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, SmallVectorImpl<CCValAssign> &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);
455 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
457 if (Subtarget->hasNEON()) {
458 addDRTypeForNEON(MVT::v2f32);
459 addDRTypeForNEON(MVT::v8i8);
460 addDRTypeForNEON(MVT::v4i16);
461 addDRTypeForNEON(MVT::v2i32);
462 addDRTypeForNEON(MVT::v1i64);
464 addQRTypeForNEON(MVT::v4f32);
465 addQRTypeForNEON(MVT::v2f64);
466 addQRTypeForNEON(MVT::v16i8);
467 addQRTypeForNEON(MVT::v8i16);
468 addQRTypeForNEON(MVT::v4i32);
469 addQRTypeForNEON(MVT::v2i64);
471 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
472 // neither Neon nor VFP support any arithmetic operations on it.
473 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
474 // supported for v4f32.
475 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
476 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
477 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
478 // FIXME: Code duplication: FDIV and FREM are expanded always, see
479 // ARMTargetLowering::addTypeForNEON method for details.
480 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
481 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
482 // FIXME: Create unittest.
483 // In another words, find a way when "copysign" appears in DAG with vector
485 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
486 // FIXME: Code duplication: SETCC has custom operation action, see
487 // ARMTargetLowering::addTypeForNEON method for details.
488 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
489 // FIXME: Create unittest for FNEG and for FABS.
490 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
491 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
493 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
494 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
496 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
499 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
501 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
502 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
503 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
504 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
505 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
507 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
508 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
510 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
511 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
512 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
514 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
519 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
520 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
521 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
522 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
524 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
526 // Mark v2f32 intrinsics.
527 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
528 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
529 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
530 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
531 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
532 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
533 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
534 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
535 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
536 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
537 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
538 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
539 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
540 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
541 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
543 // Neon does not support some operations on v1i64 and v2i64 types.
544 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
545 // Custom handling for some quad-vector types to detect VMULL.
546 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
547 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
548 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
549 // Custom handling for some vector types to avoid expensive expansions
550 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
551 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
552 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
553 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
554 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
555 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
556 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
557 // a destination type that is wider than the source, and nor does
558 // it have a FP_TO_[SU]INT instruction with a narrower destination than
560 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
561 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
563 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
565 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
566 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
568 // Custom expand long extensions to vectors.
569 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
570 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
571 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
572 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
573 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
574 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
575 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
576 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
578 // NEON does not have single instruction CTPOP for vectors with element
579 // types wider than 8-bits. However, custom lowering can leverage the
580 // v8i8/v16i8 vcnt instruction.
581 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
582 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
583 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
584 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
586 // NEON only has FMA instructions as of VFP4.
587 if (!Subtarget->hasVFP4()) {
588 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
589 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
592 setTargetDAGCombine(ISD::INTRINSIC_VOID);
593 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
594 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
595 setTargetDAGCombine(ISD::SHL);
596 setTargetDAGCombine(ISD::SRL);
597 setTargetDAGCombine(ISD::SRA);
598 setTargetDAGCombine(ISD::SIGN_EXTEND);
599 setTargetDAGCombine(ISD::ZERO_EXTEND);
600 setTargetDAGCombine(ISD::ANY_EXTEND);
601 setTargetDAGCombine(ISD::SELECT_CC);
602 setTargetDAGCombine(ISD::BUILD_VECTOR);
603 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
604 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
605 setTargetDAGCombine(ISD::STORE);
606 setTargetDAGCombine(ISD::FP_TO_SINT);
607 setTargetDAGCombine(ISD::FP_TO_UINT);
608 setTargetDAGCombine(ISD::FDIV);
610 // It is legal to extload from v4i8 to v4i16 or v4i32.
611 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
612 MVT::v4i16, MVT::v2i16,
614 for (unsigned i = 0; i < 6; ++i) {
615 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
616 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
617 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
621 // ARM and Thumb2 support UMLAL/SMLAL.
622 if (!Subtarget->isThumb1Only())
623 setTargetDAGCombine(ISD::ADDC);
626 computeRegisterProperties();
628 // ARM does not have f32 extending load.
629 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
631 // ARM does not have i1 sign extending load.
632 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
634 // ARM supports all 4 flavors of integer indexed load / store.
635 if (!Subtarget->isThumb1Only()) {
636 for (unsigned im = (unsigned)ISD::PRE_INC;
637 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
638 setIndexedLoadAction(im, MVT::i1, Legal);
639 setIndexedLoadAction(im, MVT::i8, Legal);
640 setIndexedLoadAction(im, MVT::i16, Legal);
641 setIndexedLoadAction(im, MVT::i32, Legal);
642 setIndexedStoreAction(im, MVT::i1, Legal);
643 setIndexedStoreAction(im, MVT::i8, Legal);
644 setIndexedStoreAction(im, MVT::i16, Legal);
645 setIndexedStoreAction(im, MVT::i32, Legal);
649 // i64 operation support.
650 setOperationAction(ISD::MUL, MVT::i64, Expand);
651 setOperationAction(ISD::MULHU, MVT::i32, Expand);
652 if (Subtarget->isThumb1Only()) {
653 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
654 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
656 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
657 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
658 setOperationAction(ISD::MULHS, MVT::i32, Expand);
660 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
661 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
662 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
663 setOperationAction(ISD::SRL, MVT::i64, Custom);
664 setOperationAction(ISD::SRA, MVT::i64, Custom);
666 if (!Subtarget->isThumb1Only()) {
667 // FIXME: We should do this for Thumb1 as well.
668 setOperationAction(ISD::ADDC, MVT::i32, Custom);
669 setOperationAction(ISD::ADDE, MVT::i32, Custom);
670 setOperationAction(ISD::SUBC, MVT::i32, Custom);
671 setOperationAction(ISD::SUBE, MVT::i32, Custom);
674 // ARM does not have ROTL.
675 setOperationAction(ISD::ROTL, MVT::i32, Expand);
676 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
677 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
678 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
679 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
681 // These just redirect to CTTZ and CTLZ on ARM.
682 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
683 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
685 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
687 // Only ARMv6 has BSWAP.
688 if (!Subtarget->hasV6Ops())
689 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
691 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
692 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
693 // These are expanded into libcalls if the cpu doesn't have HW divider.
694 setOperationAction(ISD::SDIV, MVT::i32, Expand);
695 setOperationAction(ISD::UDIV, MVT::i32, Expand);
698 // FIXME: Also set divmod for SREM on EABI
699 setOperationAction(ISD::SREM, MVT::i32, Expand);
700 setOperationAction(ISD::UREM, MVT::i32, Expand);
701 // Register based DivRem for AEABI (RTABI 4.2)
702 if (Subtarget->isTargetAEABI()) {
703 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
704 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
705 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
706 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
707 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
708 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
709 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
710 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
712 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
713 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
714 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
715 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
716 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
717 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
718 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
719 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
721 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
722 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
724 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
725 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
728 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
729 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
730 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
731 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
732 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
734 setOperationAction(ISD::TRAP, MVT::Other, Legal);
736 // Use the default implementation.
737 setOperationAction(ISD::VASTART, MVT::Other, Custom);
738 setOperationAction(ISD::VAARG, MVT::Other, Expand);
739 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
740 setOperationAction(ISD::VAEND, MVT::Other, Expand);
741 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
742 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
744 if (!Subtarget->isTargetDarwin()) {
745 // Non-Darwin platforms may return values in these registers via the
746 // personality function.
747 setExceptionPointerRegister(ARM::R0);
748 setExceptionSelectorRegister(ARM::R1);
751 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
752 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
753 // the default expansion.
754 // FIXME: This should be checking for v6k, not just v6.
755 if (Subtarget->hasDataBarrier() ||
756 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
757 // membarrier needs custom lowering; the rest are legal and handled
759 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
760 // Custom lowering for 64-bit ops
761 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
762 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
763 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
764 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
765 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
766 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
767 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
768 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
769 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
770 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
771 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
772 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
773 setInsertFencesForAtomic(true);
775 // Set them all for expansion, which will force libcalls.
776 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
777 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
778 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
779 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
780 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
781 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
782 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
783 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
784 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
785 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
786 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
787 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
788 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
789 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
790 // Unordered/Monotonic case.
791 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
792 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
795 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
797 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
798 if (!Subtarget->hasV6Ops()) {
799 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
800 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
802 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
804 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
805 !Subtarget->isThumb1Only()) {
806 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
807 // iff target supports vfp2.
808 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
809 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
812 // We want to custom lower some of our intrinsics.
813 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
814 if (Subtarget->isTargetDarwin()) {
815 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
816 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
817 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
820 setOperationAction(ISD::SETCC, MVT::i32, Expand);
821 setOperationAction(ISD::SETCC, MVT::f32, Expand);
822 setOperationAction(ISD::SETCC, MVT::f64, Expand);
823 setOperationAction(ISD::SELECT, MVT::i32, Custom);
824 setOperationAction(ISD::SELECT, MVT::f32, Custom);
825 setOperationAction(ISD::SELECT, MVT::f64, Custom);
826 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
827 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
828 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
830 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
831 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
832 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
833 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
834 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
836 // We don't support sin/cos/fmod/copysign/pow
837 setOperationAction(ISD::FSIN, MVT::f64, Expand);
838 setOperationAction(ISD::FSIN, MVT::f32, Expand);
839 setOperationAction(ISD::FCOS, MVT::f32, Expand);
840 setOperationAction(ISD::FCOS, MVT::f64, Expand);
841 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
842 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
843 setOperationAction(ISD::FREM, MVT::f64, Expand);
844 setOperationAction(ISD::FREM, MVT::f32, Expand);
845 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
846 !Subtarget->isThumb1Only()) {
847 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
848 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
850 setOperationAction(ISD::FPOW, MVT::f64, Expand);
851 setOperationAction(ISD::FPOW, MVT::f32, Expand);
853 if (!Subtarget->hasVFP4()) {
854 setOperationAction(ISD::FMA, MVT::f64, Expand);
855 setOperationAction(ISD::FMA, MVT::f32, Expand);
858 // Various VFP goodness
859 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
860 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
861 if (Subtarget->hasVFP2()) {
862 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
863 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
864 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
865 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
867 // Special handling for half-precision FP.
868 if (!Subtarget->hasFP16()) {
869 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
870 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
874 // We have target-specific dag combine patterns for the following nodes:
875 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
876 setTargetDAGCombine(ISD::ADD);
877 setTargetDAGCombine(ISD::SUB);
878 setTargetDAGCombine(ISD::MUL);
879 setTargetDAGCombine(ISD::AND);
880 setTargetDAGCombine(ISD::OR);
881 setTargetDAGCombine(ISD::XOR);
883 if (Subtarget->hasV6Ops())
884 setTargetDAGCombine(ISD::SRL);
886 setStackPointerRegisterToSaveRestore(ARM::SP);
888 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
889 !Subtarget->hasVFP2())
890 setSchedulingPreference(Sched::RegPressure);
892 setSchedulingPreference(Sched::Hybrid);
894 //// temporary - rewrite interface to use type
895 MaxStoresPerMemset = 8;
896 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
897 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
898 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
899 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
900 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
902 // On ARM arguments smaller than 4 bytes are extended, so all arguments
903 // are at least 4 bytes aligned.
904 setMinStackArgumentAlignment(4);
906 // Prefer likely predicted branches to selects on out-of-order cores.
907 PredictableSelectIsExpensive = Subtarget->isLikeA9();
909 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
912 // FIXME: It might make sense to define the representative register class as the
913 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
914 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
915 // SPR's representative would be DPR_VFP2. This should work well if register
916 // pressure tracking were modified such that a register use would increment the
917 // pressure of the register class's representative and all of it's super
918 // classes' representatives transitively. We have not implemented this because
919 // of the difficulty prior to coalescing of modeling operand register classes
920 // due to the common occurrence of cross class copies and subregister insertions
922 std::pair<const TargetRegisterClass*, uint8_t>
923 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
924 const TargetRegisterClass *RRC = 0;
926 switch (VT.SimpleTy) {
928 return TargetLowering::findRepresentativeClass(VT);
929 // Use DPR as representative register class for all floating point
930 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
931 // the cost is 1 for both f32 and f64.
932 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
933 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
934 RRC = &ARM::DPRRegClass;
935 // When NEON is used for SP, only half of the register file is available
936 // because operations that define both SP and DP results will be constrained
937 // to the VFP2 class (D0-D15). We currently model this constraint prior to
938 // coalescing by double-counting the SP regs. See the FIXME above.
939 if (Subtarget->useNEONForSinglePrecisionFP())
942 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
943 case MVT::v4f32: case MVT::v2f64:
944 RRC = &ARM::DPRRegClass;
948 RRC = &ARM::DPRRegClass;
952 RRC = &ARM::DPRRegClass;
956 return std::make_pair(RRC, Cost);
959 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
962 case ARMISD::Wrapper: return "ARMISD::Wrapper";
963 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
964 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
965 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
966 case ARMISD::CALL: return "ARMISD::CALL";
967 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
968 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
969 case ARMISD::tCALL: return "ARMISD::tCALL";
970 case ARMISD::BRCOND: return "ARMISD::BRCOND";
971 case ARMISD::BR_JT: return "ARMISD::BR_JT";
972 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
973 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
974 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
975 case ARMISD::CMP: return "ARMISD::CMP";
976 case ARMISD::CMN: return "ARMISD::CMN";
977 case ARMISD::CMPZ: return "ARMISD::CMPZ";
978 case ARMISD::CMPFP: return "ARMISD::CMPFP";
979 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
980 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
981 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
983 case ARMISD::CMOV: return "ARMISD::CMOV";
985 case ARMISD::RBIT: return "ARMISD::RBIT";
987 case ARMISD::FTOSI: return "ARMISD::FTOSI";
988 case ARMISD::FTOUI: return "ARMISD::FTOUI";
989 case ARMISD::SITOF: return "ARMISD::SITOF";
990 case ARMISD::UITOF: return "ARMISD::UITOF";
992 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
993 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
994 case ARMISD::RRX: return "ARMISD::RRX";
996 case ARMISD::ADDC: return "ARMISD::ADDC";
997 case ARMISD::ADDE: return "ARMISD::ADDE";
998 case ARMISD::SUBC: return "ARMISD::SUBC";
999 case ARMISD::SUBE: return "ARMISD::SUBE";
1001 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1002 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1004 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1005 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1007 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1009 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1011 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1013 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
1014 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1016 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1018 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1019 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1020 case ARMISD::VCGE: return "ARMISD::VCGE";
1021 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1022 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1023 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1024 case ARMISD::VCGT: return "ARMISD::VCGT";
1025 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1026 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1027 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1028 case ARMISD::VTST: return "ARMISD::VTST";
1030 case ARMISD::VSHL: return "ARMISD::VSHL";
1031 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1032 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1033 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1034 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1035 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1036 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1037 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1038 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1039 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1040 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1041 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1042 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1043 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1044 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1045 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1046 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1047 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1048 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1049 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1050 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1051 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1052 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1053 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1054 case ARMISD::VDUP: return "ARMISD::VDUP";
1055 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1056 case ARMISD::VEXT: return "ARMISD::VEXT";
1057 case ARMISD::VREV64: return "ARMISD::VREV64";
1058 case ARMISD::VREV32: return "ARMISD::VREV32";
1059 case ARMISD::VREV16: return "ARMISD::VREV16";
1060 case ARMISD::VZIP: return "ARMISD::VZIP";
1061 case ARMISD::VUZP: return "ARMISD::VUZP";
1062 case ARMISD::VTRN: return "ARMISD::VTRN";
1063 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1064 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1065 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1066 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1067 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1068 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1069 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1070 case ARMISD::FMAX: return "ARMISD::FMAX";
1071 case ARMISD::FMIN: return "ARMISD::FMIN";
1072 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1073 case ARMISD::VMINNM: return "ARMISD::VMIN";
1074 case ARMISD::BFI: return "ARMISD::BFI";
1075 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1076 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1077 case ARMISD::VBSL: return "ARMISD::VBSL";
1078 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1079 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1080 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1081 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1082 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1083 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1084 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1085 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1086 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1087 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1088 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1089 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1090 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1091 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1092 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1093 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1094 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1095 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1096 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1097 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1099 case ARMISD::ATOMADD64_DAG: return "ATOMADD64_DAG";
1100 case ARMISD::ATOMSUB64_DAG: return "ATOMSUB64_DAG";
1101 case ARMISD::ATOMOR64_DAG: return "ATOMOR64_DAG";
1102 case ARMISD::ATOMXOR64_DAG: return "ATOMXOR64_DAG";
1103 case ARMISD::ATOMAND64_DAG: return "ATOMAND64_DAG";
1104 case ARMISD::ATOMNAND64_DAG: return "ATOMNAND64_DAG";
1105 case ARMISD::ATOMSWAP64_DAG: return "ATOMSWAP64_DAG";
1106 case ARMISD::ATOMCMPXCHG64_DAG: return "ATOMCMPXCHG64_DAG";
1107 case ARMISD::ATOMMIN64_DAG: return "ATOMMIN64_DAG";
1108 case ARMISD::ATOMUMIN64_DAG: return "ATOMUMIN64_DAG";
1109 case ARMISD::ATOMMAX64_DAG: return "ATOMMAX64_DAG";
1110 case ARMISD::ATOMUMAX64_DAG: return "ATOMUMAX64_DAG";
1114 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1115 if (!VT.isVector()) return getPointerTy();
1116 return VT.changeVectorElementTypeToInteger();
1119 /// getRegClassFor - Return the register class that should be used for the
1120 /// specified value type.
1121 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1122 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1123 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1124 // load / store 4 to 8 consecutive D registers.
1125 if (Subtarget->hasNEON()) {
1126 if (VT == MVT::v4i64)
1127 return &ARM::QQPRRegClass;
1128 if (VT == MVT::v8i64)
1129 return &ARM::QQQQPRRegClass;
1131 return TargetLowering::getRegClassFor(VT);
1134 // Create a fast isel object.
1136 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1137 const TargetLibraryInfo *libInfo) const {
1138 return ARM::createFastISel(funcInfo, libInfo);
1141 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1142 /// be used for loads / stores from the global.
1143 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1144 return (Subtarget->isThumb1Only() ? 127 : 4095);
1147 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1148 unsigned NumVals = N->getNumValues();
1150 return Sched::RegPressure;
1152 for (unsigned i = 0; i != NumVals; ++i) {
1153 EVT VT = N->getValueType(i);
1154 if (VT == MVT::Glue || VT == MVT::Other)
1156 if (VT.isFloatingPoint() || VT.isVector())
1160 if (!N->isMachineOpcode())
1161 return Sched::RegPressure;
1163 // Load are scheduled for latency even if there instruction itinerary
1164 // is not available.
1165 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1166 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1168 if (MCID.getNumDefs() == 0)
1169 return Sched::RegPressure;
1170 if (!Itins->isEmpty() &&
1171 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1174 return Sched::RegPressure;
1177 //===----------------------------------------------------------------------===//
1179 //===----------------------------------------------------------------------===//
1181 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1182 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1184 default: llvm_unreachable("Unknown condition code!");
1185 case ISD::SETNE: return ARMCC::NE;
1186 case ISD::SETEQ: return ARMCC::EQ;
1187 case ISD::SETGT: return ARMCC::GT;
1188 case ISD::SETGE: return ARMCC::GE;
1189 case ISD::SETLT: return ARMCC::LT;
1190 case ISD::SETLE: return ARMCC::LE;
1191 case ISD::SETUGT: return ARMCC::HI;
1192 case ISD::SETUGE: return ARMCC::HS;
1193 case ISD::SETULT: return ARMCC::LO;
1194 case ISD::SETULE: return ARMCC::LS;
1198 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1199 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1200 ARMCC::CondCodes &CondCode2) {
1201 CondCode2 = ARMCC::AL;
1203 default: llvm_unreachable("Unknown FP condition!");
1205 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1207 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1209 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1210 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1211 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1212 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1213 case ISD::SETO: CondCode = ARMCC::VC; break;
1214 case ISD::SETUO: CondCode = ARMCC::VS; break;
1215 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1216 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1217 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1219 case ISD::SETULT: CondCode = ARMCC::LT; break;
1221 case ISD::SETULE: CondCode = ARMCC::LE; break;
1223 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1227 //===----------------------------------------------------------------------===//
1228 // Calling Convention Implementation
1229 //===----------------------------------------------------------------------===//
1231 #include "ARMGenCallingConv.inc"
1233 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1234 /// given CallingConvention value.
1235 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1237 bool isVarArg) const {
1240 llvm_unreachable("Unsupported calling convention");
1241 case CallingConv::Fast:
1242 if (Subtarget->hasVFP2() && !isVarArg) {
1243 if (!Subtarget->isAAPCS_ABI())
1244 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1245 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1246 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1249 case CallingConv::C: {
1250 // Use target triple & subtarget features to do actual dispatch.
1251 if (!Subtarget->isAAPCS_ABI())
1252 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1253 else if (Subtarget->hasVFP2() &&
1254 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1256 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1257 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1259 case CallingConv::ARM_AAPCS_VFP:
1261 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1263 case CallingConv::ARM_AAPCS:
1264 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1265 case CallingConv::ARM_APCS:
1266 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1267 case CallingConv::GHC:
1268 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1272 /// LowerCallResult - Lower the result values of a call into the
1273 /// appropriate copies out of appropriate physical registers.
1275 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1276 CallingConv::ID CallConv, bool isVarArg,
1277 const SmallVectorImpl<ISD::InputArg> &Ins,
1278 SDLoc dl, SelectionDAG &DAG,
1279 SmallVectorImpl<SDValue> &InVals,
1280 bool isThisReturn, SDValue ThisVal) const {
1282 // Assign locations to each value returned by this call.
1283 SmallVector<CCValAssign, 16> RVLocs;
1284 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1285 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1286 CCInfo.AnalyzeCallResult(Ins,
1287 CCAssignFnForNode(CallConv, /* Return*/ true,
1290 // Copy all of the result registers out of their specified physreg.
1291 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1292 CCValAssign VA = RVLocs[i];
1294 // Pass 'this' value directly from the argument to return value, to avoid
1295 // reg unit interference
1296 if (i == 0 && isThisReturn) {
1297 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1298 "unexpected return calling convention register assignment");
1299 InVals.push_back(ThisVal);
1304 if (VA.needsCustom()) {
1305 // Handle f64 or half of a v2f64.
1306 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1308 Chain = Lo.getValue(1);
1309 InFlag = Lo.getValue(2);
1310 VA = RVLocs[++i]; // skip ahead to next loc
1311 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1313 Chain = Hi.getValue(1);
1314 InFlag = Hi.getValue(2);
1315 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1317 if (VA.getLocVT() == MVT::v2f64) {
1318 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1319 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1320 DAG.getConstant(0, MVT::i32));
1322 VA = RVLocs[++i]; // skip ahead to next loc
1323 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1324 Chain = Lo.getValue(1);
1325 InFlag = Lo.getValue(2);
1326 VA = RVLocs[++i]; // skip ahead to next loc
1327 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1328 Chain = Hi.getValue(1);
1329 InFlag = Hi.getValue(2);
1330 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1331 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1332 DAG.getConstant(1, MVT::i32));
1335 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1337 Chain = Val.getValue(1);
1338 InFlag = Val.getValue(2);
1341 switch (VA.getLocInfo()) {
1342 default: llvm_unreachable("Unknown loc info!");
1343 case CCValAssign::Full: break;
1344 case CCValAssign::BCvt:
1345 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1349 InVals.push_back(Val);
1355 /// LowerMemOpCallTo - Store the argument to the stack.
1357 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1358 SDValue StackPtr, SDValue Arg,
1359 SDLoc dl, SelectionDAG &DAG,
1360 const CCValAssign &VA,
1361 ISD::ArgFlagsTy Flags) const {
1362 unsigned LocMemOffset = VA.getLocMemOffset();
1363 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1364 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1365 return DAG.getStore(Chain, dl, Arg, PtrOff,
1366 MachinePointerInfo::getStack(LocMemOffset),
1370 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1371 SDValue Chain, SDValue &Arg,
1372 RegsToPassVector &RegsToPass,
1373 CCValAssign &VA, CCValAssign &NextVA,
1375 SmallVectorImpl<SDValue> &MemOpChains,
1376 ISD::ArgFlagsTy Flags) const {
1378 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1379 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1380 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1382 if (NextVA.isRegLoc())
1383 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1385 assert(NextVA.isMemLoc());
1386 if (StackPtr.getNode() == 0)
1387 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1389 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1395 /// LowerCall - Lowering a call into a callseq_start <-
1396 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1399 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1400 SmallVectorImpl<SDValue> &InVals) const {
1401 SelectionDAG &DAG = CLI.DAG;
1403 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1404 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1405 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1406 SDValue Chain = CLI.Chain;
1407 SDValue Callee = CLI.Callee;
1408 bool &isTailCall = CLI.IsTailCall;
1409 CallingConv::ID CallConv = CLI.CallConv;
1410 bool doesNotRet = CLI.DoesNotReturn;
1411 bool isVarArg = CLI.IsVarArg;
1413 MachineFunction &MF = DAG.getMachineFunction();
1414 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1415 bool isThisReturn = false;
1416 bool isSibCall = false;
1417 // Disable tail calls if they're not supported.
1418 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1421 // Check if it's really possible to do a tail call.
1422 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1423 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1424 Outs, OutVals, Ins, DAG);
1425 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1426 // detected sibcalls.
1433 // Analyze operands of the call, assigning locations to each operand.
1434 SmallVector<CCValAssign, 16> ArgLocs;
1435 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1436 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1437 CCInfo.AnalyzeCallOperands(Outs,
1438 CCAssignFnForNode(CallConv, /* Return*/ false,
1441 // Get a count of how many bytes are to be pushed on the stack.
1442 unsigned NumBytes = CCInfo.getNextStackOffset();
1444 // For tail calls, memory operands are available in our caller's stack.
1448 // Adjust the stack pointer for the new arguments...
1449 // These operations are automatically eliminated by the prolog/epilog pass
1451 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1454 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1456 RegsToPassVector RegsToPass;
1457 SmallVector<SDValue, 8> MemOpChains;
1459 // Walk the register/memloc assignments, inserting copies/loads. In the case
1460 // of tail call optimization, arguments are handled later.
1461 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1463 ++i, ++realArgIdx) {
1464 CCValAssign &VA = ArgLocs[i];
1465 SDValue Arg = OutVals[realArgIdx];
1466 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1467 bool isByVal = Flags.isByVal();
1469 // Promote the value if needed.
1470 switch (VA.getLocInfo()) {
1471 default: llvm_unreachable("Unknown loc info!");
1472 case CCValAssign::Full: break;
1473 case CCValAssign::SExt:
1474 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1476 case CCValAssign::ZExt:
1477 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1479 case CCValAssign::AExt:
1480 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1482 case CCValAssign::BCvt:
1483 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1487 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1488 if (VA.needsCustom()) {
1489 if (VA.getLocVT() == MVT::v2f64) {
1490 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1491 DAG.getConstant(0, MVT::i32));
1492 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1493 DAG.getConstant(1, MVT::i32));
1495 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1496 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1498 VA = ArgLocs[++i]; // skip ahead to next loc
1499 if (VA.isRegLoc()) {
1500 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1501 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1503 assert(VA.isMemLoc());
1505 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1506 dl, DAG, VA, Flags));
1509 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1510 StackPtr, MemOpChains, Flags);
1512 } else if (VA.isRegLoc()) {
1513 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1514 assert(VA.getLocVT() == MVT::i32 &&
1515 "unexpected calling convention register assignment");
1516 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1517 "unexpected use of 'returned'");
1518 isThisReturn = true;
1520 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1521 } else if (isByVal) {
1522 assert(VA.isMemLoc());
1523 unsigned offset = 0;
1525 // True if this byval aggregate will be split between registers
1527 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1528 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1530 if (CurByValIdx < ByValArgsCount) {
1532 unsigned RegBegin, RegEnd;
1533 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1535 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1537 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1538 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1539 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1540 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1541 MachinePointerInfo(),
1542 false, false, false, 0);
1543 MemOpChains.push_back(Load.getValue(1));
1544 RegsToPass.push_back(std::make_pair(j, Load));
1547 // If parameter size outsides register area, "offset" value
1548 // helps us to calculate stack slot for remained part properly.
1549 offset = RegEnd - RegBegin;
1551 CCInfo.nextInRegsParam();
1554 if (Flags.getByValSize() > 4*offset) {
1555 unsigned LocMemOffset = VA.getLocMemOffset();
1556 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1557 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1559 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1560 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1561 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1563 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1565 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1566 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1567 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1568 Ops, array_lengthof(Ops)));
1570 } else if (!isSibCall) {
1571 assert(VA.isMemLoc());
1573 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1574 dl, DAG, VA, Flags));
1578 if (!MemOpChains.empty())
1579 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1580 &MemOpChains[0], MemOpChains.size());
1582 // Build a sequence of copy-to-reg nodes chained together with token chain
1583 // and flag operands which copy the outgoing args into the appropriate regs.
1585 // Tail call byval lowering might overwrite argument registers so in case of
1586 // tail call optimization the copies to registers are lowered later.
1588 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1589 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1590 RegsToPass[i].second, InFlag);
1591 InFlag = Chain.getValue(1);
1594 // For tail calls lower the arguments to the 'real' stack slot.
1596 // Force all the incoming stack arguments to be loaded from the stack
1597 // before any new outgoing arguments are stored to the stack, because the
1598 // outgoing stack slots may alias the incoming argument stack slots, and
1599 // the alias isn't otherwise explicit. This is slightly more conservative
1600 // than necessary, because it means that each store effectively depends
1601 // on every argument instead of just those arguments it would clobber.
1603 // Do not flag preceding copytoreg stuff together with the following stuff.
1605 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1606 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1607 RegsToPass[i].second, InFlag);
1608 InFlag = Chain.getValue(1);
1613 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1614 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1615 // node so that legalize doesn't hack it.
1616 bool isDirect = false;
1617 bool isARMFunc = false;
1618 bool isLocalARMFunc = false;
1619 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1621 if (EnableARMLongCalls) {
1622 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1623 && "long-calls with non-static relocation model!");
1624 // Handle a global address or an external symbol. If it's not one of
1625 // those, the target's already in a register, so we don't need to do
1627 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1628 const GlobalValue *GV = G->getGlobal();
1629 // Create a constant pool entry for the callee address
1630 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1631 ARMConstantPoolValue *CPV =
1632 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1634 // Get the address of the callee into a register
1635 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1636 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1637 Callee = DAG.getLoad(getPointerTy(), dl,
1638 DAG.getEntryNode(), CPAddr,
1639 MachinePointerInfo::getConstantPool(),
1640 false, false, false, 0);
1641 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1642 const char *Sym = S->getSymbol();
1644 // Create a constant pool entry for the callee address
1645 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1646 ARMConstantPoolValue *CPV =
1647 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1648 ARMPCLabelIndex, 0);
1649 // Get the address of the callee into a register
1650 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1651 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1652 Callee = DAG.getLoad(getPointerTy(), dl,
1653 DAG.getEntryNode(), CPAddr,
1654 MachinePointerInfo::getConstantPool(),
1655 false, false, false, 0);
1657 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1658 const GlobalValue *GV = G->getGlobal();
1660 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1661 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1662 getTargetMachine().getRelocationModel() != Reloc::Static;
1663 isARMFunc = !Subtarget->isThumb() || isStub;
1664 // ARM call to a local ARM function is predicable.
1665 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1666 // tBX takes a register source operand.
1667 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1668 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1669 ARMConstantPoolValue *CPV =
1670 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1671 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1672 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1673 Callee = DAG.getLoad(getPointerTy(), dl,
1674 DAG.getEntryNode(), CPAddr,
1675 MachinePointerInfo::getConstantPool(),
1676 false, false, false, 0);
1677 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1678 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1679 getPointerTy(), Callee, PICLabel);
1681 // On ELF targets for PIC code, direct calls should go through the PLT
1682 unsigned OpFlags = 0;
1683 if (Subtarget->isTargetELF() &&
1684 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1685 OpFlags = ARMII::MO_PLT;
1686 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1688 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1690 bool isStub = Subtarget->isTargetDarwin() &&
1691 getTargetMachine().getRelocationModel() != Reloc::Static;
1692 isARMFunc = !Subtarget->isThumb() || isStub;
1693 // tBX takes a register source operand.
1694 const char *Sym = S->getSymbol();
1695 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1696 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1697 ARMConstantPoolValue *CPV =
1698 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1699 ARMPCLabelIndex, 4);
1700 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1701 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1702 Callee = DAG.getLoad(getPointerTy(), dl,
1703 DAG.getEntryNode(), CPAddr,
1704 MachinePointerInfo::getConstantPool(),
1705 false, false, false, 0);
1706 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1707 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1708 getPointerTy(), Callee, PICLabel);
1710 unsigned OpFlags = 0;
1711 // On ELF targets for PIC code, direct calls should go through the PLT
1712 if (Subtarget->isTargetELF() &&
1713 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1714 OpFlags = ARMII::MO_PLT;
1715 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1719 // FIXME: handle tail calls differently.
1721 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1722 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1723 if (Subtarget->isThumb()) {
1724 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1725 CallOpc = ARMISD::CALL_NOLINK;
1727 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1729 if (!isDirect && !Subtarget->hasV5TOps())
1730 CallOpc = ARMISD::CALL_NOLINK;
1731 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1732 // Emit regular call when code size is the priority
1734 // "mov lr, pc; b _foo" to avoid confusing the RSP
1735 CallOpc = ARMISD::CALL_NOLINK;
1737 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1740 std::vector<SDValue> Ops;
1741 Ops.push_back(Chain);
1742 Ops.push_back(Callee);
1744 // Add argument registers to the end of the list so that they are known live
1746 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1747 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1748 RegsToPass[i].second.getValueType()));
1750 // Add a register mask operand representing the call-preserved registers.
1751 const uint32_t *Mask;
1752 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1753 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1755 // For 'this' returns, use the R0-preserving mask if applicable
1756 Mask = ARI->getThisReturnPreservedMask(CallConv);
1758 // Set isThisReturn to false if the calling convention is not one that
1759 // allows 'returned' to be modeled in this way, so LowerCallResult does
1760 // not try to pass 'this' straight through
1761 isThisReturn = false;
1762 Mask = ARI->getCallPreservedMask(CallConv);
1765 Mask = ARI->getCallPreservedMask(CallConv);
1767 assert(Mask && "Missing call preserved mask for calling convention");
1768 Ops.push_back(DAG.getRegisterMask(Mask));
1770 if (InFlag.getNode())
1771 Ops.push_back(InFlag);
1773 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1775 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1777 // Returns a chain and a flag for retval copy to use.
1778 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1779 InFlag = Chain.getValue(1);
1781 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1782 DAG.getIntPtrConstant(0, true), InFlag, dl);
1784 InFlag = Chain.getValue(1);
1786 // Handle result values, copying them out of physregs into vregs that we
1788 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1789 InVals, isThisReturn,
1790 isThisReturn ? OutVals[0] : SDValue());
1793 /// HandleByVal - Every parameter *after* a byval parameter is passed
1794 /// on the stack. Remember the next parameter register to allocate,
1795 /// and then confiscate the rest of the parameter registers to insure
1798 ARMTargetLowering::HandleByVal(
1799 CCState *State, unsigned &size, unsigned Align) const {
1800 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1801 assert((State->getCallOrPrologue() == Prologue ||
1802 State->getCallOrPrologue() == Call) &&
1803 "unhandled ParmContext");
1805 // For in-prologue parameters handling, we also introduce stack offset
1806 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
1807 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
1808 // NSAA should be evaluted (NSAA means "next stacked argument address").
1809 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
1810 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
1811 unsigned NSAAOffset = State->getNextStackOffset();
1812 if (State->getCallOrPrologue() != Call) {
1813 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
1815 State->getInRegsParamInfo(i, RB, RE);
1816 assert(NSAAOffset >= (RE-RB)*4 &&
1817 "Stack offset for byval regs doesn't introduced anymore?");
1818 NSAAOffset -= (RE-RB)*4;
1821 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1822 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1823 unsigned AlignInRegs = Align / 4;
1824 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1825 for (unsigned i = 0; i < Waste; ++i)
1826 reg = State->AllocateReg(GPRArgRegs, 4);
1829 unsigned excess = 4 * (ARM::R4 - reg);
1831 // Special case when NSAA != SP and parameter size greater than size of
1832 // all remained GPR regs. In that case we can't split parameter, we must
1833 // send it to stack. We also must set NCRN to R4, so waste all
1834 // remained registers.
1835 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1836 while (State->AllocateReg(GPRArgRegs, 4))
1841 // First register for byval parameter is the first register that wasn't
1842 // allocated before this method call, so it would be "reg".
1843 // If parameter is small enough to be saved in range [reg, r4), then
1844 // the end (first after last) register would be reg + param-size-in-regs,
1845 // else parameter would be splitted between registers and stack,
1846 // end register would be r4 in this case.
1847 unsigned ByValRegBegin = reg;
1848 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1849 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1850 // Note, first register is allocated in the beginning of function already,
1851 // allocate remained amount of registers we need.
1852 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1853 State->AllocateReg(GPRArgRegs, 4);
1854 // At a call site, a byval parameter that is split between
1855 // registers and memory needs its size truncated here. In a
1856 // function prologue, such byval parameters are reassembled in
1857 // memory, and are not truncated.
1858 if (State->getCallOrPrologue() == Call) {
1859 // Make remained size equal to 0 in case, when
1860 // the whole structure may be stored into registers.
1870 /// MatchingStackOffset - Return true if the given stack call argument is
1871 /// already available in the same position (relatively) of the caller's
1872 /// incoming argument stack.
1874 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1875 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1876 const TargetInstrInfo *TII) {
1877 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1879 if (Arg.getOpcode() == ISD::CopyFromReg) {
1880 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1881 if (!TargetRegisterInfo::isVirtualRegister(VR))
1883 MachineInstr *Def = MRI->getVRegDef(VR);
1886 if (!Flags.isByVal()) {
1887 if (!TII->isLoadFromStackSlot(Def, FI))
1892 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1893 if (Flags.isByVal())
1894 // ByVal argument is passed in as a pointer but it's now being
1895 // dereferenced. e.g.
1896 // define @foo(%struct.X* %A) {
1897 // tail call @bar(%struct.X* byval %A)
1900 SDValue Ptr = Ld->getBasePtr();
1901 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1904 FI = FINode->getIndex();
1908 assert(FI != INT_MAX);
1909 if (!MFI->isFixedObjectIndex(FI))
1911 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1914 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1915 /// for tail call optimization. Targets which want to do tail call
1916 /// optimization should implement this function.
1918 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1919 CallingConv::ID CalleeCC,
1921 bool isCalleeStructRet,
1922 bool isCallerStructRet,
1923 const SmallVectorImpl<ISD::OutputArg> &Outs,
1924 const SmallVectorImpl<SDValue> &OutVals,
1925 const SmallVectorImpl<ISD::InputArg> &Ins,
1926 SelectionDAG& DAG) const {
1927 const Function *CallerF = DAG.getMachineFunction().getFunction();
1928 CallingConv::ID CallerCC = CallerF->getCallingConv();
1929 bool CCMatch = CallerCC == CalleeCC;
1931 // Look for obvious safe cases to perform tail call optimization that do not
1932 // require ABI changes. This is what gcc calls sibcall.
1934 // Do not sibcall optimize vararg calls unless the call site is not passing
1936 if (isVarArg && !Outs.empty())
1939 // Also avoid sibcall optimization if either caller or callee uses struct
1940 // return semantics.
1941 if (isCalleeStructRet || isCallerStructRet)
1944 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1945 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1946 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1947 // support in the assembler and linker to be used. This would need to be
1948 // fixed to fully support tail calls in Thumb1.
1950 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1951 // LR. This means if we need to reload LR, it takes an extra instructions,
1952 // which outweighs the value of the tail call; but here we don't know yet
1953 // whether LR is going to be used. Probably the right approach is to
1954 // generate the tail call here and turn it back into CALL/RET in
1955 // emitEpilogue if LR is used.
1957 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1958 // but we need to make sure there are enough registers; the only valid
1959 // registers are the 4 used for parameters. We don't currently do this
1961 if (Subtarget->isThumb1Only())
1964 // If the calling conventions do not match, then we'd better make sure the
1965 // results are returned in the same way as what the caller expects.
1967 SmallVector<CCValAssign, 16> RVLocs1;
1968 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1969 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1970 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1972 SmallVector<CCValAssign, 16> RVLocs2;
1973 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1974 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1975 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1977 if (RVLocs1.size() != RVLocs2.size())
1979 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1980 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1982 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1984 if (RVLocs1[i].isRegLoc()) {
1985 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1988 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1994 // If Caller's vararg or byval argument has been split between registers and
1995 // stack, do not perform tail call, since part of the argument is in caller's
1997 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1998 getInfo<ARMFunctionInfo>();
1999 if (AFI_Caller->getArgRegsSaveSize())
2002 // If the callee takes no arguments then go on to check the results of the
2004 if (!Outs.empty()) {
2005 // Check if stack adjustment is needed. For now, do not do this if any
2006 // argument is passed on the stack.
2007 SmallVector<CCValAssign, 16> ArgLocs;
2008 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2009 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2010 CCInfo.AnalyzeCallOperands(Outs,
2011 CCAssignFnForNode(CalleeCC, false, isVarArg));
2012 if (CCInfo.getNextStackOffset()) {
2013 MachineFunction &MF = DAG.getMachineFunction();
2015 // Check if the arguments are already laid out in the right way as
2016 // the caller's fixed stack objects.
2017 MachineFrameInfo *MFI = MF.getFrameInfo();
2018 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2019 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2020 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2022 ++i, ++realArgIdx) {
2023 CCValAssign &VA = ArgLocs[i];
2024 EVT RegVT = VA.getLocVT();
2025 SDValue Arg = OutVals[realArgIdx];
2026 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2027 if (VA.getLocInfo() == CCValAssign::Indirect)
2029 if (VA.needsCustom()) {
2030 // f64 and vector types are split into multiple registers or
2031 // register/stack-slot combinations. The types will not match
2032 // the registers; give up on memory f64 refs until we figure
2033 // out what to do about this.
2036 if (!ArgLocs[++i].isRegLoc())
2038 if (RegVT == MVT::v2f64) {
2039 if (!ArgLocs[++i].isRegLoc())
2041 if (!ArgLocs[++i].isRegLoc())
2044 } else if (!VA.isRegLoc()) {
2045 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2057 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2058 MachineFunction &MF, bool isVarArg,
2059 const SmallVectorImpl<ISD::OutputArg> &Outs,
2060 LLVMContext &Context) const {
2061 SmallVector<CCValAssign, 16> RVLocs;
2062 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2063 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2068 ARMTargetLowering::LowerReturn(SDValue Chain,
2069 CallingConv::ID CallConv, bool isVarArg,
2070 const SmallVectorImpl<ISD::OutputArg> &Outs,
2071 const SmallVectorImpl<SDValue> &OutVals,
2072 SDLoc dl, SelectionDAG &DAG) const {
2074 // CCValAssign - represent the assignment of the return value to a location.
2075 SmallVector<CCValAssign, 16> RVLocs;
2077 // CCState - Info about the registers and stack slots.
2078 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2079 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2081 // Analyze outgoing return values.
2082 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2086 SmallVector<SDValue, 4> RetOps;
2087 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2089 // Copy the result values into the output registers.
2090 for (unsigned i = 0, realRVLocIdx = 0;
2092 ++i, ++realRVLocIdx) {
2093 CCValAssign &VA = RVLocs[i];
2094 assert(VA.isRegLoc() && "Can only return in registers!");
2096 SDValue Arg = OutVals[realRVLocIdx];
2098 switch (VA.getLocInfo()) {
2099 default: llvm_unreachable("Unknown loc info!");
2100 case CCValAssign::Full: break;
2101 case CCValAssign::BCvt:
2102 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2106 if (VA.needsCustom()) {
2107 if (VA.getLocVT() == MVT::v2f64) {
2108 // Extract the first half and return it in two registers.
2109 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2110 DAG.getConstant(0, MVT::i32));
2111 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2112 DAG.getVTList(MVT::i32, MVT::i32), Half);
2114 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2115 Flag = Chain.getValue(1);
2116 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2117 VA = RVLocs[++i]; // skip ahead to next loc
2118 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2119 HalfGPRs.getValue(1), Flag);
2120 Flag = Chain.getValue(1);
2121 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2122 VA = RVLocs[++i]; // skip ahead to next loc
2124 // Extract the 2nd half and fall through to handle it as an f64 value.
2125 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2126 DAG.getConstant(1, MVT::i32));
2128 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2130 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2131 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2132 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2133 Flag = Chain.getValue(1);
2134 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2135 VA = RVLocs[++i]; // skip ahead to next loc
2136 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2139 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2141 // Guarantee that all emitted copies are
2142 // stuck together, avoiding something bad.
2143 Flag = Chain.getValue(1);
2144 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2147 // Update chain and glue.
2150 RetOps.push_back(Flag);
2152 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2153 RetOps.data(), RetOps.size());
2156 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2157 if (N->getNumValues() != 1)
2159 if (!N->hasNUsesOfValue(1, 0))
2162 SDValue TCChain = Chain;
2163 SDNode *Copy = *N->use_begin();
2164 if (Copy->getOpcode() == ISD::CopyToReg) {
2165 // If the copy has a glue operand, we conservatively assume it isn't safe to
2166 // perform a tail call.
2167 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2169 TCChain = Copy->getOperand(0);
2170 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2171 SDNode *VMov = Copy;
2172 // f64 returned in a pair of GPRs.
2173 SmallPtrSet<SDNode*, 2> Copies;
2174 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2176 if (UI->getOpcode() != ISD::CopyToReg)
2180 if (Copies.size() > 2)
2183 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2185 SDValue UseChain = UI->getOperand(0);
2186 if (Copies.count(UseChain.getNode()))
2193 } else if (Copy->getOpcode() == ISD::BITCAST) {
2194 // f32 returned in a single GPR.
2195 if (!Copy->hasOneUse())
2197 Copy = *Copy->use_begin();
2198 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2200 TCChain = Copy->getOperand(0);
2205 bool HasRet = false;
2206 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2208 if (UI->getOpcode() != ARMISD::RET_FLAG)
2220 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2221 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2224 if (!CI->isTailCall())
2227 return !Subtarget->isThumb1Only();
2230 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2231 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2232 // one of the above mentioned nodes. It has to be wrapped because otherwise
2233 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2234 // be used to form addressing mode. These wrapped nodes will be selected
2236 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2237 EVT PtrVT = Op.getValueType();
2238 // FIXME there is no actual debug info here
2240 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2242 if (CP->isMachineConstantPoolEntry())
2243 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2244 CP->getAlignment());
2246 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2247 CP->getAlignment());
2248 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2251 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2252 return MachineJumpTableInfo::EK_Inline;
2255 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2256 SelectionDAG &DAG) const {
2257 MachineFunction &MF = DAG.getMachineFunction();
2258 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2259 unsigned ARMPCLabelIndex = 0;
2261 EVT PtrVT = getPointerTy();
2262 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2263 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2265 if (RelocM == Reloc::Static) {
2266 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2268 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2269 ARMPCLabelIndex = AFI->createPICLabelUId();
2270 ARMConstantPoolValue *CPV =
2271 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2272 ARMCP::CPBlockAddress, PCAdj);
2273 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2275 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2276 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2277 MachinePointerInfo::getConstantPool(),
2278 false, false, false, 0);
2279 if (RelocM == Reloc::Static)
2281 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2282 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2285 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2287 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2288 SelectionDAG &DAG) const {
2290 EVT PtrVT = getPointerTy();
2291 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2292 MachineFunction &MF = DAG.getMachineFunction();
2293 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2294 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2295 ARMConstantPoolValue *CPV =
2296 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2297 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2298 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2299 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2300 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2301 MachinePointerInfo::getConstantPool(),
2302 false, false, false, 0);
2303 SDValue Chain = Argument.getValue(1);
2305 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2306 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2308 // call __tls_get_addr.
2311 Entry.Node = Argument;
2312 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2313 Args.push_back(Entry);
2314 // FIXME: is there useful debug info available here?
2315 TargetLowering::CallLoweringInfo CLI(Chain,
2316 (Type *) Type::getInt32Ty(*DAG.getContext()),
2317 false, false, false, false,
2318 0, CallingConv::C, /*isTailCall=*/false,
2319 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2320 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2321 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2322 return CallResult.first;
2325 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2326 // "local exec" model.
2328 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2330 TLSModel::Model model) const {
2331 const GlobalValue *GV = GA->getGlobal();
2334 SDValue Chain = DAG.getEntryNode();
2335 EVT PtrVT = getPointerTy();
2336 // Get the Thread Pointer
2337 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2339 if (model == TLSModel::InitialExec) {
2340 MachineFunction &MF = DAG.getMachineFunction();
2341 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2342 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2343 // Initial exec model.
2344 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2345 ARMConstantPoolValue *CPV =
2346 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2347 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2349 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2350 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2351 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2352 MachinePointerInfo::getConstantPool(),
2353 false, false, false, 0);
2354 Chain = Offset.getValue(1);
2356 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2357 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2359 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2360 MachinePointerInfo::getConstantPool(),
2361 false, false, false, 0);
2364 assert(model == TLSModel::LocalExec);
2365 ARMConstantPoolValue *CPV =
2366 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2367 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2368 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2369 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2370 MachinePointerInfo::getConstantPool(),
2371 false, false, false, 0);
2374 // The address of the thread local variable is the add of the thread
2375 // pointer with the offset of the variable.
2376 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2380 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2381 // TODO: implement the "local dynamic" model
2382 assert(Subtarget->isTargetELF() &&
2383 "TLS not implemented for non-ELF targets");
2384 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2386 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2389 case TLSModel::GeneralDynamic:
2390 case TLSModel::LocalDynamic:
2391 return LowerToTLSGeneralDynamicModel(GA, DAG);
2392 case TLSModel::InitialExec:
2393 case TLSModel::LocalExec:
2394 return LowerToTLSExecModels(GA, DAG, model);
2396 llvm_unreachable("bogus TLS model");
2399 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2400 SelectionDAG &DAG) const {
2401 EVT PtrVT = getPointerTy();
2403 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2404 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2405 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2406 ARMConstantPoolValue *CPV =
2407 ARMConstantPoolConstant::Create(GV,
2408 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2409 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2410 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2411 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2413 MachinePointerInfo::getConstantPool(),
2414 false, false, false, 0);
2415 SDValue Chain = Result.getValue(1);
2416 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2417 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2419 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2420 MachinePointerInfo::getGOT(),
2421 false, false, false, 0);
2425 // If we have T2 ops, we can materialize the address directly via movt/movw
2426 // pair. This is always cheaper.
2427 if (Subtarget->useMovt()) {
2429 // FIXME: Once remat is capable of dealing with instructions with register
2430 // operands, expand this into two nodes.
2431 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2432 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2434 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2435 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2436 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2437 MachinePointerInfo::getConstantPool(),
2438 false, false, false, 0);
2442 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2443 SelectionDAG &DAG) const {
2444 EVT PtrVT = getPointerTy();
2446 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2447 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2449 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2450 // update ARMFastISel::ARMMaterializeGV.
2451 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2453 // FIXME: Once remat is capable of dealing with instructions with register
2454 // operands, expand this into two nodes.
2455 if (RelocM == Reloc::Static)
2456 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2457 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2459 unsigned Wrapper = (RelocM == Reloc::PIC_)
2460 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2461 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2462 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2463 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2464 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2465 MachinePointerInfo::getGOT(),
2466 false, false, false, 0);
2470 unsigned ARMPCLabelIndex = 0;
2472 if (RelocM == Reloc::Static) {
2473 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2475 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2476 ARMPCLabelIndex = AFI->createPICLabelUId();
2477 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2478 ARMConstantPoolValue *CPV =
2479 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2481 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2483 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2485 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2486 MachinePointerInfo::getConstantPool(),
2487 false, false, false, 0);
2488 SDValue Chain = Result.getValue(1);
2490 if (RelocM == Reloc::PIC_) {
2491 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2492 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2495 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2496 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2497 false, false, false, 0);
2502 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2503 SelectionDAG &DAG) const {
2504 assert(Subtarget->isTargetELF() &&
2505 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2506 MachineFunction &MF = DAG.getMachineFunction();
2507 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2508 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2509 EVT PtrVT = getPointerTy();
2511 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2512 ARMConstantPoolValue *CPV =
2513 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2514 ARMPCLabelIndex, PCAdj);
2515 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2516 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2517 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2518 MachinePointerInfo::getConstantPool(),
2519 false, false, false, 0);
2520 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2521 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2525 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2527 SDValue Val = DAG.getConstant(0, MVT::i32);
2528 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2529 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2530 Op.getOperand(1), Val);
2534 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2536 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2537 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2541 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2542 const ARMSubtarget *Subtarget) const {
2543 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2546 default: return SDValue(); // Don't custom lower most intrinsics.
2547 case Intrinsic::arm_thread_pointer: {
2548 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2549 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2551 case Intrinsic::eh_sjlj_lsda: {
2552 MachineFunction &MF = DAG.getMachineFunction();
2553 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2554 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2555 EVT PtrVT = getPointerTy();
2556 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2558 unsigned PCAdj = (RelocM != Reloc::PIC_)
2559 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2560 ARMConstantPoolValue *CPV =
2561 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2562 ARMCP::CPLSDA, PCAdj);
2563 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2564 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2566 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2567 MachinePointerInfo::getConstantPool(),
2568 false, false, false, 0);
2570 if (RelocM == Reloc::PIC_) {
2571 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2572 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2576 case Intrinsic::arm_neon_vmulls:
2577 case Intrinsic::arm_neon_vmullu: {
2578 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2579 ? ARMISD::VMULLs : ARMISD::VMULLu;
2580 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2581 Op.getOperand(1), Op.getOperand(2));
2586 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2587 const ARMSubtarget *Subtarget) {
2588 // FIXME: handle "fence singlethread" more efficiently.
2590 if (!Subtarget->hasDataBarrier()) {
2591 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2592 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2594 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2595 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2596 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2597 DAG.getConstant(0, MVT::i32));
2600 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2601 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2602 unsigned Domain = ARM_MB::ISH;
2603 if (Subtarget->isSwift() && Ord == Release) {
2604 // Swift happens to implement ISHST barriers in a way that's compatible with
2605 // Release semantics but weaker than ISH so we'd be fools not to use
2606 // it. Beware: other processors probably don't!
2607 Domain = ARM_MB::ISHST;
2610 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2611 DAG.getConstant(Domain, MVT::i32));
2614 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2615 const ARMSubtarget *Subtarget) {
2616 // ARM pre v5TE and Thumb1 does not have preload instructions.
2617 if (!(Subtarget->isThumb2() ||
2618 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2619 // Just preserve the chain.
2620 return Op.getOperand(0);
2623 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2625 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2626 // ARMv7 with MP extension has PLDW.
2627 return Op.getOperand(0);
2629 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2630 if (Subtarget->isThumb()) {
2632 isRead = ~isRead & 1;
2633 isData = ~isData & 1;
2636 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2637 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2638 DAG.getConstant(isData, MVT::i32));
2641 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2642 MachineFunction &MF = DAG.getMachineFunction();
2643 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2645 // vastart just stores the address of the VarArgsFrameIndex slot into the
2646 // memory location argument.
2648 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2649 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2650 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2651 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2652 MachinePointerInfo(SV), false, false, 0);
2656 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2657 SDValue &Root, SelectionDAG &DAG,
2659 MachineFunction &MF = DAG.getMachineFunction();
2660 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2662 const TargetRegisterClass *RC;
2663 if (AFI->isThumb1OnlyFunction())
2664 RC = &ARM::tGPRRegClass;
2666 RC = &ARM::GPRRegClass;
2668 // Transform the arguments stored in physical registers into virtual ones.
2669 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2670 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2673 if (NextVA.isMemLoc()) {
2674 MachineFrameInfo *MFI = MF.getFrameInfo();
2675 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2677 // Create load node to retrieve arguments from the stack.
2678 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2679 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2680 MachinePointerInfo::getFixedStack(FI),
2681 false, false, false, 0);
2683 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2684 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2687 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2691 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2692 unsigned InRegsParamRecordIdx,
2694 unsigned &ArgRegsSize,
2695 unsigned &ArgRegsSaveSize)
2698 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2699 unsigned RBegin, REnd;
2700 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2701 NumGPRs = REnd - RBegin;
2703 unsigned int firstUnalloced;
2704 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2705 sizeof(GPRArgRegs) /
2706 sizeof(GPRArgRegs[0]));
2707 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2710 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2711 ArgRegsSize = NumGPRs * 4;
2713 // If parameter is split between stack and GPRs...
2714 if (NumGPRs && Align == 8 &&
2715 (ArgRegsSize < ArgSize ||
2716 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2717 // Add padding for part of param recovered from GPRs, so
2718 // its last byte must be at address K*8 - 1.
2719 // We need to do it, since remained (stack) part of parameter has
2720 // stack alignment, and we need to "attach" "GPRs head" without gaps
2723 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2724 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2726 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2728 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
2729 (ArgRegsSize + AFI->getArgRegsSaveSize());
2730 ArgRegsSaveSize = ArgRegsSize + Padding;
2732 // We don't need to extend regs save size for byval parameters if they
2733 // are passed via GPRs only.
2734 ArgRegsSaveSize = ArgRegsSize;
2737 // The remaining GPRs hold either the beginning of variable-argument
2738 // data, or the beginning of an aggregate passed by value (usually
2739 // byval). Either way, we allocate stack slots adjacent to the data
2740 // provided by our caller, and store the unallocated registers there.
2741 // If this is a variadic function, the va_list pointer will begin with
2742 // these values; otherwise, this reassembles a (byval) structure that
2743 // was split between registers and memory.
2744 // Return: The frame index registers were stored into.
2746 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2747 SDLoc dl, SDValue &Chain,
2748 const Value *OrigArg,
2749 unsigned InRegsParamRecordIdx,
2750 unsigned OffsetFromOrigArg,
2753 bool ForceMutable) const {
2755 // Currently, two use-cases possible:
2756 // Case #1. Non var-args function, and we meet first byval parameter.
2757 // Setup first unallocated register as first byval register;
2758 // eat all remained registers
2759 // (these two actions are performed by HandleByVal method).
2760 // Then, here, we initialize stack frame with
2761 // "store-reg" instructions.
2762 // Case #2. Var-args function, that doesn't contain byval parameters.
2763 // The same: eat all remained unallocated registers,
2764 // initialize stack frame.
2766 MachineFunction &MF = DAG.getMachineFunction();
2767 MachineFrameInfo *MFI = MF.getFrameInfo();
2768 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2769 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2770 unsigned RBegin, REnd;
2771 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2772 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2773 firstRegToSaveIndex = RBegin - ARM::R0;
2774 lastRegToSaveIndex = REnd - ARM::R0;
2776 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2777 (GPRArgRegs, array_lengthof(GPRArgRegs));
2778 lastRegToSaveIndex = 4;
2781 unsigned ArgRegsSize, ArgRegsSaveSize;
2782 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2783 ArgRegsSize, ArgRegsSaveSize);
2785 // Store any by-val regs to their spots on the stack so that they may be
2786 // loaded by deferencing the result of formal parameter pointer or va_next.
2787 // Note: once stack area for byval/varargs registers
2788 // was initialized, it can't be initialized again.
2789 if (ArgRegsSaveSize) {
2791 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2794 assert(AFI->getStoredByValParamsPadding() == 0 &&
2795 "The only parameter may be padded.");
2796 AFI->setStoredByValParamsPadding(Padding);
2799 int FrameIndex = MFI->CreateFixedObject(
2801 Padding + ArgOffset,
2803 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2805 SmallVector<SDValue, 4> MemOps;
2806 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2807 ++firstRegToSaveIndex, ++i) {
2808 const TargetRegisterClass *RC;
2809 if (AFI->isThumb1OnlyFunction())
2810 RC = &ARM::tGPRRegClass;
2812 RC = &ARM::GPRRegClass;
2814 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2815 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2817 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2818 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2820 MemOps.push_back(Store);
2821 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2822 DAG.getConstant(4, getPointerTy()));
2825 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2827 if (!MemOps.empty())
2828 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2829 &MemOps[0], MemOps.size());
2832 // This will point to the next argument passed via stack.
2833 return MFI->CreateFixedObject(
2834 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
2837 // Setup stack frame, the va_list pointer will start from.
2839 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2840 SDLoc dl, SDValue &Chain,
2842 bool ForceMutable) const {
2843 MachineFunction &MF = DAG.getMachineFunction();
2844 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2846 // Try to store any remaining integer argument regs
2847 // to their spots on the stack so that they may be loaded by deferencing
2848 // the result of va_next.
2849 // If there is no regs to be stored, just point address after last
2850 // argument passed via stack.
2852 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2853 0, ArgOffset, 0, ForceMutable);
2855 AFI->setVarArgsFrameIndex(FrameIndex);
2859 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2860 CallingConv::ID CallConv, bool isVarArg,
2861 const SmallVectorImpl<ISD::InputArg>
2863 SDLoc dl, SelectionDAG &DAG,
2864 SmallVectorImpl<SDValue> &InVals)
2866 MachineFunction &MF = DAG.getMachineFunction();
2867 MachineFrameInfo *MFI = MF.getFrameInfo();
2869 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2871 // Assign locations to all of the incoming arguments.
2872 SmallVector<CCValAssign, 16> ArgLocs;
2873 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2874 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2875 CCInfo.AnalyzeFormalArguments(Ins,
2876 CCAssignFnForNode(CallConv, /* Return*/ false,
2879 SmallVector<SDValue, 16> ArgValues;
2880 int lastInsIndex = -1;
2882 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2883 unsigned CurArgIdx = 0;
2885 // Initially ArgRegsSaveSize is zero.
2886 // Then we increase this value each time we meet byval parameter.
2887 // We also increase this value in case of varargs function.
2888 AFI->setArgRegsSaveSize(0);
2890 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2891 CCValAssign &VA = ArgLocs[i];
2892 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2893 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2894 // Arguments stored in registers.
2895 if (VA.isRegLoc()) {
2896 EVT RegVT = VA.getLocVT();
2898 if (VA.needsCustom()) {
2899 // f64 and vector types are split up into multiple registers or
2900 // combinations of registers and stack slots.
2901 if (VA.getLocVT() == MVT::v2f64) {
2902 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2904 VA = ArgLocs[++i]; // skip ahead to next loc
2906 if (VA.isMemLoc()) {
2907 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2908 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2909 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2910 MachinePointerInfo::getFixedStack(FI),
2911 false, false, false, 0);
2913 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2916 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2917 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2918 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2919 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2920 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2922 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2925 const TargetRegisterClass *RC;
2927 if (RegVT == MVT::f32)
2928 RC = &ARM::SPRRegClass;
2929 else if (RegVT == MVT::f64)
2930 RC = &ARM::DPRRegClass;
2931 else if (RegVT == MVT::v2f64)
2932 RC = &ARM::QPRRegClass;
2933 else if (RegVT == MVT::i32)
2934 RC = AFI->isThumb1OnlyFunction() ?
2935 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2936 (const TargetRegisterClass*)&ARM::GPRRegClass;
2938 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2940 // Transform the arguments in physical registers into virtual ones.
2941 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2942 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2945 // If this is an 8 or 16-bit value, it is really passed promoted
2946 // to 32 bits. Insert an assert[sz]ext to capture this, then
2947 // truncate to the right size.
2948 switch (VA.getLocInfo()) {
2949 default: llvm_unreachable("Unknown loc info!");
2950 case CCValAssign::Full: break;
2951 case CCValAssign::BCvt:
2952 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2954 case CCValAssign::SExt:
2955 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2956 DAG.getValueType(VA.getValVT()));
2957 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2959 case CCValAssign::ZExt:
2960 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2961 DAG.getValueType(VA.getValVT()));
2962 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2966 InVals.push_back(ArgValue);
2968 } else { // VA.isRegLoc()
2971 assert(VA.isMemLoc());
2972 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2974 int index = ArgLocs[i].getValNo();
2976 // Some Ins[] entries become multiple ArgLoc[] entries.
2977 // Process them only once.
2978 if (index != lastInsIndex)
2980 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2981 // FIXME: For now, all byval parameter objects are marked mutable.
2982 // This can be changed with more analysis.
2983 // In case of tail call optimization mark all arguments mutable.
2984 // Since they could be overwritten by lowering of arguments in case of
2986 if (Flags.isByVal()) {
2987 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
2988 int FrameIndex = StoreByValRegs(
2989 CCInfo, DAG, dl, Chain, CurOrigArg,
2991 Ins[VA.getValNo()].PartOffset,
2992 VA.getLocMemOffset(),
2993 Flags.getByValSize(),
2994 true /*force mutable frames*/);
2995 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
2996 CCInfo.nextInRegsParam();
2998 unsigned FIOffset = VA.getLocMemOffset() +
2999 AFI->getStoredByValParamsPadding();
3000 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3003 // Create load nodes to retrieve arguments from the stack.
3004 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3005 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3006 MachinePointerInfo::getFixedStack(FI),
3007 false, false, false, 0));
3009 lastInsIndex = index;
3016 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3017 CCInfo.getNextStackOffset());
3022 /// isFloatingPointZero - Return true if this is +0.0.
3023 static bool isFloatingPointZero(SDValue Op) {
3024 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3025 return CFP->getValueAPF().isPosZero();
3026 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3027 // Maybe this has already been legalized into the constant pool?
3028 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3029 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3030 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3031 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3032 return CFP->getValueAPF().isPosZero();
3038 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3039 /// the given operands.
3041 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3042 SDValue &ARMcc, SelectionDAG &DAG,
3044 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3045 unsigned C = RHSC->getZExtValue();
3046 if (!isLegalICmpImmediate(C)) {
3047 // Constant does not fit, try adjusting it by one?
3052 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3053 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3054 RHS = DAG.getConstant(C-1, MVT::i32);
3059 if (C != 0 && isLegalICmpImmediate(C-1)) {
3060 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3061 RHS = DAG.getConstant(C-1, MVT::i32);
3066 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3067 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3068 RHS = DAG.getConstant(C+1, MVT::i32);
3073 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3074 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3075 RHS = DAG.getConstant(C+1, MVT::i32);
3082 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3083 ARMISD::NodeType CompareType;
3086 CompareType = ARMISD::CMP;
3091 CompareType = ARMISD::CMPZ;
3094 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3095 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3098 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3100 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3103 if (!isFloatingPointZero(RHS))
3104 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3106 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3107 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3110 /// duplicateCmp - Glue values can have only one use, so this function
3111 /// duplicates a comparison node.
3113 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3114 unsigned Opc = Cmp.getOpcode();
3116 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3117 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3119 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3120 Cmp = Cmp.getOperand(0);
3121 Opc = Cmp.getOpcode();
3122 if (Opc == ARMISD::CMPFP)
3123 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3125 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3126 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3128 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3131 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3132 SDValue Cond = Op.getOperand(0);
3133 SDValue SelectTrue = Op.getOperand(1);
3134 SDValue SelectFalse = Op.getOperand(2);
3139 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3140 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3142 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3143 const ConstantSDNode *CMOVTrue =
3144 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3145 const ConstantSDNode *CMOVFalse =
3146 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3148 if (CMOVTrue && CMOVFalse) {
3149 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3150 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3154 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3156 False = SelectFalse;
3157 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3162 if (True.getNode() && False.getNode()) {
3163 EVT VT = Op.getValueType();
3164 SDValue ARMcc = Cond.getOperand(2);
3165 SDValue CCR = Cond.getOperand(3);
3166 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3167 assert(True.getValueType() == VT);
3168 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3173 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3174 // undefined bits before doing a full-word comparison with zero.
3175 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3176 DAG.getConstant(1, Cond.getValueType()));
3178 return DAG.getSelectCC(dl, Cond,
3179 DAG.getConstant(0, Cond.getValueType()),
3180 SelectTrue, SelectFalse, ISD::SETNE);
3183 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3184 if (CC == ISD::SETNE)
3186 return ISD::getSetCCSwappedOperands(CC);
3189 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3190 bool &swpCmpOps, bool &swpVselOps) {
3191 // Start by selecting the GE condition code for opcodes that return true for
3193 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3195 CondCode = ARMCC::GE;
3197 // and GT for opcodes that return false for 'equality'.
3198 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3200 CondCode = ARMCC::GT;
3202 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3203 // to swap the compare operands.
3204 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3208 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3209 // If we have an unordered opcode, we need to swap the operands to the VSEL
3210 // instruction (effectively negating the condition).
3212 // This also has the effect of swapping which one of 'less' or 'greater'
3213 // returns true, so we also swap the compare operands. It also switches
3214 // whether we return true for 'equality', so we compensate by picking the
3215 // opposite condition code to our original choice.
3216 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3217 CC == ISD::SETUGT) {
3218 swpCmpOps = !swpCmpOps;
3219 swpVselOps = !swpVselOps;
3220 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3223 // 'ordered' is 'anything but unordered', so use the VS condition code and
3224 // swap the VSEL operands.
3225 if (CC == ISD::SETO) {
3226 CondCode = ARMCC::VS;
3230 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3231 // code and swap the VSEL operands.
3232 if (CC == ISD::SETUNE) {
3233 CondCode = ARMCC::EQ;
3238 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3239 EVT VT = Op.getValueType();
3240 SDValue LHS = Op.getOperand(0);
3241 SDValue RHS = Op.getOperand(1);
3242 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3243 SDValue TrueVal = Op.getOperand(2);
3244 SDValue FalseVal = Op.getOperand(3);
3247 if (LHS.getValueType() == MVT::i32) {
3248 // Try to generate VSEL on ARMv8.
3249 // The VSEL instruction can't use all the usual ARM condition
3250 // codes: it only has two bits to select the condition code, so it's
3251 // constrained to use only GE, GT, VS and EQ.
3253 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3254 // swap the operands of the previous compare instruction (effectively
3255 // inverting the compare condition, swapping 'less' and 'greater') and
3256 // sometimes need to swap the operands to the VSEL (which inverts the
3257 // condition in the sense of firing whenever the previous condition didn't)
3258 if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
3259 TrueVal.getValueType() == MVT::f64)) {
3260 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3261 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3262 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3263 CC = getInverseCCForVSEL(CC);
3264 std::swap(TrueVal, FalseVal);
3269 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3270 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3271 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3275 ARMCC::CondCodes CondCode, CondCode2;
3276 FPCCToARMCC(CC, CondCode, CondCode2);
3278 // Try to generate VSEL on ARMv8.
3279 if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
3280 TrueVal.getValueType() == MVT::f64)) {
3281 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3282 // same operands, as follows:
3283 // c = fcmp [ogt, olt, ugt, ult] a, b
3285 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3286 // handled differently than the original code sequence.
3287 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3289 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3290 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3291 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3292 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3295 bool swpCmpOps = false;
3296 bool swpVselOps = false;
3297 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3299 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3300 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3302 std::swap(LHS, RHS);
3304 std::swap(TrueVal, FalseVal);
3308 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3309 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3310 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3311 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3313 if (CondCode2 != ARMCC::AL) {
3314 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3315 // FIXME: Needs another CMP because flag can have but one use.
3316 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3317 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3318 Result, TrueVal, ARMcc2, CCR, Cmp2);
3323 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3324 /// to morph to an integer compare sequence.
3325 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3326 const ARMSubtarget *Subtarget) {
3327 SDNode *N = Op.getNode();
3328 if (!N->hasOneUse())
3329 // Otherwise it requires moving the value from fp to integer registers.
3331 if (!N->getNumValues())
3333 EVT VT = Op.getValueType();
3334 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3335 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3336 // vmrs are very slow, e.g. cortex-a8.
3339 if (isFloatingPointZero(Op)) {
3343 return ISD::isNormalLoad(N);
3346 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3347 if (isFloatingPointZero(Op))
3348 return DAG.getConstant(0, MVT::i32);
3350 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3351 return DAG.getLoad(MVT::i32, SDLoc(Op),
3352 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3353 Ld->isVolatile(), Ld->isNonTemporal(),
3354 Ld->isInvariant(), Ld->getAlignment());
3356 llvm_unreachable("Unknown VFP cmp argument!");
3359 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3360 SDValue &RetVal1, SDValue &RetVal2) {
3361 if (isFloatingPointZero(Op)) {
3362 RetVal1 = DAG.getConstant(0, MVT::i32);
3363 RetVal2 = DAG.getConstant(0, MVT::i32);
3367 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3368 SDValue Ptr = Ld->getBasePtr();
3369 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3370 Ld->getChain(), Ptr,
3371 Ld->getPointerInfo(),
3372 Ld->isVolatile(), Ld->isNonTemporal(),
3373 Ld->isInvariant(), Ld->getAlignment());
3375 EVT PtrType = Ptr.getValueType();
3376 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3377 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3378 PtrType, Ptr, DAG.getConstant(4, PtrType));
3379 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3380 Ld->getChain(), NewPtr,
3381 Ld->getPointerInfo().getWithOffset(4),
3382 Ld->isVolatile(), Ld->isNonTemporal(),
3383 Ld->isInvariant(), NewAlign);
3387 llvm_unreachable("Unknown VFP cmp argument!");
3390 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3391 /// f32 and even f64 comparisons to integer ones.
3393 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3394 SDValue Chain = Op.getOperand(0);
3395 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3396 SDValue LHS = Op.getOperand(2);
3397 SDValue RHS = Op.getOperand(3);
3398 SDValue Dest = Op.getOperand(4);
3401 bool LHSSeenZero = false;
3402 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3403 bool RHSSeenZero = false;
3404 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3405 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3406 // If unsafe fp math optimization is enabled and there are no other uses of
3407 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3408 // to an integer comparison.
3409 if (CC == ISD::SETOEQ)
3411 else if (CC == ISD::SETUNE)
3414 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3416 if (LHS.getValueType() == MVT::f32) {
3417 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3418 bitcastf32Toi32(LHS, DAG), Mask);
3419 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3420 bitcastf32Toi32(RHS, DAG), Mask);
3421 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3422 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3423 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3424 Chain, Dest, ARMcc, CCR, Cmp);
3429 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3430 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3431 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3432 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3433 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3434 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3435 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3436 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3437 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3443 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3444 SDValue Chain = Op.getOperand(0);
3445 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3446 SDValue LHS = Op.getOperand(2);
3447 SDValue RHS = Op.getOperand(3);
3448 SDValue Dest = Op.getOperand(4);
3451 if (LHS.getValueType() == MVT::i32) {
3453 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3454 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3455 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3456 Chain, Dest, ARMcc, CCR, Cmp);
3459 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3461 if (getTargetMachine().Options.UnsafeFPMath &&
3462 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3463 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3464 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3465 if (Result.getNode())
3469 ARMCC::CondCodes CondCode, CondCode2;
3470 FPCCToARMCC(CC, CondCode, CondCode2);
3472 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3473 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3474 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3475 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3476 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3477 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3478 if (CondCode2 != ARMCC::AL) {
3479 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3480 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3481 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3486 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3487 SDValue Chain = Op.getOperand(0);
3488 SDValue Table = Op.getOperand(1);
3489 SDValue Index = Op.getOperand(2);
3492 EVT PTy = getPointerTy();
3493 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3494 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3495 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3496 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3497 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3498 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3499 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3500 if (Subtarget->isThumb2()) {
3501 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3502 // which does another jump to the destination. This also makes it easier
3503 // to translate it to TBB / TBH later.
3504 // FIXME: This might not work if the function is extremely large.
3505 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3506 Addr, Op.getOperand(2), JTI, UId);
3508 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3509 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3510 MachinePointerInfo::getJumpTable(),
3511 false, false, false, 0);
3512 Chain = Addr.getValue(1);
3513 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3514 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3516 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3517 MachinePointerInfo::getJumpTable(),
3518 false, false, false, 0);
3519 Chain = Addr.getValue(1);
3520 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3524 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3525 EVT VT = Op.getValueType();
3528 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3529 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3531 return DAG.UnrollVectorOp(Op.getNode());
3534 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3535 "Invalid type for custom lowering!");
3536 if (VT != MVT::v4i16)
3537 return DAG.UnrollVectorOp(Op.getNode());
3539 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3540 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3543 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3544 EVT VT = Op.getValueType();
3546 return LowerVectorFP_TO_INT(Op, DAG);
3551 switch (Op.getOpcode()) {
3552 default: llvm_unreachable("Invalid opcode!");
3553 case ISD::FP_TO_SINT:
3554 Opc = ARMISD::FTOSI;
3556 case ISD::FP_TO_UINT:
3557 Opc = ARMISD::FTOUI;
3560 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3561 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3564 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3565 EVT VT = Op.getValueType();
3568 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3569 if (VT.getVectorElementType() == MVT::f32)
3571 return DAG.UnrollVectorOp(Op.getNode());
3574 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3575 "Invalid type for custom lowering!");
3576 if (VT != MVT::v4f32)
3577 return DAG.UnrollVectorOp(Op.getNode());
3581 switch (Op.getOpcode()) {
3582 default: llvm_unreachable("Invalid opcode!");
3583 case ISD::SINT_TO_FP:
3584 CastOpc = ISD::SIGN_EXTEND;
3585 Opc = ISD::SINT_TO_FP;
3587 case ISD::UINT_TO_FP:
3588 CastOpc = ISD::ZERO_EXTEND;
3589 Opc = ISD::UINT_TO_FP;
3593 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3594 return DAG.getNode(Opc, dl, VT, Op);
3597 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3598 EVT VT = Op.getValueType();
3600 return LowerVectorINT_TO_FP(Op, DAG);
3605 switch (Op.getOpcode()) {
3606 default: llvm_unreachable("Invalid opcode!");
3607 case ISD::SINT_TO_FP:
3608 Opc = ARMISD::SITOF;
3610 case ISD::UINT_TO_FP:
3611 Opc = ARMISD::UITOF;
3615 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3616 return DAG.getNode(Opc, dl, VT, Op);
3619 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3620 // Implement fcopysign with a fabs and a conditional fneg.
3621 SDValue Tmp0 = Op.getOperand(0);
3622 SDValue Tmp1 = Op.getOperand(1);
3624 EVT VT = Op.getValueType();
3625 EVT SrcVT = Tmp1.getValueType();
3626 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3627 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3628 bool UseNEON = !InGPR && Subtarget->hasNEON();
3631 // Use VBSL to copy the sign bit.
3632 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3633 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3634 DAG.getTargetConstant(EncodedVal, MVT::i32));
3635 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3637 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3638 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3639 DAG.getConstant(32, MVT::i32));
3640 else /*if (VT == MVT::f32)*/
3641 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3642 if (SrcVT == MVT::f32) {
3643 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3645 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3646 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3647 DAG.getConstant(32, MVT::i32));
3648 } else if (VT == MVT::f32)
3649 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3650 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3651 DAG.getConstant(32, MVT::i32));
3652 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3653 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3655 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3657 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3658 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3659 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3661 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3662 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3663 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3664 if (VT == MVT::f32) {
3665 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3666 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3667 DAG.getConstant(0, MVT::i32));
3669 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3675 // Bitcast operand 1 to i32.
3676 if (SrcVT == MVT::f64)
3677 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3678 &Tmp1, 1).getValue(1);
3679 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3681 // Or in the signbit with integer operations.
3682 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3683 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3684 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3685 if (VT == MVT::f32) {
3686 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3687 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3688 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3689 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3692 // f64: Or the high part with signbit and then combine two parts.
3693 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3695 SDValue Lo = Tmp0.getValue(0);
3696 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3697 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3698 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3701 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3702 MachineFunction &MF = DAG.getMachineFunction();
3703 MachineFrameInfo *MFI = MF.getFrameInfo();
3704 MFI->setReturnAddressIsTaken(true);
3706 EVT VT = Op.getValueType();
3708 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3710 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3711 SDValue Offset = DAG.getConstant(4, MVT::i32);
3712 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3713 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3714 MachinePointerInfo(), false, false, false, 0);
3717 // Return LR, which contains the return address. Mark it an implicit live-in.
3718 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3719 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3722 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3723 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3724 MFI->setFrameAddressIsTaken(true);
3726 EVT VT = Op.getValueType();
3727 SDLoc dl(Op); // FIXME probably not meaningful
3728 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3729 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3730 ? ARM::R7 : ARM::R11;
3731 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3733 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3734 MachinePointerInfo(),
3735 false, false, false, 0);
3739 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
3740 /// and size(DestVec) > 128-bits.
3741 /// This is achieved by doing the one extension from the SrcVec, splitting the
3742 /// result, extending these parts, and then concatenating these into the
3744 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
3745 SDValue Op = N->getOperand(0);
3746 EVT SrcVT = Op.getValueType();
3747 EVT DestVT = N->getValueType(0);
3749 assert(DestVT.getSizeInBits() > 128 &&
3750 "Custom sext/zext expansion needs >128-bit vector.");
3751 // If this is a normal length extension, use the default expansion.
3752 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
3753 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
3757 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
3758 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
3759 unsigned NumElts = SrcVT.getVectorNumElements();
3760 LLVMContext &Ctx = *DAG.getContext();
3761 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;
3763 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3765 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3767 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
3770 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
3771 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3772 DAG.getIntPtrConstant(0));
3773 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3774 DAG.getIntPtrConstant(NumElts/2));
3775 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
3776 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
3777 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
3780 /// ExpandBITCAST - If the target supports VFP, this function is called to
3781 /// expand a bit convert where either the source or destination type is i64 to
3782 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3783 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3784 /// vectors), since the legalizer won't know what to do with that.
3785 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3786 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3788 SDValue Op = N->getOperand(0);
3790 // This function is only supposed to be called for i64 types, either as the
3791 // source or destination of the bit convert.
3792 EVT SrcVT = Op.getValueType();
3793 EVT DstVT = N->getValueType(0);
3794 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3795 "ExpandBITCAST called for non-i64 type");
3797 // Turn i64->f64 into VMOVDRR.
3798 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3799 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3800 DAG.getConstant(0, MVT::i32));
3801 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3802 DAG.getConstant(1, MVT::i32));
3803 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3804 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3807 // Turn f64->i64 into VMOVRRD.
3808 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3809 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3810 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3811 // Merge the pieces into a single i64 value.
3812 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3818 /// getZeroVector - Returns a vector of specified type with all zero elements.
3819 /// Zero vectors are used to represent vector negation and in those cases
3820 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3821 /// not support i64 elements, so sometimes the zero vectors will need to be
3822 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3824 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3825 assert(VT.isVector() && "Expected a vector type");
3826 // The canonical modified immediate encoding of a zero vector is....0!
3827 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3828 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3829 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3830 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3833 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3834 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3835 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3836 SelectionDAG &DAG) const {
3837 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3838 EVT VT = Op.getValueType();
3839 unsigned VTBits = VT.getSizeInBits();
3841 SDValue ShOpLo = Op.getOperand(0);
3842 SDValue ShOpHi = Op.getOperand(1);
3843 SDValue ShAmt = Op.getOperand(2);
3845 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3847 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3849 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3850 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3851 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3852 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3853 DAG.getConstant(VTBits, MVT::i32));
3854 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3855 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3856 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3858 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3859 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3861 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3862 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3865 SDValue Ops[2] = { Lo, Hi };
3866 return DAG.getMergeValues(Ops, 2, dl);
3869 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3870 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3871 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3872 SelectionDAG &DAG) const {
3873 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3874 EVT VT = Op.getValueType();
3875 unsigned VTBits = VT.getSizeInBits();
3877 SDValue ShOpLo = Op.getOperand(0);
3878 SDValue ShOpHi = Op.getOperand(1);
3879 SDValue ShAmt = Op.getOperand(2);
3882 assert(Op.getOpcode() == ISD::SHL_PARTS);
3883 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3884 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3885 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3886 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3887 DAG.getConstant(VTBits, MVT::i32));
3888 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3889 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3891 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3892 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3893 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3895 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3896 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3899 SDValue Ops[2] = { Lo, Hi };
3900 return DAG.getMergeValues(Ops, 2, dl);
3903 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3904 SelectionDAG &DAG) const {
3905 // The rounding mode is in bits 23:22 of the FPSCR.
3906 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3907 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3908 // so that the shift + and get folded into a bitfield extract.
3910 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3911 DAG.getConstant(Intrinsic::arm_get_fpscr,
3913 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3914 DAG.getConstant(1U << 22, MVT::i32));
3915 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3916 DAG.getConstant(22, MVT::i32));
3917 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3918 DAG.getConstant(3, MVT::i32));
3921 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3922 const ARMSubtarget *ST) {
3923 EVT VT = N->getValueType(0);
3926 if (!ST->hasV6T2Ops())
3929 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3930 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3933 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3934 /// for each 16-bit element from operand, repeated. The basic idea is to
3935 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3937 /// Trace for v4i16:
3938 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3939 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3940 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3941 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3942 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3943 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3944 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3945 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3946 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3947 EVT VT = N->getValueType(0);
3950 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3951 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3952 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3953 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3954 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3955 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3958 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3959 /// bit-count for each 16-bit element from the operand. We need slightly
3960 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3961 /// 64/128-bit registers.
3963 /// Trace for v4i16:
3964 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3965 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3966 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3967 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3968 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3969 EVT VT = N->getValueType(0);
3972 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3973 if (VT.is64BitVector()) {
3974 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3975 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3976 DAG.getIntPtrConstant(0));
3978 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3979 BitCounts, DAG.getIntPtrConstant(0));
3980 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3984 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3985 /// bit-count for each 32-bit element from the operand. The idea here is
3986 /// to split the vector into 16-bit elements, leverage the 16-bit count
3987 /// routine, and then combine the results.
3989 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3990 /// input = [v0 v1 ] (vi: 32-bit elements)
3991 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3992 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3993 /// vrev: N0 = [k1 k0 k3 k2 ]
3995 /// N1 =+[k1 k0 k3 k2 ]
3997 /// N2 =+[k1 k3 k0 k2 ]
3999 /// Extended =+[k1 k3 k0 k2 ]
4001 /// Extracted=+[k1 k3 ]
4003 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4004 EVT VT = N->getValueType(0);
4007 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4009 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4010 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4011 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4012 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4013 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4015 if (VT.is64BitVector()) {
4016 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4017 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4018 DAG.getIntPtrConstant(0));
4020 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4021 DAG.getIntPtrConstant(0));
4022 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4026 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4027 const ARMSubtarget *ST) {
4028 EVT VT = N->getValueType(0);
4030 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4031 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4032 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4033 "Unexpected type for custom ctpop lowering");
4035 if (VT.getVectorElementType() == MVT::i32)
4036 return lowerCTPOP32BitElements(N, DAG);
4038 return lowerCTPOP16BitElements(N, DAG);
4041 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4042 const ARMSubtarget *ST) {
4043 EVT VT = N->getValueType(0);
4049 // Lower vector shifts on NEON to use VSHL.
4050 assert(ST->hasNEON() && "unexpected vector shift");
4052 // Left shifts translate directly to the vshiftu intrinsic.
4053 if (N->getOpcode() == ISD::SHL)
4054 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4055 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4056 N->getOperand(0), N->getOperand(1));
4058 assert((N->getOpcode() == ISD::SRA ||
4059 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4061 // NEON uses the same intrinsics for both left and right shifts. For
4062 // right shifts, the shift amounts are negative, so negate the vector of
4064 EVT ShiftVT = N->getOperand(1).getValueType();
4065 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4066 getZeroVector(ShiftVT, DAG, dl),
4068 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4069 Intrinsic::arm_neon_vshifts :
4070 Intrinsic::arm_neon_vshiftu);
4071 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4072 DAG.getConstant(vshiftInt, MVT::i32),
4073 N->getOperand(0), NegatedCount);
4076 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4077 const ARMSubtarget *ST) {
4078 EVT VT = N->getValueType(0);
4081 // We can get here for a node like i32 = ISD::SHL i32, i64
4085 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4086 "Unknown shift to lower!");
4088 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4089 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4090 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4093 // If we are in thumb mode, we don't have RRX.
4094 if (ST->isThumb1Only()) return SDValue();
4096 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4097 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4098 DAG.getConstant(0, MVT::i32));
4099 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4100 DAG.getConstant(1, MVT::i32));
4102 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4103 // captures the result into a carry flag.
4104 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4105 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4107 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4108 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4110 // Merge the pieces into a single i64 value.
4111 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4114 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4115 SDValue TmpOp0, TmpOp1;
4116 bool Invert = false;
4120 SDValue Op0 = Op.getOperand(0);
4121 SDValue Op1 = Op.getOperand(1);
4122 SDValue CC = Op.getOperand(2);
4123 EVT VT = Op.getValueType();
4124 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4127 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4128 switch (SetCCOpcode) {
4129 default: llvm_unreachable("Illegal FP comparison");
4131 case ISD::SETNE: Invert = true; // Fallthrough
4133 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4135 case ISD::SETLT: Swap = true; // Fallthrough
4137 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4139 case ISD::SETLE: Swap = true; // Fallthrough
4141 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4142 case ISD::SETUGE: Swap = true; // Fallthrough
4143 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4144 case ISD::SETUGT: Swap = true; // Fallthrough
4145 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4146 case ISD::SETUEQ: Invert = true; // Fallthrough
4148 // Expand this to (OLT | OGT).
4152 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4153 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4155 case ISD::SETUO: Invert = true; // Fallthrough
4157 // Expand this to (OLT | OGE).
4161 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4162 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4166 // Integer comparisons.
4167 switch (SetCCOpcode) {
4168 default: llvm_unreachable("Illegal integer comparison");
4169 case ISD::SETNE: Invert = true;
4170 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4171 case ISD::SETLT: Swap = true;
4172 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4173 case ISD::SETLE: Swap = true;
4174 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4175 case ISD::SETULT: Swap = true;
4176 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4177 case ISD::SETULE: Swap = true;
4178 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4181 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4182 if (Opc == ARMISD::VCEQ) {
4185 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4187 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4190 // Ignore bitconvert.
4191 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4192 AndOp = AndOp.getOperand(0);
4194 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4196 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4197 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4204 std::swap(Op0, Op1);
4206 // If one of the operands is a constant vector zero, attempt to fold the
4207 // comparison to a specialized compare-against-zero form.
4209 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4211 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4212 if (Opc == ARMISD::VCGE)
4213 Opc = ARMISD::VCLEZ;
4214 else if (Opc == ARMISD::VCGT)
4215 Opc = ARMISD::VCLTZ;
4220 if (SingleOp.getNode()) {
4223 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4225 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4227 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4229 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4231 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4233 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4236 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4240 Result = DAG.getNOT(dl, Result, VT);
4245 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4246 /// valid vector constant for a NEON instruction with a "modified immediate"
4247 /// operand (e.g., VMOV). If so, return the encoded value.
4248 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4249 unsigned SplatBitSize, SelectionDAG &DAG,
4250 EVT &VT, bool is128Bits, NEONModImmType type) {
4251 unsigned OpCmode, Imm;
4253 // SplatBitSize is set to the smallest size that splats the vector, so a
4254 // zero vector will always have SplatBitSize == 8. However, NEON modified
4255 // immediate instructions others than VMOV do not support the 8-bit encoding
4256 // of a zero vector, and the default encoding of zero is supposed to be the
4261 switch (SplatBitSize) {
4263 if (type != VMOVModImm)
4265 // Any 1-byte value is OK. Op=0, Cmode=1110.
4266 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4269 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4273 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4274 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4275 if ((SplatBits & ~0xff) == 0) {
4276 // Value = 0x00nn: Op=x, Cmode=100x.
4281 if ((SplatBits & ~0xff00) == 0) {
4282 // Value = 0xnn00: Op=x, Cmode=101x.
4284 Imm = SplatBits >> 8;
4290 // NEON's 32-bit VMOV supports splat values where:
4291 // * only one byte is nonzero, or
4292 // * the least significant byte is 0xff and the second byte is nonzero, or
4293 // * the least significant 2 bytes are 0xff and the third is nonzero.
4294 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4295 if ((SplatBits & ~0xff) == 0) {
4296 // Value = 0x000000nn: Op=x, Cmode=000x.
4301 if ((SplatBits & ~0xff00) == 0) {
4302 // Value = 0x0000nn00: Op=x, Cmode=001x.
4304 Imm = SplatBits >> 8;
4307 if ((SplatBits & ~0xff0000) == 0) {
4308 // Value = 0x00nn0000: Op=x, Cmode=010x.
4310 Imm = SplatBits >> 16;
4313 if ((SplatBits & ~0xff000000) == 0) {
4314 // Value = 0xnn000000: Op=x, Cmode=011x.
4316 Imm = SplatBits >> 24;
4320 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4321 if (type == OtherModImm) return SDValue();
4323 if ((SplatBits & ~0xffff) == 0 &&
4324 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4325 // Value = 0x0000nnff: Op=x, Cmode=1100.
4327 Imm = SplatBits >> 8;
4332 if ((SplatBits & ~0xffffff) == 0 &&
4333 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4334 // Value = 0x00nnffff: Op=x, Cmode=1101.
4336 Imm = SplatBits >> 16;
4337 SplatBits |= 0xffff;
4341 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4342 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4343 // VMOV.I32. A (very) minor optimization would be to replicate the value
4344 // and fall through here to test for a valid 64-bit splat. But, then the
4345 // caller would also need to check and handle the change in size.
4349 if (type != VMOVModImm)
4351 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4352 uint64_t BitMask = 0xff;
4354 unsigned ImmMask = 1;
4356 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4357 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4360 } else if ((SplatBits & BitMask) != 0) {
4366 // Op=1, Cmode=1110.
4369 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4374 llvm_unreachable("unexpected size for isNEONModifiedImm");
4377 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4378 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4381 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4382 const ARMSubtarget *ST) const {
4386 bool IsDouble = Op.getValueType() == MVT::f64;
4387 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4389 // Try splatting with a VMOV.f32...
4390 APFloat FPVal = CFP->getValueAPF();
4391 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4394 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4395 // We have code in place to select a valid ConstantFP already, no need to
4400 // It's a float and we are trying to use NEON operations where
4401 // possible. Lower it to a splat followed by an extract.
4403 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4404 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4406 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4407 DAG.getConstant(0, MVT::i32));
4410 // The rest of our options are NEON only, make sure that's allowed before
4412 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4416 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4418 // It wouldn't really be worth bothering for doubles except for one very
4419 // important value, which does happen to match: 0.0. So make sure we don't do
4421 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4424 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4425 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4427 if (NewVal != SDValue()) {
4429 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4432 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4434 // It's a float: cast and extract a vector element.
4435 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4437 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4438 DAG.getConstant(0, MVT::i32));
4441 // Finally, try a VMVN.i32
4442 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4444 if (NewVal != SDValue()) {
4446 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4449 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4451 // It's a float: cast and extract a vector element.
4452 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4454 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4455 DAG.getConstant(0, MVT::i32));
4461 // check if an VEXT instruction can handle the shuffle mask when the
4462 // vector sources of the shuffle are the same.
4463 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4464 unsigned NumElts = VT.getVectorNumElements();
4466 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4472 // If this is a VEXT shuffle, the immediate value is the index of the first
4473 // element. The other shuffle indices must be the successive elements after
4475 unsigned ExpectedElt = Imm;
4476 for (unsigned i = 1; i < NumElts; ++i) {
4477 // Increment the expected index. If it wraps around, just follow it
4478 // back to index zero and keep going.
4480 if (ExpectedElt == NumElts)
4483 if (M[i] < 0) continue; // ignore UNDEF indices
4484 if (ExpectedElt != static_cast<unsigned>(M[i]))
4492 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4493 bool &ReverseVEXT, unsigned &Imm) {
4494 unsigned NumElts = VT.getVectorNumElements();
4495 ReverseVEXT = false;
4497 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4503 // If this is a VEXT shuffle, the immediate value is the index of the first
4504 // element. The other shuffle indices must be the successive elements after
4506 unsigned ExpectedElt = Imm;
4507 for (unsigned i = 1; i < NumElts; ++i) {
4508 // Increment the expected index. If it wraps around, it may still be
4509 // a VEXT but the source vectors must be swapped.
4511 if (ExpectedElt == NumElts * 2) {
4516 if (M[i] < 0) continue; // ignore UNDEF indices
4517 if (ExpectedElt != static_cast<unsigned>(M[i]))
4521 // Adjust the index value if the source operands will be swapped.
4528 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4529 /// instruction with the specified blocksize. (The order of the elements
4530 /// within each block of the vector is reversed.)
4531 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4532 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4533 "Only possible block sizes for VREV are: 16, 32, 64");
4535 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4539 unsigned NumElts = VT.getVectorNumElements();
4540 unsigned BlockElts = M[0] + 1;
4541 // If the first shuffle index is UNDEF, be optimistic.
4543 BlockElts = BlockSize / EltSz;
4545 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4548 for (unsigned i = 0; i < NumElts; ++i) {
4549 if (M[i] < 0) continue; // ignore UNDEF indices
4550 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4557 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4558 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4559 // range, then 0 is placed into the resulting vector. So pretty much any mask
4560 // of 8 elements can work here.
4561 return VT == MVT::v8i8 && M.size() == 8;
4564 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4565 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4569 unsigned NumElts = VT.getVectorNumElements();
4570 WhichResult = (M[0] == 0 ? 0 : 1);
4571 for (unsigned i = 0; i < NumElts; i += 2) {
4572 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4573 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4579 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4580 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4581 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4582 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4583 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4587 unsigned NumElts = VT.getVectorNumElements();
4588 WhichResult = (M[0] == 0 ? 0 : 1);
4589 for (unsigned i = 0; i < NumElts; i += 2) {
4590 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4591 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4597 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4598 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4602 unsigned NumElts = VT.getVectorNumElements();
4603 WhichResult = (M[0] == 0 ? 0 : 1);
4604 for (unsigned i = 0; i != NumElts; ++i) {
4605 if (M[i] < 0) continue; // ignore UNDEF indices
4606 if ((unsigned) M[i] != 2 * i + WhichResult)
4610 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4611 if (VT.is64BitVector() && EltSz == 32)
4617 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4618 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4619 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4620 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4621 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4625 unsigned Half = VT.getVectorNumElements() / 2;
4626 WhichResult = (M[0] == 0 ? 0 : 1);
4627 for (unsigned j = 0; j != 2; ++j) {
4628 unsigned Idx = WhichResult;
4629 for (unsigned i = 0; i != Half; ++i) {
4630 int MIdx = M[i + j * Half];
4631 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4637 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4638 if (VT.is64BitVector() && EltSz == 32)
4644 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4645 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4649 unsigned NumElts = VT.getVectorNumElements();
4650 WhichResult = (M[0] == 0 ? 0 : 1);
4651 unsigned Idx = WhichResult * NumElts / 2;
4652 for (unsigned i = 0; i != NumElts; i += 2) {
4653 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4654 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4659 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4660 if (VT.is64BitVector() && EltSz == 32)
4666 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4667 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4668 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4669 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4670 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4674 unsigned NumElts = VT.getVectorNumElements();
4675 WhichResult = (M[0] == 0 ? 0 : 1);
4676 unsigned Idx = WhichResult * NumElts / 2;
4677 for (unsigned i = 0; i != NumElts; i += 2) {
4678 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4679 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4684 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4685 if (VT.is64BitVector() && EltSz == 32)
4691 /// \return true if this is a reverse operation on an vector.
4692 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4693 unsigned NumElts = VT.getVectorNumElements();
4694 // Make sure the mask has the right size.
4695 if (NumElts != M.size())
4698 // Look for <15, ..., 3, -1, 1, 0>.
4699 for (unsigned i = 0; i != NumElts; ++i)
4700 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4706 // If N is an integer constant that can be moved into a register in one
4707 // instruction, return an SDValue of such a constant (will become a MOV
4708 // instruction). Otherwise return null.
4709 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4710 const ARMSubtarget *ST, SDLoc dl) {
4712 if (!isa<ConstantSDNode>(N))
4714 Val = cast<ConstantSDNode>(N)->getZExtValue();
4716 if (ST->isThumb1Only()) {
4717 if (Val <= 255 || ~Val <= 255)
4718 return DAG.getConstant(Val, MVT::i32);
4720 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4721 return DAG.getConstant(Val, MVT::i32);
4726 // If this is a case we can't handle, return null and let the default
4727 // expansion code take care of it.
4728 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4729 const ARMSubtarget *ST) const {
4730 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4732 EVT VT = Op.getValueType();
4734 APInt SplatBits, SplatUndef;
4735 unsigned SplatBitSize;
4737 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4738 if (SplatBitSize <= 64) {
4739 // Check if an immediate VMOV works.
4741 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4742 SplatUndef.getZExtValue(), SplatBitSize,
4743 DAG, VmovVT, VT.is128BitVector(),
4745 if (Val.getNode()) {
4746 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4747 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4750 // Try an immediate VMVN.
4751 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4752 Val = isNEONModifiedImm(NegatedImm,
4753 SplatUndef.getZExtValue(), SplatBitSize,
4754 DAG, VmovVT, VT.is128BitVector(),
4756 if (Val.getNode()) {
4757 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4758 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4761 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4762 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4763 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4765 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4766 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4772 // Scan through the operands to see if only one value is used.
4774 // As an optimisation, even if more than one value is used it may be more
4775 // profitable to splat with one value then change some lanes.
4777 // Heuristically we decide to do this if the vector has a "dominant" value,
4778 // defined as splatted to more than half of the lanes.
4779 unsigned NumElts = VT.getVectorNumElements();
4780 bool isOnlyLowElement = true;
4781 bool usesOnlyOneValue = true;
4782 bool hasDominantValue = false;
4783 bool isConstant = true;
4785 // Map of the number of times a particular SDValue appears in the
4787 DenseMap<SDValue, unsigned> ValueCounts;
4789 for (unsigned i = 0; i < NumElts; ++i) {
4790 SDValue V = Op.getOperand(i);
4791 if (V.getOpcode() == ISD::UNDEF)
4794 isOnlyLowElement = false;
4795 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4798 ValueCounts.insert(std::make_pair(V, 0));
4799 unsigned &Count = ValueCounts[V];
4801 // Is this value dominant? (takes up more than half of the lanes)
4802 if (++Count > (NumElts / 2)) {
4803 hasDominantValue = true;
4807 if (ValueCounts.size() != 1)
4808 usesOnlyOneValue = false;
4809 if (!Value.getNode() && ValueCounts.size() > 0)
4810 Value = ValueCounts.begin()->first;
4812 if (ValueCounts.size() == 0)
4813 return DAG.getUNDEF(VT);
4815 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4816 // Keep going if we are hitting this case.
4817 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4818 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4820 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4822 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4823 // i32 and try again.
4824 if (hasDominantValue && EltSize <= 32) {
4828 // If we are VDUPing a value that comes directly from a vector, that will
4829 // cause an unnecessary move to and from a GPR, where instead we could
4830 // just use VDUPLANE. We can only do this if the lane being extracted
4831 // is at a constant index, as the VDUP from lane instructions only have
4832 // constant-index forms.
4833 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4834 isa<ConstantSDNode>(Value->getOperand(1))) {
4835 // We need to create a new undef vector to use for the VDUPLANE if the
4836 // size of the vector from which we get the value is different than the
4837 // size of the vector that we need to create. We will insert the element
4838 // such that the register coalescer will remove unnecessary copies.
4839 if (VT != Value->getOperand(0).getValueType()) {
4840 ConstantSDNode *constIndex;
4841 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4842 assert(constIndex && "The index is not a constant!");
4843 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4844 VT.getVectorNumElements();
4845 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4846 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4847 Value, DAG.getConstant(index, MVT::i32)),
4848 DAG.getConstant(index, MVT::i32));
4850 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4851 Value->getOperand(0), Value->getOperand(1));
4853 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4855 if (!usesOnlyOneValue) {
4856 // The dominant value was splatted as 'N', but we now have to insert
4857 // all differing elements.
4858 for (unsigned I = 0; I < NumElts; ++I) {
4859 if (Op.getOperand(I) == Value)
4861 SmallVector<SDValue, 3> Ops;
4863 Ops.push_back(Op.getOperand(I));
4864 Ops.push_back(DAG.getConstant(I, MVT::i32));
4865 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4870 if (VT.getVectorElementType().isFloatingPoint()) {
4871 SmallVector<SDValue, 8> Ops;
4872 for (unsigned i = 0; i < NumElts; ++i)
4873 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4875 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4876 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4877 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4879 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4881 if (usesOnlyOneValue) {
4882 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4883 if (isConstant && Val.getNode())
4884 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4888 // If all elements are constants and the case above didn't get hit, fall back
4889 // to the default expansion, which will generate a load from the constant
4894 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4896 SDValue shuffle = ReconstructShuffle(Op, DAG);
4897 if (shuffle != SDValue())
4901 // Vectors with 32- or 64-bit elements can be built by directly assigning
4902 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4903 // will be legalized.
4904 if (EltSize >= 32) {
4905 // Do the expansion with floating-point types, since that is what the VFP
4906 // registers are defined to use, and since i64 is not legal.
4907 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4908 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4909 SmallVector<SDValue, 8> Ops;
4910 for (unsigned i = 0; i < NumElts; ++i)
4911 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4912 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4913 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4916 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4917 // know the default expansion would otherwise fall back on something even
4918 // worse. For a vector with one or two non-undef values, that's
4919 // scalar_to_vector for the elements followed by a shuffle (provided the
4920 // shuffle is valid for the target) and materialization element by element
4921 // on the stack followed by a load for everything else.
4922 if (!isConstant && !usesOnlyOneValue) {
4923 SDValue Vec = DAG.getUNDEF(VT);
4924 for (unsigned i = 0 ; i < NumElts; ++i) {
4925 SDValue V = Op.getOperand(i);
4926 if (V.getOpcode() == ISD::UNDEF)
4928 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4929 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4937 // Gather data to see if the operation can be modelled as a
4938 // shuffle in combination with VEXTs.
4939 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4940 SelectionDAG &DAG) const {
4942 EVT VT = Op.getValueType();
4943 unsigned NumElts = VT.getVectorNumElements();
4945 SmallVector<SDValue, 2> SourceVecs;
4946 SmallVector<unsigned, 2> MinElts;
4947 SmallVector<unsigned, 2> MaxElts;
4949 for (unsigned i = 0; i < NumElts; ++i) {
4950 SDValue V = Op.getOperand(i);
4951 if (V.getOpcode() == ISD::UNDEF)
4953 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4954 // A shuffle can only come from building a vector from various
4955 // elements of other vectors.
4957 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4958 VT.getVectorElementType()) {
4959 // This code doesn't know how to handle shuffles where the vector
4960 // element types do not match (this happens because type legalization
4961 // promotes the return type of EXTRACT_VECTOR_ELT).
4962 // FIXME: It might be appropriate to extend this code to handle
4963 // mismatched types.
4967 // Record this extraction against the appropriate vector if possible...
4968 SDValue SourceVec = V.getOperand(0);
4969 // If the element number isn't a constant, we can't effectively
4970 // analyze what's going on.
4971 if (!isa<ConstantSDNode>(V.getOperand(1)))
4973 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4974 bool FoundSource = false;
4975 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4976 if (SourceVecs[j] == SourceVec) {
4977 if (MinElts[j] > EltNo)
4979 if (MaxElts[j] < EltNo)
4986 // Or record a new source if not...
4988 SourceVecs.push_back(SourceVec);
4989 MinElts.push_back(EltNo);
4990 MaxElts.push_back(EltNo);
4994 // Currently only do something sane when at most two source vectors
4996 if (SourceVecs.size() > 2)
4999 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5000 int VEXTOffsets[2] = {0, 0};
5002 // This loop extracts the usage patterns of the source vectors
5003 // and prepares appropriate SDValues for a shuffle if possible.
5004 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5005 if (SourceVecs[i].getValueType() == VT) {
5006 // No VEXT necessary
5007 ShuffleSrcs[i] = SourceVecs[i];
5010 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5011 // It probably isn't worth padding out a smaller vector just to
5012 // break it down again in a shuffle.
5016 // Since only 64-bit and 128-bit vectors are legal on ARM and
5017 // we've eliminated the other cases...
5018 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5019 "unexpected vector sizes in ReconstructShuffle");
5021 if (MaxElts[i] - MinElts[i] >= NumElts) {
5022 // Span too large for a VEXT to cope
5026 if (MinElts[i] >= NumElts) {
5027 // The extraction can just take the second half
5028 VEXTOffsets[i] = NumElts;
5029 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5031 DAG.getIntPtrConstant(NumElts));
5032 } else if (MaxElts[i] < NumElts) {
5033 // The extraction can just take the first half
5035 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5037 DAG.getIntPtrConstant(0));
5039 // An actual VEXT is needed
5040 VEXTOffsets[i] = MinElts[i];
5041 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5043 DAG.getIntPtrConstant(0));
5044 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5046 DAG.getIntPtrConstant(NumElts));
5047 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5048 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5052 SmallVector<int, 8> Mask;
5054 for (unsigned i = 0; i < NumElts; ++i) {
5055 SDValue Entry = Op.getOperand(i);
5056 if (Entry.getOpcode() == ISD::UNDEF) {
5061 SDValue ExtractVec = Entry.getOperand(0);
5062 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5063 .getOperand(1))->getSExtValue();
5064 if (ExtractVec == SourceVecs[0]) {
5065 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5067 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5071 // Final check before we try to produce nonsense...
5072 if (isShuffleMaskLegal(Mask, VT))
5073 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5079 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5080 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5081 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5082 /// are assumed to be legal.
5084 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5086 if (VT.getVectorNumElements() == 4 &&
5087 (VT.is128BitVector() || VT.is64BitVector())) {
5088 unsigned PFIndexes[4];
5089 for (unsigned i = 0; i != 4; ++i) {
5093 PFIndexes[i] = M[i];
5096 // Compute the index in the perfect shuffle table.
5097 unsigned PFTableIndex =
5098 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5099 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5100 unsigned Cost = (PFEntry >> 30);
5107 unsigned Imm, WhichResult;
5109 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5110 return (EltSize >= 32 ||
5111 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5112 isVREVMask(M, VT, 64) ||
5113 isVREVMask(M, VT, 32) ||
5114 isVREVMask(M, VT, 16) ||
5115 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5116 isVTBLMask(M, VT) ||
5117 isVTRNMask(M, VT, WhichResult) ||
5118 isVUZPMask(M, VT, WhichResult) ||
5119 isVZIPMask(M, VT, WhichResult) ||
5120 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5121 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5122 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5123 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5126 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5127 /// the specified operations to build the shuffle.
5128 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5129 SDValue RHS, SelectionDAG &DAG,
5131 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5132 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5133 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5136 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5145 OP_VUZPL, // VUZP, left result
5146 OP_VUZPR, // VUZP, right result
5147 OP_VZIPL, // VZIP, left result
5148 OP_VZIPR, // VZIP, right result
5149 OP_VTRNL, // VTRN, left result
5150 OP_VTRNR // VTRN, right result
5153 if (OpNum == OP_COPY) {
5154 if (LHSID == (1*9+2)*9+3) return LHS;
5155 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5159 SDValue OpLHS, OpRHS;
5160 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5161 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5162 EVT VT = OpLHS.getValueType();
5165 default: llvm_unreachable("Unknown shuffle opcode!");
5167 // VREV divides the vector in half and swaps within the half.
5168 if (VT.getVectorElementType() == MVT::i32 ||
5169 VT.getVectorElementType() == MVT::f32)
5170 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5171 // vrev <4 x i16> -> VREV32
5172 if (VT.getVectorElementType() == MVT::i16)
5173 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5174 // vrev <4 x i8> -> VREV16
5175 assert(VT.getVectorElementType() == MVT::i8);
5176 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5181 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5182 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5186 return DAG.getNode(ARMISD::VEXT, dl, VT,
5188 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5191 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5192 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5195 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5196 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5199 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5200 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5204 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5205 ArrayRef<int> ShuffleMask,
5206 SelectionDAG &DAG) {
5207 // Check to see if we can use the VTBL instruction.
5208 SDValue V1 = Op.getOperand(0);
5209 SDValue V2 = Op.getOperand(1);
5212 SmallVector<SDValue, 8> VTBLMask;
5213 for (ArrayRef<int>::iterator
5214 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5215 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5217 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5218 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5219 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5222 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5223 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5227 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5228 SelectionDAG &DAG) {
5230 SDValue OpLHS = Op.getOperand(0);
5231 EVT VT = OpLHS.getValueType();
5233 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5234 "Expect an v8i16/v16i8 type");
5235 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5236 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5237 // extract the first 8 bytes into the top double word and the last 8 bytes
5238 // into the bottom double word. The v8i16 case is similar.
5239 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5240 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5241 DAG.getConstant(ExtractNum, MVT::i32));
5244 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5245 SDValue V1 = Op.getOperand(0);
5246 SDValue V2 = Op.getOperand(1);
5248 EVT VT = Op.getValueType();
5249 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5251 // Convert shuffles that are directly supported on NEON to target-specific
5252 // DAG nodes, instead of keeping them as shuffles and matching them again
5253 // during code selection. This is more efficient and avoids the possibility
5254 // of inconsistencies between legalization and selection.
5255 // FIXME: floating-point vectors should be canonicalized to integer vectors
5256 // of the same time so that they get CSEd properly.
5257 ArrayRef<int> ShuffleMask = SVN->getMask();
5259 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5260 if (EltSize <= 32) {
5261 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5262 int Lane = SVN->getSplatIndex();
5263 // If this is undef splat, generate it via "just" vdup, if possible.
5264 if (Lane == -1) Lane = 0;
5266 // Test if V1 is a SCALAR_TO_VECTOR.
5267 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5268 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5270 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5271 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5273 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5274 !isa<ConstantSDNode>(V1.getOperand(0))) {
5275 bool IsScalarToVector = true;
5276 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5277 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5278 IsScalarToVector = false;
5281 if (IsScalarToVector)
5282 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5284 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5285 DAG.getConstant(Lane, MVT::i32));
5290 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5293 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5294 DAG.getConstant(Imm, MVT::i32));
5297 if (isVREVMask(ShuffleMask, VT, 64))
5298 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5299 if (isVREVMask(ShuffleMask, VT, 32))
5300 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5301 if (isVREVMask(ShuffleMask, VT, 16))
5302 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5304 if (V2->getOpcode() == ISD::UNDEF &&
5305 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5306 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5307 DAG.getConstant(Imm, MVT::i32));
5310 // Check for Neon shuffles that modify both input vectors in place.
5311 // If both results are used, i.e., if there are two shuffles with the same
5312 // source operands and with masks corresponding to both results of one of
5313 // these operations, DAG memoization will ensure that a single node is
5314 // used for both shuffles.
5315 unsigned WhichResult;
5316 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5317 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5318 V1, V2).getValue(WhichResult);
5319 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5320 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5321 V1, V2).getValue(WhichResult);
5322 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5323 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5324 V1, V2).getValue(WhichResult);
5326 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5327 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5328 V1, V1).getValue(WhichResult);
5329 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5330 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5331 V1, V1).getValue(WhichResult);
5332 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5333 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5334 V1, V1).getValue(WhichResult);
5337 // If the shuffle is not directly supported and it has 4 elements, use
5338 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5339 unsigned NumElts = VT.getVectorNumElements();
5341 unsigned PFIndexes[4];
5342 for (unsigned i = 0; i != 4; ++i) {
5343 if (ShuffleMask[i] < 0)
5346 PFIndexes[i] = ShuffleMask[i];
5349 // Compute the index in the perfect shuffle table.
5350 unsigned PFTableIndex =
5351 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5352 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5353 unsigned Cost = (PFEntry >> 30);
5356 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5359 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5360 if (EltSize >= 32) {
5361 // Do the expansion with floating-point types, since that is what the VFP
5362 // registers are defined to use, and since i64 is not legal.
5363 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5364 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5365 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5366 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5367 SmallVector<SDValue, 8> Ops;
5368 for (unsigned i = 0; i < NumElts; ++i) {
5369 if (ShuffleMask[i] < 0)
5370 Ops.push_back(DAG.getUNDEF(EltVT));
5372 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5373 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5374 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5377 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5378 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5381 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5382 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5384 if (VT == MVT::v8i8) {
5385 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5386 if (NewOp.getNode())
5393 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5394 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5395 SDValue Lane = Op.getOperand(2);
5396 if (!isa<ConstantSDNode>(Lane))
5402 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5403 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5404 SDValue Lane = Op.getOperand(1);
5405 if (!isa<ConstantSDNode>(Lane))
5408 SDValue Vec = Op.getOperand(0);
5409 if (Op.getValueType() == MVT::i32 &&
5410 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5412 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5418 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5419 // The only time a CONCAT_VECTORS operation can have legal types is when
5420 // two 64-bit vectors are concatenated to a 128-bit vector.
5421 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5422 "unexpected CONCAT_VECTORS");
5424 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5425 SDValue Op0 = Op.getOperand(0);
5426 SDValue Op1 = Op.getOperand(1);
5427 if (Op0.getOpcode() != ISD::UNDEF)
5428 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5429 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5430 DAG.getIntPtrConstant(0));
5431 if (Op1.getOpcode() != ISD::UNDEF)
5432 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5433 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5434 DAG.getIntPtrConstant(1));
5435 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5438 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5439 /// element has been zero/sign-extended, depending on the isSigned parameter,
5440 /// from an integer type half its size.
5441 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5443 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5444 EVT VT = N->getValueType(0);
5445 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5446 SDNode *BVN = N->getOperand(0).getNode();
5447 if (BVN->getValueType(0) != MVT::v4i32 ||
5448 BVN->getOpcode() != ISD::BUILD_VECTOR)
5450 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5451 unsigned HiElt = 1 - LoElt;
5452 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5453 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5454 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5455 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5456 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5459 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5460 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5463 if (Hi0->isNullValue() && Hi1->isNullValue())
5469 if (N->getOpcode() != ISD::BUILD_VECTOR)
5472 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5473 SDNode *Elt = N->getOperand(i).getNode();
5474 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5475 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5476 unsigned HalfSize = EltSize / 2;
5478 if (!isIntN(HalfSize, C->getSExtValue()))
5481 if (!isUIntN(HalfSize, C->getZExtValue()))
5492 /// isSignExtended - Check if a node is a vector value that is sign-extended
5493 /// or a constant BUILD_VECTOR with sign-extended elements.
5494 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5495 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5497 if (isExtendedBUILD_VECTOR(N, DAG, true))
5502 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5503 /// or a constant BUILD_VECTOR with zero-extended elements.
5504 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5505 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5507 if (isExtendedBUILD_VECTOR(N, DAG, false))
5512 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5513 if (OrigVT.getSizeInBits() >= 64)
5516 assert(OrigVT.isSimple() && "Expecting a simple value type");
5518 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5519 switch (OrigSimpleTy) {
5520 default: llvm_unreachable("Unexpected Vector Type");
5529 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5530 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5531 /// We insert the required extension here to get the vector to fill a D register.
5532 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5535 unsigned ExtOpcode) {
5536 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5537 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5538 // 64-bits we need to insert a new extension so that it will be 64-bits.
5539 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5540 if (OrigTy.getSizeInBits() >= 64)
5543 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5544 EVT NewVT = getExtensionTo64Bits(OrigTy);
5546 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5549 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5550 /// does not do any sign/zero extension. If the original vector is less
5551 /// than 64 bits, an appropriate extension will be added after the load to
5552 /// reach a total size of 64 bits. We have to add the extension separately
5553 /// because ARM does not have a sign/zero extending load for vectors.
5554 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5555 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5557 // The load already has the right type.
5558 if (ExtendedTy == LD->getMemoryVT())
5559 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5560 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5561 LD->isNonTemporal(), LD->isInvariant(),
5562 LD->getAlignment());
5564 // We need to create a zextload/sextload. We cannot just create a load
5565 // followed by a zext/zext node because LowerMUL is also run during normal
5566 // operation legalization where we can't create illegal types.
5567 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5568 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5569 LD->getMemoryVT(), LD->isVolatile(),
5570 LD->isNonTemporal(), LD->getAlignment());
5573 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5574 /// extending load, or BUILD_VECTOR with extended elements, return the
5575 /// unextended value. The unextended vector should be 64 bits so that it can
5576 /// be used as an operand to a VMULL instruction. If the original vector size
5577 /// before extension is less than 64 bits we add a an extension to resize
5578 /// the vector to 64 bits.
5579 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5580 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5581 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5582 N->getOperand(0)->getValueType(0),
5586 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5587 return SkipLoadExtensionForVMULL(LD, DAG);
5589 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5590 // have been legalized as a BITCAST from v4i32.
5591 if (N->getOpcode() == ISD::BITCAST) {
5592 SDNode *BVN = N->getOperand(0).getNode();
5593 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5594 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5595 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5596 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5597 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5599 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5600 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5601 EVT VT = N->getValueType(0);
5602 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5603 unsigned NumElts = VT.getVectorNumElements();
5604 MVT TruncVT = MVT::getIntegerVT(EltSize);
5605 SmallVector<SDValue, 8> Ops;
5606 for (unsigned i = 0; i != NumElts; ++i) {
5607 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5608 const APInt &CInt = C->getAPIntValue();
5609 // Element types smaller than 32 bits are not legal, so use i32 elements.
5610 // The values are implicitly truncated so sext vs. zext doesn't matter.
5611 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5613 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5614 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5617 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5618 unsigned Opcode = N->getOpcode();
5619 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5620 SDNode *N0 = N->getOperand(0).getNode();
5621 SDNode *N1 = N->getOperand(1).getNode();
5622 return N0->hasOneUse() && N1->hasOneUse() &&
5623 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5628 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5629 unsigned Opcode = N->getOpcode();
5630 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5631 SDNode *N0 = N->getOperand(0).getNode();
5632 SDNode *N1 = N->getOperand(1).getNode();
5633 return N0->hasOneUse() && N1->hasOneUse() &&
5634 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5639 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5640 // Multiplications are only custom-lowered for 128-bit vectors so that
5641 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5642 EVT VT = Op.getValueType();
5643 assert(VT.is128BitVector() && VT.isInteger() &&
5644 "unexpected type for custom-lowering ISD::MUL");
5645 SDNode *N0 = Op.getOperand(0).getNode();
5646 SDNode *N1 = Op.getOperand(1).getNode();
5647 unsigned NewOpc = 0;
5649 bool isN0SExt = isSignExtended(N0, DAG);
5650 bool isN1SExt = isSignExtended(N1, DAG);
5651 if (isN0SExt && isN1SExt)
5652 NewOpc = ARMISD::VMULLs;
5654 bool isN0ZExt = isZeroExtended(N0, DAG);
5655 bool isN1ZExt = isZeroExtended(N1, DAG);
5656 if (isN0ZExt && isN1ZExt)
5657 NewOpc = ARMISD::VMULLu;
5658 else if (isN1SExt || isN1ZExt) {
5659 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5660 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5661 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5662 NewOpc = ARMISD::VMULLs;
5664 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5665 NewOpc = ARMISD::VMULLu;
5667 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5669 NewOpc = ARMISD::VMULLu;
5675 if (VT == MVT::v2i64)
5676 // Fall through to expand this. It is not legal.
5679 // Other vector multiplications are legal.
5684 // Legalize to a VMULL instruction.
5687 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5689 Op0 = SkipExtensionForVMULL(N0, DAG);
5690 assert(Op0.getValueType().is64BitVector() &&
5691 Op1.getValueType().is64BitVector() &&
5692 "unexpected types for extended operands to VMULL");
5693 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5696 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5697 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5704 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5705 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5706 EVT Op1VT = Op1.getValueType();
5707 return DAG.getNode(N0->getOpcode(), DL, VT,
5708 DAG.getNode(NewOpc, DL, VT,
5709 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5710 DAG.getNode(NewOpc, DL, VT,
5711 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5715 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5717 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5718 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5719 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5720 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5721 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5722 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5723 // Get reciprocal estimate.
5724 // float4 recip = vrecpeq_f32(yf);
5725 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5726 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5727 // Because char has a smaller range than uchar, we can actually get away
5728 // without any newton steps. This requires that we use a weird bias
5729 // of 0xb000, however (again, this has been exhaustively tested).
5730 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5731 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5732 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5733 Y = DAG.getConstant(0xb000, MVT::i32);
5734 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5735 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5736 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5737 // Convert back to short.
5738 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5739 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5744 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5746 // Convert to float.
5747 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5748 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5749 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5750 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5751 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5752 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5754 // Use reciprocal estimate and one refinement step.
5755 // float4 recip = vrecpeq_f32(yf);
5756 // recip *= vrecpsq_f32(yf, recip);
5757 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5758 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5759 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5760 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5762 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5763 // Because short has a smaller range than ushort, we can actually get away
5764 // with only a single newton step. This requires that we use a weird bias
5765 // of 89, however (again, this has been exhaustively tested).
5766 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5767 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5768 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5769 N1 = DAG.getConstant(0x89, MVT::i32);
5770 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5771 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5772 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5773 // Convert back to integer and return.
5774 // return vmovn_s32(vcvt_s32_f32(result));
5775 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5776 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5780 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5781 EVT VT = Op.getValueType();
5782 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5783 "unexpected type for custom-lowering ISD::SDIV");
5786 SDValue N0 = Op.getOperand(0);
5787 SDValue N1 = Op.getOperand(1);
5790 if (VT == MVT::v8i8) {
5791 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5792 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5794 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5795 DAG.getIntPtrConstant(4));
5796 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5797 DAG.getIntPtrConstant(4));
5798 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5799 DAG.getIntPtrConstant(0));
5800 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5801 DAG.getIntPtrConstant(0));
5803 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5804 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5806 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5807 N0 = LowerCONCAT_VECTORS(N0, DAG);
5809 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5812 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5815 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5816 EVT VT = Op.getValueType();
5817 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5818 "unexpected type for custom-lowering ISD::UDIV");
5821 SDValue N0 = Op.getOperand(0);
5822 SDValue N1 = Op.getOperand(1);
5825 if (VT == MVT::v8i8) {
5826 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5827 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5829 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5830 DAG.getIntPtrConstant(4));
5831 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5832 DAG.getIntPtrConstant(4));
5833 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5834 DAG.getIntPtrConstant(0));
5835 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5836 DAG.getIntPtrConstant(0));
5838 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5839 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5841 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5842 N0 = LowerCONCAT_VECTORS(N0, DAG);
5844 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5845 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5850 // v4i16 sdiv ... Convert to float.
5851 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5852 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5853 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5854 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5855 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5856 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5858 // Use reciprocal estimate and two refinement steps.
5859 // float4 recip = vrecpeq_f32(yf);
5860 // recip *= vrecpsq_f32(yf, recip);
5861 // recip *= vrecpsq_f32(yf, recip);
5862 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5863 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5864 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5865 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5867 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5868 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5869 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5871 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5872 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5873 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5874 // and that it will never cause us to return an answer too large).
5875 // float4 result = as_float4(as_int4(xf*recip) + 2);
5876 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5877 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5878 N1 = DAG.getConstant(2, MVT::i32);
5879 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5880 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5881 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5882 // Convert back to integer and return.
5883 // return vmovn_u32(vcvt_s32_f32(result));
5884 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5885 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5889 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5890 EVT VT = Op.getNode()->getValueType(0);
5891 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5894 bool ExtraOp = false;
5895 switch (Op.getOpcode()) {
5896 default: llvm_unreachable("Invalid code");
5897 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5898 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5899 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5900 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5904 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5906 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5907 Op.getOperand(1), Op.getOperand(2));
5910 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5911 // Monotonic load/store is legal for all targets
5912 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5915 // Aquire/Release load/store is not legal for targets without a
5916 // dmb or equivalent available.
5921 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5922 SelectionDAG &DAG, unsigned NewOp) {
5924 assert (Node->getValueType(0) == MVT::i64 &&
5925 "Only know how to expand i64 atomics");
5927 SmallVector<SDValue, 6> Ops;
5928 Ops.push_back(Node->getOperand(0)); // Chain
5929 Ops.push_back(Node->getOperand(1)); // Ptr
5931 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5932 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5933 // High part of Val1
5934 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5935 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5936 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5937 // High part of Val1
5938 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5939 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5940 // High part of Val2
5941 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5942 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5944 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5946 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5947 cast<MemSDNode>(Node)->getMemOperand());
5948 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5949 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5950 Results.push_back(Result.getValue(2));
5953 static void ReplaceREADCYCLECOUNTER(SDNode *N,
5954 SmallVectorImpl<SDValue> &Results,
5956 const ARMSubtarget *Subtarget) {
5958 SDValue Cycles32, OutChain;
5960 if (Subtarget->hasPerfMon()) {
5961 // Under Power Management extensions, the cycle-count is:
5962 // mrc p15, #0, <Rt>, c9, c13, #0
5963 SDValue Ops[] = { N->getOperand(0), // Chain
5964 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
5965 DAG.getConstant(15, MVT::i32),
5966 DAG.getConstant(0, MVT::i32),
5967 DAG.getConstant(9, MVT::i32),
5968 DAG.getConstant(13, MVT::i32),
5969 DAG.getConstant(0, MVT::i32)
5972 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
5973 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
5974 array_lengthof(Ops));
5975 OutChain = Cycles32.getValue(1);
5977 // Intrinsic is defined to return 0 on unsupported platforms. Technically
5978 // there are older ARM CPUs that have implementation-specific ways of
5979 // obtaining this information (FIXME!).
5980 Cycles32 = DAG.getConstant(0, MVT::i32);
5981 OutChain = DAG.getEntryNode();
5985 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
5986 Cycles32, DAG.getConstant(0, MVT::i32));
5987 Results.push_back(Cycles64);
5988 Results.push_back(OutChain);
5991 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5992 switch (Op.getOpcode()) {
5993 default: llvm_unreachable("Don't know how to custom lower this!");
5994 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5995 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5996 case ISD::GlobalAddress:
5997 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5998 LowerGlobalAddressELF(Op, DAG);
5999 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6000 case ISD::SELECT: return LowerSELECT(Op, DAG);
6001 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6002 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6003 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6004 case ISD::VASTART: return LowerVASTART(Op, DAG);
6005 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6006 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6007 case ISD::SINT_TO_FP:
6008 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6009 case ISD::FP_TO_SINT:
6010 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6011 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6012 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6013 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6014 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6015 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6016 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6017 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6019 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6022 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6023 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6024 case ISD::SRL_PARTS:
6025 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6026 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6027 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6028 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6029 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6030 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6031 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6032 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6033 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6034 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6035 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6036 case ISD::MUL: return LowerMUL(Op, DAG);
6037 case ISD::SDIV: return LowerSDIV(Op, DAG);
6038 case ISD::UDIV: return LowerUDIV(Op, DAG);
6042 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6043 case ISD::ATOMIC_LOAD:
6044 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6046 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6050 /// ReplaceNodeResults - Replace the results of node with an illegal result
6051 /// type with new values built out of custom code.
6052 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6053 SmallVectorImpl<SDValue>&Results,
6054 SelectionDAG &DAG) const {
6056 switch (N->getOpcode()) {
6058 llvm_unreachable("Don't know how to custom expand this!");
6060 Res = ExpandBITCAST(N, DAG);
6062 case ISD::SIGN_EXTEND:
6063 case ISD::ZERO_EXTEND:
6064 Res = ExpandVectorExtension(N, DAG);
6068 Res = Expand64BitShift(N, DAG, Subtarget);
6070 case ISD::READCYCLECOUNTER:
6071 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6073 case ISD::ATOMIC_LOAD_ADD:
6074 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
6076 case ISD::ATOMIC_LOAD_AND:
6077 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
6079 case ISD::ATOMIC_LOAD_NAND:
6080 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
6082 case ISD::ATOMIC_LOAD_OR:
6083 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
6085 case ISD::ATOMIC_LOAD_SUB:
6086 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
6088 case ISD::ATOMIC_LOAD_XOR:
6089 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
6091 case ISD::ATOMIC_SWAP:
6092 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
6094 case ISD::ATOMIC_CMP_SWAP:
6095 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
6097 case ISD::ATOMIC_LOAD_MIN:
6098 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
6100 case ISD::ATOMIC_LOAD_UMIN:
6101 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
6103 case ISD::ATOMIC_LOAD_MAX:
6104 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
6106 case ISD::ATOMIC_LOAD_UMAX:
6107 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
6111 Results.push_back(Res);
6114 //===----------------------------------------------------------------------===//
6115 // ARM Scheduler Hooks
6116 //===----------------------------------------------------------------------===//
6119 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6120 MachineBasicBlock *BB,
6121 unsigned Size) const {
6122 unsigned dest = MI->getOperand(0).getReg();
6123 unsigned ptr = MI->getOperand(1).getReg();
6124 unsigned oldval = MI->getOperand(2).getReg();
6125 unsigned newval = MI->getOperand(3).getReg();
6126 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6127 DebugLoc dl = MI->getDebugLoc();
6128 bool isThumb2 = Subtarget->isThumb2();
6130 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6131 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6132 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6133 (const TargetRegisterClass*)&ARM::GPRRegClass);
6136 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6137 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6138 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6141 unsigned ldrOpc, strOpc;
6143 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6145 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6146 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6149 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6150 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6153 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6154 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6158 MachineFunction *MF = BB->getParent();
6159 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6160 MachineFunction::iterator It = BB;
6161 ++It; // insert the new blocks after the current block
6163 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6164 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6165 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6166 MF->insert(It, loop1MBB);
6167 MF->insert(It, loop2MBB);
6168 MF->insert(It, exitMBB);
6170 // Transfer the remainder of BB and its successor edges to exitMBB.
6171 exitMBB->splice(exitMBB->begin(), BB,
6172 llvm::next(MachineBasicBlock::iterator(MI)),
6174 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6178 // fallthrough --> loop1MBB
6179 BB->addSuccessor(loop1MBB);
6182 // ldrex dest, [ptr]
6186 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6187 if (ldrOpc == ARM::t2LDREX)
6189 AddDefaultPred(MIB);
6190 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6191 .addReg(dest).addReg(oldval));
6192 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6193 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6194 BB->addSuccessor(loop2MBB);
6195 BB->addSuccessor(exitMBB);
6198 // strex scratch, newval, [ptr]
6202 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6203 if (strOpc == ARM::t2STREX)
6205 AddDefaultPred(MIB);
6206 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6207 .addReg(scratch).addImm(0));
6208 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6209 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6210 BB->addSuccessor(loop1MBB);
6211 BB->addSuccessor(exitMBB);
6217 MI->eraseFromParent(); // The instruction is gone now.
6223 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6224 unsigned Size, unsigned BinOpcode) const {
6225 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6226 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6228 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6229 MachineFunction *MF = BB->getParent();
6230 MachineFunction::iterator It = BB;
6233 unsigned dest = MI->getOperand(0).getReg();
6234 unsigned ptr = MI->getOperand(1).getReg();
6235 unsigned incr = MI->getOperand(2).getReg();
6236 DebugLoc dl = MI->getDebugLoc();
6237 bool isThumb2 = Subtarget->isThumb2();
6239 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6241 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6242 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6245 unsigned ldrOpc, strOpc;
6247 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6249 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6250 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6253 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6254 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6257 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6258 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6262 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6263 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6264 MF->insert(It, loopMBB);
6265 MF->insert(It, exitMBB);
6267 // Transfer the remainder of BB and its successor edges to exitMBB.
6268 exitMBB->splice(exitMBB->begin(), BB,
6269 llvm::next(MachineBasicBlock::iterator(MI)),
6271 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6273 const TargetRegisterClass *TRC = isThumb2 ?
6274 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6275 (const TargetRegisterClass*)&ARM::GPRRegClass;
6276 unsigned scratch = MRI.createVirtualRegister(TRC);
6277 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6281 // fallthrough --> loopMBB
6282 BB->addSuccessor(loopMBB);
6286 // <binop> scratch2, dest, incr
6287 // strex scratch, scratch2, ptr
6290 // fallthrough --> exitMBB
6292 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6293 if (ldrOpc == ARM::t2LDREX)
6295 AddDefaultPred(MIB);
6297 // operand order needs to go the other way for NAND
6298 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6299 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6300 addReg(incr).addReg(dest)).addReg(0);
6302 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6303 addReg(dest).addReg(incr)).addReg(0);
6306 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6307 if (strOpc == ARM::t2STREX)
6309 AddDefaultPred(MIB);
6310 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6311 .addReg(scratch).addImm(0));
6312 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6313 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6315 BB->addSuccessor(loopMBB);
6316 BB->addSuccessor(exitMBB);
6322 MI->eraseFromParent(); // The instruction is gone now.
6328 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6329 MachineBasicBlock *BB,
6332 ARMCC::CondCodes Cond) const {
6333 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6335 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6336 MachineFunction *MF = BB->getParent();
6337 MachineFunction::iterator It = BB;
6340 unsigned dest = MI->getOperand(0).getReg();
6341 unsigned ptr = MI->getOperand(1).getReg();
6342 unsigned incr = MI->getOperand(2).getReg();
6343 unsigned oldval = dest;
6344 DebugLoc dl = MI->getDebugLoc();
6345 bool isThumb2 = Subtarget->isThumb2();
6347 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6349 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6350 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6353 unsigned ldrOpc, strOpc, extendOpc;
6355 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6357 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6358 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6359 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6362 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6363 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6364 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6367 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6368 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6373 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6374 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6375 MF->insert(It, loopMBB);
6376 MF->insert(It, exitMBB);
6378 // Transfer the remainder of BB and its successor edges to exitMBB.
6379 exitMBB->splice(exitMBB->begin(), BB,
6380 llvm::next(MachineBasicBlock::iterator(MI)),
6382 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6384 const TargetRegisterClass *TRC = isThumb2 ?
6385 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6386 (const TargetRegisterClass*)&ARM::GPRRegClass;
6387 unsigned scratch = MRI.createVirtualRegister(TRC);
6388 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6392 // fallthrough --> loopMBB
6393 BB->addSuccessor(loopMBB);
6397 // (sign extend dest, if required)
6399 // cmov.cond scratch2, incr, dest
6400 // strex scratch, scratch2, ptr
6403 // fallthrough --> exitMBB
6405 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6406 if (ldrOpc == ARM::t2LDREX)
6408 AddDefaultPred(MIB);
6410 // Sign extend the value, if necessary.
6411 if (signExtend && extendOpc) {
6412 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
6413 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6418 // Build compare and cmov instructions.
6419 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6420 .addReg(oldval).addReg(incr));
6421 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6422 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6424 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6425 if (strOpc == ARM::t2STREX)
6427 AddDefaultPred(MIB);
6428 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6429 .addReg(scratch).addImm(0));
6430 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6431 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6433 BB->addSuccessor(loopMBB);
6434 BB->addSuccessor(exitMBB);
6440 MI->eraseFromParent(); // The instruction is gone now.
6446 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6447 unsigned Op1, unsigned Op2,
6448 bool NeedsCarry, bool IsCmpxchg,
6449 bool IsMinMax, ARMCC::CondCodes CC) const {
6450 // This also handles ATOMIC_SWAP, indicated by Op1==0.
6451 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6453 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6454 MachineFunction *MF = BB->getParent();
6455 MachineFunction::iterator It = BB;
6458 unsigned destlo = MI->getOperand(0).getReg();
6459 unsigned desthi = MI->getOperand(1).getReg();
6460 unsigned ptr = MI->getOperand(2).getReg();
6461 unsigned vallo = MI->getOperand(3).getReg();
6462 unsigned valhi = MI->getOperand(4).getReg();
6463 DebugLoc dl = MI->getDebugLoc();
6464 bool isThumb2 = Subtarget->isThumb2();
6466 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6468 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6469 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6470 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6471 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6472 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6475 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6476 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6477 if (IsCmpxchg || IsMinMax)
6478 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6480 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6481 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6483 MF->insert(It, loopMBB);
6484 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6485 if (IsCmpxchg) MF->insert(It, cont2BB);
6486 MF->insert(It, exitMBB);
6488 // Transfer the remainder of BB and its successor edges to exitMBB.
6489 exitMBB->splice(exitMBB->begin(), BB,
6490 llvm::next(MachineBasicBlock::iterator(MI)),
6492 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6494 const TargetRegisterClass *TRC = isThumb2 ?
6495 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6496 (const TargetRegisterClass*)&ARM::GPRRegClass;
6497 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6501 // fallthrough --> loopMBB
6502 BB->addSuccessor(loopMBB);
6505 // ldrexd r2, r3, ptr
6506 // <binopa> r0, r2, incr
6507 // <binopb> r1, r3, incr
6508 // strexd storesuccess, r0, r1, ptr
6509 // cmp storesuccess, #0
6511 // fallthrough --> exitMBB
6516 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD))
6517 .addReg(destlo, RegState::Define)
6518 .addReg(desthi, RegState::Define)
6521 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6522 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD))
6523 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6524 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6525 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6526 .addReg(GPRPair0, 0, ARM::gsub_0);
6527 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6528 .addReg(GPRPair0, 0, ARM::gsub_1);
6531 unsigned StoreLo, StoreHi;
6534 for (unsigned i = 0; i < 2; i++) {
6535 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6537 .addReg(i == 0 ? destlo : desthi)
6538 .addReg(i == 0 ? vallo : valhi));
6539 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6540 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6541 BB->addSuccessor(exitMBB);
6542 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6543 BB = (i == 0 ? contBB : cont2BB);
6546 // Copy to physregs for strexd
6547 StoreLo = MI->getOperand(5).getReg();
6548 StoreHi = MI->getOperand(6).getReg();
6550 // Perform binary operation
6551 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6552 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6553 .addReg(destlo).addReg(vallo))
6554 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6555 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6556 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6557 .addReg(desthi).addReg(valhi))
6558 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6563 // Copy to physregs for strexd
6568 // Compare and branch to exit block.
6569 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6570 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6571 BB->addSuccessor(exitMBB);
6572 BB->addSuccessor(contBB);
6580 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6581 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6582 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess)
6583 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6585 // Marshal a pair...
6586 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6587 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6588 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6589 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6590 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6593 .addImm(ARM::gsub_0);
6594 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6597 .addImm(ARM::gsub_1);
6600 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess)
6601 .addReg(StorePair).addReg(ptr));
6604 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6605 .addReg(storesuccess).addImm(0));
6606 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6607 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6609 BB->addSuccessor(loopMBB);
6610 BB->addSuccessor(exitMBB);
6616 MI->eraseFromParent(); // The instruction is gone now.
6621 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6622 /// registers the function context.
6623 void ARMTargetLowering::
6624 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6625 MachineBasicBlock *DispatchBB, int FI) const {
6626 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6627 DebugLoc dl = MI->getDebugLoc();
6628 MachineFunction *MF = MBB->getParent();
6629 MachineRegisterInfo *MRI = &MF->getRegInfo();
6630 MachineConstantPool *MCP = MF->getConstantPool();
6631 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6632 const Function *F = MF->getFunction();
6634 bool isThumb = Subtarget->isThumb();
6635 bool isThumb2 = Subtarget->isThumb2();
6637 unsigned PCLabelId = AFI->createPICLabelUId();
6638 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6639 ARMConstantPoolValue *CPV =
6640 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6641 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6643 const TargetRegisterClass *TRC = isThumb ?
6644 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6645 (const TargetRegisterClass*)&ARM::GPRRegClass;
6647 // Grab constant pool and fixed stack memory operands.
6648 MachineMemOperand *CPMMO =
6649 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6650 MachineMemOperand::MOLoad, 4, 4);
6652 MachineMemOperand *FIMMOSt =
6653 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6654 MachineMemOperand::MOStore, 4, 4);
6656 // Load the address of the dispatch MBB into the jump buffer.
6658 // Incoming value: jbuf
6659 // ldr.n r5, LCPI1_1
6662 // str r5, [$jbuf, #+4] ; &jbuf[1]
6663 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6664 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6665 .addConstantPoolIndex(CPI)
6666 .addMemOperand(CPMMO));
6667 // Set the low bit because of thumb mode.
6668 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6670 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6671 .addReg(NewVReg1, RegState::Kill)
6673 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6674 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6675 .addReg(NewVReg2, RegState::Kill)
6677 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6678 .addReg(NewVReg3, RegState::Kill)
6680 .addImm(36) // &jbuf[1] :: pc
6681 .addMemOperand(FIMMOSt));
6682 } else if (isThumb) {
6683 // Incoming value: jbuf
6684 // ldr.n r1, LCPI1_4
6688 // add r2, $jbuf, #+4 ; &jbuf[1]
6690 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6691 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6692 .addConstantPoolIndex(CPI)
6693 .addMemOperand(CPMMO));
6694 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6695 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6696 .addReg(NewVReg1, RegState::Kill)
6698 // Set the low bit because of thumb mode.
6699 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6700 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6701 .addReg(ARM::CPSR, RegState::Define)
6703 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6704 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6705 .addReg(ARM::CPSR, RegState::Define)
6706 .addReg(NewVReg2, RegState::Kill)
6707 .addReg(NewVReg3, RegState::Kill));
6708 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6709 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6711 .addImm(36)); // &jbuf[1] :: pc
6712 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6713 .addReg(NewVReg4, RegState::Kill)
6714 .addReg(NewVReg5, RegState::Kill)
6716 .addMemOperand(FIMMOSt));
6718 // Incoming value: jbuf
6721 // str r1, [$jbuf, #+4] ; &jbuf[1]
6722 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6723 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6724 .addConstantPoolIndex(CPI)
6726 .addMemOperand(CPMMO));
6727 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6728 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6729 .addReg(NewVReg1, RegState::Kill)
6730 .addImm(PCLabelId));
6731 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6732 .addReg(NewVReg2, RegState::Kill)
6734 .addImm(36) // &jbuf[1] :: pc
6735 .addMemOperand(FIMMOSt));
6739 MachineBasicBlock *ARMTargetLowering::
6740 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6741 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6742 DebugLoc dl = MI->getDebugLoc();
6743 MachineFunction *MF = MBB->getParent();
6744 MachineRegisterInfo *MRI = &MF->getRegInfo();
6745 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6746 MachineFrameInfo *MFI = MF->getFrameInfo();
6747 int FI = MFI->getFunctionContextIndex();
6749 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6750 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6751 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6753 // Get a mapping of the call site numbers to all of the landing pads they're
6755 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6756 unsigned MaxCSNum = 0;
6757 MachineModuleInfo &MMI = MF->getMMI();
6758 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6760 if (!BB->isLandingPad()) continue;
6762 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6764 for (MachineBasicBlock::iterator
6765 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6766 if (!II->isEHLabel()) continue;
6768 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6769 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6771 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6772 for (SmallVectorImpl<unsigned>::iterator
6773 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6774 CSI != CSE; ++CSI) {
6775 CallSiteNumToLPad[*CSI].push_back(BB);
6776 MaxCSNum = std::max(MaxCSNum, *CSI);
6782 // Get an ordered list of the machine basic blocks for the jump table.
6783 std::vector<MachineBasicBlock*> LPadList;
6784 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6785 LPadList.reserve(CallSiteNumToLPad.size());
6786 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6787 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6788 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6789 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6790 LPadList.push_back(*II);
6791 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6795 assert(!LPadList.empty() &&
6796 "No landing pad destinations for the dispatch jump table!");
6798 // Create the jump table and associated information.
6799 MachineJumpTableInfo *JTI =
6800 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6801 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6802 unsigned UId = AFI->createJumpTableUId();
6803 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6805 // Create the MBBs for the dispatch code.
6807 // Shove the dispatch's address into the return slot in the function context.
6808 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6809 DispatchBB->setIsLandingPad();
6811 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6812 unsigned trap_opcode;
6813 if (Subtarget->isThumb())
6814 trap_opcode = ARM::tTRAP;
6816 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6818 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6819 DispatchBB->addSuccessor(TrapBB);
6821 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6822 DispatchBB->addSuccessor(DispContBB);
6825 MF->insert(MF->end(), DispatchBB);
6826 MF->insert(MF->end(), DispContBB);
6827 MF->insert(MF->end(), TrapBB);
6829 // Insert code into the entry block that creates and registers the function
6831 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6833 MachineMemOperand *FIMMOLd =
6834 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6835 MachineMemOperand::MOLoad |
6836 MachineMemOperand::MOVolatile, 4, 4);
6838 MachineInstrBuilder MIB;
6839 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6841 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6842 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6844 // Add a register mask with no preserved registers. This results in all
6845 // registers being marked as clobbered.
6846 MIB.addRegMask(RI.getNoPreservedMask());
6848 unsigned NumLPads = LPadList.size();
6849 if (Subtarget->isThumb2()) {
6850 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6851 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6854 .addMemOperand(FIMMOLd));
6856 if (NumLPads < 256) {
6857 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6859 .addImm(LPadList.size()));
6861 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6862 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6863 .addImm(NumLPads & 0xFFFF));
6865 unsigned VReg2 = VReg1;
6866 if ((NumLPads & 0xFFFF0000) != 0) {
6867 VReg2 = MRI->createVirtualRegister(TRC);
6868 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6870 .addImm(NumLPads >> 16));
6873 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6878 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6883 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6884 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6885 .addJumpTableIndex(MJTI)
6888 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6891 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6892 .addReg(NewVReg3, RegState::Kill)
6894 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6896 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6897 .addReg(NewVReg4, RegState::Kill)
6899 .addJumpTableIndex(MJTI)
6901 } else if (Subtarget->isThumb()) {
6902 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6903 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6906 .addMemOperand(FIMMOLd));
6908 if (NumLPads < 256) {
6909 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6913 MachineConstantPool *ConstantPool = MF->getConstantPool();
6914 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6915 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6917 // MachineConstantPool wants an explicit alignment.
6918 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6920 Align = getDataLayout()->getTypeAllocSize(C->getType());
6921 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6923 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6924 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6925 .addReg(VReg1, RegState::Define)
6926 .addConstantPoolIndex(Idx));
6927 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6932 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6937 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6938 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6939 .addReg(ARM::CPSR, RegState::Define)
6943 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6944 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6945 .addJumpTableIndex(MJTI)
6948 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6949 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6950 .addReg(ARM::CPSR, RegState::Define)
6951 .addReg(NewVReg2, RegState::Kill)
6954 MachineMemOperand *JTMMOLd =
6955 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6956 MachineMemOperand::MOLoad, 4, 4);
6958 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6959 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6960 .addReg(NewVReg4, RegState::Kill)
6962 .addMemOperand(JTMMOLd));
6964 unsigned NewVReg6 = NewVReg5;
6965 if (RelocM == Reloc::PIC_) {
6966 NewVReg6 = MRI->createVirtualRegister(TRC);
6967 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6968 .addReg(ARM::CPSR, RegState::Define)
6969 .addReg(NewVReg5, RegState::Kill)
6973 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6974 .addReg(NewVReg6, RegState::Kill)
6975 .addJumpTableIndex(MJTI)
6978 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6979 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6982 .addMemOperand(FIMMOLd));
6984 if (NumLPads < 256) {
6985 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6988 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6989 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6990 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6991 .addImm(NumLPads & 0xFFFF));
6993 unsigned VReg2 = VReg1;
6994 if ((NumLPads & 0xFFFF0000) != 0) {
6995 VReg2 = MRI->createVirtualRegister(TRC);
6996 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6998 .addImm(NumLPads >> 16));
7001 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7005 MachineConstantPool *ConstantPool = MF->getConstantPool();
7006 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7007 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7009 // MachineConstantPool wants an explicit alignment.
7010 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7012 Align = getDataLayout()->getTypeAllocSize(C->getType());
7013 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7015 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7016 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7017 .addReg(VReg1, RegState::Define)
7018 .addConstantPoolIndex(Idx)
7020 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7022 .addReg(VReg1, RegState::Kill));
7025 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7030 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7032 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7034 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7035 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7036 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7037 .addJumpTableIndex(MJTI)
7040 MachineMemOperand *JTMMOLd =
7041 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7042 MachineMemOperand::MOLoad, 4, 4);
7043 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7045 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7046 .addReg(NewVReg3, RegState::Kill)
7049 .addMemOperand(JTMMOLd));
7051 if (RelocM == Reloc::PIC_) {
7052 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7053 .addReg(NewVReg5, RegState::Kill)
7055 .addJumpTableIndex(MJTI)
7058 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7059 .addReg(NewVReg5, RegState::Kill)
7060 .addJumpTableIndex(MJTI)
7065 // Add the jump table entries as successors to the MBB.
7066 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7067 for (std::vector<MachineBasicBlock*>::iterator
7068 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7069 MachineBasicBlock *CurMBB = *I;
7070 if (SeenMBBs.insert(CurMBB))
7071 DispContBB->addSuccessor(CurMBB);
7074 // N.B. the order the invoke BBs are processed in doesn't matter here.
7075 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7076 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7077 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7078 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7079 MachineBasicBlock *BB = *I;
7081 // Remove the landing pad successor from the invoke block and replace it
7082 // with the new dispatch block.
7083 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7085 while (!Successors.empty()) {
7086 MachineBasicBlock *SMBB = Successors.pop_back_val();
7087 if (SMBB->isLandingPad()) {
7088 BB->removeSuccessor(SMBB);
7089 MBBLPads.push_back(SMBB);
7093 BB->addSuccessor(DispatchBB);
7095 // Find the invoke call and mark all of the callee-saved registers as
7096 // 'implicit defined' so that they're spilled. This prevents code from
7097 // moving instructions to before the EH block, where they will never be
7099 for (MachineBasicBlock::reverse_iterator
7100 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7101 if (!II->isCall()) continue;
7103 DenseMap<unsigned, bool> DefRegs;
7104 for (MachineInstr::mop_iterator
7105 OI = II->operands_begin(), OE = II->operands_end();
7107 if (!OI->isReg()) continue;
7108 DefRegs[OI->getReg()] = true;
7111 MachineInstrBuilder MIB(*MF, &*II);
7113 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7114 unsigned Reg = SavedRegs[i];
7115 if (Subtarget->isThumb2() &&
7116 !ARM::tGPRRegClass.contains(Reg) &&
7117 !ARM::hGPRRegClass.contains(Reg))
7119 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7121 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7124 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7131 // Mark all former landing pads as non-landing pads. The dispatch is the only
7133 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7134 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7135 (*I)->setIsLandingPad(false);
7137 // The instruction is gone now.
7138 MI->eraseFromParent();
7144 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7145 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7146 E = MBB->succ_end(); I != E; ++I)
7149 llvm_unreachable("Expecting a BB with two successors!");
7152 MachineBasicBlock *ARMTargetLowering::
7153 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
7154 // This pseudo instruction has 3 operands: dst, src, size
7155 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7156 // Otherwise, we will generate unrolled scalar copies.
7157 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7158 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7159 MachineFunction::iterator It = BB;
7162 unsigned dest = MI->getOperand(0).getReg();
7163 unsigned src = MI->getOperand(1).getReg();
7164 unsigned SizeVal = MI->getOperand(2).getImm();
7165 unsigned Align = MI->getOperand(3).getImm();
7166 DebugLoc dl = MI->getDebugLoc();
7168 bool isThumb2 = Subtarget->isThumb2();
7169 MachineFunction *MF = BB->getParent();
7170 MachineRegisterInfo &MRI = MF->getRegInfo();
7171 unsigned ldrOpc, strOpc, UnitSize = 0;
7173 const TargetRegisterClass *TRC = isThumb2 ?
7174 (const TargetRegisterClass*)&ARM::tGPRRegClass :
7175 (const TargetRegisterClass*)&ARM::GPRRegClass;
7176 const TargetRegisterClass *TRC_Vec = 0;
7179 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7180 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7182 } else if (Align & 2) {
7183 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
7184 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
7187 // Check whether we can use NEON instructions.
7188 if (!MF->getFunction()->getAttributes().
7189 hasAttribute(AttributeSet::FunctionIndex,
7190 Attribute::NoImplicitFloat) &&
7191 Subtarget->hasNEON()) {
7192 if ((Align % 16 == 0) && SizeVal >= 16) {
7193 ldrOpc = ARM::VLD1q32wb_fixed;
7194 strOpc = ARM::VST1q32wb_fixed;
7196 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
7198 else if ((Align % 8 == 0) && SizeVal >= 8) {
7199 ldrOpc = ARM::VLD1d32wb_fixed;
7200 strOpc = ARM::VST1d32wb_fixed;
7202 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
7205 // Can't use NEON instructions.
7206 if (UnitSize == 0) {
7207 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
7208 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
7213 unsigned BytesLeft = SizeVal % UnitSize;
7214 unsigned LoopSize = SizeVal - BytesLeft;
7216 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7217 // Use LDR and STR to copy.
7218 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7219 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7220 unsigned srcIn = src;
7221 unsigned destIn = dest;
7222 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7223 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7224 unsigned srcOut = MRI.createVirtualRegister(TRC);
7225 unsigned destOut = MRI.createVirtualRegister(TRC);
7226 if (UnitSize >= 8) {
7227 AddDefaultPred(BuildMI(*BB, MI, dl,
7228 TII->get(ldrOpc), scratch)
7229 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
7231 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7232 .addReg(destIn).addImm(0).addReg(scratch));
7233 } else if (isThumb2) {
7234 AddDefaultPred(BuildMI(*BB, MI, dl,
7235 TII->get(ldrOpc), scratch)
7236 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
7238 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7239 .addReg(scratch).addReg(destIn)
7242 AddDefaultPred(BuildMI(*BB, MI, dl,
7243 TII->get(ldrOpc), scratch)
7244 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
7247 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7248 .addReg(scratch).addReg(destIn)
7249 .addReg(0).addImm(UnitSize));
7255 // Handle the leftover bytes with LDRB and STRB.
7256 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7257 // [destOut] = STRB_POST(scratch, destIn, 1)
7258 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7259 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7260 for (unsigned i = 0; i < BytesLeft; i++) {
7261 unsigned scratch = MRI.createVirtualRegister(TRC);
7262 unsigned srcOut = MRI.createVirtualRegister(TRC);
7263 unsigned destOut = MRI.createVirtualRegister(TRC);
7265 AddDefaultPred(BuildMI(*BB, MI, dl,
7266 TII->get(ldrOpc),scratch)
7267 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7269 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7270 .addReg(scratch).addReg(destIn)
7271 .addReg(0).addImm(1));
7273 AddDefaultPred(BuildMI(*BB, MI, dl,
7274 TII->get(ldrOpc),scratch)
7275 .addReg(srcOut, RegState::Define).addReg(srcIn)
7276 .addReg(0).addImm(1));
7278 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7279 .addReg(scratch).addReg(destIn)
7280 .addReg(0).addImm(1));
7285 MI->eraseFromParent(); // The instruction is gone now.
7289 // Expand the pseudo op to a loop.
7292 // movw varEnd, # --> with thumb2
7294 // ldrcp varEnd, idx --> without thumb2
7295 // fallthrough --> loopMBB
7297 // PHI varPhi, varEnd, varLoop
7298 // PHI srcPhi, src, srcLoop
7299 // PHI destPhi, dst, destLoop
7300 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7301 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7302 // subs varLoop, varPhi, #UnitSize
7304 // fallthrough --> exitMBB
7306 // epilogue to handle left-over bytes
7307 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7308 // [destOut] = STRB_POST(scratch, destLoop, 1)
7309 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7310 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7311 MF->insert(It, loopMBB);
7312 MF->insert(It, exitMBB);
7314 // Transfer the remainder of BB and its successor edges to exitMBB.
7315 exitMBB->splice(exitMBB->begin(), BB,
7316 llvm::next(MachineBasicBlock::iterator(MI)),
7318 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7320 // Load an immediate to varEnd.
7321 unsigned varEnd = MRI.createVirtualRegister(TRC);
7323 unsigned VReg1 = varEnd;
7324 if ((LoopSize & 0xFFFF0000) != 0)
7325 VReg1 = MRI.createVirtualRegister(TRC);
7326 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
7327 .addImm(LoopSize & 0xFFFF));
7329 if ((LoopSize & 0xFFFF0000) != 0)
7330 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7332 .addImm(LoopSize >> 16));
7334 MachineConstantPool *ConstantPool = MF->getConstantPool();
7335 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7336 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7338 // MachineConstantPool wants an explicit alignment.
7339 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7341 Align = getDataLayout()->getTypeAllocSize(C->getType());
7342 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7344 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
7345 .addReg(varEnd, RegState::Define)
7346 .addConstantPoolIndex(Idx)
7349 BB->addSuccessor(loopMBB);
7351 // Generate the loop body:
7352 // varPhi = PHI(varLoop, varEnd)
7353 // srcPhi = PHI(srcLoop, src)
7354 // destPhi = PHI(destLoop, dst)
7355 MachineBasicBlock *entryBB = BB;
7357 unsigned varLoop = MRI.createVirtualRegister(TRC);
7358 unsigned varPhi = MRI.createVirtualRegister(TRC);
7359 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7360 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7361 unsigned destLoop = MRI.createVirtualRegister(TRC);
7362 unsigned destPhi = MRI.createVirtualRegister(TRC);
7364 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7365 .addReg(varLoop).addMBB(loopMBB)
7366 .addReg(varEnd).addMBB(entryBB);
7367 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7368 .addReg(srcLoop).addMBB(loopMBB)
7369 .addReg(src).addMBB(entryBB);
7370 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7371 .addReg(destLoop).addMBB(loopMBB)
7372 .addReg(dest).addMBB(entryBB);
7374 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7375 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7376 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7377 if (UnitSize >= 8) {
7378 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7379 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
7381 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7382 .addReg(destPhi).addImm(0).addReg(scratch));
7383 } else if (isThumb2) {
7384 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7385 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
7387 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7388 .addReg(scratch).addReg(destPhi)
7391 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7392 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
7395 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7396 .addReg(scratch).addReg(destPhi)
7397 .addReg(0).addImm(UnitSize));
7400 // Decrement loop variable by UnitSize.
7401 MachineInstrBuilder MIB = BuildMI(BB, dl,
7402 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7403 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7404 MIB->getOperand(5).setReg(ARM::CPSR);
7405 MIB->getOperand(5).setIsDef(true);
7407 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7408 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7410 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7411 BB->addSuccessor(loopMBB);
7412 BB->addSuccessor(exitMBB);
7414 // Add epilogue to handle BytesLeft.
7416 MachineInstr *StartOfExit = exitMBB->begin();
7417 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7418 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7420 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7421 // [destOut] = STRB_POST(scratch, destLoop, 1)
7422 unsigned srcIn = srcLoop;
7423 unsigned destIn = destLoop;
7424 for (unsigned i = 0; i < BytesLeft; i++) {
7425 unsigned scratch = MRI.createVirtualRegister(TRC);
7426 unsigned srcOut = MRI.createVirtualRegister(TRC);
7427 unsigned destOut = MRI.createVirtualRegister(TRC);
7429 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7430 TII->get(ldrOpc),scratch)
7431 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7433 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7434 .addReg(scratch).addReg(destIn)
7437 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7438 TII->get(ldrOpc),scratch)
7439 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
7441 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7442 .addReg(scratch).addReg(destIn)
7443 .addReg(0).addImm(1));
7449 MI->eraseFromParent(); // The instruction is gone now.
7454 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7455 MachineBasicBlock *BB) const {
7456 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7457 DebugLoc dl = MI->getDebugLoc();
7458 bool isThumb2 = Subtarget->isThumb2();
7459 switch (MI->getOpcode()) {
7462 llvm_unreachable("Unexpected instr type to insert");
7464 // The Thumb2 pre-indexed stores have the same MI operands, they just
7465 // define them differently in the .td files from the isel patterns, so
7466 // they need pseudos.
7467 case ARM::t2STR_preidx:
7468 MI->setDesc(TII->get(ARM::t2STR_PRE));
7470 case ARM::t2STRB_preidx:
7471 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7473 case ARM::t2STRH_preidx:
7474 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7477 case ARM::STRi_preidx:
7478 case ARM::STRBi_preidx: {
7479 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7480 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7481 // Decode the offset.
7482 unsigned Offset = MI->getOperand(4).getImm();
7483 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7484 Offset = ARM_AM::getAM2Offset(Offset);
7488 MachineMemOperand *MMO = *MI->memoperands_begin();
7489 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7490 .addOperand(MI->getOperand(0)) // Rn_wb
7491 .addOperand(MI->getOperand(1)) // Rt
7492 .addOperand(MI->getOperand(2)) // Rn
7493 .addImm(Offset) // offset (skip GPR==zero_reg)
7494 .addOperand(MI->getOperand(5)) // pred
7495 .addOperand(MI->getOperand(6))
7496 .addMemOperand(MMO);
7497 MI->eraseFromParent();
7500 case ARM::STRr_preidx:
7501 case ARM::STRBr_preidx:
7502 case ARM::STRH_preidx: {
7504 switch (MI->getOpcode()) {
7505 default: llvm_unreachable("unexpected opcode!");
7506 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7507 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7508 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7510 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7511 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7512 MIB.addOperand(MI->getOperand(i));
7513 MI->eraseFromParent();
7516 case ARM::ATOMIC_LOAD_ADD_I8:
7517 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7518 case ARM::ATOMIC_LOAD_ADD_I16:
7519 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7520 case ARM::ATOMIC_LOAD_ADD_I32:
7521 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7523 case ARM::ATOMIC_LOAD_AND_I8:
7524 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7525 case ARM::ATOMIC_LOAD_AND_I16:
7526 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7527 case ARM::ATOMIC_LOAD_AND_I32:
7528 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7530 case ARM::ATOMIC_LOAD_OR_I8:
7531 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7532 case ARM::ATOMIC_LOAD_OR_I16:
7533 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7534 case ARM::ATOMIC_LOAD_OR_I32:
7535 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7537 case ARM::ATOMIC_LOAD_XOR_I8:
7538 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7539 case ARM::ATOMIC_LOAD_XOR_I16:
7540 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7541 case ARM::ATOMIC_LOAD_XOR_I32:
7542 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7544 case ARM::ATOMIC_LOAD_NAND_I8:
7545 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7546 case ARM::ATOMIC_LOAD_NAND_I16:
7547 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7548 case ARM::ATOMIC_LOAD_NAND_I32:
7549 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7551 case ARM::ATOMIC_LOAD_SUB_I8:
7552 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7553 case ARM::ATOMIC_LOAD_SUB_I16:
7554 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7555 case ARM::ATOMIC_LOAD_SUB_I32:
7556 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7558 case ARM::ATOMIC_LOAD_MIN_I8:
7559 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7560 case ARM::ATOMIC_LOAD_MIN_I16:
7561 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7562 case ARM::ATOMIC_LOAD_MIN_I32:
7563 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7565 case ARM::ATOMIC_LOAD_MAX_I8:
7566 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7567 case ARM::ATOMIC_LOAD_MAX_I16:
7568 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7569 case ARM::ATOMIC_LOAD_MAX_I32:
7570 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7572 case ARM::ATOMIC_LOAD_UMIN_I8:
7573 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7574 case ARM::ATOMIC_LOAD_UMIN_I16:
7575 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7576 case ARM::ATOMIC_LOAD_UMIN_I32:
7577 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7579 case ARM::ATOMIC_LOAD_UMAX_I8:
7580 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7581 case ARM::ATOMIC_LOAD_UMAX_I16:
7582 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7583 case ARM::ATOMIC_LOAD_UMAX_I32:
7584 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7586 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7587 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7588 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7590 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7591 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7592 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7595 case ARM::ATOMADD6432:
7596 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7597 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7598 /*NeedsCarry*/ true);
7599 case ARM::ATOMSUB6432:
7600 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7601 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7602 /*NeedsCarry*/ true);
7603 case ARM::ATOMOR6432:
7604 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7605 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7606 case ARM::ATOMXOR6432:
7607 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7608 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7609 case ARM::ATOMAND6432:
7610 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7611 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7612 case ARM::ATOMSWAP6432:
7613 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7614 case ARM::ATOMCMPXCHG6432:
7615 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7616 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7617 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7618 case ARM::ATOMMIN6432:
7619 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7620 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7621 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7622 /*IsMinMax*/ true, ARMCC::LT);
7623 case ARM::ATOMMAX6432:
7624 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7625 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7626 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7627 /*IsMinMax*/ true, ARMCC::GE);
7628 case ARM::ATOMUMIN6432:
7629 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7630 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7631 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7632 /*IsMinMax*/ true, ARMCC::LO);
7633 case ARM::ATOMUMAX6432:
7634 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7635 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7636 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7637 /*IsMinMax*/ true, ARMCC::HS);
7639 case ARM::tMOVCCr_pseudo: {
7640 // To "insert" a SELECT_CC instruction, we actually have to insert the
7641 // diamond control-flow pattern. The incoming instruction knows the
7642 // destination vreg to set, the condition code register to branch on, the
7643 // true/false values to select between, and a branch opcode to use.
7644 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7645 MachineFunction::iterator It = BB;
7651 // cmpTY ccX, r1, r2
7653 // fallthrough --> copy0MBB
7654 MachineBasicBlock *thisMBB = BB;
7655 MachineFunction *F = BB->getParent();
7656 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7657 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7658 F->insert(It, copy0MBB);
7659 F->insert(It, sinkMBB);
7661 // Transfer the remainder of BB and its successor edges to sinkMBB.
7662 sinkMBB->splice(sinkMBB->begin(), BB,
7663 llvm::next(MachineBasicBlock::iterator(MI)),
7665 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7667 BB->addSuccessor(copy0MBB);
7668 BB->addSuccessor(sinkMBB);
7670 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7671 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7674 // %FalseValue = ...
7675 // # fallthrough to sinkMBB
7678 // Update machine-CFG edges
7679 BB->addSuccessor(sinkMBB);
7682 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7685 BuildMI(*BB, BB->begin(), dl,
7686 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7687 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7688 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7690 MI->eraseFromParent(); // The pseudo instruction is gone now.
7695 case ARM::BCCZi64: {
7696 // If there is an unconditional branch to the other successor, remove it.
7697 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7699 // Compare both parts that make up the double comparison separately for
7701 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7703 unsigned LHS1 = MI->getOperand(1).getReg();
7704 unsigned LHS2 = MI->getOperand(2).getReg();
7706 AddDefaultPred(BuildMI(BB, dl,
7707 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7708 .addReg(LHS1).addImm(0));
7709 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7710 .addReg(LHS2).addImm(0)
7711 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7713 unsigned RHS1 = MI->getOperand(3).getReg();
7714 unsigned RHS2 = MI->getOperand(4).getReg();
7715 AddDefaultPred(BuildMI(BB, dl,
7716 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7717 .addReg(LHS1).addReg(RHS1));
7718 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7719 .addReg(LHS2).addReg(RHS2)
7720 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7723 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7724 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7725 if (MI->getOperand(0).getImm() == ARMCC::NE)
7726 std::swap(destMBB, exitMBB);
7728 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7729 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7731 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7733 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7735 MI->eraseFromParent(); // The pseudo instruction is gone now.
7739 case ARM::Int_eh_sjlj_setjmp:
7740 case ARM::Int_eh_sjlj_setjmp_nofp:
7741 case ARM::tInt_eh_sjlj_setjmp:
7742 case ARM::t2Int_eh_sjlj_setjmp:
7743 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7744 EmitSjLjDispatchBlock(MI, BB);
7749 // To insert an ABS instruction, we have to insert the
7750 // diamond control-flow pattern. The incoming instruction knows the
7751 // source vreg to test against 0, the destination vreg to set,
7752 // the condition code register to branch on, the
7753 // true/false values to select between, and a branch opcode to use.
7758 // BCC (branch to SinkBB if V0 >= 0)
7759 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7760 // SinkBB: V1 = PHI(V2, V3)
7761 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7762 MachineFunction::iterator BBI = BB;
7764 MachineFunction *Fn = BB->getParent();
7765 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7766 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7767 Fn->insert(BBI, RSBBB);
7768 Fn->insert(BBI, SinkBB);
7770 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7771 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7772 bool isThumb2 = Subtarget->isThumb2();
7773 MachineRegisterInfo &MRI = Fn->getRegInfo();
7774 // In Thumb mode S must not be specified if source register is the SP or
7775 // PC and if destination register is the SP, so restrict register class
7776 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7777 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7778 (const TargetRegisterClass*)&ARM::GPRRegClass);
7780 // Transfer the remainder of BB and its successor edges to sinkMBB.
7781 SinkBB->splice(SinkBB->begin(), BB,
7782 llvm::next(MachineBasicBlock::iterator(MI)),
7784 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7786 BB->addSuccessor(RSBBB);
7787 BB->addSuccessor(SinkBB);
7789 // fall through to SinkMBB
7790 RSBBB->addSuccessor(SinkBB);
7792 // insert a cmp at the end of BB
7793 AddDefaultPred(BuildMI(BB, dl,
7794 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7795 .addReg(ABSSrcReg).addImm(0));
7797 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7799 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7800 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7802 // insert rsbri in RSBBB
7803 // Note: BCC and rsbri will be converted into predicated rsbmi
7804 // by if-conversion pass
7805 BuildMI(*RSBBB, RSBBB->begin(), dl,
7806 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7807 .addReg(ABSSrcReg, RegState::Kill)
7808 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7810 // insert PHI in SinkBB,
7811 // reuse ABSDstReg to not change uses of ABS instruction
7812 BuildMI(*SinkBB, SinkBB->begin(), dl,
7813 TII->get(ARM::PHI), ABSDstReg)
7814 .addReg(NewRsbDstReg).addMBB(RSBBB)
7815 .addReg(ABSSrcReg).addMBB(BB);
7817 // remove ABS instruction
7818 MI->eraseFromParent();
7820 // return last added BB
7823 case ARM::COPY_STRUCT_BYVAL_I32:
7825 return EmitStructByval(MI, BB);
7829 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7830 SDNode *Node) const {
7831 if (!MI->hasPostISelHook()) {
7832 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7833 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7837 const MCInstrDesc *MCID = &MI->getDesc();
7838 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7839 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7840 // operand is still set to noreg. If needed, set the optional operand's
7841 // register to CPSR, and remove the redundant implicit def.
7843 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7845 // Rename pseudo opcodes.
7846 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7848 const ARMBaseInstrInfo *TII =
7849 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7850 MCID = &TII->get(NewOpc);
7852 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7853 "converted opcode should be the same except for cc_out");
7857 // Add the optional cc_out operand
7858 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7860 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7862 // Any ARM instruction that sets the 's' bit should specify an optional
7863 // "cc_out" operand in the last operand position.
7864 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7865 assert(!NewOpc && "Optional cc_out operand required");
7868 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7869 // since we already have an optional CPSR def.
7870 bool definesCPSR = false;
7871 bool deadCPSR = false;
7872 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7874 const MachineOperand &MO = MI->getOperand(i);
7875 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7879 MI->RemoveOperand(i);
7884 assert(!NewOpc && "Optional cc_out operand required");
7887 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7889 assert(!MI->getOperand(ccOutIdx).getReg() &&
7890 "expect uninitialized optional cc_out operand");
7894 // If this instruction was defined with an optional CPSR def and its dag node
7895 // had a live implicit CPSR def, then activate the optional CPSR def.
7896 MachineOperand &MO = MI->getOperand(ccOutIdx);
7897 MO.setReg(ARM::CPSR);
7901 //===----------------------------------------------------------------------===//
7902 // ARM Optimization Hooks
7903 //===----------------------------------------------------------------------===//
7905 // Helper function that checks if N is a null or all ones constant.
7906 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7907 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7910 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7913 // Return true if N is conditionally 0 or all ones.
7914 // Detects these expressions where cc is an i1 value:
7916 // (select cc 0, y) [AllOnes=0]
7917 // (select cc y, 0) [AllOnes=0]
7918 // (zext cc) [AllOnes=0]
7919 // (sext cc) [AllOnes=0/1]
7920 // (select cc -1, y) [AllOnes=1]
7921 // (select cc y, -1) [AllOnes=1]
7923 // Invert is set when N is the null/all ones constant when CC is false.
7924 // OtherOp is set to the alternative value of N.
7925 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7926 SDValue &CC, bool &Invert,
7928 SelectionDAG &DAG) {
7929 switch (N->getOpcode()) {
7930 default: return false;
7932 CC = N->getOperand(0);
7933 SDValue N1 = N->getOperand(1);
7934 SDValue N2 = N->getOperand(2);
7935 if (isZeroOrAllOnes(N1, AllOnes)) {
7940 if (isZeroOrAllOnes(N2, AllOnes)) {
7947 case ISD::ZERO_EXTEND:
7948 // (zext cc) can never be the all ones value.
7952 case ISD::SIGN_EXTEND: {
7953 EVT VT = N->getValueType(0);
7954 CC = N->getOperand(0);
7955 if (CC.getValueType() != MVT::i1)
7959 // When looking for an AllOnes constant, N is an sext, and the 'other'
7961 OtherOp = DAG.getConstant(0, VT);
7962 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7963 // When looking for a 0 constant, N can be zext or sext.
7964 OtherOp = DAG.getConstant(1, VT);
7966 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7972 // Combine a constant select operand into its use:
7974 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7975 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7976 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7977 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7978 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7980 // The transform is rejected if the select doesn't have a constant operand that
7981 // is null, or all ones when AllOnes is set.
7983 // Also recognize sext/zext from i1:
7985 // (add (zext cc), x) -> (select cc (add x, 1), x)
7986 // (add (sext cc), x) -> (select cc (add x, -1), x)
7988 // These transformations eventually create predicated instructions.
7990 // @param N The node to transform.
7991 // @param Slct The N operand that is a select.
7992 // @param OtherOp The other N operand (x above).
7993 // @param DCI Context.
7994 // @param AllOnes Require the select constant to be all ones instead of null.
7995 // @returns The new node, or SDValue() on failure.
7997 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7998 TargetLowering::DAGCombinerInfo &DCI,
7999 bool AllOnes = false) {
8000 SelectionDAG &DAG = DCI.DAG;
8001 EVT VT = N->getValueType(0);
8002 SDValue NonConstantVal;
8005 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8006 NonConstantVal, DAG))
8009 // Slct is now know to be the desired identity constant when CC is true.
8010 SDValue TrueVal = OtherOp;
8011 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8012 OtherOp, NonConstantVal);
8013 // Unless SwapSelectOps says CC should be false.
8015 std::swap(TrueVal, FalseVal);
8017 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8018 CCOp, TrueVal, FalseVal);
8021 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8023 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8024 TargetLowering::DAGCombinerInfo &DCI) {
8025 SDValue N0 = N->getOperand(0);
8026 SDValue N1 = N->getOperand(1);
8027 if (N0.getNode()->hasOneUse()) {
8028 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8029 if (Result.getNode())
8032 if (N1.getNode()->hasOneUse()) {
8033 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8034 if (Result.getNode())
8040 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8041 // (only after legalization).
8042 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8043 TargetLowering::DAGCombinerInfo &DCI,
8044 const ARMSubtarget *Subtarget) {
8046 // Only perform optimization if after legalize, and if NEON is available. We
8047 // also expected both operands to be BUILD_VECTORs.
8048 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8049 || N0.getOpcode() != ISD::BUILD_VECTOR
8050 || N1.getOpcode() != ISD::BUILD_VECTOR)
8053 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8054 EVT VT = N->getValueType(0);
8055 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8058 // Check that the vector operands are of the right form.
8059 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8060 // operands, where N is the size of the formed vector.
8061 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8062 // index such that we have a pair wise add pattern.
8064 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8065 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8067 SDValue Vec = N0->getOperand(0)->getOperand(0);
8068 SDNode *V = Vec.getNode();
8069 unsigned nextIndex = 0;
8071 // For each operands to the ADD which are BUILD_VECTORs,
8072 // check to see if each of their operands are an EXTRACT_VECTOR with
8073 // the same vector and appropriate index.
8074 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8075 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8076 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8078 SDValue ExtVec0 = N0->getOperand(i);
8079 SDValue ExtVec1 = N1->getOperand(i);
8081 // First operand is the vector, verify its the same.
8082 if (V != ExtVec0->getOperand(0).getNode() ||
8083 V != ExtVec1->getOperand(0).getNode())
8086 // Second is the constant, verify its correct.
8087 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8088 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8090 // For the constant, we want to see all the even or all the odd.
8091 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8092 || C1->getZExtValue() != nextIndex+1)
8101 // Create VPADDL node.
8102 SelectionDAG &DAG = DCI.DAG;
8103 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8105 // Build operand list.
8106 SmallVector<SDValue, 8> Ops;
8107 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8108 TLI.getPointerTy()));
8110 // Input is the vector.
8113 // Get widened type and narrowed type.
8115 unsigned numElem = VT.getVectorNumElements();
8116 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8117 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8118 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8119 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8121 llvm_unreachable("Invalid vector element type for padd optimization.");
8124 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8125 widenType, &Ops[0], Ops.size());
8126 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8129 static SDValue findMUL_LOHI(SDValue V) {
8130 if (V->getOpcode() == ISD::UMUL_LOHI ||
8131 V->getOpcode() == ISD::SMUL_LOHI)
8136 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8137 TargetLowering::DAGCombinerInfo &DCI,
8138 const ARMSubtarget *Subtarget) {
8140 if (Subtarget->isThumb1Only()) return SDValue();
8142 // Only perform the checks after legalize when the pattern is available.
8143 if (DCI.isBeforeLegalize()) return SDValue();
8145 // Look for multiply add opportunities.
8146 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8147 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8148 // a glue link from the first add to the second add.
8149 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8150 // a S/UMLAL instruction.
8153 // \ / \ [no multiline comment]
8159 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8160 SDValue AddcOp0 = AddcNode->getOperand(0);
8161 SDValue AddcOp1 = AddcNode->getOperand(1);
8163 // Check if the two operands are from the same mul_lohi node.
8164 if (AddcOp0.getNode() == AddcOp1.getNode())
8167 assert(AddcNode->getNumValues() == 2 &&
8168 AddcNode->getValueType(0) == MVT::i32 &&
8169 "Expect ADDC with two result values. First: i32");
8171 // Check that we have a glued ADDC node.
8172 if (AddcNode->getValueType(1) != MVT::Glue)
8175 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8176 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8177 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8178 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8179 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8182 // Look for the glued ADDE.
8183 SDNode* AddeNode = AddcNode->getGluedUser();
8184 if (AddeNode == NULL)
8187 // Make sure it is really an ADDE.
8188 if (AddeNode->getOpcode() != ISD::ADDE)
8191 assert(AddeNode->getNumOperands() == 3 &&
8192 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8193 "ADDE node has the wrong inputs");
8195 // Check for the triangle shape.
8196 SDValue AddeOp0 = AddeNode->getOperand(0);
8197 SDValue AddeOp1 = AddeNode->getOperand(1);
8199 // Make sure that the ADDE operands are not coming from the same node.
8200 if (AddeOp0.getNode() == AddeOp1.getNode())
8203 // Find the MUL_LOHI node walking up ADDE's operands.
8204 bool IsLeftOperandMUL = false;
8205 SDValue MULOp = findMUL_LOHI(AddeOp0);
8206 if (MULOp == SDValue())
8207 MULOp = findMUL_LOHI(AddeOp1);
8209 IsLeftOperandMUL = true;
8210 if (MULOp == SDValue())
8213 // Figure out the right opcode.
8214 unsigned Opc = MULOp->getOpcode();
8215 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8217 // Figure out the high and low input values to the MLAL node.
8218 SDValue* HiMul = &MULOp;
8219 SDValue* HiAdd = NULL;
8220 SDValue* LoMul = NULL;
8221 SDValue* LowAdd = NULL;
8223 if (IsLeftOperandMUL)
8229 if (AddcOp0->getOpcode() == Opc) {
8233 if (AddcOp1->getOpcode() == Opc) {
8241 if (LoMul->getNode() != HiMul->getNode())
8244 // Create the merged node.
8245 SelectionDAG &DAG = DCI.DAG;
8247 // Build operand list.
8248 SmallVector<SDValue, 8> Ops;
8249 Ops.push_back(LoMul->getOperand(0));
8250 Ops.push_back(LoMul->getOperand(1));
8251 Ops.push_back(*LowAdd);
8252 Ops.push_back(*HiAdd);
8254 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8255 DAG.getVTList(MVT::i32, MVT::i32),
8256 &Ops[0], Ops.size());
8258 // Replace the ADDs' nodes uses by the MLA node's values.
8259 SDValue HiMLALResult(MLALNode.getNode(), 1);
8260 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8262 SDValue LoMLALResult(MLALNode.getNode(), 0);
8263 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8265 // Return original node to notify the driver to stop replacing.
8266 SDValue resNode(AddcNode, 0);
8270 /// PerformADDCCombine - Target-specific dag combine transform from
8271 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8272 static SDValue PerformADDCCombine(SDNode *N,
8273 TargetLowering::DAGCombinerInfo &DCI,
8274 const ARMSubtarget *Subtarget) {
8276 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8280 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8281 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8282 /// called with the default operands, and if that fails, with commuted
8284 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8285 TargetLowering::DAGCombinerInfo &DCI,
8286 const ARMSubtarget *Subtarget){
8288 // Attempt to create vpaddl for this add.
8289 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8290 if (Result.getNode())
8293 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8294 if (N0.getNode()->hasOneUse()) {
8295 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8296 if (Result.getNode()) return Result;
8301 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8303 static SDValue PerformADDCombine(SDNode *N,
8304 TargetLowering::DAGCombinerInfo &DCI,
8305 const ARMSubtarget *Subtarget) {
8306 SDValue N0 = N->getOperand(0);
8307 SDValue N1 = N->getOperand(1);
8309 // First try with the default operand order.
8310 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8311 if (Result.getNode())
8314 // If that didn't work, try again with the operands commuted.
8315 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8318 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8320 static SDValue PerformSUBCombine(SDNode *N,
8321 TargetLowering::DAGCombinerInfo &DCI) {
8322 SDValue N0 = N->getOperand(0);
8323 SDValue N1 = N->getOperand(1);
8325 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8326 if (N1.getNode()->hasOneUse()) {
8327 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8328 if (Result.getNode()) return Result;
8334 /// PerformVMULCombine
8335 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8336 /// special multiplier accumulator forwarding.
8342 static SDValue PerformVMULCombine(SDNode *N,
8343 TargetLowering::DAGCombinerInfo &DCI,
8344 const ARMSubtarget *Subtarget) {
8345 if (!Subtarget->hasVMLxForwarding())
8348 SelectionDAG &DAG = DCI.DAG;
8349 SDValue N0 = N->getOperand(0);
8350 SDValue N1 = N->getOperand(1);
8351 unsigned Opcode = N0.getOpcode();
8352 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8353 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8354 Opcode = N1.getOpcode();
8355 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8356 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8361 EVT VT = N->getValueType(0);
8363 SDValue N00 = N0->getOperand(0);
8364 SDValue N01 = N0->getOperand(1);
8365 return DAG.getNode(Opcode, DL, VT,
8366 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8367 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8370 static SDValue PerformMULCombine(SDNode *N,
8371 TargetLowering::DAGCombinerInfo &DCI,
8372 const ARMSubtarget *Subtarget) {
8373 SelectionDAG &DAG = DCI.DAG;
8375 if (Subtarget->isThumb1Only())
8378 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8381 EVT VT = N->getValueType(0);
8382 if (VT.is64BitVector() || VT.is128BitVector())
8383 return PerformVMULCombine(N, DCI, Subtarget);
8387 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8391 int64_t MulAmt = C->getSExtValue();
8392 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8394 ShiftAmt = ShiftAmt & (32 - 1);
8395 SDValue V = N->getOperand(0);
8399 MulAmt >>= ShiftAmt;
8402 if (isPowerOf2_32(MulAmt - 1)) {
8403 // (mul x, 2^N + 1) => (add (shl x, N), x)
8404 Res = DAG.getNode(ISD::ADD, DL, VT,
8406 DAG.getNode(ISD::SHL, DL, VT,
8408 DAG.getConstant(Log2_32(MulAmt - 1),
8410 } else if (isPowerOf2_32(MulAmt + 1)) {
8411 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8412 Res = DAG.getNode(ISD::SUB, DL, VT,
8413 DAG.getNode(ISD::SHL, DL, VT,
8415 DAG.getConstant(Log2_32(MulAmt + 1),
8421 uint64_t MulAmtAbs = -MulAmt;
8422 if (isPowerOf2_32(MulAmtAbs + 1)) {
8423 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8424 Res = DAG.getNode(ISD::SUB, DL, VT,
8426 DAG.getNode(ISD::SHL, DL, VT,
8428 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8430 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8431 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8432 Res = DAG.getNode(ISD::ADD, DL, VT,
8434 DAG.getNode(ISD::SHL, DL, VT,
8436 DAG.getConstant(Log2_32(MulAmtAbs-1),
8438 Res = DAG.getNode(ISD::SUB, DL, VT,
8439 DAG.getConstant(0, MVT::i32),Res);
8446 Res = DAG.getNode(ISD::SHL, DL, VT,
8447 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8449 // Do not add new nodes to DAG combiner worklist.
8450 DCI.CombineTo(N, Res, false);
8454 static SDValue PerformANDCombine(SDNode *N,
8455 TargetLowering::DAGCombinerInfo &DCI,
8456 const ARMSubtarget *Subtarget) {
8458 // Attempt to use immediate-form VBIC
8459 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8461 EVT VT = N->getValueType(0);
8462 SelectionDAG &DAG = DCI.DAG;
8464 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8467 APInt SplatBits, SplatUndef;
8468 unsigned SplatBitSize;
8471 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8472 if (SplatBitSize <= 64) {
8474 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8475 SplatUndef.getZExtValue(), SplatBitSize,
8476 DAG, VbicVT, VT.is128BitVector(),
8478 if (Val.getNode()) {
8480 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8481 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8482 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8487 if (!Subtarget->isThumb1Only()) {
8488 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8489 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8490 if (Result.getNode())
8497 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8498 static SDValue PerformORCombine(SDNode *N,
8499 TargetLowering::DAGCombinerInfo &DCI,
8500 const ARMSubtarget *Subtarget) {
8501 // Attempt to use immediate-form VORR
8502 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8504 EVT VT = N->getValueType(0);
8505 SelectionDAG &DAG = DCI.DAG;
8507 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8510 APInt SplatBits, SplatUndef;
8511 unsigned SplatBitSize;
8513 if (BVN && Subtarget->hasNEON() &&
8514 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8515 if (SplatBitSize <= 64) {
8517 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8518 SplatUndef.getZExtValue(), SplatBitSize,
8519 DAG, VorrVT, VT.is128BitVector(),
8521 if (Val.getNode()) {
8523 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8524 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8525 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8530 if (!Subtarget->isThumb1Only()) {
8531 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8532 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8533 if (Result.getNode())
8537 // The code below optimizes (or (and X, Y), Z).
8538 // The AND operand needs to have a single user to make these optimizations
8540 SDValue N0 = N->getOperand(0);
8541 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8543 SDValue N1 = N->getOperand(1);
8545 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8546 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8547 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8549 unsigned SplatBitSize;
8552 APInt SplatBits0, SplatBits1;
8553 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8554 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8555 // Ensure that the second operand of both ands are constants
8556 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8557 HasAnyUndefs) && !HasAnyUndefs) {
8558 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8559 HasAnyUndefs) && !HasAnyUndefs) {
8560 // Ensure that the bit width of the constants are the same and that
8561 // the splat arguments are logical inverses as per the pattern we
8562 // are trying to simplify.
8563 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8564 SplatBits0 == ~SplatBits1) {
8565 // Canonicalize the vector type to make instruction selection
8567 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8568 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8572 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8578 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8581 // BFI is only available on V6T2+
8582 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8586 // 1) or (and A, mask), val => ARMbfi A, val, mask
8587 // iff (val & mask) == val
8589 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8590 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8591 // && mask == ~mask2
8592 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8593 // && ~mask == mask2
8594 // (i.e., copy a bitfield value into another bitfield of the same width)
8599 SDValue N00 = N0.getOperand(0);
8601 // The value and the mask need to be constants so we can verify this is
8602 // actually a bitfield set. If the mask is 0xffff, we can do better
8603 // via a movt instruction, so don't use BFI in that case.
8604 SDValue MaskOp = N0.getOperand(1);
8605 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8608 unsigned Mask = MaskC->getZExtValue();
8612 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8613 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8615 unsigned Val = N1C->getZExtValue();
8616 if ((Val & ~Mask) != Val)
8619 if (ARM::isBitFieldInvertedMask(Mask)) {
8620 Val >>= countTrailingZeros(~Mask);
8622 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8623 DAG.getConstant(Val, MVT::i32),
8624 DAG.getConstant(Mask, MVT::i32));
8626 // Do not add new nodes to DAG combiner worklist.
8627 DCI.CombineTo(N, Res, false);
8630 } else if (N1.getOpcode() == ISD::AND) {
8631 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8632 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8635 unsigned Mask2 = N11C->getZExtValue();
8637 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8639 if (ARM::isBitFieldInvertedMask(Mask) &&
8641 // The pack halfword instruction works better for masks that fit it,
8642 // so use that when it's available.
8643 if (Subtarget->hasT2ExtractPack() &&
8644 (Mask == 0xffff || Mask == 0xffff0000))
8647 unsigned amt = countTrailingZeros(Mask2);
8648 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8649 DAG.getConstant(amt, MVT::i32));
8650 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8651 DAG.getConstant(Mask, MVT::i32));
8652 // Do not add new nodes to DAG combiner worklist.
8653 DCI.CombineTo(N, Res, false);
8655 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8657 // The pack halfword instruction works better for masks that fit it,
8658 // so use that when it's available.
8659 if (Subtarget->hasT2ExtractPack() &&
8660 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8663 unsigned lsb = countTrailingZeros(Mask);
8664 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8665 DAG.getConstant(lsb, MVT::i32));
8666 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8667 DAG.getConstant(Mask2, MVT::i32));
8668 // Do not add new nodes to DAG combiner worklist.
8669 DCI.CombineTo(N, Res, false);
8674 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8675 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8676 ARM::isBitFieldInvertedMask(~Mask)) {
8677 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8678 // where lsb(mask) == #shamt and masked bits of B are known zero.
8679 SDValue ShAmt = N00.getOperand(1);
8680 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8681 unsigned LSB = countTrailingZeros(Mask);
8685 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8686 DAG.getConstant(~Mask, MVT::i32));
8688 // Do not add new nodes to DAG combiner worklist.
8689 DCI.CombineTo(N, Res, false);
8695 static SDValue PerformXORCombine(SDNode *N,
8696 TargetLowering::DAGCombinerInfo &DCI,
8697 const ARMSubtarget *Subtarget) {
8698 EVT VT = N->getValueType(0);
8699 SelectionDAG &DAG = DCI.DAG;
8701 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8704 if (!Subtarget->isThumb1Only()) {
8705 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8706 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8707 if (Result.getNode())
8714 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8715 /// the bits being cleared by the AND are not demanded by the BFI.
8716 static SDValue PerformBFICombine(SDNode *N,
8717 TargetLowering::DAGCombinerInfo &DCI) {
8718 SDValue N1 = N->getOperand(1);
8719 if (N1.getOpcode() == ISD::AND) {
8720 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8723 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8724 unsigned LSB = countTrailingZeros(~InvMask);
8725 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8726 unsigned Mask = (1 << Width)-1;
8727 unsigned Mask2 = N11C->getZExtValue();
8728 if ((Mask & (~Mask2)) == 0)
8729 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8730 N->getOperand(0), N1.getOperand(0),
8736 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8737 /// ARMISD::VMOVRRD.
8738 static SDValue PerformVMOVRRDCombine(SDNode *N,
8739 TargetLowering::DAGCombinerInfo &DCI) {
8740 // vmovrrd(vmovdrr x, y) -> x,y
8741 SDValue InDouble = N->getOperand(0);
8742 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8743 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8745 // vmovrrd(load f64) -> (load i32), (load i32)
8746 SDNode *InNode = InDouble.getNode();
8747 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8748 InNode->getValueType(0) == MVT::f64 &&
8749 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8750 !cast<LoadSDNode>(InNode)->isVolatile()) {
8751 // TODO: Should this be done for non-FrameIndex operands?
8752 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8754 SelectionDAG &DAG = DCI.DAG;
8756 SDValue BasePtr = LD->getBasePtr();
8757 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8758 LD->getPointerInfo(), LD->isVolatile(),
8759 LD->isNonTemporal(), LD->isInvariant(),
8760 LD->getAlignment());
8762 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8763 DAG.getConstant(4, MVT::i32));
8764 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8765 LD->getPointerInfo(), LD->isVolatile(),
8766 LD->isNonTemporal(), LD->isInvariant(),
8767 std::min(4U, LD->getAlignment() / 2));
8769 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8770 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8771 DCI.RemoveFromWorklist(LD);
8779 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8780 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8781 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8782 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8783 SDValue Op0 = N->getOperand(0);
8784 SDValue Op1 = N->getOperand(1);
8785 if (Op0.getOpcode() == ISD::BITCAST)
8786 Op0 = Op0.getOperand(0);
8787 if (Op1.getOpcode() == ISD::BITCAST)
8788 Op1 = Op1.getOperand(0);
8789 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8790 Op0.getNode() == Op1.getNode() &&
8791 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8792 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8793 N->getValueType(0), Op0.getOperand(0));
8797 /// PerformSTORECombine - Target-specific dag combine xforms for
8799 static SDValue PerformSTORECombine(SDNode *N,
8800 TargetLowering::DAGCombinerInfo &DCI) {
8801 StoreSDNode *St = cast<StoreSDNode>(N);
8802 if (St->isVolatile())
8805 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8806 // pack all of the elements in one place. Next, store to memory in fewer
8808 SDValue StVal = St->getValue();
8809 EVT VT = StVal.getValueType();
8810 if (St->isTruncatingStore() && VT.isVector()) {
8811 SelectionDAG &DAG = DCI.DAG;
8812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8813 EVT StVT = St->getMemoryVT();
8814 unsigned NumElems = VT.getVectorNumElements();
8815 assert(StVT != VT && "Cannot truncate to the same type");
8816 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8817 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8819 // From, To sizes and ElemCount must be pow of two
8820 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8822 // We are going to use the original vector elt for storing.
8823 // Accumulated smaller vector elements must be a multiple of the store size.
8824 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8826 unsigned SizeRatio = FromEltSz / ToEltSz;
8827 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8829 // Create a type on which we perform the shuffle.
8830 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8831 NumElems*SizeRatio);
8832 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8835 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8836 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8837 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8839 // Can't shuffle using an illegal type.
8840 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8842 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8843 DAG.getUNDEF(WideVec.getValueType()),
8845 // At this point all of the data is stored at the bottom of the
8846 // register. We now need to save it to mem.
8848 // Find the largest store unit
8849 MVT StoreType = MVT::i8;
8850 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8851 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8852 MVT Tp = (MVT::SimpleValueType)tp;
8853 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8856 // Didn't find a legal store type.
8857 if (!TLI.isTypeLegal(StoreType))
8860 // Bitcast the original vector into a vector of store-size units
8861 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8862 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8863 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8864 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8865 SmallVector<SDValue, 8> Chains;
8866 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8867 TLI.getPointerTy());
8868 SDValue BasePtr = St->getBasePtr();
8870 // Perform one or more big stores into memory.
8871 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8872 for (unsigned I = 0; I < E; I++) {
8873 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8874 StoreType, ShuffWide,
8875 DAG.getIntPtrConstant(I));
8876 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8877 St->getPointerInfo(), St->isVolatile(),
8878 St->isNonTemporal(), St->getAlignment());
8879 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8881 Chains.push_back(Ch);
8883 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8887 if (!ISD::isNormalStore(St))
8890 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8891 // ARM stores of arguments in the same cache line.
8892 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8893 StVal.getNode()->hasOneUse()) {
8894 SelectionDAG &DAG = DCI.DAG;
8896 SDValue BasePtr = St->getBasePtr();
8897 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8898 StVal.getNode()->getOperand(0), BasePtr,
8899 St->getPointerInfo(), St->isVolatile(),
8900 St->isNonTemporal(), St->getAlignment());
8902 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8903 DAG.getConstant(4, MVT::i32));
8904 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8905 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8906 St->isNonTemporal(),
8907 std::min(4U, St->getAlignment() / 2));
8910 if (StVal.getValueType() != MVT::i64 ||
8911 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8914 // Bitcast an i64 store extracted from a vector to f64.
8915 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8916 SelectionDAG &DAG = DCI.DAG;
8918 SDValue IntVec = StVal.getOperand(0);
8919 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8920 IntVec.getValueType().getVectorNumElements());
8921 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8922 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8923 Vec, StVal.getOperand(1));
8925 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8926 // Make the DAGCombiner fold the bitcasts.
8927 DCI.AddToWorklist(Vec.getNode());
8928 DCI.AddToWorklist(ExtElt.getNode());
8929 DCI.AddToWorklist(V.getNode());
8930 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8931 St->getPointerInfo(), St->isVolatile(),
8932 St->isNonTemporal(), St->getAlignment(),
8936 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8937 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8938 /// i64 vector to have f64 elements, since the value can then be loaded
8939 /// directly into a VFP register.
8940 static bool hasNormalLoadOperand(SDNode *N) {
8941 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8942 for (unsigned i = 0; i < NumElts; ++i) {
8943 SDNode *Elt = N->getOperand(i).getNode();
8944 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8950 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8951 /// ISD::BUILD_VECTOR.
8952 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8953 TargetLowering::DAGCombinerInfo &DCI){
8954 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8955 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8956 // into a pair of GPRs, which is fine when the value is used as a scalar,
8957 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8958 SelectionDAG &DAG = DCI.DAG;
8959 if (N->getNumOperands() == 2) {
8960 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8965 // Load i64 elements as f64 values so that type legalization does not split
8966 // them up into i32 values.
8967 EVT VT = N->getValueType(0);
8968 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8971 SmallVector<SDValue, 8> Ops;
8972 unsigned NumElts = VT.getVectorNumElements();
8973 for (unsigned i = 0; i < NumElts; ++i) {
8974 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8976 // Make the DAGCombiner fold the bitcast.
8977 DCI.AddToWorklist(V.getNode());
8979 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8980 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8981 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8984 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8986 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8987 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8988 // At that time, we may have inserted bitcasts from integer to float.
8989 // If these bitcasts have survived DAGCombine, change the lowering of this
8990 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8991 // force to use floating point types.
8993 // Make sure we can change the type of the vector.
8994 // This is possible iff:
8995 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8996 // 1.1. Vector is used only once.
8997 // 1.2. Use is a bit convert to an integer type.
8998 // 2. The size of its operands are 32-bits (64-bits are not legal).
8999 EVT VT = N->getValueType(0);
9000 EVT EltVT = VT.getVectorElementType();
9002 // Check 1.1. and 2.
9003 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9006 // By construction, the input type must be float.
9007 assert(EltVT == MVT::f32 && "Unexpected type!");
9010 SDNode *Use = *N->use_begin();
9011 if (Use->getOpcode() != ISD::BITCAST ||
9012 Use->getValueType(0).isFloatingPoint())
9015 // Check profitability.
9016 // Model is, if more than half of the relevant operands are bitcast from
9017 // i32, turn the build_vector into a sequence of insert_vector_elt.
9018 // Relevant operands are everything that is not statically
9019 // (i.e., at compile time) bitcasted.
9020 unsigned NumOfBitCastedElts = 0;
9021 unsigned NumElts = VT.getVectorNumElements();
9022 unsigned NumOfRelevantElts = NumElts;
9023 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9024 SDValue Elt = N->getOperand(Idx);
9025 if (Elt->getOpcode() == ISD::BITCAST) {
9026 // Assume only bit cast to i32 will go away.
9027 if (Elt->getOperand(0).getValueType() == MVT::i32)
9028 ++NumOfBitCastedElts;
9029 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9030 // Constants are statically casted, thus do not count them as
9031 // relevant operands.
9032 --NumOfRelevantElts;
9035 // Check if more than half of the elements require a non-free bitcast.
9036 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9039 SelectionDAG &DAG = DCI.DAG;
9040 // Create the new vector type.
9041 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9042 // Check if the type is legal.
9043 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9044 if (!TLI.isTypeLegal(VecVT))
9048 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9049 // => BITCAST INSERT_VECTOR_ELT
9050 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9052 SDValue Vec = DAG.getUNDEF(VecVT);
9054 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9055 SDValue V = N->getOperand(Idx);
9056 if (V.getOpcode() == ISD::UNDEF)
9058 if (V.getOpcode() == ISD::BITCAST &&
9059 V->getOperand(0).getValueType() == MVT::i32)
9060 // Fold obvious case.
9061 V = V.getOperand(0);
9063 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9064 // Make the DAGCombiner fold the bitcasts.
9065 DCI.AddToWorklist(V.getNode());
9067 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9068 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9070 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9071 // Make the DAGCombiner fold the bitcasts.
9072 DCI.AddToWorklist(Vec.getNode());
9076 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9077 /// ISD::INSERT_VECTOR_ELT.
9078 static SDValue PerformInsertEltCombine(SDNode *N,
9079 TargetLowering::DAGCombinerInfo &DCI) {
9080 // Bitcast an i64 load inserted into a vector to f64.
9081 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9082 EVT VT = N->getValueType(0);
9083 SDNode *Elt = N->getOperand(1).getNode();
9084 if (VT.getVectorElementType() != MVT::i64 ||
9085 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9088 SelectionDAG &DAG = DCI.DAG;
9090 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9091 VT.getVectorNumElements());
9092 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9093 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9094 // Make the DAGCombiner fold the bitcasts.
9095 DCI.AddToWorklist(Vec.getNode());
9096 DCI.AddToWorklist(V.getNode());
9097 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9098 Vec, V, N->getOperand(2));
9099 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9102 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9103 /// ISD::VECTOR_SHUFFLE.
9104 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9105 // The LLVM shufflevector instruction does not require the shuffle mask
9106 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9107 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9108 // operands do not match the mask length, they are extended by concatenating
9109 // them with undef vectors. That is probably the right thing for other
9110 // targets, but for NEON it is better to concatenate two double-register
9111 // size vector operands into a single quad-register size vector. Do that
9112 // transformation here:
9113 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9114 // shuffle(concat(v1, v2), undef)
9115 SDValue Op0 = N->getOperand(0);
9116 SDValue Op1 = N->getOperand(1);
9117 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9118 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9119 Op0.getNumOperands() != 2 ||
9120 Op1.getNumOperands() != 2)
9122 SDValue Concat0Op1 = Op0.getOperand(1);
9123 SDValue Concat1Op1 = Op1.getOperand(1);
9124 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9125 Concat1Op1.getOpcode() != ISD::UNDEF)
9127 // Skip the transformation if any of the types are illegal.
9128 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9129 EVT VT = N->getValueType(0);
9130 if (!TLI.isTypeLegal(VT) ||
9131 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9132 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9135 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9136 Op0.getOperand(0), Op1.getOperand(0));
9137 // Translate the shuffle mask.
9138 SmallVector<int, 16> NewMask;
9139 unsigned NumElts = VT.getVectorNumElements();
9140 unsigned HalfElts = NumElts/2;
9141 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9142 for (unsigned n = 0; n < NumElts; ++n) {
9143 int MaskElt = SVN->getMaskElt(n);
9145 if (MaskElt < (int)HalfElts)
9147 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9148 NewElt = HalfElts + MaskElt - NumElts;
9149 NewMask.push_back(NewElt);
9151 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9152 DAG.getUNDEF(VT), NewMask.data());
9155 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9156 /// NEON load/store intrinsics to merge base address updates.
9157 static SDValue CombineBaseUpdate(SDNode *N,
9158 TargetLowering::DAGCombinerInfo &DCI) {
9159 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9162 SelectionDAG &DAG = DCI.DAG;
9163 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9164 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9165 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9166 SDValue Addr = N->getOperand(AddrOpIdx);
9168 // Search for a use of the address operand that is an increment.
9169 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9170 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9172 if (User->getOpcode() != ISD::ADD ||
9173 UI.getUse().getResNo() != Addr.getResNo())
9176 // Check that the add is independent of the load/store. Otherwise, folding
9177 // it would create a cycle.
9178 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9181 // Find the new opcode for the updating load/store.
9183 bool isLaneOp = false;
9184 unsigned NewOpc = 0;
9185 unsigned NumVecs = 0;
9187 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9189 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9190 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9192 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9194 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9196 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9198 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9199 NumVecs = 2; isLaneOp = true; break;
9200 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9201 NumVecs = 3; isLaneOp = true; break;
9202 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9203 NumVecs = 4; isLaneOp = true; break;
9204 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9205 NumVecs = 1; isLoad = false; break;
9206 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9207 NumVecs = 2; isLoad = false; break;
9208 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9209 NumVecs = 3; isLoad = false; break;
9210 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9211 NumVecs = 4; isLoad = false; break;
9212 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9213 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9214 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9215 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9216 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9217 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9221 switch (N->getOpcode()) {
9222 default: llvm_unreachable("unexpected opcode for Neon base update");
9223 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9224 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9225 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9229 // Find the size of memory referenced by the load/store.
9232 VecTy = N->getValueType(0);
9234 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9235 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9237 NumBytes /= VecTy.getVectorNumElements();
9239 // If the increment is a constant, it must match the memory ref size.
9240 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9241 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9242 uint64_t IncVal = CInc->getZExtValue();
9243 if (IncVal != NumBytes)
9245 } else if (NumBytes >= 3 * 16) {
9246 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9247 // separate instructions that make it harder to use a non-constant update.
9251 // Create the new updating load/store node.
9253 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9255 for (n = 0; n < NumResultVecs; ++n)
9257 Tys[n++] = MVT::i32;
9258 Tys[n] = MVT::Other;
9259 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9260 SmallVector<SDValue, 8> Ops;
9261 Ops.push_back(N->getOperand(0)); // incoming chain
9262 Ops.push_back(N->getOperand(AddrOpIdx));
9264 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9265 Ops.push_back(N->getOperand(i));
9267 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9268 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9269 Ops.data(), Ops.size(),
9270 MemInt->getMemoryVT(),
9271 MemInt->getMemOperand());
9274 std::vector<SDValue> NewResults;
9275 for (unsigned i = 0; i < NumResultVecs; ++i) {
9276 NewResults.push_back(SDValue(UpdN.getNode(), i));
9278 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9279 DCI.CombineTo(N, NewResults);
9280 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9287 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9288 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9289 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9291 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9292 SelectionDAG &DAG = DCI.DAG;
9293 EVT VT = N->getValueType(0);
9294 // vldN-dup instructions only support 64-bit vectors for N > 1.
9295 if (!VT.is64BitVector())
9298 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9299 SDNode *VLD = N->getOperand(0).getNode();
9300 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9302 unsigned NumVecs = 0;
9303 unsigned NewOpc = 0;
9304 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9305 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9307 NewOpc = ARMISD::VLD2DUP;
9308 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9310 NewOpc = ARMISD::VLD3DUP;
9311 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9313 NewOpc = ARMISD::VLD4DUP;
9318 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9319 // numbers match the load.
9320 unsigned VLDLaneNo =
9321 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9322 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9324 // Ignore uses of the chain result.
9325 if (UI.getUse().getResNo() == NumVecs)
9328 if (User->getOpcode() != ARMISD::VDUPLANE ||
9329 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9333 // Create the vldN-dup node.
9336 for (n = 0; n < NumVecs; ++n)
9338 Tys[n] = MVT::Other;
9339 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9340 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9341 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9342 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9343 Ops, 2, VLDMemInt->getMemoryVT(),
9344 VLDMemInt->getMemOperand());
9347 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9349 unsigned ResNo = UI.getUse().getResNo();
9350 // Ignore uses of the chain result.
9351 if (ResNo == NumVecs)
9354 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9357 // Now the vldN-lane intrinsic is dead except for its chain result.
9358 // Update uses of the chain.
9359 std::vector<SDValue> VLDDupResults;
9360 for (unsigned n = 0; n < NumVecs; ++n)
9361 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9362 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9363 DCI.CombineTo(VLD, VLDDupResults);
9368 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9369 /// ARMISD::VDUPLANE.
9370 static SDValue PerformVDUPLANECombine(SDNode *N,
9371 TargetLowering::DAGCombinerInfo &DCI) {
9372 SDValue Op = N->getOperand(0);
9374 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9375 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9376 if (CombineVLDDUP(N, DCI))
9377 return SDValue(N, 0);
9379 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9380 // redundant. Ignore bit_converts for now; element sizes are checked below.
9381 while (Op.getOpcode() == ISD::BITCAST)
9382 Op = Op.getOperand(0);
9383 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9386 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9387 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9388 // The canonical VMOV for a zero vector uses a 32-bit element size.
9389 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9391 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9393 EVT VT = N->getValueType(0);
9394 if (EltSize > VT.getVectorElementType().getSizeInBits())
9397 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9400 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9401 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9402 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9406 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9408 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9413 APFloat APF = C->getValueAPF();
9414 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9415 != APFloat::opOK || !isExact)
9418 c0 = (I == 0) ? cN : c0;
9419 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9426 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9427 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9428 /// when the VMUL has a constant operand that is a power of 2.
9430 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9431 /// vmul.f32 d16, d17, d16
9432 /// vcvt.s32.f32 d16, d16
9434 /// vcvt.s32.f32 d16, d16, #3
9435 static SDValue PerformVCVTCombine(SDNode *N,
9436 TargetLowering::DAGCombinerInfo &DCI,
9437 const ARMSubtarget *Subtarget) {
9438 SelectionDAG &DAG = DCI.DAG;
9439 SDValue Op = N->getOperand(0);
9441 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9442 Op.getOpcode() != ISD::FMUL)
9446 SDValue N0 = Op->getOperand(0);
9447 SDValue ConstVec = Op->getOperand(1);
9448 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9450 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9451 !isConstVecPow2(ConstVec, isSigned, C))
9454 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9455 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9456 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9457 // These instructions only exist converting from f32 to i32. We can handle
9458 // smaller integers by generating an extra truncate, but larger ones would
9463 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9464 Intrinsic::arm_neon_vcvtfp2fxu;
9465 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9466 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9467 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9468 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9469 DAG.getConstant(Log2_64(C), MVT::i32));
9471 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9472 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9477 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9478 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9479 /// when the VDIV has a constant operand that is a power of 2.
9481 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9482 /// vcvt.f32.s32 d16, d16
9483 /// vdiv.f32 d16, d17, d16
9485 /// vcvt.f32.s32 d16, d16, #3
9486 static SDValue PerformVDIVCombine(SDNode *N,
9487 TargetLowering::DAGCombinerInfo &DCI,
9488 const ARMSubtarget *Subtarget) {
9489 SelectionDAG &DAG = DCI.DAG;
9490 SDValue Op = N->getOperand(0);
9491 unsigned OpOpcode = Op.getNode()->getOpcode();
9493 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9494 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9498 SDValue ConstVec = N->getOperand(1);
9499 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9501 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9502 !isConstVecPow2(ConstVec, isSigned, C))
9505 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9506 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9507 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9508 // These instructions only exist converting from i32 to f32. We can handle
9509 // smaller integers by generating an extra extend, but larger ones would
9514 SDValue ConvInput = Op.getOperand(0);
9515 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9516 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9517 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9518 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9521 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9522 Intrinsic::arm_neon_vcvtfxu2fp;
9523 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9525 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9526 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9529 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9530 /// operand of a vector shift operation, where all the elements of the
9531 /// build_vector must have the same constant integer value.
9532 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9533 // Ignore bit_converts.
9534 while (Op.getOpcode() == ISD::BITCAST)
9535 Op = Op.getOperand(0);
9536 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9537 APInt SplatBits, SplatUndef;
9538 unsigned SplatBitSize;
9540 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9541 HasAnyUndefs, ElementBits) ||
9542 SplatBitSize > ElementBits)
9544 Cnt = SplatBits.getSExtValue();
9548 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9549 /// operand of a vector shift left operation. That value must be in the range:
9550 /// 0 <= Value < ElementBits for a left shift; or
9551 /// 0 <= Value <= ElementBits for a long left shift.
9552 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9553 assert(VT.isVector() && "vector shift count is not a vector type");
9554 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9555 if (! getVShiftImm(Op, ElementBits, Cnt))
9557 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9560 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9561 /// operand of a vector shift right operation. For a shift opcode, the value
9562 /// is positive, but for an intrinsic the value count must be negative. The
9563 /// absolute value must be in the range:
9564 /// 1 <= |Value| <= ElementBits for a right shift; or
9565 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9566 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9568 assert(VT.isVector() && "vector shift count is not a vector type");
9569 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9570 if (! getVShiftImm(Op, ElementBits, Cnt))
9574 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9577 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9578 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9579 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9582 // Don't do anything for most intrinsics.
9585 // Vector shifts: check for immediate versions and lower them.
9586 // Note: This is done during DAG combining instead of DAG legalizing because
9587 // the build_vectors for 64-bit vector element shift counts are generally
9588 // not legal, and it is hard to see their values after they get legalized to
9589 // loads from a constant pool.
9590 case Intrinsic::arm_neon_vshifts:
9591 case Intrinsic::arm_neon_vshiftu:
9592 case Intrinsic::arm_neon_vshiftls:
9593 case Intrinsic::arm_neon_vshiftlu:
9594 case Intrinsic::arm_neon_vshiftn:
9595 case Intrinsic::arm_neon_vrshifts:
9596 case Intrinsic::arm_neon_vrshiftu:
9597 case Intrinsic::arm_neon_vrshiftn:
9598 case Intrinsic::arm_neon_vqshifts:
9599 case Intrinsic::arm_neon_vqshiftu:
9600 case Intrinsic::arm_neon_vqshiftsu:
9601 case Intrinsic::arm_neon_vqshiftns:
9602 case Intrinsic::arm_neon_vqshiftnu:
9603 case Intrinsic::arm_neon_vqshiftnsu:
9604 case Intrinsic::arm_neon_vqrshiftns:
9605 case Intrinsic::arm_neon_vqrshiftnu:
9606 case Intrinsic::arm_neon_vqrshiftnsu: {
9607 EVT VT = N->getOperand(1).getValueType();
9609 unsigned VShiftOpc = 0;
9612 case Intrinsic::arm_neon_vshifts:
9613 case Intrinsic::arm_neon_vshiftu:
9614 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9615 VShiftOpc = ARMISD::VSHL;
9618 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9619 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9620 ARMISD::VSHRs : ARMISD::VSHRu);
9625 case Intrinsic::arm_neon_vshiftls:
9626 case Intrinsic::arm_neon_vshiftlu:
9627 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9629 llvm_unreachable("invalid shift count for vshll intrinsic");
9631 case Intrinsic::arm_neon_vrshifts:
9632 case Intrinsic::arm_neon_vrshiftu:
9633 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9637 case Intrinsic::arm_neon_vqshifts:
9638 case Intrinsic::arm_neon_vqshiftu:
9639 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9643 case Intrinsic::arm_neon_vqshiftsu:
9644 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9646 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9648 case Intrinsic::arm_neon_vshiftn:
9649 case Intrinsic::arm_neon_vrshiftn:
9650 case Intrinsic::arm_neon_vqshiftns:
9651 case Intrinsic::arm_neon_vqshiftnu:
9652 case Intrinsic::arm_neon_vqshiftnsu:
9653 case Intrinsic::arm_neon_vqrshiftns:
9654 case Intrinsic::arm_neon_vqrshiftnu:
9655 case Intrinsic::arm_neon_vqrshiftnsu:
9656 // Narrowing shifts require an immediate right shift.
9657 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9659 llvm_unreachable("invalid shift count for narrowing vector shift "
9663 llvm_unreachable("unhandled vector shift");
9667 case Intrinsic::arm_neon_vshifts:
9668 case Intrinsic::arm_neon_vshiftu:
9669 // Opcode already set above.
9671 case Intrinsic::arm_neon_vshiftls:
9672 case Intrinsic::arm_neon_vshiftlu:
9673 if (Cnt == VT.getVectorElementType().getSizeInBits())
9674 VShiftOpc = ARMISD::VSHLLi;
9676 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9677 ARMISD::VSHLLs : ARMISD::VSHLLu);
9679 case Intrinsic::arm_neon_vshiftn:
9680 VShiftOpc = ARMISD::VSHRN; break;
9681 case Intrinsic::arm_neon_vrshifts:
9682 VShiftOpc = ARMISD::VRSHRs; break;
9683 case Intrinsic::arm_neon_vrshiftu:
9684 VShiftOpc = ARMISD::VRSHRu; break;
9685 case Intrinsic::arm_neon_vrshiftn:
9686 VShiftOpc = ARMISD::VRSHRN; break;
9687 case Intrinsic::arm_neon_vqshifts:
9688 VShiftOpc = ARMISD::VQSHLs; break;
9689 case Intrinsic::arm_neon_vqshiftu:
9690 VShiftOpc = ARMISD::VQSHLu; break;
9691 case Intrinsic::arm_neon_vqshiftsu:
9692 VShiftOpc = ARMISD::VQSHLsu; break;
9693 case Intrinsic::arm_neon_vqshiftns:
9694 VShiftOpc = ARMISD::VQSHRNs; break;
9695 case Intrinsic::arm_neon_vqshiftnu:
9696 VShiftOpc = ARMISD::VQSHRNu; break;
9697 case Intrinsic::arm_neon_vqshiftnsu:
9698 VShiftOpc = ARMISD::VQSHRNsu; break;
9699 case Intrinsic::arm_neon_vqrshiftns:
9700 VShiftOpc = ARMISD::VQRSHRNs; break;
9701 case Intrinsic::arm_neon_vqrshiftnu:
9702 VShiftOpc = ARMISD::VQRSHRNu; break;
9703 case Intrinsic::arm_neon_vqrshiftnsu:
9704 VShiftOpc = ARMISD::VQRSHRNsu; break;
9707 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9708 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9711 case Intrinsic::arm_neon_vshiftins: {
9712 EVT VT = N->getOperand(1).getValueType();
9714 unsigned VShiftOpc = 0;
9716 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9717 VShiftOpc = ARMISD::VSLI;
9718 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9719 VShiftOpc = ARMISD::VSRI;
9721 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9724 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9725 N->getOperand(1), N->getOperand(2),
9726 DAG.getConstant(Cnt, MVT::i32));
9729 case Intrinsic::arm_neon_vqrshifts:
9730 case Intrinsic::arm_neon_vqrshiftu:
9731 // No immediate versions of these to check for.
9738 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9739 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9740 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9741 /// vector element shift counts are generally not legal, and it is hard to see
9742 /// their values after they get legalized to loads from a constant pool.
9743 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9744 const ARMSubtarget *ST) {
9745 EVT VT = N->getValueType(0);
9746 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9747 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9748 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9749 SDValue N1 = N->getOperand(1);
9750 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9751 SDValue N0 = N->getOperand(0);
9752 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9753 DAG.MaskedValueIsZero(N0.getOperand(0),
9754 APInt::getHighBitsSet(32, 16)))
9755 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9759 // Nothing to be done for scalar shifts.
9760 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9761 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9764 assert(ST->hasNEON() && "unexpected vector shift");
9767 switch (N->getOpcode()) {
9768 default: llvm_unreachable("unexpected shift opcode");
9771 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9772 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9773 DAG.getConstant(Cnt, MVT::i32));
9778 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9779 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9780 ARMISD::VSHRs : ARMISD::VSHRu);
9781 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9782 DAG.getConstant(Cnt, MVT::i32));
9788 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9789 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9790 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9791 const ARMSubtarget *ST) {
9792 SDValue N0 = N->getOperand(0);
9794 // Check for sign- and zero-extensions of vector extract operations of 8-
9795 // and 16-bit vector elements. NEON supports these directly. They are
9796 // handled during DAG combining because type legalization will promote them
9797 // to 32-bit types and it is messy to recognize the operations after that.
9798 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9799 SDValue Vec = N0.getOperand(0);
9800 SDValue Lane = N0.getOperand(1);
9801 EVT VT = N->getValueType(0);
9802 EVT EltVT = N0.getValueType();
9803 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9805 if (VT == MVT::i32 &&
9806 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9807 TLI.isTypeLegal(Vec.getValueType()) &&
9808 isa<ConstantSDNode>(Lane)) {
9811 switch (N->getOpcode()) {
9812 default: llvm_unreachable("unexpected opcode");
9813 case ISD::SIGN_EXTEND:
9814 Opc = ARMISD::VGETLANEs;
9816 case ISD::ZERO_EXTEND:
9817 case ISD::ANY_EXTEND:
9818 Opc = ARMISD::VGETLANEu;
9821 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9828 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9829 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9830 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9831 const ARMSubtarget *ST) {
9832 // If the target supports NEON, try to use vmax/vmin instructions for f32
9833 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9834 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9835 // a NaN; only do the transformation when it matches that behavior.
9837 // For now only do this when using NEON for FP operations; if using VFP, it
9838 // is not obvious that the benefit outweighs the cost of switching to the
9840 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9841 N->getValueType(0) != MVT::f32)
9844 SDValue CondLHS = N->getOperand(0);
9845 SDValue CondRHS = N->getOperand(1);
9846 SDValue LHS = N->getOperand(2);
9847 SDValue RHS = N->getOperand(3);
9848 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9850 unsigned Opcode = 0;
9852 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9853 IsReversed = false; // x CC y ? x : y
9854 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9855 IsReversed = true ; // x CC y ? y : x
9869 // If LHS is NaN, an ordered comparison will be false and the result will
9870 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9871 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9872 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9873 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9875 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9876 // will return -0, so vmin can only be used for unsafe math or if one of
9877 // the operands is known to be nonzero.
9878 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9879 !DAG.getTarget().Options.UnsafeFPMath &&
9880 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9882 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9891 // If LHS is NaN, an ordered comparison will be false and the result will
9892 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9893 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9894 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9895 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9897 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9898 // will return +0, so vmax can only be used for unsafe math or if one of
9899 // the operands is known to be nonzero.
9900 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9901 !DAG.getTarget().Options.UnsafeFPMath &&
9902 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9904 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9910 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9913 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9915 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9916 SDValue Cmp = N->getOperand(4);
9917 if (Cmp.getOpcode() != ARMISD::CMPZ)
9918 // Only looking at EQ and NE cases.
9921 EVT VT = N->getValueType(0);
9923 SDValue LHS = Cmp.getOperand(0);
9924 SDValue RHS = Cmp.getOperand(1);
9925 SDValue FalseVal = N->getOperand(0);
9926 SDValue TrueVal = N->getOperand(1);
9927 SDValue ARMcc = N->getOperand(2);
9928 ARMCC::CondCodes CC =
9929 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9947 /// FIXME: Turn this into a target neutral optimization?
9949 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9950 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9951 N->getOperand(3), Cmp);
9952 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9954 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9955 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9956 N->getOperand(3), NewCmp);
9959 if (Res.getNode()) {
9960 APInt KnownZero, KnownOne;
9961 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9962 // Capture demanded bits information that would be otherwise lost.
9963 if (KnownZero == 0xfffffffe)
9964 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9965 DAG.getValueType(MVT::i1));
9966 else if (KnownZero == 0xffffff00)
9967 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9968 DAG.getValueType(MVT::i8));
9969 else if (KnownZero == 0xffff0000)
9970 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9971 DAG.getValueType(MVT::i16));
9977 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9978 DAGCombinerInfo &DCI) const {
9979 switch (N->getOpcode()) {
9981 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9982 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9983 case ISD::SUB: return PerformSUBCombine(N, DCI);
9984 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9985 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9986 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9987 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9988 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9989 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9990 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9991 case ISD::STORE: return PerformSTORECombine(N, DCI);
9992 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9993 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9994 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9995 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9996 case ISD::FP_TO_SINT:
9997 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9998 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9999 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10002 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10003 case ISD::SIGN_EXTEND:
10004 case ISD::ZERO_EXTEND:
10005 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10006 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10007 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10008 case ARMISD::VLD2DUP:
10009 case ARMISD::VLD3DUP:
10010 case ARMISD::VLD4DUP:
10011 return CombineBaseUpdate(N, DCI);
10012 case ARMISD::BUILD_VECTOR:
10013 return PerformARMBUILD_VECTORCombine(N, DCI);
10014 case ISD::INTRINSIC_VOID:
10015 case ISD::INTRINSIC_W_CHAIN:
10016 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10017 case Intrinsic::arm_neon_vld1:
10018 case Intrinsic::arm_neon_vld2:
10019 case Intrinsic::arm_neon_vld3:
10020 case Intrinsic::arm_neon_vld4:
10021 case Intrinsic::arm_neon_vld2lane:
10022 case Intrinsic::arm_neon_vld3lane:
10023 case Intrinsic::arm_neon_vld4lane:
10024 case Intrinsic::arm_neon_vst1:
10025 case Intrinsic::arm_neon_vst2:
10026 case Intrinsic::arm_neon_vst3:
10027 case Intrinsic::arm_neon_vst4:
10028 case Intrinsic::arm_neon_vst2lane:
10029 case Intrinsic::arm_neon_vst3lane:
10030 case Intrinsic::arm_neon_vst4lane:
10031 return CombineBaseUpdate(N, DCI);
10039 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10041 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10044 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
10045 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10046 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10048 switch (VT.getSimpleVT().SimpleTy) {
10054 // Unaligned access can use (for example) LRDB, LRDH, LDR
10055 if (AllowsUnaligned) {
10057 *Fast = Subtarget->hasV7Ops();
10064 // For any little-endian targets with neon, we can support unaligned ld/st
10065 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10066 // A big-endian target may also explictly support unaligned accesses
10067 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10077 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10078 unsigned AlignCheck) {
10079 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10080 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10083 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10084 unsigned DstAlign, unsigned SrcAlign,
10085 bool IsMemset, bool ZeroMemset,
10087 MachineFunction &MF) const {
10088 const Function *F = MF.getFunction();
10090 // See if we can use NEON instructions for this...
10091 if ((!IsMemset || ZeroMemset) &&
10092 Subtarget->hasNEON() &&
10093 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10094 Attribute::NoImplicitFloat)) {
10097 (memOpAlign(SrcAlign, DstAlign, 16) ||
10098 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
10100 } else if (Size >= 8 &&
10101 (memOpAlign(SrcAlign, DstAlign, 8) ||
10102 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
10107 // Lowering to i32/i16 if the size permits.
10110 else if (Size >= 2)
10113 // Let the target-independent logic figure it out.
10117 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10118 if (Val.getOpcode() != ISD::LOAD)
10121 EVT VT1 = Val.getValueType();
10122 if (!VT1.isSimple() || !VT1.isInteger() ||
10123 !VT2.isSimple() || !VT2.isInteger())
10126 switch (VT1.getSimpleVT().SimpleTy) {
10131 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10138 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10139 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10142 if (!isTypeLegal(EVT::getEVT(Ty1)))
10145 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10147 // Assuming the caller doesn't have a zeroext or signext return parameter,
10148 // truncation all the way down to i1 is valid.
10153 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10157 unsigned Scale = 1;
10158 switch (VT.getSimpleVT().SimpleTy) {
10159 default: return false;
10174 if ((V & (Scale - 1)) != 0)
10177 return V == (V & ((1LL << 5) - 1));
10180 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10181 const ARMSubtarget *Subtarget) {
10182 bool isNeg = false;
10188 switch (VT.getSimpleVT().SimpleTy) {
10189 default: return false;
10194 // + imm12 or - imm8
10196 return V == (V & ((1LL << 8) - 1));
10197 return V == (V & ((1LL << 12) - 1));
10200 // Same as ARM mode. FIXME: NEON?
10201 if (!Subtarget->hasVFP2())
10206 return V == (V & ((1LL << 8) - 1));
10210 /// isLegalAddressImmediate - Return true if the integer value can be used
10211 /// as the offset of the target addressing mode for load / store of the
10213 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10214 const ARMSubtarget *Subtarget) {
10218 if (!VT.isSimple())
10221 if (Subtarget->isThumb1Only())
10222 return isLegalT1AddressImmediate(V, VT);
10223 else if (Subtarget->isThumb2())
10224 return isLegalT2AddressImmediate(V, VT, Subtarget);
10229 switch (VT.getSimpleVT().SimpleTy) {
10230 default: return false;
10235 return V == (V & ((1LL << 12) - 1));
10238 return V == (V & ((1LL << 8) - 1));
10241 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10246 return V == (V & ((1LL << 8) - 1));
10250 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10252 int Scale = AM.Scale;
10256 switch (VT.getSimpleVT().SimpleTy) {
10257 default: return false;
10265 Scale = Scale & ~1;
10266 return Scale == 2 || Scale == 4 || Scale == 8;
10269 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10273 // Note, we allow "void" uses (basically, uses that aren't loads or
10274 // stores), because arm allows folding a scale into many arithmetic
10275 // operations. This should be made more precise and revisited later.
10277 // Allow r << imm, but the imm has to be a multiple of two.
10278 if (Scale & 1) return false;
10279 return isPowerOf2_32(Scale);
10283 /// isLegalAddressingMode - Return true if the addressing mode represented
10284 /// by AM is legal for this target, for a load/store of the specified type.
10285 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10287 EVT VT = getValueType(Ty, true);
10288 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10291 // Can never fold addr of global into load/store.
10295 switch (AM.Scale) {
10296 case 0: // no scale reg, must be "r+i" or "r", or "i".
10299 if (Subtarget->isThumb1Only())
10303 // ARM doesn't support any R+R*scale+imm addr modes.
10307 if (!VT.isSimple())
10310 if (Subtarget->isThumb2())
10311 return isLegalT2ScaledAddressingMode(AM, VT);
10313 int Scale = AM.Scale;
10314 switch (VT.getSimpleVT().SimpleTy) {
10315 default: return false;
10319 if (Scale < 0) Scale = -Scale;
10323 return isPowerOf2_32(Scale & ~1);
10327 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10332 // Note, we allow "void" uses (basically, uses that aren't loads or
10333 // stores), because arm allows folding a scale into many arithmetic
10334 // operations. This should be made more precise and revisited later.
10336 // Allow r << imm, but the imm has to be a multiple of two.
10337 if (Scale & 1) return false;
10338 return isPowerOf2_32(Scale);
10344 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10345 /// icmp immediate, that is the target has icmp instructions which can compare
10346 /// a register against the immediate without having to materialize the
10347 /// immediate into a register.
10348 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10349 // Thumb2 and ARM modes can use cmn for negative immediates.
10350 if (!Subtarget->isThumb())
10351 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10352 if (Subtarget->isThumb2())
10353 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10354 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10355 return Imm >= 0 && Imm <= 255;
10358 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10359 /// *or sub* immediate, that is the target has add or sub instructions which can
10360 /// add a register with the immediate without having to materialize the
10361 /// immediate into a register.
10362 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10363 // Same encoding for add/sub, just flip the sign.
10364 int64_t AbsImm = llvm::abs64(Imm);
10365 if (!Subtarget->isThumb())
10366 return ARM_AM::getSOImmVal(AbsImm) != -1;
10367 if (Subtarget->isThumb2())
10368 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10369 // Thumb1 only has 8-bit unsigned immediate.
10370 return AbsImm >= 0 && AbsImm <= 255;
10373 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10374 bool isSEXTLoad, SDValue &Base,
10375 SDValue &Offset, bool &isInc,
10376 SelectionDAG &DAG) {
10377 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10380 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10381 // AddressingMode 3
10382 Base = Ptr->getOperand(0);
10383 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10384 int RHSC = (int)RHS->getZExtValue();
10385 if (RHSC < 0 && RHSC > -256) {
10386 assert(Ptr->getOpcode() == ISD::ADD);
10388 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10392 isInc = (Ptr->getOpcode() == ISD::ADD);
10393 Offset = Ptr->getOperand(1);
10395 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10396 // AddressingMode 2
10397 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10398 int RHSC = (int)RHS->getZExtValue();
10399 if (RHSC < 0 && RHSC > -0x1000) {
10400 assert(Ptr->getOpcode() == ISD::ADD);
10402 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10403 Base = Ptr->getOperand(0);
10408 if (Ptr->getOpcode() == ISD::ADD) {
10410 ARM_AM::ShiftOpc ShOpcVal=
10411 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10412 if (ShOpcVal != ARM_AM::no_shift) {
10413 Base = Ptr->getOperand(1);
10414 Offset = Ptr->getOperand(0);
10416 Base = Ptr->getOperand(0);
10417 Offset = Ptr->getOperand(1);
10422 isInc = (Ptr->getOpcode() == ISD::ADD);
10423 Base = Ptr->getOperand(0);
10424 Offset = Ptr->getOperand(1);
10428 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10432 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10433 bool isSEXTLoad, SDValue &Base,
10434 SDValue &Offset, bool &isInc,
10435 SelectionDAG &DAG) {
10436 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10439 Base = Ptr->getOperand(0);
10440 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10441 int RHSC = (int)RHS->getZExtValue();
10442 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10443 assert(Ptr->getOpcode() == ISD::ADD);
10445 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10447 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10448 isInc = Ptr->getOpcode() == ISD::ADD;
10449 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10457 /// getPreIndexedAddressParts - returns true by value, base pointer and
10458 /// offset pointer and addressing mode by reference if the node's address
10459 /// can be legally represented as pre-indexed load / store address.
10461 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10463 ISD::MemIndexedMode &AM,
10464 SelectionDAG &DAG) const {
10465 if (Subtarget->isThumb1Only())
10470 bool isSEXTLoad = false;
10471 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10472 Ptr = LD->getBasePtr();
10473 VT = LD->getMemoryVT();
10474 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10475 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10476 Ptr = ST->getBasePtr();
10477 VT = ST->getMemoryVT();
10482 bool isLegal = false;
10483 if (Subtarget->isThumb2())
10484 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10485 Offset, isInc, DAG);
10487 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10488 Offset, isInc, DAG);
10492 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10496 /// getPostIndexedAddressParts - returns true by value, base pointer and
10497 /// offset pointer and addressing mode by reference if this node can be
10498 /// combined with a load / store to form a post-indexed load / store.
10499 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10502 ISD::MemIndexedMode &AM,
10503 SelectionDAG &DAG) const {
10504 if (Subtarget->isThumb1Only())
10509 bool isSEXTLoad = false;
10510 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10511 VT = LD->getMemoryVT();
10512 Ptr = LD->getBasePtr();
10513 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10514 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10515 VT = ST->getMemoryVT();
10516 Ptr = ST->getBasePtr();
10521 bool isLegal = false;
10522 if (Subtarget->isThumb2())
10523 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10526 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10532 // Swap base ptr and offset to catch more post-index load / store when
10533 // it's legal. In Thumb2 mode, offset must be an immediate.
10534 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10535 !Subtarget->isThumb2())
10536 std::swap(Base, Offset);
10538 // Post-indexed load / store update the base pointer.
10543 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10547 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10550 const SelectionDAG &DAG,
10551 unsigned Depth) const {
10552 unsigned BitWidth = KnownOne.getBitWidth();
10553 KnownZero = KnownOne = APInt(BitWidth, 0);
10554 switch (Op.getOpcode()) {
10560 // These nodes' second result is a boolean
10561 if (Op.getResNo() == 0)
10563 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10565 case ARMISD::CMOV: {
10566 // Bits are known zero/one if known on the LHS and RHS.
10567 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10568 if (KnownZero == 0 && KnownOne == 0) return;
10570 APInt KnownZeroRHS, KnownOneRHS;
10571 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10572 KnownZero &= KnownZeroRHS;
10573 KnownOne &= KnownOneRHS;
10579 //===----------------------------------------------------------------------===//
10580 // ARM Inline Assembly Support
10581 //===----------------------------------------------------------------------===//
10583 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10584 // Looking for "rev" which is V6+.
10585 if (!Subtarget->hasV6Ops())
10588 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10589 std::string AsmStr = IA->getAsmString();
10590 SmallVector<StringRef, 4> AsmPieces;
10591 SplitString(AsmStr, AsmPieces, ";\n");
10593 switch (AsmPieces.size()) {
10594 default: return false;
10596 AsmStr = AsmPieces[0];
10598 SplitString(AsmStr, AsmPieces, " \t,");
10601 if (AsmPieces.size() == 3 &&
10602 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10603 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10604 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10605 if (Ty && Ty->getBitWidth() == 32)
10606 return IntrinsicLowering::LowerToByteSwap(CI);
10614 /// getConstraintType - Given a constraint letter, return the type of
10615 /// constraint it is for this target.
10616 ARMTargetLowering::ConstraintType
10617 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10618 if (Constraint.size() == 1) {
10619 switch (Constraint[0]) {
10621 case 'l': return C_RegisterClass;
10622 case 'w': return C_RegisterClass;
10623 case 'h': return C_RegisterClass;
10624 case 'x': return C_RegisterClass;
10625 case 't': return C_RegisterClass;
10626 case 'j': return C_Other; // Constant for movw.
10627 // An address with a single base register. Due to the way we
10628 // currently handle addresses it is the same as an 'r' memory constraint.
10629 case 'Q': return C_Memory;
10631 } else if (Constraint.size() == 2) {
10632 switch (Constraint[0]) {
10634 // All 'U+' constraints are addresses.
10635 case 'U': return C_Memory;
10638 return TargetLowering::getConstraintType(Constraint);
10641 /// Examine constraint type and operand type and determine a weight value.
10642 /// This object must already have been set up with the operand type
10643 /// and the current alternative constraint selected.
10644 TargetLowering::ConstraintWeight
10645 ARMTargetLowering::getSingleConstraintMatchWeight(
10646 AsmOperandInfo &info, const char *constraint) const {
10647 ConstraintWeight weight = CW_Invalid;
10648 Value *CallOperandVal = info.CallOperandVal;
10649 // If we don't have a value, we can't do a match,
10650 // but allow it at the lowest weight.
10651 if (CallOperandVal == NULL)
10653 Type *type = CallOperandVal->getType();
10654 // Look at the constraint type.
10655 switch (*constraint) {
10657 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10660 if (type->isIntegerTy()) {
10661 if (Subtarget->isThumb())
10662 weight = CW_SpecificReg;
10664 weight = CW_Register;
10668 if (type->isFloatingPointTy())
10669 weight = CW_Register;
10675 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10677 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10679 if (Constraint.size() == 1) {
10680 // GCC ARM Constraint Letters
10681 switch (Constraint[0]) {
10682 case 'l': // Low regs or general regs.
10683 if (Subtarget->isThumb())
10684 return RCPair(0U, &ARM::tGPRRegClass);
10685 return RCPair(0U, &ARM::GPRRegClass);
10686 case 'h': // High regs or no regs.
10687 if (Subtarget->isThumb())
10688 return RCPair(0U, &ARM::hGPRRegClass);
10691 return RCPair(0U, &ARM::GPRRegClass);
10693 if (VT == MVT::f32)
10694 return RCPair(0U, &ARM::SPRRegClass);
10695 if (VT.getSizeInBits() == 64)
10696 return RCPair(0U, &ARM::DPRRegClass);
10697 if (VT.getSizeInBits() == 128)
10698 return RCPair(0U, &ARM::QPRRegClass);
10701 if (VT == MVT::f32)
10702 return RCPair(0U, &ARM::SPR_8RegClass);
10703 if (VT.getSizeInBits() == 64)
10704 return RCPair(0U, &ARM::DPR_8RegClass);
10705 if (VT.getSizeInBits() == 128)
10706 return RCPair(0U, &ARM::QPR_8RegClass);
10709 if (VT == MVT::f32)
10710 return RCPair(0U, &ARM::SPRRegClass);
10714 if (StringRef("{cc}").equals_lower(Constraint))
10715 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10717 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10720 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10721 /// vector. If it is invalid, don't add anything to Ops.
10722 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10723 std::string &Constraint,
10724 std::vector<SDValue>&Ops,
10725 SelectionDAG &DAG) const {
10726 SDValue Result(0, 0);
10728 // Currently only support length 1 constraints.
10729 if (Constraint.length() != 1) return;
10731 char ConstraintLetter = Constraint[0];
10732 switch (ConstraintLetter) {
10735 case 'I': case 'J': case 'K': case 'L':
10736 case 'M': case 'N': case 'O':
10737 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10741 int64_t CVal64 = C->getSExtValue();
10742 int CVal = (int) CVal64;
10743 // None of these constraints allow values larger than 32 bits. Check
10744 // that the value fits in an int.
10745 if (CVal != CVal64)
10748 switch (ConstraintLetter) {
10750 // Constant suitable for movw, must be between 0 and
10752 if (Subtarget->hasV6T2Ops())
10753 if (CVal >= 0 && CVal <= 65535)
10757 if (Subtarget->isThumb1Only()) {
10758 // This must be a constant between 0 and 255, for ADD
10760 if (CVal >= 0 && CVal <= 255)
10762 } else if (Subtarget->isThumb2()) {
10763 // A constant that can be used as an immediate value in a
10764 // data-processing instruction.
10765 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10768 // A constant that can be used as an immediate value in a
10769 // data-processing instruction.
10770 if (ARM_AM::getSOImmVal(CVal) != -1)
10776 if (Subtarget->isThumb()) { // FIXME thumb2
10777 // This must be a constant between -255 and -1, for negated ADD
10778 // immediates. This can be used in GCC with an "n" modifier that
10779 // prints the negated value, for use with SUB instructions. It is
10780 // not useful otherwise but is implemented for compatibility.
10781 if (CVal >= -255 && CVal <= -1)
10784 // This must be a constant between -4095 and 4095. It is not clear
10785 // what this constraint is intended for. Implemented for
10786 // compatibility with GCC.
10787 if (CVal >= -4095 && CVal <= 4095)
10793 if (Subtarget->isThumb1Only()) {
10794 // A 32-bit value where only one byte has a nonzero value. Exclude
10795 // zero to match GCC. This constraint is used by GCC internally for
10796 // constants that can be loaded with a move/shift combination.
10797 // It is not useful otherwise but is implemented for compatibility.
10798 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10800 } else if (Subtarget->isThumb2()) {
10801 // A constant whose bitwise inverse can be used as an immediate
10802 // value in a data-processing instruction. This can be used in GCC
10803 // with a "B" modifier that prints the inverted value, for use with
10804 // BIC and MVN instructions. It is not useful otherwise but is
10805 // implemented for compatibility.
10806 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10809 // A constant whose bitwise inverse can be used as an immediate
10810 // value in a data-processing instruction. This can be used in GCC
10811 // with a "B" modifier that prints the inverted value, for use with
10812 // BIC and MVN instructions. It is not useful otherwise but is
10813 // implemented for compatibility.
10814 if (ARM_AM::getSOImmVal(~CVal) != -1)
10820 if (Subtarget->isThumb1Only()) {
10821 // This must be a constant between -7 and 7,
10822 // for 3-operand ADD/SUB immediate instructions.
10823 if (CVal >= -7 && CVal < 7)
10825 } else if (Subtarget->isThumb2()) {
10826 // A constant whose negation can be used as an immediate value in a
10827 // data-processing instruction. This can be used in GCC with an "n"
10828 // modifier that prints the negated value, for use with SUB
10829 // instructions. It is not useful otherwise but is implemented for
10831 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10834 // A constant whose negation can be used as an immediate value in a
10835 // data-processing instruction. This can be used in GCC with an "n"
10836 // modifier that prints the negated value, for use with SUB
10837 // instructions. It is not useful otherwise but is implemented for
10839 if (ARM_AM::getSOImmVal(-CVal) != -1)
10845 if (Subtarget->isThumb()) { // FIXME thumb2
10846 // This must be a multiple of 4 between 0 and 1020, for
10847 // ADD sp + immediate.
10848 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10851 // A power of two or a constant between 0 and 32. This is used in
10852 // GCC for the shift amount on shifted register operands, but it is
10853 // useful in general for any shift amounts.
10854 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10860 if (Subtarget->isThumb()) { // FIXME thumb2
10861 // This must be a constant between 0 and 31, for shift amounts.
10862 if (CVal >= 0 && CVal <= 31)
10868 if (Subtarget->isThumb()) { // FIXME thumb2
10869 // This must be a multiple of 4 between -508 and 508, for
10870 // ADD/SUB sp = sp + immediate.
10871 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10876 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10880 if (Result.getNode()) {
10881 Ops.push_back(Result);
10884 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10887 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10888 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10889 unsigned Opcode = Op->getOpcode();
10890 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10891 "Invalid opcode for Div/Rem lowering");
10892 bool isSigned = (Opcode == ISD::SDIVREM);
10893 EVT VT = Op->getValueType(0);
10894 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10897 switch (VT.getSimpleVT().SimpleTy) {
10898 default: llvm_unreachable("Unexpected request for libcall!");
10899 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10900 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10901 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10902 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10905 SDValue InChain = DAG.getEntryNode();
10907 TargetLowering::ArgListTy Args;
10908 TargetLowering::ArgListEntry Entry;
10909 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10910 EVT ArgVT = Op->getOperand(i).getValueType();
10911 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10912 Entry.Node = Op->getOperand(i);
10914 Entry.isSExt = isSigned;
10915 Entry.isZExt = !isSigned;
10916 Args.push_back(Entry);
10919 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10922 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10926 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
10927 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
10928 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
10929 Callee, Args, DAG, dl);
10930 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10932 return CallInfo.first;
10936 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10937 // The ARM target isn't yet aware of offsets.
10941 bool ARM::isBitFieldInvertedMask(unsigned v) {
10942 if (v == 0xffffffff)
10945 // there can be 1's on either or both "outsides", all the "inside"
10946 // bits must be 0's
10947 unsigned TO = CountTrailingOnes_32(v);
10948 unsigned LO = CountLeadingOnes_32(v);
10949 v = (v >> TO) << TO;
10950 v = (v << LO) >> LO;
10954 /// isFPImmLegal - Returns true if the target can instruction select the
10955 /// specified FP immediate natively. If false, the legalizer will
10956 /// materialize the FP immediate as a load from a constant pool.
10957 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10958 if (!Subtarget->hasVFP3())
10960 if (VT == MVT::f32)
10961 return ARM_AM::getFP32Imm(Imm) != -1;
10962 if (VT == MVT::f64)
10963 return ARM_AM::getFP64Imm(Imm) != -1;
10967 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10968 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10969 /// specified in the intrinsic calls.
10970 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10972 unsigned Intrinsic) const {
10973 switch (Intrinsic) {
10974 case Intrinsic::arm_neon_vld1:
10975 case Intrinsic::arm_neon_vld2:
10976 case Intrinsic::arm_neon_vld3:
10977 case Intrinsic::arm_neon_vld4:
10978 case Intrinsic::arm_neon_vld2lane:
10979 case Intrinsic::arm_neon_vld3lane:
10980 case Intrinsic::arm_neon_vld4lane: {
10981 Info.opc = ISD::INTRINSIC_W_CHAIN;
10982 // Conservatively set memVT to the entire set of vectors loaded.
10983 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10984 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10985 Info.ptrVal = I.getArgOperand(0);
10987 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10988 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10989 Info.vol = false; // volatile loads with NEON intrinsics not supported
10990 Info.readMem = true;
10991 Info.writeMem = false;
10994 case Intrinsic::arm_neon_vst1:
10995 case Intrinsic::arm_neon_vst2:
10996 case Intrinsic::arm_neon_vst3:
10997 case Intrinsic::arm_neon_vst4:
10998 case Intrinsic::arm_neon_vst2lane:
10999 case Intrinsic::arm_neon_vst3lane:
11000 case Intrinsic::arm_neon_vst4lane: {
11001 Info.opc = ISD::INTRINSIC_VOID;
11002 // Conservatively set memVT to the entire set of vectors stored.
11003 unsigned NumElts = 0;
11004 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11005 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11006 if (!ArgTy->isVectorTy())
11008 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11010 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11011 Info.ptrVal = I.getArgOperand(0);
11013 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11014 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11015 Info.vol = false; // volatile stores with NEON intrinsics not supported
11016 Info.readMem = false;
11017 Info.writeMem = true;
11020 case Intrinsic::arm_ldrex: {
11021 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11022 Info.opc = ISD::INTRINSIC_W_CHAIN;
11023 Info.memVT = MVT::getVT(PtrTy->getElementType());
11024 Info.ptrVal = I.getArgOperand(0);
11026 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11028 Info.readMem = true;
11029 Info.writeMem = false;
11032 case Intrinsic::arm_strex: {
11033 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11034 Info.opc = ISD::INTRINSIC_W_CHAIN;
11035 Info.memVT = MVT::getVT(PtrTy->getElementType());
11036 Info.ptrVal = I.getArgOperand(1);
11038 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11040 Info.readMem = false;
11041 Info.writeMem = true;
11044 case Intrinsic::arm_strexd: {
11045 Info.opc = ISD::INTRINSIC_W_CHAIN;
11046 Info.memVT = MVT::i64;
11047 Info.ptrVal = I.getArgOperand(2);
11051 Info.readMem = false;
11052 Info.writeMem = true;
11055 case Intrinsic::arm_ldrexd: {
11056 Info.opc = ISD::INTRINSIC_W_CHAIN;
11057 Info.memVT = MVT::i64;
11058 Info.ptrVal = I.getArgOperand(0);
11062 Info.readMem = true;
11063 Info.writeMem = false;