(ARM::R1);
724 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
725 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
726 // the default expansion.
727 // FIXME: This should be checking for v6k, not just v6.
728 if (Subtarget->hasDataBarrier() ||
729 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
730 // membarrier needs custom lowering; the rest are legal and handled
732 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
733 // Custom lowering for 64-bit ops
734 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
735 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
736 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
737 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
738 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
739 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
740 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
741 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
742 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
743 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
744 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
745 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
746 setInsertFencesForAtomic(true);
748 // Set them all for expansion, which will force libcalls.
749 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
750 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
751 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
752 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
753 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
754 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
755 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
756 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
757 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
758 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
759 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
760 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
761 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
762 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
763 // Unordered/Monotonic case.
764 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
765 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
768 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
770 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
771 if (!Subtarget->hasV6Ops()) {
772 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
773 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
775 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
777 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
778 !Subtarget->isThumb1Only()) {
779 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
780 // iff target supports vfp2.
781 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
782 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
785 // We want to custom lower some of our intrinsics.
786 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
787 if (Subtarget->isTargetDarwin()) {
788 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
789 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
790 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
793 setOperationAction(ISD::SETCC, MVT::i32, Expand);
794 setOperationAction(ISD::SETCC, MVT::f32, Expand);
795 setOperationAction(ISD::SETCC, MVT::f64, Expand);
796 setOperationAction(ISD::SELECT, MVT::i32, Custom);
797 setOperationAction(ISD::SELECT, MVT::f32, Custom);
798 setOperationAction(ISD::SELECT, MVT::f64, Custom);
799 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
800 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
801 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
803 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
804 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
805 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
806 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
807 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
809 // We don't support sin/cos/fmod/copysign/pow
810 setOperationAction(ISD::FSIN, MVT::f64, Expand);
811 setOperationAction(ISD::FSIN, MVT::f32, Expand);
812 setOperationAction(ISD::FCOS, MVT::f32, Expand);
813 setOperationAction(ISD::FCOS, MVT::f64, Expand);
814 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
815 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
816 setOperationAction(ISD::FREM, MVT::f64, Expand);
817 setOperationAction(ISD::FREM, MVT::f32, Expand);
818 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
819 !Subtarget->isThumb1Only()) {
820 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
821 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
823 setOperationAction(ISD::FPOW, MVT::f64, Expand);
824 setOperationAction(ISD::FPOW, MVT::f32, Expand);
826 if (!Subtarget->hasVFP4()) {
827 setOperationAction(ISD::FMA, MVT::f64, Expand);
828 setOperationAction(ISD::FMA, MVT::f32, Expand);
831 // Various VFP goodness
832 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
833 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
834 if (Subtarget->hasVFP2()) {
835 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
836 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
837 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
838 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
840 // Special handling for half-precision FP.
841 if (!Subtarget->hasFP16()) {
842 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
843 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
847 // We have target-specific dag combine patterns for the following nodes:
848 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
849 setTargetDAGCombine(ISD::ADD);
850 setTargetDAGCombine(ISD::SUB);
851 setTargetDAGCombine(ISD::MUL);
852 setTargetDAGCombine(ISD::AND);
853 setTargetDAGCombine(ISD::OR);
854 setTargetDAGCombine(ISD::XOR);
856 if (Subtarget->hasV6Ops())
857 setTargetDAGCombine(ISD::SRL);
859 setStackPointerRegisterToSaveRestore(ARM::SP);
861 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
862 !Subtarget->hasVFP2())
863 setSchedulingPreference(Sched::RegPressure);
865 setSchedulingPreference(Sched::Hybrid);
867 //// temporary - rewrite interface to use type
868 MaxStoresPerMemset = 8;
869 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
870 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
871 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
872 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
873 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
875 // On ARM arguments smaller than 4 bytes are extended, so all arguments
876 // are at least 4 bytes aligned.
877 setMinStackArgumentAlignment(4);
879 // Prefer likely predicted branches to selects on out-of-order cores.
880 PredictableSelectIsExpensive = Subtarget->isLikeA9();
882 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
885 // FIXME: It might make sense to define the representative register class as the
886 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
887 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
888 // SPR's representative would be DPR_VFP2. This should work well if register
889 // pressure tracking were modified such that a register use would increment the
890 // pressure of the register class's representative and all of it's super
891 // classes' representatives transitively. We have not implemented this because
892 // of the difficulty prior to coalescing of modeling operand register classes
893 // due to the common occurrence of cross class copies and subregister insertions
895 std::pair<const TargetRegisterClass*, uint8_t>
896 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
897 const TargetRegisterClass *RRC = 0;
899 switch (VT.SimpleTy) {
901 return TargetLowering::findRepresentativeClass(VT);
902 // Use DPR as representative register class for all floating point
903 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
904 // the cost is 1 for both f32 and f64.
905 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
906 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
907 RRC = &ARM::DPRRegClass;
908 // When NEON is used for SP, only half of the register file is available
909 // because operations that define both SP and DP results will be constrained
910 // to the VFP2 class (D0-D15). We currently model this constraint prior to
911 // coalescing by double-counting the SP regs. See the FIXME above.
912 if (Subtarget->useNEONForSinglePrecisionFP())
915 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
916 case MVT::v4f32: case MVT::v2f64:
917 RRC = &ARM::DPRRegClass;
921 RRC = &ARM::DPRRegClass;
925 RRC = &ARM::DPRRegClass;
929 return std::make_pair(RRC, Cost);
932 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
935 case ARMISD::Wrapper: return "ARMISD::Wrapper";
936 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
937 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
938 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
939 case ARMISD::CALL: return "ARMISD::CALL";
940 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
941 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
942 case ARMISD::tCALL: return "ARMISD::tCALL";
943 case ARMISD::BRCOND: return "ARMISD::BRCOND";
944 case ARMISD::BR_JT: return "ARMISD::BR_JT";
945 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
946 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
947 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
948 case ARMISD::CMP: return "ARMISD::CMP";
949 case ARMISD::CMN: return "ARMISD::CMN";
950 case ARMISD::CMPZ: return "ARMISD::CMPZ";
951 case ARMISD::CMPFP: return "ARMISD::CMPFP";
952 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
953 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
954 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
956 case ARMISD::CMOV: return "ARMISD::CMOV";
958 case ARMISD::RBIT: return "ARMISD::RBIT";
960 case ARMISD::FTOSI: return "ARMISD::FTOSI";
961 case ARMISD::FTOUI: return "ARMISD::FTOUI";
962 case ARMISD::SITOF: return "ARMISD::SITOF";
963 case ARMISD::UITOF: return "ARMISD::UITOF";
965 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
966 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
967 case ARMISD::RRX: return "ARMISD::RRX";
969 case ARMISD::ADDC: return "ARMISD::ADDC";
970 case ARMISD::ADDE: return "ARMISD::ADDE";
971 case ARMISD::SUBC: return "ARMISD::SUBC";
972 case ARMISD::SUBE: return "ARMISD::SUBE";
974 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
975 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
977 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
978 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
980 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
982 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
984 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
986 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
987 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
989 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
991 case ARMISD::VCEQ: return "ARMISD::VCEQ";
992 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
993 case ARMISD::VCGE: return "ARMISD::VCGE";
994 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
995 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
996 case ARMISD::VCGEU: return "ARMISD::VCGEU";
997 case ARMISD::VCGT: return "ARMISD::VCGT";
998 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
999 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1000 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1001 case ARMISD::VTST: return "ARMISD::VTST";
1003 case ARMISD::VSHL: return "ARMISD::VSHL";
1004 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1005 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1006 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1007 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1008 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1009 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1010 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1011 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1012 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1013 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1014 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1015 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1016 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1017 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1018 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1019 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1020 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1021 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1022 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1023 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1024 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1025 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1026 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1027 case ARMISD::VDUP: return "ARMISD::VDUP";
1028 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1029 case ARMISD::VEXT: return "ARMISD::VEXT";
1030 case ARMISD::VREV64: return "ARMISD::VREV64";
1031 case ARMISD::VREV32: return "ARMISD::VREV32";
1032 case ARMISD::VREV16: return "ARMISD::VREV16";
1033 case ARMISD::VZIP: return "ARMISD::VZIP";
1034 case ARMISD::VUZP: return "ARMISD::VUZP";
1035 case ARMISD::VTRN: return "ARMISD::VTRN";
1036 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1037 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1038 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1039 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1040 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1041 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1042 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1043 case ARMISD::FMAX: return "ARMISD::FMAX";
1044 case ARMISD::FMIN: return "ARMISD::FMIN";
1045 case ARMISD::BFI: return "ARMISD::BFI";
1046 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1047 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1048 case ARMISD::VBSL: return "ARMISD::VBSL";
1049 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1050 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1051 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1052 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1053 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1054 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1055 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1056 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1057 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1058 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1059 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1060 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1061 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1062 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1063 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1064 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1065 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1066 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1067 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1068 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1072 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1073 if (!VT.isVector()) return getPointerTy();
1074 return VT.changeVectorElementTypeToInteger();
1077 /// getRegClassFor - Return the register class that should be used for the
1078 /// specified value type.
1079 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1080 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1081 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1082 // load / store 4 to 8 consecutive D registers.
1083 if (Subtarget->hasNEON()) {
1084 if (VT == MVT::v4i64)
1085 return &ARM::QQPRRegClass;
1086 if (VT == MVT::v8i64)
1087 return &ARM::QQQQPRRegClass;
1089 return TargetLowering::getRegClassFor(VT);
1092 // Create a fast isel object.
1094 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1095 const TargetLibraryInfo *libInfo) const {
1096 return ARM::createFastISel(funcInfo, libInfo);
1099 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1100 /// be used for loads / stores from the global.
1101 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1102 return (Subtarget->isThumb1Only() ? 127 : 4095);
1105 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1106 unsigned NumVals = N->getNumValues();
1108 return Sched::RegPressure;
1110 for (unsigned i = 0; i != NumVals; ++i) {
1111 EVT VT = N->getValueType(i);
1112 if (VT == MVT::Glue || VT == MVT::Other)
1114 if (VT.isFloatingPoint() || VT.isVector())
1118 if (!N->isMachineOpcode())
1119 return Sched::RegPressure;
1121 // Load are scheduled for latency even if there instruction itinerary
1122 // is not available.
1123 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1124 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1126 if (MCID.getNumDefs() == 0)
1127 return Sched::RegPressure;
1128 if (!Itins->isEmpty() &&
1129 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1132 return Sched::RegPressure;
1135 //===----------------------------------------------------------------------===//
1137 //===----------------------------------------------------------------------===//
1139 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1140 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1142 default: llvm_unreachable("Unknown condition code!");
1143 case ISD::SETNE: return ARMCC::NE;
1144 case ISD::SETEQ: return ARMCC::EQ;
1145 case ISD::SETGT: return ARMCC::GT;
1146 case ISD::SETGE: return ARMCC::GE;
1147 case ISD::SETLT: return ARMCC::LT;
1148 case ISD::SETLE: return ARMCC::LE;
1149 case ISD::SETUGT: return ARMCC::HI;
1150 case ISD::SETUGE: return ARMCC::HS;
1151 case ISD::SETULT: return ARMCC::LO;
1152 case ISD::SETULE: return ARMCC::LS;
1156 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1157 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1158 ARMCC::CondCodes &CondCode2) {
1159 CondCode2 = ARMCC::AL;
1161 default: llvm_unreachable("Unknown FP condition!");
1163 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1165 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1167 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1168 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1169 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1170 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1171 case ISD::SETO: CondCode = ARMCC::VC; break;
1172 case ISD::SETUO: CondCode = ARMCC::VS; break;
1173 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1174 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1175 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1177 case ISD::SETULT: CondCode = ARMCC::LT; break;
1179 case ISD::SETULE: CondCode = ARMCC::LE; break;
1181 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1185 //===----------------------------------------------------------------------===//
1186 // Calling Convention Implementation
1187 //===----------------------------------------------------------------------===//
1189 #include "ARMGenCallingConv.inc"
1191 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1192 /// given CallingConvention value.
1193 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1195 bool isVarArg) const {
1198 llvm_unreachable("Unsupported calling convention");
1199 case CallingConv::Fast:
1200 if (Subtarget->hasVFP2() && !isVarArg) {
1201 if (!Subtarget->isAAPCS_ABI())
1202 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1203 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1204 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1207 case CallingConv::C: {
1208 // Use target triple & subtarget features to do actual dispatch.
1209 if (!Subtarget->isAAPCS_ABI())
1210 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1211 else if (Subtarget->hasVFP2() &&
1212 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1214 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1215 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1217 case CallingConv::ARM_AAPCS_VFP:
1219 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1221 case CallingConv::ARM_AAPCS:
1222 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1223 case CallingConv::ARM_APCS:
1224 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1225 case CallingConv::GHC:
1226 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1230 /// LowerCallResult - Lower the result values of a call into the
1231 /// appropriate copies out of appropriate physical registers.
1233 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1234 CallingConv::ID CallConv, bool isVarArg,
1235 const SmallVectorImpl<ISD::InputArg> &Ins,
1236 DebugLoc dl, SelectionDAG &DAG,
1237 SmallVectorImpl<SDValue> &InVals,
1238 bool isThisReturn, SDValue ThisVal) const {
1240 // Assign locations to each value returned by this call.
1241 SmallVector<CCValAssign, 16> RVLocs;
1242 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1243 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1244 CCInfo.AnalyzeCallResult(Ins,
1245 CCAssignFnForNode(CallConv, /* Return*/ true,
1248 // Copy all of the result registers out of their specified physreg.
1249 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1250 CCValAssign VA = RVLocs[i];
1252 // Pass 'this' value directly from the argument to return value, to avoid
1253 // reg unit interference
1254 if (i == 0 && isThisReturn) {
1255 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32);
1256 InVals.push_back(ThisVal);
1261 if (VA.needsCustom()) {
1262 // Handle f64 or half of a v2f64.
1263 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1265 Chain = Lo.getValue(1);
1266 InFlag = Lo.getValue(2);
1267 VA = RVLocs[++i]; // skip ahead to next loc
1268 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1270 Chain = Hi.getValue(1);
1271 InFlag = Hi.getValue(2);
1272 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1274 if (VA.getLocVT() == MVT::v2f64) {
1275 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1276 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1277 DAG.getConstant(0, MVT::i32));
1279 VA = RVLocs[++i]; // skip ahead to next loc
1280 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1281 Chain = Lo.getValue(1);
1282 InFlag = Lo.getValue(2);
1283 VA = RVLocs[++i]; // skip ahead to next loc
1284 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1285 Chain = Hi.getValue(1);
1286 InFlag = Hi.getValue(2);
1287 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1288 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1289 DAG.getConstant(1, MVT::i32));
1292 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1294 Chain = Val.getValue(1);
1295 InFlag = Val.getValue(2);
1298 switch (VA.getLocInfo()) {
1299 default: llvm_unreachable("Unknown loc info!");
1300 case CCValAssign::Full: break;
1301 case CCValAssign::BCvt:
1302 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1306 InVals.push_back(Val);
1312 /// LowerMemOpCallTo - Store the argument to the stack.
1314 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1315 SDValue StackPtr, SDValue Arg,
1316 DebugLoc dl, SelectionDAG &DAG,
1317 const CCValAssign &VA,
1318 ISD::ArgFlagsTy Flags) const {
1319 unsigned LocMemOffset = VA.getLocMemOffset();
1320 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1321 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1322 return DAG.getStore(Chain, dl, Arg, PtrOff,
1323 MachinePointerInfo::getStack(LocMemOffset),
1327 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1328 SDValue Chain, SDValue &Arg,
1329 RegsToPassVector &RegsToPass,
1330 CCValAssign &VA, CCValAssign &NextVA,
1332 SmallVector<SDValue, 8> &MemOpChains,
1333 ISD::ArgFlagsTy Flags) const {
1335 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1336 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1337 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1339 if (NextVA.isRegLoc())
1340 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1342 assert(NextVA.isMemLoc());
1343 if (StackPtr.getNode() == 0)
1344 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1346 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1352 /// LowerCall - Lowering a call into a callseq_start <-
1353 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1356 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1357 SmallVectorImpl<SDValue> &InVals) const {
1358 SelectionDAG &DAG = CLI.DAG;
1359 DebugLoc &dl = CLI.DL;
1360 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1361 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1362 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1363 SDValue Chain = CLI.Chain;
1364 SDValue Callee = CLI.Callee;
1365 bool &isTailCall = CLI.IsTailCall;
1366 CallingConv::ID CallConv = CLI.CallConv;
1367 bool doesNotRet = CLI.DoesNotReturn;
1368 bool isVarArg = CLI.IsVarArg;
1370 MachineFunction &MF = DAG.getMachineFunction();
1371 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1372 bool IsThisReturn = false;
1373 bool IsSibCall = false;
1374 // Disable tail calls if they're not supported.
1375 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1378 // Check if it's really possible to do a tail call.
1379 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1380 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1381 Outs, OutVals, Ins, DAG);
1382 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1383 // detected sibcalls.
1390 // Analyze operands of the call, assigning locations to each operand.
1391 SmallVector<CCValAssign, 16> ArgLocs;
1392 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1393 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1394 CCInfo.AnalyzeCallOperands(Outs,
1395 CCAssignFnForNode(CallConv, /* Return*/ false,
1398 // Get a count of how many bytes are to be pushed on the stack.
1399 unsigned NumBytes = CCInfo.getNextStackOffset();
1401 // For tail calls, memory operands are available in our caller's stack.
1405 // Adjust the stack pointer for the new arguments...
1406 // These operations are automatically eliminated by the prolog/epilog pass
1408 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1410 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1412 RegsToPassVector RegsToPass;
1413 SmallVector<SDValue, 8> MemOpChains;
1415 // Walk the register/memloc assignments, inserting copies/loads. In the case
1416 // of tail call optimization, arguments are handled later.
1417 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1419 ++i, ++realArgIdx) {
1420 CCValAssign &VA = ArgLocs[i];
1421 SDValue Arg = OutVals[realArgIdx];
1422 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1423 bool isByVal = Flags.isByVal();
1425 // Promote the value if needed.
1426 switch (VA.getLocInfo()) {
1427 default: llvm_unreachable("Unknown loc info!");
1428 case CCValAssign::Full: break;
1429 case CCValAssign::SExt:
1430 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1432 case CCValAssign::ZExt:
1433 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1435 case CCValAssign::AExt:
1436 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1438 case CCValAssign::BCvt:
1439 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1443 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1444 if (VA.needsCustom()) {
1445 if (VA.getLocVT() == MVT::v2f64) {
1446 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1447 DAG.getConstant(0, MVT::i32));
1448 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1449 DAG.getConstant(1, MVT::i32));
1451 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1452 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1454 VA = ArgLocs[++i]; // skip ahead to next loc
1455 if (VA.isRegLoc()) {
1456 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1457 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1459 assert(VA.isMemLoc());
1461 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1462 dl, DAG, VA, Flags));
1465 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1466 StackPtr, MemOpChains, Flags);
1468 } else if (VA.isRegLoc()) {
1469 if (realArgIdx == 0 && Flags.isReturned() && VA.getLocVT() == MVT::i32) {
1470 assert(!Ins.empty() && Ins[0].VT == Outs[0].VT &&
1471 "unexpected use of 'returned'");
1472 IsThisReturn = true;
1474 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1475 } else if (isByVal) {
1476 assert(VA.isMemLoc());
1477 unsigned offset = 0;
1479 // True if this byval aggregate will be split between registers
1481 if (CCInfo.isFirstByValRegValid()) {
1482 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1484 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1485 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1486 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1487 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1488 MachinePointerInfo(),
1489 false, false, false, 0);
1490 MemOpChains.push_back(Load.getValue(1));
1491 RegsToPass.push_back(std::make_pair(j, Load));
1493 offset = ARM::R4 - CCInfo.getFirstByValReg();
1494 CCInfo.clearFirstByValReg();
1497 if (Flags.getByValSize() - 4*offset > 0) {
1498 unsigned LocMemOffset = VA.getLocMemOffset();
1499 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1500 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1502 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1503 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1504 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1506 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1508 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1509 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1510 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1511 Ops, array_lengthof(Ops)));
1513 } else if (!IsSibCall) {
1514 assert(VA.isMemLoc());
1516 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1517 dl, DAG, VA, Flags));
1521 if (!MemOpChains.empty())
1522 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1523 &MemOpChains[0], MemOpChains.size());
1525 // Build a sequence of copy-to-reg nodes chained together with token chain
1526 // and flag operands which copy the outgoing args into the appropriate regs.
1528 // Tail call byval lowering might overwrite argument registers so in case of
1529 // tail call optimization the copies to registers are lowered later.
1531 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1532 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1533 RegsToPass[i].second, InFlag);
1534 InFlag = Chain.getValue(1);
1537 // For tail calls lower the arguments to the 'real' stack slot.
1539 // Force all the incoming stack arguments to be loaded from the stack
1540 // before any new outgoing arguments are stored to the stack, because the
1541 // outgoing stack slots may alias the incoming argument stack slots, and
1542 // the alias isn't otherwise explicit. This is slightly more conservative
1543 // than necessary, because it means that each store effectively depends
1544 // on every argument instead of just those arguments it would clobber.
1546 // Do not flag preceding copytoreg stuff together with the following stuff.
1548 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1549 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1550 RegsToPass[i].second, InFlag);
1551 InFlag = Chain.getValue(1);
1556 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1557 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1558 // node so that legalize doesn't hack it.
1559 bool isDirect = false;
1560 bool isARMFunc = false;
1561 bool isLocalARMFunc = false;
1562 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1564 if (EnableARMLongCalls) {
1565 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1566 && "long-calls with non-static relocation model!");
1567 // Handle a global address or an external symbol. If it's not one of
1568 // those, the target's already in a register, so we don't need to do
1570 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1571 const GlobalValue *GV = G->getGlobal();
1572 // Create a constant pool entry for the callee address
1573 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1574 ARMConstantPoolValue *CPV =
1575 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1577 // Get the address of the callee into a register
1578 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1579 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1580 Callee = DAG.getLoad(getPointerTy(), dl,
1581 DAG.getEntryNode(), CPAddr,
1582 MachinePointerInfo::getConstantPool(),
1583 false, false, false, 0);
1584 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1585 const char *Sym = S->getSymbol();
1587 // Create a constant pool entry for the callee address
1588 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1589 ARMConstantPoolValue *CPV =
1590 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1591 ARMPCLabelIndex, 0);
1592 // Get the address of the callee into a register
1593 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1594 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1595 Callee = DAG.getLoad(getPointerTy(), dl,
1596 DAG.getEntryNode(), CPAddr,
1597 MachinePointerInfo::getConstantPool(),
1598 false, false, false, 0);
1600 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1601 const GlobalValue *GV = G->getGlobal();
1603 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1604 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1605 getTargetMachine().getRelocationModel() != Reloc::Static;
1606 isARMFunc = !Subtarget->isThumb() || isStub;
1607 // ARM call to a local ARM function is predicable.
1608 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1609 // tBX takes a register source operand.
1610 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1611 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1612 ARMConstantPoolValue *CPV =
1613 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1614 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1615 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1616 Callee = DAG.getLoad(getPointerTy(), dl,
1617 DAG.getEntryNode(), CPAddr,
1618 MachinePointerInfo::getConstantPool(),
1619 false, false, false, 0);
1620 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1621 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1622 getPointerTy(), Callee, PICLabel);
1624 // On ELF targets for PIC code, direct calls should go through the PLT
1625 unsigned OpFlags = 0;
1626 if (Subtarget->isTargetELF() &&
1627 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1628 OpFlags = ARMII::MO_PLT;
1629 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1631 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1633 bool isStub = Subtarget->isTargetDarwin() &&
1634 getTargetMachine().getRelocationModel() != Reloc::Static;
1635 isARMFunc = !Subtarget->isThumb() || isStub;
1636 // tBX takes a register source operand.
1637 const char *Sym = S->getSymbol();
1638 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1639 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1640 ARMConstantPoolValue *CPV =
1641 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1642 ARMPCLabelIndex, 4);
1643 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1644 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1645 Callee = DAG.getLoad(getPointerTy(), dl,
1646 DAG.getEntryNode(), CPAddr,
1647 MachinePointerInfo::getConstantPool(),
1648 false, false, false, 0);
1649 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1650 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1651 getPointerTy(), Callee, PICLabel);
1653 unsigned OpFlags = 0;
1654 // On ELF targets for PIC code, direct calls should go through the PLT
1655 if (Subtarget->isTargetELF() &&
1656 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1657 OpFlags = ARMII::MO_PLT;
1658 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1662 // FIXME: handle tail calls differently.
1664 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1665 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1666 if (Subtarget->isThumb()) {
1667 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1668 CallOpc = ARMISD::CALL_NOLINK;
1670 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1672 if (!isDirect && !Subtarget->hasV5TOps())
1673 CallOpc = ARMISD::CALL_NOLINK;
1674 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1675 // Emit regular call when code size is the priority
1677 // "mov lr, pc; b _foo" to avoid confusing the RSP
1678 CallOpc = ARMISD::CALL_NOLINK;
1680 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1683 std::vector<SDValue> Ops;
1684 Ops.push_back(Chain);
1685 Ops.push_back(Callee);
1687 // Add argument registers to the end of the list so that they are known live
1689 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1690 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1691 RegsToPass[i].second.getValueType()));
1693 // Add a register mask operand representing the call-preserved registers.
1694 const uint32_t *Mask;
1695 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1696 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1698 // For 'this' returns, use the R0-preserving mask
1699 Mask = ARI->getThisReturnPreservedMask(CallConv);
1701 Mask = ARI->getCallPreservedMask(CallConv);
1703 assert(Mask && "Missing call preserved mask for calling convention");
1704 Ops.push_back(DAG.getRegisterMask(Mask));
1706 if (InFlag.getNode())
1707 Ops.push_back(InFlag);
1709 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1711 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1713 // Returns a chain and a flag for retval copy to use.
1714 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1715 InFlag = Chain.getValue(1);
1717 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1718 DAG.getIntPtrConstant(0, true), InFlag);
1720 InFlag = Chain.getValue(1);
1722 // Handle result values, copying them out of physregs into vregs that we
1724 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1725 InVals, IsThisReturn,
1726 IsThisReturn ? OutVals[0] : SDValue());
1729 /// HandleByVal - Every parameter *after* a byval parameter is passed
1730 /// on the stack. Remember the next parameter register to allocate,
1731 /// and then confiscate the rest of the parameter registers to insure
1734 ARMTargetLowering::HandleByVal(
1735 CCState *State, unsigned &size, unsigned Align) const {
1736 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1737 assert((State->getCallOrPrologue() == Prologue ||
1738 State->getCallOrPrologue() == Call) &&
1739 "unhandled ParmContext");
1740 if ((!State->isFirstByValRegValid()) &&
1741 (!Subtarget->isAAPCS_ABI() || State->getNextStackOffset() == 0) &&
1742 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1743 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1744 unsigned AlignInRegs = Align / 4;
1745 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1746 for (unsigned i = 0; i < Waste; ++i)
1747 reg = State->AllocateReg(GPRArgRegs, 4);
1750 State->setFirstByValReg(reg);
1751 // At a call site, a byval parameter that is split between
1752 // registers and memory needs its size truncated here. In a
1753 // function prologue, such byval parameters are reassembled in
1754 // memory, and are not truncated.
1755 if (State->getCallOrPrologue() == Call) {
1756 unsigned excess = 4 * (ARM::R4 - reg);
1757 assert(size >= excess && "expected larger existing stack allocation");
1762 // Confiscate any remaining parameter registers to preclude their
1763 // assignment to subsequent parameters.
1764 while (State->AllocateReg(GPRArgRegs, 4))
1768 /// MatchingStackOffset - Return true if the given stack call argument is
1769 /// already available in the same position (relatively) of the caller's
1770 /// incoming argument stack.
1772 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1773 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1774 const TargetInstrInfo *TII) {
1775 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1777 if (Arg.getOpcode() == ISD::CopyFromReg) {
1778 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1779 if (!TargetRegisterInfo::isVirtualRegister(VR))
1781 MachineInstr *Def = MRI->getVRegDef(VR);
1784 if (!Flags.isByVal()) {
1785 if (!TII->isLoadFromStackSlot(Def, FI))
1790 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1791 if (Flags.isByVal())
1792 // ByVal argument is passed in as a pointer but it's now being
1793 // dereferenced. e.g.
1794 // define @foo(%struct.X* %A) {
1795 // tail call @bar(%struct.X* byval %A)
1798 SDValue Ptr = Ld->getBasePtr();
1799 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1802 FI = FINode->getIndex();
1806 assert(FI != INT_MAX);
1807 if (!MFI->isFixedObjectIndex(FI))
1809 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1812 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1813 /// for tail call optimization. Targets which want to do tail call
1814 /// optimization should implement this function.
1816 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1817 CallingConv::ID CalleeCC,
1819 bool isCalleeStructRet,
1820 bool isCallerStructRet,
1821 const SmallVectorImpl<ISD::OutputArg> &Outs,
1822 const SmallVectorImpl<SDValue> &OutVals,
1823 const SmallVectorImpl<ISD::InputArg> &Ins,
1824 SelectionDAG& DAG) const {
1825 const Function *CallerF = DAG.getMachineFunction().getFunction();
1826 CallingConv::ID CallerCC = CallerF->getCallingConv();
1827 bool CCMatch = CallerCC == CalleeCC;
1829 // Look for obvious safe cases to perform tail call optimization that do not
1830 // require ABI changes. This is what gcc calls sibcall.
1832 // Do not sibcall optimize vararg calls unless the call site is not passing
1834 if (isVarArg && !Outs.empty())
1837 // Also avoid sibcall optimization if either caller or callee uses struct
1838 // return semantics.
1839 if (isCalleeStructRet || isCallerStructRet)
1842 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1843 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1844 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1845 // support in the assembler and linker to be used. This would need to be
1846 // fixed to fully support tail calls in Thumb1.
1848 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1849 // LR. This means if we need to reload LR, it takes an extra instructions,
1850 // which outweighs the value of the tail call; but here we don't know yet
1851 // whether LR is going to be used. Probably the right approach is to
1852 // generate the tail call here and turn it back into CALL/RET in
1853 // emitEpilogue if LR is used.
1855 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1856 // but we need to make sure there are enough registers; the only valid
1857 // registers are the 4 used for parameters. We don't currently do this
1859 if (Subtarget->isThumb1Only())
1862 // If the calling conventions do not match, then we'd better make sure the
1863 // results are returned in the same way as what the caller expects.
1865 SmallVector<CCValAssign, 16> RVLocs1;
1866 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1867 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1868 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1870 SmallVector<CCValAssign, 16> RVLocs2;
1871 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1872 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1873 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1875 if (RVLocs1.size() != RVLocs2.size())
1877 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1878 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1880 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1882 if (RVLocs1[i].isRegLoc()) {
1883 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1886 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1892 // If Caller's vararg or byval argument has been split between registers and
1893 // stack, do not perform tail call, since part of the argument is in caller's
1895 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1896 getInfo<ARMFunctionInfo>();
1897 if (AFI_Caller->getVarArgsRegSaveSize())
1900 // If the callee takes no arguments then go on to check the results of the
1902 if (!Outs.empty()) {
1903 // Check if stack adjustment is needed. For now, do not do this if any
1904 // argument is passed on the stack.
1905 SmallVector<CCValAssign, 16> ArgLocs;
1906 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1907 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1908 CCInfo.AnalyzeCallOperands(Outs,
1909 CCAssignFnForNode(CalleeCC, false, isVarArg));
1910 if (CCInfo.getNextStackOffset()) {
1911 MachineFunction &MF = DAG.getMachineFunction();
1913 // Check if the arguments are already laid out in the right way as
1914 // the caller's fixed stack objects.
1915 MachineFrameInfo *MFI = MF.getFrameInfo();
1916 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1917 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1918 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1920 ++i, ++realArgIdx) {
1921 CCValAssign &VA = ArgLocs[i];
1922 EVT RegVT = VA.getLocVT();
1923 SDValue Arg = OutVals[realArgIdx];
1924 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1925 if (VA.getLocInfo() == CCValAssign::Indirect)
1927 if (VA.needsCustom()) {
1928 // f64 and vector types are split into multiple registers or
1929 // register/stack-slot combinations. The types will not match
1930 // the registers; give up on memory f64 refs until we figure
1931 // out what to do about this.
1934 if (!ArgLocs[++i].isRegLoc())
1936 if (RegVT == MVT::v2f64) {
1937 if (!ArgLocs[++i].isRegLoc())
1939 if (!ArgLocs[++i].isRegLoc())
1942 } else if (!VA.isRegLoc()) {
1943 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1955 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1956 MachineFunction &MF, bool isVarArg,
1957 const SmallVectorImpl<ISD::OutputArg> &Outs,
1958 LLVMContext &Context) const {
1959 SmallVector<CCValAssign, 16> RVLocs;
1960 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
1961 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
1966 ARMTargetLowering::LowerReturn(SDValue Chain,
1967 CallingConv::ID CallConv, bool isVarArg,
1968 const SmallVectorImpl<ISD::OutputArg> &Outs,
1969 const SmallVectorImpl<SDValue> &OutVals,
1970 DebugLoc dl, SelectionDAG &DAG) const {
1972 // CCValAssign - represent the assignment of the return value to a location.
1973 SmallVector<CCValAssign, 16> RVLocs;
1975 // CCState - Info about the registers and stack slots.
1976 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1977 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1979 // Analyze outgoing return values.
1980 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1984 SmallVector<SDValue, 4> RetOps;
1985 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1987 // Copy the result values into the output registers.
1988 for (unsigned i = 0, realRVLocIdx = 0;
1990 ++i, ++realRVLocIdx) {
1991 CCValAssign &VA = RVLocs[i];
1992 assert(VA.isRegLoc() && "Can only return in registers!");
1994 SDValue Arg = OutVals[realRVLocIdx];
1996 switch (VA.getLocInfo()) {
1997 default: llvm_unreachable("Unknown loc info!");
1998 case CCValAssign::Full: break;
1999 case CCValAssign::BCvt:
2000 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2004 if (VA.needsCustom()) {
2005 if (VA.getLocVT() == MVT::v2f64) {
2006 // Extract the first half and return it in two registers.
2007 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2008 DAG.getConstant(0, MVT::i32));
2009 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2010 DAG.getVTList(MVT::i32, MVT::i32), Half);
2012 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2013 Flag = Chain.getValue(1);
2014 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2015 VA = RVLocs[++i]; // skip ahead to next loc
2016 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2017 HalfGPRs.getValue(1), Flag);
2018 Flag = Chain.getValue(1);
2019 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2020 VA = RVLocs[++i]; // skip ahead to next loc
2022 // Extract the 2nd half and fall through to handle it as an f64 value.
2023 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2024 DAG.getConstant(1, MVT::i32));
2026 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2028 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2029 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2030 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2031 Flag = Chain.getValue(1);
2032 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2033 VA = RVLocs[++i]; // skip ahead to next loc
2034 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2037 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2039 // Guarantee that all emitted copies are
2040 // stuck together, avoiding something bad.
2041 Flag = Chain.getValue(1);
2042 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2045 // Update chain and glue.
2048 RetOps.push_back(Flag);
2050 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2051 RetOps.data(), RetOps.size());
2054 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2055 if (N->getNumValues() != 1)
2057 if (!N->hasNUsesOfValue(1, 0))
2060 SDValue TCChain = Chain;
2061 SDNode *Copy = *N->use_begin();
2062 if (Copy->getOpcode() == ISD::CopyToReg) {
2063 // If the copy has a glue operand, we conservatively assume it isn't safe to
2064 // perform a tail call.
2065 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2067 TCChain = Copy->getOperand(0);
2068 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2069 SDNode *VMov = Copy;
2070 // f64 returned in a pair of GPRs.
2071 SmallPtrSet<SDNode*, 2> Copies;
2072 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2074 if (UI->getOpcode() != ISD::CopyToReg)
2078 if (Copies.size() > 2)
2081 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2083 SDValue UseChain = UI->getOperand(0);
2084 if (Copies.count(UseChain.getNode()))
2091 } else if (Copy->getOpcode() == ISD::BITCAST) {
2092 // f32 returned in a single GPR.
2093 if (!Copy->hasOneUse())
2095 Copy = *Copy->use_begin();
2096 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2098 Chain = Copy->getOperand(0);
2103 bool HasRet = false;
2104 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2106 if (UI->getOpcode() != ARMISD::RET_FLAG)
2118 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2119 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2122 if (!CI->isTailCall())
2125 return !Subtarget->isThumb1Only();
2128 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2129 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2130 // one of the above mentioned nodes. It has to be wrapped because otherwise
2131 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2132 // be used to form addressing mode. These wrapped nodes will be selected
2134 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2135 EVT PtrVT = Op.getValueType();
2136 // FIXME there is no actual debug info here
2137 DebugLoc dl = Op.getDebugLoc();
2138 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2140 if (CP->isMachineConstantPoolEntry())
2141 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2142 CP->getAlignment());
2144 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2145 CP->getAlignment());
2146 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2149 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2150 return MachineJumpTableInfo::EK_Inline;
2153 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2154 SelectionDAG &DAG) const {
2155 MachineFunction &MF = DAG.getMachineFunction();
2156 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2157 unsigned ARMPCLabelIndex = 0;
2158 DebugLoc DL = Op.getDebugLoc();
2159 EVT PtrVT = getPointerTy();
2160 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2161 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2163 if (RelocM == Reloc::Static) {
2164 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2166 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2167 ARMPCLabelIndex = AFI->createPICLabelUId();
2168 ARMConstantPoolValue *CPV =
2169 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2170 ARMCP::CPBlockAddress, PCAdj);
2171 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2173 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2174 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2175 MachinePointerInfo::getConstantPool(),
2176 false, false, false, 0);
2177 if (RelocM == Reloc::Static)
2179 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2180 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2183 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2185 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2186 SelectionDAG &DAG) const {
2187 DebugLoc dl = GA->getDebugLoc();
2188 EVT PtrVT = getPointerTy();
2189 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2190 MachineFunction &MF = DAG.getMachineFunction();
2191 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2192 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2193 ARMConstantPoolValue *CPV =
2194 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2195 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2196 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2197 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2198 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2199 MachinePointerInfo::getConstantPool(),
2200 false, false, false, 0);
2201 SDValue Chain = Argument.getValue(1);
2203 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2204 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2206 // call __tls_get_addr.
2209 Entry.Node = Argument;
2210 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2211 Args.push_back(Entry);
2212 // FIXME: is there useful debug info available here?
2213 TargetLowering::CallLoweringInfo CLI(Chain,
2214 (Type *) Type::getInt32Ty(*DAG.getContext()),
2215 false, false, false, false,
2216 0, CallingConv::C, /*isTailCall=*/false,
2217 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2218 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2219 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2220 return CallResult.first;
2223 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2224 // "local exec" model.
2226 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2228 TLSModel::Model model) const {
2229 const GlobalValue *GV = GA->getGlobal();
2230 DebugLoc dl = GA->getDebugLoc();
2232 SDValue Chain = DAG.getEntryNode();
2233 EVT PtrVT = getPointerTy();
2234 // Get the Thread Pointer
2235 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2237 if (model == TLSModel::InitialExec) {
2238 MachineFunction &MF = DAG.getMachineFunction();
2239 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2240 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2241 // Initial exec model.
2242 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2243 ARMConstantPoolValue *CPV =
2244 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2245 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2247 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2248 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2249 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2250 MachinePointerInfo::getConstantPool(),
2251 false, false, false, 0);
2252 Chain = Offset.getValue(1);
2254 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2255 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2257 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2258 MachinePointerInfo::getConstantPool(),
2259 false, false, false, 0);
2262 assert(model == TLSModel::LocalExec);
2263 ARMConstantPoolValue *CPV =
2264 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2265 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2266 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2267 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2268 MachinePointerInfo::getConstantPool(),
2269 false, false, false, 0);
2272 // The address of the thread local variable is the add of the thread
2273 // pointer with the offset of the variable.
2274 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2278 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2279 // TODO: implement the "local dynamic" model
2280 assert(Subtarget->isTargetELF() &&
2281 "TLS not implemented for non-ELF targets");
2282 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2284 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2287 case TLSModel::GeneralDynamic:
2288 case TLSModel::LocalDynamic:
2289 return LowerToTLSGeneralDynamicModel(GA, DAG);
2290 case TLSModel::InitialExec:
2291 case TLSModel::LocalExec:
2292 return LowerToTLSExecModels(GA, DAG, model);
2294 llvm_unreachable("bogus TLS model");
2297 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2298 SelectionDAG &DAG) const {
2299 EVT PtrVT = getPointerTy();
2300 DebugLoc dl = Op.getDebugLoc();
2301 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2302 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2303 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2304 ARMConstantPoolValue *CPV =
2305 ARMConstantPoolConstant::Create(GV,
2306 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2307 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2308 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2309 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2311 MachinePointerInfo::getConstantPool(),
2312 false, false, false, 0);
2313 SDValue Chain = Result.getValue(1);
2314 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2315 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2317 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2318 MachinePointerInfo::getGOT(),
2319 false, false, false, 0);
2323 // If we have T2 ops, we can materialize the address directly via movt/movw
2324 // pair. This is always cheaper.
2325 if (Subtarget->useMovt()) {
2327 // FIXME: Once remat is capable of dealing with instructions with register
2328 // operands, expand this into two nodes.
2329 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2330 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2332 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2333 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2334 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2335 MachinePointerInfo::getConstantPool(),
2336 false, false, false, 0);
2340 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2341 SelectionDAG &DAG) const {
2342 EVT PtrVT = getPointerTy();
2343 DebugLoc dl = Op.getDebugLoc();
2344 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2345 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2347 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2348 // update ARMFastISel::ARMMaterializeGV.
2349 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2351 // FIXME: Once remat is capable of dealing with instructions with register
2352 // operands, expand this into two nodes.
2353 if (RelocM == Reloc::Static)
2354 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2355 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2357 unsigned Wrapper = (RelocM == Reloc::PIC_)
2358 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2359 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2360 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2361 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2362 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2363 MachinePointerInfo::getGOT(),
2364 false, false, false, 0);
2368 unsigned ARMPCLabelIndex = 0;
2370 if (RelocM == Reloc::Static) {
2371 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2373 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2374 ARMPCLabelIndex = AFI->createPICLabelUId();
2375 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2376 ARMConstantPoolValue *CPV =
2377 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2379 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2381 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2383 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2384 MachinePointerInfo::getConstantPool(),
2385 false, false, false, 0);
2386 SDValue Chain = Result.getValue(1);
2388 if (RelocM == Reloc::PIC_) {
2389 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2390 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2393 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2394 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2395 false, false, false, 0);
2400 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2401 SelectionDAG &DAG) const {
2402 assert(Subtarget->isTargetELF() &&
2403 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2404 MachineFunction &MF = DAG.getMachineFunction();
2405 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2406 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2407 EVT PtrVT = getPointerTy();
2408 DebugLoc dl = Op.getDebugLoc();
2409 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2410 ARMConstantPoolValue *CPV =
2411 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2412 ARMPCLabelIndex, PCAdj);
2413 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2414 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2415 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2416 MachinePointerInfo::getConstantPool(),
2417 false, false, false, 0);
2418 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2419 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2423 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2424 DebugLoc dl = Op.getDebugLoc();
2425 SDValue Val = DAG.getConstant(0, MVT::i32);
2426 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2427 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2428 Op.getOperand(1), Val);
2432 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2433 DebugLoc dl = Op.getDebugLoc();
2434 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2435 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2439 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2440 const ARMSubtarget *Subtarget) const {
2441 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2442 DebugLoc dl = Op.getDebugLoc();
2444 default: return SDValue(); // Don't custom lower most intrinsics.
2445 case Intrinsic::arm_thread_pointer: {
2446 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2447 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2449 case Intrinsic::eh_sjlj_lsda: {
2450 MachineFunction &MF = DAG.getMachineFunction();
2451 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2452 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2453 EVT PtrVT = getPointerTy();
2454 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2456 unsigned PCAdj = (RelocM != Reloc::PIC_)
2457 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2458 ARMConstantPoolValue *CPV =
2459 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2460 ARMCP::CPLSDA, PCAdj);
2461 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2462 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2464 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2465 MachinePointerInfo::getConstantPool(),
2466 false, false, false, 0);
2468 if (RelocM == Reloc::PIC_) {
2469 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2470 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2474 case Intrinsic::arm_neon_vmulls:
2475 case Intrinsic::arm_neon_vmullu: {
2476 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2477 ? ARMISD::VMULLs : ARMISD::VMULLu;
2478 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2479 Op.getOperand(1), Op.getOperand(2));
2484 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2485 const ARMSubtarget *Subtarget) {
2486 // FIXME: handle "fence singlethread" more efficiently.
2487 DebugLoc dl = Op.getDebugLoc();
2488 if (!Subtarget->hasDataBarrier()) {
2489 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2490 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2492 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2493 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2494 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2495 DAG.getConstant(0, MVT::i32));
2498 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2499 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2502 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2503 const ARMSubtarget *Subtarget) {
2504 // ARM pre v5TE and Thumb1 does not have preload instructions.
2505 if (!(Subtarget->isThumb2() ||
2506 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2507 // Just preserve the chain.
2508 return Op.getOperand(0);
2510 DebugLoc dl = Op.getDebugLoc();
2511 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2513 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2514 // ARMv7 with MP extension has PLDW.
2515 return Op.getOperand(0);
2517 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2518 if (Subtarget->isThumb()) {
2520 isRead = ~isRead & 1;
2521 isData = ~isData & 1;
2524 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2525 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2526 DAG.getConstant(isData, MVT::i32));
2529 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2530 MachineFunction &MF = DAG.getMachineFunction();
2531 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2533 // vastart just stores the address of the VarArgsFrameIndex slot into the
2534 // memory location argument.
2535 DebugLoc dl = Op.getDebugLoc();
2536 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2537 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2538 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2539 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2540 MachinePointerInfo(SV), false, false, 0);
2544 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2545 SDValue &Root, SelectionDAG &DAG,
2546 DebugLoc dl) const {
2547 MachineFunction &MF = DAG.getMachineFunction();
2548 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2550 const TargetRegisterClass *RC;
2551 if (AFI->isThumb1OnlyFunction())
2552 RC = &ARM::tGPRRegClass;
2554 RC = &ARM::GPRRegClass;
2556 // Transform the arguments stored in physical registers into virtual ones.
2557 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2558 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2561 if (NextVA.isMemLoc()) {
2562 MachineFrameInfo *MFI = MF.getFrameInfo();
2563 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2565 // Create load node to retrieve arguments from the stack.
2566 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2567 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2568 MachinePointerInfo::getFixedStack(FI),
2569 false, false, false, 0);
2571 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2572 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2575 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2579 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2580 unsigned &VARegSize, unsigned &VARegSaveSize)
2583 if (CCInfo.isFirstByValRegValid())
2584 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2586 unsigned int firstUnalloced;
2587 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2588 sizeof(GPRArgRegs) /
2589 sizeof(GPRArgRegs[0]));
2590 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2593 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2594 VARegSize = NumGPRs * 4;
2595 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2598 // The remaining GPRs hold either the beginning of variable-argument
2599 // data, or the beginning of an aggregate passed by value (usually
2600 // byval). Either way, we allocate stack slots adjacent to the data
2601 // provided by our caller, and store the unallocated registers there.
2602 // If this is a variadic function, the va_list pointer will begin with
2603 // these values; otherwise, this reassembles a (byval) structure that
2604 // was split between registers and memory.
2606 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2607 DebugLoc dl, SDValue &Chain,
2608 const Value *OrigArg,
2609 unsigned OffsetFromOrigArg,
2611 bool ForceMutable) const {
2612 MachineFunction &MF = DAG.getMachineFunction();
2613 MachineFrameInfo *MFI = MF.getFrameInfo();
2614 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2615 unsigned firstRegToSaveIndex;
2616 if (CCInfo.isFirstByValRegValid())
2617 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2619 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2620 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2623 unsigned VARegSize, VARegSaveSize;
2624 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2625 if (VARegSaveSize) {
2626 // If this function is vararg, store any remaining integer argument regs
2627 // to their spots on the stack so that they may be loaded by deferencing
2628 // the result of va_next.
2629 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2630 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2631 ArgOffset + VARegSaveSize
2634 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2637 SmallVector<SDValue, 4> MemOps;
2638 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2639 const TargetRegisterClass *RC;
2640 if (AFI->isThumb1OnlyFunction())
2641 RC = &ARM::tGPRRegClass;
2643 RC = &ARM::GPRRegClass;
2645 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2646 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2648 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2649 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2651 MemOps.push_back(Store);
2652 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2653 DAG.getConstant(4, getPointerTy()));
2655 if (!MemOps.empty())
2656 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2657 &MemOps[0], MemOps.size());
2659 // This will point to the next argument passed via stack.
2660 AFI->setVarArgsFrameIndex(
2661 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2665 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2666 CallingConv::ID CallConv, bool isVarArg,
2667 const SmallVectorImpl<ISD::InputArg>
2669 DebugLoc dl, SelectionDAG &DAG,
2670 SmallVectorImpl<SDValue> &InVals)
2672 MachineFunction &MF = DAG.getMachineFunction();
2673 MachineFrameInfo *MFI = MF.getFrameInfo();
2675 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2677 // Assign locations to all of the incoming arguments.
2678 SmallVector<CCValAssign, 16> ArgLocs;
2679 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2680 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2681 CCInfo.AnalyzeFormalArguments(Ins,
2682 CCAssignFnForNode(CallConv, /* Return*/ false,
2685 SmallVector<SDValue, 16> ArgValues;
2686 int lastInsIndex = -1;
2688 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2689 unsigned CurArgIdx = 0;
2690 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2691 CCValAssign &VA = ArgLocs[i];
2692 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2693 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2694 // Arguments stored in registers.
2695 if (VA.isRegLoc()) {
2696 EVT RegVT = VA.getLocVT();
2698 if (VA.needsCustom()) {
2699 // f64 and vector types are split up into multiple registers or
2700 // combinations of registers and stack slots.
2701 if (VA.getLocVT() == MVT::v2f64) {
2702 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2704 VA = ArgLocs[++i]; // skip ahead to next loc
2706 if (VA.isMemLoc()) {
2707 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2708 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2709 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2710 MachinePointerInfo::getFixedStack(FI),
2711 false, false, false, 0);
2713 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2716 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2717 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2718 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2719 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2720 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2722 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2725 const TargetRegisterClass *RC;
2727 if (RegVT == MVT::f32)
2728 RC = &ARM::SPRRegClass;
2729 else if (RegVT == MVT::f64)
2730 RC = &ARM::DPRRegClass;
2731 else if (RegVT == MVT::v2f64)
2732 RC = &ARM::QPRRegClass;
2733 else if (RegVT == MVT::i32)
2734 RC = AFI->isThumb1OnlyFunction() ?
2735 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2736 (const TargetRegisterClass*)&ARM::GPRRegClass;
2738 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2740 // Transform the arguments in physical registers into virtual ones.
2741 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2742 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2745 // If this is an 8 or 16-bit value, it is really passed promoted
2746 // to 32 bits. Insert an assert[sz]ext to capture this, then
2747 // truncate to the right size.
2748 switch (VA.getLocInfo()) {
2749 default: llvm_unreachable("Unknown loc info!");
2750 case CCValAssign::Full: break;
2751 case CCValAssign::BCvt:
2752 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2754 case CCValAssign::SExt:
2755 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2756 DAG.getValueType(VA.getValVT()));
2757 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2759 case CCValAssign::ZExt:
2760 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2761 DAG.getValueType(VA.getValVT()));
2762 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2766 InVals.push_back(ArgValue);
2768 } else { // VA.isRegLoc()
2771 assert(VA.isMemLoc());
2772 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2774 int index = ArgLocs[i].getValNo();
2776 // Some Ins[] entries become multiple ArgLoc[] entries.
2777 // Process them only once.
2778 if (index != lastInsIndex)
2780 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2781 // FIXME: For now, all byval parameter objects are marked mutable.
2782 // This can be changed with more analysis.
2783 // In case of tail call optimization mark all arguments mutable.
2784 // Since they could be overwritten by lowering of arguments in case of
2786 if (Flags.isByVal()) {
2787 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2788 if (!AFI->getVarArgsFrameIndex()) {
2789 VarArgStyleRegisters(CCInfo, DAG,
2790 dl, Chain, CurOrigArg,
2791 Ins[VA.getValNo()].PartOffset,
2792 VA.getLocMemOffset(),
2793 true /*force mutable frames*/);
2794 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2795 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2797 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2798 VA.getLocMemOffset(), false);
2799 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2802 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2803 VA.getLocMemOffset(), true);
2805 // Create load nodes to retrieve arguments from the stack.
2806 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2807 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2808 MachinePointerInfo::getFixedStack(FI),
2809 false, false, false, 0));
2811 lastInsIndex = index;
2818 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2819 CCInfo.getNextStackOffset());
2824 /// isFloatingPointZero - Return true if this is +0.0.
2825 static bool isFloatingPointZero(SDValue Op) {
2826 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2827 return CFP->getValueAPF().isPosZero();
2828 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2829 // Maybe this has already been legalized into the constant pool?
2830 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2831 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2832 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2833 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2834 return CFP->getValueAPF().isPosZero();
2840 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2841 /// the given operands.
2843 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2844 SDValue &ARMcc, SelectionDAG &DAG,
2845 DebugLoc dl) const {
2846 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2847 unsigned C = RHSC->getZExtValue();
2848 if (!isLegalICmpImmediate(C)) {
2849 // Constant does not fit, try adjusting it by one?
2854 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2855 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2856 RHS = DAG.getConstant(C-1, MVT::i32);
2861 if (C != 0 && isLegalICmpImmediate(C-1)) {
2862 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2863 RHS = DAG.getConstant(C-1, MVT::i32);
2868 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2869 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2870 RHS = DAG.getConstant(C+1, MVT::i32);
2875 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2876 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2877 RHS = DAG.getConstant(C+1, MVT::i32);
2884 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2885 ARMISD::NodeType CompareType;
2888 CompareType = ARMISD::CMP;
2893 CompareType = ARMISD::CMPZ;
2896 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2897 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2900 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2902 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2903 DebugLoc dl) const {
2905 if (!isFloatingPointZero(RHS))
2906 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2908 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2909 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2912 /// duplicateCmp - Glue values can have only one use, so this function
2913 /// duplicates a comparison node.
2915 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2916 unsigned Opc = Cmp.getOpcode();
2917 DebugLoc DL = Cmp.getDebugLoc();
2918 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2919 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2921 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2922 Cmp = Cmp.getOperand(0);
2923 Opc = Cmp.getOpcode();
2924 if (Opc == ARMISD::CMPFP)
2925 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2927 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2928 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2930 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2933 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2934 SDValue Cond = Op.getOperand(0);
2935 SDValue SelectTrue = Op.getOperand(1);
2936 SDValue SelectFalse = Op.getOperand(2);
2937 DebugLoc dl = Op.getDebugLoc();
2941 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2942 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2944 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2945 const ConstantSDNode *CMOVTrue =
2946 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2947 const ConstantSDNode *CMOVFalse =
2948 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2950 if (CMOVTrue && CMOVFalse) {
2951 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2952 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2956 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2958 False = SelectFalse;
2959 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2964 if (True.getNode() && False.getNode()) {
2965 EVT VT = Op.getValueType();
2966 SDValue ARMcc = Cond.getOperand(2);
2967 SDValue CCR = Cond.getOperand(3);
2968 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2969 assert(True.getValueType() == VT);
2970 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2975 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2976 // undefined bits before doing a full-word comparison with zero.
2977 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2978 DAG.getConstant(1, Cond.getValueType()));
2980 return DAG.getSelectCC(dl, Cond,
2981 DAG.getConstant(0, Cond.getValueType()),
2982 SelectTrue, SelectFalse, ISD::SETNE);
2985 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2986 EVT VT = Op.getValueType();
2987 SDValue LHS = Op.getOperand(0);
2988 SDValue RHS = Op.getOperand(1);
2989 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2990 SDValue TrueVal = Op.getOperand(2);
2991 SDValue FalseVal = Op.getOperand(3);
2992 DebugLoc dl = Op.getDebugLoc();
2994 if (LHS.getValueType() == MVT::i32) {
2996 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2997 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2998 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
3001 ARMCC::CondCodes CondCode, CondCode2;
3002 FPCCToARMCC(CC, CondCode, CondCode2);
3004 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3005 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3006 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3007 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3009 if (CondCode2 != ARMCC::AL) {
3010 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3011 // FIXME: Needs another CMP because flag can have but one use.
3012 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3013 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3014 Result, TrueVal, ARMcc2, CCR, Cmp2);
3019 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3020 /// to morph to an integer compare sequence.
3021 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3022 const ARMSubtarget *Subtarget) {
3023 SDNode *N = Op.getNode();
3024 if (!N->hasOneUse())
3025 // Otherwise it requires moving the value from fp to integer registers.
3027 if (!N->getNumValues())
3029 EVT VT = Op.getValueType();
3030 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3031 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3032 // vmrs are very slow, e.g. cortex-a8.
3035 if (isFloatingPointZero(Op)) {
3039 return ISD::isNormalLoad(N);
3042 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3043 if (isFloatingPointZero(Op))
3044 return DAG.getConstant(0, MVT::i32);
3046 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3047 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3048 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3049 Ld->isVolatile(), Ld->isNonTemporal(),
3050 Ld->isInvariant(), Ld->getAlignment());
3052 llvm_unreachable("Unknown VFP cmp argument!");
3055 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3056 SDValue &RetVal1, SDValue &RetVal2) {
3057 if (isFloatingPointZero(Op)) {
3058 RetVal1 = DAG.getConstant(0, MVT::i32);
3059 RetVal2 = DAG.getConstant(0, MVT::i32);
3063 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3064 SDValue Ptr = Ld->getBasePtr();
3065 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3066 Ld->getChain(), Ptr,
3067 Ld->getPointerInfo(),
3068 Ld->isVolatile(), Ld->isNonTemporal(),
3069 Ld->isInvariant(), Ld->getAlignment());
3071 EVT PtrType = Ptr.getValueType();
3072 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3073 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3074 PtrType, Ptr, DAG.getConstant(4, PtrType));
3075 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3076 Ld->getChain(), NewPtr,
3077 Ld->getPointerInfo().getWithOffset(4),
3078 Ld->isVolatile(), Ld->isNonTemporal(),
3079 Ld->isInvariant(), NewAlign);
3083 llvm_unreachable("Unknown VFP cmp argument!");
3086 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3087 /// f32 and even f64 comparisons to integer ones.
3089 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3090 SDValue Chain = Op.getOperand(0);
3091 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3092 SDValue LHS = Op.getOperand(2);
3093 SDValue RHS = Op.getOperand(3);
3094 SDValue Dest = Op.getOperand(4);
3095 DebugLoc dl = Op.getDebugLoc();
3097 bool LHSSeenZero = false;
3098 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3099 bool RHSSeenZero = false;
3100 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3101 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3102 // If unsafe fp math optimization is enabled and there are no other uses of
3103 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3104 // to an integer comparison.
3105 if (CC == ISD::SETOEQ)
3107 else if (CC == ISD::SETUNE)
3110 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3112 if (LHS.getValueType() == MVT::f32) {
3113 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3114 bitcastf32Toi32(LHS, DAG), Mask);
3115 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3116 bitcastf32Toi32(RHS, DAG), Mask);
3117 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3118 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3119 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3120 Chain, Dest, ARMcc, CCR, Cmp);
3125 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3126 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3127 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3128 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3129 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3130 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3131 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3132 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3133 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3139 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3140 SDValue Chain = Op.getOperand(0);
3141 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3142 SDValue LHS = Op.getOperand(2);
3143 SDValue RHS = Op.getOperand(3);
3144 SDValue Dest = Op.getOperand(4);
3145 DebugLoc dl = Op.getDebugLoc();
3147 if (LHS.getValueType() == MVT::i32) {
3149 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3150 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3151 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3152 Chain, Dest, ARMcc, CCR, Cmp);
3155 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3157 if (getTargetMachine().Options.UnsafeFPMath &&
3158 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3159 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3160 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3161 if (Result.getNode())
3165 ARMCC::CondCodes CondCode, CondCode2;
3166 FPCCToARMCC(CC, CondCode, CondCode2);
3168 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3169 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3170 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3171 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3172 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3173 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3174 if (CondCode2 != ARMCC::AL) {
3175 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3176 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3177 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3182 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3183 SDValue Chain = Op.getOperand(0);
3184 SDValue Table = Op.getOperand(1);
3185 SDValue Index = Op.getOperand(2);
3186 DebugLoc dl = Op.getDebugLoc();
3188 EVT PTy = getPointerTy();
3189 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3190 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3191 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3192 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3193 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3194 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3195 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3196 if (Subtarget->isThumb2()) {
3197 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3198 // which does another jump to the destination. This also makes it easier
3199 // to translate it to TBB / TBH later.
3200 // FIXME: This might not work if the function is extremely large.
3201 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3202 Addr, Op.getOperand(2), JTI, UId);
3204 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3205 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3206 MachinePointerInfo::getJumpTable(),
3207 false, false, false, 0);
3208 Chain = Addr.getValue(1);
3209 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3210 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3212 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3213 MachinePointerInfo::getJumpTable(),
3214 false, false, false, 0);
3215 Chain = Addr.getValue(1);
3216 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3220 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3221 EVT VT = Op.getValueType();
3222 DebugLoc dl = Op.getDebugLoc();
3224 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3225 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3227 return DAG.UnrollVectorOp(Op.getNode());
3230 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3231 "Invalid type for custom lowering!");
3232 if (VT != MVT::v4i16)
3233 return DAG.UnrollVectorOp(Op.getNode());
3235 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3236 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3239 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3240 EVT VT = Op.getValueType();
3242 return LowerVectorFP_TO_INT(Op, DAG);
3244 DebugLoc dl = Op.getDebugLoc();
3247 switch (Op.getOpcode()) {
3248 default: llvm_unreachable("Invalid opcode!");
3249 case ISD::FP_TO_SINT:
3250 Opc = ARMISD::FTOSI;
3252 case ISD::FP_TO_UINT:
3253 Opc = ARMISD::FTOUI;
3256 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3257 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3260 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3261 EVT VT = Op.getValueType();
3262 DebugLoc dl = Op.getDebugLoc();
3264 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3265 if (VT.getVectorElementType() == MVT::f32)
3267 return DAG.UnrollVectorOp(Op.getNode());
3270 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3271 "Invalid type for custom lowering!");
3272 if (VT != MVT::v4f32)
3273 return DAG.UnrollVectorOp(Op.getNode());
3277 switch (Op.getOpcode()) {
3278 default: llvm_unreachable("Invalid opcode!");
3279 case ISD::SINT_TO_FP:
3280 CastOpc = ISD::SIGN_EXTEND;
3281 Opc = ISD::SINT_TO_FP;
3283 case ISD::UINT_TO_FP:
3284 CastOpc = ISD::ZERO_EXTEND;
3285 Opc = ISD::UINT_TO_FP;
3289 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3290 return DAG.getNode(Opc, dl, VT, Op);
3293 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3294 EVT VT = Op.getValueType();
3296 return LowerVectorINT_TO_FP(Op, DAG);
3298 DebugLoc dl = Op.getDebugLoc();
3301 switch (Op.getOpcode()) {
3302 default: llvm_unreachable("Invalid opcode!");
3303 case ISD::SINT_TO_FP:
3304 Opc = ARMISD::SITOF;
3306 case ISD::UINT_TO_FP:
3307 Opc = ARMISD::UITOF;
3311 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3312 return DAG.getNode(Opc, dl, VT, Op);
3315 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3316 // Implement fcopysign with a fabs and a conditional fneg.
3317 SDValue Tmp0 = Op.getOperand(0);
3318 SDValue Tmp1 = Op.getOperand(1);
3319 DebugLoc dl = Op.getDebugLoc();
3320 EVT VT = Op.getValueType();
3321 EVT SrcVT = Tmp1.getValueType();
3322 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3323 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3324 bool UseNEON = !InGPR && Subtarget->hasNEON();
3327 // Use VBSL to copy the sign bit.
3328 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3329 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3330 DAG.getTargetConstant(EncodedVal, MVT::i32));
3331 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3333 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3334 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3335 DAG.getConstant(32, MVT::i32));
3336 else /*if (VT == MVT::f32)*/
3337 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3338 if (SrcVT == MVT::f32) {
3339 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3341 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3342 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3343 DAG.getConstant(32, MVT::i32));
3344 } else if (VT == MVT::f32)
3345 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3346 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3347 DAG.getConstant(32, MVT::i32));
3348 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3349 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3351 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3353 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3354 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3355 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3357 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3358 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3359 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3360 if (VT == MVT::f32) {
3361 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3362 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3363 DAG.getConstant(0, MVT::i32));
3365 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3371 // Bitcast operand 1 to i32.
3372 if (SrcVT == MVT::f64)
3373 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3374 &Tmp1, 1).getValue(1);
3375 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3377 // Or in the signbit with integer operations.
3378 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3379 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3380 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3381 if (VT == MVT::f32) {
3382 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3383 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3384 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3385 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3388 // f64: Or the high part with signbit and then combine two parts.
3389 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3391 SDValue Lo = Tmp0.getValue(0);
3392 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3393 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3394 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3397 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3398 MachineFunction &MF = DAG.getMachineFunction();
3399 MachineFrameInfo *MFI = MF.getFrameInfo();
3400 MFI->setReturnAddressIsTaken(true);
3402 EVT VT = Op.getValueType();
3403 DebugLoc dl = Op.getDebugLoc();
3404 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3406 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3407 SDValue Offset = DAG.getConstant(4, MVT::i32);
3408 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3409 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3410 MachinePointerInfo(), false, false, false, 0);
3413 // Return LR, which contains the return address. Mark it an implicit live-in.
3414 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3415 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3418 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3419 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3420 MFI->setFrameAddressIsTaken(true);
3422 EVT VT = Op.getValueType();
3423 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3424 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3425 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3426 ? ARM::R7 : ARM::R11;
3427 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3429 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3430 MachinePointerInfo(),
3431 false, false, false, 0);
3435 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
3436 /// and size(DestVec) > 128-bits.
3437 /// This is achieved by doing the one extension from the SrcVec, splitting the
3438 /// result, extending these parts, and then concatenating these into the
3440 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
3441 SDValue Op = N->getOperand(0);
3442 EVT SrcVT = Op.getValueType();
3443 EVT DestVT = N->getValueType(0);
3445 assert(DestVT.getSizeInBits() > 128 &&
3446 "Custom sext/zext expansion needs >128-bit vector.");
3447 // If this is a normal length extension, use the default expansion.
3448 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
3449 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
3452 DebugLoc dl = N->getDebugLoc();
3453 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
3454 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
3455 unsigned NumElts = SrcVT.getVectorNumElements();
3456 LLVMContext &Ctx = *DAG.getContext();
3457 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;
3459 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3461 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3463 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
3466 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
3467 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3468 DAG.getIntPtrConstant(0));
3469 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3470 DAG.getIntPtrConstant(NumElts/2));
3471 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
3472 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
3473 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
3476 /// ExpandBITCAST - If the target supports VFP, this function is called to
3477 /// expand a bit convert where either the source or destination type is i64 to
3478 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3479 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3480 /// vectors), since the legalizer won't know what to do with that.
3481 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3482 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3483 DebugLoc dl = N->getDebugLoc();
3484 SDValue Op = N->getOperand(0);
3486 // This function is only supposed to be called for i64 types, either as the
3487 // source or destination of the bit convert.
3488 EVT SrcVT = Op.getValueType();
3489 EVT DstVT = N->getValueType(0);
3490 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3491 "ExpandBITCAST called for non-i64 type");
3493 // Turn i64->f64 into VMOVDRR.
3494 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3495 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3496 DAG.getConstant(0, MVT::i32));
3497 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3498 DAG.getConstant(1, MVT::i32));
3499 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3500 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3503 // Turn f64->i64 into VMOVRRD.
3504 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3505 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3506 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3507 // Merge the pieces into a single i64 value.
3508 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3514 /// getZeroVector - Returns a vector of specified type with all zero elements.
3515 /// Zero vectors are used to represent vector negation and in those cases
3516 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3517 /// not support i64 elements, so sometimes the zero vectors will need to be
3518 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3520 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3521 assert(VT.isVector() && "Expected a vector type");
3522 // The canonical modified immediate encoding of a zero vector is....0!
3523 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3524 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3525 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3526 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3529 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3530 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3531 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3532 SelectionDAG &DAG) const {
3533 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3534 EVT VT = Op.getValueType();
3535 unsigned VTBits = VT.getSizeInBits();
3536 DebugLoc dl = Op.getDebugLoc();
3537 SDValue ShOpLo = Op.getOperand(0);
3538 SDValue ShOpHi = Op.getOperand(1);
3539 SDValue ShAmt = Op.getOperand(2);
3541 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3543 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3545 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3546 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3547 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3548 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3549 DAG.getConstant(VTBits, MVT::i32));
3550 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3551 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3552 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3554 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3555 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3557 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3558 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3561 SDValue Ops[2] = { Lo, Hi };
3562 return DAG.getMergeValues(Ops, 2, dl);
3565 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3566 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3567 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3568 SelectionDAG &DAG) const {
3569 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3570 EVT VT = Op.getValueType();
3571 unsigned VTBits = VT.getSizeInBits();
3572 DebugLoc dl = Op.getDebugLoc();
3573 SDValue ShOpLo = Op.getOperand(0);
3574 SDValue ShOpHi = Op.getOperand(1);
3575 SDValue ShAmt = Op.getOperand(2);
3578 assert(Op.getOpcode() == ISD::SHL_PARTS);
3579 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3580 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3581 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3582 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3583 DAG.getConstant(VTBits, MVT::i32));
3584 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3585 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3587 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3588 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3589 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3591 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3592 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3595 SDValue Ops[2] = { Lo, Hi };
3596 return DAG.getMergeValues(Ops, 2, dl);
3599 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3600 SelectionDAG &DAG) const {
3601 // The rounding mode is in bits 23:22 of the FPSCR.
3602 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3603 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3604 // so that the shift + and get folded into a bitfield extract.
3605 DebugLoc dl = Op.getDebugLoc();
3606 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3607 DAG.getConstant(Intrinsic::arm_get_fpscr,
3609 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3610 DAG.getConstant(1U << 22, MVT::i32));
3611 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3612 DAG.getConstant(22, MVT::i32));
3613 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3614 DAG.getConstant(3, MVT::i32));
3617 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3618 const ARMSubtarget *ST) {
3619 EVT VT = N->getValueType(0);
3620 DebugLoc dl = N->getDebugLoc();
3622 if (!ST->hasV6T2Ops())
3625 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3626 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3629 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3630 /// for each 16-bit element from operand, repeated. The basic idea is to
3631 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3633 /// Trace for v4i16:
3634 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3635 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3636 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3637 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3638 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3639 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3640 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3641 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3642 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3643 EVT VT = N->getValueType(0);
3644 DebugLoc DL = N->getDebugLoc();
3646 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3647 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3648 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3649 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3650 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3651 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3654 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3655 /// bit-count for each 16-bit element from the operand. We need slightly
3656 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3657 /// 64/128-bit registers.
3659 /// Trace for v4i16:
3660 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3661 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3662 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3663 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3664 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3665 EVT VT = N->getValueType(0);
3666 DebugLoc DL = N->getDebugLoc();
3668 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3669 if (VT.is64BitVector()) {
3670 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3671 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3672 DAG.getIntPtrConstant(0));
3674 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3675 BitCounts, DAG.getIntPtrConstant(0));
3676 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3680 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3681 /// bit-count for each 32-bit element from the operand. The idea here is
3682 /// to split the vector into 16-bit elements, leverage the 16-bit count
3683 /// routine, and then combine the results.
3685 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3686 /// input = [v0 v1 ] (vi: 32-bit elements)
3687 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3688 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3689 /// vrev: N0 = [k1 k0 k3 k2 ]
3691 /// N1 =+[k1 k0 k3 k2 ]
3693 /// N2 =+[k1 k3 k0 k2 ]
3695 /// Extended =+[k1 k3 k0 k2 ]
3697 /// Extracted=+[k1 k3 ]
3699 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
3700 EVT VT = N->getValueType(0);
3701 DebugLoc DL = N->getDebugLoc();
3703 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
3705 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
3706 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
3707 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
3708 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
3709 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
3711 if (VT.is64BitVector()) {
3712 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
3713 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
3714 DAG.getIntPtrConstant(0));
3716 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
3717 DAG.getIntPtrConstant(0));
3718 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
3722 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
3723 const ARMSubtarget *ST) {
3724 EVT VT = N->getValueType(0);
3726 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
3727 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
3728 VT == MVT::v4i16 || VT == MVT::v8i16) &&
3729 "Unexpected type for custom ctpop lowering");
3731 if (VT.getVectorElementType() == MVT::i32)
3732 return lowerCTPOP32BitElements(N, DAG);
3734 return lowerCTPOP16BitElements(N, DAG);
3737 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3738 const ARMSubtarget *ST) {
3739 EVT VT = N->getValueType(0);
3740 DebugLoc dl = N->getDebugLoc();
3745 // Lower vector shifts on NEON to use VSHL.
3746 assert(ST->hasNEON() && "unexpected vector shift");
3748 // Left shifts translate directly to the vshiftu intrinsic.
3749 if (N->getOpcode() == ISD::SHL)
3750 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3751 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3752 N->getOperand(0), N->getOperand(1));
3754 assert((N->getOpcode() == ISD::SRA ||
3755 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3757 // NEON uses the same intrinsics for both left and right shifts. For
3758 // right shifts, the shift amounts are negative, so negate the vector of
3760 EVT ShiftVT = N->getOperand(1).getValueType();
3761 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3762 getZeroVector(ShiftVT, DAG, dl),
3764 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3765 Intrinsic::arm_neon_vshifts :
3766 Intrinsic::arm_neon_vshiftu);
3767 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3768 DAG.getConstant(vshiftInt, MVT::i32),
3769 N->getOperand(0), NegatedCount);
3772 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3773 const ARMSubtarget *ST) {
3774 EVT VT = N->getValueType(0);
3775 DebugLoc dl = N->getDebugLoc();
3777 // We can get here for a node like i32 = ISD::SHL i32, i64
3781 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3782 "Unknown shift to lower!");
3784 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3785 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3786 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3789 // If we are in thumb mode, we don't have RRX.
3790 if (ST->isThumb1Only()) return SDValue();
3792 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3793 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3794 DAG.getConstant(0, MVT::i32));
3795 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3796 DAG.getConstant(1, MVT::i32));
3798 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3799 // captures the result into a carry flag.
3800 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3801 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3803 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3804 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3806 // Merge the pieces into a single i64 value.
3807 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3810 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3811 SDValue TmpOp0, TmpOp1;
3812 bool Invert = false;
3816 SDValue Op0 = Op.getOperand(0);
3817 SDValue Op1 = Op.getOperand(1);
3818 SDValue CC = Op.getOperand(2);
3819 EVT VT = Op.getValueType();
3820 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3821 DebugLoc dl = Op.getDebugLoc();
3823 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3824 switch (SetCCOpcode) {
3825 default: llvm_unreachable("Illegal FP comparison");
3827 case ISD::SETNE: Invert = true; // Fallthrough
3829 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3831 case ISD::SETLT: Swap = true; // Fallthrough
3833 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3835 case ISD::SETLE: Swap = true; // Fallthrough
3837 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3838 case ISD::SETUGE: Swap = true; // Fallthrough
3839 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3840 case ISD::SETUGT: Swap = true; // Fallthrough
3841 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3842 case ISD::SETUEQ: Invert = true; // Fallthrough
3844 // Expand this to (OLT | OGT).
3848 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3849 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3851 case ISD::SETUO: Invert = true; // Fallthrough
3853 // Expand this to (OLT | OGE).
3857 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3858 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3862 // Integer comparisons.
3863 switch (SetCCOpcode) {
3864 default: llvm_unreachable("Illegal integer comparison");
3865 case ISD::SETNE: Invert = true;
3866 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3867 case ISD::SETLT: Swap = true;
3868 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3869 case ISD::SETLE: Swap = true;
3870 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3871 case ISD::SETULT: Swap = true;
3872 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3873 case ISD::SETULE: Swap = true;
3874 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3877 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3878 if (Opc == ARMISD::VCEQ) {
3881 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3883 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3886 // Ignore bitconvert.
3887 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3888 AndOp = AndOp.getOperand(0);
3890 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3892 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3893 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3900 std::swap(Op0, Op1);
3902 // If one of the operands is a constant vector zero, attempt to fold the
3903 // comparison to a specialized compare-against-zero form.
3905 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3907 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3908 if (Opc == ARMISD::VCGE)
3909 Opc = ARMISD::VCLEZ;
3910 else if (Opc == ARMISD::VCGT)
3911 Opc = ARMISD::VCLTZ;
3916 if (SingleOp.getNode()) {
3919 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3921 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3923 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3925 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3927 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3929 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3932 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3936 Result = DAG.getNOT(dl, Result, VT);
3941 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3942 /// valid vector constant for a NEON instruction with a "modified immediate"
3943 /// operand (e.g., VMOV). If so, return the encoded value.
3944 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3945 unsigned SplatBitSize, SelectionDAG &DAG,
3946 EVT &VT, bool is128Bits, NEONModImmType type) {
3947 unsigned OpCmode, Imm;
3949 // SplatBitSize is set to the smallest size that splats the vector, so a
3950 // zero vector will always have SplatBitSize == 8. However, NEON modified
3951 // immediate instructions others than VMOV do not support the 8-bit encoding
3952 // of a zero vector, and the default encoding of zero is supposed to be the
3957 switch (SplatBitSize) {
3959 if (type != VMOVModImm)
3961 // Any 1-byte value is OK. Op=0, Cmode=1110.
3962 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3965 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3969 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3970 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3971 if ((SplatBits & ~0xff) == 0) {
3972 // Value = 0x00nn: Op=x, Cmode=100x.
3977 if ((SplatBits & ~0xff00) == 0) {
3978 // Value = 0xnn00: Op=x, Cmode=101x.
3980 Imm = SplatBits >> 8;
3986 // NEON's 32-bit VMOV supports splat values where:
3987 // * only one byte is nonzero, or
3988 // * the least significant byte is 0xff and the second byte is nonzero, or
3989 // * the least significant 2 bytes are 0xff and the third is nonzero.
3990 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3991 if ((SplatBits & ~0xff) == 0) {
3992 // Value = 0x000000nn: Op=x, Cmode=000x.
3997 if ((SplatBits & ~0xff00) == 0) {
3998 // Value = 0x0000nn00: Op=x, Cmode=001x.
4000 Imm = SplatBits >> 8;
4003 if ((SplatBits & ~0xff0000) == 0) {
4004 // Value = 0x00nn0000: Op=x, Cmode=010x.
4006 Imm = SplatBits >> 16;
4009 if ((SplatBits & ~0xff000000) == 0) {
4010 // Value = 0xnn000000: Op=x, Cmode=011x.
4012 Imm = SplatBits >> 24;
4016 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4017 if (type == OtherModImm) return SDValue();
4019 if ((SplatBits & ~0xffff) == 0 &&
4020 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4021 // Value = 0x0000nnff: Op=x, Cmode=1100.
4023 Imm = SplatBits >> 8;
4028 if ((SplatBits & ~0xffffff) == 0 &&
4029 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4030 // Value = 0x00nnffff: Op=x, Cmode=1101.
4032 Imm = SplatBits >> 16;
4033 SplatBits |= 0xffff;
4037 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4038 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4039 // VMOV.I32. A (very) minor optimization would be to replicate the value
4040 // and fall through here to test for a valid 64-bit splat. But, then the
4041 // caller would also need to check and handle the change in size.
4045 if (type != VMOVModImm)
4047 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4048 uint64_t BitMask = 0xff;
4050 unsigned ImmMask = 1;
4052 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4053 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4056 } else if ((SplatBits & BitMask) != 0) {
4062 // Op=1, Cmode=1110.
4065 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4070 llvm_unreachable("unexpected size for isNEONModifiedImm");
4073 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4074 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4077 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4078 const ARMSubtarget *ST) const {
4079 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
4082 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4083 assert(Op.getValueType() == MVT::f32 &&
4084 "ConstantFP custom lowering should only occur for f32.");
4086 // Try splatting with a VMOV.f32...
4087 APFloat FPVal = CFP->getValueAPF();
4088 int ImmVal = ARM_AM::getFP32Imm(FPVal);
4090 DebugLoc DL = Op.getDebugLoc();
4091 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4092 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4094 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4095 DAG.getConstant(0, MVT::i32));
4098 // If that fails, try a VMOV.i32
4100 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
4101 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
4103 if (NewVal != SDValue()) {
4104 DebugLoc DL = Op.getDebugLoc();
4105 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4107 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4109 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4110 DAG.getConstant(0, MVT::i32));
4113 // Finally, try a VMVN.i32
4114 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
4116 if (NewVal != SDValue()) {
4117 DebugLoc DL = Op.getDebugLoc();
4118 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4119 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4121 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4122 DAG.getConstant(0, MVT::i32));
4128 // check if an VEXT instruction can handle the shuffle mask when the
4129 // vector sources of the shuffle are the same.
4130 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4131 unsigned NumElts = VT.getVectorNumElements();
4133 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4139 // If this is a VEXT shuffle, the immediate value is the index of the first
4140 // element. The other shuffle indices must be the successive elements after
4142 unsigned ExpectedElt = Imm;
4143 for (unsigned i = 1; i < NumElts; ++i) {
4144 // Increment the expected index. If it wraps around, just follow it
4145 // back to index zero and keep going.
4147 if (ExpectedElt == NumElts)
4150 if (M[i] < 0) continue; // ignore UNDEF indices
4151 if (ExpectedElt != static_cast<unsigned>(M[i]))
4159 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4160 bool &ReverseVEXT, unsigned &Imm) {
4161 unsigned NumElts = VT.getVectorNumElements();
4162 ReverseVEXT = false;
4164 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4170 // If this is a VEXT shuffle, the immediate value is the index of the first
4171 // element. The other shuffle indices must be the successive elements after
4173 unsigned ExpectedElt = Imm;
4174 for (unsigned i = 1; i < NumElts; ++i) {
4175 // Increment the expected index. If it wraps around, it may still be
4176 // a VEXT but the source vectors must be swapped.
4178 if (ExpectedElt == NumElts * 2) {
4183 if (M[i] < 0) continue; // ignore UNDEF indices
4184 if (ExpectedElt != static_cast<unsigned>(M[i]))
4188 // Adjust the index value if the source operands will be swapped.
4195 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4196 /// instruction with the specified blocksize. (The order of the elements
4197 /// within each block of the vector is reversed.)
4198 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4199 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4200 "Only possible block sizes for VREV are: 16, 32, 64");
4202 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4206 unsigned NumElts = VT.getVectorNumElements();
4207 unsigned BlockElts = M[0] + 1;
4208 // If the first shuffle index is UNDEF, be optimistic.
4210 BlockElts = BlockSize / EltSz;
4212 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4215 for (unsigned i = 0; i < NumElts; ++i) {
4216 if (M[i] < 0) continue; // ignore UNDEF indices
4217 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4224 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4225 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4226 // range, then 0 is placed into the resulting vector. So pretty much any mask
4227 // of 8 elements can work here.
4228 return VT == MVT::v8i8 && M.size() == 8;
4231 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4232 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4236 unsigned NumElts = VT.getVectorNumElements();
4237 WhichResult = (M[0] == 0 ? 0 : 1);
4238 for (unsigned i = 0; i < NumElts; i += 2) {
4239 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4240 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4246 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4247 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4248 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4249 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4250 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4254 unsigned NumElts = VT.getVectorNumElements();
4255 WhichResult = (M[0] == 0 ? 0 : 1);
4256 for (unsigned i = 0; i < NumElts; i += 2) {
4257 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4258 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4264 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4265 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4269 unsigned NumElts = VT.getVectorNumElements();
4270 WhichResult = (M[0] == 0 ? 0 : 1);
4271 for (unsigned i = 0; i != NumElts; ++i) {
4272 if (M[i] < 0) continue; // ignore UNDEF indices
4273 if ((unsigned) M[i] != 2 * i + WhichResult)
4277 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4278 if (VT.is64BitVector() && EltSz == 32)
4284 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4285 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4286 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4287 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4288 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4292 unsigned Half = VT.getVectorNumElements() / 2;
4293 WhichResult = (M[0] == 0 ? 0 : 1);
4294 for (unsigned j = 0; j != 2; ++j) {
4295 unsigned Idx = WhichResult;
4296 for (unsigned i = 0; i != Half; ++i) {
4297 int MIdx = M[i + j * Half];
4298 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4304 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4305 if (VT.is64BitVector() && EltSz == 32)
4311 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4312 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4316 unsigned NumElts = VT.getVectorNumElements();
4317 WhichResult = (M[0] == 0 ? 0 : 1);
4318 unsigned Idx = WhichResult * NumElts / 2;
4319 for (unsigned i = 0; i != NumElts; i += 2) {
4320 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4321 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4326 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4327 if (VT.is64BitVector() && EltSz == 32)
4333 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4334 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4335 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4336 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4337 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4341 unsigned NumElts = VT.getVectorNumElements();
4342 WhichResult = (M[0] == 0 ? 0 : 1);
4343 unsigned Idx = WhichResult * NumElts / 2;
4344 for (unsigned i = 0; i != NumElts; i += 2) {
4345 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4346 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4351 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4352 if (VT.is64BitVector() && EltSz == 32)
4358 /// \return true if this is a reverse operation on an vector.
4359 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4360 unsigned NumElts = VT.getVectorNumElements();
4361 // Make sure the mask has the right size.
4362 if (NumElts != M.size())
4365 // Look for <15, ..., 3, -1, 1, 0>.
4366 for (unsigned i = 0; i != NumElts; ++i)
4367 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4373 // If N is an integer constant that can be moved into a register in one
4374 // instruction, return an SDValue of such a constant (will become a MOV
4375 // instruction). Otherwise return null.
4376 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4377 const ARMSubtarget *ST, DebugLoc dl) {
4379 if (!isa<ConstantSDNode>(N))
4381 Val = cast<ConstantSDNode>(N)->getZExtValue();
4383 if (ST->isThumb1Only()) {
4384 if (Val <= 255 || ~Val <= 255)
4385 return DAG.getConstant(Val, MVT::i32);
4387 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4388 return DAG.getConstant(Val, MVT::i32);
4393 // If this is a case we can't handle, return null and let the default
4394 // expansion code take care of it.
4395 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4396 const ARMSubtarget *ST) const {
4397 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4398 DebugLoc dl = Op.getDebugLoc();
4399 EVT VT = Op.getValueType();
4401 APInt SplatBits, SplatUndef;
4402 unsigned SplatBitSize;
4404 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4405 if (SplatBitSize <= 64) {
4406 // Check if an immediate VMOV works.
4408 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4409 SplatUndef.getZExtValue(), SplatBitSize,
4410 DAG, VmovVT, VT.is128BitVector(),
4412 if (Val.getNode()) {
4413 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4414 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4417 // Try an immediate VMVN.
4418 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4419 Val = isNEONModifiedImm(NegatedImm,
4420 SplatUndef.getZExtValue(), SplatBitSize,
4421 DAG, VmovVT, VT.is128BitVector(),
4423 if (Val.getNode()) {
4424 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4425 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4428 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4429 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4430 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4432 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4433 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4439 // Scan through the operands to see if only one value is used.
4441 // As an optimisation, even if more than one value is used it may be more
4442 // profitable to splat with one value then change some lanes.
4444 // Heuristically we decide to do this if the vector has a "dominant" value,
4445 // defined as splatted to more than half of the lanes.
4446 unsigned NumElts = VT.getVectorNumElements();
4447 bool isOnlyLowElement = true;
4448 bool usesOnlyOneValue = true;
4449 bool hasDominantValue = false;
4450 bool isConstant = true;
4452 // Map of the number of times a particular SDValue appears in the
4454 DenseMap<SDValue, unsigned> ValueCounts;
4456 for (unsigned i = 0; i < NumElts; ++i) {
4457 SDValue V = Op.getOperand(i);
4458 if (V.getOpcode() == ISD::UNDEF)
4461 isOnlyLowElement = false;
4462 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4465 ValueCounts.insert(std::make_pair(V, 0));
4466 unsigned &Count = ValueCounts[V];
4468 // Is this value dominant? (takes up more than half of the lanes)
4469 if (++Count > (NumElts / 2)) {
4470 hasDominantValue = true;
4474 if (ValueCounts.size() != 1)
4475 usesOnlyOneValue = false;
4476 if (!Value.getNode() && ValueCounts.size() > 0)
4477 Value = ValueCounts.begin()->first;
4479 if (ValueCounts.size() == 0)
4480 return DAG.getUNDEF(VT);
4482 if (isOnlyLowElement)
4483 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4485 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4487 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4488 // i32 and try again.
4489 if (hasDominantValue && EltSize <= 32) {
4493 // If we are VDUPing a value that comes directly from a vector, that will
4494 // cause an unnecessary move to and from a GPR, where instead we could
4495 // just use VDUPLANE. We can only do this if the lane being extracted
4496 // is at a constant index, as the VDUP from lane instructions only have
4497 // constant-index forms.
4498 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4499 isa<ConstantSDNode>(Value->getOperand(1))) {
4500 // We need to create a new undef vector to use for the VDUPLANE if the
4501 // size of the vector from which we get the value is different than the
4502 // size of the vector that we need to create. We will insert the element
4503 // such that the register coalescer will remove unnecessary copies.
4504 if (VT != Value->getOperand(0).getValueType()) {
4505 ConstantSDNode *constIndex;
4506 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4507 assert(constIndex && "The index is not a constant!");
4508 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4509 VT.getVectorNumElements();
4510 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4511 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4512 Value, DAG.getConstant(index, MVT::i32)),
4513 DAG.getConstant(index, MVT::i32));
4515 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4516 Value->getOperand(0), Value->getOperand(1));
4518 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4520 if (!usesOnlyOneValue) {
4521 // The dominant value was splatted as 'N', but we now have to insert
4522 // all differing elements.
4523 for (unsigned I = 0; I < NumElts; ++I) {
4524 if (Op.getOperand(I) == Value)
4526 SmallVector<SDValue, 3> Ops;
4528 Ops.push_back(Op.getOperand(I));
4529 Ops.push_back(DAG.getConstant(I, MVT::i32));
4530 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4535 if (VT.getVectorElementType().isFloatingPoint()) {
4536 SmallVector<SDValue, 8> Ops;
4537 for (unsigned i = 0; i < NumElts; ++i)
4538 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4540 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4541 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4542 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4544 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4546 if (usesOnlyOneValue) {
4547 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4548 if (isConstant && Val.getNode())
4549 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4553 // If all elements are constants and the case above didn't get hit, fall back
4554 // to the default expansion, which will generate a load from the constant
4559 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4561 SDValue shuffle = ReconstructShuffle(Op, DAG);
4562 if (shuffle != SDValue())
4566 // Vectors with 32- or 64-bit elements can be built by directly assigning
4567 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4568 // will be legalized.
4569 if (EltSize >= 32) {
4570 // Do the expansion with floating-point types, since that is what the VFP
4571 // registers are defined to use, and since i64 is not legal.
4572 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4573 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4574 SmallVector<SDValue, 8> Ops;
4575 for (unsigned i = 0; i < NumElts; ++i)
4576 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4577 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4578 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4584 // Gather data to see if the operation can be modelled as a
4585 // shuffle in combination with VEXTs.
4586 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4587 SelectionDAG &DAG) const {
4588 DebugLoc dl = Op.getDebugLoc();
4589 EVT VT = Op.getValueType();
4590 unsigned NumElts = VT.getVectorNumElements();
4592 SmallVector<SDValue, 2> SourceVecs;
4593 SmallVector<unsigned, 2> MinElts;
4594 SmallVector<unsigned, 2> MaxElts;
4596 for (unsigned i = 0; i < NumElts; ++i) {
4597 SDValue V = Op.getOperand(i);
4598 if (V.getOpcode() == ISD::UNDEF)
4600 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4601 // A shuffle can only come from building a vector from various
4602 // elements of other vectors.
4604 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4605 VT.getVectorElementType()) {
4606 // This code doesn't know how to handle shuffles where the vector
4607 // element types do not match (this happens because type legalization
4608 // promotes the return type of EXTRACT_VECTOR_ELT).
4609 // FIXME: It might be appropriate to extend this code to handle
4610 // mismatched types.
4614 // Record this extraction against the appropriate vector if possible...
4615 SDValue SourceVec = V.getOperand(0);
4616 // If the element number isn't a constant, we can't effectively
4617 // analyze what's going on.
4618 if (!isa<ConstantSDNode>(V.getOperand(1)))
4620 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4621 bool FoundSource = false;
4622 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4623 if (SourceVecs[j] == SourceVec) {
4624 if (MinElts[j] > EltNo)
4626 if (MaxElts[j] < EltNo)
4633 // Or record a new source if not...
4635 SourceVecs.push_back(SourceVec);
4636 MinElts.push_back(EltNo);
4637 MaxElts.push_back(EltNo);
4641 // Currently only do something sane when at most two source vectors
4643 if (SourceVecs.size() > 2)
4646 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4647 int VEXTOffsets[2] = {0, 0};
4649 // This loop extracts the usage patterns of the source vectors
4650 // and prepares appropriate SDValues for a shuffle if possible.
4651 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4652 if (SourceVecs[i].getValueType() == VT) {
4653 // No VEXT necessary
4654 ShuffleSrcs[i] = SourceVecs[i];
4657 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4658 // It probably isn't worth padding out a smaller vector just to
4659 // break it down again in a shuffle.
4663 // Since only 64-bit and 128-bit vectors are legal on ARM and
4664 // we've eliminated the other cases...
4665 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4666 "unexpected vector sizes in ReconstructShuffle");
4668 if (MaxElts[i] - MinElts[i] >= NumElts) {
4669 // Span too large for a VEXT to cope
4673 if (MinElts[i] >= NumElts) {
4674 // The extraction can just take the second half
4675 VEXTOffsets[i] = NumElts;
4676 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4678 DAG.getIntPtrConstant(NumElts));
4679 } else if (MaxElts[i] < NumElts) {
4680 // The extraction can just take the first half
4682 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4684 DAG.getIntPtrConstant(0));
4686 // An actual VEXT is needed
4687 VEXTOffsets[i] = MinElts[i];
4688 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4690 DAG.getIntPtrConstant(0));
4691 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4693 DAG.getIntPtrConstant(NumElts));
4694 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4695 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4699 SmallVector<int, 8> Mask;
4701 for (unsigned i = 0; i < NumElts; ++i) {
4702 SDValue Entry = Op.getOperand(i);
4703 if (Entry.getOpcode() == ISD::UNDEF) {
4708 SDValue ExtractVec = Entry.getOperand(0);
4709 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4710 .getOperand(1))->getSExtValue();
4711 if (ExtractVec == SourceVecs[0]) {
4712 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4714 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4718 // Final check before we try to produce nonsense...
4719 if (isShuffleMaskLegal(Mask, VT))
4720 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4726 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4727 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4728 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4729 /// are assumed to be legal.
4731 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4733 if (VT.getVectorNumElements() == 4 &&
4734 (VT.is128BitVector() || VT.is64BitVector())) {
4735 unsigned PFIndexes[4];
4736 for (unsigned i = 0; i != 4; ++i) {
4740 PFIndexes[i] = M[i];
4743 // Compute the index in the perfect shuffle table.
4744 unsigned PFTableIndex =
4745 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4746 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4747 unsigned Cost = (PFEntry >> 30);
4754 unsigned Imm, WhichResult;
4756 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4757 return (EltSize >= 32 ||
4758 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4759 isVREVMask(M, VT, 64) ||
4760 isVREVMask(M, VT, 32) ||
4761 isVREVMask(M, VT, 16) ||
4762 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4763 isVTBLMask(M, VT) ||
4764 isVTRNMask(M, VT, WhichResult) ||
4765 isVUZPMask(M, VT, WhichResult) ||
4766 isVZIPMask(M, VT, WhichResult) ||
4767 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4768 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4769 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
4770 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
4773 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4774 /// the specified operations to build the shuffle.
4775 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4776 SDValue RHS, SelectionDAG &DAG,
4778 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4779 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4780 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4783 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4792 OP_VUZPL, // VUZP, left result
4793 OP_VUZPR, // VUZP, right result
4794 OP_VZIPL, // VZIP, left result
4795 OP_VZIPR, // VZIP, right result
4796 OP_VTRNL, // VTRN, left result
4797 OP_VTRNR // VTRN, right result
4800 if (OpNum == OP_COPY) {
4801 if (LHSID == (1*9+2)*9+3) return LHS;
4802 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4806 SDValue OpLHS, OpRHS;
4807 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4808 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4809 EVT VT = OpLHS.getValueType();
4812 default: llvm_unreachable("Unknown shuffle opcode!");
4814 // VREV divides the vector in half and swaps within the half.
4815 if (VT.getVectorElementType() == MVT::i32 ||
4816 VT.getVectorElementType() == MVT::f32)
4817 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4818 // vrev <4 x i16> -> VREV32
4819 if (VT.getVectorElementType() == MVT::i16)
4820 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4821 // vrev <4 x i8> -> VREV16
4822 assert(VT.getVectorElementType() == MVT::i8);
4823 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4828 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4829 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4833 return DAG.getNode(ARMISD::VEXT, dl, VT,
4835 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4838 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4839 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4842 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4843 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4846 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4847 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4851 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4852 ArrayRef<int> ShuffleMask,
4853 SelectionDAG &DAG) {
4854 // Check to see if we can use the VTBL instruction.
4855 SDValue V1 = Op.getOperand(0);
4856 SDValue V2 = Op.getOperand(1);
4857 DebugLoc DL = Op.getDebugLoc();
4859 SmallVector<SDValue, 8> VTBLMask;
4860 for (ArrayRef<int>::iterator
4861 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4862 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4864 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4865 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4866 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4869 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4870 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4874 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
4875 SelectionDAG &DAG) {
4876 DebugLoc DL = Op.getDebugLoc();
4877 SDValue OpLHS = Op.getOperand(0);
4878 EVT VT = OpLHS.getValueType();
4880 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
4881 "Expect an v8i16/v16i8 type");
4882 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
4883 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
4884 // extract the first 8 bytes into the top double word and the last 8 bytes
4885 // into the bottom double word. The v8i16 case is similar.
4886 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
4887 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
4888 DAG.getConstant(ExtractNum, MVT::i32));
4891 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4892 SDValue V1 = Op.getOperand(0);
4893 SDValue V2 = Op.getOperand(1);
4894 DebugLoc dl = Op.getDebugLoc();
4895 EVT VT = Op.getValueType();
4896 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4898 // Convert shuffles that are directly supported on NEON to target-specific
4899 // DAG nodes, instead of keeping them as shuffles and matching them again
4900 // during code selection. This is more efficient and avoids the possibility
4901 // of inconsistencies between legalization and selection.
4902 // FIXME: floating-point vectors should be canonicalized to integer vectors
4903 // of the same time so that they get CSEd properly.
4904 ArrayRef<int> ShuffleMask = SVN->getMask();
4906 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4907 if (EltSize <= 32) {
4908 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4909 int Lane = SVN->getSplatIndex();
4910 // If this is undef splat, generate it via "just" vdup, if possible.
4911 if (Lane == -1) Lane = 0;
4913 // Test if V1 is a SCALAR_TO_VECTOR.
4914 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4915 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4917 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4918 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4920 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4921 !isa<ConstantSDNode>(V1.getOperand(0))) {
4922 bool IsScalarToVector = true;
4923 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4924 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4925 IsScalarToVector = false;
4928 if (IsScalarToVector)
4929 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4931 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4932 DAG.getConstant(Lane, MVT::i32));
4937 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4940 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4941 DAG.getConstant(Imm, MVT::i32));
4944 if (isVREVMask(ShuffleMask, VT, 64))
4945 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4946 if (isVREVMask(ShuffleMask, VT, 32))
4947 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4948 if (isVREVMask(ShuffleMask, VT, 16))
4949 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4951 if (V2->getOpcode() == ISD::UNDEF &&
4952 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4953 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4954 DAG.getConstant(Imm, MVT::i32));
4957 // Check for Neon shuffles that modify both input vectors in place.
4958 // If both results are used, i.e., if there are two shuffles with the same
4959 // source operands and with masks corresponding to both results of one of
4960 // these operations, DAG memoization will ensure that a single node is
4961 // used for both shuffles.
4962 unsigned WhichResult;
4963 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4964 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4965 V1, V2).getValue(WhichResult);
4966 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4967 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4968 V1, V2).getValue(WhichResult);
4969 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4970 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4971 V1, V2).getValue(WhichResult);
4973 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4974 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4975 V1, V1).getValue(WhichResult);
4976 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4977 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4978 V1, V1).getValue(WhichResult);
4979 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4980 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4981 V1, V1).getValue(WhichResult);
4984 // If the shuffle is not directly supported and it has 4 elements, use
4985 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4986 unsigned NumElts = VT.getVectorNumElements();
4988 unsigned PFIndexes[4];
4989 for (unsigned i = 0; i != 4; ++i) {
4990 if (ShuffleMask[i] < 0)
4993 PFIndexes[i] = ShuffleMask[i];
4996 // Compute the index in the perfect shuffle table.
4997 unsigned PFTableIndex =
4998 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4999 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5000 unsigned Cost = (PFEntry >> 30);
5003 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5006 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5007 if (EltSize >= 32) {
5008 // Do the expansion with floating-point types, since that is what the VFP
5009 // registers are defined to use, and since i64 is not legal.
5010 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5011 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5012 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5013 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5014 SmallVector<SDValue, 8> Ops;
5015 for (unsigned i = 0; i < NumElts; ++i) {
5016 if (ShuffleMask[i] < 0)
5017 Ops.push_back(DAG.getUNDEF(EltVT));
5019 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5020 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5021 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5024 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5025 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5028 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5029 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5031 if (VT == MVT::v8i8) {
5032 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5033 if (NewOp.getNode())
5040 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5041 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5042 SDValue Lane = Op.getOperand(2);
5043 if (!isa<ConstantSDNode>(Lane))
5049 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5050 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5051 SDValue Lane = Op.getOperand(1);
5052 if (!isa<ConstantSDNode>(Lane))
5055 SDValue Vec = Op.getOperand(0);
5056 if (Op.getValueType() == MVT::i32 &&
5057 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5058 DebugLoc dl = Op.getDebugLoc();
5059 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5065 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5066 // The only time a CONCAT_VECTORS operation can have legal types is when
5067 // two 64-bit vectors are concatenated to a 128-bit vector.
5068 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5069 "unexpected CONCAT_VECTORS");
5070 DebugLoc dl = Op.getDebugLoc();
5071 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5072 SDValue Op0 = Op.getOperand(0);
5073 SDValue Op1 = Op.getOperand(1);
5074 if (Op0.getOpcode() != ISD::UNDEF)
5075 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5076 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5077 DAG.getIntPtrConstant(0));
5078 if (Op1.getOpcode() != ISD::UNDEF)
5079 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5080 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5081 DAG.getIntPtrConstant(1));
5082 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5085 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5086 /// element has been zero/sign-extended, depending on the isSigned parameter,
5087 /// from an integer type half its size.
5088 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5090 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5091 EVT VT = N->getValueType(0);
5092 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5093 SDNode *BVN = N->getOperand(0).getNode();
5094 if (BVN->getValueType(0) != MVT::v4i32 ||
5095 BVN->getOpcode() != ISD::BUILD_VECTOR)
5097 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5098 unsigned HiElt = 1 - LoElt;
5099 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5100 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5101 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5102 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5103 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5106 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5107 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5110 if (Hi0->isNullValue() && Hi1->isNullValue())
5116 if (N->getOpcode() != ISD::BUILD_VECTOR)
5119 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5120 SDNode *Elt = N->getOperand(i).getNode();
5121 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5122 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5123 unsigned HalfSize = EltSize / 2;
5125 if (!isIntN(HalfSize, C->getSExtValue()))
5128 if (!isUIntN(HalfSize, C->getZExtValue()))
5139 /// isSignExtended - Check if a node is a vector value that is sign-extended
5140 /// or a constant BUILD_VECTOR with sign-extended elements.
5141 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5142 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5144 if (isExtendedBUILD_VECTOR(N, DAG, true))
5149 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5150 /// or a constant BUILD_VECTOR with zero-extended elements.
5151 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5152 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5154 if (isExtendedBUILD_VECTOR(N, DAG, false))
5159 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5160 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5161 /// We insert the required extension here to get the vector to fill a D register.
5162 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5165 unsigned ExtOpcode) {
5166 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5167 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5168 // 64-bits we need to insert a new extension so that it will be 64-bits.
5169 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5170 if (OrigTy.getSizeInBits() >= 64)
5173 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5174 MVT::SimpleValueType OrigSimpleTy = OrigTy.getSimpleVT().SimpleTy;
5176 switch (OrigSimpleTy) {
5177 default: llvm_unreachable("Unexpected Orig Vector Type");
5186 return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N);
5189 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5190 /// does not do any sign/zero extension. If the original vector is less
5191 /// than 64 bits, an appropriate extension will be added after the load to
5192 /// reach a total size of 64 bits. We have to add the extension separately
5193 /// because ARM does not have a sign/zero extending load for vectors.
5194 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5195 SDValue NonExtendingLoad =
5196 DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(),
5197 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5198 LD->isNonTemporal(), LD->isInvariant(),
5199 LD->getAlignment());
5201 switch (LD->getExtensionType()) {
5202 default: llvm_unreachable("Unexpected LoadExtType");
5204 case ISD::SEXTLOAD: ExtOp = ISD::SIGN_EXTEND; break;
5205 case ISD::ZEXTLOAD: ExtOp = ISD::ZERO_EXTEND; break;
5207 MVT::SimpleValueType MemType = LD->getMemoryVT().getSimpleVT().SimpleTy;
5208 MVT::SimpleValueType ExtType = LD->getValueType(0).getSimpleVT().SimpleTy;
5209 return AddRequiredExtensionForVMULL(NonExtendingLoad, DAG,
5210 MemType, ExtType, ExtOp);
5213 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5214 /// extending load, or BUILD_VECTOR with extended elements, return the
5215 /// unextended value. The unextended vector should be 64 bits so that it can
5216 /// be used as an operand to a VMULL instruction. If the original vector size
5217 /// before extension is less than 64 bits we add a an extension to resize
5218 /// the vector to 64 bits.
5219 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5220 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5221 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5222 N->getOperand(0)->getValueType(0),
5226 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5227 return SkipLoadExtensionForVMULL(LD, DAG);
5229 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5230 // have been legalized as a BITCAST from v4i32.
5231 if (N->getOpcode() == ISD::BITCAST) {
5232 SDNode *BVN = N->getOperand(0).getNode();
5233 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5234 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5235 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5236 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
5237 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5239 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5240 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5241 EVT VT = N->getValueType(0);
5242 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5243 unsigned NumElts = VT.getVectorNumElements();
5244 MVT TruncVT = MVT::getIntegerVT(EltSize);
5245 SmallVector<SDValue, 8> Ops;
5246 for (unsigned i = 0; i != NumElts; ++i) {
5247 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5248 const APInt &CInt = C->getAPIntValue();
5249 // Element types smaller than 32 bits are not legal, so use i32 elements.
5250 // The values are implicitly truncated so sext vs. zext doesn't matter.
5251 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5253 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5254 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5257 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5258 unsigned Opcode = N->getOpcode();
5259 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5260 SDNode *N0 = N->getOperand(0).getNode();
5261 SDNode *N1 = N->getOperand(1).getNode();
5262 return N0->hasOneUse() && N1->hasOneUse() &&
5263 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5268 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5269 unsigned Opcode = N->getOpcode();
5270 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5271 SDNode *N0 = N->getOperand(0).getNode();
5272 SDNode *N1 = N->getOperand(1).getNode();
5273 return N0->hasOneUse() && N1->hasOneUse() &&
5274 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5279 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5280 // Multiplications are only custom-lowered for 128-bit vectors so that
5281 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5282 EVT VT = Op.getValueType();
5283 assert(VT.is128BitVector() && VT.isInteger() &&
5284 "unexpected type for custom-lowering ISD::MUL");
5285 SDNode *N0 = Op.getOperand(0).getNode();
5286 SDNode *N1 = Op.getOperand(1).getNode();
5287 unsigned NewOpc = 0;
5289 bool isN0SExt = isSignExtended(N0, DAG);
5290 bool isN1SExt = isSignExtended(N1, DAG);
5291 if (isN0SExt && isN1SExt)
5292 NewOpc = ARMISD::VMULLs;
5294 bool isN0ZExt = isZeroExtended(N0, DAG);
5295 bool isN1ZExt = isZeroExtended(N1, DAG);
5296 if (isN0ZExt && isN1ZExt)
5297 NewOpc = ARMISD::VMULLu;
5298 else if (isN1SExt || isN1ZExt) {
5299 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5300 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5301 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5302 NewOpc = ARMISD::VMULLs;
5304 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5305 NewOpc = ARMISD::VMULLu;
5307 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5309 NewOpc = ARMISD::VMULLu;
5315 if (VT == MVT::v2i64)
5316 // Fall through to expand this. It is not legal.
5319 // Other vector multiplications are legal.
5324 // Legalize to a VMULL instruction.
5325 DebugLoc DL = Op.getDebugLoc();
5327 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5329 Op0 = SkipExtensionForVMULL(N0, DAG);
5330 assert(Op0.getValueType().is64BitVector() &&
5331 Op1.getValueType().is64BitVector() &&
5332 "unexpected types for extended operands to VMULL");
5333 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5336 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5337 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5344 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5345 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5346 EVT Op1VT = Op1.getValueType();
5347 return DAG.getNode(N0->getOpcode(), DL, VT,
5348 DAG.getNode(NewOpc, DL, VT,
5349 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5350 DAG.getNode(NewOpc, DL, VT,
5351 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5355 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5357 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5358 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5359 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5360 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5361 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5362 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5363 // Get reciprocal estimate.
5364 // float4 recip = vrecpeq_f32(yf);
5365 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5366 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5367 // Because char has a smaller range than uchar, we can actually get away
5368 // without any newton steps. This requires that we use a weird bias
5369 // of 0xb000, however (again, this has been exhaustively tested).
5370 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5371 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5372 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5373 Y = DAG.getConstant(0xb000, MVT::i32);
5374 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5375 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5376 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5377 // Convert back to short.
5378 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5379 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5384 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5386 // Convert to float.
5387 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5388 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5389 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5390 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5391 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5392 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5394 // Use reciprocal estimate and one refinement step.
5395 // float4 recip = vrecpeq_f32(yf);
5396 // recip *= vrecpsq_f32(yf, recip);
5397 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5398 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5399 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5400 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5402 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5403 // Because short has a smaller range than ushort, we can actually get away
5404 // with only a single newton step. This requires that we use a weird bias
5405 // of 89, however (again, this has been exhaustively tested).
5406 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5407 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5408 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5409 N1 = DAG.getConstant(0x89, MVT::i32);
5410 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5411 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5412 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5413 // Convert back to integer and return.
5414 // return vmovn_s32(vcvt_s32_f32(result));
5415 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5416 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5420 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5421 EVT VT = Op.getValueType();
5422 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5423 "unexpected type for custom-lowering ISD::SDIV");
5425 DebugLoc dl = Op.getDebugLoc();
5426 SDValue N0 = Op.getOperand(0);
5427 SDValue N1 = Op.getOperand(1);
5430 if (VT == MVT::v8i8) {
5431 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5432 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5434 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5435 DAG.getIntPtrConstant(4));
5436 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5437 DAG.getIntPtrConstant(4));
5438 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5439 DAG.getIntPtrConstant(0));
5440 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5441 DAG.getIntPtrConstant(0));
5443 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5444 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5446 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5447 N0 = LowerCONCAT_VECTORS(N0, DAG);
5449 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5452 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5455 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5456 EVT VT = Op.getValueType();
5457 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5458 "unexpected type for custom-lowering ISD::UDIV");
5460 DebugLoc dl = Op.getDebugLoc();
5461 SDValue N0 = Op.getOperand(0);
5462 SDValue N1 = Op.getOperand(1);
5465 if (VT == MVT::v8i8) {
5466 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5467 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5469 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5470 DAG.getIntPtrConstant(4));
5471 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5472 DAG.getIntPtrConstant(4));
5473 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5474 DAG.getIntPtrConstant(0));
5475 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5476 DAG.getIntPtrConstant(0));
5478 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5479 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5481 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5482 N0 = LowerCONCAT_VECTORS(N0, DAG);
5484 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5485 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5490 // v4i16 sdiv ... Convert to float.
5491 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5492 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5493 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5494 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5495 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5496 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5498 // Use reciprocal estimate and two refinement steps.
5499 // float4 recip = vrecpeq_f32(yf);
5500 // recip *= vrecpsq_f32(yf, recip);
5501 // recip *= vrecpsq_f32(yf, recip);
5502 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5503 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5504 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5505 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5507 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5508 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5509 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5511 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5512 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5513 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5514 // and that it will never cause us to return an answer too large).
5515 // float4 result = as_float4(as_int4(xf*recip) + 2);
5516 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5517 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5518 N1 = DAG.getConstant(2, MVT::i32);
5519 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5520 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5521 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5522 // Convert back to integer and return.
5523 // return vmovn_u32(vcvt_s32_f32(result));
5524 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5525 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5529 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5530 EVT VT = Op.getNode()->getValueType(0);
5531 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5534 bool ExtraOp = false;
5535 switch (Op.getOpcode()) {
5536 default: llvm_unreachable("Invalid code");
5537 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5538 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5539 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5540 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5544 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5546 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5547 Op.getOperand(1), Op.getOperand(2));
5550 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5551 // Monotonic load/store is legal for all targets
5552 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5555 // Aquire/Release load/store is not legal for targets without a
5556 // dmb or equivalent available.
5562 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5563 SelectionDAG &DAG, unsigned NewOp) {
5564 DebugLoc dl = Node->getDebugLoc();
5565 assert (Node->getValueType(0) == MVT::i64 &&
5566 "Only know how to expand i64 atomics");
5568 SmallVector<SDValue, 6> Ops;
5569 Ops.push_back(Node->getOperand(0)); // Chain
5570 Ops.push_back(Node->getOperand(1)); // Ptr
5572 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5573 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5574 // High part of Val1
5575 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5576 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5577 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5578 // High part of Val1
5579 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5580 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5581 // High part of Val2
5582 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5583 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5585 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5587 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5588 cast<MemSDNode>(Node)->getMemOperand());
5589 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5590 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5591 Results.push_back(Result.getValue(2));
5594 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5595 switch (Op.getOpcode()) {
5596 default: llvm_unreachable("Don't know how to custom lower this!");
5597 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5598 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5599 case ISD::GlobalAddress:
5600 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5601 LowerGlobalAddressELF(Op, DAG);
5602 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5603 case ISD::SELECT: return LowerSELECT(Op, DAG);
5604 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5605 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5606 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5607 case ISD::VASTART: return LowerVASTART(Op, DAG);
5608 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5609 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5610 case ISD::SINT_TO_FP:
5611 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5612 case ISD::FP_TO_SINT:
5613 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5614 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5615 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5616 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5617 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5618 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5619 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5620 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5622 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5625 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5626 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5627 case ISD::SRL_PARTS:
5628 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5629 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5630 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
5631 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5632 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5633 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5634 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5635 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5636 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5637 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5638 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5639 case ISD::MUL: return LowerMUL(Op, DAG);
5640 case ISD::SDIV: return LowerSDIV(Op, DAG);
5641 case ISD::UDIV: return LowerUDIV(Op, DAG);
5645 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5646 case ISD::ATOMIC_LOAD:
5647 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5651 /// ReplaceNodeResults - Replace the results of node with an illegal result
5652 /// type with new values built out of custom code.
5653 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5654 SmallVectorImpl<SDValue>&Results,
5655 SelectionDAG &DAG) const {
5657 switch (N->getOpcode()) {
5659 llvm_unreachable("Don't know how to custom expand this!");
5661 Res = ExpandBITCAST(N, DAG);
5663 case ISD::SIGN_EXTEND:
5664 case ISD::ZERO_EXTEND:
5665 Res = ExpandVectorExtension(N, DAG);
5669 Res = Expand64BitShift(N, DAG, Subtarget);
5671 case ISD::ATOMIC_LOAD_ADD:
5672 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5674 case ISD::ATOMIC_LOAD_AND:
5675 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5677 case ISD::ATOMIC_LOAD_NAND:
5678 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5680 case ISD::ATOMIC_LOAD_OR:
5681 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5683 case ISD::ATOMIC_LOAD_SUB:
5684 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5686 case ISD::ATOMIC_LOAD_XOR:
5687 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5689 case ISD::ATOMIC_SWAP:
5690 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5692 case ISD::ATOMIC_CMP_SWAP:
5693 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5695 case ISD::ATOMIC_LOAD_MIN:
5696 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
5698 case ISD::ATOMIC_LOAD_UMIN:
5699 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
5701 case ISD::ATOMIC_LOAD_MAX:
5702 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
5704 case ISD::ATOMIC_LOAD_UMAX:
5705 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
5709 Results.push_back(Res);
5712 //===----------------------------------------------------------------------===//
5713 // ARM Scheduler Hooks
5714 //===----------------------------------------------------------------------===//
5717 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5718 MachineBasicBlock *BB,
5719 unsigned Size) const {
5720 unsigned dest = MI->getOperand(0).getReg();
5721 unsigned ptr = MI->getOperand(1).getReg();
5722 unsigned oldval = MI->getOperand(2).getReg();
5723 unsigned newval = MI->getOperand(3).getReg();
5724 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5725 DebugLoc dl = MI->getDebugLoc();
5726 bool isThumb2 = Subtarget->isThumb2();
5728 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5729 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5730 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5731 (const TargetRegisterClass*)&ARM::GPRRegClass);
5734 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5735 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5736 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5739 unsigned ldrOpc, strOpc;
5741 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5743 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5744 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5747 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5748 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5751 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5752 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5756 MachineFunction *MF = BB->getParent();
5757 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5758 MachineFunction::iterator It = BB;
5759 ++It; // insert the new blocks after the current block
5761 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5762 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5763 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5764 MF->insert(It, loop1MBB);
5765 MF->insert(It, loop2MBB);
5766 MF->insert(It, exitMBB);
5768 // Transfer the remainder of BB and its successor edges to exitMBB.
5769 exitMBB->splice(exitMBB->begin(), BB,
5770 llvm::next(MachineBasicBlock::iterator(MI)),
5772 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5776 // fallthrough --> loop1MBB
5777 BB->addSuccessor(loop1MBB);
5780 // ldrex dest, [ptr]
5784 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5785 if (ldrOpc == ARM::t2LDREX)
5787 AddDefaultPred(MIB);
5788 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5789 .addReg(dest).addReg(oldval));
5790 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5791 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5792 BB->addSuccessor(loop2MBB);
5793 BB->addSuccessor(exitMBB);
5796 // strex scratch, newval, [ptr]
5800 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5801 if (strOpc == ARM::t2STREX)
5803 AddDefaultPred(MIB);
5804 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5805 .addReg(scratch).addImm(0));
5806 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5807 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5808 BB->addSuccessor(loop1MBB);
5809 BB->addSuccessor(exitMBB);
5815 MI->eraseFromParent(); // The instruction is gone now.
5821 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5822 unsigned Size, unsigned BinOpcode) const {
5823 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5824 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5826 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5827 MachineFunction *MF = BB->getParent();
5828 MachineFunction::iterator It = BB;
5831 unsigned dest = MI->getOperand(0).getReg();
5832 unsigned ptr = MI->getOperand(1).getReg();
5833 unsigned incr = MI->getOperand(2).getReg();
5834 DebugLoc dl = MI->getDebugLoc();
5835 bool isThumb2 = Subtarget->isThumb2();
5837 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5839 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5840 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5843 unsigned ldrOpc, strOpc;
5845 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5847 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5848 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5851 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5852 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5855 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5856 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5860 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5861 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5862 MF->insert(It, loopMBB);
5863 MF->insert(It, exitMBB);
5865 // Transfer the remainder of BB and its successor edges to exitMBB.
5866 exitMBB->splice(exitMBB->begin(), BB,
5867 llvm::next(MachineBasicBlock::iterator(MI)),
5869 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5871 const TargetRegisterClass *TRC = isThumb2 ?
5872 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5873 (const TargetRegisterClass*)&ARM::GPRRegClass;
5874 unsigned scratch = MRI.createVirtualRegister(TRC);
5875 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5879 // fallthrough --> loopMBB
5880 BB->addSuccessor(loopMBB);
5884 // <binop> scratch2, dest, incr
5885 // strex scratch, scratch2, ptr
5888 // fallthrough --> exitMBB
5890 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5891 if (ldrOpc == ARM::t2LDREX)
5893 AddDefaultPred(MIB);
5895 // operand order needs to go the other way for NAND
5896 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5897 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5898 addReg(incr).addReg(dest)).addReg(0);
5900 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5901 addReg(dest).addReg(incr)).addReg(0);
5904 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5905 if (strOpc == ARM::t2STREX)
5907 AddDefaultPred(MIB);
5908 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5909 .addReg(scratch).addImm(0));
5910 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5911 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5913 BB->addSuccessor(loopMBB);
5914 BB->addSuccessor(exitMBB);
5920 MI->eraseFromParent(); // The instruction is gone now.
5926 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5927 MachineBasicBlock *BB,
5930 ARMCC::CondCodes Cond) const {
5931 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5933 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5934 MachineFunction *MF = BB->getParent();
5935 MachineFunction::iterator It = BB;
5938 unsigned dest = MI->getOperand(0).getReg();
5939 unsigned ptr = MI->getOperand(1).getReg();
5940 unsigned incr = MI->getOperand(2).getReg();
5941 unsigned oldval = dest;
5942 DebugLoc dl = MI->getDebugLoc();
5943 bool isThumb2 = Subtarget->isThumb2();
5945 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5947 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5948 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5951 unsigned ldrOpc, strOpc, extendOpc;
5953 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5955 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5956 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5957 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5960 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5961 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5962 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5965 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5966 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5971 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5972 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5973 MF->insert(It, loopMBB);
5974 MF->insert(It, exitMBB);
5976 // Transfer the remainder of BB and its successor edges to exitMBB.
5977 exitMBB->splice(exitMBB->begin(), BB,
5978 llvm::next(MachineBasicBlock::iterator(MI)),
5980 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5982 const TargetRegisterClass *TRC = isThumb2 ?
5983 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5984 (const TargetRegisterClass*)&ARM::GPRRegClass;
5985 unsigned scratch = MRI.createVirtualRegister(TRC);
5986 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5990 // fallthrough --> loopMBB
5991 BB->addSuccessor(loopMBB);
5995 // (sign extend dest, if required)
5997 // cmov.cond scratch2, incr, dest
5998 // strex scratch, scratch2, ptr
6001 // fallthrough --> exitMBB
6003 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6004 if (ldrOpc == ARM::t2LDREX)
6006 AddDefaultPred(MIB);
6008 // Sign extend the value, if necessary.
6009 if (signExtend && extendOpc) {
6010 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
6011 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6016 // Build compare and cmov instructions.
6017 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6018 .addReg(oldval).addReg(incr));
6019 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6020 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6022 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6023 if (strOpc == ARM::t2STREX)
6025 AddDefaultPred(MIB);
6026 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6027 .addReg(scratch).addImm(0));
6028 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6029 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6031 BB->addSuccessor(loopMBB);
6032 BB->addSuccessor(exitMBB);
6038 MI->eraseFromParent(); // The instruction is gone now.
6044 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6045 unsigned Op1, unsigned Op2,
6046 bool NeedsCarry, bool IsCmpxchg,
6047 bool IsMinMax, ARMCC::CondCodes CC) const {
6048 // This also handles ATOMIC_SWAP, indicated by Op1==0.
6049 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6051 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6052 MachineFunction *MF = BB->getParent();
6053 MachineFunction::iterator It = BB;
6056 unsigned destlo = MI->getOperand(0).getReg();
6057 unsigned desthi = MI->getOperand(1).getReg();
6058 unsigned ptr = MI->getOperand(2).getReg();
6059 unsigned vallo = MI->getOperand(3).getReg();
6060 unsigned valhi = MI->getOperand(4).getReg();
6061 DebugLoc dl = MI->getDebugLoc();
6062 bool isThumb2 = Subtarget->isThumb2();
6064 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6066 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6067 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6068 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6071 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6072 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6073 if (IsCmpxchg || IsMinMax)
6074 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6076 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6077 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6079 MF->insert(It, loopMBB);
6080 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6081 if (IsCmpxchg) MF->insert(It, cont2BB);
6082 MF->insert(It, exitMBB);
6084 // Transfer the remainder of BB and its successor edges to exitMBB.
6085 exitMBB->splice(exitMBB->begin(), BB,
6086 llvm::next(MachineBasicBlock::iterator(MI)),
6088 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6090 const TargetRegisterClass *TRC = isThumb2 ?
6091 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6092 (const TargetRegisterClass*)&ARM::GPRRegClass;
6093 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6097 // fallthrough --> loopMBB
6098 BB->addSuccessor(loopMBB);
6101 // ldrexd r2, r3, ptr
6102 // <binopa> r0, r2, incr
6103 // <binopb> r1, r3, incr
6104 // strexd storesuccess, r0, r1, ptr
6105 // cmp storesuccess, #0
6107 // fallthrough --> exitMBB
6112 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD))
6113 .addReg(destlo, RegState::Define)
6114 .addReg(desthi, RegState::Define)
6117 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6118 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD))
6119 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6120 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6121 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6122 .addReg(GPRPair0, 0, ARM::gsub_0);
6123 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6124 .addReg(GPRPair0, 0, ARM::gsub_1);
6127 unsigned StoreLo, StoreHi;
6130 for (unsigned i = 0; i < 2; i++) {
6131 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6133 .addReg(i == 0 ? destlo : desthi)
6134 .addReg(i == 0 ? vallo : valhi));
6135 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6136 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6137 BB->addSuccessor(exitMBB);
6138 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6139 BB = (i == 0 ? contBB : cont2BB);
6142 // Copy to physregs for strexd
6143 StoreLo = MI->getOperand(5).getReg();
6144 StoreHi = MI->getOperand(6).getReg();
6146 // Perform binary operation
6147 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6148 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6149 .addReg(destlo).addReg(vallo))
6150 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6151 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6152 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6153 .addReg(desthi).addReg(valhi))
6154 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6159 // Copy to physregs for strexd
6164 // Compare and branch to exit block.
6165 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6166 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6167 BB->addSuccessor(exitMBB);
6168 BB->addSuccessor(contBB);
6176 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess)
6177 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6179 // Marshal a pair...
6180 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6181 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6182 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6183 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6184 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6187 .addImm(ARM::gsub_0);
6188 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6191 .addImm(ARM::gsub_1);
6194 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess)
6195 .addReg(StorePair).addReg(ptr));
6198 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6199 .addReg(storesuccess).addImm(0));
6200 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6201 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6203 BB->addSuccessor(loopMBB);
6204 BB->addSuccessor(exitMBB);
6210 MI->eraseFromParent(); // The instruction is gone now.
6215 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6216 /// registers the function context.
6217 void ARMTargetLowering::
6218 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6219 MachineBasicBlock *DispatchBB, int FI) const {
6220 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6221 DebugLoc dl = MI->getDebugLoc();
6222 MachineFunction *MF = MBB->getParent();
6223 MachineRegisterInfo *MRI = &MF->getRegInfo();
6224 MachineConstantPool *MCP = MF->getConstantPool();
6225 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6226 const Function *F = MF->getFunction();
6228 bool isThumb = Subtarget->isThumb();
6229 bool isThumb2 = Subtarget->isThumb2();
6231 unsigned PCLabelId = AFI->createPICLabelUId();
6232 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6233 ARMConstantPoolValue *CPV =
6234 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6235 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6237 const TargetRegisterClass *TRC = isThumb ?
6238 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6239 (const TargetRegisterClass*)&ARM::GPRRegClass;
6241 // Grab constant pool and fixed stack memory operands.
6242 MachineMemOperand *CPMMO =
6243 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6244 MachineMemOperand::MOLoad, 4, 4);
6246 MachineMemOperand *FIMMOSt =
6247 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6248 MachineMemOperand::MOStore, 4, 4);
6250 // Load the address of the dispatch MBB into the jump buffer.
6252 // Incoming value: jbuf
6253 // ldr.n r5, LCPI1_1
6256 // str r5, [$jbuf, #+4] ; &jbuf[1]
6257 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6258 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6259 .addConstantPoolIndex(CPI)
6260 .addMemOperand(CPMMO));
6261 // Set the low bit because of thumb mode.
6262 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6264 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6265 .addReg(NewVReg1, RegState::Kill)
6267 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6268 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6269 .addReg(NewVReg2, RegState::Kill)
6271 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6272 .addReg(NewVReg3, RegState::Kill)
6274 .addImm(36) // &jbuf[1] :: pc
6275 .addMemOperand(FIMMOSt));
6276 } else if (isThumb) {
6277 // Incoming value: jbuf
6278 // ldr.n r1, LCPI1_4
6282 // add r2, $jbuf, #+4 ; &jbuf[1]
6284 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6285 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6286 .addConstantPoolIndex(CPI)
6287 .addMemOperand(CPMMO));
6288 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6289 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6290 .addReg(NewVReg1, RegState::Kill)
6292 // Set the low bit because of thumb mode.
6293 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6294 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6295 .addReg(ARM::CPSR, RegState::Define)
6297 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6298 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6299 .addReg(ARM::CPSR, RegState::Define)
6300 .addReg(NewVReg2, RegState::Kill)
6301 .addReg(NewVReg3, RegState::Kill));
6302 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6303 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6305 .addImm(36)); // &jbuf[1] :: pc
6306 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6307 .addReg(NewVReg4, RegState::Kill)
6308 .addReg(NewVReg5, RegState::Kill)
6310 .addMemOperand(FIMMOSt));
6312 // Incoming value: jbuf
6315 // str r1, [$jbuf, #+4] ; &jbuf[1]
6316 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6317 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6318 .addConstantPoolIndex(CPI)
6320 .addMemOperand(CPMMO));
6321 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6322 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6323 .addReg(NewVReg1, RegState::Kill)
6324 .addImm(PCLabelId));
6325 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6326 .addReg(NewVReg2, RegState::Kill)
6328 .addImm(36) // &jbuf[1] :: pc
6329 .addMemOperand(FIMMOSt));
6333 MachineBasicBlock *ARMTargetLowering::
6334 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6335 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6336 DebugLoc dl = MI->getDebugLoc();
6337 MachineFunction *MF = MBB->getParent();
6338 MachineRegisterInfo *MRI = &MF->getRegInfo();
6339 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6340 MachineFrameInfo *MFI = MF->getFrameInfo();
6341 int FI = MFI->getFunctionContextIndex();
6343 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6344 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6345 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6347 // Get a mapping of the call site numbers to all of the landing pads they're
6349 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6350 unsigned MaxCSNum = 0;
6351 MachineModuleInfo &MMI = MF->getMMI();
6352 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6354 if (!BB->isLandingPad()) continue;
6356 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6358 for (MachineBasicBlock::iterator
6359 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6360 if (!II->isEHLabel()) continue;
6362 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6363 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6365 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6366 for (SmallVectorImpl<unsigned>::iterator
6367 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6368 CSI != CSE; ++CSI) {
6369 CallSiteNumToLPad[*CSI].push_back(BB);
6370 MaxCSNum = std::max(MaxCSNum, *CSI);
6376 // Get an ordered list of the machine basic blocks for the jump table.
6377 std::vector<MachineBasicBlock*> LPadList;
6378 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6379 LPadList.reserve(CallSiteNumToLPad.size());
6380 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6381 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6382 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6383 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6384 LPadList.push_back(*II);
6385 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6389 assert(!LPadList.empty() &&
6390 "No landing pad destinations for the dispatch jump table!");
6392 // Create the jump table and associated information.
6393 MachineJumpTableInfo *JTI =
6394 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6395 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6396 unsigned UId = AFI->createJumpTableUId();
6397 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6399 // Create the MBBs for the dispatch code.
6401 // Shove the dispatch's address into the return slot in the function context.
6402 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6403 DispatchBB->setIsLandingPad();
6405 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6406 unsigned trap_opcode;
6407 if (Subtarget->isThumb())
6408 trap_opcode = ARM::tTRAP;
6410 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6412 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6413 DispatchBB->addSuccessor(TrapBB);
6415 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6416 DispatchBB->addSuccessor(DispContBB);
6419 MF->insert(MF->end(), DispatchBB);
6420 MF->insert(MF->end(), DispContBB);
6421 MF->insert(MF->end(), TrapBB);
6423 // Insert code into the entry block that creates and registers the function
6425 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6427 MachineMemOperand *FIMMOLd =
6428 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6429 MachineMemOperand::MOLoad |
6430 MachineMemOperand::MOVolatile, 4, 4);
6432 MachineInstrBuilder MIB;
6433 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6435 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6436 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6438 // Add a register mask with no preserved registers. This results in all
6439 // registers being marked as clobbered.
6440 MIB.addRegMask(RI.getNoPreservedMask());
6442 unsigned NumLPads = LPadList.size();
6443 if (Subtarget->isThumb2()) {
6444 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6445 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6448 .addMemOperand(FIMMOLd));
6450 if (NumLPads < 256) {
6451 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6453 .addImm(LPadList.size()));
6455 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6456 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6457 .addImm(NumLPads & 0xFFFF));
6459 unsigned VReg2 = VReg1;
6460 if ((NumLPads & 0xFFFF0000) != 0) {
6461 VReg2 = MRI->createVirtualRegister(TRC);
6462 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6464 .addImm(NumLPads >> 16));
6467 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6472 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6477 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6478 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6479 .addJumpTableIndex(MJTI)
6482 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6485 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6486 .addReg(NewVReg3, RegState::Kill)
6488 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6490 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6491 .addReg(NewVReg4, RegState::Kill)
6493 .addJumpTableIndex(MJTI)
6495 } else if (Subtarget->isThumb()) {
6496 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6497 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6500 .addMemOperand(FIMMOLd));
6502 if (NumLPads < 256) {
6503 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6507 MachineConstantPool *ConstantPool = MF->getConstantPool();
6508 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6509 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6511 // MachineConstantPool wants an explicit alignment.
6512 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6514 Align = getDataLayout()->getTypeAllocSize(C->getType());
6515 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6517 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6518 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6519 .addReg(VReg1, RegState::Define)
6520 .addConstantPoolIndex(Idx));
6521 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6526 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6531 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6532 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6533 .addReg(ARM::CPSR, RegState::Define)
6537 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6538 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6539 .addJumpTableIndex(MJTI)
6542 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6543 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6544 .addReg(ARM::CPSR, RegState::Define)
6545 .addReg(NewVReg2, RegState::Kill)
6548 MachineMemOperand *JTMMOLd =
6549 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6550 MachineMemOperand::MOLoad, 4, 4);
6552 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6553 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6554 .addReg(NewVReg4, RegState::Kill)
6556 .addMemOperand(JTMMOLd));
6558 unsigned NewVReg6 = NewVReg5;
6559 if (RelocM == Reloc::PIC_) {
6560 NewVReg6 = MRI->createVirtualRegister(TRC);
6561 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6562 .addReg(ARM::CPSR, RegState::Define)
6563 .addReg(NewVReg5, RegState::Kill)
6567 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6568 .addReg(NewVReg6, RegState::Kill)
6569 .addJumpTableIndex(MJTI)
6572 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6573 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6576 .addMemOperand(FIMMOLd));
6578 if (NumLPads < 256) {
6579 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6582 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6583 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6584 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6585 .addImm(NumLPads & 0xFFFF));
6587 unsigned VReg2 = VReg1;
6588 if ((NumLPads & 0xFFFF0000) != 0) {
6589 VReg2 = MRI->createVirtualRegister(TRC);
6590 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6592 .addImm(NumLPads >> 16));
6595 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6599 MachineConstantPool *ConstantPool = MF->getConstantPool();
6600 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6601 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6603 // MachineConstantPool wants an explicit alignment.
6604 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6606 Align = getDataLayout()->getTypeAllocSize(C->getType());
6607 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6609 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6610 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6611 .addReg(VReg1, RegState::Define)
6612 .addConstantPoolIndex(Idx)
6614 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6616 .addReg(VReg1, RegState::Kill));
6619 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6624 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6626 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6628 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6629 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6630 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6631 .addJumpTableIndex(MJTI)
6634 MachineMemOperand *JTMMOLd =
6635 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6636 MachineMemOperand::MOLoad, 4, 4);
6637 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6639 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6640 .addReg(NewVReg3, RegState::Kill)
6643 .addMemOperand(JTMMOLd));
6645 if (RelocM == Reloc::PIC_) {
6646 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6647 .addReg(NewVReg5, RegState::Kill)
6649 .addJumpTableIndex(MJTI)
6652 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6653 .addReg(NewVReg5, RegState::Kill)
6654 .addJumpTableIndex(MJTI)
6659 // Add the jump table entries as successors to the MBB.
6660 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6661 for (std::vector<MachineBasicBlock*>::iterator
6662 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6663 MachineBasicBlock *CurMBB = *I;
6664 if (SeenMBBs.insert(CurMBB))
6665 DispContBB->addSuccessor(CurMBB);
6668 // N.B. the order the invoke BBs are processed in doesn't matter here.
6669 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6670 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6671 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6672 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6673 MachineBasicBlock *BB = *I;
6675 // Remove the landing pad successor from the invoke block and replace it
6676 // with the new dispatch block.
6677 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6679 while (!Successors.empty()) {
6680 MachineBasicBlock *SMBB = Successors.pop_back_val();
6681 if (SMBB->isLandingPad()) {
6682 BB->removeSuccessor(SMBB);
6683 MBBLPads.push_back(SMBB);
6687 BB->addSuccessor(DispatchBB);
6689 // Find the invoke call and mark all of the callee-saved registers as
6690 // 'implicit defined' so that they're spilled. This prevents code from
6691 // moving instructions to before the EH block, where they will never be
6693 for (MachineBasicBlock::reverse_iterator
6694 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6695 if (!II->isCall()) continue;
6697 DenseMap<unsigned, bool> DefRegs;
6698 for (MachineInstr::mop_iterator
6699 OI = II->operands_begin(), OE = II->operands_end();
6701 if (!OI->isReg()) continue;
6702 DefRegs[OI->getReg()] = true;
6705 MachineInstrBuilder MIB(*MF, &*II);
6707 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6708 unsigned Reg = SavedRegs[i];
6709 if (Subtarget->isThumb2() &&
6710 !ARM::tGPRRegClass.contains(Reg) &&
6711 !ARM::hGPRRegClass.contains(Reg))
6713 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6715 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6718 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6725 // Mark all former landing pads as non-landing pads. The dispatch is the only
6727 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6728 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6729 (*I)->setIsLandingPad(false);
6731 // The instruction is gone now.
6732 MI->eraseFromParent();
6738 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6739 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6740 E = MBB->succ_end(); I != E; ++I)
6743 llvm_unreachable("Expecting a BB with two successors!");
6746 MachineBasicBlock *ARMTargetLowering::
6747 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6748 // This pseudo instruction has 3 operands: dst, src, size
6749 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6750 // Otherwise, we will generate unrolled scalar copies.
6751 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6752 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6753 MachineFunction::iterator It = BB;
6756 unsigned dest = MI->getOperand(0).getReg();
6757 unsigned src = MI->getOperand(1).getReg();
6758 unsigned SizeVal = MI->getOperand(2).getImm();
6759 unsigned Align = MI->getOperand(3).getImm();
6760 DebugLoc dl = MI->getDebugLoc();
6762 bool isThumb2 = Subtarget->isThumb2();
6763 MachineFunction *MF = BB->getParent();
6764 MachineRegisterInfo &MRI = MF->getRegInfo();
6765 unsigned ldrOpc, strOpc, UnitSize = 0;
6767 const TargetRegisterClass *TRC = isThumb2 ?
6768 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6769 (const TargetRegisterClass*)&ARM::GPRRegClass;
6770 const TargetRegisterClass *TRC_Vec = 0;
6773 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6774 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6776 } else if (Align & 2) {
6777 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6778 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6781 // Check whether we can use NEON instructions.
6782 if (!MF->getFunction()->getAttributes().
6783 hasAttribute(AttributeSet::FunctionIndex,
6784 Attribute::NoImplicitFloat) &&
6785 Subtarget->hasNEON()) {
6786 if ((Align % 16 == 0) && SizeVal >= 16) {
6787 ldrOpc = ARM::VLD1q32wb_fixed;
6788 strOpc = ARM::VST1q32wb_fixed;
6790 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6792 else if ((Align % 8 == 0) && SizeVal >= 8) {
6793 ldrOpc = ARM::VLD1d32wb_fixed;
6794 strOpc = ARM::VST1d32wb_fixed;
6796 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6799 // Can't use NEON instructions.
6800 if (UnitSize == 0) {
6801 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6802 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6807 unsigned BytesLeft = SizeVal % UnitSize;
6808 unsigned LoopSize = SizeVal - BytesLeft;
6810 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6811 // Use LDR and STR to copy.
6812 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6813 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6814 unsigned srcIn = src;
6815 unsigned destIn = dest;
6816 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6817 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6818 unsigned srcOut = MRI.createVirtualRegister(TRC);
6819 unsigned destOut = MRI.createVirtualRegister(TRC);
6820 if (UnitSize >= 8) {
6821 AddDefaultPred(BuildMI(*BB, MI, dl,
6822 TII->get(ldrOpc), scratch)
6823 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6825 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6826 .addReg(destIn).addImm(0).addReg(scratch));
6827 } else if (isThumb2) {
6828 AddDefaultPred(BuildMI(*BB, MI, dl,
6829 TII->get(ldrOpc), scratch)
6830 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6832 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6833 .addReg(scratch).addReg(destIn)
6836 AddDefaultPred(BuildMI(*BB, MI, dl,
6837 TII->get(ldrOpc), scratch)
6838 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6841 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6842 .addReg(scratch).addReg(destIn)
6843 .addReg(0).addImm(UnitSize));
6849 // Handle the leftover bytes with LDRB and STRB.
6850 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6851 // [destOut] = STRB_POST(scratch, destIn, 1)
6852 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6853 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6854 for (unsigned i = 0; i < BytesLeft; i++) {
6855 unsigned scratch = MRI.createVirtualRegister(TRC);
6856 unsigned srcOut = MRI.createVirtualRegister(TRC);
6857 unsigned destOut = MRI.createVirtualRegister(TRC);
6859 AddDefaultPred(BuildMI(*BB, MI, dl,
6860 TII->get(ldrOpc),scratch)
6861 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6863 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6864 .addReg(scratch).addReg(destIn)
6865 .addReg(0).addImm(1));
6867 AddDefaultPred(BuildMI(*BB, MI, dl,
6868 TII->get(ldrOpc),scratch)
6869 .addReg(srcOut, RegState::Define).addReg(srcIn)
6870 .addReg(0).addImm(1));
6872 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6873 .addReg(scratch).addReg(destIn)
6874 .addReg(0).addImm(1));
6879 MI->eraseFromParent(); // The instruction is gone now.
6883 // Expand the pseudo op to a loop.
6886 // movw varEnd, # --> with thumb2
6888 // ldrcp varEnd, idx --> without thumb2
6889 // fallthrough --> loopMBB
6891 // PHI varPhi, varEnd, varLoop
6892 // PHI srcPhi, src, srcLoop
6893 // PHI destPhi, dst, destLoop
6894 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6895 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6896 // subs varLoop, varPhi, #UnitSize
6898 // fallthrough --> exitMBB
6900 // epilogue to handle left-over bytes
6901 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6902 // [destOut] = STRB_POST(scratch, destLoop, 1)
6903 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6904 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6905 MF->insert(It, loopMBB);
6906 MF->insert(It, exitMBB);
6908 // Transfer the remainder of BB and its successor edges to exitMBB.
6909 exitMBB->splice(exitMBB->begin(), BB,
6910 llvm::next(MachineBasicBlock::iterator(MI)),
6912 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6914 // Load an immediate to varEnd.
6915 unsigned varEnd = MRI.createVirtualRegister(TRC);
6917 unsigned VReg1 = varEnd;
6918 if ((LoopSize & 0xFFFF0000) != 0)
6919 VReg1 = MRI.createVirtualRegister(TRC);
6920 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6921 .addImm(LoopSize & 0xFFFF));
6923 if ((LoopSize & 0xFFFF0000) != 0)
6924 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6926 .addImm(LoopSize >> 16));
6928 MachineConstantPool *ConstantPool = MF->getConstantPool();
6929 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6930 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6932 // MachineConstantPool wants an explicit alignment.
6933 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6935 Align = getDataLayout()->getTypeAllocSize(C->getType());
6936 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6938 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6939 .addReg(varEnd, RegState::Define)
6940 .addConstantPoolIndex(Idx)
6943 BB->addSuccessor(loopMBB);
6945 // Generate the loop body:
6946 // varPhi = PHI(varLoop, varEnd)
6947 // srcPhi = PHI(srcLoop, src)
6948 // destPhi = PHI(destLoop, dst)
6949 MachineBasicBlock *entryBB = BB;
6951 unsigned varLoop = MRI.createVirtualRegister(TRC);
6952 unsigned varPhi = MRI.createVirtualRegister(TRC);
6953 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6954 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6955 unsigned destLoop = MRI.createVirtualRegister(TRC);
6956 unsigned destPhi = MRI.createVirtualRegister(TRC);
6958 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6959 .addReg(varLoop).addMBB(loopMBB)
6960 .addReg(varEnd).addMBB(entryBB);
6961 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6962 .addReg(srcLoop).addMBB(loopMBB)
6963 .addReg(src).addMBB(entryBB);
6964 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6965 .addReg(destLoop).addMBB(loopMBB)
6966 .addReg(dest).addMBB(entryBB);
6968 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6969 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6970 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6971 if (UnitSize >= 8) {
6972 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6973 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6975 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6976 .addReg(destPhi).addImm(0).addReg(scratch));
6977 } else if (isThumb2) {
6978 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6979 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6981 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6982 .addReg(scratch).addReg(destPhi)
6985 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6986 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6989 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6990 .addReg(scratch).addReg(destPhi)
6991 .addReg(0).addImm(UnitSize));
6994 // Decrement loop variable by UnitSize.
6995 MachineInstrBuilder MIB = BuildMI(BB, dl,
6996 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6997 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6998 MIB->getOperand(5).setReg(ARM::CPSR);
6999 MIB->getOperand(5).setIsDef(true);
7001 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7002 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7004 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7005 BB->addSuccessor(loopMBB);
7006 BB->addSuccessor(exitMBB);
7008 // Add epilogue to handle BytesLeft.
7010 MachineInstr *StartOfExit = exitMBB->begin();
7011 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7012 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7014 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7015 // [destOut] = STRB_POST(scratch, destLoop, 1)
7016 unsigned srcIn = srcLoop;
7017 unsigned destIn = destLoop;
7018 for (unsigned i = 0; i < BytesLeft; i++) {
7019 unsigned scratch = MRI.createVirtualRegister(TRC);
7020 unsigned srcOut = MRI.createVirtualRegister(TRC);
7021 unsigned destOut = MRI.createVirtualRegister(TRC);
7023 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7024 TII->get(ldrOpc),scratch)
7025 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7027 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7028 .addReg(scratch).addReg(destIn)
7031 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7032 TII->get(ldrOpc),scratch)
7033 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
7035 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7036 .addReg(scratch).addReg(destIn)
7037 .addReg(0).addImm(1));
7043 MI->eraseFromParent(); // The instruction is gone now.
7048 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7049 MachineBasicBlock *BB) const {
7050 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7051 DebugLoc dl = MI->getDebugLoc();
7052 bool isThumb2 = Subtarget->isThumb2();
7053 switch (MI->getOpcode()) {
7056 llvm_unreachable("Unexpected instr type to insert");
7058 // The Thumb2 pre-indexed stores have the same MI operands, they just
7059 // define them differently in the .td files from the isel patterns, so
7060 // they need pseudos.
7061 case ARM::t2STR_preidx:
7062 MI->setDesc(TII->get(ARM::t2STR_PRE));
7064 case ARM::t2STRB_preidx:
7065 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7067 case ARM::t2STRH_preidx:
7068 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7071 case ARM::STRi_preidx:
7072 case ARM::STRBi_preidx: {
7073 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7074 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7075 // Decode the offset.
7076 unsigned Offset = MI->getOperand(4).getImm();
7077 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7078 Offset = ARM_AM::getAM2Offset(Offset);
7082 MachineMemOperand *MMO = *MI->memoperands_begin();
7083 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7084 .addOperand(MI->getOperand(0)) // Rn_wb
7085 .addOperand(MI->getOperand(1)) // Rt
7086 .addOperand(MI->getOperand(2)) // Rn
7087 .addImm(Offset) // offset (skip GPR==zero_reg)
7088 .addOperand(MI->getOperand(5)) // pred
7089 .addOperand(MI->getOperand(6))
7090 .addMemOperand(MMO);
7091 MI->eraseFromParent();
7094 case ARM::STRr_preidx:
7095 case ARM::STRBr_preidx:
7096 case ARM::STRH_preidx: {
7098 switch (MI->getOpcode()) {
7099 default: llvm_unreachable("unexpected opcode!");
7100 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7101 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7102 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7104 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7105 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7106 MIB.addOperand(MI->getOperand(i));
7107 MI->eraseFromParent();
7110 case ARM::ATOMIC_LOAD_ADD_I8:
7111 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7112 case ARM::ATOMIC_LOAD_ADD_I16:
7113 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7114 case ARM::ATOMIC_LOAD_ADD_I32:
7115 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7117 case ARM::ATOMIC_LOAD_AND_I8:
7118 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7119 case ARM::ATOMIC_LOAD_AND_I16:
7120 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7121 case ARM::ATOMIC_LOAD_AND_I32:
7122 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7124 case ARM::ATOMIC_LOAD_OR_I8:
7125 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7126 case ARM::ATOMIC_LOAD_OR_I16:
7127 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7128 case ARM::ATOMIC_LOAD_OR_I32:
7129 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7131 case ARM::ATOMIC_LOAD_XOR_I8:
7132 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7133 case ARM::ATOMIC_LOAD_XOR_I16:
7134 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7135 case ARM::ATOMIC_LOAD_XOR_I32:
7136 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7138 case ARM::ATOMIC_LOAD_NAND_I8:
7139 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7140 case ARM::ATOMIC_LOAD_NAND_I16:
7141 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7142 case ARM::ATOMIC_LOAD_NAND_I32:
7143 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7145 case ARM::ATOMIC_LOAD_SUB_I8:
7146 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7147 case ARM::ATOMIC_LOAD_SUB_I16:
7148 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7149 case ARM::ATOMIC_LOAD_SUB_I32:
7150 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7152 case ARM::ATOMIC_LOAD_MIN_I8:
7153 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7154 case ARM::ATOMIC_LOAD_MIN_I16:
7155 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7156 case ARM::ATOMIC_LOAD_MIN_I32:
7157 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7159 case ARM::ATOMIC_LOAD_MAX_I8:
7160 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7161 case ARM::ATOMIC_LOAD_MAX_I16:
7162 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7163 case ARM::ATOMIC_LOAD_MAX_I32:
7164 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7166 case ARM::ATOMIC_LOAD_UMIN_I8:
7167 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7168 case ARM::ATOMIC_LOAD_UMIN_I16:
7169 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7170 case ARM::ATOMIC_LOAD_UMIN_I32:
7171 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7173 case ARM::ATOMIC_LOAD_UMAX_I8:
7174 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7175 case ARM::ATOMIC_LOAD_UMAX_I16:
7176 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7177 case ARM::ATOMIC_LOAD_UMAX_I32:
7178 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7180 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7181 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7182 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7184 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7185 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7186 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7189 case ARM::ATOMADD6432:
7190 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7191 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7192 /*NeedsCarry*/ true);
7193 case ARM::ATOMSUB6432:
7194 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7195 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7196 /*NeedsCarry*/ true);
7197 case ARM::ATOMOR6432:
7198 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7199 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7200 case ARM::ATOMXOR6432:
7201 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7202 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7203 case ARM::ATOMAND6432:
7204 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7205 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7206 case ARM::ATOMSWAP6432:
7207 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7208 case ARM::ATOMCMPXCHG6432:
7209 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7210 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7211 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7212 case ARM::ATOMMIN6432:
7213 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7214 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7215 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7216 /*IsMinMax*/ true, ARMCC::LT);
7217 case ARM::ATOMMAX6432:
7218 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7219 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7220 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7221 /*IsMinMax*/ true, ARMCC::GE);
7222 case ARM::ATOMUMIN6432:
7223 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7224 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7225 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7226 /*IsMinMax*/ true, ARMCC::LO);
7227 case ARM::ATOMUMAX6432:
7228 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7229 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7230 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7231 /*IsMinMax*/ true, ARMCC::HS);
7233 case ARM::tMOVCCr_pseudo: {
7234 // To "insert" a SELECT_CC instruction, we actually have to insert the
7235 // diamond control-flow pattern. The incoming instruction knows the
7236 // destination vreg to set, the condition code register to branch on, the
7237 // true/false values to select between, and a branch opcode to use.
7238 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7239 MachineFunction::iterator It = BB;
7245 // cmpTY ccX, r1, r2
7247 // fallthrough --> copy0MBB
7248 MachineBasicBlock *thisMBB = BB;
7249 MachineFunction *F = BB->getParent();
7250 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7251 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7252 F->insert(It, copy0MBB);
7253 F->insert(It, sinkMBB);
7255 // Transfer the remainder of BB and its successor edges to sinkMBB.
7256 sinkMBB->splice(sinkMBB->begin(), BB,
7257 llvm::next(MachineBasicBlock::iterator(MI)),
7259 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7261 BB->addSuccessor(copy0MBB);
7262 BB->addSuccessor(sinkMBB);
7264 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7265 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7268 // %FalseValue = ...
7269 // # fallthrough to sinkMBB
7272 // Update machine-CFG edges
7273 BB->addSuccessor(sinkMBB);
7276 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7279 BuildMI(*BB, BB->begin(), dl,
7280 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7281 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7282 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7284 MI->eraseFromParent(); // The pseudo instruction is gone now.
7289 case ARM::BCCZi64: {
7290 // If there is an unconditional branch to the other successor, remove it.
7291 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7293 // Compare both parts that make up the double comparison separately for
7295 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7297 unsigned LHS1 = MI->getOperand(1).getReg();
7298 unsigned LHS2 = MI->getOperand(2).getReg();
7300 AddDefaultPred(BuildMI(BB, dl,
7301 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7302 .addReg(LHS1).addImm(0));
7303 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7304 .addReg(LHS2).addImm(0)
7305 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7307 unsigned RHS1 = MI->getOperand(3).getReg();
7308 unsigned RHS2 = MI->getOperand(4).getReg();
7309 AddDefaultPred(BuildMI(BB, dl,
7310 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7311 .addReg(LHS1).addReg(RHS1));
7312 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7313 .addReg(LHS2).addReg(RHS2)
7314 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7317 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7318 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7319 if (MI->getOperand(0).getImm() == ARMCC::NE)
7320 std::swap(destMBB, exitMBB);
7322 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7323 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7325 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7327 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7329 MI->eraseFromParent(); // The pseudo instruction is gone now.
7333 case ARM::Int_eh_sjlj_setjmp:
7334 case ARM::Int_eh_sjlj_setjmp_nofp:
7335 case ARM::tInt_eh_sjlj_setjmp:
7336 case ARM::t2Int_eh_sjlj_setjmp:
7337 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7338 EmitSjLjDispatchBlock(MI, BB);
7343 // To insert an ABS instruction, we have to insert the
7344 // diamond control-flow pattern. The incoming instruction knows the
7345 // source vreg to test against 0, the destination vreg to set,
7346 // the condition code register to branch on, the
7347 // true/false values to select between, and a branch opcode to use.
7352 // BCC (branch to SinkBB if V0 >= 0)
7353 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7354 // SinkBB: V1 = PHI(V2, V3)
7355 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7356 MachineFunction::iterator BBI = BB;
7358 MachineFunction *Fn = BB->getParent();
7359 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7360 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7361 Fn->insert(BBI, RSBBB);
7362 Fn->insert(BBI, SinkBB);
7364 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7365 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7366 bool isThumb2 = Subtarget->isThumb2();
7367 MachineRegisterInfo &MRI = Fn->getRegInfo();
7368 // In Thumb mode S must not be specified if source register is the SP or
7369 // PC and if destination register is the SP, so restrict register class
7370 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7371 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7372 (const TargetRegisterClass*)&ARM::GPRRegClass);
7374 // Transfer the remainder of BB and its successor edges to sinkMBB.
7375 SinkBB->splice(SinkBB->begin(), BB,
7376 llvm::next(MachineBasicBlock::iterator(MI)),
7378 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7380 BB->addSuccessor(RSBBB);
7381 BB->addSuccessor(SinkBB);
7383 // fall through to SinkMBB
7384 RSBBB->addSuccessor(SinkBB);
7386 // insert a cmp at the end of BB
7387 AddDefaultPred(BuildMI(BB, dl,
7388 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7389 .addReg(ABSSrcReg).addImm(0));
7391 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7393 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7394 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7396 // insert rsbri in RSBBB
7397 // Note: BCC and rsbri will be converted into predicated rsbmi
7398 // by if-conversion pass
7399 BuildMI(*RSBBB, RSBBB->begin(), dl,
7400 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7401 .addReg(ABSSrcReg, RegState::Kill)
7402 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7404 // insert PHI in SinkBB,
7405 // reuse ABSDstReg to not change uses of ABS instruction
7406 BuildMI(*SinkBB, SinkBB->begin(), dl,
7407 TII->get(ARM::PHI), ABSDstReg)
7408 .addReg(NewRsbDstReg).addMBB(RSBBB)
7409 .addReg(ABSSrcReg).addMBB(BB);
7411 // remove ABS instruction
7412 MI->eraseFromParent();
7414 // return last added BB
7417 case ARM::COPY_STRUCT_BYVAL_I32:
7419 return EmitStructByval(MI, BB);
7423 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7424 SDNode *Node) const {
7425 if (!MI->hasPostISelHook()) {
7426 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7427 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7431 const MCInstrDesc *MCID = &MI->getDesc();
7432 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7433 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7434 // operand is still set to noreg. If needed, set the optional operand's
7435 // register to CPSR, and remove the redundant implicit def.
7437 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7439 // Rename pseudo opcodes.
7440 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7442 const ARMBaseInstrInfo *TII =
7443 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7444 MCID = &TII->get(NewOpc);
7446 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7447 "converted opcode should be the same except for cc_out");
7451 // Add the optional cc_out operand
7452 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7454 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7456 // Any ARM instruction that sets the 's' bit should specify an optional
7457 // "cc_out" operand in the last operand position.
7458 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7459 assert(!NewOpc && "Optional cc_out operand required");
7462 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7463 // since we already have an optional CPSR def.
7464 bool definesCPSR = false;
7465 bool deadCPSR = false;
7466 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7468 const MachineOperand &MO = MI->getOperand(i);
7469 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7473 MI->RemoveOperand(i);
7478 assert(!NewOpc && "Optional cc_out operand required");
7481 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7483 assert(!MI->getOperand(ccOutIdx).getReg() &&
7484 "expect uninitialized optional cc_out operand");
7488 // If this instruction was defined with an optional CPSR def and its dag node
7489 // had a live implicit CPSR def, then activate the optional CPSR def.
7490 MachineOperand &MO = MI->getOperand(ccOutIdx);
7491 MO.setReg(ARM::CPSR);
7495 //===----------------------------------------------------------------------===//
7496 // ARM Optimization Hooks
7497 //===----------------------------------------------------------------------===//
7499 // Helper function that checks if N is a null or all ones constant.
7500 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7501 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7504 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7507 // Return true if N is conditionally 0 or all ones.
7508 // Detects these expressions where cc is an i1 value:
7510 // (select cc 0, y) [AllOnes=0]
7511 // (select cc y, 0) [AllOnes=0]
7512 // (zext cc) [AllOnes=0]
7513 // (sext cc) [AllOnes=0/1]
7514 // (select cc -1, y) [AllOnes=1]
7515 // (select cc y, -1) [AllOnes=1]
7517 // Invert is set when N is the null/all ones constant when CC is false.
7518 // OtherOp is set to the alternative value of N.
7519 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7520 SDValue &CC, bool &Invert,
7522 SelectionDAG &DAG) {
7523 switch (N->getOpcode()) {
7524 default: return false;
7526 CC = N->getOperand(0);
7527 SDValue N1 = N->getOperand(1);
7528 SDValue N2 = N->getOperand(2);
7529 if (isZeroOrAllOnes(N1, AllOnes)) {
7534 if (isZeroOrAllOnes(N2, AllOnes)) {
7541 case ISD::ZERO_EXTEND:
7542 // (zext cc) can never be the all ones value.
7546 case ISD::SIGN_EXTEND: {
7547 EVT VT = N->getValueType(0);
7548 CC = N->getOperand(0);
7549 if (CC.getValueType() != MVT::i1)
7553 // When looking for an AllOnes constant, N is an sext, and the 'other'
7555 OtherOp = DAG.getConstant(0, VT);
7556 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7557 // When looking for a 0 constant, N can be zext or sext.
7558 OtherOp = DAG.getConstant(1, VT);
7560 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7566 // Combine a constant select operand into its use:
7568 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7569 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7570 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7571 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7572 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7574 // The transform is rejected if the select doesn't have a constant operand that
7575 // is null, or all ones when AllOnes is set.
7577 // Also recognize sext/zext from i1:
7579 // (add (zext cc), x) -> (select cc (add x, 1), x)
7580 // (add (sext cc), x) -> (select cc (add x, -1), x)
7582 // These transformations eventually create predicated instructions.
7584 // @param N The node to transform.
7585 // @param Slct The N operand that is a select.
7586 // @param OtherOp The other N operand (x above).
7587 // @param DCI Context.
7588 // @param AllOnes Require the select constant to be all ones instead of null.
7589 // @returns The new node, or SDValue() on failure.
7591 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7592 TargetLowering::DAGCombinerInfo &DCI,
7593 bool AllOnes = false) {
7594 SelectionDAG &DAG = DCI.DAG;
7595 EVT VT = N->getValueType(0);
7596 SDValue NonConstantVal;
7599 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7600 NonConstantVal, DAG))
7603 // Slct is now know to be the desired identity constant when CC is true.
7604 SDValue TrueVal = OtherOp;
7605 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7606 OtherOp, NonConstantVal);
7607 // Unless SwapSelectOps says CC should be false.
7609 std::swap(TrueVal, FalseVal);
7611 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7612 CCOp, TrueVal, FalseVal);
7615 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7617 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7618 TargetLowering::DAGCombinerInfo &DCI) {
7619 SDValue N0 = N->getOperand(0);
7620 SDValue N1 = N->getOperand(1);
7621 if (N0.getNode()->hasOneUse()) {
7622 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7623 if (Result.getNode())
7626 if (N1.getNode()->hasOneUse()) {
7627 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7628 if (Result.getNode())
7634 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7635 // (only after legalization).
7636 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7637 TargetLowering::DAGCombinerInfo &DCI,
7638 const ARMSubtarget *Subtarget) {
7640 // Only perform optimization if after legalize, and if NEON is available. We
7641 // also expected both operands to be BUILD_VECTORs.
7642 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7643 || N0.getOpcode() != ISD::BUILD_VECTOR
7644 || N1.getOpcode() != ISD::BUILD_VECTOR)
7647 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7648 EVT VT = N->getValueType(0);
7649 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7652 // Check that the vector operands are of the right form.
7653 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7654 // operands, where N is the size of the formed vector.
7655 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7656 // index such that we have a pair wise add pattern.
7658 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7659 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7661 SDValue Vec = N0->getOperand(0)->getOperand(0);
7662 SDNode *V = Vec.getNode();
7663 unsigned nextIndex = 0;
7665 // For each operands to the ADD which are BUILD_VECTORs,
7666 // check to see if each of their operands are an EXTRACT_VECTOR with
7667 // the same vector and appropriate index.
7668 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7669 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7670 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7672 SDValue ExtVec0 = N0->getOperand(i);
7673 SDValue ExtVec1 = N1->getOperand(i);
7675 // First operand is the vector, verify its the same.
7676 if (V != ExtVec0->getOperand(0).getNode() ||
7677 V != ExtVec1->getOperand(0).getNode())
7680 // Second is the constant, verify its correct.
7681 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7682 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7684 // For the constant, we want to see all the even or all the odd.
7685 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7686 || C1->getZExtValue() != nextIndex+1)
7695 // Create VPADDL node.
7696 SelectionDAG &DAG = DCI.DAG;
7697 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7699 // Build operand list.
7700 SmallVector<SDValue, 8> Ops;
7701 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7702 TLI.getPointerTy()));
7704 // Input is the vector.
7707 // Get widened type and narrowed type.
7709 unsigned numElem = VT.getVectorNumElements();
7710 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7711 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7712 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7713 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7715 llvm_unreachable("Invalid vector element type for padd optimization.");
7718 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7719 widenType, &Ops[0], Ops.size());
7720 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7723 static SDValue findMUL_LOHI(SDValue V) {
7724 if (V->getOpcode() == ISD::UMUL_LOHI ||
7725 V->getOpcode() == ISD::SMUL_LOHI)
7730 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7731 TargetLowering::DAGCombinerInfo &DCI,
7732 const ARMSubtarget *Subtarget) {
7734 if (Subtarget->isThumb1Only()) return SDValue();
7736 // Only perform the checks after legalize when the pattern is available.
7737 if (DCI.isBeforeLegalize()) return SDValue();
7739 // Look for multiply add opportunities.
7740 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7741 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7742 // a glue link from the first add to the second add.
7743 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7744 // a S/UMLAL instruction.
7747 // \ / \ [no multiline comment]
7753 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7754 SDValue AddcOp0 = AddcNode->getOperand(0);
7755 SDValue AddcOp1 = AddcNode->getOperand(1);
7757 // Check if the two operands are from the same mul_lohi node.
7758 if (AddcOp0.getNode() == AddcOp1.getNode())
7761 assert(AddcNode->getNumValues() == 2 &&
7762 AddcNode->getValueType(0) == MVT::i32 &&
7763 AddcNode->getValueType(1) == MVT::Glue &&
7764 "Expect ADDC with two result values: i32, glue");
7766 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7767 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7768 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7769 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7770 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7773 // Look for the glued ADDE.
7774 SDNode* AddeNode = AddcNode->getGluedUser();
7775 if (AddeNode == NULL)
7778 // Make sure it is really an ADDE.
7779 if (AddeNode->getOpcode() != ISD::ADDE)
7782 assert(AddeNode->getNumOperands() == 3 &&
7783 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7784 "ADDE node has the wrong inputs");
7786 // Check for the triangle shape.
7787 SDValue AddeOp0 = AddeNode->getOperand(0);
7788 SDValue AddeOp1 = AddeNode->getOperand(1);
7790 // Make sure that the ADDE operands are not coming from the same node.
7791 if (AddeOp0.getNode() == AddeOp1.getNode())
7794 // Find the MUL_LOHI node walking up ADDE's operands.
7795 bool IsLeftOperandMUL = false;
7796 SDValue MULOp = findMUL_LOHI(AddeOp0);
7797 if (MULOp == SDValue())
7798 MULOp = findMUL_LOHI(AddeOp1);
7800 IsLeftOperandMUL = true;
7801 if (MULOp == SDValue())
7804 // Figure out the right opcode.
7805 unsigned Opc = MULOp->getOpcode();
7806 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7808 // Figure out the high and low input values to the MLAL node.
7809 SDValue* HiMul = &MULOp;
7810 SDValue* HiAdd = NULL;
7811 SDValue* LoMul = NULL;
7812 SDValue* LowAdd = NULL;
7814 if (IsLeftOperandMUL)
7820 if (AddcOp0->getOpcode() == Opc) {
7824 if (AddcOp1->getOpcode() == Opc) {
7832 if (LoMul->getNode() != HiMul->getNode())
7835 // Create the merged node.
7836 SelectionDAG &DAG = DCI.DAG;
7838 // Build operand list.
7839 SmallVector<SDValue, 8> Ops;
7840 Ops.push_back(LoMul->getOperand(0));
7841 Ops.push_back(LoMul->getOperand(1));
7842 Ops.push_back(*LowAdd);
7843 Ops.push_back(*HiAdd);
7845 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7846 DAG.getVTList(MVT::i32, MVT::i32),
7847 &Ops[0], Ops.size());
7849 // Replace the ADDs' nodes uses by the MLA node's values.
7850 SDValue HiMLALResult(MLALNode.getNode(), 1);
7851 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7853 SDValue LoMLALResult(MLALNode.getNode(), 0);
7854 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7856 // Return original node to notify the driver to stop replacing.
7857 SDValue resNode(AddcNode, 0);
7861 /// PerformADDCCombine - Target-specific dag combine transform from
7862 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7863 static SDValue PerformADDCCombine(SDNode *N,
7864 TargetLowering::DAGCombinerInfo &DCI,
7865 const ARMSubtarget *Subtarget) {
7867 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7871 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7872 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7873 /// called with the default operands, and if that fails, with commuted
7875 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7876 TargetLowering::DAGCombinerInfo &DCI,
7877 const ARMSubtarget *Subtarget){
7879 // Attempt to create vpaddl for this add.
7880 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7881 if (Result.getNode())
7884 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7885 if (N0.getNode()->hasOneUse()) {
7886 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7887 if (Result.getNode()) return Result;
7892 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7894 static SDValue PerformADDCombine(SDNode *N,
7895 TargetLowering::DAGCombinerInfo &DCI,
7896 const ARMSubtarget *Subtarget) {
7897 SDValue N0 = N->getOperand(0);
7898 SDValue N1 = N->getOperand(1);
7900 // First try with the default operand order.
7901 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7902 if (Result.getNode())
7905 // If that didn't work, try again with the operands commuted.
7906 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7909 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7911 static SDValue PerformSUBCombine(SDNode *N,
7912 TargetLowering::DAGCombinerInfo &DCI) {
7913 SDValue N0 = N->getOperand(0);
7914 SDValue N1 = N->getOperand(1);
7916 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7917 if (N1.getNode()->hasOneUse()) {
7918 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7919 if (Result.getNode()) return Result;
7925 /// PerformVMULCombine
7926 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7927 /// special multiplier accumulator forwarding.
7933 static SDValue PerformVMULCombine(SDNode *N,
7934 TargetLowering::DAGCombinerInfo &DCI,
7935 const ARMSubtarget *Subtarget) {
7936 if (!Subtarget->hasVMLxForwarding())
7939 SelectionDAG &DAG = DCI.DAG;
7940 SDValue N0 = N->getOperand(0);
7941 SDValue N1 = N->getOperand(1);
7942 unsigned Opcode = N0.getOpcode();
7943 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7944 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7945 Opcode = N1.getOpcode();
7946 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7947 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7952 EVT VT = N->getValueType(0);
7953 DebugLoc DL = N->getDebugLoc();
7954 SDValue N00 = N0->getOperand(0);
7955 SDValue N01 = N0->getOperand(1);
7956 return DAG.getNode(Opcode, DL, VT,
7957 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7958 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7961 static SDValue PerformMULCombine(SDNode *N,
7962 TargetLowering::DAGCombinerInfo &DCI,
7963 const ARMSubtarget *Subtarget) {
7964 SelectionDAG &DAG = DCI.DAG;
7966 if (Subtarget->isThumb1Only())
7969 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7972 EVT VT = N->getValueType(0);
7973 if (VT.is64BitVector() || VT.is128BitVector())
7974 return PerformVMULCombine(N, DCI, Subtarget);
7978 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7982 int64_t MulAmt = C->getSExtValue();
7983 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7985 ShiftAmt = ShiftAmt & (32 - 1);
7986 SDValue V = N->getOperand(0);
7987 DebugLoc DL = N->getDebugLoc();
7990 MulAmt >>= ShiftAmt;
7993 if (isPowerOf2_32(MulAmt - 1)) {
7994 // (mul x, 2^N + 1) => (add (shl x, N), x)
7995 Res = DAG.getNode(ISD::ADD, DL, VT,
7997 DAG.getNode(ISD::SHL, DL, VT,
7999 DAG.getConstant(Log2_32(MulAmt - 1),
8001 } else if (isPowerOf2_32(MulAmt + 1)) {
8002 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8003 Res = DAG.getNode(ISD::SUB, DL, VT,
8004 DAG.getNode(ISD::SHL, DL, VT,
8006 DAG.getConstant(Log2_32(MulAmt + 1),
8012 uint64_t MulAmtAbs = -MulAmt;
8013 if (isPowerOf2_32(MulAmtAbs + 1)) {
8014 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8015 Res = DAG.getNode(ISD::SUB, DL, VT,
8017 DAG.getNode(ISD::SHL, DL, VT,
8019 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8021 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8022 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8023 Res = DAG.getNode(ISD::ADD, DL, VT,
8025 DAG.getNode(ISD::SHL, DL, VT,
8027 DAG.getConstant(Log2_32(MulAmtAbs-1),
8029 Res = DAG.getNode(ISD::SUB, DL, VT,
8030 DAG.getConstant(0, MVT::i32),Res);
8037 Res = DAG.getNode(ISD::SHL, DL, VT,
8038 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8040 // Do not add new nodes to DAG combiner worklist.
8041 DCI.CombineTo(N, Res, false);
8045 static SDValue PerformANDCombine(SDNode *N,
8046 TargetLowering::DAGCombinerInfo &DCI,
8047 const ARMSubtarget *Subtarget) {
8049 // Attempt to use immediate-form VBIC
8050 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8051 DebugLoc dl = N->getDebugLoc();
8052 EVT VT = N->getValueType(0);
8053 SelectionDAG &DAG = DCI.DAG;
8055 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8058 APInt SplatBits, SplatUndef;
8059 unsigned SplatBitSize;
8062 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8063 if (SplatBitSize <= 64) {
8065 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8066 SplatUndef.getZExtValue(), SplatBitSize,
8067 DAG, VbicVT, VT.is128BitVector(),
8069 if (Val.getNode()) {
8071 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8072 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8073 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8078 if (!Subtarget->isThumb1Only()) {
8079 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8080 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8081 if (Result.getNode())
8088 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8089 static SDValue PerformORCombine(SDNode *N,
8090 TargetLowering::DAGCombinerInfo &DCI,
8091 const ARMSubtarget *Subtarget) {
8092 // Attempt to use immediate-form VORR
8093 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8094 DebugLoc dl = N->getDebugLoc();
8095 EVT VT = N->getValueType(0);
8096 SelectionDAG &DAG = DCI.DAG;
8098 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8101 APInt SplatBits, SplatUndef;
8102 unsigned SplatBitSize;
8104 if (BVN && Subtarget->hasNEON() &&
8105 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8106 if (SplatBitSize <= 64) {
8108 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8109 SplatUndef.getZExtValue(), SplatBitSize,
8110 DAG, VorrVT, VT.is128BitVector(),
8112 if (Val.getNode()) {
8114 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8115 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8116 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8121 if (!Subtarget->isThumb1Only()) {
8122 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8123 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8124 if (Result.getNode())
8128 // The code below optimizes (or (and X, Y), Z).
8129 // The AND operand needs to have a single user to make these optimizations
8131 SDValue N0 = N->getOperand(0);
8132 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8134 SDValue N1 = N->getOperand(1);
8136 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8137 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8138 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8140 unsigned SplatBitSize;
8143 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8145 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8146 HasAnyUndefs) && !HasAnyUndefs) {
8147 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8149 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8150 HasAnyUndefs) && !HasAnyUndefs &&
8151 SplatBits0 == ~SplatBits1) {
8152 // Canonicalize the vector type to make instruction selection simpler.
8153 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8154 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8155 N0->getOperand(1), N0->getOperand(0),
8157 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8162 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8165 // BFI is only available on V6T2+
8166 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8169 DebugLoc DL = N->getDebugLoc();
8170 // 1) or (and A, mask), val => ARMbfi A, val, mask
8171 // iff (val & mask) == val
8173 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8174 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8175 // && mask == ~mask2
8176 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8177 // && ~mask == mask2
8178 // (i.e., copy a bitfield value into another bitfield of the same width)
8183 SDValue N00 = N0.getOperand(0);
8185 // The value and the mask need to be constants so we can verify this is
8186 // actually a bitfield set. If the mask is 0xffff, we can do better
8187 // via a movt instruction, so don't use BFI in that case.
8188 SDValue MaskOp = N0.getOperand(1);
8189 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8192 unsigned Mask = MaskC->getZExtValue();
8196 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8197 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8199 unsigned Val = N1C->getZExtValue();
8200 if ((Val & ~Mask) != Val)
8203 if (ARM::isBitFieldInvertedMask(Mask)) {
8204 Val >>= CountTrailingZeros_32(~Mask);
8206 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8207 DAG.getConstant(Val, MVT::i32),
8208 DAG.getConstant(Mask, MVT::i32));
8210 // Do not add new nodes to DAG combiner worklist.
8211 DCI.CombineTo(N, Res, false);
8214 } else if (N1.getOpcode() == ISD::AND) {
8215 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8216 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8219 unsigned Mask2 = N11C->getZExtValue();
8221 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8223 if (ARM::isBitFieldInvertedMask(Mask) &&
8225 // The pack halfword instruction works better for masks that fit it,
8226 // so use that when it's available.
8227 if (Subtarget->hasT2ExtractPack() &&
8228 (Mask == 0xffff || Mask == 0xffff0000))
8231 unsigned amt = CountTrailingZeros_32(Mask2);
8232 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8233 DAG.getConstant(amt, MVT::i32));
8234 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8235 DAG.getConstant(Mask, MVT::i32));
8236 // Do not add new nodes to DAG combiner worklist.
8237 DCI.CombineTo(N, Res, false);
8239 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8241 // The pack halfword instruction works better for masks that fit it,
8242 // so use that when it's available.
8243 if (Subtarget->hasT2ExtractPack() &&
8244 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8247 unsigned lsb = CountTrailingZeros_32(Mask);
8248 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8249 DAG.getConstant(lsb, MVT::i32));
8250 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8251 DAG.getConstant(Mask2, MVT::i32));
8252 // Do not add new nodes to DAG combiner worklist.
8253 DCI.CombineTo(N, Res, false);
8258 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8259 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8260 ARM::isBitFieldInvertedMask(~Mask)) {
8261 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8262 // where lsb(mask) == #shamt and masked bits of B are known zero.
8263 SDValue ShAmt = N00.getOperand(1);
8264 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8265 unsigned LSB = CountTrailingZeros_32(Mask);
8269 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8270 DAG.getConstant(~Mask, MVT::i32));
8272 // Do not add new nodes to DAG combiner worklist.
8273 DCI.CombineTo(N, Res, false);
8279 static SDValue PerformXORCombine(SDNode *N,
8280 TargetLowering::DAGCombinerInfo &DCI,
8281 const ARMSubtarget *Subtarget) {
8282 EVT VT = N->getValueType(0);
8283 SelectionDAG &DAG = DCI.DAG;
8285 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8288 if (!Subtarget->isThumb1Only()) {
8289 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8290 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8291 if (Result.getNode())
8298 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8299 /// the bits being cleared by the AND are not demanded by the BFI.
8300 static SDValue PerformBFICombine(SDNode *N,
8301 TargetLowering::DAGCombinerInfo &DCI) {
8302 SDValue N1 = N->getOperand(1);
8303 if (N1.getOpcode() == ISD::AND) {
8304 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8307 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8308 unsigned LSB = CountTrailingZeros_32(~InvMask);
8309 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
8310 unsigned Mask = (1 << Width)-1;
8311 unsigned Mask2 = N11C->getZExtValue();
8312 if ((Mask & (~Mask2)) == 0)
8313 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
8314 N->getOperand(0), N1.getOperand(0),
8320 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8321 /// ARMISD::VMOVRRD.
8322 static SDValue PerformVMOVRRDCombine(SDNode *N,
8323 TargetLowering::DAGCombinerInfo &DCI) {
8324 // vmovrrd(vmovdrr x, y) -> x,y
8325 SDValue InDouble = N->getOperand(0);
8326 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8327 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8329 // vmovrrd(load f64) -> (load i32), (load i32)
8330 SDNode *InNode = InDouble.getNode();
8331 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8332 InNode->getValueType(0) == MVT::f64 &&
8333 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8334 !cast<LoadSDNode>(InNode)->isVolatile()) {
8335 // TODO: Should this be done for non-FrameIndex operands?
8336 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8338 SelectionDAG &DAG = DCI.DAG;
8339 DebugLoc DL = LD->getDebugLoc();
8340 SDValue BasePtr = LD->getBasePtr();
8341 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8342 LD->getPointerInfo(), LD->isVolatile(),
8343 LD->isNonTemporal(), LD->isInvariant(),
8344 LD->getAlignment());
8346 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8347 DAG.getConstant(4, MVT::i32));
8348 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8349 LD->getPointerInfo(), LD->isVolatile(),
8350 LD->isNonTemporal(), LD->isInvariant(),
8351 std::min(4U, LD->getAlignment() / 2));
8353 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8354 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8355 DCI.RemoveFromWorklist(LD);
8363 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8364 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8365 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8366 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8367 SDValue Op0 = N->getOperand(0);
8368 SDValue Op1 = N->getOperand(1);
8369 if (Op0.getOpcode() == ISD::BITCAST)
8370 Op0 = Op0.getOperand(0);
8371 if (Op1.getOpcode() == ISD::BITCAST)
8372 Op1 = Op1.getOperand(0);
8373 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8374 Op0.getNode() == Op1.getNode() &&
8375 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8376 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
8377 N->getValueType(0), Op0.getOperand(0));
8381 /// PerformSTORECombine - Target-specific dag combine xforms for
8383 static SDValue PerformSTORECombine(SDNode *N,
8384 TargetLowering::DAGCombinerInfo &DCI) {
8385 StoreSDNode *St = cast<StoreSDNode>(N);
8386 if (St->isVolatile())
8389 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8390 // pack all of the elements in one place. Next, store to memory in fewer
8392 SDValue StVal = St->getValue();
8393 EVT VT = StVal.getValueType();
8394 if (St->isTruncatingStore() && VT.isVector()) {
8395 SelectionDAG &DAG = DCI.DAG;
8396 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8397 EVT StVT = St->getMemoryVT();
8398 unsigned NumElems = VT.getVectorNumElements();
8399 assert(StVT != VT && "Cannot truncate to the same type");
8400 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8401 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8403 // From, To sizes and ElemCount must be pow of two
8404 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8406 // We are going to use the original vector elt for storing.
8407 // Accumulated smaller vector elements must be a multiple of the store size.
8408 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8410 unsigned SizeRatio = FromEltSz / ToEltSz;
8411 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8413 // Create a type on which we perform the shuffle.
8414 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8415 NumElems*SizeRatio);
8416 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8418 DebugLoc DL = St->getDebugLoc();
8419 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8420 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8421 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8423 // Can't shuffle using an illegal type.
8424 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8426 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8427 DAG.getUNDEF(WideVec.getValueType()),
8429 // At this point all of the data is stored at the bottom of the
8430 // register. We now need to save it to mem.
8432 // Find the largest store unit
8433 MVT StoreType = MVT::i8;
8434 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8435 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8436 MVT Tp = (MVT::SimpleValueType)tp;
8437 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8440 // Didn't find a legal store type.
8441 if (!TLI.isTypeLegal(StoreType))
8444 // Bitcast the original vector into a vector of store-size units
8445 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8446 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8447 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8448 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8449 SmallVector<SDValue, 8> Chains;
8450 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8451 TLI.getPointerTy());
8452 SDValue BasePtr = St->getBasePtr();
8454 // Perform one or more big stores into memory.
8455 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8456 for (unsigned I = 0; I < E; I++) {
8457 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8458 StoreType, ShuffWide,
8459 DAG.getIntPtrConstant(I));
8460 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8461 St->getPointerInfo(), St->isVolatile(),
8462 St->isNonTemporal(), St->getAlignment());
8463 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8465 Chains.push_back(Ch);
8467 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8471 if (!ISD::isNormalStore(St))
8474 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8475 // ARM stores of arguments in the same cache line.
8476 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8477 StVal.getNode()->hasOneUse()) {
8478 SelectionDAG &DAG = DCI.DAG;
8479 DebugLoc DL = St->getDebugLoc();
8480 SDValue BasePtr = St->getBasePtr();
8481 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8482 StVal.getNode()->getOperand(0), BasePtr,
8483 St->getPointerInfo(), St->isVolatile(),
8484 St->isNonTemporal(), St->getAlignment());
8486 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8487 DAG.getConstant(4, MVT::i32));
8488 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8489 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8490 St->isNonTemporal(),
8491 std::min(4U, St->getAlignment() / 2));
8494 if (StVal.getValueType() != MVT::i64 ||
8495 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8498 // Bitcast an i64 store extracted from a vector to f64.
8499 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8500 SelectionDAG &DAG = DCI.DAG;
8501 DebugLoc dl = StVal.getDebugLoc();
8502 SDValue IntVec = StVal.getOperand(0);
8503 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8504 IntVec.getValueType().getVectorNumElements());
8505 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8506 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8507 Vec, StVal.getOperand(1));
8508 dl = N->getDebugLoc();
8509 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8510 // Make the DAGCombiner fold the bitcasts.
8511 DCI.AddToWorklist(Vec.getNode());
8512 DCI.AddToWorklist(ExtElt.getNode());
8513 DCI.AddToWorklist(V.getNode());
8514 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8515 St->getPointerInfo(), St->isVolatile(),
8516 St->isNonTemporal(), St->getAlignment(),
8520 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8521 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8522 /// i64 vector to have f64 elements, since the value can then be loaded
8523 /// directly into a VFP register.
8524 static bool hasNormalLoadOperand(SDNode *N) {
8525 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8526 for (unsigned i = 0; i < NumElts; ++i) {
8527 SDNode *Elt = N->getOperand(i).getNode();
8528 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8534 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8535 /// ISD::BUILD_VECTOR.
8536 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8537 TargetLowering::DAGCombinerInfo &DCI){
8538 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8539 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8540 // into a pair of GPRs, which is fine when the value is used as a scalar,
8541 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8542 SelectionDAG &DAG = DCI.DAG;
8543 if (N->getNumOperands() == 2) {
8544 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8549 // Load i64 elements as f64 values so that type legalization does not split
8550 // them up into i32 values.
8551 EVT VT = N->getValueType(0);
8552 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8554 DebugLoc dl = N->getDebugLoc();
8555 SmallVector<SDValue, 8> Ops;
8556 unsigned NumElts = VT.getVectorNumElements();
8557 for (unsigned i = 0; i < NumElts; ++i) {
8558 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8560 // Make the DAGCombiner fold the bitcast.
8561 DCI.AddToWorklist(V.getNode());
8563 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8564 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8565 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8568 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8569 /// ISD::INSERT_VECTOR_ELT.
8570 static SDValue PerformInsertEltCombine(SDNode *N,
8571 TargetLowering::DAGCombinerInfo &DCI) {
8572 // Bitcast an i64 load inserted into a vector to f64.
8573 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8574 EVT VT = N->getValueType(0);
8575 SDNode *Elt = N->getOperand(1).getNode();
8576 if (VT.getVectorElementType() != MVT::i64 ||
8577 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8580 SelectionDAG &DAG = DCI.DAG;
8581 DebugLoc dl = N->getDebugLoc();
8582 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8583 VT.getVectorNumElements());
8584 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8585 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8586 // Make the DAGCombiner fold the bitcasts.
8587 DCI.AddToWorklist(Vec.getNode());
8588 DCI.AddToWorklist(V.getNode());
8589 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8590 Vec, V, N->getOperand(2));
8591 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8594 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8595 /// ISD::VECTOR_SHUFFLE.
8596 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8597 // The LLVM shufflevector instruction does not require the shuffle mask
8598 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8599 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8600 // operands do not match the mask length, they are extended by concatenating
8601 // them with undef vectors. That is probably the right thing for other
8602 // targets, but for NEON it is better to concatenate two double-register
8603 // size vector operands into a single quad-register size vector. Do that
8604 // transformation here:
8605 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8606 // shuffle(concat(v1, v2), undef)
8607 SDValue Op0 = N->getOperand(0);
8608 SDValue Op1 = N->getOperand(1);
8609 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8610 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8611 Op0.getNumOperands() != 2 ||
8612 Op1.getNumOperands() != 2)
8614 SDValue Concat0Op1 = Op0.getOperand(1);
8615 SDValue Concat1Op1 = Op1.getOperand(1);
8616 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8617 Concat1Op1.getOpcode() != ISD::UNDEF)
8619 // Skip the transformation if any of the types are illegal.
8620 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8621 EVT VT = N->getValueType(0);
8622 if (!TLI.isTypeLegal(VT) ||
8623 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8624 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8627 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8628 Op0.getOperand(0), Op1.getOperand(0));
8629 // Translate the shuffle mask.
8630 SmallVector<int, 16> NewMask;
8631 unsigned NumElts = VT.getVectorNumElements();
8632 unsigned HalfElts = NumElts/2;
8633 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8634 for (unsigned n = 0; n < NumElts; ++n) {
8635 int MaskElt = SVN->getMaskElt(n);
8637 if (MaskElt < (int)HalfElts)
8639 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8640 NewElt = HalfElts + MaskElt - NumElts;
8641 NewMask.push_back(NewElt);
8643 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8644 DAG.getUNDEF(VT), NewMask.data());
8647 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8648 /// NEON load/store intrinsics to merge base address updates.
8649 static SDValue CombineBaseUpdate(SDNode *N,
8650 TargetLowering::DAGCombinerInfo &DCI) {
8651 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8654 SelectionDAG &DAG = DCI.DAG;
8655 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8656 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8657 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8658 SDValue Addr = N->getOperand(AddrOpIdx);
8660 // Search for a use of the address operand that is an increment.
8661 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8662 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8664 if (User->getOpcode() != ISD::ADD ||
8665 UI.getUse().getResNo() != Addr.getResNo())
8668 // Check that the add is independent of the load/store. Otherwise, folding
8669 // it would create a cycle.
8670 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8673 // Find the new opcode for the updating load/store.
8675 bool isLaneOp = false;
8676 unsigned NewOpc = 0;
8677 unsigned NumVecs = 0;
8679 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8681 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8682 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8684 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8686 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8688 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8690 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8691 NumVecs = 2; isLaneOp = true; break;
8692 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8693 NumVecs = 3; isLaneOp = true; break;
8694 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8695 NumVecs = 4; isLaneOp = true; break;
8696 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8697 NumVecs = 1; isLoad = false; break;
8698 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8699 NumVecs = 2; isLoad = false; break;
8700 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8701 NumVecs = 3; isLoad = false; break;
8702 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8703 NumVecs = 4; isLoad = false; break;
8704 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8705 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8706 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8707 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8708 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8709 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8713 switch (N->getOpcode()) {
8714 default: llvm_unreachable("unexpected opcode for Neon base update");
8715 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8716 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8717 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8721 // Find the size of memory referenced by the load/store.
8724 VecTy = N->getValueType(0);
8726 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8727 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8729 NumBytes /= VecTy.getVectorNumElements();
8731 // If the increment is a constant, it must match the memory ref size.
8732 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8733 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8734 uint64_t IncVal = CInc->getZExtValue();
8735 if (IncVal != NumBytes)
8737 } else if (NumBytes >= 3 * 16) {
8738 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8739 // separate instructions that make it harder to use a non-constant update.
8743 // Create the new updating load/store node.
8745 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8747 for (n = 0; n < NumResultVecs; ++n)
8749 Tys[n++] = MVT::i32;
8750 Tys[n] = MVT::Other;
8751 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8752 SmallVector<SDValue, 8> Ops;
8753 Ops.push_back(N->getOperand(0)); // incoming chain
8754 Ops.push_back(N->getOperand(AddrOpIdx));
8756 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8757 Ops.push_back(N->getOperand(i));
8759 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8760 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8761 Ops.data(), Ops.size(),
8762 MemInt->getMemoryVT(),
8763 MemInt->getMemOperand());
8766 std::vector<SDValue> NewResults;
8767 for (unsigned i = 0; i < NumResultVecs; ++i) {
8768 NewResults.push_back(SDValue(UpdN.getNode(), i));
8770 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8771 DCI.CombineTo(N, NewResults);
8772 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8779 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8780 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8781 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8783 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8784 SelectionDAG &DAG = DCI.DAG;
8785 EVT VT = N->getValueType(0);
8786 // vldN-dup instructions only support 64-bit vectors for N > 1.
8787 if (!VT.is64BitVector())
8790 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8791 SDNode *VLD = N->getOperand(0).getNode();
8792 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8794 unsigned NumVecs = 0;
8795 unsigned NewOpc = 0;
8796 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8797 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8799 NewOpc = ARMISD::VLD2DUP;
8800 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8802 NewOpc = ARMISD::VLD3DUP;
8803 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8805 NewOpc = ARMISD::VLD4DUP;
8810 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8811 // numbers match the load.
8812 unsigned VLDLaneNo =
8813 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8814 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8816 // Ignore uses of the chain result.
8817 if (UI.getUse().getResNo() == NumVecs)
8820 if (User->getOpcode() != ARMISD::VDUPLANE ||
8821 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8825 // Create the vldN-dup node.
8828 for (n = 0; n < NumVecs; ++n)
8830 Tys[n] = MVT::Other;
8831 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8832 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8833 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8834 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8835 Ops, 2, VLDMemInt->getMemoryVT(),
8836 VLDMemInt->getMemOperand());
8839 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8841 unsigned ResNo = UI.getUse().getResNo();
8842 // Ignore uses of the chain result.
8843 if (ResNo == NumVecs)
8846 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8849 // Now the vldN-lane intrinsic is dead except for its chain result.
8850 // Update uses of the chain.
8851 std::vector<SDValue> VLDDupResults;
8852 for (unsigned n = 0; n < NumVecs; ++n)
8853 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8854 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8855 DCI.CombineTo(VLD, VLDDupResults);
8860 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8861 /// ARMISD::VDUPLANE.
8862 static SDValue PerformVDUPLANECombine(SDNode *N,
8863 TargetLowering::DAGCombinerInfo &DCI) {
8864 SDValue Op = N->getOperand(0);
8866 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8867 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8868 if (CombineVLDDUP(N, DCI))
8869 return SDValue(N, 0);
8871 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8872 // redundant. Ignore bit_converts for now; element sizes are checked below.
8873 while (Op.getOpcode() == ISD::BITCAST)
8874 Op = Op.getOperand(0);
8875 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8878 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8879 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8880 // The canonical VMOV for a zero vector uses a 32-bit element size.
8881 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8883 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8885 EVT VT = N->getValueType(0);
8886 if (EltSize > VT.getVectorElementType().getSizeInBits())
8889 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8892 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8893 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8894 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8898 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8900 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8905 APFloat APF = C->getValueAPF();
8906 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8907 != APFloat::opOK || !isExact)
8910 c0 = (I == 0) ? cN : c0;
8911 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8918 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8919 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8920 /// when the VMUL has a constant operand that is a power of 2.
8922 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8923 /// vmul.f32 d16, d17, d16
8924 /// vcvt.s32.f32 d16, d16
8926 /// vcvt.s32.f32 d16, d16, #3
8927 static SDValue PerformVCVTCombine(SDNode *N,
8928 TargetLowering::DAGCombinerInfo &DCI,
8929 const ARMSubtarget *Subtarget) {
8930 SelectionDAG &DAG = DCI.DAG;
8931 SDValue Op = N->getOperand(0);
8933 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8934 Op.getOpcode() != ISD::FMUL)
8938 SDValue N0 = Op->getOperand(0);
8939 SDValue ConstVec = Op->getOperand(1);
8940 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8942 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8943 !isConstVecPow2(ConstVec, isSigned, C))
8946 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8947 Intrinsic::arm_neon_vcvtfp2fxu;
8948 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8950 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8951 DAG.getConstant(Log2_64(C), MVT::i32));
8954 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8955 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8956 /// when the VDIV has a constant operand that is a power of 2.
8958 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8959 /// vcvt.f32.s32 d16, d16
8960 /// vdiv.f32 d16, d17, d16
8962 /// vcvt.f32.s32 d16, d16, #3
8963 static SDValue PerformVDIVCombine(SDNode *N,
8964 TargetLowering::DAGCombinerInfo &DCI,
8965 const ARMSubtarget *Subtarget) {
8966 SelectionDAG &DAG = DCI.DAG;
8967 SDValue Op = N->getOperand(0);
8968 unsigned OpOpcode = Op.getNode()->getOpcode();
8970 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8971 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8975 SDValue ConstVec = N->getOperand(1);
8976 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8978 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8979 !isConstVecPow2(ConstVec, isSigned, C))
8982 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8983 Intrinsic::arm_neon_vcvtfxu2fp;
8984 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8986 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8987 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8990 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8991 /// operand of a vector shift operation, where all the elements of the
8992 /// build_vector must have the same constant integer value.
8993 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8994 // Ignore bit_converts.
8995 while (Op.getOpcode() == ISD::BITCAST)
8996 Op = Op.getOperand(0);
8997 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8998 APInt SplatBits, SplatUndef;
8999 unsigned SplatBitSize;
9001 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9002 HasAnyUndefs, ElementBits) ||
9003 SplatBitSize > ElementBits)
9005 Cnt = SplatBits.getSExtValue();
9009 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9010 /// operand of a vector shift left operation. That value must be in the range:
9011 /// 0 <= Value < ElementBits for a left shift; or
9012 /// 0 <= Value <= ElementBits for a long left shift.
9013 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9014 assert(VT.isVector() && "vector shift count is not a vector type");
9015 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9016 if (! getVShiftImm(Op, ElementBits, Cnt))
9018 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9021 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9022 /// operand of a vector shift right operation. For a shift opcode, the value
9023 /// is positive, but for an intrinsic the value count must be negative. The
9024 /// absolute value must be in the range:
9025 /// 1 <= |Value| <= ElementBits for a right shift; or
9026 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9027 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9029 assert(VT.isVector() && "vector shift count is not a vector type");
9030 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9031 if (! getVShiftImm(Op, ElementBits, Cnt))
9035 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9038 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9039 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9040 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9043 // Don't do anything for most intrinsics.
9046 // Vector shifts: check for immediate versions and lower them.
9047 // Note: This is done during DAG combining instead of DAG legalizing because
9048 // the build_vectors for 64-bit vector element shift counts are generally
9049 // not legal, and it is hard to see their values after they get legalized to
9050 // loads from a constant pool.
9051 case Intrinsic::arm_neon_vshifts:
9052 case Intrinsic::arm_neon_vshiftu:
9053 case Intrinsic::arm_neon_vshiftls:
9054 case Intrinsic::arm_neon_vshiftlu:
9055 case Intrinsic::arm_neon_vshiftn:
9056 case Intrinsic::arm_neon_vrshifts:
9057 case Intrinsic::arm_neon_vrshiftu:
9058 case Intrinsic::arm_neon_vrshiftn:
9059 case Intrinsic::arm_neon_vqshifts:
9060 case Intrinsic::arm_neon_vqshiftu:
9061 case Intrinsic::arm_neon_vqshiftsu:
9062 case Intrinsic::arm_neon_vqshiftns:
9063 case Intrinsic::arm_neon_vqshiftnu:
9064 case Intrinsic::arm_neon_vqshiftnsu:
9065 case Intrinsic::arm_neon_vqrshiftns:
9066 case Intrinsic::arm_neon_vqrshiftnu:
9067 case Intrinsic::arm_neon_vqrshiftnsu: {
9068 EVT VT = N->getOperand(1).getValueType();
9070 unsigned VShiftOpc = 0;
9073 case Intrinsic::arm_neon_vshifts:
9074 case Intrinsic::arm_neon_vshiftu:
9075 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9076 VShiftOpc = ARMISD::VSHL;
9079 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9080 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9081 ARMISD::VSHRs : ARMISD::VSHRu);
9086 case Intrinsic::arm_neon_vshiftls:
9087 case Intrinsic::arm_neon_vshiftlu:
9088 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9090 llvm_unreachable("invalid shift count for vshll intrinsic");
9092 case Intrinsic::arm_neon_vrshifts:
9093 case Intrinsic::arm_neon_vrshiftu:
9094 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9098 case Intrinsic::arm_neon_vqshifts:
9099 case Intrinsic::arm_neon_vqshiftu:
9100 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9104 case Intrinsic::arm_neon_vqshiftsu:
9105 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9107 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9109 case Intrinsic::arm_neon_vshiftn:
9110 case Intrinsic::arm_neon_vrshiftn:
9111 case Intrinsic::arm_neon_vqshiftns:
9112 case Intrinsic::arm_neon_vqshiftnu:
9113 case Intrinsic::arm_neon_vqshiftnsu:
9114 case Intrinsic::arm_neon_vqrshiftns:
9115 case Intrinsic::arm_neon_vqrshiftnu:
9116 case Intrinsic::arm_neon_vqrshiftnsu:
9117 // Narrowing shifts require an immediate right shift.
9118 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9120 llvm_unreachable("invalid shift count for narrowing vector shift "
9124 llvm_unreachable("unhandled vector shift");
9128 case Intrinsic::arm_neon_vshifts:
9129 case Intrinsic::arm_neon_vshiftu:
9130 // Opcode already set above.
9132 case Intrinsic::arm_neon_vshiftls:
9133 case Intrinsic::arm_neon_vshiftlu:
9134 if (Cnt == VT.getVectorElementType().getSizeInBits())
9135 VShiftOpc = ARMISD::VSHLLi;
9137 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9138 ARMISD::VSHLLs : ARMISD::VSHLLu);
9140 case Intrinsic::arm_neon_vshiftn:
9141 VShiftOpc = ARMISD::VSHRN; break;
9142 case Intrinsic::arm_neon_vrshifts:
9143 VShiftOpc = ARMISD::VRSHRs; break;
9144 case Intrinsic::arm_neon_vrshiftu:
9145 VShiftOpc = ARMISD::VRSHRu; break;
9146 case Intrinsic::arm_neon_vrshiftn:
9147 VShiftOpc = ARMISD::VRSHRN; break;
9148 case Intrinsic::arm_neon_vqshifts:
9149 VShiftOpc = ARMISD::VQSHLs; break;
9150 case Intrinsic::arm_neon_vqshiftu:
9151 VShiftOpc = ARMISD::VQSHLu; break;
9152 case Intrinsic::arm_neon_vqshiftsu:
9153 VShiftOpc = ARMISD::VQSHLsu; break;
9154 case Intrinsic::arm_neon_vqshiftns:
9155 VShiftOpc = ARMISD::VQSHRNs; break;
9156 case Intrinsic::arm_neon_vqshiftnu:
9157 VShiftOpc = ARMISD::VQSHRNu; break;
9158 case Intrinsic::arm_neon_vqshiftnsu:
9159 VShiftOpc = ARMISD::VQSHRNsu; break;
9160 case Intrinsic::arm_neon_vqrshiftns:
9161 VShiftOpc = ARMISD::VQRSHRNs; break;
9162 case Intrinsic::arm_neon_vqrshiftnu:
9163 VShiftOpc = ARMISD::VQRSHRNu; break;
9164 case Intrinsic::arm_neon_vqrshiftnsu:
9165 VShiftOpc = ARMISD::VQRSHRNsu; break;
9168 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9169 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9172 case Intrinsic::arm_neon_vshiftins: {
9173 EVT VT = N->getOperand(1).getValueType();
9175 unsigned VShiftOpc = 0;
9177 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9178 VShiftOpc = ARMISD::VSLI;
9179 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9180 VShiftOpc = ARMISD::VSRI;
9182 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9185 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9186 N->getOperand(1), N->getOperand(2),
9187 DAG.getConstant(Cnt, MVT::i32));
9190 case Intrinsic::arm_neon_vqrshifts:
9191 case Intrinsic::arm_neon_vqrshiftu:
9192 // No immediate versions of these to check for.
9199 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9200 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9201 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9202 /// vector element shift counts are generally not legal, and it is hard to see
9203 /// their values after they get legalized to loads from a constant pool.
9204 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9205 const ARMSubtarget *ST) {
9206 EVT VT = N->getValueType(0);
9207 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9208 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9209 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9210 SDValue N1 = N->getOperand(1);
9211 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9212 SDValue N0 = N->getOperand(0);
9213 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9214 DAG.MaskedValueIsZero(N0.getOperand(0),
9215 APInt::getHighBitsSet(32, 16)))
9216 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
9220 // Nothing to be done for scalar shifts.
9221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9222 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9225 assert(ST->hasNEON() && "unexpected vector shift");
9228 switch (N->getOpcode()) {
9229 default: llvm_unreachable("unexpected shift opcode");
9232 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9233 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
9234 DAG.getConstant(Cnt, MVT::i32));
9239 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9240 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9241 ARMISD::VSHRs : ARMISD::VSHRu);
9242 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
9243 DAG.getConstant(Cnt, MVT::i32));
9249 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9250 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9251 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9252 const ARMSubtarget *ST) {
9253 SDValue N0 = N->getOperand(0);
9255 // Check for sign- and zero-extensions of vector extract operations of 8-
9256 // and 16-bit vector elements. NEON supports these directly. They are
9257 // handled during DAG combining because type legalization will promote them
9258 // to 32-bit types and it is messy to recognize the operations after that.
9259 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9260 SDValue Vec = N0.getOperand(0);
9261 SDValue Lane = N0.getOperand(1);
9262 EVT VT = N->getValueType(0);
9263 EVT EltVT = N0.getValueType();
9264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9266 if (VT == MVT::i32 &&
9267 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9268 TLI.isTypeLegal(Vec.getValueType()) &&
9269 isa<ConstantSDNode>(Lane)) {
9272 switch (N->getOpcode()) {
9273 default: llvm_unreachable("unexpected opcode");
9274 case ISD::SIGN_EXTEND:
9275 Opc = ARMISD::VGETLANEs;
9277 case ISD::ZERO_EXTEND:
9278 case ISD::ANY_EXTEND:
9279 Opc = ARMISD::VGETLANEu;
9282 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
9289 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9290 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9291 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9292 const ARMSubtarget *ST) {
9293 // If the target supports NEON, try to use vmax/vmin instructions for f32
9294 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9295 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9296 // a NaN; only do the transformation when it matches that behavior.
9298 // For now only do this when using NEON for FP operations; if using VFP, it
9299 // is not obvious that the benefit outweighs the cost of switching to the
9301 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9302 N->getValueType(0) != MVT::f32)
9305 SDValue CondLHS = N->getOperand(0);
9306 SDValue CondRHS = N->getOperand(1);
9307 SDValue LHS = N->getOperand(2);
9308 SDValue RHS = N->getOperand(3);
9309 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9311 unsigned Opcode = 0;
9313 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9314 IsReversed = false; // x CC y ? x : y
9315 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9316 IsReversed = true ; // x CC y ? y : x
9330 // If LHS is NaN, an ordered comparison will be false and the result will
9331 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9332 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9333 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9334 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9336 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9337 // will return -0, so vmin can only be used for unsafe math or if one of
9338 // the operands is known to be nonzero.
9339 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9340 !DAG.getTarget().Options.UnsafeFPMath &&
9341 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9343 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9352 // If LHS is NaN, an ordered comparison will be false and the result will
9353 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9354 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9355 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9356 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9358 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9359 // will return +0, so vmax can only be used for unsafe math or if one of
9360 // the operands is known to be nonzero.
9361 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9362 !DAG.getTarget().Options.UnsafeFPMath &&
9363 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9365 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9371 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
9374 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9376 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9377 SDValue Cmp = N->getOperand(4);
9378 if (Cmp.getOpcode() != ARMISD::CMPZ)
9379 // Only looking at EQ and NE cases.
9382 EVT VT = N->getValueType(0);
9383 DebugLoc dl = N->getDebugLoc();
9384 SDValue LHS = Cmp.getOperand(0);
9385 SDValue RHS = Cmp.getOperand(1);
9386 SDValue FalseVal = N->getOperand(0);
9387 SDValue TrueVal = N->getOperand(1);
9388 SDValue ARMcc = N->getOperand(2);
9389 ARMCC::CondCodes CC =
9390 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9408 /// FIXME: Turn this into a target neutral optimization?
9410 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9411 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9412 N->getOperand(3), Cmp);
9413 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9415 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9416 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9417 N->getOperand(3), NewCmp);
9420 if (Res.getNode()) {
9421 APInt KnownZero, KnownOne;
9422 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9423 // Capture demanded bits information that would be otherwise lost.
9424 if (KnownZero == 0xfffffffe)
9425 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9426 DAG.getValueType(MVT::i1));
9427 else if (KnownZero == 0xffffff00)
9428 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9429 DAG.getValueType(MVT::i8));
9430 else if (KnownZero == 0xffff0000)
9431 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9432 DAG.getValueType(MVT::i16));
9438 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9439 DAGCombinerInfo &DCI) const {
9440 switch (N->getOpcode()) {
9442 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9443 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9444 case ISD::SUB: return PerformSUBCombine(N, DCI);
9445 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9446 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9447 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9448 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9449 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9450 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9451 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9452 case ISD::STORE: return PerformSTORECombine(N, DCI);
9453 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9454 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9455 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9456 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9457 case ISD::FP_TO_SINT:
9458 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9459 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9460 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9463 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9464 case ISD::SIGN_EXTEND:
9465 case ISD::ZERO_EXTEND:
9466 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9467 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9468 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9469 case ARMISD::VLD2DUP:
9470 case ARMISD::VLD3DUP:
9471 case ARMISD::VLD4DUP:
9472 return CombineBaseUpdate(N, DCI);
9473 case ISD::INTRINSIC_VOID:
9474 case ISD::INTRINSIC_W_CHAIN:
9475 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9476 case Intrinsic::arm_neon_vld1:
9477 case Intrinsic::arm_neon_vld2:
9478 case Intrinsic::arm_neon_vld3:
9479 case Intrinsic::arm_neon_vld4:
9480 case Intrinsic::arm_neon_vld2lane:
9481 case Intrinsic::arm_neon_vld3lane:
9482 case Intrinsic::arm_neon_vld4lane:
9483 case Intrinsic::arm_neon_vst1:
9484 case Intrinsic::arm_neon_vst2:
9485 case Intrinsic::arm_neon_vst3:
9486 case Intrinsic::arm_neon_vst4:
9487 case Intrinsic::arm_neon_vst2lane:
9488 case Intrinsic::arm_neon_vst3lane:
9489 case Intrinsic::arm_neon_vst4lane:
9490 return CombineBaseUpdate(N, DCI);
9498 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9500 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9503 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
9504 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9505 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9507 switch (VT.getSimpleVT().SimpleTy) {
9513 // Unaligned access can use (for example) LRDB, LRDH, LDR
9514 if (AllowsUnaligned) {
9516 *Fast = Subtarget->hasV7Ops();
9523 // For any little-endian targets with neon, we can support unaligned ld/st
9524 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9525 // A big-endian target may also explictly support unaligned accesses
9526 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9536 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9537 unsigned AlignCheck) {
9538 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9539 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9542 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9543 unsigned DstAlign, unsigned SrcAlign,
9544 bool IsMemset, bool ZeroMemset,
9546 MachineFunction &MF) const {
9547 const Function *F = MF.getFunction();
9549 // See if we can use NEON instructions for this...
9550 if ((!IsMemset || ZeroMemset) &&
9551 Subtarget->hasNEON() &&
9552 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9553 Attribute::NoImplicitFloat)) {
9556 (memOpAlign(SrcAlign, DstAlign, 16) ||
9557 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
9559 } else if (Size >= 8 &&
9560 (memOpAlign(SrcAlign, DstAlign, 8) ||
9561 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
9566 // Lowering to i32/i16 if the size permits.
9572 // Let the target-independent logic figure it out.
9576 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9577 if (Val.getOpcode() != ISD::LOAD)
9580 EVT VT1 = Val.getValueType();
9581 if (!VT1.isSimple() || !VT1.isInteger() ||
9582 !VT2.isSimple() || !VT2.isInteger())
9585 switch (VT1.getSimpleVT().SimpleTy) {
9590 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9597 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9602 switch (VT.getSimpleVT().SimpleTy) {
9603 default: return false;
9618 if ((V & (Scale - 1)) != 0)
9621 return V == (V & ((1LL << 5) - 1));
9624 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9625 const ARMSubtarget *Subtarget) {
9632 switch (VT.getSimpleVT().SimpleTy) {
9633 default: return false;
9638 // + imm12 or - imm8
9640 return V == (V & ((1LL << 8) - 1));
9641 return V == (V & ((1LL << 12) - 1));
9644 // Same as ARM mode. FIXME: NEON?
9645 if (!Subtarget->hasVFP2())
9650 return V == (V & ((1LL << 8) - 1));
9654 /// isLegalAddressImmediate - Return true if the integer value can be used
9655 /// as the offset of the target addressing mode for load / store of the
9657 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9658 const ARMSubtarget *Subtarget) {
9665 if (Subtarget->isThumb1Only())
9666 return isLegalT1AddressImmediate(V, VT);
9667 else if (Subtarget->isThumb2())
9668 return isLegalT2AddressImmediate(V, VT, Subtarget);
9673 switch (VT.getSimpleVT().SimpleTy) {
9674 default: return false;
9679 return V == (V & ((1LL << 12) - 1));
9682 return V == (V & ((1LL << 8) - 1));
9685 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9690 return V == (V & ((1LL << 8) - 1));
9694 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9696 int Scale = AM.Scale;
9700 switch (VT.getSimpleVT().SimpleTy) {
9701 default: return false;
9710 return Scale == 2 || Scale == 4 || Scale == 8;
9713 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9717 // Note, we allow "void" uses (basically, uses that aren't loads or
9718 // stores), because arm allows folding a scale into many arithmetic
9719 // operations. This should be made more precise and revisited later.
9721 // Allow r << imm, but the imm has to be a multiple of two.
9722 if (Scale & 1) return false;
9723 return isPowerOf2_32(Scale);
9727 /// isLegalAddressingMode - Return true if the addressing mode represented
9728 /// by AM is legal for this target, for a load/store of the specified type.
9729 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9731 EVT VT = getValueType(Ty, true);
9732 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9735 // Can never fold addr of global into load/store.
9740 case 0: // no scale reg, must be "r+i" or "r", or "i".
9743 if (Subtarget->isThumb1Only())
9747 // ARM doesn't support any R+R*scale+imm addr modes.
9754 if (Subtarget->isThumb2())
9755 return isLegalT2ScaledAddressingMode(AM, VT);
9757 int Scale = AM.Scale;
9758 switch (VT.getSimpleVT().SimpleTy) {
9759 default: return false;
9763 if (Scale < 0) Scale = -Scale;
9767 return isPowerOf2_32(Scale & ~1);
9771 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9776 // Note, we allow "void" uses (basically, uses that aren't loads or
9777 // stores), because arm allows folding a scale into many arithmetic
9778 // operations. This should be made more precise and revisited later.
9780 // Allow r << imm, but the imm has to be a multiple of two.
9781 if (Scale & 1) return false;
9782 return isPowerOf2_32(Scale);
9788 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9789 /// icmp immediate, that is the target has icmp instructions which can compare
9790 /// a register against the immediate without having to materialize the
9791 /// immediate into a register.
9792 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9793 // Thumb2 and ARM modes can use cmn for negative immediates.
9794 if (!Subtarget->isThumb())
9795 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9796 if (Subtarget->isThumb2())
9797 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9798 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9799 return Imm >= 0 && Imm <= 255;
9802 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9803 /// *or sub* immediate, that is the target has add or sub instructions which can
9804 /// add a register with the immediate without having to materialize the
9805 /// immediate into a register.
9806 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9807 // Same encoding for add/sub, just flip the sign.
9808 int64_t AbsImm = llvm::abs64(Imm);
9809 if (!Subtarget->isThumb())
9810 return ARM_AM::getSOImmVal(AbsImm) != -1;
9811 if (Subtarget->isThumb2())
9812 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9813 // Thumb1 only has 8-bit unsigned immediate.
9814 return AbsImm >= 0 && AbsImm <= 255;
9817 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9818 bool isSEXTLoad, SDValue &Base,
9819 SDValue &Offset, bool &isInc,
9820 SelectionDAG &DAG) {
9821 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9824 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9826 Base = Ptr->getOperand(0);
9827 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9828 int RHSC = (int)RHS->getZExtValue();
9829 if (RHSC < 0 && RHSC > -256) {
9830 assert(Ptr->getOpcode() == ISD::ADD);
9832 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9836 isInc = (Ptr->getOpcode() == ISD::ADD);
9837 Offset = Ptr->getOperand(1);
9839 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9841 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9842 int RHSC = (int)RHS->getZExtValue();
9843 if (RHSC < 0 && RHSC > -0x1000) {
9844 assert(Ptr->getOpcode() == ISD::ADD);
9846 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9847 Base = Ptr->getOperand(0);
9852 if (Ptr->getOpcode() == ISD::ADD) {
9854 ARM_AM::ShiftOpc ShOpcVal=
9855 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9856 if (ShOpcVal != ARM_AM::no_shift) {
9857 Base = Ptr->getOperand(1);
9858 Offset = Ptr->getOperand(0);
9860 Base = Ptr->getOperand(0);
9861 Offset = Ptr->getOperand(1);
9866 isInc = (Ptr->getOpcode() == ISD::ADD);
9867 Base = Ptr->getOperand(0);
9868 Offset = Ptr->getOperand(1);
9872 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9876 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9877 bool isSEXTLoad, SDValue &Base,
9878 SDValue &Offset, bool &isInc,
9879 SelectionDAG &DAG) {
9880 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9883 Base = Ptr->getOperand(0);
9884 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9885 int RHSC = (int)RHS->getZExtValue();
9886 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9887 assert(Ptr->getOpcode() == ISD::ADD);
9889 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9891 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9892 isInc = Ptr->getOpcode() == ISD::ADD;
9893 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9901 /// getPreIndexedAddressParts - returns true by value, base pointer and
9902 /// offset pointer and addressing mode by reference if the node's address
9903 /// can be legally represented as pre-indexed load / store address.
9905 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9907 ISD::MemIndexedMode &AM,
9908 SelectionDAG &DAG) const {
9909 if (Subtarget->isThumb1Only())
9914 bool isSEXTLoad = false;
9915 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9916 Ptr = LD->getBasePtr();
9917 VT = LD->getMemoryVT();
9918 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9919 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9920 Ptr = ST->getBasePtr();
9921 VT = ST->getMemoryVT();
9926 bool isLegal = false;
9927 if (Subtarget->isThumb2())
9928 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9929 Offset, isInc, DAG);
9931 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9932 Offset, isInc, DAG);
9936 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9940 /// getPostIndexedAddressParts - returns true by value, base pointer and
9941 /// offset pointer and addressing mode by reference if this node can be
9942 /// combined with a load / store to form a post-indexed load / store.
9943 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9946 ISD::MemIndexedMode &AM,
9947 SelectionDAG &DAG) const {
9948 if (Subtarget->isThumb1Only())
9953 bool isSEXTLoad = false;
9954 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9955 VT = LD->getMemoryVT();
9956 Ptr = LD->getBasePtr();
9957 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9958 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9959 VT = ST->getMemoryVT();
9960 Ptr = ST->getBasePtr();
9965 bool isLegal = false;
9966 if (Subtarget->isThumb2())
9967 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9970 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9976 // Swap base ptr and offset to catch more post-index load / store when
9977 // it's legal. In Thumb2 mode, offset must be an immediate.
9978 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9979 !Subtarget->isThumb2())
9980 std::swap(Base, Offset);
9982 // Post-indexed load / store update the base pointer.
9987 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9991 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9994 const SelectionDAG &DAG,
9995 unsigned Depth) const {
9996 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9997 switch (Op.getOpcode()) {
9999 case ARMISD::CMOV: {
10000 // Bits are known zero/one if known on the LHS and RHS.
10001 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10002 if (KnownZero == 0 && KnownOne == 0) return;
10004 APInt KnownZeroRHS, KnownOneRHS;
10005 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10006 KnownZero &= KnownZeroRHS;
10007 KnownOne &= KnownOneRHS;
10013 //===----------------------------------------------------------------------===//
10014 // ARM Inline Assembly Support
10015 //===----------------------------------------------------------------------===//
10017 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10018 // Looking for "rev" which is V6+.
10019 if (!Subtarget->hasV6Ops())
10022 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10023 std::string AsmStr = IA->getAsmString();
10024 SmallVector<StringRef, 4> AsmPieces;
10025 SplitString(AsmStr, AsmPieces, ";\n");
10027 switch (AsmPieces.size()) {
10028 default: return false;
10030 AsmStr = AsmPieces[0];
10032 SplitString(AsmStr, AsmPieces, " \t,");
10035 if (AsmPieces.size() == 3 &&
10036 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10037 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10038 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10039 if (Ty && Ty->getBitWidth() == 32)
10040 return IntrinsicLowering::LowerToByteSwap(CI);
10048 /// getConstraintType - Given a constraint letter, return the type of
10049 /// constraint it is for this target.
10050 ARMTargetLowering::ConstraintType
10051 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10052 if (Constraint.size() == 1) {
10053 switch (Constraint[0]) {
10055 case 'l': return C_RegisterClass;
10056 case 'w': return C_RegisterClass;
10057 case 'h': return C_RegisterClass;
10058 case 'x': return C_RegisterClass;
10059 case 't': return C_RegisterClass;
10060 case 'j': return C_Other; // Constant for movw.
10061 // An address with a single base register. Due to the way we
10062 // currently handle addresses it is the same as an 'r' memory constraint.
10063 case 'Q': return C_Memory;
10065 } else if (Constraint.size() == 2) {
10066 switch (Constraint[0]) {
10068 // All 'U+' constraints are addresses.
10069 case 'U': return C_Memory;
10072 return TargetLowering::getConstraintType(Constraint);
10075 /// Examine constraint type and operand type and determine a weight value.
10076 /// This object must already have been set up with the operand type
10077 /// and the current alternative constraint selected.
10078 TargetLowering::ConstraintWeight
10079 ARMTargetLowering::getSingleConstraintMatchWeight(
10080 AsmOperandInfo &info, const char *constraint) const {
10081 ConstraintWeight weight = CW_Invalid;
10082 Value *CallOperandVal = info.CallOperandVal;
10083 // If we don't have a value, we can't do a match,
10084 // but allow it at the lowest weight.
10085 if (CallOperandVal == NULL)
10087 Type *type = CallOperandVal->getType();
10088 // Look at the constraint type.
10089 switch (*constraint) {
10091 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10094 if (type->isIntegerTy()) {
10095 if (Subtarget->isThumb())
10096 weight = CW_SpecificReg;
10098 weight = CW_Register;
10102 if (type->isFloatingPointTy())
10103 weight = CW_Register;
10109 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10111 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10113 if (Constraint.size() == 1) {
10114 // GCC ARM Constraint Letters
10115 switch (Constraint[0]) {
10116 case 'l': // Low regs or general regs.
10117 if (Subtarget->isThumb())
10118 return RCPair(0U, &ARM::tGPRRegClass);
10119 return RCPair(0U, &ARM::GPRRegClass);
10120 case 'h': // High regs or no regs.
10121 if (Subtarget->isThumb())
10122 return RCPair(0U, &ARM::hGPRRegClass);
10125 return RCPair(0U, &ARM::GPRRegClass);
10127 if (VT == MVT::f32)
10128 return RCPair(0U, &ARM::SPRRegClass);
10129 if (VT.getSizeInBits() == 64)
10130 return RCPair(0U, &ARM::DPRRegClass);
10131 if (VT.getSizeInBits() == 128)
10132 return RCPair(0U, &ARM::QPRRegClass);
10135 if (VT == MVT::f32)
10136 return RCPair(0U, &ARM::SPR_8RegClass);
10137 if (VT.getSizeInBits() == 64)
10138 return RCPair(0U, &ARM::DPR_8RegClass);
10139 if (VT.getSizeInBits() == 128)
10140 return RCPair(0U, &ARM::QPR_8RegClass);
10143 if (VT == MVT::f32)
10144 return RCPair(0U, &ARM::SPRRegClass);
10148 if (StringRef("{cc}").equals_lower(Constraint))
10149 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10151 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10154 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10155 /// vector. If it is invalid, don't add anything to Ops.
10156 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10157 std::string &Constraint,
10158 std::vector<SDValue>&Ops,
10159 SelectionDAG &DAG) const {
10160 SDValue Result(0, 0);
10162 // Currently only support length 1 constraints.
10163 if (Constraint.length() != 1) return;
10165 char ConstraintLetter = Constraint[0];
10166 switch (ConstraintLetter) {
10169 case 'I': case 'J': case 'K': case 'L':
10170 case 'M': case 'N': case 'O':
10171 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10175 int64_t CVal64 = C->getSExtValue();
10176 int CVal = (int) CVal64;
10177 // None of these constraints allow values larger than 32 bits. Check
10178 // that the value fits in an int.
10179 if (CVal != CVal64)
10182 switch (ConstraintLetter) {
10184 // Constant suitable for movw, must be between 0 and
10186 if (Subtarget->hasV6T2Ops())
10187 if (CVal >= 0 && CVal <= 65535)
10191 if (Subtarget->isThumb1Only()) {
10192 // This must be a constant between 0 and 255, for ADD
10194 if (CVal >= 0 && CVal <= 255)
10196 } else if (Subtarget->isThumb2()) {
10197 // A constant that can be used as an immediate value in a
10198 // data-processing instruction.
10199 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10202 // A constant that can be used as an immediate value in a
10203 // data-processing instruction.
10204 if (ARM_AM::getSOImmVal(CVal) != -1)
10210 if (Subtarget->isThumb()) { // FIXME thumb2
10211 // This must be a constant between -255 and -1, for negated ADD
10212 // immediates. This can be used in GCC with an "n" modifier that
10213 // prints the negated value, for use with SUB instructions. It is
10214 // not useful otherwise but is implemented for compatibility.
10215 if (CVal >= -255 && CVal <= -1)
10218 // This must be a constant between -4095 and 4095. It is not clear
10219 // what this constraint is intended for. Implemented for
10220 // compatibility with GCC.
10221 if (CVal >= -4095 && CVal <= 4095)
10227 if (Subtarget->isThumb1Only()) {
10228 // A 32-bit value where only one byte has a nonzero value. Exclude
10229 // zero to match GCC. This constraint is used by GCC internally for
10230 // constants that can be loaded with a move/shift combination.
10231 // It is not useful otherwise but is implemented for compatibility.
10232 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10234 } else if (Subtarget->isThumb2()) {
10235 // A constant whose bitwise inverse can be used as an immediate
10236 // value in a data-processing instruction. This can be used in GCC
10237 // with a "B" modifier that prints the inverted value, for use with
10238 // BIC and MVN instructions. It is not useful otherwise but is
10239 // implemented for compatibility.
10240 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10243 // A constant whose bitwise inverse can be used as an immediate
10244 // value in a data-processing instruction. This can be used in GCC
10245 // with a "B" modifier that prints the inverted value, for use with
10246 // BIC and MVN instructions. It is not useful otherwise but is
10247 // implemented for compatibility.
10248 if (ARM_AM::getSOImmVal(~CVal) != -1)
10254 if (Subtarget->isThumb1Only()) {
10255 // This must be a constant between -7 and 7,
10256 // for 3-operand ADD/SUB immediate instructions.
10257 if (CVal >= -7 && CVal < 7)
10259 } else if (Subtarget->isThumb2()) {
10260 // A constant whose negation can be used as an immediate value in a
10261 // data-processing instruction. This can be used in GCC with an "n"
10262 // modifier that prints the negated value, for use with SUB
10263 // instructions. It is not useful otherwise but is implemented for
10265 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10268 // A constant whose negation can be used as an immediate value in a
10269 // data-processing instruction. This can be used in GCC with an "n"
10270 // modifier that prints the negated value, for use with SUB
10271 // instructions. It is not useful otherwise but is implemented for
10273 if (ARM_AM::getSOImmVal(-CVal) != -1)
10279 if (Subtarget->isThumb()) { // FIXME thumb2
10280 // This must be a multiple of 4 between 0 and 1020, for
10281 // ADD sp + immediate.
10282 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10285 // A power of two or a constant between 0 and 32. This is used in
10286 // GCC for the shift amount on shifted register operands, but it is
10287 // useful in general for any shift amounts.
10288 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10294 if (Subtarget->isThumb()) { // FIXME thumb2
10295 // This must be a constant between 0 and 31, for shift amounts.
10296 if (CVal >= 0 && CVal <= 31)
10302 if (Subtarget->isThumb()) { // FIXME thumb2
10303 // This must be a multiple of 4 between -508 and 508, for
10304 // ADD/SUB sp = sp + immediate.
10305 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10310 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10314 if (Result.getNode()) {
10315 Ops.push_back(Result);
10318 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10322 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10323 // The ARM target isn't yet aware of offsets.
10327 bool ARM::isBitFieldInvertedMask(unsigned v) {
10328 if (v == 0xffffffff)
10330 // there can be 1's on either or both "outsides", all the "inside"
10331 // bits must be 0's
10332 unsigned int lsb = 0, msb = 31;
10333 while (v & (1 << msb)) --msb;
10334 while (v & (1 << lsb)) ++lsb;
10335 for (unsigned int i = lsb; i <= msb; ++i) {
10342 /// isFPImmLegal - Returns true if the target can instruction select the
10343 /// specified FP immediate natively. If false, the legalizer will
10344 /// materialize the FP immediate as a load from a constant pool.
10345 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10346 if (!Subtarget->hasVFP3())
10348 if (VT == MVT::f32)
10349 return ARM_AM::getFP32Imm(Imm) != -1;
10350 if (VT == MVT::f64)
10351 return ARM_AM::getFP64Imm(Imm) != -1;
10355 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10356 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10357 /// specified in the intrinsic calls.
10358 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10360 unsigned Intrinsic) const {
10361 switch (Intrinsic) {
10362 case Intrinsic::arm_neon_vld1:
10363 case Intrinsic::arm_neon_vld2:
10364 case Intrinsic::arm_neon_vld3:
10365 case Intrinsic::arm_neon_vld4:
10366 case Intrinsic::arm_neon_vld2lane:
10367 case Intrinsic::arm_neon_vld3lane:
10368 case Intrinsic::arm_neon_vld4lane: {
10369 Info.opc = ISD::INTRINSIC_W_CHAIN;
10370 // Conservatively set memVT to the entire set of vectors loaded.
10371 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10372 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10373 Info.ptrVal = I.getArgOperand(0);
10375 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10376 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10377 Info.vol = false; // volatile loads with NEON intrinsics not supported
10378 Info.readMem = true;
10379 Info.writeMem = false;
10382 case Intrinsic::arm_neon_vst1:
10383 case Intrinsic::arm_neon_vst2:
10384 case Intrinsic::arm_neon_vst3:
10385 case Intrinsic::arm_neon_vst4:
10386 case Intrinsic::arm_neon_vst2lane:
10387 case Intrinsic::arm_neon_vst3lane:
10388 case Intrinsic::arm_neon_vst4lane: {
10389 Info.opc = ISD::INTRINSIC_VOID;
10390 // Conservatively set memVT to the entire set of vectors stored.
10391 unsigned NumElts = 0;
10392 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10393 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10394 if (!ArgTy->isVectorTy())
10396 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10398 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10399 Info.ptrVal = I.getArgOperand(0);
10401 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10402 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10403 Info.vol = false; // volatile stores with NEON intrinsics not supported
10404 Info.readMem = false;
10405 Info.writeMem = true;
10408 case Intrinsic::arm_strexd: {
10409 Info.opc = ISD::INTRINSIC_W_CHAIN;
10410 Info.memVT = MVT::i64;
10411 Info.ptrVal = I.getArgOperand(2);
10415 Info.readMem = false;
10416 Info.writeMem = true;
10419 case Intrinsic::arm_ldrexd: {
10420 Info.opc = ISD::INTRINSIC_W_CHAIN;
10421 Info.memVT = MVT::i64;
10422 Info.ptrVal = I.getArgOperand(0);
10426 Info.readMem = true;
10427 Info.writeMem = false;