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::BFI: return "ARMISD::BFI";
1073 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1074 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1075 case ARMISD::VBSL: return "ARMISD::VBSL";
1076 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1077 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1078 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1079 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1080 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1081 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1082 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1083 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1084 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1085 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1086 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1087 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1088 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1089 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1090 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1091 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1092 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1093 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1094 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1095 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1097 case ARMISD::ATOMADD64_DAG: return "ATOMADD64_DAG";
1098 case ARMISD::ATOMSUB64_DAG: return "ATOMSUB64_DAG";
1099 case ARMISD::ATOMOR64_DAG: return "ATOMOR64_DAG";
1100 case ARMISD::ATOMXOR64_DAG: return "ATOMXOR64_DAG";
1101 case ARMISD::ATOMAND64_DAG: return "ATOMAND64_DAG";
1102 case ARMISD::ATOMNAND64_DAG: return "ATOMNAND64_DAG";
1103 case ARMISD::ATOMSWAP64_DAG: return "ATOMSWAP64_DAG";
1104 case ARMISD::ATOMCMPXCHG64_DAG: return "ATOMCMPXCHG64_DAG";
1105 case ARMISD::ATOMMIN64_DAG: return "ATOMMIN64_DAG";
1106 case ARMISD::ATOMUMIN64_DAG: return "ATOMUMIN64_DAG";
1107 case ARMISD::ATOMMAX64_DAG: return "ATOMMAX64_DAG";
1108 case ARMISD::ATOMUMAX64_DAG: return "ATOMUMAX64_DAG";
1112 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1113 if (!VT.isVector()) return getPointerTy();
1114 return VT.changeVectorElementTypeToInteger();
1117 /// getRegClassFor - Return the register class that should be used for the
1118 /// specified value type.
1119 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1120 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1121 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1122 // load / store 4 to 8 consecutive D registers.
1123 if (Subtarget->hasNEON()) {
1124 if (VT == MVT::v4i64)
1125 return &ARM::QQPRRegClass;
1126 if (VT == MVT::v8i64)
1127 return &ARM::QQQQPRRegClass;
1129 return TargetLowering::getRegClassFor(VT);
1132 // Create a fast isel object.
1134 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1135 const TargetLibraryInfo *libInfo) const {
1136 return ARM::createFastISel(funcInfo, libInfo);
1139 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1140 /// be used for loads / stores from the global.
1141 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1142 return (Subtarget->isThumb1Only() ? 127 : 4095);
1145 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1146 unsigned NumVals = N->getNumValues();
1148 return Sched::RegPressure;
1150 for (unsigned i = 0; i != NumVals; ++i) {
1151 EVT VT = N->getValueType(i);
1152 if (VT == MVT::Glue || VT == MVT::Other)
1154 if (VT.isFloatingPoint() || VT.isVector())
1158 if (!N->isMachineOpcode())
1159 return Sched::RegPressure;
1161 // Load are scheduled for latency even if there instruction itinerary
1162 // is not available.
1163 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1164 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1166 if (MCID.getNumDefs() == 0)
1167 return Sched::RegPressure;
1168 if (!Itins->isEmpty() &&
1169 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1172 return Sched::RegPressure;
1175 //===----------------------------------------------------------------------===//
1177 //===----------------------------------------------------------------------===//
1179 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1180 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1182 default: llvm_unreachable("Unknown condition code!");
1183 case ISD::SETNE: return ARMCC::NE;
1184 case ISD::SETEQ: return ARMCC::EQ;
1185 case ISD::SETGT: return ARMCC::GT;
1186 case ISD::SETGE: return ARMCC::GE;
1187 case ISD::SETLT: return ARMCC::LT;
1188 case ISD::SETLE: return ARMCC::LE;
1189 case ISD::SETUGT: return ARMCC::HI;
1190 case ISD::SETUGE: return ARMCC::HS;
1191 case ISD::SETULT: return ARMCC::LO;
1192 case ISD::SETULE: return ARMCC::LS;
1196 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1197 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1198 ARMCC::CondCodes &CondCode2) {
1199 CondCode2 = ARMCC::AL;
1201 default: llvm_unreachable("Unknown FP condition!");
1203 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1205 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1207 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1208 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1209 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1210 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1211 case ISD::SETO: CondCode = ARMCC::VC; break;
1212 case ISD::SETUO: CondCode = ARMCC::VS; break;
1213 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1214 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1215 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1217 case ISD::SETULT: CondCode = ARMCC::LT; break;
1219 case ISD::SETULE: CondCode = ARMCC::LE; break;
1221 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1225 //===----------------------------------------------------------------------===//
1226 // Calling Convention Implementation
1227 //===----------------------------------------------------------------------===//
1229 #include "ARMGenCallingConv.inc"
1231 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1232 /// given CallingConvention value.
1233 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1235 bool isVarArg) const {
1238 llvm_unreachable("Unsupported calling convention");
1239 case CallingConv::Fast:
1240 if (Subtarget->hasVFP2() && !isVarArg) {
1241 if (!Subtarget->isAAPCS_ABI())
1242 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1243 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1244 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1247 case CallingConv::C: {
1248 // Use target triple & subtarget features to do actual dispatch.
1249 if (!Subtarget->isAAPCS_ABI())
1250 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1251 else if (Subtarget->hasVFP2() &&
1252 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1254 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1255 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1257 case CallingConv::ARM_AAPCS_VFP:
1259 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1261 case CallingConv::ARM_AAPCS:
1262 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1263 case CallingConv::ARM_APCS:
1264 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1265 case CallingConv::GHC:
1266 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1270 /// LowerCallResult - Lower the result values of a call into the
1271 /// appropriate copies out of appropriate physical registers.
1273 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1274 CallingConv::ID CallConv, bool isVarArg,
1275 const SmallVectorImpl<ISD::InputArg> &Ins,
1276 SDLoc dl, SelectionDAG &DAG,
1277 SmallVectorImpl<SDValue> &InVals,
1278 bool isThisReturn, SDValue ThisVal) const {
1280 // Assign locations to each value returned by this call.
1281 SmallVector<CCValAssign, 16> RVLocs;
1282 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1283 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1284 CCInfo.AnalyzeCallResult(Ins,
1285 CCAssignFnForNode(CallConv, /* Return*/ true,
1288 // Copy all of the result registers out of their specified physreg.
1289 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1290 CCValAssign VA = RVLocs[i];
1292 // Pass 'this' value directly from the argument to return value, to avoid
1293 // reg unit interference
1294 if (i == 0 && isThisReturn) {
1295 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1296 "unexpected return calling convention register assignment");
1297 InVals.push_back(ThisVal);
1302 if (VA.needsCustom()) {
1303 // Handle f64 or half of a v2f64.
1304 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1306 Chain = Lo.getValue(1);
1307 InFlag = Lo.getValue(2);
1308 VA = RVLocs[++i]; // skip ahead to next loc
1309 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1311 Chain = Hi.getValue(1);
1312 InFlag = Hi.getValue(2);
1313 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1315 if (VA.getLocVT() == MVT::v2f64) {
1316 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1317 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1318 DAG.getConstant(0, MVT::i32));
1320 VA = RVLocs[++i]; // skip ahead to next loc
1321 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1322 Chain = Lo.getValue(1);
1323 InFlag = Lo.getValue(2);
1324 VA = RVLocs[++i]; // skip ahead to next loc
1325 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1326 Chain = Hi.getValue(1);
1327 InFlag = Hi.getValue(2);
1328 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1329 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1330 DAG.getConstant(1, MVT::i32));
1333 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1335 Chain = Val.getValue(1);
1336 InFlag = Val.getValue(2);
1339 switch (VA.getLocInfo()) {
1340 default: llvm_unreachable("Unknown loc info!");
1341 case CCValAssign::Full: break;
1342 case CCValAssign::BCvt:
1343 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1347 InVals.push_back(Val);
1353 /// LowerMemOpCallTo - Store the argument to the stack.
1355 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1356 SDValue StackPtr, SDValue Arg,
1357 SDLoc dl, SelectionDAG &DAG,
1358 const CCValAssign &VA,
1359 ISD::ArgFlagsTy Flags) const {
1360 unsigned LocMemOffset = VA.getLocMemOffset();
1361 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1362 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1363 return DAG.getStore(Chain, dl, Arg, PtrOff,
1364 MachinePointerInfo::getStack(LocMemOffset),
1368 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1369 SDValue Chain, SDValue &Arg,
1370 RegsToPassVector &RegsToPass,
1371 CCValAssign &VA, CCValAssign &NextVA,
1373 SmallVectorImpl<SDValue> &MemOpChains,
1374 ISD::ArgFlagsTy Flags) const {
1376 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1377 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1378 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1380 if (NextVA.isRegLoc())
1381 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1383 assert(NextVA.isMemLoc());
1384 if (StackPtr.getNode() == 0)
1385 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1387 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1393 /// LowerCall - Lowering a call into a callseq_start <-
1394 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1397 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1398 SmallVectorImpl<SDValue> &InVals) const {
1399 SelectionDAG &DAG = CLI.DAG;
1401 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1402 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1403 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1404 SDValue Chain = CLI.Chain;
1405 SDValue Callee = CLI.Callee;
1406 bool &isTailCall = CLI.IsTailCall;
1407 CallingConv::ID CallConv = CLI.CallConv;
1408 bool doesNotRet = CLI.DoesNotReturn;
1409 bool isVarArg = CLI.IsVarArg;
1411 MachineFunction &MF = DAG.getMachineFunction();
1412 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1413 bool isThisReturn = false;
1414 bool isSibCall = false;
1415 // Disable tail calls if they're not supported.
1416 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1419 // Check if it's really possible to do a tail call.
1420 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1421 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1422 Outs, OutVals, Ins, DAG);
1423 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1424 // detected sibcalls.
1431 // Analyze operands of the call, assigning locations to each operand.
1432 SmallVector<CCValAssign, 16> ArgLocs;
1433 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1434 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1435 CCInfo.AnalyzeCallOperands(Outs,
1436 CCAssignFnForNode(CallConv, /* Return*/ false,
1439 // Get a count of how many bytes are to be pushed on the stack.
1440 unsigned NumBytes = CCInfo.getNextStackOffset();
1442 // For tail calls, memory operands are available in our caller's stack.
1446 // Adjust the stack pointer for the new arguments...
1447 // These operations are automatically eliminated by the prolog/epilog pass
1449 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1452 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1454 RegsToPassVector RegsToPass;
1455 SmallVector<SDValue, 8> MemOpChains;
1457 // Walk the register/memloc assignments, inserting copies/loads. In the case
1458 // of tail call optimization, arguments are handled later.
1459 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1461 ++i, ++realArgIdx) {
1462 CCValAssign &VA = ArgLocs[i];
1463 SDValue Arg = OutVals[realArgIdx];
1464 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1465 bool isByVal = Flags.isByVal();
1467 // Promote the value if needed.
1468 switch (VA.getLocInfo()) {
1469 default: llvm_unreachable("Unknown loc info!");
1470 case CCValAssign::Full: break;
1471 case CCValAssign::SExt:
1472 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1474 case CCValAssign::ZExt:
1475 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1477 case CCValAssign::AExt:
1478 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1480 case CCValAssign::BCvt:
1481 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1485 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1486 if (VA.needsCustom()) {
1487 if (VA.getLocVT() == MVT::v2f64) {
1488 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1489 DAG.getConstant(0, MVT::i32));
1490 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1491 DAG.getConstant(1, MVT::i32));
1493 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1494 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1496 VA = ArgLocs[++i]; // skip ahead to next loc
1497 if (VA.isRegLoc()) {
1498 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1499 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1501 assert(VA.isMemLoc());
1503 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1504 dl, DAG, VA, Flags));
1507 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1508 StackPtr, MemOpChains, Flags);
1510 } else if (VA.isRegLoc()) {
1511 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1512 assert(VA.getLocVT() == MVT::i32 &&
1513 "unexpected calling convention register assignment");
1514 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1515 "unexpected use of 'returned'");
1516 isThisReturn = true;
1518 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1519 } else if (isByVal) {
1520 assert(VA.isMemLoc());
1521 unsigned offset = 0;
1523 // True if this byval aggregate will be split between registers
1525 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1526 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1528 if (CurByValIdx < ByValArgsCount) {
1530 unsigned RegBegin, RegEnd;
1531 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1533 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1535 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1536 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1537 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1538 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1539 MachinePointerInfo(),
1540 false, false, false, 0);
1541 MemOpChains.push_back(Load.getValue(1));
1542 RegsToPass.push_back(std::make_pair(j, Load));
1545 // If parameter size outsides register area, "offset" value
1546 // helps us to calculate stack slot for remained part properly.
1547 offset = RegEnd - RegBegin;
1549 CCInfo.nextInRegsParam();
1552 if (Flags.getByValSize() > 4*offset) {
1553 unsigned LocMemOffset = VA.getLocMemOffset();
1554 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1555 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1557 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1558 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1559 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1561 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1563 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1564 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1565 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1566 Ops, array_lengthof(Ops)));
1568 } else if (!isSibCall) {
1569 assert(VA.isMemLoc());
1571 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1572 dl, DAG, VA, Flags));
1576 if (!MemOpChains.empty())
1577 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1578 &MemOpChains[0], MemOpChains.size());
1580 // Build a sequence of copy-to-reg nodes chained together with token chain
1581 // and flag operands which copy the outgoing args into the appropriate regs.
1583 // Tail call byval lowering might overwrite argument registers so in case of
1584 // tail call optimization the copies to registers are lowered later.
1586 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1587 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1588 RegsToPass[i].second, InFlag);
1589 InFlag = Chain.getValue(1);
1592 // For tail calls lower the arguments to the 'real' stack slot.
1594 // Force all the incoming stack arguments to be loaded from the stack
1595 // before any new outgoing arguments are stored to the stack, because the
1596 // outgoing stack slots may alias the incoming argument stack slots, and
1597 // the alias isn't otherwise explicit. This is slightly more conservative
1598 // than necessary, because it means that each store effectively depends
1599 // on every argument instead of just those arguments it would clobber.
1601 // Do not flag preceding copytoreg stuff together with the following stuff.
1603 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1604 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1605 RegsToPass[i].second, InFlag);
1606 InFlag = Chain.getValue(1);
1611 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1612 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1613 // node so that legalize doesn't hack it.
1614 bool isDirect = false;
1615 bool isARMFunc = false;
1616 bool isLocalARMFunc = false;
1617 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1619 if (EnableARMLongCalls) {
1620 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1621 && "long-calls with non-static relocation model!");
1622 // Handle a global address or an external symbol. If it's not one of
1623 // those, the target's already in a register, so we don't need to do
1625 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1626 const GlobalValue *GV = G->getGlobal();
1627 // Create a constant pool entry for the callee address
1628 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1629 ARMConstantPoolValue *CPV =
1630 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1632 // Get the address of the callee into a register
1633 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1634 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1635 Callee = DAG.getLoad(getPointerTy(), dl,
1636 DAG.getEntryNode(), CPAddr,
1637 MachinePointerInfo::getConstantPool(),
1638 false, false, false, 0);
1639 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1640 const char *Sym = S->getSymbol();
1642 // Create a constant pool entry for the callee address
1643 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1644 ARMConstantPoolValue *CPV =
1645 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1646 ARMPCLabelIndex, 0);
1647 // Get the address of the callee into a register
1648 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1649 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1650 Callee = DAG.getLoad(getPointerTy(), dl,
1651 DAG.getEntryNode(), CPAddr,
1652 MachinePointerInfo::getConstantPool(),
1653 false, false, false, 0);
1655 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1656 const GlobalValue *GV = G->getGlobal();
1658 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1659 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1660 getTargetMachine().getRelocationModel() != Reloc::Static;
1661 isARMFunc = !Subtarget->isThumb() || isStub;
1662 // ARM call to a local ARM function is predicable.
1663 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1664 // tBX takes a register source operand.
1665 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1666 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1667 ARMConstantPoolValue *CPV =
1668 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1669 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1670 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1671 Callee = DAG.getLoad(getPointerTy(), dl,
1672 DAG.getEntryNode(), CPAddr,
1673 MachinePointerInfo::getConstantPool(),
1674 false, false, false, 0);
1675 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1676 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1677 getPointerTy(), Callee, PICLabel);
1679 // On ELF targets for PIC code, direct calls should go through the PLT
1680 unsigned OpFlags = 0;
1681 if (Subtarget->isTargetELF() &&
1682 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1683 OpFlags = ARMII::MO_PLT;
1684 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1686 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1688 bool isStub = Subtarget->isTargetDarwin() &&
1689 getTargetMachine().getRelocationModel() != Reloc::Static;
1690 isARMFunc = !Subtarget->isThumb() || isStub;
1691 // tBX takes a register source operand.
1692 const char *Sym = S->getSymbol();
1693 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1694 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1695 ARMConstantPoolValue *CPV =
1696 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1697 ARMPCLabelIndex, 4);
1698 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1699 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1700 Callee = DAG.getLoad(getPointerTy(), dl,
1701 DAG.getEntryNode(), CPAddr,
1702 MachinePointerInfo::getConstantPool(),
1703 false, false, false, 0);
1704 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1705 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1706 getPointerTy(), Callee, PICLabel);
1708 unsigned OpFlags = 0;
1709 // On ELF targets for PIC code, direct calls should go through the PLT
1710 if (Subtarget->isTargetELF() &&
1711 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1712 OpFlags = ARMII::MO_PLT;
1713 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1717 // FIXME: handle tail calls differently.
1719 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1720 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1721 if (Subtarget->isThumb()) {
1722 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1723 CallOpc = ARMISD::CALL_NOLINK;
1725 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1727 if (!isDirect && !Subtarget->hasV5TOps())
1728 CallOpc = ARMISD::CALL_NOLINK;
1729 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1730 // Emit regular call when code size is the priority
1732 // "mov lr, pc; b _foo" to avoid confusing the RSP
1733 CallOpc = ARMISD::CALL_NOLINK;
1735 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1738 std::vector<SDValue> Ops;
1739 Ops.push_back(Chain);
1740 Ops.push_back(Callee);
1742 // Add argument registers to the end of the list so that they are known live
1744 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1745 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1746 RegsToPass[i].second.getValueType()));
1748 // Add a register mask operand representing the call-preserved registers.
1749 const uint32_t *Mask;
1750 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1751 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1753 // For 'this' returns, use the R0-preserving mask if applicable
1754 Mask = ARI->getThisReturnPreservedMask(CallConv);
1756 // Set isThisReturn to false if the calling convention is not one that
1757 // allows 'returned' to be modeled in this way, so LowerCallResult does
1758 // not try to pass 'this' straight through
1759 isThisReturn = false;
1760 Mask = ARI->getCallPreservedMask(CallConv);
1763 Mask = ARI->getCallPreservedMask(CallConv);
1765 assert(Mask && "Missing call preserved mask for calling convention");
1766 Ops.push_back(DAG.getRegisterMask(Mask));
1768 if (InFlag.getNode())
1769 Ops.push_back(InFlag);
1771 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1773 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1775 // Returns a chain and a flag for retval copy to use.
1776 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1777 InFlag = Chain.getValue(1);
1779 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1780 DAG.getIntPtrConstant(0, true), InFlag, dl);
1782 InFlag = Chain.getValue(1);
1784 // Handle result values, copying them out of physregs into vregs that we
1786 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1787 InVals, isThisReturn,
1788 isThisReturn ? OutVals[0] : SDValue());
1791 /// HandleByVal - Every parameter *after* a byval parameter is passed
1792 /// on the stack. Remember the next parameter register to allocate,
1793 /// and then confiscate the rest of the parameter registers to insure
1796 ARMTargetLowering::HandleByVal(
1797 CCState *State, unsigned &size, unsigned Align) const {
1798 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1799 assert((State->getCallOrPrologue() == Prologue ||
1800 State->getCallOrPrologue() == Call) &&
1801 "unhandled ParmContext");
1803 // For in-prologue parameters handling, we also introduce stack offset
1804 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
1805 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
1806 // NSAA should be evaluted (NSAA means "next stacked argument address").
1807 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
1808 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
1809 unsigned NSAAOffset = State->getNextStackOffset();
1810 if (State->getCallOrPrologue() != Call) {
1811 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
1813 State->getInRegsParamInfo(i, RB, RE);
1814 assert(NSAAOffset >= (RE-RB)*4 &&
1815 "Stack offset for byval regs doesn't introduced anymore?");
1816 NSAAOffset -= (RE-RB)*4;
1819 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1820 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1821 unsigned AlignInRegs = Align / 4;
1822 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1823 for (unsigned i = 0; i < Waste; ++i)
1824 reg = State->AllocateReg(GPRArgRegs, 4);
1827 unsigned excess = 4 * (ARM::R4 - reg);
1829 // Special case when NSAA != SP and parameter size greater than size of
1830 // all remained GPR regs. In that case we can't split parameter, we must
1831 // send it to stack. We also must set NCRN to R4, so waste all
1832 // remained registers.
1833 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1834 while (State->AllocateReg(GPRArgRegs, 4))
1839 // First register for byval parameter is the first register that wasn't
1840 // allocated before this method call, so it would be "reg".
1841 // If parameter is small enough to be saved in range [reg, r4), then
1842 // the end (first after last) register would be reg + param-size-in-regs,
1843 // else parameter would be splitted between registers and stack,
1844 // end register would be r4 in this case.
1845 unsigned ByValRegBegin = reg;
1846 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1847 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1848 // Note, first register is allocated in the beginning of function already,
1849 // allocate remained amount of registers we need.
1850 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1851 State->AllocateReg(GPRArgRegs, 4);
1852 // At a call site, a byval parameter that is split between
1853 // registers and memory needs its size truncated here. In a
1854 // function prologue, such byval parameters are reassembled in
1855 // memory, and are not truncated.
1856 if (State->getCallOrPrologue() == Call) {
1857 // Make remained size equal to 0 in case, when
1858 // the whole structure may be stored into registers.
1868 /// MatchingStackOffset - Return true if the given stack call argument is
1869 /// already available in the same position (relatively) of the caller's
1870 /// incoming argument stack.
1872 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1873 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1874 const TargetInstrInfo *TII) {
1875 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1877 if (Arg.getOpcode() == ISD::CopyFromReg) {
1878 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1879 if (!TargetRegisterInfo::isVirtualRegister(VR))
1881 MachineInstr *Def = MRI->getVRegDef(VR);
1884 if (!Flags.isByVal()) {
1885 if (!TII->isLoadFromStackSlot(Def, FI))
1890 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1891 if (Flags.isByVal())
1892 // ByVal argument is passed in as a pointer but it's now being
1893 // dereferenced. e.g.
1894 // define @foo(%struct.X* %A) {
1895 // tail call @bar(%struct.X* byval %A)
1898 SDValue Ptr = Ld->getBasePtr();
1899 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1902 FI = FINode->getIndex();
1906 assert(FI != INT_MAX);
1907 if (!MFI->isFixedObjectIndex(FI))
1909 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1912 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1913 /// for tail call optimization. Targets which want to do tail call
1914 /// optimization should implement this function.
1916 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1917 CallingConv::ID CalleeCC,
1919 bool isCalleeStructRet,
1920 bool isCallerStructRet,
1921 const SmallVectorImpl<ISD::OutputArg> &Outs,
1922 const SmallVectorImpl<SDValue> &OutVals,
1923 const SmallVectorImpl<ISD::InputArg> &Ins,
1924 SelectionDAG& DAG) const {
1925 const Function *CallerF = DAG.getMachineFunction().getFunction();
1926 CallingConv::ID CallerCC = CallerF->getCallingConv();
1927 bool CCMatch = CallerCC == CalleeCC;
1929 // Look for obvious safe cases to perform tail call optimization that do not
1930 // require ABI changes. This is what gcc calls sibcall.
1932 // Do not sibcall optimize vararg calls unless the call site is not passing
1934 if (isVarArg && !Outs.empty())
1937 // Also avoid sibcall optimization if either caller or callee uses struct
1938 // return semantics.
1939 if (isCalleeStructRet || isCallerStructRet)
1942 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1943 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1944 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1945 // support in the assembler and linker to be used. This would need to be
1946 // fixed to fully support tail calls in Thumb1.
1948 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1949 // LR. This means if we need to reload LR, it takes an extra instructions,
1950 // which outweighs the value of the tail call; but here we don't know yet
1951 // whether LR is going to be used. Probably the right approach is to
1952 // generate the tail call here and turn it back into CALL/RET in
1953 // emitEpilogue if LR is used.
1955 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1956 // but we need to make sure there are enough registers; the only valid
1957 // registers are the 4 used for parameters. We don't currently do this
1959 if (Subtarget->isThumb1Only())
1962 // If the calling conventions do not match, then we'd better make sure the
1963 // results are returned in the same way as what the caller expects.
1965 SmallVector<CCValAssign, 16> RVLocs1;
1966 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1967 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1968 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1970 SmallVector<CCValAssign, 16> RVLocs2;
1971 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1972 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1973 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1975 if (RVLocs1.size() != RVLocs2.size())
1977 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1978 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1980 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1982 if (RVLocs1[i].isRegLoc()) {
1983 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1986 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1992 // If Caller's vararg or byval argument has been split between registers and
1993 // stack, do not perform tail call, since part of the argument is in caller's
1995 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1996 getInfo<ARMFunctionInfo>();
1997 if (AFI_Caller->getArgRegsSaveSize())
2000 // If the callee takes no arguments then go on to check the results of the
2002 if (!Outs.empty()) {
2003 // Check if stack adjustment is needed. For now, do not do this if any
2004 // argument is passed on the stack.
2005 SmallVector<CCValAssign, 16> ArgLocs;
2006 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2007 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2008 CCInfo.AnalyzeCallOperands(Outs,
2009 CCAssignFnForNode(CalleeCC, false, isVarArg));
2010 if (CCInfo.getNextStackOffset()) {
2011 MachineFunction &MF = DAG.getMachineFunction();
2013 // Check if the arguments are already laid out in the right way as
2014 // the caller's fixed stack objects.
2015 MachineFrameInfo *MFI = MF.getFrameInfo();
2016 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2017 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2018 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2020 ++i, ++realArgIdx) {
2021 CCValAssign &VA = ArgLocs[i];
2022 EVT RegVT = VA.getLocVT();
2023 SDValue Arg = OutVals[realArgIdx];
2024 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2025 if (VA.getLocInfo() == CCValAssign::Indirect)
2027 if (VA.needsCustom()) {
2028 // f64 and vector types are split into multiple registers or
2029 // register/stack-slot combinations. The types will not match
2030 // the registers; give up on memory f64 refs until we figure
2031 // out what to do about this.
2034 if (!ArgLocs[++i].isRegLoc())
2036 if (RegVT == MVT::v2f64) {
2037 if (!ArgLocs[++i].isRegLoc())
2039 if (!ArgLocs[++i].isRegLoc())
2042 } else if (!VA.isRegLoc()) {
2043 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2055 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2056 MachineFunction &MF, bool isVarArg,
2057 const SmallVectorImpl<ISD::OutputArg> &Outs,
2058 LLVMContext &Context) const {
2059 SmallVector<CCValAssign, 16> RVLocs;
2060 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2061 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2066 ARMTargetLowering::LowerReturn(SDValue Chain,
2067 CallingConv::ID CallConv, bool isVarArg,
2068 const SmallVectorImpl<ISD::OutputArg> &Outs,
2069 const SmallVectorImpl<SDValue> &OutVals,
2070 SDLoc dl, SelectionDAG &DAG) const {
2072 // CCValAssign - represent the assignment of the return value to a location.
2073 SmallVector<CCValAssign, 16> RVLocs;
2075 // CCState - Info about the registers and stack slots.
2076 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2077 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2079 // Analyze outgoing return values.
2080 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2084 SmallVector<SDValue, 4> RetOps;
2085 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2087 // Copy the result values into the output registers.
2088 for (unsigned i = 0, realRVLocIdx = 0;
2090 ++i, ++realRVLocIdx) {
2091 CCValAssign &VA = RVLocs[i];
2092 assert(VA.isRegLoc() && "Can only return in registers!");
2094 SDValue Arg = OutVals[realRVLocIdx];
2096 switch (VA.getLocInfo()) {
2097 default: llvm_unreachable("Unknown loc info!");
2098 case CCValAssign::Full: break;
2099 case CCValAssign::BCvt:
2100 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2104 if (VA.needsCustom()) {
2105 if (VA.getLocVT() == MVT::v2f64) {
2106 // Extract the first half and return it in two registers.
2107 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2108 DAG.getConstant(0, MVT::i32));
2109 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2110 DAG.getVTList(MVT::i32, MVT::i32), Half);
2112 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2113 Flag = Chain.getValue(1);
2114 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2115 VA = RVLocs[++i]; // skip ahead to next loc
2116 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2117 HalfGPRs.getValue(1), Flag);
2118 Flag = Chain.getValue(1);
2119 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2120 VA = RVLocs[++i]; // skip ahead to next loc
2122 // Extract the 2nd half and fall through to handle it as an f64 value.
2123 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2124 DAG.getConstant(1, MVT::i32));
2126 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2128 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2129 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2130 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2131 Flag = Chain.getValue(1);
2132 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2133 VA = RVLocs[++i]; // skip ahead to next loc
2134 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2137 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2139 // Guarantee that all emitted copies are
2140 // stuck together, avoiding something bad.
2141 Flag = Chain.getValue(1);
2142 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2145 // Update chain and glue.
2148 RetOps.push_back(Flag);
2150 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2151 RetOps.data(), RetOps.size());
2154 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2155 if (N->getNumValues() != 1)
2157 if (!N->hasNUsesOfValue(1, 0))
2160 SDValue TCChain = Chain;
2161 SDNode *Copy = *N->use_begin();
2162 if (Copy->getOpcode() == ISD::CopyToReg) {
2163 // If the copy has a glue operand, we conservatively assume it isn't safe to
2164 // perform a tail call.
2165 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2167 TCChain = Copy->getOperand(0);
2168 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2169 SDNode *VMov = Copy;
2170 // f64 returned in a pair of GPRs.
2171 SmallPtrSet<SDNode*, 2> Copies;
2172 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2174 if (UI->getOpcode() != ISD::CopyToReg)
2178 if (Copies.size() > 2)
2181 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2183 SDValue UseChain = UI->getOperand(0);
2184 if (Copies.count(UseChain.getNode()))
2191 } else if (Copy->getOpcode() == ISD::BITCAST) {
2192 // f32 returned in a single GPR.
2193 if (!Copy->hasOneUse())
2195 Copy = *Copy->use_begin();
2196 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2198 TCChain = Copy->getOperand(0);
2203 bool HasRet = false;
2204 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2206 if (UI->getOpcode() != ARMISD::RET_FLAG)
2218 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2219 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2222 if (!CI->isTailCall())
2225 return !Subtarget->isThumb1Only();
2228 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2229 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2230 // one of the above mentioned nodes. It has to be wrapped because otherwise
2231 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2232 // be used to form addressing mode. These wrapped nodes will be selected
2234 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2235 EVT PtrVT = Op.getValueType();
2236 // FIXME there is no actual debug info here
2238 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2240 if (CP->isMachineConstantPoolEntry())
2241 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2242 CP->getAlignment());
2244 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2245 CP->getAlignment());
2246 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2249 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2250 return MachineJumpTableInfo::EK_Inline;
2253 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2254 SelectionDAG &DAG) const {
2255 MachineFunction &MF = DAG.getMachineFunction();
2256 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2257 unsigned ARMPCLabelIndex = 0;
2259 EVT PtrVT = getPointerTy();
2260 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2261 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2263 if (RelocM == Reloc::Static) {
2264 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2266 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2267 ARMPCLabelIndex = AFI->createPICLabelUId();
2268 ARMConstantPoolValue *CPV =
2269 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2270 ARMCP::CPBlockAddress, PCAdj);
2271 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2273 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2274 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2275 MachinePointerInfo::getConstantPool(),
2276 false, false, false, 0);
2277 if (RelocM == Reloc::Static)
2279 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2280 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2283 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2285 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2286 SelectionDAG &DAG) const {
2288 EVT PtrVT = getPointerTy();
2289 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2290 MachineFunction &MF = DAG.getMachineFunction();
2291 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2292 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2293 ARMConstantPoolValue *CPV =
2294 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2295 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2296 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2297 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2298 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2299 MachinePointerInfo::getConstantPool(),
2300 false, false, false, 0);
2301 SDValue Chain = Argument.getValue(1);
2303 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2304 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2306 // call __tls_get_addr.
2309 Entry.Node = Argument;
2310 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2311 Args.push_back(Entry);
2312 // FIXME: is there useful debug info available here?
2313 TargetLowering::CallLoweringInfo CLI(Chain,
2314 (Type *) Type::getInt32Ty(*DAG.getContext()),
2315 false, false, false, false,
2316 0, CallingConv::C, /*isTailCall=*/false,
2317 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2318 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2319 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2320 return CallResult.first;
2323 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2324 // "local exec" model.
2326 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2328 TLSModel::Model model) const {
2329 const GlobalValue *GV = GA->getGlobal();
2332 SDValue Chain = DAG.getEntryNode();
2333 EVT PtrVT = getPointerTy();
2334 // Get the Thread Pointer
2335 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2337 if (model == TLSModel::InitialExec) {
2338 MachineFunction &MF = DAG.getMachineFunction();
2339 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2340 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2341 // Initial exec model.
2342 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2343 ARMConstantPoolValue *CPV =
2344 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2345 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2347 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2348 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2349 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2350 MachinePointerInfo::getConstantPool(),
2351 false, false, false, 0);
2352 Chain = Offset.getValue(1);
2354 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2355 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2357 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2358 MachinePointerInfo::getConstantPool(),
2359 false, false, false, 0);
2362 assert(model == TLSModel::LocalExec);
2363 ARMConstantPoolValue *CPV =
2364 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2365 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2366 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2367 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2368 MachinePointerInfo::getConstantPool(),
2369 false, false, false, 0);
2372 // The address of the thread local variable is the add of the thread
2373 // pointer with the offset of the variable.
2374 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2378 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2379 // TODO: implement the "local dynamic" model
2380 assert(Subtarget->isTargetELF() &&
2381 "TLS not implemented for non-ELF targets");
2382 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2384 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2387 case TLSModel::GeneralDynamic:
2388 case TLSModel::LocalDynamic:
2389 return LowerToTLSGeneralDynamicModel(GA, DAG);
2390 case TLSModel::InitialExec:
2391 case TLSModel::LocalExec:
2392 return LowerToTLSExecModels(GA, DAG, model);
2394 llvm_unreachable("bogus TLS model");
2397 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2398 SelectionDAG &DAG) const {
2399 EVT PtrVT = getPointerTy();
2401 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2402 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2403 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2404 ARMConstantPoolValue *CPV =
2405 ARMConstantPoolConstant::Create(GV,
2406 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2407 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2408 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2409 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2411 MachinePointerInfo::getConstantPool(),
2412 false, false, false, 0);
2413 SDValue Chain = Result.getValue(1);
2414 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2415 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2417 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2418 MachinePointerInfo::getGOT(),
2419 false, false, false, 0);
2423 // If we have T2 ops, we can materialize the address directly via movt/movw
2424 // pair. This is always cheaper.
2425 if (Subtarget->useMovt()) {
2427 // FIXME: Once remat is capable of dealing with instructions with register
2428 // operands, expand this into two nodes.
2429 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2430 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2432 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2433 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2434 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2435 MachinePointerInfo::getConstantPool(),
2436 false, false, false, 0);
2440 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2441 SelectionDAG &DAG) const {
2442 EVT PtrVT = getPointerTy();
2444 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2445 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2447 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2448 // update ARMFastISel::ARMMaterializeGV.
2449 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2451 // FIXME: Once remat is capable of dealing with instructions with register
2452 // operands, expand this into two nodes.
2453 if (RelocM == Reloc::Static)
2454 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2455 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2457 unsigned Wrapper = (RelocM == Reloc::PIC_)
2458 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2459 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2460 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2461 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2462 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2463 MachinePointerInfo::getGOT(),
2464 false, false, false, 0);
2468 unsigned ARMPCLabelIndex = 0;
2470 if (RelocM == Reloc::Static) {
2471 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2473 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2474 ARMPCLabelIndex = AFI->createPICLabelUId();
2475 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2476 ARMConstantPoolValue *CPV =
2477 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2479 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2481 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2483 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2484 MachinePointerInfo::getConstantPool(),
2485 false, false, false, 0);
2486 SDValue Chain = Result.getValue(1);
2488 if (RelocM == Reloc::PIC_) {
2489 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2490 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2493 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2494 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2495 false, false, false, 0);
2500 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2501 SelectionDAG &DAG) const {
2502 assert(Subtarget->isTargetELF() &&
2503 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2504 MachineFunction &MF = DAG.getMachineFunction();
2505 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2507 EVT PtrVT = getPointerTy();
2509 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2510 ARMConstantPoolValue *CPV =
2511 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2512 ARMPCLabelIndex, PCAdj);
2513 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2514 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2515 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2516 MachinePointerInfo::getConstantPool(),
2517 false, false, false, 0);
2518 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2519 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2523 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2525 SDValue Val = DAG.getConstant(0, MVT::i32);
2526 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2527 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2528 Op.getOperand(1), Val);
2532 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2534 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2535 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2539 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2540 const ARMSubtarget *Subtarget) const {
2541 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2544 default: return SDValue(); // Don't custom lower most intrinsics.
2545 case Intrinsic::arm_thread_pointer: {
2546 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2547 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2549 case Intrinsic::eh_sjlj_lsda: {
2550 MachineFunction &MF = DAG.getMachineFunction();
2551 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2552 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2553 EVT PtrVT = getPointerTy();
2554 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2556 unsigned PCAdj = (RelocM != Reloc::PIC_)
2557 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2558 ARMConstantPoolValue *CPV =
2559 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2560 ARMCP::CPLSDA, PCAdj);
2561 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2562 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2564 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2565 MachinePointerInfo::getConstantPool(),
2566 false, false, false, 0);
2568 if (RelocM == Reloc::PIC_) {
2569 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2570 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2574 case Intrinsic::arm_neon_vmulls:
2575 case Intrinsic::arm_neon_vmullu: {
2576 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2577 ? ARMISD::VMULLs : ARMISD::VMULLu;
2578 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2579 Op.getOperand(1), Op.getOperand(2));
2584 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2585 const ARMSubtarget *Subtarget) {
2586 // FIXME: handle "fence singlethread" more efficiently.
2588 if (!Subtarget->hasDataBarrier()) {
2589 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2590 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2592 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2593 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2594 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2595 DAG.getConstant(0, MVT::i32));
2598 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2599 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2600 unsigned Domain = ARM_MB::ISH;
2601 if (Subtarget->isSwift() && Ord == Release) {
2602 // Swift happens to implement ISHST barriers in a way that's compatible with
2603 // Release semantics but weaker than ISH so we'd be fools not to use
2604 // it. Beware: other processors probably don't!
2605 Domain = ARM_MB::ISHST;
2608 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2609 DAG.getConstant(Domain, MVT::i32));
2612 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2613 const ARMSubtarget *Subtarget) {
2614 // ARM pre v5TE and Thumb1 does not have preload instructions.
2615 if (!(Subtarget->isThumb2() ||
2616 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2617 // Just preserve the chain.
2618 return Op.getOperand(0);
2621 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2623 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2624 // ARMv7 with MP extension has PLDW.
2625 return Op.getOperand(0);
2627 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2628 if (Subtarget->isThumb()) {
2630 isRead = ~isRead & 1;
2631 isData = ~isData & 1;
2634 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2635 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2636 DAG.getConstant(isData, MVT::i32));
2639 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2640 MachineFunction &MF = DAG.getMachineFunction();
2641 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2643 // vastart just stores the address of the VarArgsFrameIndex slot into the
2644 // memory location argument.
2646 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2647 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2648 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2649 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2650 MachinePointerInfo(SV), false, false, 0);
2654 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2655 SDValue &Root, SelectionDAG &DAG,
2657 MachineFunction &MF = DAG.getMachineFunction();
2658 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2660 const TargetRegisterClass *RC;
2661 if (AFI->isThumb1OnlyFunction())
2662 RC = &ARM::tGPRRegClass;
2664 RC = &ARM::GPRRegClass;
2666 // Transform the arguments stored in physical registers into virtual ones.
2667 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2668 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2671 if (NextVA.isMemLoc()) {
2672 MachineFrameInfo *MFI = MF.getFrameInfo();
2673 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2675 // Create load node to retrieve arguments from the stack.
2676 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2677 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2678 MachinePointerInfo::getFixedStack(FI),
2679 false, false, false, 0);
2681 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2682 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2685 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2689 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2690 unsigned InRegsParamRecordIdx,
2692 unsigned &ArgRegsSize,
2693 unsigned &ArgRegsSaveSize)
2696 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2697 unsigned RBegin, REnd;
2698 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2699 NumGPRs = REnd - RBegin;
2701 unsigned int firstUnalloced;
2702 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2703 sizeof(GPRArgRegs) /
2704 sizeof(GPRArgRegs[0]));
2705 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2708 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2709 ArgRegsSize = NumGPRs * 4;
2711 // If parameter is split between stack and GPRs...
2712 if (NumGPRs && Align == 8 &&
2713 (ArgRegsSize < ArgSize ||
2714 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2715 // Add padding for part of param recovered from GPRs, so
2716 // its last byte must be at address K*8 - 1.
2717 // We need to do it, since remained (stack) part of parameter has
2718 // stack alignment, and we need to "attach" "GPRs head" without gaps
2721 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2722 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2724 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2726 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
2727 (ArgRegsSize + AFI->getArgRegsSaveSize());
2728 ArgRegsSaveSize = ArgRegsSize + Padding;
2730 // We don't need to extend regs save size for byval parameters if they
2731 // are passed via GPRs only.
2732 ArgRegsSaveSize = ArgRegsSize;
2735 // The remaining GPRs hold either the beginning of variable-argument
2736 // data, or the beginning of an aggregate passed by value (usually
2737 // byval). Either way, we allocate stack slots adjacent to the data
2738 // provided by our caller, and store the unallocated registers there.
2739 // If this is a variadic function, the va_list pointer will begin with
2740 // these values; otherwise, this reassembles a (byval) structure that
2741 // was split between registers and memory.
2742 // Return: The frame index registers were stored into.
2744 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2745 SDLoc dl, SDValue &Chain,
2746 const Value *OrigArg,
2747 unsigned InRegsParamRecordIdx,
2748 unsigned OffsetFromOrigArg,
2751 bool ForceMutable) const {
2753 // Currently, two use-cases possible:
2754 // Case #1. Non var-args function, and we meet first byval parameter.
2755 // Setup first unallocated register as first byval register;
2756 // eat all remained registers
2757 // (these two actions are performed by HandleByVal method).
2758 // Then, here, we initialize stack frame with
2759 // "store-reg" instructions.
2760 // Case #2. Var-args function, that doesn't contain byval parameters.
2761 // The same: eat all remained unallocated registers,
2762 // initialize stack frame.
2764 MachineFunction &MF = DAG.getMachineFunction();
2765 MachineFrameInfo *MFI = MF.getFrameInfo();
2766 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2767 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2768 unsigned RBegin, REnd;
2769 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2770 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2771 firstRegToSaveIndex = RBegin - ARM::R0;
2772 lastRegToSaveIndex = REnd - ARM::R0;
2774 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2775 (GPRArgRegs, array_lengthof(GPRArgRegs));
2776 lastRegToSaveIndex = 4;
2779 unsigned ArgRegsSize, ArgRegsSaveSize;
2780 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2781 ArgRegsSize, ArgRegsSaveSize);
2783 // Store any by-val regs to their spots on the stack so that they may be
2784 // loaded by deferencing the result of formal parameter pointer or va_next.
2785 // Note: once stack area for byval/varargs registers
2786 // was initialized, it can't be initialized again.
2787 if (ArgRegsSaveSize) {
2789 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2792 assert(AFI->getStoredByValParamsPadding() == 0 &&
2793 "The only parameter may be padded.");
2794 AFI->setStoredByValParamsPadding(Padding);
2797 int FrameIndex = MFI->CreateFixedObject(
2799 Padding + ArgOffset,
2801 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2803 SmallVector<SDValue, 4> MemOps;
2804 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2805 ++firstRegToSaveIndex, ++i) {
2806 const TargetRegisterClass *RC;
2807 if (AFI->isThumb1OnlyFunction())
2808 RC = &ARM::tGPRRegClass;
2810 RC = &ARM::GPRRegClass;
2812 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2813 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2815 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2816 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2818 MemOps.push_back(Store);
2819 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2820 DAG.getConstant(4, getPointerTy()));
2823 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2825 if (!MemOps.empty())
2826 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2827 &MemOps[0], MemOps.size());
2830 // This will point to the next argument passed via stack.
2831 return MFI->CreateFixedObject(
2832 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
2835 // Setup stack frame, the va_list pointer will start from.
2837 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2838 SDLoc dl, SDValue &Chain,
2840 bool ForceMutable) const {
2841 MachineFunction &MF = DAG.getMachineFunction();
2842 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2844 // Try to store any remaining integer argument regs
2845 // to their spots on the stack so that they may be loaded by deferencing
2846 // the result of va_next.
2847 // If there is no regs to be stored, just point address after last
2848 // argument passed via stack.
2850 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2851 0, ArgOffset, 0, ForceMutable);
2853 AFI->setVarArgsFrameIndex(FrameIndex);
2857 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2858 CallingConv::ID CallConv, bool isVarArg,
2859 const SmallVectorImpl<ISD::InputArg>
2861 SDLoc dl, SelectionDAG &DAG,
2862 SmallVectorImpl<SDValue> &InVals)
2864 MachineFunction &MF = DAG.getMachineFunction();
2865 MachineFrameInfo *MFI = MF.getFrameInfo();
2867 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2869 // Assign locations to all of the incoming arguments.
2870 SmallVector<CCValAssign, 16> ArgLocs;
2871 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2872 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2873 CCInfo.AnalyzeFormalArguments(Ins,
2874 CCAssignFnForNode(CallConv, /* Return*/ false,
2877 SmallVector<SDValue, 16> ArgValues;
2878 int lastInsIndex = -1;
2880 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2881 unsigned CurArgIdx = 0;
2883 // Initially ArgRegsSaveSize is zero.
2884 // Then we increase this value each time we meet byval parameter.
2885 // We also increase this value in case of varargs function.
2886 AFI->setArgRegsSaveSize(0);
2888 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2889 CCValAssign &VA = ArgLocs[i];
2890 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2891 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2892 // Arguments stored in registers.
2893 if (VA.isRegLoc()) {
2894 EVT RegVT = VA.getLocVT();
2896 if (VA.needsCustom()) {
2897 // f64 and vector types are split up into multiple registers or
2898 // combinations of registers and stack slots.
2899 if (VA.getLocVT() == MVT::v2f64) {
2900 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2902 VA = ArgLocs[++i]; // skip ahead to next loc
2904 if (VA.isMemLoc()) {
2905 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2906 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2907 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2908 MachinePointerInfo::getFixedStack(FI),
2909 false, false, false, 0);
2911 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2914 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2915 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2916 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2917 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2918 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2920 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2923 const TargetRegisterClass *RC;
2925 if (RegVT == MVT::f32)
2926 RC = &ARM::SPRRegClass;
2927 else if (RegVT == MVT::f64)
2928 RC = &ARM::DPRRegClass;
2929 else if (RegVT == MVT::v2f64)
2930 RC = &ARM::QPRRegClass;
2931 else if (RegVT == MVT::i32)
2932 RC = AFI->isThumb1OnlyFunction() ?
2933 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2934 (const TargetRegisterClass*)&ARM::GPRRegClass;
2936 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2938 // Transform the arguments in physical registers into virtual ones.
2939 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2940 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2943 // If this is an 8 or 16-bit value, it is really passed promoted
2944 // to 32 bits. Insert an assert[sz]ext to capture this, then
2945 // truncate to the right size.
2946 switch (VA.getLocInfo()) {
2947 default: llvm_unreachable("Unknown loc info!");
2948 case CCValAssign::Full: break;
2949 case CCValAssign::BCvt:
2950 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2952 case CCValAssign::SExt:
2953 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2954 DAG.getValueType(VA.getValVT()));
2955 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2957 case CCValAssign::ZExt:
2958 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2959 DAG.getValueType(VA.getValVT()));
2960 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2964 InVals.push_back(ArgValue);
2966 } else { // VA.isRegLoc()
2969 assert(VA.isMemLoc());
2970 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2972 int index = ArgLocs[i].getValNo();
2974 // Some Ins[] entries become multiple ArgLoc[] entries.
2975 // Process them only once.
2976 if (index != lastInsIndex)
2978 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2979 // FIXME: For now, all byval parameter objects are marked mutable.
2980 // This can be changed with more analysis.
2981 // In case of tail call optimization mark all arguments mutable.
2982 // Since they could be overwritten by lowering of arguments in case of
2984 if (Flags.isByVal()) {
2985 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
2986 int FrameIndex = StoreByValRegs(
2987 CCInfo, DAG, dl, Chain, CurOrigArg,
2989 Ins[VA.getValNo()].PartOffset,
2990 VA.getLocMemOffset(),
2991 Flags.getByValSize(),
2992 true /*force mutable frames*/);
2993 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
2994 CCInfo.nextInRegsParam();
2996 unsigned FIOffset = VA.getLocMemOffset() +
2997 AFI->getStoredByValParamsPadding();
2998 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3001 // Create load nodes to retrieve arguments from the stack.
3002 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3003 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3004 MachinePointerInfo::getFixedStack(FI),
3005 false, false, false, 0));
3007 lastInsIndex = index;
3014 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3015 CCInfo.getNextStackOffset());
3020 /// isFloatingPointZero - Return true if this is +0.0.
3021 static bool isFloatingPointZero(SDValue Op) {
3022 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3023 return CFP->getValueAPF().isPosZero();
3024 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3025 // Maybe this has already been legalized into the constant pool?
3026 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3027 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3028 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3029 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3030 return CFP->getValueAPF().isPosZero();
3036 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3037 /// the given operands.
3039 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3040 SDValue &ARMcc, SelectionDAG &DAG,
3042 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3043 unsigned C = RHSC->getZExtValue();
3044 if (!isLegalICmpImmediate(C)) {
3045 // Constant does not fit, try adjusting it by one?
3050 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3051 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3052 RHS = DAG.getConstant(C-1, MVT::i32);
3057 if (C != 0 && isLegalICmpImmediate(C-1)) {
3058 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3059 RHS = DAG.getConstant(C-1, MVT::i32);
3064 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3065 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3066 RHS = DAG.getConstant(C+1, MVT::i32);
3071 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3072 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3073 RHS = DAG.getConstant(C+1, MVT::i32);
3080 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3081 ARMISD::NodeType CompareType;
3084 CompareType = ARMISD::CMP;
3089 CompareType = ARMISD::CMPZ;
3092 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3093 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3096 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3098 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3101 if (!isFloatingPointZero(RHS))
3102 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3104 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3105 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3108 /// duplicateCmp - Glue values can have only one use, so this function
3109 /// duplicates a comparison node.
3111 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3112 unsigned Opc = Cmp.getOpcode();
3114 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3115 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3117 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3118 Cmp = Cmp.getOperand(0);
3119 Opc = Cmp.getOpcode();
3120 if (Opc == ARMISD::CMPFP)
3121 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3123 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3124 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3126 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3129 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3130 SDValue Cond = Op.getOperand(0);
3131 SDValue SelectTrue = Op.getOperand(1);
3132 SDValue SelectFalse = Op.getOperand(2);
3137 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3138 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3140 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3141 const ConstantSDNode *CMOVTrue =
3142 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3143 const ConstantSDNode *CMOVFalse =
3144 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3146 if (CMOVTrue && CMOVFalse) {
3147 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3148 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3152 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3154 False = SelectFalse;
3155 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3160 if (True.getNode() && False.getNode()) {
3161 EVT VT = Op.getValueType();
3162 SDValue ARMcc = Cond.getOperand(2);
3163 SDValue CCR = Cond.getOperand(3);
3164 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3165 assert(True.getValueType() == VT);
3166 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3171 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3172 // undefined bits before doing a full-word comparison with zero.
3173 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3174 DAG.getConstant(1, Cond.getValueType()));
3176 return DAG.getSelectCC(dl, Cond,
3177 DAG.getConstant(0, Cond.getValueType()),
3178 SelectTrue, SelectFalse, ISD::SETNE);
3181 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3182 if (CC == ISD::SETNE)
3184 return ISD::getSetCCSwappedOperands(CC);
3187 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3188 bool &swpCmpOps, bool &swpVselOps) {
3189 // Start by selecting the GE condition code for opcodes that return true for
3191 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3193 CondCode = ARMCC::GE;
3195 // and GT for opcodes that return false for 'equality'.
3196 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3198 CondCode = ARMCC::GT;
3200 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3201 // to swap the compare operands.
3202 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3206 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3207 // If we have an unordered opcode, we need to swap the operands to the VSEL
3208 // instruction (effectively negating the condition).
3210 // This also has the effect of swapping which one of 'less' or 'greater'
3211 // returns true, so we also swap the compare operands. It also switches
3212 // whether we return true for 'equality', so we compensate by picking the
3213 // opposite condition code to our original choice.
3214 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3215 CC == ISD::SETUGT) {
3216 swpCmpOps = !swpCmpOps;
3217 swpVselOps = !swpVselOps;
3218 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3221 // 'ordered' is 'anything but unordered', so use the VS condition code and
3222 // swap the VSEL operands.
3223 if (CC == ISD::SETO) {
3224 CondCode = ARMCC::VS;
3228 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3229 // code and swap the VSEL operands.
3230 if (CC == ISD::SETUNE) {
3231 CondCode = ARMCC::EQ;
3236 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3237 EVT VT = Op.getValueType();
3238 SDValue LHS = Op.getOperand(0);
3239 SDValue RHS = Op.getOperand(1);
3240 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3241 SDValue TrueVal = Op.getOperand(2);
3242 SDValue FalseVal = Op.getOperand(3);
3245 if (LHS.getValueType() == MVT::i32) {
3246 // Try to generate VSEL on ARMv8.
3247 // The VSEL instruction can't use all the usual ARM condition
3248 // codes: it only has two bits to select the condition code, so it's
3249 // constrained to use only GE, GT, VS and EQ.
3251 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3252 // swap the operands of the previous compare instruction (effectively
3253 // inverting the compare condition, swapping 'less' and 'greater') and
3254 // sometimes need to swap the operands to the VSEL (which inverts the
3255 // condition in the sense of firing whenever the previous condition didn't)
3256 if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
3257 TrueVal.getValueType() == MVT::f64)) {
3258 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3259 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3260 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3261 CC = getInverseCCForVSEL(CC);
3262 std::swap(TrueVal, FalseVal);
3267 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3268 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3269 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3273 ARMCC::CondCodes CondCode, CondCode2;
3274 FPCCToARMCC(CC, CondCode, CondCode2);
3276 // Try to generate VSEL on ARMv8.
3277 if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
3278 TrueVal.getValueType() == MVT::f64)) {
3279 bool swpCmpOps = false;
3280 bool swpVselOps = false;
3281 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3283 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3284 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3286 std::swap(LHS, RHS);
3288 std::swap(TrueVal, FalseVal);
3292 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3293 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3294 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3295 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3297 if (CondCode2 != ARMCC::AL) {
3298 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3299 // FIXME: Needs another CMP because flag can have but one use.
3300 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3301 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3302 Result, TrueVal, ARMcc2, CCR, Cmp2);
3307 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3308 /// to morph to an integer compare sequence.
3309 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3310 const ARMSubtarget *Subtarget) {
3311 SDNode *N = Op.getNode();
3312 if (!N->hasOneUse())
3313 // Otherwise it requires moving the value from fp to integer registers.
3315 if (!N->getNumValues())
3317 EVT VT = Op.getValueType();
3318 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3319 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3320 // vmrs are very slow, e.g. cortex-a8.
3323 if (isFloatingPointZero(Op)) {
3327 return ISD::isNormalLoad(N);
3330 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3331 if (isFloatingPointZero(Op))
3332 return DAG.getConstant(0, MVT::i32);
3334 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3335 return DAG.getLoad(MVT::i32, SDLoc(Op),
3336 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3337 Ld->isVolatile(), Ld->isNonTemporal(),
3338 Ld->isInvariant(), Ld->getAlignment());
3340 llvm_unreachable("Unknown VFP cmp argument!");
3343 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3344 SDValue &RetVal1, SDValue &RetVal2) {
3345 if (isFloatingPointZero(Op)) {
3346 RetVal1 = DAG.getConstant(0, MVT::i32);
3347 RetVal2 = DAG.getConstant(0, MVT::i32);
3351 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3352 SDValue Ptr = Ld->getBasePtr();
3353 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3354 Ld->getChain(), Ptr,
3355 Ld->getPointerInfo(),
3356 Ld->isVolatile(), Ld->isNonTemporal(),
3357 Ld->isInvariant(), Ld->getAlignment());
3359 EVT PtrType = Ptr.getValueType();
3360 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3361 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3362 PtrType, Ptr, DAG.getConstant(4, PtrType));
3363 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3364 Ld->getChain(), NewPtr,
3365 Ld->getPointerInfo().getWithOffset(4),
3366 Ld->isVolatile(), Ld->isNonTemporal(),
3367 Ld->isInvariant(), NewAlign);
3371 llvm_unreachable("Unknown VFP cmp argument!");
3374 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3375 /// f32 and even f64 comparisons to integer ones.
3377 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3378 SDValue Chain = Op.getOperand(0);
3379 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3380 SDValue LHS = Op.getOperand(2);
3381 SDValue RHS = Op.getOperand(3);
3382 SDValue Dest = Op.getOperand(4);
3385 bool LHSSeenZero = false;
3386 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3387 bool RHSSeenZero = false;
3388 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3389 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3390 // If unsafe fp math optimization is enabled and there are no other uses of
3391 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3392 // to an integer comparison.
3393 if (CC == ISD::SETOEQ)
3395 else if (CC == ISD::SETUNE)
3398 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3400 if (LHS.getValueType() == MVT::f32) {
3401 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3402 bitcastf32Toi32(LHS, DAG), Mask);
3403 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3404 bitcastf32Toi32(RHS, DAG), Mask);
3405 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3406 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3407 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3408 Chain, Dest, ARMcc, CCR, Cmp);
3413 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3414 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3415 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3416 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3417 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3418 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3419 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3420 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3421 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3427 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3428 SDValue Chain = Op.getOperand(0);
3429 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3430 SDValue LHS = Op.getOperand(2);
3431 SDValue RHS = Op.getOperand(3);
3432 SDValue Dest = Op.getOperand(4);
3435 if (LHS.getValueType() == MVT::i32) {
3437 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3438 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3439 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3440 Chain, Dest, ARMcc, CCR, Cmp);
3443 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3445 if (getTargetMachine().Options.UnsafeFPMath &&
3446 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3447 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3448 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3449 if (Result.getNode())
3453 ARMCC::CondCodes CondCode, CondCode2;
3454 FPCCToARMCC(CC, CondCode, CondCode2);
3456 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3457 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3458 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3459 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3460 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3461 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3462 if (CondCode2 != ARMCC::AL) {
3463 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3464 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3465 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3470 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3471 SDValue Chain = Op.getOperand(0);
3472 SDValue Table = Op.getOperand(1);
3473 SDValue Index = Op.getOperand(2);
3476 EVT PTy = getPointerTy();
3477 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3478 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3479 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3480 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3481 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3482 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3483 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3484 if (Subtarget->isThumb2()) {
3485 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3486 // which does another jump to the destination. This also makes it easier
3487 // to translate it to TBB / TBH later.
3488 // FIXME: This might not work if the function is extremely large.
3489 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3490 Addr, Op.getOperand(2), JTI, UId);
3492 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3493 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3494 MachinePointerInfo::getJumpTable(),
3495 false, false, false, 0);
3496 Chain = Addr.getValue(1);
3497 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3498 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3500 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3501 MachinePointerInfo::getJumpTable(),
3502 false, false, false, 0);
3503 Chain = Addr.getValue(1);
3504 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3508 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3509 EVT VT = Op.getValueType();
3512 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3513 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3515 return DAG.UnrollVectorOp(Op.getNode());
3518 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3519 "Invalid type for custom lowering!");
3520 if (VT != MVT::v4i16)
3521 return DAG.UnrollVectorOp(Op.getNode());
3523 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3524 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3527 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3528 EVT VT = Op.getValueType();
3530 return LowerVectorFP_TO_INT(Op, DAG);
3535 switch (Op.getOpcode()) {
3536 default: llvm_unreachable("Invalid opcode!");
3537 case ISD::FP_TO_SINT:
3538 Opc = ARMISD::FTOSI;
3540 case ISD::FP_TO_UINT:
3541 Opc = ARMISD::FTOUI;
3544 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3545 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3548 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3549 EVT VT = Op.getValueType();
3552 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3553 if (VT.getVectorElementType() == MVT::f32)
3555 return DAG.UnrollVectorOp(Op.getNode());
3558 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3559 "Invalid type for custom lowering!");
3560 if (VT != MVT::v4f32)
3561 return DAG.UnrollVectorOp(Op.getNode());
3565 switch (Op.getOpcode()) {
3566 default: llvm_unreachable("Invalid opcode!");
3567 case ISD::SINT_TO_FP:
3568 CastOpc = ISD::SIGN_EXTEND;
3569 Opc = ISD::SINT_TO_FP;
3571 case ISD::UINT_TO_FP:
3572 CastOpc = ISD::ZERO_EXTEND;
3573 Opc = ISD::UINT_TO_FP;
3577 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3578 return DAG.getNode(Opc, dl, VT, Op);
3581 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3582 EVT VT = Op.getValueType();
3584 return LowerVectorINT_TO_FP(Op, DAG);
3589 switch (Op.getOpcode()) {
3590 default: llvm_unreachable("Invalid opcode!");
3591 case ISD::SINT_TO_FP:
3592 Opc = ARMISD::SITOF;
3594 case ISD::UINT_TO_FP:
3595 Opc = ARMISD::UITOF;
3599 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3600 return DAG.getNode(Opc, dl, VT, Op);
3603 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3604 // Implement fcopysign with a fabs and a conditional fneg.
3605 SDValue Tmp0 = Op.getOperand(0);
3606 SDValue Tmp1 = Op.getOperand(1);
3608 EVT VT = Op.getValueType();
3609 EVT SrcVT = Tmp1.getValueType();
3610 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3611 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3612 bool UseNEON = !InGPR && Subtarget->hasNEON();
3615 // Use VBSL to copy the sign bit.
3616 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3617 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3618 DAG.getTargetConstant(EncodedVal, MVT::i32));
3619 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3621 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3622 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3623 DAG.getConstant(32, MVT::i32));
3624 else /*if (VT == MVT::f32)*/
3625 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3626 if (SrcVT == MVT::f32) {
3627 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3629 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3630 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3631 DAG.getConstant(32, MVT::i32));
3632 } else if (VT == MVT::f32)
3633 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3634 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3635 DAG.getConstant(32, MVT::i32));
3636 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3637 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3639 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3641 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3642 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3643 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3645 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3646 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3647 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3648 if (VT == MVT::f32) {
3649 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3650 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3651 DAG.getConstant(0, MVT::i32));
3653 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3659 // Bitcast operand 1 to i32.
3660 if (SrcVT == MVT::f64)
3661 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3662 &Tmp1, 1).getValue(1);
3663 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3665 // Or in the signbit with integer operations.
3666 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3667 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3668 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3669 if (VT == MVT::f32) {
3670 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3671 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3672 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3673 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3676 // f64: Or the high part with signbit and then combine two parts.
3677 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3679 SDValue Lo = Tmp0.getValue(0);
3680 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3681 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3682 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3685 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3686 MachineFunction &MF = DAG.getMachineFunction();
3687 MachineFrameInfo *MFI = MF.getFrameInfo();
3688 MFI->setReturnAddressIsTaken(true);
3690 EVT VT = Op.getValueType();
3692 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3694 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3695 SDValue Offset = DAG.getConstant(4, MVT::i32);
3696 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3697 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3698 MachinePointerInfo(), false, false, false, 0);
3701 // Return LR, which contains the return address. Mark it an implicit live-in.
3702 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3703 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3706 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3707 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3708 MFI->setFrameAddressIsTaken(true);
3710 EVT VT = Op.getValueType();
3711 SDLoc dl(Op); // FIXME probably not meaningful
3712 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3713 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3714 ? ARM::R7 : ARM::R11;
3715 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3717 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3718 MachinePointerInfo(),
3719 false, false, false, 0);
3723 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
3724 /// and size(DestVec) > 128-bits.
3725 /// This is achieved by doing the one extension from the SrcVec, splitting the
3726 /// result, extending these parts, and then concatenating these into the
3728 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
3729 SDValue Op = N->getOperand(0);
3730 EVT SrcVT = Op.getValueType();
3731 EVT DestVT = N->getValueType(0);
3733 assert(DestVT.getSizeInBits() > 128 &&
3734 "Custom sext/zext expansion needs >128-bit vector.");
3735 // If this is a normal length extension, use the default expansion.
3736 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
3737 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
3741 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
3742 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
3743 unsigned NumElts = SrcVT.getVectorNumElements();
3744 LLVMContext &Ctx = *DAG.getContext();
3745 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;
3747 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3749 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3751 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
3754 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
3755 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3756 DAG.getIntPtrConstant(0));
3757 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3758 DAG.getIntPtrConstant(NumElts/2));
3759 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
3760 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
3761 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
3764 /// ExpandBITCAST - If the target supports VFP, this function is called to
3765 /// expand a bit convert where either the source or destination type is i64 to
3766 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3767 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3768 /// vectors), since the legalizer won't know what to do with that.
3769 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3770 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3772 SDValue Op = N->getOperand(0);
3774 // This function is only supposed to be called for i64 types, either as the
3775 // source or destination of the bit convert.
3776 EVT SrcVT = Op.getValueType();
3777 EVT DstVT = N->getValueType(0);
3778 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3779 "ExpandBITCAST called for non-i64 type");
3781 // Turn i64->f64 into VMOVDRR.
3782 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3783 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3784 DAG.getConstant(0, MVT::i32));
3785 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3786 DAG.getConstant(1, MVT::i32));
3787 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3788 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3791 // Turn f64->i64 into VMOVRRD.
3792 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3793 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3794 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3795 // Merge the pieces into a single i64 value.
3796 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3802 /// getZeroVector - Returns a vector of specified type with all zero elements.
3803 /// Zero vectors are used to represent vector negation and in those cases
3804 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3805 /// not support i64 elements, so sometimes the zero vectors will need to be
3806 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3808 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3809 assert(VT.isVector() && "Expected a vector type");
3810 // The canonical modified immediate encoding of a zero vector is....0!
3811 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3812 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3813 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3814 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3817 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3818 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3819 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3820 SelectionDAG &DAG) const {
3821 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3822 EVT VT = Op.getValueType();
3823 unsigned VTBits = VT.getSizeInBits();
3825 SDValue ShOpLo = Op.getOperand(0);
3826 SDValue ShOpHi = Op.getOperand(1);
3827 SDValue ShAmt = Op.getOperand(2);
3829 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3831 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3833 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3834 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3835 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3836 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3837 DAG.getConstant(VTBits, MVT::i32));
3838 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3839 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3840 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3842 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3843 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3845 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3846 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3849 SDValue Ops[2] = { Lo, Hi };
3850 return DAG.getMergeValues(Ops, 2, dl);
3853 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3854 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3855 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3856 SelectionDAG &DAG) const {
3857 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3858 EVT VT = Op.getValueType();
3859 unsigned VTBits = VT.getSizeInBits();
3861 SDValue ShOpLo = Op.getOperand(0);
3862 SDValue ShOpHi = Op.getOperand(1);
3863 SDValue ShAmt = Op.getOperand(2);
3866 assert(Op.getOpcode() == ISD::SHL_PARTS);
3867 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3868 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3869 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3870 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3871 DAG.getConstant(VTBits, MVT::i32));
3872 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3873 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3875 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3876 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3877 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3879 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3880 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3883 SDValue Ops[2] = { Lo, Hi };
3884 return DAG.getMergeValues(Ops, 2, dl);
3887 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3888 SelectionDAG &DAG) const {
3889 // The rounding mode is in bits 23:22 of the FPSCR.
3890 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3891 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3892 // so that the shift + and get folded into a bitfield extract.
3894 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3895 DAG.getConstant(Intrinsic::arm_get_fpscr,
3897 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3898 DAG.getConstant(1U << 22, MVT::i32));
3899 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3900 DAG.getConstant(22, MVT::i32));
3901 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3902 DAG.getConstant(3, MVT::i32));
3905 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3906 const ARMSubtarget *ST) {
3907 EVT VT = N->getValueType(0);
3910 if (!ST->hasV6T2Ops())
3913 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3914 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3917 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3918 /// for each 16-bit element from operand, repeated. The basic idea is to
3919 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3921 /// Trace for v4i16:
3922 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3923 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3924 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3925 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3926 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3927 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3928 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3929 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3930 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3931 EVT VT = N->getValueType(0);
3934 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3935 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3936 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3937 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3938 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3939 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3942 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3943 /// bit-count for each 16-bit element from the operand. We need slightly
3944 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3945 /// 64/128-bit registers.
3947 /// Trace for v4i16:
3948 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3949 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3950 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3951 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3952 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3953 EVT VT = N->getValueType(0);
3956 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3957 if (VT.is64BitVector()) {
3958 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3959 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3960 DAG.getIntPtrConstant(0));
3962 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3963 BitCounts, DAG.getIntPtrConstant(0));
3964 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3968 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3969 /// bit-count for each 32-bit element from the operand. The idea here is
3970 /// to split the vector into 16-bit elements, leverage the 16-bit count
3971 /// routine, and then combine the results.
3973 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3974 /// input = [v0 v1 ] (vi: 32-bit elements)
3975 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3976 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3977 /// vrev: N0 = [k1 k0 k3 k2 ]
3979 /// N1 =+[k1 k0 k3 k2 ]
3981 /// N2 =+[k1 k3 k0 k2 ]
3983 /// Extended =+[k1 k3 k0 k2 ]
3985 /// Extracted=+[k1 k3 ]
3987 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
3988 EVT VT = N->getValueType(0);
3991 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
3993 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
3994 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
3995 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
3996 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
3997 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
3999 if (VT.is64BitVector()) {
4000 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4001 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4002 DAG.getIntPtrConstant(0));
4004 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4005 DAG.getIntPtrConstant(0));
4006 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4010 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4011 const ARMSubtarget *ST) {
4012 EVT VT = N->getValueType(0);
4014 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4015 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4016 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4017 "Unexpected type for custom ctpop lowering");
4019 if (VT.getVectorElementType() == MVT::i32)
4020 return lowerCTPOP32BitElements(N, DAG);
4022 return lowerCTPOP16BitElements(N, DAG);
4025 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4026 const ARMSubtarget *ST) {
4027 EVT VT = N->getValueType(0);
4033 // Lower vector shifts on NEON to use VSHL.
4034 assert(ST->hasNEON() && "unexpected vector shift");
4036 // Left shifts translate directly to the vshiftu intrinsic.
4037 if (N->getOpcode() == ISD::SHL)
4038 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4039 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4040 N->getOperand(0), N->getOperand(1));
4042 assert((N->getOpcode() == ISD::SRA ||
4043 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4045 // NEON uses the same intrinsics for both left and right shifts. For
4046 // right shifts, the shift amounts are negative, so negate the vector of
4048 EVT ShiftVT = N->getOperand(1).getValueType();
4049 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4050 getZeroVector(ShiftVT, DAG, dl),
4052 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4053 Intrinsic::arm_neon_vshifts :
4054 Intrinsic::arm_neon_vshiftu);
4055 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4056 DAG.getConstant(vshiftInt, MVT::i32),
4057 N->getOperand(0), NegatedCount);
4060 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4061 const ARMSubtarget *ST) {
4062 EVT VT = N->getValueType(0);
4065 // We can get here for a node like i32 = ISD::SHL i32, i64
4069 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4070 "Unknown shift to lower!");
4072 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4073 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4074 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4077 // If we are in thumb mode, we don't have RRX.
4078 if (ST->isThumb1Only()) return SDValue();
4080 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4081 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4082 DAG.getConstant(0, MVT::i32));
4083 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4084 DAG.getConstant(1, MVT::i32));
4086 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4087 // captures the result into a carry flag.
4088 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4089 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4091 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4092 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4094 // Merge the pieces into a single i64 value.
4095 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4098 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4099 SDValue TmpOp0, TmpOp1;
4100 bool Invert = false;
4104 SDValue Op0 = Op.getOperand(0);
4105 SDValue Op1 = Op.getOperand(1);
4106 SDValue CC = Op.getOperand(2);
4107 EVT VT = Op.getValueType();
4108 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4111 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4112 switch (SetCCOpcode) {
4113 default: llvm_unreachable("Illegal FP comparison");
4115 case ISD::SETNE: Invert = true; // Fallthrough
4117 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4119 case ISD::SETLT: Swap = true; // Fallthrough
4121 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4123 case ISD::SETLE: Swap = true; // Fallthrough
4125 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4126 case ISD::SETUGE: Swap = true; // Fallthrough
4127 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4128 case ISD::SETUGT: Swap = true; // Fallthrough
4129 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4130 case ISD::SETUEQ: Invert = true; // Fallthrough
4132 // Expand this to (OLT | OGT).
4136 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4137 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4139 case ISD::SETUO: Invert = true; // Fallthrough
4141 // Expand this to (OLT | OGE).
4145 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4146 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4150 // Integer comparisons.
4151 switch (SetCCOpcode) {
4152 default: llvm_unreachable("Illegal integer comparison");
4153 case ISD::SETNE: Invert = true;
4154 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4155 case ISD::SETLT: Swap = true;
4156 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4157 case ISD::SETLE: Swap = true;
4158 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4159 case ISD::SETULT: Swap = true;
4160 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4161 case ISD::SETULE: Swap = true;
4162 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4165 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4166 if (Opc == ARMISD::VCEQ) {
4169 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4171 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4174 // Ignore bitconvert.
4175 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4176 AndOp = AndOp.getOperand(0);
4178 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4180 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4181 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4188 std::swap(Op0, Op1);
4190 // If one of the operands is a constant vector zero, attempt to fold the
4191 // comparison to a specialized compare-against-zero form.
4193 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4195 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4196 if (Opc == ARMISD::VCGE)
4197 Opc = ARMISD::VCLEZ;
4198 else if (Opc == ARMISD::VCGT)
4199 Opc = ARMISD::VCLTZ;
4204 if (SingleOp.getNode()) {
4207 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4209 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4211 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4213 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4215 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4217 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4220 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4224 Result = DAG.getNOT(dl, Result, VT);
4229 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4230 /// valid vector constant for a NEON instruction with a "modified immediate"
4231 /// operand (e.g., VMOV). If so, return the encoded value.
4232 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4233 unsigned SplatBitSize, SelectionDAG &DAG,
4234 EVT &VT, bool is128Bits, NEONModImmType type) {
4235 unsigned OpCmode, Imm;
4237 // SplatBitSize is set to the smallest size that splats the vector, so a
4238 // zero vector will always have SplatBitSize == 8. However, NEON modified
4239 // immediate instructions others than VMOV do not support the 8-bit encoding
4240 // of a zero vector, and the default encoding of zero is supposed to be the
4245 switch (SplatBitSize) {
4247 if (type != VMOVModImm)
4249 // Any 1-byte value is OK. Op=0, Cmode=1110.
4250 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4253 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4257 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4258 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4259 if ((SplatBits & ~0xff) == 0) {
4260 // Value = 0x00nn: Op=x, Cmode=100x.
4265 if ((SplatBits & ~0xff00) == 0) {
4266 // Value = 0xnn00: Op=x, Cmode=101x.
4268 Imm = SplatBits >> 8;
4274 // NEON's 32-bit VMOV supports splat values where:
4275 // * only one byte is nonzero, or
4276 // * the least significant byte is 0xff and the second byte is nonzero, or
4277 // * the least significant 2 bytes are 0xff and the third is nonzero.
4278 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4279 if ((SplatBits & ~0xff) == 0) {
4280 // Value = 0x000000nn: Op=x, Cmode=000x.
4285 if ((SplatBits & ~0xff00) == 0) {
4286 // Value = 0x0000nn00: Op=x, Cmode=001x.
4288 Imm = SplatBits >> 8;
4291 if ((SplatBits & ~0xff0000) == 0) {
4292 // Value = 0x00nn0000: Op=x, Cmode=010x.
4294 Imm = SplatBits >> 16;
4297 if ((SplatBits & ~0xff000000) == 0) {
4298 // Value = 0xnn000000: Op=x, Cmode=011x.
4300 Imm = SplatBits >> 24;
4304 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4305 if (type == OtherModImm) return SDValue();
4307 if ((SplatBits & ~0xffff) == 0 &&
4308 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4309 // Value = 0x0000nnff: Op=x, Cmode=1100.
4311 Imm = SplatBits >> 8;
4316 if ((SplatBits & ~0xffffff) == 0 &&
4317 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4318 // Value = 0x00nnffff: Op=x, Cmode=1101.
4320 Imm = SplatBits >> 16;
4321 SplatBits |= 0xffff;
4325 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4326 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4327 // VMOV.I32. A (very) minor optimization would be to replicate the value
4328 // and fall through here to test for a valid 64-bit splat. But, then the
4329 // caller would also need to check and handle the change in size.
4333 if (type != VMOVModImm)
4335 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4336 uint64_t BitMask = 0xff;
4338 unsigned ImmMask = 1;
4340 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4341 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4344 } else if ((SplatBits & BitMask) != 0) {
4350 // Op=1, Cmode=1110.
4353 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4358 llvm_unreachable("unexpected size for isNEONModifiedImm");
4361 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4362 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4365 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4366 const ARMSubtarget *ST) const {
4370 bool IsDouble = Op.getValueType() == MVT::f64;
4371 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4373 // Try splatting with a VMOV.f32...
4374 APFloat FPVal = CFP->getValueAPF();
4375 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4378 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4379 // We have code in place to select a valid ConstantFP already, no need to
4384 // It's a float and we are trying to use NEON operations where
4385 // possible. Lower it to a splat followed by an extract.
4387 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4388 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4390 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4391 DAG.getConstant(0, MVT::i32));
4394 // The rest of our options are NEON only, make sure that's allowed before
4396 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4400 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4402 // It wouldn't really be worth bothering for doubles except for one very
4403 // important value, which does happen to match: 0.0. So make sure we don't do
4405 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4408 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4409 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4411 if (NewVal != SDValue()) {
4413 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4416 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4418 // It's a float: cast and extract a vector element.
4419 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4421 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4422 DAG.getConstant(0, MVT::i32));
4425 // Finally, try a VMVN.i32
4426 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4428 if (NewVal != SDValue()) {
4430 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4433 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4435 // It's a float: cast and extract a vector element.
4436 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4438 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4439 DAG.getConstant(0, MVT::i32));
4445 // check if an VEXT instruction can handle the shuffle mask when the
4446 // vector sources of the shuffle are the same.
4447 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4448 unsigned NumElts = VT.getVectorNumElements();
4450 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4456 // If this is a VEXT shuffle, the immediate value is the index of the first
4457 // element. The other shuffle indices must be the successive elements after
4459 unsigned ExpectedElt = Imm;
4460 for (unsigned i = 1; i < NumElts; ++i) {
4461 // Increment the expected index. If it wraps around, just follow it
4462 // back to index zero and keep going.
4464 if (ExpectedElt == NumElts)
4467 if (M[i] < 0) continue; // ignore UNDEF indices
4468 if (ExpectedElt != static_cast<unsigned>(M[i]))
4476 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4477 bool &ReverseVEXT, unsigned &Imm) {
4478 unsigned NumElts = VT.getVectorNumElements();
4479 ReverseVEXT = false;
4481 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4487 // If this is a VEXT shuffle, the immediate value is the index of the first
4488 // element. The other shuffle indices must be the successive elements after
4490 unsigned ExpectedElt = Imm;
4491 for (unsigned i = 1; i < NumElts; ++i) {
4492 // Increment the expected index. If it wraps around, it may still be
4493 // a VEXT but the source vectors must be swapped.
4495 if (ExpectedElt == NumElts * 2) {
4500 if (M[i] < 0) continue; // ignore UNDEF indices
4501 if (ExpectedElt != static_cast<unsigned>(M[i]))
4505 // Adjust the index value if the source operands will be swapped.
4512 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4513 /// instruction with the specified blocksize. (The order of the elements
4514 /// within each block of the vector is reversed.)
4515 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4516 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4517 "Only possible block sizes for VREV are: 16, 32, 64");
4519 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4523 unsigned NumElts = VT.getVectorNumElements();
4524 unsigned BlockElts = M[0] + 1;
4525 // If the first shuffle index is UNDEF, be optimistic.
4527 BlockElts = BlockSize / EltSz;
4529 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4532 for (unsigned i = 0; i < NumElts; ++i) {
4533 if (M[i] < 0) continue; // ignore UNDEF indices
4534 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4541 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4542 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4543 // range, then 0 is placed into the resulting vector. So pretty much any mask
4544 // of 8 elements can work here.
4545 return VT == MVT::v8i8 && M.size() == 8;
4548 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4549 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4553 unsigned NumElts = VT.getVectorNumElements();
4554 WhichResult = (M[0] == 0 ? 0 : 1);
4555 for (unsigned i = 0; i < NumElts; i += 2) {
4556 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4557 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4563 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4564 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4565 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4566 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4567 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4571 unsigned NumElts = VT.getVectorNumElements();
4572 WhichResult = (M[0] == 0 ? 0 : 1);
4573 for (unsigned i = 0; i < NumElts; i += 2) {
4574 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4575 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4581 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4582 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4586 unsigned NumElts = VT.getVectorNumElements();
4587 WhichResult = (M[0] == 0 ? 0 : 1);
4588 for (unsigned i = 0; i != NumElts; ++i) {
4589 if (M[i] < 0) continue; // ignore UNDEF indices
4590 if ((unsigned) M[i] != 2 * i + WhichResult)
4594 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4595 if (VT.is64BitVector() && EltSz == 32)
4601 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4602 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4603 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4604 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4605 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4609 unsigned Half = VT.getVectorNumElements() / 2;
4610 WhichResult = (M[0] == 0 ? 0 : 1);
4611 for (unsigned j = 0; j != 2; ++j) {
4612 unsigned Idx = WhichResult;
4613 for (unsigned i = 0; i != Half; ++i) {
4614 int MIdx = M[i + j * Half];
4615 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4621 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4622 if (VT.is64BitVector() && EltSz == 32)
4628 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4629 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4633 unsigned NumElts = VT.getVectorNumElements();
4634 WhichResult = (M[0] == 0 ? 0 : 1);
4635 unsigned Idx = WhichResult * NumElts / 2;
4636 for (unsigned i = 0; i != NumElts; i += 2) {
4637 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4638 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4643 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4644 if (VT.is64BitVector() && EltSz == 32)
4650 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4651 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4652 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4653 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4654 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4658 unsigned NumElts = VT.getVectorNumElements();
4659 WhichResult = (M[0] == 0 ? 0 : 1);
4660 unsigned Idx = WhichResult * NumElts / 2;
4661 for (unsigned i = 0; i != NumElts; i += 2) {
4662 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4663 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4668 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4669 if (VT.is64BitVector() && EltSz == 32)
4675 /// \return true if this is a reverse operation on an vector.
4676 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4677 unsigned NumElts = VT.getVectorNumElements();
4678 // Make sure the mask has the right size.
4679 if (NumElts != M.size())
4682 // Look for <15, ..., 3, -1, 1, 0>.
4683 for (unsigned i = 0; i != NumElts; ++i)
4684 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4690 // If N is an integer constant that can be moved into a register in one
4691 // instruction, return an SDValue of such a constant (will become a MOV
4692 // instruction). Otherwise return null.
4693 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4694 const ARMSubtarget *ST, SDLoc dl) {
4696 if (!isa<ConstantSDNode>(N))
4698 Val = cast<ConstantSDNode>(N)->getZExtValue();
4700 if (ST->isThumb1Only()) {
4701 if (Val <= 255 || ~Val <= 255)
4702 return DAG.getConstant(Val, MVT::i32);
4704 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4705 return DAG.getConstant(Val, MVT::i32);
4710 // If this is a case we can't handle, return null and let the default
4711 // expansion code take care of it.
4712 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4713 const ARMSubtarget *ST) const {
4714 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4716 EVT VT = Op.getValueType();
4718 APInt SplatBits, SplatUndef;
4719 unsigned SplatBitSize;
4721 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4722 if (SplatBitSize <= 64) {
4723 // Check if an immediate VMOV works.
4725 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4726 SplatUndef.getZExtValue(), SplatBitSize,
4727 DAG, VmovVT, VT.is128BitVector(),
4729 if (Val.getNode()) {
4730 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4731 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4734 // Try an immediate VMVN.
4735 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4736 Val = isNEONModifiedImm(NegatedImm,
4737 SplatUndef.getZExtValue(), SplatBitSize,
4738 DAG, VmovVT, VT.is128BitVector(),
4740 if (Val.getNode()) {
4741 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4742 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4745 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4746 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4747 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4749 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4750 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4756 // Scan through the operands to see if only one value is used.
4758 // As an optimisation, even if more than one value is used it may be more
4759 // profitable to splat with one value then change some lanes.
4761 // Heuristically we decide to do this if the vector has a "dominant" value,
4762 // defined as splatted to more than half of the lanes.
4763 unsigned NumElts = VT.getVectorNumElements();
4764 bool isOnlyLowElement = true;
4765 bool usesOnlyOneValue = true;
4766 bool hasDominantValue = false;
4767 bool isConstant = true;
4769 // Map of the number of times a particular SDValue appears in the
4771 DenseMap<SDValue, unsigned> ValueCounts;
4773 for (unsigned i = 0; i < NumElts; ++i) {
4774 SDValue V = Op.getOperand(i);
4775 if (V.getOpcode() == ISD::UNDEF)
4778 isOnlyLowElement = false;
4779 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4782 ValueCounts.insert(std::make_pair(V, 0));
4783 unsigned &Count = ValueCounts[V];
4785 // Is this value dominant? (takes up more than half of the lanes)
4786 if (++Count > (NumElts / 2)) {
4787 hasDominantValue = true;
4791 if (ValueCounts.size() != 1)
4792 usesOnlyOneValue = false;
4793 if (!Value.getNode() && ValueCounts.size() > 0)
4794 Value = ValueCounts.begin()->first;
4796 if (ValueCounts.size() == 0)
4797 return DAG.getUNDEF(VT);
4799 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4800 // Keep going if we are hitting this case.
4801 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4802 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4804 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4806 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4807 // i32 and try again.
4808 if (hasDominantValue && EltSize <= 32) {
4812 // If we are VDUPing a value that comes directly from a vector, that will
4813 // cause an unnecessary move to and from a GPR, where instead we could
4814 // just use VDUPLANE. We can only do this if the lane being extracted
4815 // is at a constant index, as the VDUP from lane instructions only have
4816 // constant-index forms.
4817 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4818 isa<ConstantSDNode>(Value->getOperand(1))) {
4819 // We need to create a new undef vector to use for the VDUPLANE if the
4820 // size of the vector from which we get the value is different than the
4821 // size of the vector that we need to create. We will insert the element
4822 // such that the register coalescer will remove unnecessary copies.
4823 if (VT != Value->getOperand(0).getValueType()) {
4824 ConstantSDNode *constIndex;
4825 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4826 assert(constIndex && "The index is not a constant!");
4827 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4828 VT.getVectorNumElements();
4829 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4830 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4831 Value, DAG.getConstant(index, MVT::i32)),
4832 DAG.getConstant(index, MVT::i32));
4834 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4835 Value->getOperand(0), Value->getOperand(1));
4837 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4839 if (!usesOnlyOneValue) {
4840 // The dominant value was splatted as 'N', but we now have to insert
4841 // all differing elements.
4842 for (unsigned I = 0; I < NumElts; ++I) {
4843 if (Op.getOperand(I) == Value)
4845 SmallVector<SDValue, 3> Ops;
4847 Ops.push_back(Op.getOperand(I));
4848 Ops.push_back(DAG.getConstant(I, MVT::i32));
4849 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4854 if (VT.getVectorElementType().isFloatingPoint()) {
4855 SmallVector<SDValue, 8> Ops;
4856 for (unsigned i = 0; i < NumElts; ++i)
4857 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4859 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4860 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4861 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4863 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4865 if (usesOnlyOneValue) {
4866 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4867 if (isConstant && Val.getNode())
4868 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4872 // If all elements are constants and the case above didn't get hit, fall back
4873 // to the default expansion, which will generate a load from the constant
4878 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4880 SDValue shuffle = ReconstructShuffle(Op, DAG);
4881 if (shuffle != SDValue())
4885 // Vectors with 32- or 64-bit elements can be built by directly assigning
4886 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4887 // will be legalized.
4888 if (EltSize >= 32) {
4889 // Do the expansion with floating-point types, since that is what the VFP
4890 // registers are defined to use, and since i64 is not legal.
4891 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4892 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4893 SmallVector<SDValue, 8> Ops;
4894 for (unsigned i = 0; i < NumElts; ++i)
4895 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4896 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4897 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4900 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4901 // know the default expansion would otherwise fall back on something even
4902 // worse. For a vector with one or two non-undef values, that's
4903 // scalar_to_vector for the elements followed by a shuffle (provided the
4904 // shuffle is valid for the target) and materialization element by element
4905 // on the stack followed by a load for everything else.
4906 if (!isConstant && !usesOnlyOneValue) {
4907 SDValue Vec = DAG.getUNDEF(VT);
4908 for (unsigned i = 0 ; i < NumElts; ++i) {
4909 SDValue V = Op.getOperand(i);
4910 if (V.getOpcode() == ISD::UNDEF)
4912 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4913 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4921 // Gather data to see if the operation can be modelled as a
4922 // shuffle in combination with VEXTs.
4923 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4924 SelectionDAG &DAG) const {
4926 EVT VT = Op.getValueType();
4927 unsigned NumElts = VT.getVectorNumElements();
4929 SmallVector<SDValue, 2> SourceVecs;
4930 SmallVector<unsigned, 2> MinElts;
4931 SmallVector<unsigned, 2> MaxElts;
4933 for (unsigned i = 0; i < NumElts; ++i) {
4934 SDValue V = Op.getOperand(i);
4935 if (V.getOpcode() == ISD::UNDEF)
4937 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4938 // A shuffle can only come from building a vector from various
4939 // elements of other vectors.
4941 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4942 VT.getVectorElementType()) {
4943 // This code doesn't know how to handle shuffles where the vector
4944 // element types do not match (this happens because type legalization
4945 // promotes the return type of EXTRACT_VECTOR_ELT).
4946 // FIXME: It might be appropriate to extend this code to handle
4947 // mismatched types.
4951 // Record this extraction against the appropriate vector if possible...
4952 SDValue SourceVec = V.getOperand(0);
4953 // If the element number isn't a constant, we can't effectively
4954 // analyze what's going on.
4955 if (!isa<ConstantSDNode>(V.getOperand(1)))
4957 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4958 bool FoundSource = false;
4959 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4960 if (SourceVecs[j] == SourceVec) {
4961 if (MinElts[j] > EltNo)
4963 if (MaxElts[j] < EltNo)
4970 // Or record a new source if not...
4972 SourceVecs.push_back(SourceVec);
4973 MinElts.push_back(EltNo);
4974 MaxElts.push_back(EltNo);
4978 // Currently only do something sane when at most two source vectors
4980 if (SourceVecs.size() > 2)
4983 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4984 int VEXTOffsets[2] = {0, 0};
4986 // This loop extracts the usage patterns of the source vectors
4987 // and prepares appropriate SDValues for a shuffle if possible.
4988 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4989 if (SourceVecs[i].getValueType() == VT) {
4990 // No VEXT necessary
4991 ShuffleSrcs[i] = SourceVecs[i];
4994 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4995 // It probably isn't worth padding out a smaller vector just to
4996 // break it down again in a shuffle.
5000 // Since only 64-bit and 128-bit vectors are legal on ARM and
5001 // we've eliminated the other cases...
5002 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5003 "unexpected vector sizes in ReconstructShuffle");
5005 if (MaxElts[i] - MinElts[i] >= NumElts) {
5006 // Span too large for a VEXT to cope
5010 if (MinElts[i] >= NumElts) {
5011 // The extraction can just take the second half
5012 VEXTOffsets[i] = NumElts;
5013 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5015 DAG.getIntPtrConstant(NumElts));
5016 } else if (MaxElts[i] < NumElts) {
5017 // The extraction can just take the first half
5019 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5021 DAG.getIntPtrConstant(0));
5023 // An actual VEXT is needed
5024 VEXTOffsets[i] = MinElts[i];
5025 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5027 DAG.getIntPtrConstant(0));
5028 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5030 DAG.getIntPtrConstant(NumElts));
5031 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5032 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5036 SmallVector<int, 8> Mask;
5038 for (unsigned i = 0; i < NumElts; ++i) {
5039 SDValue Entry = Op.getOperand(i);
5040 if (Entry.getOpcode() == ISD::UNDEF) {
5045 SDValue ExtractVec = Entry.getOperand(0);
5046 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5047 .getOperand(1))->getSExtValue();
5048 if (ExtractVec == SourceVecs[0]) {
5049 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5051 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5055 // Final check before we try to produce nonsense...
5056 if (isShuffleMaskLegal(Mask, VT))
5057 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5063 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5064 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5065 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5066 /// are assumed to be legal.
5068 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5070 if (VT.getVectorNumElements() == 4 &&
5071 (VT.is128BitVector() || VT.is64BitVector())) {
5072 unsigned PFIndexes[4];
5073 for (unsigned i = 0; i != 4; ++i) {
5077 PFIndexes[i] = M[i];
5080 // Compute the index in the perfect shuffle table.
5081 unsigned PFTableIndex =
5082 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5083 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5084 unsigned Cost = (PFEntry >> 30);
5091 unsigned Imm, WhichResult;
5093 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5094 return (EltSize >= 32 ||
5095 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5096 isVREVMask(M, VT, 64) ||
5097 isVREVMask(M, VT, 32) ||
5098 isVREVMask(M, VT, 16) ||
5099 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5100 isVTBLMask(M, VT) ||
5101 isVTRNMask(M, VT, WhichResult) ||
5102 isVUZPMask(M, VT, WhichResult) ||
5103 isVZIPMask(M, VT, WhichResult) ||
5104 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5105 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5106 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5107 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5110 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5111 /// the specified operations to build the shuffle.
5112 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5113 SDValue RHS, SelectionDAG &DAG,
5115 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5116 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5117 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5120 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5129 OP_VUZPL, // VUZP, left result
5130 OP_VUZPR, // VUZP, right result
5131 OP_VZIPL, // VZIP, left result
5132 OP_VZIPR, // VZIP, right result
5133 OP_VTRNL, // VTRN, left result
5134 OP_VTRNR // VTRN, right result
5137 if (OpNum == OP_COPY) {
5138 if (LHSID == (1*9+2)*9+3) return LHS;
5139 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5143 SDValue OpLHS, OpRHS;
5144 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5145 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5146 EVT VT = OpLHS.getValueType();
5149 default: llvm_unreachable("Unknown shuffle opcode!");
5151 // VREV divides the vector in half and swaps within the half.
5152 if (VT.getVectorElementType() == MVT::i32 ||
5153 VT.getVectorElementType() == MVT::f32)
5154 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5155 // vrev <4 x i16> -> VREV32
5156 if (VT.getVectorElementType() == MVT::i16)
5157 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5158 // vrev <4 x i8> -> VREV16
5159 assert(VT.getVectorElementType() == MVT::i8);
5160 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5165 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5166 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5170 return DAG.getNode(ARMISD::VEXT, dl, VT,
5172 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5175 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5176 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5179 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5180 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5183 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5184 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5188 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5189 ArrayRef<int> ShuffleMask,
5190 SelectionDAG &DAG) {
5191 // Check to see if we can use the VTBL instruction.
5192 SDValue V1 = Op.getOperand(0);
5193 SDValue V2 = Op.getOperand(1);
5196 SmallVector<SDValue, 8> VTBLMask;
5197 for (ArrayRef<int>::iterator
5198 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5199 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5201 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5202 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5203 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5206 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5207 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5211 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5212 SelectionDAG &DAG) {
5214 SDValue OpLHS = Op.getOperand(0);
5215 EVT VT = OpLHS.getValueType();
5217 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5218 "Expect an v8i16/v16i8 type");
5219 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5220 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5221 // extract the first 8 bytes into the top double word and the last 8 bytes
5222 // into the bottom double word. The v8i16 case is similar.
5223 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5224 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5225 DAG.getConstant(ExtractNum, MVT::i32));
5228 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5229 SDValue V1 = Op.getOperand(0);
5230 SDValue V2 = Op.getOperand(1);
5232 EVT VT = Op.getValueType();
5233 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5235 // Convert shuffles that are directly supported on NEON to target-specific
5236 // DAG nodes, instead of keeping them as shuffles and matching them again
5237 // during code selection. This is more efficient and avoids the possibility
5238 // of inconsistencies between legalization and selection.
5239 // FIXME: floating-point vectors should be canonicalized to integer vectors
5240 // of the same time so that they get CSEd properly.
5241 ArrayRef<int> ShuffleMask = SVN->getMask();
5243 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5244 if (EltSize <= 32) {
5245 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5246 int Lane = SVN->getSplatIndex();
5247 // If this is undef splat, generate it via "just" vdup, if possible.
5248 if (Lane == -1) Lane = 0;
5250 // Test if V1 is a SCALAR_TO_VECTOR.
5251 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5252 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5254 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5255 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5257 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5258 !isa<ConstantSDNode>(V1.getOperand(0))) {
5259 bool IsScalarToVector = true;
5260 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5261 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5262 IsScalarToVector = false;
5265 if (IsScalarToVector)
5266 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5268 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5269 DAG.getConstant(Lane, MVT::i32));
5274 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5277 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5278 DAG.getConstant(Imm, MVT::i32));
5281 if (isVREVMask(ShuffleMask, VT, 64))
5282 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5283 if (isVREVMask(ShuffleMask, VT, 32))
5284 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5285 if (isVREVMask(ShuffleMask, VT, 16))
5286 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5288 if (V2->getOpcode() == ISD::UNDEF &&
5289 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5290 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5291 DAG.getConstant(Imm, MVT::i32));
5294 // Check for Neon shuffles that modify both input vectors in place.
5295 // If both results are used, i.e., if there are two shuffles with the same
5296 // source operands and with masks corresponding to both results of one of
5297 // these operations, DAG memoization will ensure that a single node is
5298 // used for both shuffles.
5299 unsigned WhichResult;
5300 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5301 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5302 V1, V2).getValue(WhichResult);
5303 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5304 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5305 V1, V2).getValue(WhichResult);
5306 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5307 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5308 V1, V2).getValue(WhichResult);
5310 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5311 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5312 V1, V1).getValue(WhichResult);
5313 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5314 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5315 V1, V1).getValue(WhichResult);
5316 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5317 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5318 V1, V1).getValue(WhichResult);
5321 // If the shuffle is not directly supported and it has 4 elements, use
5322 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5323 unsigned NumElts = VT.getVectorNumElements();
5325 unsigned PFIndexes[4];
5326 for (unsigned i = 0; i != 4; ++i) {
5327 if (ShuffleMask[i] < 0)
5330 PFIndexes[i] = ShuffleMask[i];
5333 // Compute the index in the perfect shuffle table.
5334 unsigned PFTableIndex =
5335 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5336 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5337 unsigned Cost = (PFEntry >> 30);
5340 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5343 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5344 if (EltSize >= 32) {
5345 // Do the expansion with floating-point types, since that is what the VFP
5346 // registers are defined to use, and since i64 is not legal.
5347 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5348 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5349 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5350 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5351 SmallVector<SDValue, 8> Ops;
5352 for (unsigned i = 0; i < NumElts; ++i) {
5353 if (ShuffleMask[i] < 0)
5354 Ops.push_back(DAG.getUNDEF(EltVT));
5356 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5357 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5358 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5361 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5362 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5365 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5366 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5368 if (VT == MVT::v8i8) {
5369 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5370 if (NewOp.getNode())
5377 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5378 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5379 SDValue Lane = Op.getOperand(2);
5380 if (!isa<ConstantSDNode>(Lane))
5386 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5387 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5388 SDValue Lane = Op.getOperand(1);
5389 if (!isa<ConstantSDNode>(Lane))
5392 SDValue Vec = Op.getOperand(0);
5393 if (Op.getValueType() == MVT::i32 &&
5394 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5396 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5402 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5403 // The only time a CONCAT_VECTORS operation can have legal types is when
5404 // two 64-bit vectors are concatenated to a 128-bit vector.
5405 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5406 "unexpected CONCAT_VECTORS");
5408 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5409 SDValue Op0 = Op.getOperand(0);
5410 SDValue Op1 = Op.getOperand(1);
5411 if (Op0.getOpcode() != ISD::UNDEF)
5412 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5413 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5414 DAG.getIntPtrConstant(0));
5415 if (Op1.getOpcode() != ISD::UNDEF)
5416 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5417 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5418 DAG.getIntPtrConstant(1));
5419 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5422 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5423 /// element has been zero/sign-extended, depending on the isSigned parameter,
5424 /// from an integer type half its size.
5425 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5427 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5428 EVT VT = N->getValueType(0);
5429 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5430 SDNode *BVN = N->getOperand(0).getNode();
5431 if (BVN->getValueType(0) != MVT::v4i32 ||
5432 BVN->getOpcode() != ISD::BUILD_VECTOR)
5434 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5435 unsigned HiElt = 1 - LoElt;
5436 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5437 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5438 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5439 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5440 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5443 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5444 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5447 if (Hi0->isNullValue() && Hi1->isNullValue())
5453 if (N->getOpcode() != ISD::BUILD_VECTOR)
5456 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5457 SDNode *Elt = N->getOperand(i).getNode();
5458 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5459 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5460 unsigned HalfSize = EltSize / 2;
5462 if (!isIntN(HalfSize, C->getSExtValue()))
5465 if (!isUIntN(HalfSize, C->getZExtValue()))
5476 /// isSignExtended - Check if a node is a vector value that is sign-extended
5477 /// or a constant BUILD_VECTOR with sign-extended elements.
5478 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5479 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5481 if (isExtendedBUILD_VECTOR(N, DAG, true))
5486 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5487 /// or a constant BUILD_VECTOR with zero-extended elements.
5488 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5489 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5491 if (isExtendedBUILD_VECTOR(N, DAG, false))
5496 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5497 if (OrigVT.getSizeInBits() >= 64)
5500 assert(OrigVT.isSimple() && "Expecting a simple value type");
5502 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5503 switch (OrigSimpleTy) {
5504 default: llvm_unreachable("Unexpected Vector Type");
5513 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5514 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5515 /// We insert the required extension here to get the vector to fill a D register.
5516 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5519 unsigned ExtOpcode) {
5520 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5521 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5522 // 64-bits we need to insert a new extension so that it will be 64-bits.
5523 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5524 if (OrigTy.getSizeInBits() >= 64)
5527 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5528 EVT NewVT = getExtensionTo64Bits(OrigTy);
5530 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5533 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5534 /// does not do any sign/zero extension. If the original vector is less
5535 /// than 64 bits, an appropriate extension will be added after the load to
5536 /// reach a total size of 64 bits. We have to add the extension separately
5537 /// because ARM does not have a sign/zero extending load for vectors.
5538 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5539 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5541 // The load already has the right type.
5542 if (ExtendedTy == LD->getMemoryVT())
5543 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5544 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5545 LD->isNonTemporal(), LD->isInvariant(),
5546 LD->getAlignment());
5548 // We need to create a zextload/sextload. We cannot just create a load
5549 // followed by a zext/zext node because LowerMUL is also run during normal
5550 // operation legalization where we can't create illegal types.
5551 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5552 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5553 LD->getMemoryVT(), LD->isVolatile(),
5554 LD->isNonTemporal(), LD->getAlignment());
5557 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5558 /// extending load, or BUILD_VECTOR with extended elements, return the
5559 /// unextended value. The unextended vector should be 64 bits so that it can
5560 /// be used as an operand to a VMULL instruction. If the original vector size
5561 /// before extension is less than 64 bits we add a an extension to resize
5562 /// the vector to 64 bits.
5563 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5564 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5565 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5566 N->getOperand(0)->getValueType(0),
5570 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5571 return SkipLoadExtensionForVMULL(LD, DAG);
5573 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5574 // have been legalized as a BITCAST from v4i32.
5575 if (N->getOpcode() == ISD::BITCAST) {
5576 SDNode *BVN = N->getOperand(0).getNode();
5577 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5578 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5579 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5580 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5581 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5583 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5584 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5585 EVT VT = N->getValueType(0);
5586 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5587 unsigned NumElts = VT.getVectorNumElements();
5588 MVT TruncVT = MVT::getIntegerVT(EltSize);
5589 SmallVector<SDValue, 8> Ops;
5590 for (unsigned i = 0; i != NumElts; ++i) {
5591 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5592 const APInt &CInt = C->getAPIntValue();
5593 // Element types smaller than 32 bits are not legal, so use i32 elements.
5594 // The values are implicitly truncated so sext vs. zext doesn't matter.
5595 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5597 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5598 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5601 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5602 unsigned Opcode = N->getOpcode();
5603 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5604 SDNode *N0 = N->getOperand(0).getNode();
5605 SDNode *N1 = N->getOperand(1).getNode();
5606 return N0->hasOneUse() && N1->hasOneUse() &&
5607 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5612 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5613 unsigned Opcode = N->getOpcode();
5614 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5615 SDNode *N0 = N->getOperand(0).getNode();
5616 SDNode *N1 = N->getOperand(1).getNode();
5617 return N0->hasOneUse() && N1->hasOneUse() &&
5618 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5623 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5624 // Multiplications are only custom-lowered for 128-bit vectors so that
5625 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5626 EVT VT = Op.getValueType();
5627 assert(VT.is128BitVector() && VT.isInteger() &&
5628 "unexpected type for custom-lowering ISD::MUL");
5629 SDNode *N0 = Op.getOperand(0).getNode();
5630 SDNode *N1 = Op.getOperand(1).getNode();
5631 unsigned NewOpc = 0;
5633 bool isN0SExt = isSignExtended(N0, DAG);
5634 bool isN1SExt = isSignExtended(N1, DAG);
5635 if (isN0SExt && isN1SExt)
5636 NewOpc = ARMISD::VMULLs;
5638 bool isN0ZExt = isZeroExtended(N0, DAG);
5639 bool isN1ZExt = isZeroExtended(N1, DAG);
5640 if (isN0ZExt && isN1ZExt)
5641 NewOpc = ARMISD::VMULLu;
5642 else if (isN1SExt || isN1ZExt) {
5643 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5644 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5645 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5646 NewOpc = ARMISD::VMULLs;
5648 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5649 NewOpc = ARMISD::VMULLu;
5651 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5653 NewOpc = ARMISD::VMULLu;
5659 if (VT == MVT::v2i64)
5660 // Fall through to expand this. It is not legal.
5663 // Other vector multiplications are legal.
5668 // Legalize to a VMULL instruction.
5671 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5673 Op0 = SkipExtensionForVMULL(N0, DAG);
5674 assert(Op0.getValueType().is64BitVector() &&
5675 Op1.getValueType().is64BitVector() &&
5676 "unexpected types for extended operands to VMULL");
5677 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5680 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5681 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5688 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5689 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5690 EVT Op1VT = Op1.getValueType();
5691 return DAG.getNode(N0->getOpcode(), DL, VT,
5692 DAG.getNode(NewOpc, DL, VT,
5693 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5694 DAG.getNode(NewOpc, DL, VT,
5695 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5699 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5701 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5702 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5703 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5704 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5705 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5706 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5707 // Get reciprocal estimate.
5708 // float4 recip = vrecpeq_f32(yf);
5709 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5710 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5711 // Because char has a smaller range than uchar, we can actually get away
5712 // without any newton steps. This requires that we use a weird bias
5713 // of 0xb000, however (again, this has been exhaustively tested).
5714 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5715 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5716 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5717 Y = DAG.getConstant(0xb000, MVT::i32);
5718 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5719 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5720 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5721 // Convert back to short.
5722 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5723 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5728 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5730 // Convert to float.
5731 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5732 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5733 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5734 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5735 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5736 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5738 // Use reciprocal estimate and one refinement step.
5739 // float4 recip = vrecpeq_f32(yf);
5740 // recip *= vrecpsq_f32(yf, recip);
5741 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5742 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5743 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5744 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5746 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5747 // Because short has a smaller range than ushort, we can actually get away
5748 // with only a single newton step. This requires that we use a weird bias
5749 // of 89, however (again, this has been exhaustively tested).
5750 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5751 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5752 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5753 N1 = DAG.getConstant(0x89, MVT::i32);
5754 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5755 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5756 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5757 // Convert back to integer and return.
5758 // return vmovn_s32(vcvt_s32_f32(result));
5759 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5760 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5764 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5765 EVT VT = Op.getValueType();
5766 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5767 "unexpected type for custom-lowering ISD::SDIV");
5770 SDValue N0 = Op.getOperand(0);
5771 SDValue N1 = Op.getOperand(1);
5774 if (VT == MVT::v8i8) {
5775 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5776 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5778 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5779 DAG.getIntPtrConstant(4));
5780 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5781 DAG.getIntPtrConstant(4));
5782 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5783 DAG.getIntPtrConstant(0));
5784 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5785 DAG.getIntPtrConstant(0));
5787 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5788 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5790 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5791 N0 = LowerCONCAT_VECTORS(N0, DAG);
5793 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5796 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5799 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5800 EVT VT = Op.getValueType();
5801 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5802 "unexpected type for custom-lowering ISD::UDIV");
5805 SDValue N0 = Op.getOperand(0);
5806 SDValue N1 = Op.getOperand(1);
5809 if (VT == MVT::v8i8) {
5810 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5811 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5813 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5814 DAG.getIntPtrConstant(4));
5815 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5816 DAG.getIntPtrConstant(4));
5817 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5818 DAG.getIntPtrConstant(0));
5819 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5820 DAG.getIntPtrConstant(0));
5822 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5823 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5825 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5826 N0 = LowerCONCAT_VECTORS(N0, DAG);
5828 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5829 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5834 // v4i16 sdiv ... Convert to float.
5835 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5836 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5837 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5838 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5839 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5840 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5842 // Use reciprocal estimate and two refinement steps.
5843 // float4 recip = vrecpeq_f32(yf);
5844 // recip *= vrecpsq_f32(yf, recip);
5845 // recip *= vrecpsq_f32(yf, recip);
5846 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5847 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5848 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5849 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5851 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5852 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5853 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5855 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5856 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5857 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5858 // and that it will never cause us to return an answer too large).
5859 // float4 result = as_float4(as_int4(xf*recip) + 2);
5860 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5861 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5862 N1 = DAG.getConstant(2, MVT::i32);
5863 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5864 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5865 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5866 // Convert back to integer and return.
5867 // return vmovn_u32(vcvt_s32_f32(result));
5868 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5869 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5873 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5874 EVT VT = Op.getNode()->getValueType(0);
5875 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5878 bool ExtraOp = false;
5879 switch (Op.getOpcode()) {
5880 default: llvm_unreachable("Invalid code");
5881 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5882 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5883 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5884 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5888 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5890 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5891 Op.getOperand(1), Op.getOperand(2));
5894 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5895 // Monotonic load/store is legal for all targets
5896 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5899 // Aquire/Release load/store is not legal for targets without a
5900 // dmb or equivalent available.
5905 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5906 SelectionDAG &DAG, unsigned NewOp) {
5908 assert (Node->getValueType(0) == MVT::i64 &&
5909 "Only know how to expand i64 atomics");
5911 SmallVector<SDValue, 6> Ops;
5912 Ops.push_back(Node->getOperand(0)); // Chain
5913 Ops.push_back(Node->getOperand(1)); // Ptr
5915 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5916 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5917 // High part of Val1
5918 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5919 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5920 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5921 // High part of Val1
5922 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5923 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5924 // High part of Val2
5925 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5926 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5928 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5930 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5931 cast<MemSDNode>(Node)->getMemOperand());
5932 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5933 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5934 Results.push_back(Result.getValue(2));
5937 static void ReplaceREADCYCLECOUNTER(SDNode *N,
5938 SmallVectorImpl<SDValue> &Results,
5940 const ARMSubtarget *Subtarget) {
5942 SDValue Cycles32, OutChain;
5944 if (Subtarget->hasPerfMon()) {
5945 // Under Power Management extensions, the cycle-count is:
5946 // mrc p15, #0, <Rt>, c9, c13, #0
5947 SDValue Ops[] = { N->getOperand(0), // Chain
5948 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
5949 DAG.getConstant(15, MVT::i32),
5950 DAG.getConstant(0, MVT::i32),
5951 DAG.getConstant(9, MVT::i32),
5952 DAG.getConstant(13, MVT::i32),
5953 DAG.getConstant(0, MVT::i32)
5956 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
5957 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
5958 array_lengthof(Ops));
5959 OutChain = Cycles32.getValue(1);
5961 // Intrinsic is defined to return 0 on unsupported platforms. Technically
5962 // there are older ARM CPUs that have implementation-specific ways of
5963 // obtaining this information (FIXME!).
5964 Cycles32 = DAG.getConstant(0, MVT::i32);
5965 OutChain = DAG.getEntryNode();
5969 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
5970 Cycles32, DAG.getConstant(0, MVT::i32));
5971 Results.push_back(Cycles64);
5972 Results.push_back(OutChain);
5975 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5976 switch (Op.getOpcode()) {
5977 default: llvm_unreachable("Don't know how to custom lower this!");
5978 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5979 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5980 case ISD::GlobalAddress:
5981 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5982 LowerGlobalAddressELF(Op, DAG);
5983 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5984 case ISD::SELECT: return LowerSELECT(Op, DAG);
5985 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5986 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5987 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5988 case ISD::VASTART: return LowerVASTART(Op, DAG);
5989 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5990 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5991 case ISD::SINT_TO_FP:
5992 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5993 case ISD::FP_TO_SINT:
5994 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5995 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5996 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5997 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5998 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5999 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6000 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6001 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6003 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6006 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6007 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6008 case ISD::SRL_PARTS:
6009 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6010 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6011 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6012 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6013 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6014 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6015 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6016 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6017 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6018 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6019 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6020 case ISD::MUL: return LowerMUL(Op, DAG);
6021 case ISD::SDIV: return LowerSDIV(Op, DAG);
6022 case ISD::UDIV: return LowerUDIV(Op, DAG);
6026 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6027 case ISD::ATOMIC_LOAD:
6028 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6030 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6034 /// ReplaceNodeResults - Replace the results of node with an illegal result
6035 /// type with new values built out of custom code.
6036 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6037 SmallVectorImpl<SDValue>&Results,
6038 SelectionDAG &DAG) const {
6040 switch (N->getOpcode()) {
6042 llvm_unreachable("Don't know how to custom expand this!");
6044 Res = ExpandBITCAST(N, DAG);
6046 case ISD::SIGN_EXTEND:
6047 case ISD::ZERO_EXTEND:
6048 Res = ExpandVectorExtension(N, DAG);
6052 Res = Expand64BitShift(N, DAG, Subtarget);
6054 case ISD::READCYCLECOUNTER:
6055 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6057 case ISD::ATOMIC_LOAD_ADD:
6058 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
6060 case ISD::ATOMIC_LOAD_AND:
6061 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
6063 case ISD::ATOMIC_LOAD_NAND:
6064 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
6066 case ISD::ATOMIC_LOAD_OR:
6067 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
6069 case ISD::ATOMIC_LOAD_SUB:
6070 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
6072 case ISD::ATOMIC_LOAD_XOR:
6073 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
6075 case ISD::ATOMIC_SWAP:
6076 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
6078 case ISD::ATOMIC_CMP_SWAP:
6079 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
6081 case ISD::ATOMIC_LOAD_MIN:
6082 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
6084 case ISD::ATOMIC_LOAD_UMIN:
6085 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
6087 case ISD::ATOMIC_LOAD_MAX:
6088 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
6090 case ISD::ATOMIC_LOAD_UMAX:
6091 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
6095 Results.push_back(Res);
6098 //===----------------------------------------------------------------------===//
6099 // ARM Scheduler Hooks
6100 //===----------------------------------------------------------------------===//
6103 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6104 MachineBasicBlock *BB,
6105 unsigned Size) const {
6106 unsigned dest = MI->getOperand(0).getReg();
6107 unsigned ptr = MI->getOperand(1).getReg();
6108 unsigned oldval = MI->getOperand(2).getReg();
6109 unsigned newval = MI->getOperand(3).getReg();
6110 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6111 DebugLoc dl = MI->getDebugLoc();
6112 bool isThumb2 = Subtarget->isThumb2();
6114 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6115 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6116 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6117 (const TargetRegisterClass*)&ARM::GPRRegClass);
6120 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6121 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6122 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6125 unsigned ldrOpc, strOpc;
6127 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6129 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6130 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6133 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6134 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6137 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6138 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6142 MachineFunction *MF = BB->getParent();
6143 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6144 MachineFunction::iterator It = BB;
6145 ++It; // insert the new blocks after the current block
6147 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6148 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6149 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6150 MF->insert(It, loop1MBB);
6151 MF->insert(It, loop2MBB);
6152 MF->insert(It, exitMBB);
6154 // Transfer the remainder of BB and its successor edges to exitMBB.
6155 exitMBB->splice(exitMBB->begin(), BB,
6156 llvm::next(MachineBasicBlock::iterator(MI)),
6158 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6162 // fallthrough --> loop1MBB
6163 BB->addSuccessor(loop1MBB);
6166 // ldrex dest, [ptr]
6170 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6171 if (ldrOpc == ARM::t2LDREX)
6173 AddDefaultPred(MIB);
6174 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6175 .addReg(dest).addReg(oldval));
6176 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6177 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6178 BB->addSuccessor(loop2MBB);
6179 BB->addSuccessor(exitMBB);
6182 // strex scratch, newval, [ptr]
6186 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6187 if (strOpc == ARM::t2STREX)
6189 AddDefaultPred(MIB);
6190 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6191 .addReg(scratch).addImm(0));
6192 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6193 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6194 BB->addSuccessor(loop1MBB);
6195 BB->addSuccessor(exitMBB);
6201 MI->eraseFromParent(); // The instruction is gone now.
6207 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6208 unsigned Size, unsigned BinOpcode) const {
6209 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6210 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6212 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6213 MachineFunction *MF = BB->getParent();
6214 MachineFunction::iterator It = BB;
6217 unsigned dest = MI->getOperand(0).getReg();
6218 unsigned ptr = MI->getOperand(1).getReg();
6219 unsigned incr = MI->getOperand(2).getReg();
6220 DebugLoc dl = MI->getDebugLoc();
6221 bool isThumb2 = Subtarget->isThumb2();
6223 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6225 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6226 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6229 unsigned ldrOpc, strOpc;
6231 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6233 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6234 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6237 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6238 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6241 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6242 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6246 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6247 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6248 MF->insert(It, loopMBB);
6249 MF->insert(It, exitMBB);
6251 // Transfer the remainder of BB and its successor edges to exitMBB.
6252 exitMBB->splice(exitMBB->begin(), BB,
6253 llvm::next(MachineBasicBlock::iterator(MI)),
6255 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6257 const TargetRegisterClass *TRC = isThumb2 ?
6258 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6259 (const TargetRegisterClass*)&ARM::GPRRegClass;
6260 unsigned scratch = MRI.createVirtualRegister(TRC);
6261 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6265 // fallthrough --> loopMBB
6266 BB->addSuccessor(loopMBB);
6270 // <binop> scratch2, dest, incr
6271 // strex scratch, scratch2, ptr
6274 // fallthrough --> exitMBB
6276 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6277 if (ldrOpc == ARM::t2LDREX)
6279 AddDefaultPred(MIB);
6281 // operand order needs to go the other way for NAND
6282 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6283 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6284 addReg(incr).addReg(dest)).addReg(0);
6286 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6287 addReg(dest).addReg(incr)).addReg(0);
6290 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6291 if (strOpc == ARM::t2STREX)
6293 AddDefaultPred(MIB);
6294 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6295 .addReg(scratch).addImm(0));
6296 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6297 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6299 BB->addSuccessor(loopMBB);
6300 BB->addSuccessor(exitMBB);
6306 MI->eraseFromParent(); // The instruction is gone now.
6312 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6313 MachineBasicBlock *BB,
6316 ARMCC::CondCodes Cond) const {
6317 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6319 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6320 MachineFunction *MF = BB->getParent();
6321 MachineFunction::iterator It = BB;
6324 unsigned dest = MI->getOperand(0).getReg();
6325 unsigned ptr = MI->getOperand(1).getReg();
6326 unsigned incr = MI->getOperand(2).getReg();
6327 unsigned oldval = dest;
6328 DebugLoc dl = MI->getDebugLoc();
6329 bool isThumb2 = Subtarget->isThumb2();
6331 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6333 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6334 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6337 unsigned ldrOpc, strOpc, extendOpc;
6339 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
6341 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
6342 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
6343 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6346 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
6347 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
6348 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6351 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
6352 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
6357 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6358 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6359 MF->insert(It, loopMBB);
6360 MF->insert(It, exitMBB);
6362 // Transfer the remainder of BB and its successor edges to exitMBB.
6363 exitMBB->splice(exitMBB->begin(), BB,
6364 llvm::next(MachineBasicBlock::iterator(MI)),
6366 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6368 const TargetRegisterClass *TRC = isThumb2 ?
6369 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6370 (const TargetRegisterClass*)&ARM::GPRRegClass;
6371 unsigned scratch = MRI.createVirtualRegister(TRC);
6372 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6376 // fallthrough --> loopMBB
6377 BB->addSuccessor(loopMBB);
6381 // (sign extend dest, if required)
6383 // cmov.cond scratch2, incr, dest
6384 // strex scratch, scratch2, ptr
6387 // fallthrough --> exitMBB
6389 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6390 if (ldrOpc == ARM::t2LDREX)
6392 AddDefaultPred(MIB);
6394 // Sign extend the value, if necessary.
6395 if (signExtend && extendOpc) {
6396 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
6397 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6402 // Build compare and cmov instructions.
6403 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6404 .addReg(oldval).addReg(incr));
6405 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6406 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6408 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6409 if (strOpc == ARM::t2STREX)
6411 AddDefaultPred(MIB);
6412 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6413 .addReg(scratch).addImm(0));
6414 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6415 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6417 BB->addSuccessor(loopMBB);
6418 BB->addSuccessor(exitMBB);
6424 MI->eraseFromParent(); // The instruction is gone now.
6430 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6431 unsigned Op1, unsigned Op2,
6432 bool NeedsCarry, bool IsCmpxchg,
6433 bool IsMinMax, ARMCC::CondCodes CC) const {
6434 // This also handles ATOMIC_SWAP, indicated by Op1==0.
6435 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6437 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6438 MachineFunction *MF = BB->getParent();
6439 MachineFunction::iterator It = BB;
6442 unsigned destlo = MI->getOperand(0).getReg();
6443 unsigned desthi = MI->getOperand(1).getReg();
6444 unsigned ptr = MI->getOperand(2).getReg();
6445 unsigned vallo = MI->getOperand(3).getReg();
6446 unsigned valhi = MI->getOperand(4).getReg();
6447 DebugLoc dl = MI->getDebugLoc();
6448 bool isThumb2 = Subtarget->isThumb2();
6450 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6452 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6453 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6454 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6455 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6456 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6459 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6460 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6461 if (IsCmpxchg || IsMinMax)
6462 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6464 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6465 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6467 MF->insert(It, loopMBB);
6468 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6469 if (IsCmpxchg) MF->insert(It, cont2BB);
6470 MF->insert(It, exitMBB);
6472 // Transfer the remainder of BB and its successor edges to exitMBB.
6473 exitMBB->splice(exitMBB->begin(), BB,
6474 llvm::next(MachineBasicBlock::iterator(MI)),
6476 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6478 const TargetRegisterClass *TRC = isThumb2 ?
6479 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6480 (const TargetRegisterClass*)&ARM::GPRRegClass;
6481 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6485 // fallthrough --> loopMBB
6486 BB->addSuccessor(loopMBB);
6489 // ldrexd r2, r3, ptr
6490 // <binopa> r0, r2, incr
6491 // <binopb> r1, r3, incr
6492 // strexd storesuccess, r0, r1, ptr
6493 // cmp storesuccess, #0
6495 // fallthrough --> exitMBB
6500 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD))
6501 .addReg(destlo, RegState::Define)
6502 .addReg(desthi, RegState::Define)
6505 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6506 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD))
6507 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6508 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6509 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6510 .addReg(GPRPair0, 0, ARM::gsub_0);
6511 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6512 .addReg(GPRPair0, 0, ARM::gsub_1);
6515 unsigned StoreLo, StoreHi;
6518 for (unsigned i = 0; i < 2; i++) {
6519 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6521 .addReg(i == 0 ? destlo : desthi)
6522 .addReg(i == 0 ? vallo : valhi));
6523 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6524 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6525 BB->addSuccessor(exitMBB);
6526 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6527 BB = (i == 0 ? contBB : cont2BB);
6530 // Copy to physregs for strexd
6531 StoreLo = MI->getOperand(5).getReg();
6532 StoreHi = MI->getOperand(6).getReg();
6534 // Perform binary operation
6535 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6536 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6537 .addReg(destlo).addReg(vallo))
6538 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6539 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6540 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6541 .addReg(desthi).addReg(valhi))
6542 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6547 // Copy to physregs for strexd
6552 // Compare and branch to exit block.
6553 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6554 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6555 BB->addSuccessor(exitMBB);
6556 BB->addSuccessor(contBB);
6564 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6565 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6566 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess)
6567 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6569 // Marshal a pair...
6570 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6571 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6572 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6573 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6574 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6577 .addImm(ARM::gsub_0);
6578 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6581 .addImm(ARM::gsub_1);
6584 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess)
6585 .addReg(StorePair).addReg(ptr));
6588 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6589 .addReg(storesuccess).addImm(0));
6590 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6591 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6593 BB->addSuccessor(loopMBB);
6594 BB->addSuccessor(exitMBB);
6600 MI->eraseFromParent(); // The instruction is gone now.
6605 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6606 /// registers the function context.
6607 void ARMTargetLowering::
6608 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6609 MachineBasicBlock *DispatchBB, int FI) const {
6610 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6611 DebugLoc dl = MI->getDebugLoc();
6612 MachineFunction *MF = MBB->getParent();
6613 MachineRegisterInfo *MRI = &MF->getRegInfo();
6614 MachineConstantPool *MCP = MF->getConstantPool();
6615 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6616 const Function *F = MF->getFunction();
6618 bool isThumb = Subtarget->isThumb();
6619 bool isThumb2 = Subtarget->isThumb2();
6621 unsigned PCLabelId = AFI->createPICLabelUId();
6622 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6623 ARMConstantPoolValue *CPV =
6624 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6625 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6627 const TargetRegisterClass *TRC = isThumb ?
6628 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6629 (const TargetRegisterClass*)&ARM::GPRRegClass;
6631 // Grab constant pool and fixed stack memory operands.
6632 MachineMemOperand *CPMMO =
6633 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6634 MachineMemOperand::MOLoad, 4, 4);
6636 MachineMemOperand *FIMMOSt =
6637 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6638 MachineMemOperand::MOStore, 4, 4);
6640 // Load the address of the dispatch MBB into the jump buffer.
6642 // Incoming value: jbuf
6643 // ldr.n r5, LCPI1_1
6646 // str r5, [$jbuf, #+4] ; &jbuf[1]
6647 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6648 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6649 .addConstantPoolIndex(CPI)
6650 .addMemOperand(CPMMO));
6651 // Set the low bit because of thumb mode.
6652 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6654 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6655 .addReg(NewVReg1, RegState::Kill)
6657 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6658 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6659 .addReg(NewVReg2, RegState::Kill)
6661 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6662 .addReg(NewVReg3, RegState::Kill)
6664 .addImm(36) // &jbuf[1] :: pc
6665 .addMemOperand(FIMMOSt));
6666 } else if (isThumb) {
6667 // Incoming value: jbuf
6668 // ldr.n r1, LCPI1_4
6672 // add r2, $jbuf, #+4 ; &jbuf[1]
6674 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6675 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6676 .addConstantPoolIndex(CPI)
6677 .addMemOperand(CPMMO));
6678 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6679 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6680 .addReg(NewVReg1, RegState::Kill)
6682 // Set the low bit because of thumb mode.
6683 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6684 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6685 .addReg(ARM::CPSR, RegState::Define)
6687 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6688 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6689 .addReg(ARM::CPSR, RegState::Define)
6690 .addReg(NewVReg2, RegState::Kill)
6691 .addReg(NewVReg3, RegState::Kill));
6692 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6693 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6695 .addImm(36)); // &jbuf[1] :: pc
6696 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6697 .addReg(NewVReg4, RegState::Kill)
6698 .addReg(NewVReg5, RegState::Kill)
6700 .addMemOperand(FIMMOSt));
6702 // Incoming value: jbuf
6705 // str r1, [$jbuf, #+4] ; &jbuf[1]
6706 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6707 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6708 .addConstantPoolIndex(CPI)
6710 .addMemOperand(CPMMO));
6711 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6712 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6713 .addReg(NewVReg1, RegState::Kill)
6714 .addImm(PCLabelId));
6715 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6716 .addReg(NewVReg2, RegState::Kill)
6718 .addImm(36) // &jbuf[1] :: pc
6719 .addMemOperand(FIMMOSt));
6723 MachineBasicBlock *ARMTargetLowering::
6724 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6725 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6726 DebugLoc dl = MI->getDebugLoc();
6727 MachineFunction *MF = MBB->getParent();
6728 MachineRegisterInfo *MRI = &MF->getRegInfo();
6729 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6730 MachineFrameInfo *MFI = MF->getFrameInfo();
6731 int FI = MFI->getFunctionContextIndex();
6733 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6734 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6735 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6737 // Get a mapping of the call site numbers to all of the landing pads they're
6739 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6740 unsigned MaxCSNum = 0;
6741 MachineModuleInfo &MMI = MF->getMMI();
6742 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6744 if (!BB->isLandingPad()) continue;
6746 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6748 for (MachineBasicBlock::iterator
6749 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6750 if (!II->isEHLabel()) continue;
6752 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6753 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6755 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6756 for (SmallVectorImpl<unsigned>::iterator
6757 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6758 CSI != CSE; ++CSI) {
6759 CallSiteNumToLPad[*CSI].push_back(BB);
6760 MaxCSNum = std::max(MaxCSNum, *CSI);
6766 // Get an ordered list of the machine basic blocks for the jump table.
6767 std::vector<MachineBasicBlock*> LPadList;
6768 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6769 LPadList.reserve(CallSiteNumToLPad.size());
6770 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6771 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6772 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6773 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6774 LPadList.push_back(*II);
6775 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6779 assert(!LPadList.empty() &&
6780 "No landing pad destinations for the dispatch jump table!");
6782 // Create the jump table and associated information.
6783 MachineJumpTableInfo *JTI =
6784 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6785 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6786 unsigned UId = AFI->createJumpTableUId();
6787 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6789 // Create the MBBs for the dispatch code.
6791 // Shove the dispatch's address into the return slot in the function context.
6792 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6793 DispatchBB->setIsLandingPad();
6795 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6796 unsigned trap_opcode;
6797 if (Subtarget->isThumb())
6798 trap_opcode = ARM::tTRAP;
6800 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6802 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6803 DispatchBB->addSuccessor(TrapBB);
6805 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6806 DispatchBB->addSuccessor(DispContBB);
6809 MF->insert(MF->end(), DispatchBB);
6810 MF->insert(MF->end(), DispContBB);
6811 MF->insert(MF->end(), TrapBB);
6813 // Insert code into the entry block that creates and registers the function
6815 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6817 MachineMemOperand *FIMMOLd =
6818 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6819 MachineMemOperand::MOLoad |
6820 MachineMemOperand::MOVolatile, 4, 4);
6822 MachineInstrBuilder MIB;
6823 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6825 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6826 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6828 // Add a register mask with no preserved registers. This results in all
6829 // registers being marked as clobbered.
6830 MIB.addRegMask(RI.getNoPreservedMask());
6832 unsigned NumLPads = LPadList.size();
6833 if (Subtarget->isThumb2()) {
6834 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6835 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6838 .addMemOperand(FIMMOLd));
6840 if (NumLPads < 256) {
6841 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6843 .addImm(LPadList.size()));
6845 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6846 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6847 .addImm(NumLPads & 0xFFFF));
6849 unsigned VReg2 = VReg1;
6850 if ((NumLPads & 0xFFFF0000) != 0) {
6851 VReg2 = MRI->createVirtualRegister(TRC);
6852 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6854 .addImm(NumLPads >> 16));
6857 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6862 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6867 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6868 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6869 .addJumpTableIndex(MJTI)
6872 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6875 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6876 .addReg(NewVReg3, RegState::Kill)
6878 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6880 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6881 .addReg(NewVReg4, RegState::Kill)
6883 .addJumpTableIndex(MJTI)
6885 } else if (Subtarget->isThumb()) {
6886 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6887 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6890 .addMemOperand(FIMMOLd));
6892 if (NumLPads < 256) {
6893 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6897 MachineConstantPool *ConstantPool = MF->getConstantPool();
6898 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6899 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6901 // MachineConstantPool wants an explicit alignment.
6902 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6904 Align = getDataLayout()->getTypeAllocSize(C->getType());
6905 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6907 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6908 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6909 .addReg(VReg1, RegState::Define)
6910 .addConstantPoolIndex(Idx));
6911 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6916 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6921 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6922 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6923 .addReg(ARM::CPSR, RegState::Define)
6927 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6928 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6929 .addJumpTableIndex(MJTI)
6932 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6933 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6934 .addReg(ARM::CPSR, RegState::Define)
6935 .addReg(NewVReg2, RegState::Kill)
6938 MachineMemOperand *JTMMOLd =
6939 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6940 MachineMemOperand::MOLoad, 4, 4);
6942 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6943 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6944 .addReg(NewVReg4, RegState::Kill)
6946 .addMemOperand(JTMMOLd));
6948 unsigned NewVReg6 = NewVReg5;
6949 if (RelocM == Reloc::PIC_) {
6950 NewVReg6 = MRI->createVirtualRegister(TRC);
6951 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6952 .addReg(ARM::CPSR, RegState::Define)
6953 .addReg(NewVReg5, RegState::Kill)
6957 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6958 .addReg(NewVReg6, RegState::Kill)
6959 .addJumpTableIndex(MJTI)
6962 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6963 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6966 .addMemOperand(FIMMOLd));
6968 if (NumLPads < 256) {
6969 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6972 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6973 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6974 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6975 .addImm(NumLPads & 0xFFFF));
6977 unsigned VReg2 = VReg1;
6978 if ((NumLPads & 0xFFFF0000) != 0) {
6979 VReg2 = MRI->createVirtualRegister(TRC);
6980 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6982 .addImm(NumLPads >> 16));
6985 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6989 MachineConstantPool *ConstantPool = MF->getConstantPool();
6990 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6991 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6993 // MachineConstantPool wants an explicit alignment.
6994 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6996 Align = getDataLayout()->getTypeAllocSize(C->getType());
6997 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6999 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7000 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7001 .addReg(VReg1, RegState::Define)
7002 .addConstantPoolIndex(Idx)
7004 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7006 .addReg(VReg1, RegState::Kill));
7009 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7014 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7016 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7018 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7019 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7020 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7021 .addJumpTableIndex(MJTI)
7024 MachineMemOperand *JTMMOLd =
7025 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7026 MachineMemOperand::MOLoad, 4, 4);
7027 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7029 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7030 .addReg(NewVReg3, RegState::Kill)
7033 .addMemOperand(JTMMOLd));
7035 if (RelocM == Reloc::PIC_) {
7036 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7037 .addReg(NewVReg5, RegState::Kill)
7039 .addJumpTableIndex(MJTI)
7042 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7043 .addReg(NewVReg5, RegState::Kill)
7044 .addJumpTableIndex(MJTI)
7049 // Add the jump table entries as successors to the MBB.
7050 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7051 for (std::vector<MachineBasicBlock*>::iterator
7052 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7053 MachineBasicBlock *CurMBB = *I;
7054 if (SeenMBBs.insert(CurMBB))
7055 DispContBB->addSuccessor(CurMBB);
7058 // N.B. the order the invoke BBs are processed in doesn't matter here.
7059 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7060 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7061 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7062 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7063 MachineBasicBlock *BB = *I;
7065 // Remove the landing pad successor from the invoke block and replace it
7066 // with the new dispatch block.
7067 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7069 while (!Successors.empty()) {
7070 MachineBasicBlock *SMBB = Successors.pop_back_val();
7071 if (SMBB->isLandingPad()) {
7072 BB->removeSuccessor(SMBB);
7073 MBBLPads.push_back(SMBB);
7077 BB->addSuccessor(DispatchBB);
7079 // Find the invoke call and mark all of the callee-saved registers as
7080 // 'implicit defined' so that they're spilled. This prevents code from
7081 // moving instructions to before the EH block, where they will never be
7083 for (MachineBasicBlock::reverse_iterator
7084 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7085 if (!II->isCall()) continue;
7087 DenseMap<unsigned, bool> DefRegs;
7088 for (MachineInstr::mop_iterator
7089 OI = II->operands_begin(), OE = II->operands_end();
7091 if (!OI->isReg()) continue;
7092 DefRegs[OI->getReg()] = true;
7095 MachineInstrBuilder MIB(*MF, &*II);
7097 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7098 unsigned Reg = SavedRegs[i];
7099 if (Subtarget->isThumb2() &&
7100 !ARM::tGPRRegClass.contains(Reg) &&
7101 !ARM::hGPRRegClass.contains(Reg))
7103 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7105 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7108 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7115 // Mark all former landing pads as non-landing pads. The dispatch is the only
7117 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7118 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7119 (*I)->setIsLandingPad(false);
7121 // The instruction is gone now.
7122 MI->eraseFromParent();
7128 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7129 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7130 E = MBB->succ_end(); I != E; ++I)
7133 llvm_unreachable("Expecting a BB with two successors!");
7136 MachineBasicBlock *ARMTargetLowering::
7137 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
7138 // This pseudo instruction has 3 operands: dst, src, size
7139 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7140 // Otherwise, we will generate unrolled scalar copies.
7141 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7142 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7143 MachineFunction::iterator It = BB;
7146 unsigned dest = MI->getOperand(0).getReg();
7147 unsigned src = MI->getOperand(1).getReg();
7148 unsigned SizeVal = MI->getOperand(2).getImm();
7149 unsigned Align = MI->getOperand(3).getImm();
7150 DebugLoc dl = MI->getDebugLoc();
7152 bool isThumb2 = Subtarget->isThumb2();
7153 MachineFunction *MF = BB->getParent();
7154 MachineRegisterInfo &MRI = MF->getRegInfo();
7155 unsigned ldrOpc, strOpc, UnitSize = 0;
7157 const TargetRegisterClass *TRC = isThumb2 ?
7158 (const TargetRegisterClass*)&ARM::tGPRRegClass :
7159 (const TargetRegisterClass*)&ARM::GPRRegClass;
7160 const TargetRegisterClass *TRC_Vec = 0;
7163 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7164 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7166 } else if (Align & 2) {
7167 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
7168 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
7171 // Check whether we can use NEON instructions.
7172 if (!MF->getFunction()->getAttributes().
7173 hasAttribute(AttributeSet::FunctionIndex,
7174 Attribute::NoImplicitFloat) &&
7175 Subtarget->hasNEON()) {
7176 if ((Align % 16 == 0) && SizeVal >= 16) {
7177 ldrOpc = ARM::VLD1q32wb_fixed;
7178 strOpc = ARM::VST1q32wb_fixed;
7180 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
7182 else if ((Align % 8 == 0) && SizeVal >= 8) {
7183 ldrOpc = ARM::VLD1d32wb_fixed;
7184 strOpc = ARM::VST1d32wb_fixed;
7186 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
7189 // Can't use NEON instructions.
7190 if (UnitSize == 0) {
7191 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
7192 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
7197 unsigned BytesLeft = SizeVal % UnitSize;
7198 unsigned LoopSize = SizeVal - BytesLeft;
7200 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7201 // Use LDR and STR to copy.
7202 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7203 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7204 unsigned srcIn = src;
7205 unsigned destIn = dest;
7206 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7207 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7208 unsigned srcOut = MRI.createVirtualRegister(TRC);
7209 unsigned destOut = MRI.createVirtualRegister(TRC);
7210 if (UnitSize >= 8) {
7211 AddDefaultPred(BuildMI(*BB, MI, dl,
7212 TII->get(ldrOpc), scratch)
7213 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
7215 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7216 .addReg(destIn).addImm(0).addReg(scratch));
7217 } else if (isThumb2) {
7218 AddDefaultPred(BuildMI(*BB, MI, dl,
7219 TII->get(ldrOpc), scratch)
7220 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
7222 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7223 .addReg(scratch).addReg(destIn)
7226 AddDefaultPred(BuildMI(*BB, MI, dl,
7227 TII->get(ldrOpc), scratch)
7228 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
7231 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7232 .addReg(scratch).addReg(destIn)
7233 .addReg(0).addImm(UnitSize));
7239 // Handle the leftover bytes with LDRB and STRB.
7240 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7241 // [destOut] = STRB_POST(scratch, destIn, 1)
7242 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7243 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7244 for (unsigned i = 0; i < BytesLeft; i++) {
7245 unsigned scratch = MRI.createVirtualRegister(TRC);
7246 unsigned srcOut = MRI.createVirtualRegister(TRC);
7247 unsigned destOut = MRI.createVirtualRegister(TRC);
7249 AddDefaultPred(BuildMI(*BB, MI, dl,
7250 TII->get(ldrOpc),scratch)
7251 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7253 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7254 .addReg(scratch).addReg(destIn)
7255 .addReg(0).addImm(1));
7257 AddDefaultPred(BuildMI(*BB, MI, dl,
7258 TII->get(ldrOpc),scratch)
7259 .addReg(srcOut, RegState::Define).addReg(srcIn)
7260 .addReg(0).addImm(1));
7262 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7263 .addReg(scratch).addReg(destIn)
7264 .addReg(0).addImm(1));
7269 MI->eraseFromParent(); // The instruction is gone now.
7273 // Expand the pseudo op to a loop.
7276 // movw varEnd, # --> with thumb2
7278 // ldrcp varEnd, idx --> without thumb2
7279 // fallthrough --> loopMBB
7281 // PHI varPhi, varEnd, varLoop
7282 // PHI srcPhi, src, srcLoop
7283 // PHI destPhi, dst, destLoop
7284 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7285 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7286 // subs varLoop, varPhi, #UnitSize
7288 // fallthrough --> exitMBB
7290 // epilogue to handle left-over bytes
7291 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7292 // [destOut] = STRB_POST(scratch, destLoop, 1)
7293 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7294 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7295 MF->insert(It, loopMBB);
7296 MF->insert(It, exitMBB);
7298 // Transfer the remainder of BB and its successor edges to exitMBB.
7299 exitMBB->splice(exitMBB->begin(), BB,
7300 llvm::next(MachineBasicBlock::iterator(MI)),
7302 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7304 // Load an immediate to varEnd.
7305 unsigned varEnd = MRI.createVirtualRegister(TRC);
7307 unsigned VReg1 = varEnd;
7308 if ((LoopSize & 0xFFFF0000) != 0)
7309 VReg1 = MRI.createVirtualRegister(TRC);
7310 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
7311 .addImm(LoopSize & 0xFFFF));
7313 if ((LoopSize & 0xFFFF0000) != 0)
7314 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7316 .addImm(LoopSize >> 16));
7318 MachineConstantPool *ConstantPool = MF->getConstantPool();
7319 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7320 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7322 // MachineConstantPool wants an explicit alignment.
7323 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7325 Align = getDataLayout()->getTypeAllocSize(C->getType());
7326 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7328 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
7329 .addReg(varEnd, RegState::Define)
7330 .addConstantPoolIndex(Idx)
7333 BB->addSuccessor(loopMBB);
7335 // Generate the loop body:
7336 // varPhi = PHI(varLoop, varEnd)
7337 // srcPhi = PHI(srcLoop, src)
7338 // destPhi = PHI(destLoop, dst)
7339 MachineBasicBlock *entryBB = BB;
7341 unsigned varLoop = MRI.createVirtualRegister(TRC);
7342 unsigned varPhi = MRI.createVirtualRegister(TRC);
7343 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7344 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7345 unsigned destLoop = MRI.createVirtualRegister(TRC);
7346 unsigned destPhi = MRI.createVirtualRegister(TRC);
7348 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7349 .addReg(varLoop).addMBB(loopMBB)
7350 .addReg(varEnd).addMBB(entryBB);
7351 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7352 .addReg(srcLoop).addMBB(loopMBB)
7353 .addReg(src).addMBB(entryBB);
7354 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7355 .addReg(destLoop).addMBB(loopMBB)
7356 .addReg(dest).addMBB(entryBB);
7358 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7359 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7360 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7361 if (UnitSize >= 8) {
7362 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7363 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
7365 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7366 .addReg(destPhi).addImm(0).addReg(scratch));
7367 } else if (isThumb2) {
7368 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7369 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
7371 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7372 .addReg(scratch).addReg(destPhi)
7375 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7376 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
7379 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7380 .addReg(scratch).addReg(destPhi)
7381 .addReg(0).addImm(UnitSize));
7384 // Decrement loop variable by UnitSize.
7385 MachineInstrBuilder MIB = BuildMI(BB, dl,
7386 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7387 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7388 MIB->getOperand(5).setReg(ARM::CPSR);
7389 MIB->getOperand(5).setIsDef(true);
7391 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7392 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7394 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7395 BB->addSuccessor(loopMBB);
7396 BB->addSuccessor(exitMBB);
7398 // Add epilogue to handle BytesLeft.
7400 MachineInstr *StartOfExit = exitMBB->begin();
7401 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7402 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7404 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7405 // [destOut] = STRB_POST(scratch, destLoop, 1)
7406 unsigned srcIn = srcLoop;
7407 unsigned destIn = destLoop;
7408 for (unsigned i = 0; i < BytesLeft; i++) {
7409 unsigned scratch = MRI.createVirtualRegister(TRC);
7410 unsigned srcOut = MRI.createVirtualRegister(TRC);
7411 unsigned destOut = MRI.createVirtualRegister(TRC);
7413 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7414 TII->get(ldrOpc),scratch)
7415 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7417 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7418 .addReg(scratch).addReg(destIn)
7421 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7422 TII->get(ldrOpc),scratch)
7423 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
7425 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7426 .addReg(scratch).addReg(destIn)
7427 .addReg(0).addImm(1));
7433 MI->eraseFromParent(); // The instruction is gone now.
7438 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7439 MachineBasicBlock *BB) const {
7440 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7441 DebugLoc dl = MI->getDebugLoc();
7442 bool isThumb2 = Subtarget->isThumb2();
7443 switch (MI->getOpcode()) {
7446 llvm_unreachable("Unexpected instr type to insert");
7448 // The Thumb2 pre-indexed stores have the same MI operands, they just
7449 // define them differently in the .td files from the isel patterns, so
7450 // they need pseudos.
7451 case ARM::t2STR_preidx:
7452 MI->setDesc(TII->get(ARM::t2STR_PRE));
7454 case ARM::t2STRB_preidx:
7455 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7457 case ARM::t2STRH_preidx:
7458 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7461 case ARM::STRi_preidx:
7462 case ARM::STRBi_preidx: {
7463 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7464 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7465 // Decode the offset.
7466 unsigned Offset = MI->getOperand(4).getImm();
7467 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7468 Offset = ARM_AM::getAM2Offset(Offset);
7472 MachineMemOperand *MMO = *MI->memoperands_begin();
7473 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7474 .addOperand(MI->getOperand(0)) // Rn_wb
7475 .addOperand(MI->getOperand(1)) // Rt
7476 .addOperand(MI->getOperand(2)) // Rn
7477 .addImm(Offset) // offset (skip GPR==zero_reg)
7478 .addOperand(MI->getOperand(5)) // pred
7479 .addOperand(MI->getOperand(6))
7480 .addMemOperand(MMO);
7481 MI->eraseFromParent();
7484 case ARM::STRr_preidx:
7485 case ARM::STRBr_preidx:
7486 case ARM::STRH_preidx: {
7488 switch (MI->getOpcode()) {
7489 default: llvm_unreachable("unexpected opcode!");
7490 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7491 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7492 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7494 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7495 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7496 MIB.addOperand(MI->getOperand(i));
7497 MI->eraseFromParent();
7500 case ARM::ATOMIC_LOAD_ADD_I8:
7501 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7502 case ARM::ATOMIC_LOAD_ADD_I16:
7503 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7504 case ARM::ATOMIC_LOAD_ADD_I32:
7505 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7507 case ARM::ATOMIC_LOAD_AND_I8:
7508 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7509 case ARM::ATOMIC_LOAD_AND_I16:
7510 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7511 case ARM::ATOMIC_LOAD_AND_I32:
7512 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7514 case ARM::ATOMIC_LOAD_OR_I8:
7515 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7516 case ARM::ATOMIC_LOAD_OR_I16:
7517 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7518 case ARM::ATOMIC_LOAD_OR_I32:
7519 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7521 case ARM::ATOMIC_LOAD_XOR_I8:
7522 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7523 case ARM::ATOMIC_LOAD_XOR_I16:
7524 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7525 case ARM::ATOMIC_LOAD_XOR_I32:
7526 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7528 case ARM::ATOMIC_LOAD_NAND_I8:
7529 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7530 case ARM::ATOMIC_LOAD_NAND_I16:
7531 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7532 case ARM::ATOMIC_LOAD_NAND_I32:
7533 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7535 case ARM::ATOMIC_LOAD_SUB_I8:
7536 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7537 case ARM::ATOMIC_LOAD_SUB_I16:
7538 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7539 case ARM::ATOMIC_LOAD_SUB_I32:
7540 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7542 case ARM::ATOMIC_LOAD_MIN_I8:
7543 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7544 case ARM::ATOMIC_LOAD_MIN_I16:
7545 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7546 case ARM::ATOMIC_LOAD_MIN_I32:
7547 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7549 case ARM::ATOMIC_LOAD_MAX_I8:
7550 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7551 case ARM::ATOMIC_LOAD_MAX_I16:
7552 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7553 case ARM::ATOMIC_LOAD_MAX_I32:
7554 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7556 case ARM::ATOMIC_LOAD_UMIN_I8:
7557 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7558 case ARM::ATOMIC_LOAD_UMIN_I16:
7559 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7560 case ARM::ATOMIC_LOAD_UMIN_I32:
7561 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7563 case ARM::ATOMIC_LOAD_UMAX_I8:
7564 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7565 case ARM::ATOMIC_LOAD_UMAX_I16:
7566 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7567 case ARM::ATOMIC_LOAD_UMAX_I32:
7568 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7570 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7571 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7572 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7574 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7575 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7576 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7579 case ARM::ATOMADD6432:
7580 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7581 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7582 /*NeedsCarry*/ true);
7583 case ARM::ATOMSUB6432:
7584 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7585 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7586 /*NeedsCarry*/ true);
7587 case ARM::ATOMOR6432:
7588 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7589 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7590 case ARM::ATOMXOR6432:
7591 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7592 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7593 case ARM::ATOMAND6432:
7594 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7595 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7596 case ARM::ATOMSWAP6432:
7597 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7598 case ARM::ATOMCMPXCHG6432:
7599 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7600 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7601 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7602 case ARM::ATOMMIN6432:
7603 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7604 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7605 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7606 /*IsMinMax*/ true, ARMCC::LT);
7607 case ARM::ATOMMAX6432:
7608 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7609 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7610 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7611 /*IsMinMax*/ true, ARMCC::GE);
7612 case ARM::ATOMUMIN6432:
7613 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7614 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7615 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7616 /*IsMinMax*/ true, ARMCC::LO);
7617 case ARM::ATOMUMAX6432:
7618 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7619 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7620 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7621 /*IsMinMax*/ true, ARMCC::HS);
7623 case ARM::tMOVCCr_pseudo: {
7624 // To "insert" a SELECT_CC instruction, we actually have to insert the
7625 // diamond control-flow pattern. The incoming instruction knows the
7626 // destination vreg to set, the condition code register to branch on, the
7627 // true/false values to select between, and a branch opcode to use.
7628 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7629 MachineFunction::iterator It = BB;
7635 // cmpTY ccX, r1, r2
7637 // fallthrough --> copy0MBB
7638 MachineBasicBlock *thisMBB = BB;
7639 MachineFunction *F = BB->getParent();
7640 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7641 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7642 F->insert(It, copy0MBB);
7643 F->insert(It, sinkMBB);
7645 // Transfer the remainder of BB and its successor edges to sinkMBB.
7646 sinkMBB->splice(sinkMBB->begin(), BB,
7647 llvm::next(MachineBasicBlock::iterator(MI)),
7649 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7651 BB->addSuccessor(copy0MBB);
7652 BB->addSuccessor(sinkMBB);
7654 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7655 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7658 // %FalseValue = ...
7659 // # fallthrough to sinkMBB
7662 // Update machine-CFG edges
7663 BB->addSuccessor(sinkMBB);
7666 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7669 BuildMI(*BB, BB->begin(), dl,
7670 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7671 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7672 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7674 MI->eraseFromParent(); // The pseudo instruction is gone now.
7679 case ARM::BCCZi64: {
7680 // If there is an unconditional branch to the other successor, remove it.
7681 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7683 // Compare both parts that make up the double comparison separately for
7685 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7687 unsigned LHS1 = MI->getOperand(1).getReg();
7688 unsigned LHS2 = MI->getOperand(2).getReg();
7690 AddDefaultPred(BuildMI(BB, dl,
7691 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7692 .addReg(LHS1).addImm(0));
7693 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7694 .addReg(LHS2).addImm(0)
7695 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7697 unsigned RHS1 = MI->getOperand(3).getReg();
7698 unsigned RHS2 = MI->getOperand(4).getReg();
7699 AddDefaultPred(BuildMI(BB, dl,
7700 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7701 .addReg(LHS1).addReg(RHS1));
7702 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7703 .addReg(LHS2).addReg(RHS2)
7704 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7707 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7708 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7709 if (MI->getOperand(0).getImm() == ARMCC::NE)
7710 std::swap(destMBB, exitMBB);
7712 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7713 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7715 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7717 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7719 MI->eraseFromParent(); // The pseudo instruction is gone now.
7723 case ARM::Int_eh_sjlj_setjmp:
7724 case ARM::Int_eh_sjlj_setjmp_nofp:
7725 case ARM::tInt_eh_sjlj_setjmp:
7726 case ARM::t2Int_eh_sjlj_setjmp:
7727 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7728 EmitSjLjDispatchBlock(MI, BB);
7733 // To insert an ABS instruction, we have to insert the
7734 // diamond control-flow pattern. The incoming instruction knows the
7735 // source vreg to test against 0, the destination vreg to set,
7736 // the condition code register to branch on, the
7737 // true/false values to select between, and a branch opcode to use.
7742 // BCC (branch to SinkBB if V0 >= 0)
7743 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7744 // SinkBB: V1 = PHI(V2, V3)
7745 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7746 MachineFunction::iterator BBI = BB;
7748 MachineFunction *Fn = BB->getParent();
7749 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7750 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7751 Fn->insert(BBI, RSBBB);
7752 Fn->insert(BBI, SinkBB);
7754 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7755 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7756 bool isThumb2 = Subtarget->isThumb2();
7757 MachineRegisterInfo &MRI = Fn->getRegInfo();
7758 // In Thumb mode S must not be specified if source register is the SP or
7759 // PC and if destination register is the SP, so restrict register class
7760 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7761 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7762 (const TargetRegisterClass*)&ARM::GPRRegClass);
7764 // Transfer the remainder of BB and its successor edges to sinkMBB.
7765 SinkBB->splice(SinkBB->begin(), BB,
7766 llvm::next(MachineBasicBlock::iterator(MI)),
7768 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7770 BB->addSuccessor(RSBBB);
7771 BB->addSuccessor(SinkBB);
7773 // fall through to SinkMBB
7774 RSBBB->addSuccessor(SinkBB);
7776 // insert a cmp at the end of BB
7777 AddDefaultPred(BuildMI(BB, dl,
7778 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7779 .addReg(ABSSrcReg).addImm(0));
7781 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7783 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7784 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7786 // insert rsbri in RSBBB
7787 // Note: BCC and rsbri will be converted into predicated rsbmi
7788 // by if-conversion pass
7789 BuildMI(*RSBBB, RSBBB->begin(), dl,
7790 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7791 .addReg(ABSSrcReg, RegState::Kill)
7792 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7794 // insert PHI in SinkBB,
7795 // reuse ABSDstReg to not change uses of ABS instruction
7796 BuildMI(*SinkBB, SinkBB->begin(), dl,
7797 TII->get(ARM::PHI), ABSDstReg)
7798 .addReg(NewRsbDstReg).addMBB(RSBBB)
7799 .addReg(ABSSrcReg).addMBB(BB);
7801 // remove ABS instruction
7802 MI->eraseFromParent();
7804 // return last added BB
7807 case ARM::COPY_STRUCT_BYVAL_I32:
7809 return EmitStructByval(MI, BB);
7813 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7814 SDNode *Node) const {
7815 if (!MI->hasPostISelHook()) {
7816 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7817 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7821 const MCInstrDesc *MCID = &MI->getDesc();
7822 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7823 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7824 // operand is still set to noreg. If needed, set the optional operand's
7825 // register to CPSR, and remove the redundant implicit def.
7827 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7829 // Rename pseudo opcodes.
7830 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7832 const ARMBaseInstrInfo *TII =
7833 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7834 MCID = &TII->get(NewOpc);
7836 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7837 "converted opcode should be the same except for cc_out");
7841 // Add the optional cc_out operand
7842 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7844 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7846 // Any ARM instruction that sets the 's' bit should specify an optional
7847 // "cc_out" operand in the last operand position.
7848 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7849 assert(!NewOpc && "Optional cc_out operand required");
7852 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7853 // since we already have an optional CPSR def.
7854 bool definesCPSR = false;
7855 bool deadCPSR = false;
7856 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7858 const MachineOperand &MO = MI->getOperand(i);
7859 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7863 MI->RemoveOperand(i);
7868 assert(!NewOpc && "Optional cc_out operand required");
7871 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7873 assert(!MI->getOperand(ccOutIdx).getReg() &&
7874 "expect uninitialized optional cc_out operand");
7878 // If this instruction was defined with an optional CPSR def and its dag node
7879 // had a live implicit CPSR def, then activate the optional CPSR def.
7880 MachineOperand &MO = MI->getOperand(ccOutIdx);
7881 MO.setReg(ARM::CPSR);
7885 //===----------------------------------------------------------------------===//
7886 // ARM Optimization Hooks
7887 //===----------------------------------------------------------------------===//
7889 // Helper function that checks if N is a null or all ones constant.
7890 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7891 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7894 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7897 // Return true if N is conditionally 0 or all ones.
7898 // Detects these expressions where cc is an i1 value:
7900 // (select cc 0, y) [AllOnes=0]
7901 // (select cc y, 0) [AllOnes=0]
7902 // (zext cc) [AllOnes=0]
7903 // (sext cc) [AllOnes=0/1]
7904 // (select cc -1, y) [AllOnes=1]
7905 // (select cc y, -1) [AllOnes=1]
7907 // Invert is set when N is the null/all ones constant when CC is false.
7908 // OtherOp is set to the alternative value of N.
7909 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7910 SDValue &CC, bool &Invert,
7912 SelectionDAG &DAG) {
7913 switch (N->getOpcode()) {
7914 default: return false;
7916 CC = N->getOperand(0);
7917 SDValue N1 = N->getOperand(1);
7918 SDValue N2 = N->getOperand(2);
7919 if (isZeroOrAllOnes(N1, AllOnes)) {
7924 if (isZeroOrAllOnes(N2, AllOnes)) {
7931 case ISD::ZERO_EXTEND:
7932 // (zext cc) can never be the all ones value.
7936 case ISD::SIGN_EXTEND: {
7937 EVT VT = N->getValueType(0);
7938 CC = N->getOperand(0);
7939 if (CC.getValueType() != MVT::i1)
7943 // When looking for an AllOnes constant, N is an sext, and the 'other'
7945 OtherOp = DAG.getConstant(0, VT);
7946 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7947 // When looking for a 0 constant, N can be zext or sext.
7948 OtherOp = DAG.getConstant(1, VT);
7950 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7956 // Combine a constant select operand into its use:
7958 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7959 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7960 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7961 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7962 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7964 // The transform is rejected if the select doesn't have a constant operand that
7965 // is null, or all ones when AllOnes is set.
7967 // Also recognize sext/zext from i1:
7969 // (add (zext cc), x) -> (select cc (add x, 1), x)
7970 // (add (sext cc), x) -> (select cc (add x, -1), x)
7972 // These transformations eventually create predicated instructions.
7974 // @param N The node to transform.
7975 // @param Slct The N operand that is a select.
7976 // @param OtherOp The other N operand (x above).
7977 // @param DCI Context.
7978 // @param AllOnes Require the select constant to be all ones instead of null.
7979 // @returns The new node, or SDValue() on failure.
7981 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7982 TargetLowering::DAGCombinerInfo &DCI,
7983 bool AllOnes = false) {
7984 SelectionDAG &DAG = DCI.DAG;
7985 EVT VT = N->getValueType(0);
7986 SDValue NonConstantVal;
7989 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7990 NonConstantVal, DAG))
7993 // Slct is now know to be the desired identity constant when CC is true.
7994 SDValue TrueVal = OtherOp;
7995 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7996 OtherOp, NonConstantVal);
7997 // Unless SwapSelectOps says CC should be false.
7999 std::swap(TrueVal, FalseVal);
8001 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8002 CCOp, TrueVal, FalseVal);
8005 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8007 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8008 TargetLowering::DAGCombinerInfo &DCI) {
8009 SDValue N0 = N->getOperand(0);
8010 SDValue N1 = N->getOperand(1);
8011 if (N0.getNode()->hasOneUse()) {
8012 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8013 if (Result.getNode())
8016 if (N1.getNode()->hasOneUse()) {
8017 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8018 if (Result.getNode())
8024 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8025 // (only after legalization).
8026 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8027 TargetLowering::DAGCombinerInfo &DCI,
8028 const ARMSubtarget *Subtarget) {
8030 // Only perform optimization if after legalize, and if NEON is available. We
8031 // also expected both operands to be BUILD_VECTORs.
8032 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8033 || N0.getOpcode() != ISD::BUILD_VECTOR
8034 || N1.getOpcode() != ISD::BUILD_VECTOR)
8037 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8038 EVT VT = N->getValueType(0);
8039 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8042 // Check that the vector operands are of the right form.
8043 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8044 // operands, where N is the size of the formed vector.
8045 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8046 // index such that we have a pair wise add pattern.
8048 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8049 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8051 SDValue Vec = N0->getOperand(0)->getOperand(0);
8052 SDNode *V = Vec.getNode();
8053 unsigned nextIndex = 0;
8055 // For each operands to the ADD which are BUILD_VECTORs,
8056 // check to see if each of their operands are an EXTRACT_VECTOR with
8057 // the same vector and appropriate index.
8058 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8059 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8060 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8062 SDValue ExtVec0 = N0->getOperand(i);
8063 SDValue ExtVec1 = N1->getOperand(i);
8065 // First operand is the vector, verify its the same.
8066 if (V != ExtVec0->getOperand(0).getNode() ||
8067 V != ExtVec1->getOperand(0).getNode())
8070 // Second is the constant, verify its correct.
8071 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8072 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8074 // For the constant, we want to see all the even or all the odd.
8075 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8076 || C1->getZExtValue() != nextIndex+1)
8085 // Create VPADDL node.
8086 SelectionDAG &DAG = DCI.DAG;
8087 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8089 // Build operand list.
8090 SmallVector<SDValue, 8> Ops;
8091 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8092 TLI.getPointerTy()));
8094 // Input is the vector.
8097 // Get widened type and narrowed type.
8099 unsigned numElem = VT.getVectorNumElements();
8100 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8101 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8102 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8103 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8105 llvm_unreachable("Invalid vector element type for padd optimization.");
8108 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8109 widenType, &Ops[0], Ops.size());
8110 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8113 static SDValue findMUL_LOHI(SDValue V) {
8114 if (V->getOpcode() == ISD::UMUL_LOHI ||
8115 V->getOpcode() == ISD::SMUL_LOHI)
8120 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8121 TargetLowering::DAGCombinerInfo &DCI,
8122 const ARMSubtarget *Subtarget) {
8124 if (Subtarget->isThumb1Only()) return SDValue();
8126 // Only perform the checks after legalize when the pattern is available.
8127 if (DCI.isBeforeLegalize()) return SDValue();
8129 // Look for multiply add opportunities.
8130 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8131 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8132 // a glue link from the first add to the second add.
8133 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8134 // a S/UMLAL instruction.
8137 // \ / \ [no multiline comment]
8143 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8144 SDValue AddcOp0 = AddcNode->getOperand(0);
8145 SDValue AddcOp1 = AddcNode->getOperand(1);
8147 // Check if the two operands are from the same mul_lohi node.
8148 if (AddcOp0.getNode() == AddcOp1.getNode())
8151 assert(AddcNode->getNumValues() == 2 &&
8152 AddcNode->getValueType(0) == MVT::i32 &&
8153 "Expect ADDC with two result values. First: i32");
8155 // Check that we have a glued ADDC node.
8156 if (AddcNode->getValueType(1) != MVT::Glue)
8159 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8160 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8161 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8162 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8163 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8166 // Look for the glued ADDE.
8167 SDNode* AddeNode = AddcNode->getGluedUser();
8168 if (AddeNode == NULL)
8171 // Make sure it is really an ADDE.
8172 if (AddeNode->getOpcode() != ISD::ADDE)
8175 assert(AddeNode->getNumOperands() == 3 &&
8176 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8177 "ADDE node has the wrong inputs");
8179 // Check for the triangle shape.
8180 SDValue AddeOp0 = AddeNode->getOperand(0);
8181 SDValue AddeOp1 = AddeNode->getOperand(1);
8183 // Make sure that the ADDE operands are not coming from the same node.
8184 if (AddeOp0.getNode() == AddeOp1.getNode())
8187 // Find the MUL_LOHI node walking up ADDE's operands.
8188 bool IsLeftOperandMUL = false;
8189 SDValue MULOp = findMUL_LOHI(AddeOp0);
8190 if (MULOp == SDValue())
8191 MULOp = findMUL_LOHI(AddeOp1);
8193 IsLeftOperandMUL = true;
8194 if (MULOp == SDValue())
8197 // Figure out the right opcode.
8198 unsigned Opc = MULOp->getOpcode();
8199 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8201 // Figure out the high and low input values to the MLAL node.
8202 SDValue* HiMul = &MULOp;
8203 SDValue* HiAdd = NULL;
8204 SDValue* LoMul = NULL;
8205 SDValue* LowAdd = NULL;
8207 if (IsLeftOperandMUL)
8213 if (AddcOp0->getOpcode() == Opc) {
8217 if (AddcOp1->getOpcode() == Opc) {
8225 if (LoMul->getNode() != HiMul->getNode())
8228 // Create the merged node.
8229 SelectionDAG &DAG = DCI.DAG;
8231 // Build operand list.
8232 SmallVector<SDValue, 8> Ops;
8233 Ops.push_back(LoMul->getOperand(0));
8234 Ops.push_back(LoMul->getOperand(1));
8235 Ops.push_back(*LowAdd);
8236 Ops.push_back(*HiAdd);
8238 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8239 DAG.getVTList(MVT::i32, MVT::i32),
8240 &Ops[0], Ops.size());
8242 // Replace the ADDs' nodes uses by the MLA node's values.
8243 SDValue HiMLALResult(MLALNode.getNode(), 1);
8244 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8246 SDValue LoMLALResult(MLALNode.getNode(), 0);
8247 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8249 // Return original node to notify the driver to stop replacing.
8250 SDValue resNode(AddcNode, 0);
8254 /// PerformADDCCombine - Target-specific dag combine transform from
8255 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8256 static SDValue PerformADDCCombine(SDNode *N,
8257 TargetLowering::DAGCombinerInfo &DCI,
8258 const ARMSubtarget *Subtarget) {
8260 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8264 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8265 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8266 /// called with the default operands, and if that fails, with commuted
8268 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8269 TargetLowering::DAGCombinerInfo &DCI,
8270 const ARMSubtarget *Subtarget){
8272 // Attempt to create vpaddl for this add.
8273 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8274 if (Result.getNode())
8277 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8278 if (N0.getNode()->hasOneUse()) {
8279 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8280 if (Result.getNode()) return Result;
8285 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8287 static SDValue PerformADDCombine(SDNode *N,
8288 TargetLowering::DAGCombinerInfo &DCI,
8289 const ARMSubtarget *Subtarget) {
8290 SDValue N0 = N->getOperand(0);
8291 SDValue N1 = N->getOperand(1);
8293 // First try with the default operand order.
8294 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8295 if (Result.getNode())
8298 // If that didn't work, try again with the operands commuted.
8299 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8302 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8304 static SDValue PerformSUBCombine(SDNode *N,
8305 TargetLowering::DAGCombinerInfo &DCI) {
8306 SDValue N0 = N->getOperand(0);
8307 SDValue N1 = N->getOperand(1);
8309 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8310 if (N1.getNode()->hasOneUse()) {
8311 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8312 if (Result.getNode()) return Result;
8318 /// PerformVMULCombine
8319 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8320 /// special multiplier accumulator forwarding.
8326 static SDValue PerformVMULCombine(SDNode *N,
8327 TargetLowering::DAGCombinerInfo &DCI,
8328 const ARMSubtarget *Subtarget) {
8329 if (!Subtarget->hasVMLxForwarding())
8332 SelectionDAG &DAG = DCI.DAG;
8333 SDValue N0 = N->getOperand(0);
8334 SDValue N1 = N->getOperand(1);
8335 unsigned Opcode = N0.getOpcode();
8336 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8337 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8338 Opcode = N1.getOpcode();
8339 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8340 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8345 EVT VT = N->getValueType(0);
8347 SDValue N00 = N0->getOperand(0);
8348 SDValue N01 = N0->getOperand(1);
8349 return DAG.getNode(Opcode, DL, VT,
8350 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8351 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8354 static SDValue PerformMULCombine(SDNode *N,
8355 TargetLowering::DAGCombinerInfo &DCI,
8356 const ARMSubtarget *Subtarget) {
8357 SelectionDAG &DAG = DCI.DAG;
8359 if (Subtarget->isThumb1Only())
8362 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8365 EVT VT = N->getValueType(0);
8366 if (VT.is64BitVector() || VT.is128BitVector())
8367 return PerformVMULCombine(N, DCI, Subtarget);
8371 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8375 int64_t MulAmt = C->getSExtValue();
8376 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8378 ShiftAmt = ShiftAmt & (32 - 1);
8379 SDValue V = N->getOperand(0);
8383 MulAmt >>= ShiftAmt;
8386 if (isPowerOf2_32(MulAmt - 1)) {
8387 // (mul x, 2^N + 1) => (add (shl x, N), x)
8388 Res = DAG.getNode(ISD::ADD, DL, VT,
8390 DAG.getNode(ISD::SHL, DL, VT,
8392 DAG.getConstant(Log2_32(MulAmt - 1),
8394 } else if (isPowerOf2_32(MulAmt + 1)) {
8395 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8396 Res = DAG.getNode(ISD::SUB, DL, VT,
8397 DAG.getNode(ISD::SHL, DL, VT,
8399 DAG.getConstant(Log2_32(MulAmt + 1),
8405 uint64_t MulAmtAbs = -MulAmt;
8406 if (isPowerOf2_32(MulAmtAbs + 1)) {
8407 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8408 Res = DAG.getNode(ISD::SUB, DL, VT,
8410 DAG.getNode(ISD::SHL, DL, VT,
8412 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8414 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8415 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8416 Res = DAG.getNode(ISD::ADD, DL, VT,
8418 DAG.getNode(ISD::SHL, DL, VT,
8420 DAG.getConstant(Log2_32(MulAmtAbs-1),
8422 Res = DAG.getNode(ISD::SUB, DL, VT,
8423 DAG.getConstant(0, MVT::i32),Res);
8430 Res = DAG.getNode(ISD::SHL, DL, VT,
8431 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8433 // Do not add new nodes to DAG combiner worklist.
8434 DCI.CombineTo(N, Res, false);
8438 static SDValue PerformANDCombine(SDNode *N,
8439 TargetLowering::DAGCombinerInfo &DCI,
8440 const ARMSubtarget *Subtarget) {
8442 // Attempt to use immediate-form VBIC
8443 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8445 EVT VT = N->getValueType(0);
8446 SelectionDAG &DAG = DCI.DAG;
8448 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8451 APInt SplatBits, SplatUndef;
8452 unsigned SplatBitSize;
8455 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8456 if (SplatBitSize <= 64) {
8458 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8459 SplatUndef.getZExtValue(), SplatBitSize,
8460 DAG, VbicVT, VT.is128BitVector(),
8462 if (Val.getNode()) {
8464 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8465 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8466 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8471 if (!Subtarget->isThumb1Only()) {
8472 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8473 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8474 if (Result.getNode())
8481 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8482 static SDValue PerformORCombine(SDNode *N,
8483 TargetLowering::DAGCombinerInfo &DCI,
8484 const ARMSubtarget *Subtarget) {
8485 // Attempt to use immediate-form VORR
8486 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8488 EVT VT = N->getValueType(0);
8489 SelectionDAG &DAG = DCI.DAG;
8491 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8494 APInt SplatBits, SplatUndef;
8495 unsigned SplatBitSize;
8497 if (BVN && Subtarget->hasNEON() &&
8498 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8499 if (SplatBitSize <= 64) {
8501 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8502 SplatUndef.getZExtValue(), SplatBitSize,
8503 DAG, VorrVT, VT.is128BitVector(),
8505 if (Val.getNode()) {
8507 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8508 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8509 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8514 if (!Subtarget->isThumb1Only()) {
8515 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8516 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8517 if (Result.getNode())
8521 // The code below optimizes (or (and X, Y), Z).
8522 // The AND operand needs to have a single user to make these optimizations
8524 SDValue N0 = N->getOperand(0);
8525 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8527 SDValue N1 = N->getOperand(1);
8529 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8530 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8531 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8533 unsigned SplatBitSize;
8536 APInt SplatBits0, SplatBits1;
8537 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8538 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8539 // Ensure that the second operand of both ands are constants
8540 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8541 HasAnyUndefs) && !HasAnyUndefs) {
8542 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8543 HasAnyUndefs) && !HasAnyUndefs) {
8544 // Ensure that the bit width of the constants are the same and that
8545 // the splat arguments are logical inverses as per the pattern we
8546 // are trying to simplify.
8547 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8548 SplatBits0 == ~SplatBits1) {
8549 // Canonicalize the vector type to make instruction selection
8551 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8552 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8556 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8562 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8565 // BFI is only available on V6T2+
8566 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8570 // 1) or (and A, mask), val => ARMbfi A, val, mask
8571 // iff (val & mask) == val
8573 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8574 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8575 // && mask == ~mask2
8576 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8577 // && ~mask == mask2
8578 // (i.e., copy a bitfield value into another bitfield of the same width)
8583 SDValue N00 = N0.getOperand(0);
8585 // The value and the mask need to be constants so we can verify this is
8586 // actually a bitfield set. If the mask is 0xffff, we can do better
8587 // via a movt instruction, so don't use BFI in that case.
8588 SDValue MaskOp = N0.getOperand(1);
8589 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8592 unsigned Mask = MaskC->getZExtValue();
8596 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8597 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8599 unsigned Val = N1C->getZExtValue();
8600 if ((Val & ~Mask) != Val)
8603 if (ARM::isBitFieldInvertedMask(Mask)) {
8604 Val >>= countTrailingZeros(~Mask);
8606 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8607 DAG.getConstant(Val, MVT::i32),
8608 DAG.getConstant(Mask, MVT::i32));
8610 // Do not add new nodes to DAG combiner worklist.
8611 DCI.CombineTo(N, Res, false);
8614 } else if (N1.getOpcode() == ISD::AND) {
8615 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8616 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8619 unsigned Mask2 = N11C->getZExtValue();
8621 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8623 if (ARM::isBitFieldInvertedMask(Mask) &&
8625 // The pack halfword instruction works better for masks that fit it,
8626 // so use that when it's available.
8627 if (Subtarget->hasT2ExtractPack() &&
8628 (Mask == 0xffff || Mask == 0xffff0000))
8631 unsigned amt = countTrailingZeros(Mask2);
8632 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8633 DAG.getConstant(amt, MVT::i32));
8634 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8635 DAG.getConstant(Mask, MVT::i32));
8636 // Do not add new nodes to DAG combiner worklist.
8637 DCI.CombineTo(N, Res, false);
8639 } else 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 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8647 unsigned lsb = countTrailingZeros(Mask);
8648 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8649 DAG.getConstant(lsb, MVT::i32));
8650 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8651 DAG.getConstant(Mask2, MVT::i32));
8652 // Do not add new nodes to DAG combiner worklist.
8653 DCI.CombineTo(N, Res, false);
8658 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8659 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8660 ARM::isBitFieldInvertedMask(~Mask)) {
8661 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8662 // where lsb(mask) == #shamt and masked bits of B are known zero.
8663 SDValue ShAmt = N00.getOperand(1);
8664 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8665 unsigned LSB = countTrailingZeros(Mask);
8669 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8670 DAG.getConstant(~Mask, MVT::i32));
8672 // Do not add new nodes to DAG combiner worklist.
8673 DCI.CombineTo(N, Res, false);
8679 static SDValue PerformXORCombine(SDNode *N,
8680 TargetLowering::DAGCombinerInfo &DCI,
8681 const ARMSubtarget *Subtarget) {
8682 EVT VT = N->getValueType(0);
8683 SelectionDAG &DAG = DCI.DAG;
8685 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8688 if (!Subtarget->isThumb1Only()) {
8689 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8690 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8691 if (Result.getNode())
8698 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8699 /// the bits being cleared by the AND are not demanded by the BFI.
8700 static SDValue PerformBFICombine(SDNode *N,
8701 TargetLowering::DAGCombinerInfo &DCI) {
8702 SDValue N1 = N->getOperand(1);
8703 if (N1.getOpcode() == ISD::AND) {
8704 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8707 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8708 unsigned LSB = countTrailingZeros(~InvMask);
8709 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8710 unsigned Mask = (1 << Width)-1;
8711 unsigned Mask2 = N11C->getZExtValue();
8712 if ((Mask & (~Mask2)) == 0)
8713 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8714 N->getOperand(0), N1.getOperand(0),
8720 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8721 /// ARMISD::VMOVRRD.
8722 static SDValue PerformVMOVRRDCombine(SDNode *N,
8723 TargetLowering::DAGCombinerInfo &DCI) {
8724 // vmovrrd(vmovdrr x, y) -> x,y
8725 SDValue InDouble = N->getOperand(0);
8726 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8727 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8729 // vmovrrd(load f64) -> (load i32), (load i32)
8730 SDNode *InNode = InDouble.getNode();
8731 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8732 InNode->getValueType(0) == MVT::f64 &&
8733 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8734 !cast<LoadSDNode>(InNode)->isVolatile()) {
8735 // TODO: Should this be done for non-FrameIndex operands?
8736 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8738 SelectionDAG &DAG = DCI.DAG;
8740 SDValue BasePtr = LD->getBasePtr();
8741 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8742 LD->getPointerInfo(), LD->isVolatile(),
8743 LD->isNonTemporal(), LD->isInvariant(),
8744 LD->getAlignment());
8746 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8747 DAG.getConstant(4, MVT::i32));
8748 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8749 LD->getPointerInfo(), LD->isVolatile(),
8750 LD->isNonTemporal(), LD->isInvariant(),
8751 std::min(4U, LD->getAlignment() / 2));
8753 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8754 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8755 DCI.RemoveFromWorklist(LD);
8763 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8764 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8765 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8766 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8767 SDValue Op0 = N->getOperand(0);
8768 SDValue Op1 = N->getOperand(1);
8769 if (Op0.getOpcode() == ISD::BITCAST)
8770 Op0 = Op0.getOperand(0);
8771 if (Op1.getOpcode() == ISD::BITCAST)
8772 Op1 = Op1.getOperand(0);
8773 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8774 Op0.getNode() == Op1.getNode() &&
8775 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8776 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8777 N->getValueType(0), Op0.getOperand(0));
8781 /// PerformSTORECombine - Target-specific dag combine xforms for
8783 static SDValue PerformSTORECombine(SDNode *N,
8784 TargetLowering::DAGCombinerInfo &DCI) {
8785 StoreSDNode *St = cast<StoreSDNode>(N);
8786 if (St->isVolatile())
8789 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8790 // pack all of the elements in one place. Next, store to memory in fewer
8792 SDValue StVal = St->getValue();
8793 EVT VT = StVal.getValueType();
8794 if (St->isTruncatingStore() && VT.isVector()) {
8795 SelectionDAG &DAG = DCI.DAG;
8796 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8797 EVT StVT = St->getMemoryVT();
8798 unsigned NumElems = VT.getVectorNumElements();
8799 assert(StVT != VT && "Cannot truncate to the same type");
8800 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8801 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8803 // From, To sizes and ElemCount must be pow of two
8804 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8806 // We are going to use the original vector elt for storing.
8807 // Accumulated smaller vector elements must be a multiple of the store size.
8808 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8810 unsigned SizeRatio = FromEltSz / ToEltSz;
8811 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8813 // Create a type on which we perform the shuffle.
8814 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8815 NumElems*SizeRatio);
8816 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8819 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8820 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8821 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8823 // Can't shuffle using an illegal type.
8824 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8826 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8827 DAG.getUNDEF(WideVec.getValueType()),
8829 // At this point all of the data is stored at the bottom of the
8830 // register. We now need to save it to mem.
8832 // Find the largest store unit
8833 MVT StoreType = MVT::i8;
8834 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8835 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8836 MVT Tp = (MVT::SimpleValueType)tp;
8837 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8840 // Didn't find a legal store type.
8841 if (!TLI.isTypeLegal(StoreType))
8844 // Bitcast the original vector into a vector of store-size units
8845 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8846 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8847 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8848 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8849 SmallVector<SDValue, 8> Chains;
8850 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8851 TLI.getPointerTy());
8852 SDValue BasePtr = St->getBasePtr();
8854 // Perform one or more big stores into memory.
8855 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8856 for (unsigned I = 0; I < E; I++) {
8857 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8858 StoreType, ShuffWide,
8859 DAG.getIntPtrConstant(I));
8860 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8861 St->getPointerInfo(), St->isVolatile(),
8862 St->isNonTemporal(), St->getAlignment());
8863 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8865 Chains.push_back(Ch);
8867 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8871 if (!ISD::isNormalStore(St))
8874 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8875 // ARM stores of arguments in the same cache line.
8876 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8877 StVal.getNode()->hasOneUse()) {
8878 SelectionDAG &DAG = DCI.DAG;
8880 SDValue BasePtr = St->getBasePtr();
8881 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8882 StVal.getNode()->getOperand(0), BasePtr,
8883 St->getPointerInfo(), St->isVolatile(),
8884 St->isNonTemporal(), St->getAlignment());
8886 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8887 DAG.getConstant(4, MVT::i32));
8888 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8889 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8890 St->isNonTemporal(),
8891 std::min(4U, St->getAlignment() / 2));
8894 if (StVal.getValueType() != MVT::i64 ||
8895 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8898 // Bitcast an i64 store extracted from a vector to f64.
8899 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8900 SelectionDAG &DAG = DCI.DAG;
8902 SDValue IntVec = StVal.getOperand(0);
8903 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8904 IntVec.getValueType().getVectorNumElements());
8905 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8906 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8907 Vec, StVal.getOperand(1));
8909 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8910 // Make the DAGCombiner fold the bitcasts.
8911 DCI.AddToWorklist(Vec.getNode());
8912 DCI.AddToWorklist(ExtElt.getNode());
8913 DCI.AddToWorklist(V.getNode());
8914 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8915 St->getPointerInfo(), St->isVolatile(),
8916 St->isNonTemporal(), St->getAlignment(),
8920 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8921 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8922 /// i64 vector to have f64 elements, since the value can then be loaded
8923 /// directly into a VFP register.
8924 static bool hasNormalLoadOperand(SDNode *N) {
8925 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8926 for (unsigned i = 0; i < NumElts; ++i) {
8927 SDNode *Elt = N->getOperand(i).getNode();
8928 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8934 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8935 /// ISD::BUILD_VECTOR.
8936 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8937 TargetLowering::DAGCombinerInfo &DCI){
8938 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8939 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8940 // into a pair of GPRs, which is fine when the value is used as a scalar,
8941 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8942 SelectionDAG &DAG = DCI.DAG;
8943 if (N->getNumOperands() == 2) {
8944 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8949 // Load i64 elements as f64 values so that type legalization does not split
8950 // them up into i32 values.
8951 EVT VT = N->getValueType(0);
8952 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8955 SmallVector<SDValue, 8> Ops;
8956 unsigned NumElts = VT.getVectorNumElements();
8957 for (unsigned i = 0; i < NumElts; ++i) {
8958 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8960 // Make the DAGCombiner fold the bitcast.
8961 DCI.AddToWorklist(V.getNode());
8963 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8964 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8965 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8968 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8970 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8971 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8972 // At that time, we may have inserted bitcasts from integer to float.
8973 // If these bitcasts have survived DAGCombine, change the lowering of this
8974 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8975 // force to use floating point types.
8977 // Make sure we can change the type of the vector.
8978 // This is possible iff:
8979 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8980 // 1.1. Vector is used only once.
8981 // 1.2. Use is a bit convert to an integer type.
8982 // 2. The size of its operands are 32-bits (64-bits are not legal).
8983 EVT VT = N->getValueType(0);
8984 EVT EltVT = VT.getVectorElementType();
8986 // Check 1.1. and 2.
8987 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8990 // By construction, the input type must be float.
8991 assert(EltVT == MVT::f32 && "Unexpected type!");
8994 SDNode *Use = *N->use_begin();
8995 if (Use->getOpcode() != ISD::BITCAST ||
8996 Use->getValueType(0).isFloatingPoint())
8999 // Check profitability.
9000 // Model is, if more than half of the relevant operands are bitcast from
9001 // i32, turn the build_vector into a sequence of insert_vector_elt.
9002 // Relevant operands are everything that is not statically
9003 // (i.e., at compile time) bitcasted.
9004 unsigned NumOfBitCastedElts = 0;
9005 unsigned NumElts = VT.getVectorNumElements();
9006 unsigned NumOfRelevantElts = NumElts;
9007 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9008 SDValue Elt = N->getOperand(Idx);
9009 if (Elt->getOpcode() == ISD::BITCAST) {
9010 // Assume only bit cast to i32 will go away.
9011 if (Elt->getOperand(0).getValueType() == MVT::i32)
9012 ++NumOfBitCastedElts;
9013 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9014 // Constants are statically casted, thus do not count them as
9015 // relevant operands.
9016 --NumOfRelevantElts;
9019 // Check if more than half of the elements require a non-free bitcast.
9020 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9023 SelectionDAG &DAG = DCI.DAG;
9024 // Create the new vector type.
9025 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9026 // Check if the type is legal.
9027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9028 if (!TLI.isTypeLegal(VecVT))
9032 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9033 // => BITCAST INSERT_VECTOR_ELT
9034 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9036 SDValue Vec = DAG.getUNDEF(VecVT);
9038 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9039 SDValue V = N->getOperand(Idx);
9040 if (V.getOpcode() == ISD::UNDEF)
9042 if (V.getOpcode() == ISD::BITCAST &&
9043 V->getOperand(0).getValueType() == MVT::i32)
9044 // Fold obvious case.
9045 V = V.getOperand(0);
9047 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9048 // Make the DAGCombiner fold the bitcasts.
9049 DCI.AddToWorklist(V.getNode());
9051 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9052 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9054 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9055 // Make the DAGCombiner fold the bitcasts.
9056 DCI.AddToWorklist(Vec.getNode());
9060 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9061 /// ISD::INSERT_VECTOR_ELT.
9062 static SDValue PerformInsertEltCombine(SDNode *N,
9063 TargetLowering::DAGCombinerInfo &DCI) {
9064 // Bitcast an i64 load inserted into a vector to f64.
9065 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9066 EVT VT = N->getValueType(0);
9067 SDNode *Elt = N->getOperand(1).getNode();
9068 if (VT.getVectorElementType() != MVT::i64 ||
9069 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9072 SelectionDAG &DAG = DCI.DAG;
9074 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9075 VT.getVectorNumElements());
9076 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9077 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9078 // Make the DAGCombiner fold the bitcasts.
9079 DCI.AddToWorklist(Vec.getNode());
9080 DCI.AddToWorklist(V.getNode());
9081 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9082 Vec, V, N->getOperand(2));
9083 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9086 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9087 /// ISD::VECTOR_SHUFFLE.
9088 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9089 // The LLVM shufflevector instruction does not require the shuffle mask
9090 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9091 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9092 // operands do not match the mask length, they are extended by concatenating
9093 // them with undef vectors. That is probably the right thing for other
9094 // targets, but for NEON it is better to concatenate two double-register
9095 // size vector operands into a single quad-register size vector. Do that
9096 // transformation here:
9097 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9098 // shuffle(concat(v1, v2), undef)
9099 SDValue Op0 = N->getOperand(0);
9100 SDValue Op1 = N->getOperand(1);
9101 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9102 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9103 Op0.getNumOperands() != 2 ||
9104 Op1.getNumOperands() != 2)
9106 SDValue Concat0Op1 = Op0.getOperand(1);
9107 SDValue Concat1Op1 = Op1.getOperand(1);
9108 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9109 Concat1Op1.getOpcode() != ISD::UNDEF)
9111 // Skip the transformation if any of the types are illegal.
9112 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9113 EVT VT = N->getValueType(0);
9114 if (!TLI.isTypeLegal(VT) ||
9115 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9116 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9119 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9120 Op0.getOperand(0), Op1.getOperand(0));
9121 // Translate the shuffle mask.
9122 SmallVector<int, 16> NewMask;
9123 unsigned NumElts = VT.getVectorNumElements();
9124 unsigned HalfElts = NumElts/2;
9125 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9126 for (unsigned n = 0; n < NumElts; ++n) {
9127 int MaskElt = SVN->getMaskElt(n);
9129 if (MaskElt < (int)HalfElts)
9131 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9132 NewElt = HalfElts + MaskElt - NumElts;
9133 NewMask.push_back(NewElt);
9135 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9136 DAG.getUNDEF(VT), NewMask.data());
9139 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9140 /// NEON load/store intrinsics to merge base address updates.
9141 static SDValue CombineBaseUpdate(SDNode *N,
9142 TargetLowering::DAGCombinerInfo &DCI) {
9143 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9146 SelectionDAG &DAG = DCI.DAG;
9147 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9148 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9149 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9150 SDValue Addr = N->getOperand(AddrOpIdx);
9152 // Search for a use of the address operand that is an increment.
9153 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9154 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9156 if (User->getOpcode() != ISD::ADD ||
9157 UI.getUse().getResNo() != Addr.getResNo())
9160 // Check that the add is independent of the load/store. Otherwise, folding
9161 // it would create a cycle.
9162 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9165 // Find the new opcode for the updating load/store.
9167 bool isLaneOp = false;
9168 unsigned NewOpc = 0;
9169 unsigned NumVecs = 0;
9171 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9173 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9174 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9176 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9178 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9180 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9182 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9183 NumVecs = 2; isLaneOp = true; break;
9184 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9185 NumVecs = 3; isLaneOp = true; break;
9186 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9187 NumVecs = 4; isLaneOp = true; break;
9188 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9189 NumVecs = 1; isLoad = false; break;
9190 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9191 NumVecs = 2; isLoad = false; break;
9192 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9193 NumVecs = 3; isLoad = false; break;
9194 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9195 NumVecs = 4; isLoad = false; break;
9196 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9197 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9198 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9199 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9200 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9201 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9205 switch (N->getOpcode()) {
9206 default: llvm_unreachable("unexpected opcode for Neon base update");
9207 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9208 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9209 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9213 // Find the size of memory referenced by the load/store.
9216 VecTy = N->getValueType(0);
9218 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9219 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9221 NumBytes /= VecTy.getVectorNumElements();
9223 // If the increment is a constant, it must match the memory ref size.
9224 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9225 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9226 uint64_t IncVal = CInc->getZExtValue();
9227 if (IncVal != NumBytes)
9229 } else if (NumBytes >= 3 * 16) {
9230 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9231 // separate instructions that make it harder to use a non-constant update.
9235 // Create the new updating load/store node.
9237 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9239 for (n = 0; n < NumResultVecs; ++n)
9241 Tys[n++] = MVT::i32;
9242 Tys[n] = MVT::Other;
9243 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9244 SmallVector<SDValue, 8> Ops;
9245 Ops.push_back(N->getOperand(0)); // incoming chain
9246 Ops.push_back(N->getOperand(AddrOpIdx));
9248 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9249 Ops.push_back(N->getOperand(i));
9251 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9252 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9253 Ops.data(), Ops.size(),
9254 MemInt->getMemoryVT(),
9255 MemInt->getMemOperand());
9258 std::vector<SDValue> NewResults;
9259 for (unsigned i = 0; i < NumResultVecs; ++i) {
9260 NewResults.push_back(SDValue(UpdN.getNode(), i));
9262 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9263 DCI.CombineTo(N, NewResults);
9264 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9271 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9272 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9273 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9275 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9276 SelectionDAG &DAG = DCI.DAG;
9277 EVT VT = N->getValueType(0);
9278 // vldN-dup instructions only support 64-bit vectors for N > 1.
9279 if (!VT.is64BitVector())
9282 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9283 SDNode *VLD = N->getOperand(0).getNode();
9284 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9286 unsigned NumVecs = 0;
9287 unsigned NewOpc = 0;
9288 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9289 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9291 NewOpc = ARMISD::VLD2DUP;
9292 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9294 NewOpc = ARMISD::VLD3DUP;
9295 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9297 NewOpc = ARMISD::VLD4DUP;
9302 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9303 // numbers match the load.
9304 unsigned VLDLaneNo =
9305 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9306 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9308 // Ignore uses of the chain result.
9309 if (UI.getUse().getResNo() == NumVecs)
9312 if (User->getOpcode() != ARMISD::VDUPLANE ||
9313 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9317 // Create the vldN-dup node.
9320 for (n = 0; n < NumVecs; ++n)
9322 Tys[n] = MVT::Other;
9323 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9324 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9325 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9326 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9327 Ops, 2, VLDMemInt->getMemoryVT(),
9328 VLDMemInt->getMemOperand());
9331 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9333 unsigned ResNo = UI.getUse().getResNo();
9334 // Ignore uses of the chain result.
9335 if (ResNo == NumVecs)
9338 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9341 // Now the vldN-lane intrinsic is dead except for its chain result.
9342 // Update uses of the chain.
9343 std::vector<SDValue> VLDDupResults;
9344 for (unsigned n = 0; n < NumVecs; ++n)
9345 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9346 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9347 DCI.CombineTo(VLD, VLDDupResults);
9352 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9353 /// ARMISD::VDUPLANE.
9354 static SDValue PerformVDUPLANECombine(SDNode *N,
9355 TargetLowering::DAGCombinerInfo &DCI) {
9356 SDValue Op = N->getOperand(0);
9358 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9359 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9360 if (CombineVLDDUP(N, DCI))
9361 return SDValue(N, 0);
9363 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9364 // redundant. Ignore bit_converts for now; element sizes are checked below.
9365 while (Op.getOpcode() == ISD::BITCAST)
9366 Op = Op.getOperand(0);
9367 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9370 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9371 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9372 // The canonical VMOV for a zero vector uses a 32-bit element size.
9373 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9375 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9377 EVT VT = N->getValueType(0);
9378 if (EltSize > VT.getVectorElementType().getSizeInBits())
9381 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9384 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9385 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9386 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9390 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9392 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9397 APFloat APF = C->getValueAPF();
9398 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9399 != APFloat::opOK || !isExact)
9402 c0 = (I == 0) ? cN : c0;
9403 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9410 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9411 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9412 /// when the VMUL has a constant operand that is a power of 2.
9414 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9415 /// vmul.f32 d16, d17, d16
9416 /// vcvt.s32.f32 d16, d16
9418 /// vcvt.s32.f32 d16, d16, #3
9419 static SDValue PerformVCVTCombine(SDNode *N,
9420 TargetLowering::DAGCombinerInfo &DCI,
9421 const ARMSubtarget *Subtarget) {
9422 SelectionDAG &DAG = DCI.DAG;
9423 SDValue Op = N->getOperand(0);
9425 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9426 Op.getOpcode() != ISD::FMUL)
9430 SDValue N0 = Op->getOperand(0);
9431 SDValue ConstVec = Op->getOperand(1);
9432 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9434 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9435 !isConstVecPow2(ConstVec, isSigned, C))
9438 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9439 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9440 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9441 // These instructions only exist converting from f32 to i32. We can handle
9442 // smaller integers by generating an extra truncate, but larger ones would
9447 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9448 Intrinsic::arm_neon_vcvtfp2fxu;
9449 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9450 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9451 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9452 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9453 DAG.getConstant(Log2_64(C), MVT::i32));
9455 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9456 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9461 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9462 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9463 /// when the VDIV has a constant operand that is a power of 2.
9465 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9466 /// vcvt.f32.s32 d16, d16
9467 /// vdiv.f32 d16, d17, d16
9469 /// vcvt.f32.s32 d16, d16, #3
9470 static SDValue PerformVDIVCombine(SDNode *N,
9471 TargetLowering::DAGCombinerInfo &DCI,
9472 const ARMSubtarget *Subtarget) {
9473 SelectionDAG &DAG = DCI.DAG;
9474 SDValue Op = N->getOperand(0);
9475 unsigned OpOpcode = Op.getNode()->getOpcode();
9477 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9478 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9482 SDValue ConstVec = N->getOperand(1);
9483 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9485 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9486 !isConstVecPow2(ConstVec, isSigned, C))
9489 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9490 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9491 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9492 // These instructions only exist converting from i32 to f32. We can handle
9493 // smaller integers by generating an extra extend, but larger ones would
9498 SDValue ConvInput = Op.getOperand(0);
9499 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9500 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9501 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9502 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9505 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9506 Intrinsic::arm_neon_vcvtfxu2fp;
9507 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9509 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9510 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9513 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9514 /// operand of a vector shift operation, where all the elements of the
9515 /// build_vector must have the same constant integer value.
9516 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9517 // Ignore bit_converts.
9518 while (Op.getOpcode() == ISD::BITCAST)
9519 Op = Op.getOperand(0);
9520 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9521 APInt SplatBits, SplatUndef;
9522 unsigned SplatBitSize;
9524 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9525 HasAnyUndefs, ElementBits) ||
9526 SplatBitSize > ElementBits)
9528 Cnt = SplatBits.getSExtValue();
9532 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9533 /// operand of a vector shift left operation. That value must be in the range:
9534 /// 0 <= Value < ElementBits for a left shift; or
9535 /// 0 <= Value <= ElementBits for a long left shift.
9536 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9537 assert(VT.isVector() && "vector shift count is not a vector type");
9538 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9539 if (! getVShiftImm(Op, ElementBits, Cnt))
9541 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9544 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9545 /// operand of a vector shift right operation. For a shift opcode, the value
9546 /// is positive, but for an intrinsic the value count must be negative. The
9547 /// absolute value must be in the range:
9548 /// 1 <= |Value| <= ElementBits for a right shift; or
9549 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9550 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9552 assert(VT.isVector() && "vector shift count is not a vector type");
9553 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9554 if (! getVShiftImm(Op, ElementBits, Cnt))
9558 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9561 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9562 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9563 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9566 // Don't do anything for most intrinsics.
9569 // Vector shifts: check for immediate versions and lower them.
9570 // Note: This is done during DAG combining instead of DAG legalizing because
9571 // the build_vectors for 64-bit vector element shift counts are generally
9572 // not legal, and it is hard to see their values after they get legalized to
9573 // loads from a constant pool.
9574 case Intrinsic::arm_neon_vshifts:
9575 case Intrinsic::arm_neon_vshiftu:
9576 case Intrinsic::arm_neon_vshiftls:
9577 case Intrinsic::arm_neon_vshiftlu:
9578 case Intrinsic::arm_neon_vshiftn:
9579 case Intrinsic::arm_neon_vrshifts:
9580 case Intrinsic::arm_neon_vrshiftu:
9581 case Intrinsic::arm_neon_vrshiftn:
9582 case Intrinsic::arm_neon_vqshifts:
9583 case Intrinsic::arm_neon_vqshiftu:
9584 case Intrinsic::arm_neon_vqshiftsu:
9585 case Intrinsic::arm_neon_vqshiftns:
9586 case Intrinsic::arm_neon_vqshiftnu:
9587 case Intrinsic::arm_neon_vqshiftnsu:
9588 case Intrinsic::arm_neon_vqrshiftns:
9589 case Intrinsic::arm_neon_vqrshiftnu:
9590 case Intrinsic::arm_neon_vqrshiftnsu: {
9591 EVT VT = N->getOperand(1).getValueType();
9593 unsigned VShiftOpc = 0;
9596 case Intrinsic::arm_neon_vshifts:
9597 case Intrinsic::arm_neon_vshiftu:
9598 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9599 VShiftOpc = ARMISD::VSHL;
9602 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9603 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9604 ARMISD::VSHRs : ARMISD::VSHRu);
9609 case Intrinsic::arm_neon_vshiftls:
9610 case Intrinsic::arm_neon_vshiftlu:
9611 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9613 llvm_unreachable("invalid shift count for vshll intrinsic");
9615 case Intrinsic::arm_neon_vrshifts:
9616 case Intrinsic::arm_neon_vrshiftu:
9617 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9621 case Intrinsic::arm_neon_vqshifts:
9622 case Intrinsic::arm_neon_vqshiftu:
9623 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9627 case Intrinsic::arm_neon_vqshiftsu:
9628 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9630 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9632 case Intrinsic::arm_neon_vshiftn:
9633 case Intrinsic::arm_neon_vrshiftn:
9634 case Intrinsic::arm_neon_vqshiftns:
9635 case Intrinsic::arm_neon_vqshiftnu:
9636 case Intrinsic::arm_neon_vqshiftnsu:
9637 case Intrinsic::arm_neon_vqrshiftns:
9638 case Intrinsic::arm_neon_vqrshiftnu:
9639 case Intrinsic::arm_neon_vqrshiftnsu:
9640 // Narrowing shifts require an immediate right shift.
9641 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9643 llvm_unreachable("invalid shift count for narrowing vector shift "
9647 llvm_unreachable("unhandled vector shift");
9651 case Intrinsic::arm_neon_vshifts:
9652 case Intrinsic::arm_neon_vshiftu:
9653 // Opcode already set above.
9655 case Intrinsic::arm_neon_vshiftls:
9656 case Intrinsic::arm_neon_vshiftlu:
9657 if (Cnt == VT.getVectorElementType().getSizeInBits())
9658 VShiftOpc = ARMISD::VSHLLi;
9660 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9661 ARMISD::VSHLLs : ARMISD::VSHLLu);
9663 case Intrinsic::arm_neon_vshiftn:
9664 VShiftOpc = ARMISD::VSHRN; break;
9665 case Intrinsic::arm_neon_vrshifts:
9666 VShiftOpc = ARMISD::VRSHRs; break;
9667 case Intrinsic::arm_neon_vrshiftu:
9668 VShiftOpc = ARMISD::VRSHRu; break;
9669 case Intrinsic::arm_neon_vrshiftn:
9670 VShiftOpc = ARMISD::VRSHRN; break;
9671 case Intrinsic::arm_neon_vqshifts:
9672 VShiftOpc = ARMISD::VQSHLs; break;
9673 case Intrinsic::arm_neon_vqshiftu:
9674 VShiftOpc = ARMISD::VQSHLu; break;
9675 case Intrinsic::arm_neon_vqshiftsu:
9676 VShiftOpc = ARMISD::VQSHLsu; break;
9677 case Intrinsic::arm_neon_vqshiftns:
9678 VShiftOpc = ARMISD::VQSHRNs; break;
9679 case Intrinsic::arm_neon_vqshiftnu:
9680 VShiftOpc = ARMISD::VQSHRNu; break;
9681 case Intrinsic::arm_neon_vqshiftnsu:
9682 VShiftOpc = ARMISD::VQSHRNsu; break;
9683 case Intrinsic::arm_neon_vqrshiftns:
9684 VShiftOpc = ARMISD::VQRSHRNs; break;
9685 case Intrinsic::arm_neon_vqrshiftnu:
9686 VShiftOpc = ARMISD::VQRSHRNu; break;
9687 case Intrinsic::arm_neon_vqrshiftnsu:
9688 VShiftOpc = ARMISD::VQRSHRNsu; break;
9691 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9692 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9695 case Intrinsic::arm_neon_vshiftins: {
9696 EVT VT = N->getOperand(1).getValueType();
9698 unsigned VShiftOpc = 0;
9700 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9701 VShiftOpc = ARMISD::VSLI;
9702 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9703 VShiftOpc = ARMISD::VSRI;
9705 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9708 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9709 N->getOperand(1), N->getOperand(2),
9710 DAG.getConstant(Cnt, MVT::i32));
9713 case Intrinsic::arm_neon_vqrshifts:
9714 case Intrinsic::arm_neon_vqrshiftu:
9715 // No immediate versions of these to check for.
9722 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9723 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9724 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9725 /// vector element shift counts are generally not legal, and it is hard to see
9726 /// their values after they get legalized to loads from a constant pool.
9727 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9728 const ARMSubtarget *ST) {
9729 EVT VT = N->getValueType(0);
9730 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9731 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9732 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9733 SDValue N1 = N->getOperand(1);
9734 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9735 SDValue N0 = N->getOperand(0);
9736 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9737 DAG.MaskedValueIsZero(N0.getOperand(0),
9738 APInt::getHighBitsSet(32, 16)))
9739 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9743 // Nothing to be done for scalar shifts.
9744 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9745 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9748 assert(ST->hasNEON() && "unexpected vector shift");
9751 switch (N->getOpcode()) {
9752 default: llvm_unreachable("unexpected shift opcode");
9755 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9756 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9757 DAG.getConstant(Cnt, MVT::i32));
9762 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9763 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9764 ARMISD::VSHRs : ARMISD::VSHRu);
9765 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9766 DAG.getConstant(Cnt, MVT::i32));
9772 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9773 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9774 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9775 const ARMSubtarget *ST) {
9776 SDValue N0 = N->getOperand(0);
9778 // Check for sign- and zero-extensions of vector extract operations of 8-
9779 // and 16-bit vector elements. NEON supports these directly. They are
9780 // handled during DAG combining because type legalization will promote them
9781 // to 32-bit types and it is messy to recognize the operations after that.
9782 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9783 SDValue Vec = N0.getOperand(0);
9784 SDValue Lane = N0.getOperand(1);
9785 EVT VT = N->getValueType(0);
9786 EVT EltVT = N0.getValueType();
9787 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9789 if (VT == MVT::i32 &&
9790 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9791 TLI.isTypeLegal(Vec.getValueType()) &&
9792 isa<ConstantSDNode>(Lane)) {
9795 switch (N->getOpcode()) {
9796 default: llvm_unreachable("unexpected opcode");
9797 case ISD::SIGN_EXTEND:
9798 Opc = ARMISD::VGETLANEs;
9800 case ISD::ZERO_EXTEND:
9801 case ISD::ANY_EXTEND:
9802 Opc = ARMISD::VGETLANEu;
9805 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9812 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9813 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9814 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9815 const ARMSubtarget *ST) {
9816 // If the target supports NEON, try to use vmax/vmin instructions for f32
9817 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9818 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9819 // a NaN; only do the transformation when it matches that behavior.
9821 // For now only do this when using NEON for FP operations; if using VFP, it
9822 // is not obvious that the benefit outweighs the cost of switching to the
9824 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9825 N->getValueType(0) != MVT::f32)
9828 SDValue CondLHS = N->getOperand(0);
9829 SDValue CondRHS = N->getOperand(1);
9830 SDValue LHS = N->getOperand(2);
9831 SDValue RHS = N->getOperand(3);
9832 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9834 unsigned Opcode = 0;
9836 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9837 IsReversed = false; // x CC y ? x : y
9838 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9839 IsReversed = true ; // x CC y ? y : x
9853 // If LHS is NaN, an ordered comparison will be false and the result will
9854 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9855 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9856 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9857 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9859 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9860 // will return -0, so vmin can only be used for unsafe math or if one of
9861 // the operands is known to be nonzero.
9862 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9863 !DAG.getTarget().Options.UnsafeFPMath &&
9864 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9866 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9875 // If LHS is NaN, an ordered comparison will be false and the result will
9876 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9877 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9878 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9879 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9881 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9882 // will return +0, so vmax can only be used for unsafe math or if one of
9883 // the operands is known to be nonzero.
9884 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9885 !DAG.getTarget().Options.UnsafeFPMath &&
9886 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9888 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9894 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9897 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9899 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9900 SDValue Cmp = N->getOperand(4);
9901 if (Cmp.getOpcode() != ARMISD::CMPZ)
9902 // Only looking at EQ and NE cases.
9905 EVT VT = N->getValueType(0);
9907 SDValue LHS = Cmp.getOperand(0);
9908 SDValue RHS = Cmp.getOperand(1);
9909 SDValue FalseVal = N->getOperand(0);
9910 SDValue TrueVal = N->getOperand(1);
9911 SDValue ARMcc = N->getOperand(2);
9912 ARMCC::CondCodes CC =
9913 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9931 /// FIXME: Turn this into a target neutral optimization?
9933 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9934 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9935 N->getOperand(3), Cmp);
9936 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9938 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9939 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9940 N->getOperand(3), NewCmp);
9943 if (Res.getNode()) {
9944 APInt KnownZero, KnownOne;
9945 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9946 // Capture demanded bits information that would be otherwise lost.
9947 if (KnownZero == 0xfffffffe)
9948 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9949 DAG.getValueType(MVT::i1));
9950 else if (KnownZero == 0xffffff00)
9951 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9952 DAG.getValueType(MVT::i8));
9953 else if (KnownZero == 0xffff0000)
9954 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9955 DAG.getValueType(MVT::i16));
9961 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9962 DAGCombinerInfo &DCI) const {
9963 switch (N->getOpcode()) {
9965 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9966 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9967 case ISD::SUB: return PerformSUBCombine(N, DCI);
9968 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9969 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9970 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9971 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9972 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9973 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9974 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9975 case ISD::STORE: return PerformSTORECombine(N, DCI);
9976 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9977 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9978 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9979 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9980 case ISD::FP_TO_SINT:
9981 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9982 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9983 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9986 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9987 case ISD::SIGN_EXTEND:
9988 case ISD::ZERO_EXTEND:
9989 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9990 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9991 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9992 case ARMISD::VLD2DUP:
9993 case ARMISD::VLD3DUP:
9994 case ARMISD::VLD4DUP:
9995 return CombineBaseUpdate(N, DCI);
9996 case ARMISD::BUILD_VECTOR:
9997 return PerformARMBUILD_VECTORCombine(N, DCI);
9998 case ISD::INTRINSIC_VOID:
9999 case ISD::INTRINSIC_W_CHAIN:
10000 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10001 case Intrinsic::arm_neon_vld1:
10002 case Intrinsic::arm_neon_vld2:
10003 case Intrinsic::arm_neon_vld3:
10004 case Intrinsic::arm_neon_vld4:
10005 case Intrinsic::arm_neon_vld2lane:
10006 case Intrinsic::arm_neon_vld3lane:
10007 case Intrinsic::arm_neon_vld4lane:
10008 case Intrinsic::arm_neon_vst1:
10009 case Intrinsic::arm_neon_vst2:
10010 case Intrinsic::arm_neon_vst3:
10011 case Intrinsic::arm_neon_vst4:
10012 case Intrinsic::arm_neon_vst2lane:
10013 case Intrinsic::arm_neon_vst3lane:
10014 case Intrinsic::arm_neon_vst4lane:
10015 return CombineBaseUpdate(N, DCI);
10023 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10025 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10028 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
10029 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10030 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10032 switch (VT.getSimpleVT().SimpleTy) {
10038 // Unaligned access can use (for example) LRDB, LRDH, LDR
10039 if (AllowsUnaligned) {
10041 *Fast = Subtarget->hasV7Ops();
10048 // For any little-endian targets with neon, we can support unaligned ld/st
10049 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10050 // A big-endian target may also explictly support unaligned accesses
10051 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10061 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10062 unsigned AlignCheck) {
10063 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10064 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10067 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10068 unsigned DstAlign, unsigned SrcAlign,
10069 bool IsMemset, bool ZeroMemset,
10071 MachineFunction &MF) const {
10072 const Function *F = MF.getFunction();
10074 // See if we can use NEON instructions for this...
10075 if ((!IsMemset || ZeroMemset) &&
10076 Subtarget->hasNEON() &&
10077 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10078 Attribute::NoImplicitFloat)) {
10081 (memOpAlign(SrcAlign, DstAlign, 16) ||
10082 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
10084 } else if (Size >= 8 &&
10085 (memOpAlign(SrcAlign, DstAlign, 8) ||
10086 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
10091 // Lowering to i32/i16 if the size permits.
10094 else if (Size >= 2)
10097 // Let the target-independent logic figure it out.
10101 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10102 if (Val.getOpcode() != ISD::LOAD)
10105 EVT VT1 = Val.getValueType();
10106 if (!VT1.isSimple() || !VT1.isInteger() ||
10107 !VT2.isSimple() || !VT2.isInteger())
10110 switch (VT1.getSimpleVT().SimpleTy) {
10115 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10122 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10123 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10126 if (!isTypeLegal(EVT::getEVT(Ty1)))
10129 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10131 // Assuming the caller doesn't have a zeroext or signext return parameter,
10132 // truncation all the way down to i1 is valid.
10137 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10141 unsigned Scale = 1;
10142 switch (VT.getSimpleVT().SimpleTy) {
10143 default: return false;
10158 if ((V & (Scale - 1)) != 0)
10161 return V == (V & ((1LL << 5) - 1));
10164 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10165 const ARMSubtarget *Subtarget) {
10166 bool isNeg = false;
10172 switch (VT.getSimpleVT().SimpleTy) {
10173 default: return false;
10178 // + imm12 or - imm8
10180 return V == (V & ((1LL << 8) - 1));
10181 return V == (V & ((1LL << 12) - 1));
10184 // Same as ARM mode. FIXME: NEON?
10185 if (!Subtarget->hasVFP2())
10190 return V == (V & ((1LL << 8) - 1));
10194 /// isLegalAddressImmediate - Return true if the integer value can be used
10195 /// as the offset of the target addressing mode for load / store of the
10197 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10198 const ARMSubtarget *Subtarget) {
10202 if (!VT.isSimple())
10205 if (Subtarget->isThumb1Only())
10206 return isLegalT1AddressImmediate(V, VT);
10207 else if (Subtarget->isThumb2())
10208 return isLegalT2AddressImmediate(V, VT, Subtarget);
10213 switch (VT.getSimpleVT().SimpleTy) {
10214 default: return false;
10219 return V == (V & ((1LL << 12) - 1));
10222 return V == (V & ((1LL << 8) - 1));
10225 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10230 return V == (V & ((1LL << 8) - 1));
10234 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10236 int Scale = AM.Scale;
10240 switch (VT.getSimpleVT().SimpleTy) {
10241 default: return false;
10249 Scale = Scale & ~1;
10250 return Scale == 2 || Scale == 4 || Scale == 8;
10253 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10257 // Note, we allow "void" uses (basically, uses that aren't loads or
10258 // stores), because arm allows folding a scale into many arithmetic
10259 // operations. This should be made more precise and revisited later.
10261 // Allow r << imm, but the imm has to be a multiple of two.
10262 if (Scale & 1) return false;
10263 return isPowerOf2_32(Scale);
10267 /// isLegalAddressingMode - Return true if the addressing mode represented
10268 /// by AM is legal for this target, for a load/store of the specified type.
10269 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10271 EVT VT = getValueType(Ty, true);
10272 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10275 // Can never fold addr of global into load/store.
10279 switch (AM.Scale) {
10280 case 0: // no scale reg, must be "r+i" or "r", or "i".
10283 if (Subtarget->isThumb1Only())
10287 // ARM doesn't support any R+R*scale+imm addr modes.
10291 if (!VT.isSimple())
10294 if (Subtarget->isThumb2())
10295 return isLegalT2ScaledAddressingMode(AM, VT);
10297 int Scale = AM.Scale;
10298 switch (VT.getSimpleVT().SimpleTy) {
10299 default: return false;
10303 if (Scale < 0) Scale = -Scale;
10307 return isPowerOf2_32(Scale & ~1);
10311 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10316 // Note, we allow "void" uses (basically, uses that aren't loads or
10317 // stores), because arm allows folding a scale into many arithmetic
10318 // operations. This should be made more precise and revisited later.
10320 // Allow r << imm, but the imm has to be a multiple of two.
10321 if (Scale & 1) return false;
10322 return isPowerOf2_32(Scale);
10328 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10329 /// icmp immediate, that is the target has icmp instructions which can compare
10330 /// a register against the immediate without having to materialize the
10331 /// immediate into a register.
10332 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10333 // Thumb2 and ARM modes can use cmn for negative immediates.
10334 if (!Subtarget->isThumb())
10335 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10336 if (Subtarget->isThumb2())
10337 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10338 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10339 return Imm >= 0 && Imm <= 255;
10342 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10343 /// *or sub* immediate, that is the target has add or sub instructions which can
10344 /// add a register with the immediate without having to materialize the
10345 /// immediate into a register.
10346 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10347 // Same encoding for add/sub, just flip the sign.
10348 int64_t AbsImm = llvm::abs64(Imm);
10349 if (!Subtarget->isThumb())
10350 return ARM_AM::getSOImmVal(AbsImm) != -1;
10351 if (Subtarget->isThumb2())
10352 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10353 // Thumb1 only has 8-bit unsigned immediate.
10354 return AbsImm >= 0 && AbsImm <= 255;
10357 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10358 bool isSEXTLoad, SDValue &Base,
10359 SDValue &Offset, bool &isInc,
10360 SelectionDAG &DAG) {
10361 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10364 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10365 // AddressingMode 3
10366 Base = Ptr->getOperand(0);
10367 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10368 int RHSC = (int)RHS->getZExtValue();
10369 if (RHSC < 0 && RHSC > -256) {
10370 assert(Ptr->getOpcode() == ISD::ADD);
10372 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10376 isInc = (Ptr->getOpcode() == ISD::ADD);
10377 Offset = Ptr->getOperand(1);
10379 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10380 // AddressingMode 2
10381 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10382 int RHSC = (int)RHS->getZExtValue();
10383 if (RHSC < 0 && RHSC > -0x1000) {
10384 assert(Ptr->getOpcode() == ISD::ADD);
10386 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10387 Base = Ptr->getOperand(0);
10392 if (Ptr->getOpcode() == ISD::ADD) {
10394 ARM_AM::ShiftOpc ShOpcVal=
10395 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10396 if (ShOpcVal != ARM_AM::no_shift) {
10397 Base = Ptr->getOperand(1);
10398 Offset = Ptr->getOperand(0);
10400 Base = Ptr->getOperand(0);
10401 Offset = Ptr->getOperand(1);
10406 isInc = (Ptr->getOpcode() == ISD::ADD);
10407 Base = Ptr->getOperand(0);
10408 Offset = Ptr->getOperand(1);
10412 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10416 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10417 bool isSEXTLoad, SDValue &Base,
10418 SDValue &Offset, bool &isInc,
10419 SelectionDAG &DAG) {
10420 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10423 Base = Ptr->getOperand(0);
10424 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10425 int RHSC = (int)RHS->getZExtValue();
10426 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10427 assert(Ptr->getOpcode() == ISD::ADD);
10429 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10431 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10432 isInc = Ptr->getOpcode() == ISD::ADD;
10433 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10441 /// getPreIndexedAddressParts - returns true by value, base pointer and
10442 /// offset pointer and addressing mode by reference if the node's address
10443 /// can be legally represented as pre-indexed load / store address.
10445 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10447 ISD::MemIndexedMode &AM,
10448 SelectionDAG &DAG) const {
10449 if (Subtarget->isThumb1Only())
10454 bool isSEXTLoad = false;
10455 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10456 Ptr = LD->getBasePtr();
10457 VT = LD->getMemoryVT();
10458 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10459 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10460 Ptr = ST->getBasePtr();
10461 VT = ST->getMemoryVT();
10466 bool isLegal = false;
10467 if (Subtarget->isThumb2())
10468 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10469 Offset, isInc, DAG);
10471 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10472 Offset, isInc, DAG);
10476 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10480 /// getPostIndexedAddressParts - returns true by value, base pointer and
10481 /// offset pointer and addressing mode by reference if this node can be
10482 /// combined with a load / store to form a post-indexed load / store.
10483 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10486 ISD::MemIndexedMode &AM,
10487 SelectionDAG &DAG) const {
10488 if (Subtarget->isThumb1Only())
10493 bool isSEXTLoad = false;
10494 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10495 VT = LD->getMemoryVT();
10496 Ptr = LD->getBasePtr();
10497 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10498 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10499 VT = ST->getMemoryVT();
10500 Ptr = ST->getBasePtr();
10505 bool isLegal = false;
10506 if (Subtarget->isThumb2())
10507 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10510 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10516 // Swap base ptr and offset to catch more post-index load / store when
10517 // it's legal. In Thumb2 mode, offset must be an immediate.
10518 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10519 !Subtarget->isThumb2())
10520 std::swap(Base, Offset);
10522 // Post-indexed load / store update the base pointer.
10527 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10531 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10534 const SelectionDAG &DAG,
10535 unsigned Depth) const {
10536 unsigned BitWidth = KnownOne.getBitWidth();
10537 KnownZero = KnownOne = APInt(BitWidth, 0);
10538 switch (Op.getOpcode()) {
10544 // These nodes' second result is a boolean
10545 if (Op.getResNo() == 0)
10547 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10549 case ARMISD::CMOV: {
10550 // Bits are known zero/one if known on the LHS and RHS.
10551 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10552 if (KnownZero == 0 && KnownOne == 0) return;
10554 APInt KnownZeroRHS, KnownOneRHS;
10555 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10556 KnownZero &= KnownZeroRHS;
10557 KnownOne &= KnownOneRHS;
10563 //===----------------------------------------------------------------------===//
10564 // ARM Inline Assembly Support
10565 //===----------------------------------------------------------------------===//
10567 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10568 // Looking for "rev" which is V6+.
10569 if (!Subtarget->hasV6Ops())
10572 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10573 std::string AsmStr = IA->getAsmString();
10574 SmallVector<StringRef, 4> AsmPieces;
10575 SplitString(AsmStr, AsmPieces, ";\n");
10577 switch (AsmPieces.size()) {
10578 default: return false;
10580 AsmStr = AsmPieces[0];
10582 SplitString(AsmStr, AsmPieces, " \t,");
10585 if (AsmPieces.size() == 3 &&
10586 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10587 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10588 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10589 if (Ty && Ty->getBitWidth() == 32)
10590 return IntrinsicLowering::LowerToByteSwap(CI);
10598 /// getConstraintType - Given a constraint letter, return the type of
10599 /// constraint it is for this target.
10600 ARMTargetLowering::ConstraintType
10601 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10602 if (Constraint.size() == 1) {
10603 switch (Constraint[0]) {
10605 case 'l': return C_RegisterClass;
10606 case 'w': return C_RegisterClass;
10607 case 'h': return C_RegisterClass;
10608 case 'x': return C_RegisterClass;
10609 case 't': return C_RegisterClass;
10610 case 'j': return C_Other; // Constant for movw.
10611 // An address with a single base register. Due to the way we
10612 // currently handle addresses it is the same as an 'r' memory constraint.
10613 case 'Q': return C_Memory;
10615 } else if (Constraint.size() == 2) {
10616 switch (Constraint[0]) {
10618 // All 'U+' constraints are addresses.
10619 case 'U': return C_Memory;
10622 return TargetLowering::getConstraintType(Constraint);
10625 /// Examine constraint type and operand type and determine a weight value.
10626 /// This object must already have been set up with the operand type
10627 /// and the current alternative constraint selected.
10628 TargetLowering::ConstraintWeight
10629 ARMTargetLowering::getSingleConstraintMatchWeight(
10630 AsmOperandInfo &info, const char *constraint) const {
10631 ConstraintWeight weight = CW_Invalid;
10632 Value *CallOperandVal = info.CallOperandVal;
10633 // If we don't have a value, we can't do a match,
10634 // but allow it at the lowest weight.
10635 if (CallOperandVal == NULL)
10637 Type *type = CallOperandVal->getType();
10638 // Look at the constraint type.
10639 switch (*constraint) {
10641 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10644 if (type->isIntegerTy()) {
10645 if (Subtarget->isThumb())
10646 weight = CW_SpecificReg;
10648 weight = CW_Register;
10652 if (type->isFloatingPointTy())
10653 weight = CW_Register;
10659 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10661 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10663 if (Constraint.size() == 1) {
10664 // GCC ARM Constraint Letters
10665 switch (Constraint[0]) {
10666 case 'l': // Low regs or general regs.
10667 if (Subtarget->isThumb())
10668 return RCPair(0U, &ARM::tGPRRegClass);
10669 return RCPair(0U, &ARM::GPRRegClass);
10670 case 'h': // High regs or no regs.
10671 if (Subtarget->isThumb())
10672 return RCPair(0U, &ARM::hGPRRegClass);
10675 return RCPair(0U, &ARM::GPRRegClass);
10677 if (VT == MVT::f32)
10678 return RCPair(0U, &ARM::SPRRegClass);
10679 if (VT.getSizeInBits() == 64)
10680 return RCPair(0U, &ARM::DPRRegClass);
10681 if (VT.getSizeInBits() == 128)
10682 return RCPair(0U, &ARM::QPRRegClass);
10685 if (VT == MVT::f32)
10686 return RCPair(0U, &ARM::SPR_8RegClass);
10687 if (VT.getSizeInBits() == 64)
10688 return RCPair(0U, &ARM::DPR_8RegClass);
10689 if (VT.getSizeInBits() == 128)
10690 return RCPair(0U, &ARM::QPR_8RegClass);
10693 if (VT == MVT::f32)
10694 return RCPair(0U, &ARM::SPRRegClass);
10698 if (StringRef("{cc}").equals_lower(Constraint))
10699 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10701 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10704 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10705 /// vector. If it is invalid, don't add anything to Ops.
10706 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10707 std::string &Constraint,
10708 std::vector<SDValue>&Ops,
10709 SelectionDAG &DAG) const {
10710 SDValue Result(0, 0);
10712 // Currently only support length 1 constraints.
10713 if (Constraint.length() != 1) return;
10715 char ConstraintLetter = Constraint[0];
10716 switch (ConstraintLetter) {
10719 case 'I': case 'J': case 'K': case 'L':
10720 case 'M': case 'N': case 'O':
10721 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10725 int64_t CVal64 = C->getSExtValue();
10726 int CVal = (int) CVal64;
10727 // None of these constraints allow values larger than 32 bits. Check
10728 // that the value fits in an int.
10729 if (CVal != CVal64)
10732 switch (ConstraintLetter) {
10734 // Constant suitable for movw, must be between 0 and
10736 if (Subtarget->hasV6T2Ops())
10737 if (CVal >= 0 && CVal <= 65535)
10741 if (Subtarget->isThumb1Only()) {
10742 // This must be a constant between 0 and 255, for ADD
10744 if (CVal >= 0 && CVal <= 255)
10746 } else if (Subtarget->isThumb2()) {
10747 // A constant that can be used as an immediate value in a
10748 // data-processing instruction.
10749 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10752 // A constant that can be used as an immediate value in a
10753 // data-processing instruction.
10754 if (ARM_AM::getSOImmVal(CVal) != -1)
10760 if (Subtarget->isThumb()) { // FIXME thumb2
10761 // This must be a constant between -255 and -1, for negated ADD
10762 // immediates. This can be used in GCC with an "n" modifier that
10763 // prints the negated value, for use with SUB instructions. It is
10764 // not useful otherwise but is implemented for compatibility.
10765 if (CVal >= -255 && CVal <= -1)
10768 // This must be a constant between -4095 and 4095. It is not clear
10769 // what this constraint is intended for. Implemented for
10770 // compatibility with GCC.
10771 if (CVal >= -4095 && CVal <= 4095)
10777 if (Subtarget->isThumb1Only()) {
10778 // A 32-bit value where only one byte has a nonzero value. Exclude
10779 // zero to match GCC. This constraint is used by GCC internally for
10780 // constants that can be loaded with a move/shift combination.
10781 // It is not useful otherwise but is implemented for compatibility.
10782 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10784 } else if (Subtarget->isThumb2()) {
10785 // A constant whose bitwise inverse can be used as an immediate
10786 // value in a data-processing instruction. This can be used in GCC
10787 // with a "B" modifier that prints the inverted value, for use with
10788 // BIC and MVN instructions. It is not useful otherwise but is
10789 // implemented for compatibility.
10790 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10793 // A constant whose bitwise inverse can be used as an immediate
10794 // value in a data-processing instruction. This can be used in GCC
10795 // with a "B" modifier that prints the inverted value, for use with
10796 // BIC and MVN instructions. It is not useful otherwise but is
10797 // implemented for compatibility.
10798 if (ARM_AM::getSOImmVal(~CVal) != -1)
10804 if (Subtarget->isThumb1Only()) {
10805 // This must be a constant between -7 and 7,
10806 // for 3-operand ADD/SUB immediate instructions.
10807 if (CVal >= -7 && CVal < 7)
10809 } else if (Subtarget->isThumb2()) {
10810 // A constant whose negation can be used as an immediate value in a
10811 // data-processing instruction. This can be used in GCC with an "n"
10812 // modifier that prints the negated value, for use with SUB
10813 // instructions. It is not useful otherwise but is implemented for
10815 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10818 // A constant whose negation can be used as an immediate value in a
10819 // data-processing instruction. This can be used in GCC with an "n"
10820 // modifier that prints the negated value, for use with SUB
10821 // instructions. It is not useful otherwise but is implemented for
10823 if (ARM_AM::getSOImmVal(-CVal) != -1)
10829 if (Subtarget->isThumb()) { // FIXME thumb2
10830 // This must be a multiple of 4 between 0 and 1020, for
10831 // ADD sp + immediate.
10832 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10835 // A power of two or a constant between 0 and 32. This is used in
10836 // GCC for the shift amount on shifted register operands, but it is
10837 // useful in general for any shift amounts.
10838 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10844 if (Subtarget->isThumb()) { // FIXME thumb2
10845 // This must be a constant between 0 and 31, for shift amounts.
10846 if (CVal >= 0 && CVal <= 31)
10852 if (Subtarget->isThumb()) { // FIXME thumb2
10853 // This must be a multiple of 4 between -508 and 508, for
10854 // ADD/SUB sp = sp + immediate.
10855 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10860 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10864 if (Result.getNode()) {
10865 Ops.push_back(Result);
10868 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10871 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10872 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10873 unsigned Opcode = Op->getOpcode();
10874 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10875 "Invalid opcode for Div/Rem lowering");
10876 bool isSigned = (Opcode == ISD::SDIVREM);
10877 EVT VT = Op->getValueType(0);
10878 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10881 switch (VT.getSimpleVT().SimpleTy) {
10882 default: llvm_unreachable("Unexpected request for libcall!");
10883 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10884 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10885 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10886 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10889 SDValue InChain = DAG.getEntryNode();
10891 TargetLowering::ArgListTy Args;
10892 TargetLowering::ArgListEntry Entry;
10893 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10894 EVT ArgVT = Op->getOperand(i).getValueType();
10895 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10896 Entry.Node = Op->getOperand(i);
10898 Entry.isSExt = isSigned;
10899 Entry.isZExt = !isSigned;
10900 Args.push_back(Entry);
10903 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10906 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10910 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
10911 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
10912 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
10913 Callee, Args, DAG, dl);
10914 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10916 return CallInfo.first;
10920 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10921 // The ARM target isn't yet aware of offsets.
10925 bool ARM::isBitFieldInvertedMask(unsigned v) {
10926 if (v == 0xffffffff)
10929 // there can be 1's on either or both "outsides", all the "inside"
10930 // bits must be 0's
10931 unsigned TO = CountTrailingOnes_32(v);
10932 unsigned LO = CountLeadingOnes_32(v);
10933 v = (v >> TO) << TO;
10934 v = (v << LO) >> LO;
10938 /// isFPImmLegal - Returns true if the target can instruction select the
10939 /// specified FP immediate natively. If false, the legalizer will
10940 /// materialize the FP immediate as a load from a constant pool.
10941 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10942 if (!Subtarget->hasVFP3())
10944 if (VT == MVT::f32)
10945 return ARM_AM::getFP32Imm(Imm) != -1;
10946 if (VT == MVT::f64)
10947 return ARM_AM::getFP64Imm(Imm) != -1;
10951 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10952 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10953 /// specified in the intrinsic calls.
10954 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10956 unsigned Intrinsic) const {
10957 switch (Intrinsic) {
10958 case Intrinsic::arm_neon_vld1:
10959 case Intrinsic::arm_neon_vld2:
10960 case Intrinsic::arm_neon_vld3:
10961 case Intrinsic::arm_neon_vld4:
10962 case Intrinsic::arm_neon_vld2lane:
10963 case Intrinsic::arm_neon_vld3lane:
10964 case Intrinsic::arm_neon_vld4lane: {
10965 Info.opc = ISD::INTRINSIC_W_CHAIN;
10966 // Conservatively set memVT to the entire set of vectors loaded.
10967 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10968 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10969 Info.ptrVal = I.getArgOperand(0);
10971 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10972 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10973 Info.vol = false; // volatile loads with NEON intrinsics not supported
10974 Info.readMem = true;
10975 Info.writeMem = false;
10978 case Intrinsic::arm_neon_vst1:
10979 case Intrinsic::arm_neon_vst2:
10980 case Intrinsic::arm_neon_vst3:
10981 case Intrinsic::arm_neon_vst4:
10982 case Intrinsic::arm_neon_vst2lane:
10983 case Intrinsic::arm_neon_vst3lane:
10984 case Intrinsic::arm_neon_vst4lane: {
10985 Info.opc = ISD::INTRINSIC_VOID;
10986 // Conservatively set memVT to the entire set of vectors stored.
10987 unsigned NumElts = 0;
10988 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10989 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10990 if (!ArgTy->isVectorTy())
10992 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10994 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10995 Info.ptrVal = I.getArgOperand(0);
10997 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10998 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10999 Info.vol = false; // volatile stores with NEON intrinsics not supported
11000 Info.readMem = false;
11001 Info.writeMem = true;
11004 case Intrinsic::arm_ldrex: {
11005 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11006 Info.opc = ISD::INTRINSIC_W_CHAIN;
11007 Info.memVT = MVT::getVT(PtrTy->getElementType());
11008 Info.ptrVal = I.getArgOperand(0);
11010 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11012 Info.readMem = true;
11013 Info.writeMem = false;
11016 case Intrinsic::arm_strex: {
11017 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11018 Info.opc = ISD::INTRINSIC_W_CHAIN;
11019 Info.memVT = MVT::getVT(PtrTy->getElementType());
11020 Info.ptrVal = I.getArgOperand(1);
11022 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11024 Info.readMem = false;
11025 Info.writeMem = true;
11028 case Intrinsic::arm_strexd: {
11029 Info.opc = ISD::INTRINSIC_W_CHAIN;
11030 Info.memVT = MVT::i64;
11031 Info.ptrVal = I.getArgOperand(2);
11035 Info.readMem = false;
11036 Info.writeMem = true;
11039 case Intrinsic::arm_ldrexd: {
11040 Info.opc = ISD::INTRINSIC_W_CHAIN;
11041 Info.memVT = MVT::i64;
11042 Info.ptrVal = I.getArgOperand(0);
11046 Info.readMem = true;
11047 Info.writeMem = false;