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().isiOS() &&
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 // On v8, we have particularly efficient implementations of atomic fences
773 // if they can be combined with nearby atomic loads and stores.
774 if (!Subtarget->hasV8Ops()) {
775 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
776 setInsertFencesForAtomic(true);
778 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
779 //setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Custom);
781 // Set them all for expansion, which will force libcalls.
782 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
783 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
784 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
785 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
786 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
787 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
788 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
789 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
790 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
791 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
792 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
793 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
794 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
795 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
796 // Unordered/Monotonic case.
797 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
798 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
801 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
803 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
804 if (!Subtarget->hasV6Ops()) {
805 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
806 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
808 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
810 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
811 !Subtarget->isThumb1Only()) {
812 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
813 // iff target supports vfp2.
814 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
815 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
818 // We want to custom lower some of our intrinsics.
819 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
820 if (Subtarget->isTargetDarwin()) {
821 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
822 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
823 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
826 setOperationAction(ISD::SETCC, MVT::i32, Expand);
827 setOperationAction(ISD::SETCC, MVT::f32, Expand);
828 setOperationAction(ISD::SETCC, MVT::f64, Expand);
829 setOperationAction(ISD::SELECT, MVT::i32, Custom);
830 setOperationAction(ISD::SELECT, MVT::f32, Custom);
831 setOperationAction(ISD::SELECT, MVT::f64, Custom);
832 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
833 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
834 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
836 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
837 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
838 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
839 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
840 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
842 // We don't support sin/cos/fmod/copysign/pow
843 setOperationAction(ISD::FSIN, MVT::f64, Expand);
844 setOperationAction(ISD::FSIN, MVT::f32, Expand);
845 setOperationAction(ISD::FCOS, MVT::f32, Expand);
846 setOperationAction(ISD::FCOS, MVT::f64, Expand);
847 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
848 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
849 setOperationAction(ISD::FREM, MVT::f64, Expand);
850 setOperationAction(ISD::FREM, MVT::f32, Expand);
851 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
852 !Subtarget->isThumb1Only()) {
853 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
854 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
856 setOperationAction(ISD::FPOW, MVT::f64, Expand);
857 setOperationAction(ISD::FPOW, MVT::f32, Expand);
859 if (!Subtarget->hasVFP4()) {
860 setOperationAction(ISD::FMA, MVT::f64, Expand);
861 setOperationAction(ISD::FMA, MVT::f32, Expand);
864 // Various VFP goodness
865 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
866 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
867 if (Subtarget->hasVFP2()) {
868 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
869 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
870 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
871 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
873 // Special handling for half-precision FP.
874 if (!Subtarget->hasFP16()) {
875 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
876 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
880 // We have target-specific dag combine patterns for the following nodes:
881 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
882 setTargetDAGCombine(ISD::ADD);
883 setTargetDAGCombine(ISD::SUB);
884 setTargetDAGCombine(ISD::MUL);
885 setTargetDAGCombine(ISD::AND);
886 setTargetDAGCombine(ISD::OR);
887 setTargetDAGCombine(ISD::XOR);
889 if (Subtarget->hasV6Ops())
890 setTargetDAGCombine(ISD::SRL);
892 setStackPointerRegisterToSaveRestore(ARM::SP);
894 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
895 !Subtarget->hasVFP2())
896 setSchedulingPreference(Sched::RegPressure);
898 setSchedulingPreference(Sched::Hybrid);
900 //// temporary - rewrite interface to use type
901 MaxStoresPerMemset = 8;
902 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
903 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
904 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
905 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
906 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
908 // On ARM arguments smaller than 4 bytes are extended, so all arguments
909 // are at least 4 bytes aligned.
910 setMinStackArgumentAlignment(4);
912 // Prefer likely predicted branches to selects on out-of-order cores.
913 PredictableSelectIsExpensive = Subtarget->isLikeA9();
915 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
918 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
919 bool isThumb2, unsigned &LdrOpc,
921 static const unsigned LoadBares[4][2] = {{ARM::LDREXB, ARM::t2LDREXB},
922 {ARM::LDREXH, ARM::t2LDREXH},
923 {ARM::LDREX, ARM::t2LDREX},
924 {ARM::LDREXD, ARM::t2LDREXD}};
925 static const unsigned LoadAcqs[4][2] = {{ARM::LDAEXB, ARM::t2LDAEXB},
926 {ARM::LDAEXH, ARM::t2LDAEXH},
927 {ARM::LDAEX, ARM::t2LDAEX},
928 {ARM::LDAEXD, ARM::t2LDAEXD}};
929 static const unsigned StoreBares[4][2] = {{ARM::STREXB, ARM::t2STREXB},
930 {ARM::STREXH, ARM::t2STREXH},
931 {ARM::STREX, ARM::t2STREX},
932 {ARM::STREXD, ARM::t2STREXD}};
933 static const unsigned StoreRels[4][2] = {{ARM::STLEXB, ARM::t2STLEXB},
934 {ARM::STLEXH, ARM::t2STLEXH},
935 {ARM::STLEX, ARM::t2STLEX},
936 {ARM::STLEXD, ARM::t2STLEXD}};
938 const unsigned (*LoadOps)[2], (*StoreOps)[2];
939 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
944 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
945 StoreOps = StoreRels;
947 StoreOps = StoreBares;
949 assert(isPowerOf2_32(Size) && Size <= 8 &&
950 "unsupported size for atomic binary op!");
952 LdrOpc = LoadOps[Log2_32(Size)][isThumb2];
953 StrOpc = StoreOps[Log2_32(Size)][isThumb2];
956 // FIXME: It might make sense to define the representative register class as the
957 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
958 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
959 // SPR's representative would be DPR_VFP2. This should work well if register
960 // pressure tracking were modified such that a register use would increment the
961 // pressure of the register class's representative and all of it's super
962 // classes' representatives transitively. We have not implemented this because
963 // of the difficulty prior to coalescing of modeling operand register classes
964 // due to the common occurrence of cross class copies and subregister insertions
966 std::pair<const TargetRegisterClass*, uint8_t>
967 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
968 const TargetRegisterClass *RRC = 0;
970 switch (VT.SimpleTy) {
972 return TargetLowering::findRepresentativeClass(VT);
973 // Use DPR as representative register class for all floating point
974 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
975 // the cost is 1 for both f32 and f64.
976 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
977 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
978 RRC = &ARM::DPRRegClass;
979 // When NEON is used for SP, only half of the register file is available
980 // because operations that define both SP and DP results will be constrained
981 // to the VFP2 class (D0-D15). We currently model this constraint prior to
982 // coalescing by double-counting the SP regs. See the FIXME above.
983 if (Subtarget->useNEONForSinglePrecisionFP())
986 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
987 case MVT::v4f32: case MVT::v2f64:
988 RRC = &ARM::DPRRegClass;
992 RRC = &ARM::DPRRegClass;
996 RRC = &ARM::DPRRegClass;
1000 return std::make_pair(RRC, Cost);
1003 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1006 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1007 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
1008 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1009 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1010 case ARMISD::CALL: return "ARMISD::CALL";
1011 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1012 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1013 case ARMISD::tCALL: return "ARMISD::tCALL";
1014 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1015 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1016 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1017 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1018 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1019 case ARMISD::CMP: return "ARMISD::CMP";
1020 case ARMISD::CMN: return "ARMISD::CMN";
1021 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1022 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1023 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1024 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1025 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1027 case ARMISD::CMOV: return "ARMISD::CMOV";
1029 case ARMISD::RBIT: return "ARMISD::RBIT";
1031 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1032 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1033 case ARMISD::SITOF: return "ARMISD::SITOF";
1034 case ARMISD::UITOF: return "ARMISD::UITOF";
1036 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1037 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1038 case ARMISD::RRX: return "ARMISD::RRX";
1040 case ARMISD::ADDC: return "ARMISD::ADDC";
1041 case ARMISD::ADDE: return "ARMISD::ADDE";
1042 case ARMISD::SUBC: return "ARMISD::SUBC";
1043 case ARMISD::SUBE: return "ARMISD::SUBE";
1045 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1046 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1048 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1049 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1051 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1053 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1055 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1057 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1059 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1061 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1062 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1063 case ARMISD::VCGE: return "ARMISD::VCGE";
1064 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1065 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1066 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1067 case ARMISD::VCGT: return "ARMISD::VCGT";
1068 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1069 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1070 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1071 case ARMISD::VTST: return "ARMISD::VTST";
1073 case ARMISD::VSHL: return "ARMISD::VSHL";
1074 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1075 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1076 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1077 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1078 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1079 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1080 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1081 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1082 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1083 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1084 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1085 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1086 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1087 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1088 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1089 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1090 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1091 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1092 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1093 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1094 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1095 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1096 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1097 case ARMISD::VDUP: return "ARMISD::VDUP";
1098 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1099 case ARMISD::VEXT: return "ARMISD::VEXT";
1100 case ARMISD::VREV64: return "ARMISD::VREV64";
1101 case ARMISD::VREV32: return "ARMISD::VREV32";
1102 case ARMISD::VREV16: return "ARMISD::VREV16";
1103 case ARMISD::VZIP: return "ARMISD::VZIP";
1104 case ARMISD::VUZP: return "ARMISD::VUZP";
1105 case ARMISD::VTRN: return "ARMISD::VTRN";
1106 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1107 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1108 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1109 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1110 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1111 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1112 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1113 case ARMISD::FMAX: return "ARMISD::FMAX";
1114 case ARMISD::FMIN: return "ARMISD::FMIN";
1115 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1116 case ARMISD::VMINNM: return "ARMISD::VMIN";
1117 case ARMISD::BFI: return "ARMISD::BFI";
1118 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1119 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1120 case ARMISD::VBSL: return "ARMISD::VBSL";
1121 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1122 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1123 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1124 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1125 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1126 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1127 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1128 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1129 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1130 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1131 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1132 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1133 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1134 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1135 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1136 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1137 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1138 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1139 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1140 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1144 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1145 if (!VT.isVector()) return getPointerTy();
1146 return VT.changeVectorElementTypeToInteger();
1149 /// getRegClassFor - Return the register class that should be used for the
1150 /// specified value type.
1151 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1152 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1153 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1154 // load / store 4 to 8 consecutive D registers.
1155 if (Subtarget->hasNEON()) {
1156 if (VT == MVT::v4i64)
1157 return &ARM::QQPRRegClass;
1158 if (VT == MVT::v8i64)
1159 return &ARM::QQQQPRRegClass;
1161 return TargetLowering::getRegClassFor(VT);
1164 // Create a fast isel object.
1166 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1167 const TargetLibraryInfo *libInfo) const {
1168 return ARM::createFastISel(funcInfo, libInfo);
1171 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1172 /// be used for loads / stores from the global.
1173 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1174 return (Subtarget->isThumb1Only() ? 127 : 4095);
1177 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1178 unsigned NumVals = N->getNumValues();
1180 return Sched::RegPressure;
1182 for (unsigned i = 0; i != NumVals; ++i) {
1183 EVT VT = N->getValueType(i);
1184 if (VT == MVT::Glue || VT == MVT::Other)
1186 if (VT.isFloatingPoint() || VT.isVector())
1190 if (!N->isMachineOpcode())
1191 return Sched::RegPressure;
1193 // Load are scheduled for latency even if there instruction itinerary
1194 // is not available.
1195 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1196 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1198 if (MCID.getNumDefs() == 0)
1199 return Sched::RegPressure;
1200 if (!Itins->isEmpty() &&
1201 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1204 return Sched::RegPressure;
1207 //===----------------------------------------------------------------------===//
1209 //===----------------------------------------------------------------------===//
1211 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1212 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1214 default: llvm_unreachable("Unknown condition code!");
1215 case ISD::SETNE: return ARMCC::NE;
1216 case ISD::SETEQ: return ARMCC::EQ;
1217 case ISD::SETGT: return ARMCC::GT;
1218 case ISD::SETGE: return ARMCC::GE;
1219 case ISD::SETLT: return ARMCC::LT;
1220 case ISD::SETLE: return ARMCC::LE;
1221 case ISD::SETUGT: return ARMCC::HI;
1222 case ISD::SETUGE: return ARMCC::HS;
1223 case ISD::SETULT: return ARMCC::LO;
1224 case ISD::SETULE: return ARMCC::LS;
1228 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1229 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1230 ARMCC::CondCodes &CondCode2) {
1231 CondCode2 = ARMCC::AL;
1233 default: llvm_unreachable("Unknown FP condition!");
1235 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1237 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1239 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1240 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1241 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1242 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1243 case ISD::SETO: CondCode = ARMCC::VC; break;
1244 case ISD::SETUO: CondCode = ARMCC::VS; break;
1245 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1246 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1247 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1249 case ISD::SETULT: CondCode = ARMCC::LT; break;
1251 case ISD::SETULE: CondCode = ARMCC::LE; break;
1253 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1257 //===----------------------------------------------------------------------===//
1258 // Calling Convention Implementation
1259 //===----------------------------------------------------------------------===//
1261 #include "ARMGenCallingConv.inc"
1263 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1264 /// given CallingConvention value.
1265 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1267 bool isVarArg) const {
1270 llvm_unreachable("Unsupported calling convention");
1271 case CallingConv::Fast:
1272 if (Subtarget->hasVFP2() && !isVarArg) {
1273 if (!Subtarget->isAAPCS_ABI())
1274 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1275 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1276 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1279 case CallingConv::C: {
1280 // Use target triple & subtarget features to do actual dispatch.
1281 if (!Subtarget->isAAPCS_ABI())
1282 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1283 else if (Subtarget->hasVFP2() &&
1284 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1286 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1287 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1289 case CallingConv::ARM_AAPCS_VFP:
1291 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1293 case CallingConv::ARM_AAPCS:
1294 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1295 case CallingConv::ARM_APCS:
1296 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1297 case CallingConv::GHC:
1298 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1302 /// LowerCallResult - Lower the result values of a call into the
1303 /// appropriate copies out of appropriate physical registers.
1305 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1306 CallingConv::ID CallConv, bool isVarArg,
1307 const SmallVectorImpl<ISD::InputArg> &Ins,
1308 SDLoc dl, SelectionDAG &DAG,
1309 SmallVectorImpl<SDValue> &InVals,
1310 bool isThisReturn, SDValue ThisVal) const {
1312 // Assign locations to each value returned by this call.
1313 SmallVector<CCValAssign, 16> RVLocs;
1314 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1315 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1316 CCInfo.AnalyzeCallResult(Ins,
1317 CCAssignFnForNode(CallConv, /* Return*/ true,
1320 // Copy all of the result registers out of their specified physreg.
1321 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1322 CCValAssign VA = RVLocs[i];
1324 // Pass 'this' value directly from the argument to return value, to avoid
1325 // reg unit interference
1326 if (i == 0 && isThisReturn) {
1327 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1328 "unexpected return calling convention register assignment");
1329 InVals.push_back(ThisVal);
1334 if (VA.needsCustom()) {
1335 // Handle f64 or half of a v2f64.
1336 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1338 Chain = Lo.getValue(1);
1339 InFlag = Lo.getValue(2);
1340 VA = RVLocs[++i]; // skip ahead to next loc
1341 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1343 Chain = Hi.getValue(1);
1344 InFlag = Hi.getValue(2);
1345 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1347 if (VA.getLocVT() == MVT::v2f64) {
1348 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1349 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1350 DAG.getConstant(0, MVT::i32));
1352 VA = RVLocs[++i]; // skip ahead to next loc
1353 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1354 Chain = Lo.getValue(1);
1355 InFlag = Lo.getValue(2);
1356 VA = RVLocs[++i]; // skip ahead to next loc
1357 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1358 Chain = Hi.getValue(1);
1359 InFlag = Hi.getValue(2);
1360 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1361 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1362 DAG.getConstant(1, MVT::i32));
1365 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1367 Chain = Val.getValue(1);
1368 InFlag = Val.getValue(2);
1371 switch (VA.getLocInfo()) {
1372 default: llvm_unreachable("Unknown loc info!");
1373 case CCValAssign::Full: break;
1374 case CCValAssign::BCvt:
1375 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1379 InVals.push_back(Val);
1385 /// LowerMemOpCallTo - Store the argument to the stack.
1387 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1388 SDValue StackPtr, SDValue Arg,
1389 SDLoc dl, SelectionDAG &DAG,
1390 const CCValAssign &VA,
1391 ISD::ArgFlagsTy Flags) const {
1392 unsigned LocMemOffset = VA.getLocMemOffset();
1393 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1394 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1395 return DAG.getStore(Chain, dl, Arg, PtrOff,
1396 MachinePointerInfo::getStack(LocMemOffset),
1400 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1401 SDValue Chain, SDValue &Arg,
1402 RegsToPassVector &RegsToPass,
1403 CCValAssign &VA, CCValAssign &NextVA,
1405 SmallVectorImpl<SDValue> &MemOpChains,
1406 ISD::ArgFlagsTy Flags) const {
1408 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1409 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1410 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1412 if (NextVA.isRegLoc())
1413 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1415 assert(NextVA.isMemLoc());
1416 if (StackPtr.getNode() == 0)
1417 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1419 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1425 /// LowerCall - Lowering a call into a callseq_start <-
1426 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1429 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1430 SmallVectorImpl<SDValue> &InVals) const {
1431 SelectionDAG &DAG = CLI.DAG;
1433 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1434 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1435 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1436 SDValue Chain = CLI.Chain;
1437 SDValue Callee = CLI.Callee;
1438 bool &isTailCall = CLI.IsTailCall;
1439 CallingConv::ID CallConv = CLI.CallConv;
1440 bool doesNotRet = CLI.DoesNotReturn;
1441 bool isVarArg = CLI.IsVarArg;
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1445 bool isThisReturn = false;
1446 bool isSibCall = false;
1447 // Disable tail calls if they're not supported.
1448 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1451 // Check if it's really possible to do a tail call.
1452 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1453 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1454 Outs, OutVals, Ins, DAG);
1455 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1456 // detected sibcalls.
1463 // Analyze operands of the call, assigning locations to each operand.
1464 SmallVector<CCValAssign, 16> ArgLocs;
1465 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1466 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1467 CCInfo.AnalyzeCallOperands(Outs,
1468 CCAssignFnForNode(CallConv, /* Return*/ false,
1471 // Get a count of how many bytes are to be pushed on the stack.
1472 unsigned NumBytes = CCInfo.getNextStackOffset();
1474 // For tail calls, memory operands are available in our caller's stack.
1478 // Adjust the stack pointer for the new arguments...
1479 // These operations are automatically eliminated by the prolog/epilog pass
1481 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1484 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1486 RegsToPassVector RegsToPass;
1487 SmallVector<SDValue, 8> MemOpChains;
1489 // Walk the register/memloc assignments, inserting copies/loads. In the case
1490 // of tail call optimization, arguments are handled later.
1491 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1493 ++i, ++realArgIdx) {
1494 CCValAssign &VA = ArgLocs[i];
1495 SDValue Arg = OutVals[realArgIdx];
1496 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1497 bool isByVal = Flags.isByVal();
1499 // Promote the value if needed.
1500 switch (VA.getLocInfo()) {
1501 default: llvm_unreachable("Unknown loc info!");
1502 case CCValAssign::Full: break;
1503 case CCValAssign::SExt:
1504 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1506 case CCValAssign::ZExt:
1507 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1509 case CCValAssign::AExt:
1510 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1512 case CCValAssign::BCvt:
1513 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1517 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1518 if (VA.needsCustom()) {
1519 if (VA.getLocVT() == MVT::v2f64) {
1520 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1521 DAG.getConstant(0, MVT::i32));
1522 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1523 DAG.getConstant(1, MVT::i32));
1525 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1526 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1528 VA = ArgLocs[++i]; // skip ahead to next loc
1529 if (VA.isRegLoc()) {
1530 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1531 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1533 assert(VA.isMemLoc());
1535 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1536 dl, DAG, VA, Flags));
1539 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1540 StackPtr, MemOpChains, Flags);
1542 } else if (VA.isRegLoc()) {
1543 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1544 assert(VA.getLocVT() == MVT::i32 &&
1545 "unexpected calling convention register assignment");
1546 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1547 "unexpected use of 'returned'");
1548 isThisReturn = true;
1550 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1551 } else if (isByVal) {
1552 assert(VA.isMemLoc());
1553 unsigned offset = 0;
1555 // True if this byval aggregate will be split between registers
1557 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1558 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1560 if (CurByValIdx < ByValArgsCount) {
1562 unsigned RegBegin, RegEnd;
1563 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1565 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1567 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1568 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1569 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1570 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1571 MachinePointerInfo(),
1572 false, false, false, 0);
1573 MemOpChains.push_back(Load.getValue(1));
1574 RegsToPass.push_back(std::make_pair(j, Load));
1577 // If parameter size outsides register area, "offset" value
1578 // helps us to calculate stack slot for remained part properly.
1579 offset = RegEnd - RegBegin;
1581 CCInfo.nextInRegsParam();
1584 if (Flags.getByValSize() > 4*offset) {
1585 unsigned LocMemOffset = VA.getLocMemOffset();
1586 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1587 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1589 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1590 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1591 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1593 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1595 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1596 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1597 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1598 Ops, array_lengthof(Ops)));
1600 } else if (!isSibCall) {
1601 assert(VA.isMemLoc());
1603 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1604 dl, DAG, VA, Flags));
1608 if (!MemOpChains.empty())
1609 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1610 &MemOpChains[0], MemOpChains.size());
1612 // Build a sequence of copy-to-reg nodes chained together with token chain
1613 // and flag operands which copy the outgoing args into the appropriate regs.
1615 // Tail call byval lowering might overwrite argument registers so in case of
1616 // tail call optimization the copies to registers are lowered later.
1618 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1619 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1620 RegsToPass[i].second, InFlag);
1621 InFlag = Chain.getValue(1);
1624 // For tail calls lower the arguments to the 'real' stack slot.
1626 // Force all the incoming stack arguments to be loaded from the stack
1627 // before any new outgoing arguments are stored to the stack, because the
1628 // outgoing stack slots may alias the incoming argument stack slots, and
1629 // the alias isn't otherwise explicit. This is slightly more conservative
1630 // than necessary, because it means that each store effectively depends
1631 // on every argument instead of just those arguments it would clobber.
1633 // Do not flag preceding copytoreg stuff together with the following stuff.
1635 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1636 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1637 RegsToPass[i].second, InFlag);
1638 InFlag = Chain.getValue(1);
1643 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1644 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1645 // node so that legalize doesn't hack it.
1646 bool isDirect = false;
1647 bool isARMFunc = false;
1648 bool isLocalARMFunc = false;
1649 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1651 if (EnableARMLongCalls) {
1652 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1653 && "long-calls with non-static relocation model!");
1654 // Handle a global address or an external symbol. If it's not one of
1655 // those, the target's already in a register, so we don't need to do
1657 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1658 const GlobalValue *GV = G->getGlobal();
1659 // Create a constant pool entry for the callee address
1660 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1661 ARMConstantPoolValue *CPV =
1662 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1664 // Get the address of the callee into a register
1665 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1666 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1667 Callee = DAG.getLoad(getPointerTy(), dl,
1668 DAG.getEntryNode(), CPAddr,
1669 MachinePointerInfo::getConstantPool(),
1670 false, false, false, 0);
1671 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1672 const char *Sym = S->getSymbol();
1674 // Create a constant pool entry for the callee address
1675 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1676 ARMConstantPoolValue *CPV =
1677 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1678 ARMPCLabelIndex, 0);
1679 // Get the address of the callee into a register
1680 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1681 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1682 Callee = DAG.getLoad(getPointerTy(), dl,
1683 DAG.getEntryNode(), CPAddr,
1684 MachinePointerInfo::getConstantPool(),
1685 false, false, false, 0);
1687 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1688 const GlobalValue *GV = G->getGlobal();
1690 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1691 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1692 getTargetMachine().getRelocationModel() != Reloc::Static;
1693 isARMFunc = !Subtarget->isThumb() || isStub;
1694 // ARM call to a local ARM function is predicable.
1695 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1696 // tBX takes a register source operand.
1697 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1698 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1699 ARMConstantPoolValue *CPV =
1700 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1701 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1702 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1703 Callee = DAG.getLoad(getPointerTy(), dl,
1704 DAG.getEntryNode(), CPAddr,
1705 MachinePointerInfo::getConstantPool(),
1706 false, false, false, 0);
1707 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1708 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1709 getPointerTy(), Callee, PICLabel);
1711 // On ELF targets for PIC code, direct calls should go through the PLT
1712 unsigned OpFlags = 0;
1713 if (Subtarget->isTargetELF() &&
1714 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1715 OpFlags = ARMII::MO_PLT;
1716 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1718 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1720 bool isStub = Subtarget->isTargetDarwin() &&
1721 getTargetMachine().getRelocationModel() != Reloc::Static;
1722 isARMFunc = !Subtarget->isThumb() || isStub;
1723 // tBX takes a register source operand.
1724 const char *Sym = S->getSymbol();
1725 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1726 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1727 ARMConstantPoolValue *CPV =
1728 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1729 ARMPCLabelIndex, 4);
1730 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1731 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1732 Callee = DAG.getLoad(getPointerTy(), dl,
1733 DAG.getEntryNode(), CPAddr,
1734 MachinePointerInfo::getConstantPool(),
1735 false, false, false, 0);
1736 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1737 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1738 getPointerTy(), Callee, PICLabel);
1740 unsigned OpFlags = 0;
1741 // On ELF targets for PIC code, direct calls should go through the PLT
1742 if (Subtarget->isTargetELF() &&
1743 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1744 OpFlags = ARMII::MO_PLT;
1745 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1749 // FIXME: handle tail calls differently.
1751 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1752 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1753 if (Subtarget->isThumb()) {
1754 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1755 CallOpc = ARMISD::CALL_NOLINK;
1757 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1759 if (!isDirect && !Subtarget->hasV5TOps())
1760 CallOpc = ARMISD::CALL_NOLINK;
1761 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1762 // Emit regular call when code size is the priority
1764 // "mov lr, pc; b _foo" to avoid confusing the RSP
1765 CallOpc = ARMISD::CALL_NOLINK;
1767 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1770 std::vector<SDValue> Ops;
1771 Ops.push_back(Chain);
1772 Ops.push_back(Callee);
1774 // Add argument registers to the end of the list so that they are known live
1776 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1777 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1778 RegsToPass[i].second.getValueType()));
1780 // Add a register mask operand representing the call-preserved registers.
1781 const uint32_t *Mask;
1782 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1783 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1785 // For 'this' returns, use the R0-preserving mask if applicable
1786 Mask = ARI->getThisReturnPreservedMask(CallConv);
1788 // Set isThisReturn to false if the calling convention is not one that
1789 // allows 'returned' to be modeled in this way, so LowerCallResult does
1790 // not try to pass 'this' straight through
1791 isThisReturn = false;
1792 Mask = ARI->getCallPreservedMask(CallConv);
1795 Mask = ARI->getCallPreservedMask(CallConv);
1797 assert(Mask && "Missing call preserved mask for calling convention");
1798 Ops.push_back(DAG.getRegisterMask(Mask));
1800 if (InFlag.getNode())
1801 Ops.push_back(InFlag);
1803 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1805 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1807 // Returns a chain and a flag for retval copy to use.
1808 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1809 InFlag = Chain.getValue(1);
1811 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1812 DAG.getIntPtrConstant(0, true), InFlag, dl);
1814 InFlag = Chain.getValue(1);
1816 // Handle result values, copying them out of physregs into vregs that we
1818 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1819 InVals, isThisReturn,
1820 isThisReturn ? OutVals[0] : SDValue());
1823 /// HandleByVal - Every parameter *after* a byval parameter is passed
1824 /// on the stack. Remember the next parameter register to allocate,
1825 /// and then confiscate the rest of the parameter registers to insure
1828 ARMTargetLowering::HandleByVal(
1829 CCState *State, unsigned &size, unsigned Align) const {
1830 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1831 assert((State->getCallOrPrologue() == Prologue ||
1832 State->getCallOrPrologue() == Call) &&
1833 "unhandled ParmContext");
1835 // For in-prologue parameters handling, we also introduce stack offset
1836 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
1837 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
1838 // NSAA should be evaluted (NSAA means "next stacked argument address").
1839 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
1840 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
1841 unsigned NSAAOffset = State->getNextStackOffset();
1842 if (State->getCallOrPrologue() != Call) {
1843 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
1845 State->getInRegsParamInfo(i, RB, RE);
1846 assert(NSAAOffset >= (RE-RB)*4 &&
1847 "Stack offset for byval regs doesn't introduced anymore?");
1848 NSAAOffset -= (RE-RB)*4;
1851 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1852 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1853 unsigned AlignInRegs = Align / 4;
1854 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1855 for (unsigned i = 0; i < Waste; ++i)
1856 reg = State->AllocateReg(GPRArgRegs, 4);
1859 unsigned excess = 4 * (ARM::R4 - reg);
1861 // Special case when NSAA != SP and parameter size greater than size of
1862 // all remained GPR regs. In that case we can't split parameter, we must
1863 // send it to stack. We also must set NCRN to R4, so waste all
1864 // remained registers.
1865 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1866 while (State->AllocateReg(GPRArgRegs, 4))
1871 // First register for byval parameter is the first register that wasn't
1872 // allocated before this method call, so it would be "reg".
1873 // If parameter is small enough to be saved in range [reg, r4), then
1874 // the end (first after last) register would be reg + param-size-in-regs,
1875 // else parameter would be splitted between registers and stack,
1876 // end register would be r4 in this case.
1877 unsigned ByValRegBegin = reg;
1878 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1879 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1880 // Note, first register is allocated in the beginning of function already,
1881 // allocate remained amount of registers we need.
1882 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1883 State->AllocateReg(GPRArgRegs, 4);
1884 // At a call site, a byval parameter that is split between
1885 // registers and memory needs its size truncated here. In a
1886 // function prologue, such byval parameters are reassembled in
1887 // memory, and are not truncated.
1888 if (State->getCallOrPrologue() == Call) {
1889 // Make remained size equal to 0 in case, when
1890 // the whole structure may be stored into registers.
1900 /// MatchingStackOffset - Return true if the given stack call argument is
1901 /// already available in the same position (relatively) of the caller's
1902 /// incoming argument stack.
1904 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1905 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1906 const TargetInstrInfo *TII) {
1907 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1909 if (Arg.getOpcode() == ISD::CopyFromReg) {
1910 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1911 if (!TargetRegisterInfo::isVirtualRegister(VR))
1913 MachineInstr *Def = MRI->getVRegDef(VR);
1916 if (!Flags.isByVal()) {
1917 if (!TII->isLoadFromStackSlot(Def, FI))
1922 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1923 if (Flags.isByVal())
1924 // ByVal argument is passed in as a pointer but it's now being
1925 // dereferenced. e.g.
1926 // define @foo(%struct.X* %A) {
1927 // tail call @bar(%struct.X* byval %A)
1930 SDValue Ptr = Ld->getBasePtr();
1931 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1934 FI = FINode->getIndex();
1938 assert(FI != INT_MAX);
1939 if (!MFI->isFixedObjectIndex(FI))
1941 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1944 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1945 /// for tail call optimization. Targets which want to do tail call
1946 /// optimization should implement this function.
1948 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1949 CallingConv::ID CalleeCC,
1951 bool isCalleeStructRet,
1952 bool isCallerStructRet,
1953 const SmallVectorImpl<ISD::OutputArg> &Outs,
1954 const SmallVectorImpl<SDValue> &OutVals,
1955 const SmallVectorImpl<ISD::InputArg> &Ins,
1956 SelectionDAG& DAG) const {
1957 const Function *CallerF = DAG.getMachineFunction().getFunction();
1958 CallingConv::ID CallerCC = CallerF->getCallingConv();
1959 bool CCMatch = CallerCC == CalleeCC;
1961 // Look for obvious safe cases to perform tail call optimization that do not
1962 // require ABI changes. This is what gcc calls sibcall.
1964 // Do not sibcall optimize vararg calls unless the call site is not passing
1966 if (isVarArg && !Outs.empty())
1969 // Also avoid sibcall optimization if either caller or callee uses struct
1970 // return semantics.
1971 if (isCalleeStructRet || isCallerStructRet)
1974 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1975 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1976 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1977 // support in the assembler and linker to be used. This would need to be
1978 // fixed to fully support tail calls in Thumb1.
1980 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1981 // LR. This means if we need to reload LR, it takes an extra instructions,
1982 // which outweighs the value of the tail call; but here we don't know yet
1983 // whether LR is going to be used. Probably the right approach is to
1984 // generate the tail call here and turn it back into CALL/RET in
1985 // emitEpilogue if LR is used.
1987 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1988 // but we need to make sure there are enough registers; the only valid
1989 // registers are the 4 used for parameters. We don't currently do this
1991 if (Subtarget->isThumb1Only())
1994 // If the calling conventions do not match, then we'd better make sure the
1995 // results are returned in the same way as what the caller expects.
1997 SmallVector<CCValAssign, 16> RVLocs1;
1998 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1999 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
2000 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2002 SmallVector<CCValAssign, 16> RVLocs2;
2003 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2004 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
2005 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2007 if (RVLocs1.size() != RVLocs2.size())
2009 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2010 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2012 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2014 if (RVLocs1[i].isRegLoc()) {
2015 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2018 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2024 // If Caller's vararg or byval argument has been split between registers and
2025 // stack, do not perform tail call, since part of the argument is in caller's
2027 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2028 getInfo<ARMFunctionInfo>();
2029 if (AFI_Caller->getArgRegsSaveSize())
2032 // If the callee takes no arguments then go on to check the results of the
2034 if (!Outs.empty()) {
2035 // Check if stack adjustment is needed. For now, do not do this if any
2036 // argument is passed on the stack.
2037 SmallVector<CCValAssign, 16> ArgLocs;
2038 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2039 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2040 CCInfo.AnalyzeCallOperands(Outs,
2041 CCAssignFnForNode(CalleeCC, false, isVarArg));
2042 if (CCInfo.getNextStackOffset()) {
2043 MachineFunction &MF = DAG.getMachineFunction();
2045 // Check if the arguments are already laid out in the right way as
2046 // the caller's fixed stack objects.
2047 MachineFrameInfo *MFI = MF.getFrameInfo();
2048 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2049 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2050 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2052 ++i, ++realArgIdx) {
2053 CCValAssign &VA = ArgLocs[i];
2054 EVT RegVT = VA.getLocVT();
2055 SDValue Arg = OutVals[realArgIdx];
2056 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2057 if (VA.getLocInfo() == CCValAssign::Indirect)
2059 if (VA.needsCustom()) {
2060 // f64 and vector types are split into multiple registers or
2061 // register/stack-slot combinations. The types will not match
2062 // the registers; give up on memory f64 refs until we figure
2063 // out what to do about this.
2066 if (!ArgLocs[++i].isRegLoc())
2068 if (RegVT == MVT::v2f64) {
2069 if (!ArgLocs[++i].isRegLoc())
2071 if (!ArgLocs[++i].isRegLoc())
2074 } else if (!VA.isRegLoc()) {
2075 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2087 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2088 MachineFunction &MF, bool isVarArg,
2089 const SmallVectorImpl<ISD::OutputArg> &Outs,
2090 LLVMContext &Context) const {
2091 SmallVector<CCValAssign, 16> RVLocs;
2092 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2093 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2098 ARMTargetLowering::LowerReturn(SDValue Chain,
2099 CallingConv::ID CallConv, bool isVarArg,
2100 const SmallVectorImpl<ISD::OutputArg> &Outs,
2101 const SmallVectorImpl<SDValue> &OutVals,
2102 SDLoc dl, SelectionDAG &DAG) const {
2104 // CCValAssign - represent the assignment of the return value to a location.
2105 SmallVector<CCValAssign, 16> RVLocs;
2107 // CCState - Info about the registers and stack slots.
2108 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2109 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2111 // Analyze outgoing return values.
2112 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2116 SmallVector<SDValue, 4> RetOps;
2117 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2119 // Copy the result values into the output registers.
2120 for (unsigned i = 0, realRVLocIdx = 0;
2122 ++i, ++realRVLocIdx) {
2123 CCValAssign &VA = RVLocs[i];
2124 assert(VA.isRegLoc() && "Can only return in registers!");
2126 SDValue Arg = OutVals[realRVLocIdx];
2128 switch (VA.getLocInfo()) {
2129 default: llvm_unreachable("Unknown loc info!");
2130 case CCValAssign::Full: break;
2131 case CCValAssign::BCvt:
2132 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2136 if (VA.needsCustom()) {
2137 if (VA.getLocVT() == MVT::v2f64) {
2138 // Extract the first half and return it in two registers.
2139 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2140 DAG.getConstant(0, MVT::i32));
2141 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2142 DAG.getVTList(MVT::i32, MVT::i32), Half);
2144 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2145 Flag = Chain.getValue(1);
2146 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2147 VA = RVLocs[++i]; // skip ahead to next loc
2148 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2149 HalfGPRs.getValue(1), Flag);
2150 Flag = Chain.getValue(1);
2151 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2152 VA = RVLocs[++i]; // skip ahead to next loc
2154 // Extract the 2nd half and fall through to handle it as an f64 value.
2155 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2156 DAG.getConstant(1, MVT::i32));
2158 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2160 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2161 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2162 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2163 Flag = Chain.getValue(1);
2164 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2165 VA = RVLocs[++i]; // skip ahead to next loc
2166 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2169 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2171 // Guarantee that all emitted copies are
2172 // stuck together, avoiding something bad.
2173 Flag = Chain.getValue(1);
2174 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2177 // Update chain and glue.
2180 RetOps.push_back(Flag);
2182 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2183 RetOps.data(), RetOps.size());
2186 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2187 if (N->getNumValues() != 1)
2189 if (!N->hasNUsesOfValue(1, 0))
2192 SDValue TCChain = Chain;
2193 SDNode *Copy = *N->use_begin();
2194 if (Copy->getOpcode() == ISD::CopyToReg) {
2195 // If the copy has a glue operand, we conservatively assume it isn't safe to
2196 // perform a tail call.
2197 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2199 TCChain = Copy->getOperand(0);
2200 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2201 SDNode *VMov = Copy;
2202 // f64 returned in a pair of GPRs.
2203 SmallPtrSet<SDNode*, 2> Copies;
2204 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2206 if (UI->getOpcode() != ISD::CopyToReg)
2210 if (Copies.size() > 2)
2213 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2215 SDValue UseChain = UI->getOperand(0);
2216 if (Copies.count(UseChain.getNode()))
2223 } else if (Copy->getOpcode() == ISD::BITCAST) {
2224 // f32 returned in a single GPR.
2225 if (!Copy->hasOneUse())
2227 Copy = *Copy->use_begin();
2228 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2230 TCChain = Copy->getOperand(0);
2235 bool HasRet = false;
2236 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2238 if (UI->getOpcode() != ARMISD::RET_FLAG)
2250 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2251 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2254 if (!CI->isTailCall())
2257 return !Subtarget->isThumb1Only();
2260 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2261 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2262 // one of the above mentioned nodes. It has to be wrapped because otherwise
2263 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2264 // be used to form addressing mode. These wrapped nodes will be selected
2266 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2267 EVT PtrVT = Op.getValueType();
2268 // FIXME there is no actual debug info here
2270 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2272 if (CP->isMachineConstantPoolEntry())
2273 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2274 CP->getAlignment());
2276 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2277 CP->getAlignment());
2278 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2281 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2282 return MachineJumpTableInfo::EK_Inline;
2285 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2286 SelectionDAG &DAG) const {
2287 MachineFunction &MF = DAG.getMachineFunction();
2288 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2289 unsigned ARMPCLabelIndex = 0;
2291 EVT PtrVT = getPointerTy();
2292 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2293 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2295 if (RelocM == Reloc::Static) {
2296 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2298 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2299 ARMPCLabelIndex = AFI->createPICLabelUId();
2300 ARMConstantPoolValue *CPV =
2301 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2302 ARMCP::CPBlockAddress, PCAdj);
2303 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2305 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2306 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2307 MachinePointerInfo::getConstantPool(),
2308 false, false, false, 0);
2309 if (RelocM == Reloc::Static)
2311 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2312 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2315 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2317 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2318 SelectionDAG &DAG) const {
2320 EVT PtrVT = getPointerTy();
2321 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2322 MachineFunction &MF = DAG.getMachineFunction();
2323 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2324 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2325 ARMConstantPoolValue *CPV =
2326 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2327 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2328 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2329 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2330 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2331 MachinePointerInfo::getConstantPool(),
2332 false, false, false, 0);
2333 SDValue Chain = Argument.getValue(1);
2335 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2336 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2338 // call __tls_get_addr.
2341 Entry.Node = Argument;
2342 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2343 Args.push_back(Entry);
2344 // FIXME: is there useful debug info available here?
2345 TargetLowering::CallLoweringInfo CLI(Chain,
2346 (Type *) Type::getInt32Ty(*DAG.getContext()),
2347 false, false, false, false,
2348 0, CallingConv::C, /*isTailCall=*/false,
2349 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2350 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2351 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2352 return CallResult.first;
2355 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2356 // "local exec" model.
2358 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2360 TLSModel::Model model) const {
2361 const GlobalValue *GV = GA->getGlobal();
2364 SDValue Chain = DAG.getEntryNode();
2365 EVT PtrVT = getPointerTy();
2366 // Get the Thread Pointer
2367 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2369 if (model == TLSModel::InitialExec) {
2370 MachineFunction &MF = DAG.getMachineFunction();
2371 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2372 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2373 // Initial exec model.
2374 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2375 ARMConstantPoolValue *CPV =
2376 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2377 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2379 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2380 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2381 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2382 MachinePointerInfo::getConstantPool(),
2383 false, false, false, 0);
2384 Chain = Offset.getValue(1);
2386 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2387 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2389 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2390 MachinePointerInfo::getConstantPool(),
2391 false, false, false, 0);
2394 assert(model == TLSModel::LocalExec);
2395 ARMConstantPoolValue *CPV =
2396 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2397 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2398 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2399 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2400 MachinePointerInfo::getConstantPool(),
2401 false, false, false, 0);
2404 // The address of the thread local variable is the add of the thread
2405 // pointer with the offset of the variable.
2406 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2410 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2411 // TODO: implement the "local dynamic" model
2412 assert(Subtarget->isTargetELF() &&
2413 "TLS not implemented for non-ELF targets");
2414 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2416 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2419 case TLSModel::GeneralDynamic:
2420 case TLSModel::LocalDynamic:
2421 return LowerToTLSGeneralDynamicModel(GA, DAG);
2422 case TLSModel::InitialExec:
2423 case TLSModel::LocalExec:
2424 return LowerToTLSExecModels(GA, DAG, model);
2426 llvm_unreachable("bogus TLS model");
2429 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2430 SelectionDAG &DAG) const {
2431 EVT PtrVT = getPointerTy();
2433 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2434 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2435 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2436 ARMConstantPoolValue *CPV =
2437 ARMConstantPoolConstant::Create(GV,
2438 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2439 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2440 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2441 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2443 MachinePointerInfo::getConstantPool(),
2444 false, false, false, 0);
2445 SDValue Chain = Result.getValue(1);
2446 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2447 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2449 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2450 MachinePointerInfo::getGOT(),
2451 false, false, false, 0);
2455 // If we have T2 ops, we can materialize the address directly via movt/movw
2456 // pair. This is always cheaper.
2457 if (Subtarget->useMovt()) {
2459 // FIXME: Once remat is capable of dealing with instructions with register
2460 // operands, expand this into two nodes.
2461 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2462 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2464 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2465 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2466 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2467 MachinePointerInfo::getConstantPool(),
2468 false, false, false, 0);
2472 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2473 SelectionDAG &DAG) const {
2474 EVT PtrVT = getPointerTy();
2476 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2477 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2479 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2480 // update ARMFastISel::ARMMaterializeGV.
2481 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2483 // FIXME: Once remat is capable of dealing with instructions with register
2484 // operands, expand this into two nodes.
2485 if (RelocM == Reloc::Static)
2486 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2487 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2489 unsigned Wrapper = (RelocM == Reloc::PIC_)
2490 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2491 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2492 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2493 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2494 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2495 MachinePointerInfo::getGOT(),
2496 false, false, false, 0);
2500 unsigned ARMPCLabelIndex = 0;
2502 if (RelocM == Reloc::Static) {
2503 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2505 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2506 ARMPCLabelIndex = AFI->createPICLabelUId();
2507 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2508 ARMConstantPoolValue *CPV =
2509 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2511 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2513 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 Chain = Result.getValue(1);
2520 if (RelocM == Reloc::PIC_) {
2521 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2522 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2525 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2526 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2527 false, false, false, 0);
2532 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2533 SelectionDAG &DAG) const {
2534 assert(Subtarget->isTargetELF() &&
2535 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2536 MachineFunction &MF = DAG.getMachineFunction();
2537 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2538 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2539 EVT PtrVT = getPointerTy();
2541 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2542 ARMConstantPoolValue *CPV =
2543 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2544 ARMPCLabelIndex, PCAdj);
2545 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2546 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2547 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2548 MachinePointerInfo::getConstantPool(),
2549 false, false, false, 0);
2550 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2551 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2555 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2557 SDValue Val = DAG.getConstant(0, MVT::i32);
2558 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2559 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2560 Op.getOperand(1), Val);
2564 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2566 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2567 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2571 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2572 const ARMSubtarget *Subtarget) const {
2573 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2576 default: return SDValue(); // Don't custom lower most intrinsics.
2577 case Intrinsic::arm_thread_pointer: {
2578 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2579 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2581 case Intrinsic::eh_sjlj_lsda: {
2582 MachineFunction &MF = DAG.getMachineFunction();
2583 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2584 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2585 EVT PtrVT = getPointerTy();
2586 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2588 unsigned PCAdj = (RelocM != Reloc::PIC_)
2589 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2590 ARMConstantPoolValue *CPV =
2591 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2592 ARMCP::CPLSDA, PCAdj);
2593 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2594 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2596 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2597 MachinePointerInfo::getConstantPool(),
2598 false, false, false, 0);
2600 if (RelocM == Reloc::PIC_) {
2601 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2602 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2606 case Intrinsic::arm_neon_vmulls:
2607 case Intrinsic::arm_neon_vmullu: {
2608 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2609 ? ARMISD::VMULLs : ARMISD::VMULLu;
2610 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2611 Op.getOperand(1), Op.getOperand(2));
2616 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2617 const ARMSubtarget *Subtarget) {
2618 // FIXME: handle "fence singlethread" more efficiently.
2620 if (!Subtarget->hasDataBarrier()) {
2621 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2622 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2624 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2625 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2626 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2627 DAG.getConstant(0, MVT::i32));
2630 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2631 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2632 unsigned Domain = ARM_MB::ISH;
2633 if (Subtarget->isMClass()) {
2634 // Only a full system barrier exists in the M-class architectures.
2635 Domain = ARM_MB::SY;
2636 } else if (Subtarget->isSwift() && Ord == Release) {
2637 // Swift happens to implement ISHST barriers in a way that's compatible with
2638 // Release semantics but weaker than ISH so we'd be fools not to use
2639 // it. Beware: other processors probably don't!
2640 Domain = ARM_MB::ISHST;
2643 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2644 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2645 DAG.getConstant(Domain, MVT::i32));
2648 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2649 const ARMSubtarget *Subtarget) {
2650 // ARM pre v5TE and Thumb1 does not have preload instructions.
2651 if (!(Subtarget->isThumb2() ||
2652 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2653 // Just preserve the chain.
2654 return Op.getOperand(0);
2657 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2659 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2660 // ARMv7 with MP extension has PLDW.
2661 return Op.getOperand(0);
2663 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2664 if (Subtarget->isThumb()) {
2666 isRead = ~isRead & 1;
2667 isData = ~isData & 1;
2670 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2671 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2672 DAG.getConstant(isData, MVT::i32));
2675 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2676 MachineFunction &MF = DAG.getMachineFunction();
2677 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2679 // vastart just stores the address of the VarArgsFrameIndex slot into the
2680 // memory location argument.
2682 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2683 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2684 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2685 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2686 MachinePointerInfo(SV), false, false, 0);
2690 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2691 SDValue &Root, SelectionDAG &DAG,
2693 MachineFunction &MF = DAG.getMachineFunction();
2694 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2696 const TargetRegisterClass *RC;
2697 if (AFI->isThumb1OnlyFunction())
2698 RC = &ARM::tGPRRegClass;
2700 RC = &ARM::GPRRegClass;
2702 // Transform the arguments stored in physical registers into virtual ones.
2703 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2704 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2707 if (NextVA.isMemLoc()) {
2708 MachineFrameInfo *MFI = MF.getFrameInfo();
2709 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2711 // Create load node to retrieve arguments from the stack.
2712 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2713 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2714 MachinePointerInfo::getFixedStack(FI),
2715 false, false, false, 0);
2717 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2718 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2721 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2725 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2726 unsigned InRegsParamRecordIdx,
2728 unsigned &ArgRegsSize,
2729 unsigned &ArgRegsSaveSize)
2732 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2733 unsigned RBegin, REnd;
2734 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2735 NumGPRs = REnd - RBegin;
2737 unsigned int firstUnalloced;
2738 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2739 sizeof(GPRArgRegs) /
2740 sizeof(GPRArgRegs[0]));
2741 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2744 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2745 ArgRegsSize = NumGPRs * 4;
2747 // If parameter is split between stack and GPRs...
2748 if (NumGPRs && Align == 8 &&
2749 (ArgRegsSize < ArgSize ||
2750 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2751 // Add padding for part of param recovered from GPRs, so
2752 // its last byte must be at address K*8 - 1.
2753 // We need to do it, since remained (stack) part of parameter has
2754 // stack alignment, and we need to "attach" "GPRs head" without gaps
2757 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2758 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2760 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2762 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
2763 (ArgRegsSize + AFI->getArgRegsSaveSize());
2764 ArgRegsSaveSize = ArgRegsSize + Padding;
2766 // We don't need to extend regs save size for byval parameters if they
2767 // are passed via GPRs only.
2768 ArgRegsSaveSize = ArgRegsSize;
2771 // The remaining GPRs hold either the beginning of variable-argument
2772 // data, or the beginning of an aggregate passed by value (usually
2773 // byval). Either way, we allocate stack slots adjacent to the data
2774 // provided by our caller, and store the unallocated registers there.
2775 // If this is a variadic function, the va_list pointer will begin with
2776 // these values; otherwise, this reassembles a (byval) structure that
2777 // was split between registers and memory.
2778 // Return: The frame index registers were stored into.
2780 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2781 SDLoc dl, SDValue &Chain,
2782 const Value *OrigArg,
2783 unsigned InRegsParamRecordIdx,
2784 unsigned OffsetFromOrigArg,
2787 bool ForceMutable) const {
2789 // Currently, two use-cases possible:
2790 // Case #1. Non var-args function, and we meet first byval parameter.
2791 // Setup first unallocated register as first byval register;
2792 // eat all remained registers
2793 // (these two actions are performed by HandleByVal method).
2794 // Then, here, we initialize stack frame with
2795 // "store-reg" instructions.
2796 // Case #2. Var-args function, that doesn't contain byval parameters.
2797 // The same: eat all remained unallocated registers,
2798 // initialize stack frame.
2800 MachineFunction &MF = DAG.getMachineFunction();
2801 MachineFrameInfo *MFI = MF.getFrameInfo();
2802 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2803 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2804 unsigned RBegin, REnd;
2805 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2806 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2807 firstRegToSaveIndex = RBegin - ARM::R0;
2808 lastRegToSaveIndex = REnd - ARM::R0;
2810 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2811 (GPRArgRegs, array_lengthof(GPRArgRegs));
2812 lastRegToSaveIndex = 4;
2815 unsigned ArgRegsSize, ArgRegsSaveSize;
2816 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2817 ArgRegsSize, ArgRegsSaveSize);
2819 // Store any by-val regs to their spots on the stack so that they may be
2820 // loaded by deferencing the result of formal parameter pointer or va_next.
2821 // Note: once stack area for byval/varargs registers
2822 // was initialized, it can't be initialized again.
2823 if (ArgRegsSaveSize) {
2825 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2828 assert(AFI->getStoredByValParamsPadding() == 0 &&
2829 "The only parameter may be padded.");
2830 AFI->setStoredByValParamsPadding(Padding);
2833 int FrameIndex = MFI->CreateFixedObject(
2835 Padding + ArgOffset,
2837 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2839 SmallVector<SDValue, 4> MemOps;
2840 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2841 ++firstRegToSaveIndex, ++i) {
2842 const TargetRegisterClass *RC;
2843 if (AFI->isThumb1OnlyFunction())
2844 RC = &ARM::tGPRRegClass;
2846 RC = &ARM::GPRRegClass;
2848 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2849 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2851 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2852 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2854 MemOps.push_back(Store);
2855 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2856 DAG.getConstant(4, getPointerTy()));
2859 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2861 if (!MemOps.empty())
2862 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2863 &MemOps[0], MemOps.size());
2866 // This will point to the next argument passed via stack.
2867 return MFI->CreateFixedObject(
2868 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
2871 // Setup stack frame, the va_list pointer will start from.
2873 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2874 SDLoc dl, SDValue &Chain,
2876 bool ForceMutable) const {
2877 MachineFunction &MF = DAG.getMachineFunction();
2878 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2880 // Try to store any remaining integer argument regs
2881 // to their spots on the stack so that they may be loaded by deferencing
2882 // the result of va_next.
2883 // If there is no regs to be stored, just point address after last
2884 // argument passed via stack.
2886 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2887 0, ArgOffset, 0, ForceMutable);
2889 AFI->setVarArgsFrameIndex(FrameIndex);
2893 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2894 CallingConv::ID CallConv, bool isVarArg,
2895 const SmallVectorImpl<ISD::InputArg>
2897 SDLoc dl, SelectionDAG &DAG,
2898 SmallVectorImpl<SDValue> &InVals)
2900 MachineFunction &MF = DAG.getMachineFunction();
2901 MachineFrameInfo *MFI = MF.getFrameInfo();
2903 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2905 // Assign locations to all of the incoming arguments.
2906 SmallVector<CCValAssign, 16> ArgLocs;
2907 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2908 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2909 CCInfo.AnalyzeFormalArguments(Ins,
2910 CCAssignFnForNode(CallConv, /* Return*/ false,
2913 SmallVector<SDValue, 16> ArgValues;
2914 int lastInsIndex = -1;
2916 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2917 unsigned CurArgIdx = 0;
2919 // Initially ArgRegsSaveSize is zero.
2920 // Then we increase this value each time we meet byval parameter.
2921 // We also increase this value in case of varargs function.
2922 AFI->setArgRegsSaveSize(0);
2924 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2925 CCValAssign &VA = ArgLocs[i];
2926 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2927 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2928 // Arguments stored in registers.
2929 if (VA.isRegLoc()) {
2930 EVT RegVT = VA.getLocVT();
2932 if (VA.needsCustom()) {
2933 // f64 and vector types are split up into multiple registers or
2934 // combinations of registers and stack slots.
2935 if (VA.getLocVT() == MVT::v2f64) {
2936 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2938 VA = ArgLocs[++i]; // skip ahead to next loc
2940 if (VA.isMemLoc()) {
2941 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2942 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2943 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2944 MachinePointerInfo::getFixedStack(FI),
2945 false, false, false, 0);
2947 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2950 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2951 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2952 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2953 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2954 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2956 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2959 const TargetRegisterClass *RC;
2961 if (RegVT == MVT::f32)
2962 RC = &ARM::SPRRegClass;
2963 else if (RegVT == MVT::f64)
2964 RC = &ARM::DPRRegClass;
2965 else if (RegVT == MVT::v2f64)
2966 RC = &ARM::QPRRegClass;
2967 else if (RegVT == MVT::i32)
2968 RC = AFI->isThumb1OnlyFunction() ?
2969 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2970 (const TargetRegisterClass*)&ARM::GPRRegClass;
2972 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2974 // Transform the arguments in physical registers into virtual ones.
2975 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2976 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2979 // If this is an 8 or 16-bit value, it is really passed promoted
2980 // to 32 bits. Insert an assert[sz]ext to capture this, then
2981 // truncate to the right size.
2982 switch (VA.getLocInfo()) {
2983 default: llvm_unreachable("Unknown loc info!");
2984 case CCValAssign::Full: break;
2985 case CCValAssign::BCvt:
2986 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2988 case CCValAssign::SExt:
2989 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2990 DAG.getValueType(VA.getValVT()));
2991 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2993 case CCValAssign::ZExt:
2994 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2995 DAG.getValueType(VA.getValVT()));
2996 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3000 InVals.push_back(ArgValue);
3002 } else { // VA.isRegLoc()
3005 assert(VA.isMemLoc());
3006 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3008 int index = ArgLocs[i].getValNo();
3010 // Some Ins[] entries become multiple ArgLoc[] entries.
3011 // Process them only once.
3012 if (index != lastInsIndex)
3014 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3015 // FIXME: For now, all byval parameter objects are marked mutable.
3016 // This can be changed with more analysis.
3017 // In case of tail call optimization mark all arguments mutable.
3018 // Since they could be overwritten by lowering of arguments in case of
3020 if (Flags.isByVal()) {
3021 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3022 int FrameIndex = StoreByValRegs(
3023 CCInfo, DAG, dl, Chain, CurOrigArg,
3025 Ins[VA.getValNo()].PartOffset,
3026 VA.getLocMemOffset(),
3027 Flags.getByValSize(),
3028 true /*force mutable frames*/);
3029 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3030 CCInfo.nextInRegsParam();
3032 unsigned FIOffset = VA.getLocMemOffset() +
3033 AFI->getStoredByValParamsPadding();
3034 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3037 // Create load nodes to retrieve arguments from the stack.
3038 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3039 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3040 MachinePointerInfo::getFixedStack(FI),
3041 false, false, false, 0));
3043 lastInsIndex = index;
3050 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3051 CCInfo.getNextStackOffset());
3056 /// isFloatingPointZero - Return true if this is +0.0.
3057 static bool isFloatingPointZero(SDValue Op) {
3058 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3059 return CFP->getValueAPF().isPosZero();
3060 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3061 // Maybe this has already been legalized into the constant pool?
3062 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3063 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3064 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3065 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3066 return CFP->getValueAPF().isPosZero();
3072 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3073 /// the given operands.
3075 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3076 SDValue &ARMcc, SelectionDAG &DAG,
3078 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3079 unsigned C = RHSC->getZExtValue();
3080 if (!isLegalICmpImmediate(C)) {
3081 // Constant does not fit, try adjusting it by one?
3086 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3087 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3088 RHS = DAG.getConstant(C-1, MVT::i32);
3093 if (C != 0 && isLegalICmpImmediate(C-1)) {
3094 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3095 RHS = DAG.getConstant(C-1, MVT::i32);
3100 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3101 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3102 RHS = DAG.getConstant(C+1, MVT::i32);
3107 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3108 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3109 RHS = DAG.getConstant(C+1, MVT::i32);
3116 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3117 ARMISD::NodeType CompareType;
3120 CompareType = ARMISD::CMP;
3125 CompareType = ARMISD::CMPZ;
3128 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3129 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3132 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3134 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3137 if (!isFloatingPointZero(RHS))
3138 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3140 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3141 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3144 /// duplicateCmp - Glue values can have only one use, so this function
3145 /// duplicates a comparison node.
3147 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3148 unsigned Opc = Cmp.getOpcode();
3150 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3151 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3153 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3154 Cmp = Cmp.getOperand(0);
3155 Opc = Cmp.getOpcode();
3156 if (Opc == ARMISD::CMPFP)
3157 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3159 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3160 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3162 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3165 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3166 SDValue Cond = Op.getOperand(0);
3167 SDValue SelectTrue = Op.getOperand(1);
3168 SDValue SelectFalse = Op.getOperand(2);
3173 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3174 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3176 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3177 const ConstantSDNode *CMOVTrue =
3178 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3179 const ConstantSDNode *CMOVFalse =
3180 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3182 if (CMOVTrue && CMOVFalse) {
3183 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3184 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3188 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3190 False = SelectFalse;
3191 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3196 if (True.getNode() && False.getNode()) {
3197 EVT VT = Op.getValueType();
3198 SDValue ARMcc = Cond.getOperand(2);
3199 SDValue CCR = Cond.getOperand(3);
3200 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3201 assert(True.getValueType() == VT);
3202 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3207 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3208 // undefined bits before doing a full-word comparison with zero.
3209 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3210 DAG.getConstant(1, Cond.getValueType()));
3212 return DAG.getSelectCC(dl, Cond,
3213 DAG.getConstant(0, Cond.getValueType()),
3214 SelectTrue, SelectFalse, ISD::SETNE);
3217 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3218 if (CC == ISD::SETNE)
3220 return ISD::getSetCCSwappedOperands(CC);
3223 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3224 bool &swpCmpOps, bool &swpVselOps) {
3225 // Start by selecting the GE condition code for opcodes that return true for
3227 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3229 CondCode = ARMCC::GE;
3231 // and GT for opcodes that return false for 'equality'.
3232 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3234 CondCode = ARMCC::GT;
3236 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3237 // to swap the compare operands.
3238 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3242 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3243 // If we have an unordered opcode, we need to swap the operands to the VSEL
3244 // instruction (effectively negating the condition).
3246 // This also has the effect of swapping which one of 'less' or 'greater'
3247 // returns true, so we also swap the compare operands. It also switches
3248 // whether we return true for 'equality', so we compensate by picking the
3249 // opposite condition code to our original choice.
3250 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3251 CC == ISD::SETUGT) {
3252 swpCmpOps = !swpCmpOps;
3253 swpVselOps = !swpVselOps;
3254 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3257 // 'ordered' is 'anything but unordered', so use the VS condition code and
3258 // swap the VSEL operands.
3259 if (CC == ISD::SETO) {
3260 CondCode = ARMCC::VS;
3264 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3265 // code and swap the VSEL operands.
3266 if (CC == ISD::SETUNE) {
3267 CondCode = ARMCC::EQ;
3272 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3273 EVT VT = Op.getValueType();
3274 SDValue LHS = Op.getOperand(0);
3275 SDValue RHS = Op.getOperand(1);
3276 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3277 SDValue TrueVal = Op.getOperand(2);
3278 SDValue FalseVal = Op.getOperand(3);
3281 if (LHS.getValueType() == MVT::i32) {
3282 // Try to generate VSEL on ARMv8.
3283 // The VSEL instruction can't use all the usual ARM condition
3284 // codes: it only has two bits to select the condition code, so it's
3285 // constrained to use only GE, GT, VS and EQ.
3287 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3288 // swap the operands of the previous compare instruction (effectively
3289 // inverting the compare condition, swapping 'less' and 'greater') and
3290 // sometimes need to swap the operands to the VSEL (which inverts the
3291 // condition in the sense of firing whenever the previous condition didn't)
3292 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3293 TrueVal.getValueType() == MVT::f64)) {
3294 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3295 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3296 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3297 CC = getInverseCCForVSEL(CC);
3298 std::swap(TrueVal, FalseVal);
3303 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3304 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3305 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3309 ARMCC::CondCodes CondCode, CondCode2;
3310 FPCCToARMCC(CC, CondCode, CondCode2);
3312 // Try to generate VSEL on ARMv8.
3313 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3314 TrueVal.getValueType() == MVT::f64)) {
3315 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3316 // same operands, as follows:
3317 // c = fcmp [ogt, olt, ugt, ult] a, b
3319 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3320 // handled differently than the original code sequence.
3321 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3323 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3324 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3325 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3326 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3329 bool swpCmpOps = false;
3330 bool swpVselOps = false;
3331 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3333 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3334 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3336 std::swap(LHS, RHS);
3338 std::swap(TrueVal, FalseVal);
3342 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3343 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3344 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3345 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3347 if (CondCode2 != ARMCC::AL) {
3348 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3349 // FIXME: Needs another CMP because flag can have but one use.
3350 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3351 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3352 Result, TrueVal, ARMcc2, CCR, Cmp2);
3357 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3358 /// to morph to an integer compare sequence.
3359 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3360 const ARMSubtarget *Subtarget) {
3361 SDNode *N = Op.getNode();
3362 if (!N->hasOneUse())
3363 // Otherwise it requires moving the value from fp to integer registers.
3365 if (!N->getNumValues())
3367 EVT VT = Op.getValueType();
3368 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3369 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3370 // vmrs are very slow, e.g. cortex-a8.
3373 if (isFloatingPointZero(Op)) {
3377 return ISD::isNormalLoad(N);
3380 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3381 if (isFloatingPointZero(Op))
3382 return DAG.getConstant(0, MVT::i32);
3384 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3385 return DAG.getLoad(MVT::i32, SDLoc(Op),
3386 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3387 Ld->isVolatile(), Ld->isNonTemporal(),
3388 Ld->isInvariant(), Ld->getAlignment());
3390 llvm_unreachable("Unknown VFP cmp argument!");
3393 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3394 SDValue &RetVal1, SDValue &RetVal2) {
3395 if (isFloatingPointZero(Op)) {
3396 RetVal1 = DAG.getConstant(0, MVT::i32);
3397 RetVal2 = DAG.getConstant(0, MVT::i32);
3401 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3402 SDValue Ptr = Ld->getBasePtr();
3403 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3404 Ld->getChain(), Ptr,
3405 Ld->getPointerInfo(),
3406 Ld->isVolatile(), Ld->isNonTemporal(),
3407 Ld->isInvariant(), Ld->getAlignment());
3409 EVT PtrType = Ptr.getValueType();
3410 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3411 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3412 PtrType, Ptr, DAG.getConstant(4, PtrType));
3413 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3414 Ld->getChain(), NewPtr,
3415 Ld->getPointerInfo().getWithOffset(4),
3416 Ld->isVolatile(), Ld->isNonTemporal(),
3417 Ld->isInvariant(), NewAlign);
3421 llvm_unreachable("Unknown VFP cmp argument!");
3424 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3425 /// f32 and even f64 comparisons to integer ones.
3427 ARMTargetLowering::OptimizeVFPBrcond(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 bool LHSSeenZero = false;
3436 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3437 bool RHSSeenZero = false;
3438 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3439 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3440 // If unsafe fp math optimization is enabled and there are no other uses of
3441 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3442 // to an integer comparison.
3443 if (CC == ISD::SETOEQ)
3445 else if (CC == ISD::SETUNE)
3448 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3450 if (LHS.getValueType() == MVT::f32) {
3451 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3452 bitcastf32Toi32(LHS, DAG), Mask);
3453 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3454 bitcastf32Toi32(RHS, DAG), Mask);
3455 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3456 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3457 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3458 Chain, Dest, ARMcc, CCR, Cmp);
3463 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3464 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3465 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3466 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3467 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3468 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3469 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3470 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3471 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3477 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3478 SDValue Chain = Op.getOperand(0);
3479 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3480 SDValue LHS = Op.getOperand(2);
3481 SDValue RHS = Op.getOperand(3);
3482 SDValue Dest = Op.getOperand(4);
3485 if (LHS.getValueType() == MVT::i32) {
3487 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3488 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3489 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3490 Chain, Dest, ARMcc, CCR, Cmp);
3493 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3495 if (getTargetMachine().Options.UnsafeFPMath &&
3496 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3497 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3498 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3499 if (Result.getNode())
3503 ARMCC::CondCodes CondCode, CondCode2;
3504 FPCCToARMCC(CC, CondCode, CondCode2);
3506 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3507 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3508 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3509 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3510 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3511 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3512 if (CondCode2 != ARMCC::AL) {
3513 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3514 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3515 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3520 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3521 SDValue Chain = Op.getOperand(0);
3522 SDValue Table = Op.getOperand(1);
3523 SDValue Index = Op.getOperand(2);
3526 EVT PTy = getPointerTy();
3527 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3528 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3529 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3530 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3531 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3532 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3533 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3534 if (Subtarget->isThumb2()) {
3535 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3536 // which does another jump to the destination. This also makes it easier
3537 // to translate it to TBB / TBH later.
3538 // FIXME: This might not work if the function is extremely large.
3539 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3540 Addr, Op.getOperand(2), JTI, UId);
3542 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3543 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3544 MachinePointerInfo::getJumpTable(),
3545 false, false, false, 0);
3546 Chain = Addr.getValue(1);
3547 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3548 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3550 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3551 MachinePointerInfo::getJumpTable(),
3552 false, false, false, 0);
3553 Chain = Addr.getValue(1);
3554 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3558 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3559 EVT VT = Op.getValueType();
3562 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3563 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3565 return DAG.UnrollVectorOp(Op.getNode());
3568 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3569 "Invalid type for custom lowering!");
3570 if (VT != MVT::v4i16)
3571 return DAG.UnrollVectorOp(Op.getNode());
3573 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3574 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3577 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3578 EVT VT = Op.getValueType();
3580 return LowerVectorFP_TO_INT(Op, DAG);
3585 switch (Op.getOpcode()) {
3586 default: llvm_unreachable("Invalid opcode!");
3587 case ISD::FP_TO_SINT:
3588 Opc = ARMISD::FTOSI;
3590 case ISD::FP_TO_UINT:
3591 Opc = ARMISD::FTOUI;
3594 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3595 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3598 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3599 EVT VT = Op.getValueType();
3602 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3603 if (VT.getVectorElementType() == MVT::f32)
3605 return DAG.UnrollVectorOp(Op.getNode());
3608 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3609 "Invalid type for custom lowering!");
3610 if (VT != MVT::v4f32)
3611 return DAG.UnrollVectorOp(Op.getNode());
3615 switch (Op.getOpcode()) {
3616 default: llvm_unreachable("Invalid opcode!");
3617 case ISD::SINT_TO_FP:
3618 CastOpc = ISD::SIGN_EXTEND;
3619 Opc = ISD::SINT_TO_FP;
3621 case ISD::UINT_TO_FP:
3622 CastOpc = ISD::ZERO_EXTEND;
3623 Opc = ISD::UINT_TO_FP;
3627 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3628 return DAG.getNode(Opc, dl, VT, Op);
3631 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3632 EVT VT = Op.getValueType();
3634 return LowerVectorINT_TO_FP(Op, DAG);
3639 switch (Op.getOpcode()) {
3640 default: llvm_unreachable("Invalid opcode!");
3641 case ISD::SINT_TO_FP:
3642 Opc = ARMISD::SITOF;
3644 case ISD::UINT_TO_FP:
3645 Opc = ARMISD::UITOF;
3649 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3650 return DAG.getNode(Opc, dl, VT, Op);
3653 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3654 // Implement fcopysign with a fabs and a conditional fneg.
3655 SDValue Tmp0 = Op.getOperand(0);
3656 SDValue Tmp1 = Op.getOperand(1);
3658 EVT VT = Op.getValueType();
3659 EVT SrcVT = Tmp1.getValueType();
3660 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3661 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3662 bool UseNEON = !InGPR && Subtarget->hasNEON();
3665 // Use VBSL to copy the sign bit.
3666 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3667 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3668 DAG.getTargetConstant(EncodedVal, MVT::i32));
3669 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3671 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3672 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3673 DAG.getConstant(32, MVT::i32));
3674 else /*if (VT == MVT::f32)*/
3675 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3676 if (SrcVT == MVT::f32) {
3677 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3679 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3680 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3681 DAG.getConstant(32, MVT::i32));
3682 } else if (VT == MVT::f32)
3683 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3684 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3685 DAG.getConstant(32, MVT::i32));
3686 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3687 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3689 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3691 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3692 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3693 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3695 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3696 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3697 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3698 if (VT == MVT::f32) {
3699 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3700 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3701 DAG.getConstant(0, MVT::i32));
3703 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3709 // Bitcast operand 1 to i32.
3710 if (SrcVT == MVT::f64)
3711 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3712 &Tmp1, 1).getValue(1);
3713 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3715 // Or in the signbit with integer operations.
3716 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3717 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3718 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3719 if (VT == MVT::f32) {
3720 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3721 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3722 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3723 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3726 // f64: Or the high part with signbit and then combine two parts.
3727 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3729 SDValue Lo = Tmp0.getValue(0);
3730 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3731 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3732 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3735 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3736 MachineFunction &MF = DAG.getMachineFunction();
3737 MachineFrameInfo *MFI = MF.getFrameInfo();
3738 MFI->setReturnAddressIsTaken(true);
3740 EVT VT = Op.getValueType();
3742 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3744 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3745 SDValue Offset = DAG.getConstant(4, MVT::i32);
3746 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3747 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3748 MachinePointerInfo(), false, false, false, 0);
3751 // Return LR, which contains the return address. Mark it an implicit live-in.
3752 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3753 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3756 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3757 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3758 MFI->setFrameAddressIsTaken(true);
3760 EVT VT = Op.getValueType();
3761 SDLoc dl(Op); // FIXME probably not meaningful
3762 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3763 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3764 ? ARM::R7 : ARM::R11;
3765 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3767 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3768 MachinePointerInfo(),
3769 false, false, false, 0);
3773 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
3774 /// and size(DestVec) > 128-bits.
3775 /// This is achieved by doing the one extension from the SrcVec, splitting the
3776 /// result, extending these parts, and then concatenating these into the
3778 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
3779 SDValue Op = N->getOperand(0);
3780 EVT SrcVT = Op.getValueType();
3781 EVT DestVT = N->getValueType(0);
3783 assert(DestVT.getSizeInBits() > 128 &&
3784 "Custom sext/zext expansion needs >128-bit vector.");
3785 // If this is a normal length extension, use the default expansion.
3786 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
3787 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
3791 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
3792 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
3793 unsigned NumElts = SrcVT.getVectorNumElements();
3794 LLVMContext &Ctx = *DAG.getContext();
3795 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;
3797 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3799 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
3801 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
3804 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
3805 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3806 DAG.getIntPtrConstant(0));
3807 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
3808 DAG.getIntPtrConstant(NumElts/2));
3809 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
3810 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
3811 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
3814 /// ExpandBITCAST - If the target supports VFP, this function is called to
3815 /// expand a bit convert where either the source or destination type is i64 to
3816 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3817 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3818 /// vectors), since the legalizer won't know what to do with that.
3819 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3820 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3822 SDValue Op = N->getOperand(0);
3824 // This function is only supposed to be called for i64 types, either as the
3825 // source or destination of the bit convert.
3826 EVT SrcVT = Op.getValueType();
3827 EVT DstVT = N->getValueType(0);
3828 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3829 "ExpandBITCAST called for non-i64 type");
3831 // Turn i64->f64 into VMOVDRR.
3832 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3833 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3834 DAG.getConstant(0, MVT::i32));
3835 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3836 DAG.getConstant(1, MVT::i32));
3837 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3838 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3841 // Turn f64->i64 into VMOVRRD.
3842 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3843 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3844 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3845 // Merge the pieces into a single i64 value.
3846 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3852 /// getZeroVector - Returns a vector of specified type with all zero elements.
3853 /// Zero vectors are used to represent vector negation and in those cases
3854 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3855 /// not support i64 elements, so sometimes the zero vectors will need to be
3856 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3858 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3859 assert(VT.isVector() && "Expected a vector type");
3860 // The canonical modified immediate encoding of a zero vector is....0!
3861 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3862 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3863 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3864 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3867 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3868 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3869 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3870 SelectionDAG &DAG) const {
3871 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3872 EVT VT = Op.getValueType();
3873 unsigned VTBits = VT.getSizeInBits();
3875 SDValue ShOpLo = Op.getOperand(0);
3876 SDValue ShOpHi = Op.getOperand(1);
3877 SDValue ShAmt = Op.getOperand(2);
3879 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3881 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3883 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3884 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3885 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3886 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3887 DAG.getConstant(VTBits, MVT::i32));
3888 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3889 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3890 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3892 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3893 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3895 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3896 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3899 SDValue Ops[2] = { Lo, Hi };
3900 return DAG.getMergeValues(Ops, 2, dl);
3903 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3904 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3905 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3906 SelectionDAG &DAG) const {
3907 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3908 EVT VT = Op.getValueType();
3909 unsigned VTBits = VT.getSizeInBits();
3911 SDValue ShOpLo = Op.getOperand(0);
3912 SDValue ShOpHi = Op.getOperand(1);
3913 SDValue ShAmt = Op.getOperand(2);
3916 assert(Op.getOpcode() == ISD::SHL_PARTS);
3917 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3918 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3919 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3920 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3921 DAG.getConstant(VTBits, MVT::i32));
3922 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3923 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3925 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3926 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3927 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3929 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3930 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3933 SDValue Ops[2] = { Lo, Hi };
3934 return DAG.getMergeValues(Ops, 2, dl);
3937 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3938 SelectionDAG &DAG) const {
3939 // The rounding mode is in bits 23:22 of the FPSCR.
3940 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3941 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3942 // so that the shift + and get folded into a bitfield extract.
3944 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3945 DAG.getConstant(Intrinsic::arm_get_fpscr,
3947 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3948 DAG.getConstant(1U << 22, MVT::i32));
3949 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3950 DAG.getConstant(22, MVT::i32));
3951 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3952 DAG.getConstant(3, MVT::i32));
3955 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3956 const ARMSubtarget *ST) {
3957 EVT VT = N->getValueType(0);
3960 if (!ST->hasV6T2Ops())
3963 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3964 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3967 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3968 /// for each 16-bit element from operand, repeated. The basic idea is to
3969 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3971 /// Trace for v4i16:
3972 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3973 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3974 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3975 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3976 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3977 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3978 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3979 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3980 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3981 EVT VT = N->getValueType(0);
3984 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3985 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3986 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3987 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3988 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3989 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3992 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3993 /// bit-count for each 16-bit element from the operand. We need slightly
3994 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3995 /// 64/128-bit registers.
3997 /// Trace for v4i16:
3998 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3999 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4000 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4001 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4002 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4003 EVT VT = N->getValueType(0);
4006 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4007 if (VT.is64BitVector()) {
4008 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4009 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4010 DAG.getIntPtrConstant(0));
4012 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4013 BitCounts, DAG.getIntPtrConstant(0));
4014 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4018 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4019 /// bit-count for each 32-bit element from the operand. The idea here is
4020 /// to split the vector into 16-bit elements, leverage the 16-bit count
4021 /// routine, and then combine the results.
4023 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4024 /// input = [v0 v1 ] (vi: 32-bit elements)
4025 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4026 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4027 /// vrev: N0 = [k1 k0 k3 k2 ]
4029 /// N1 =+[k1 k0 k3 k2 ]
4031 /// N2 =+[k1 k3 k0 k2 ]
4033 /// Extended =+[k1 k3 k0 k2 ]
4035 /// Extracted=+[k1 k3 ]
4037 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4038 EVT VT = N->getValueType(0);
4041 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4043 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4044 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4045 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4046 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4047 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4049 if (VT.is64BitVector()) {
4050 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4051 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4052 DAG.getIntPtrConstant(0));
4054 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4055 DAG.getIntPtrConstant(0));
4056 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4060 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4061 const ARMSubtarget *ST) {
4062 EVT VT = N->getValueType(0);
4064 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4065 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4066 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4067 "Unexpected type for custom ctpop lowering");
4069 if (VT.getVectorElementType() == MVT::i32)
4070 return lowerCTPOP32BitElements(N, DAG);
4072 return lowerCTPOP16BitElements(N, DAG);
4075 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4076 const ARMSubtarget *ST) {
4077 EVT VT = N->getValueType(0);
4083 // Lower vector shifts on NEON to use VSHL.
4084 assert(ST->hasNEON() && "unexpected vector shift");
4086 // Left shifts translate directly to the vshiftu intrinsic.
4087 if (N->getOpcode() == ISD::SHL)
4088 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4089 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4090 N->getOperand(0), N->getOperand(1));
4092 assert((N->getOpcode() == ISD::SRA ||
4093 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4095 // NEON uses the same intrinsics for both left and right shifts. For
4096 // right shifts, the shift amounts are negative, so negate the vector of
4098 EVT ShiftVT = N->getOperand(1).getValueType();
4099 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4100 getZeroVector(ShiftVT, DAG, dl),
4102 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4103 Intrinsic::arm_neon_vshifts :
4104 Intrinsic::arm_neon_vshiftu);
4105 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4106 DAG.getConstant(vshiftInt, MVT::i32),
4107 N->getOperand(0), NegatedCount);
4110 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4111 const ARMSubtarget *ST) {
4112 EVT VT = N->getValueType(0);
4115 // We can get here for a node like i32 = ISD::SHL i32, i64
4119 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4120 "Unknown shift to lower!");
4122 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4123 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4124 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4127 // If we are in thumb mode, we don't have RRX.
4128 if (ST->isThumb1Only()) return SDValue();
4130 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4131 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4132 DAG.getConstant(0, MVT::i32));
4133 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4134 DAG.getConstant(1, MVT::i32));
4136 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4137 // captures the result into a carry flag.
4138 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4139 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4141 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4142 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4144 // Merge the pieces into a single i64 value.
4145 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4148 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4149 SDValue TmpOp0, TmpOp1;
4150 bool Invert = false;
4154 SDValue Op0 = Op.getOperand(0);
4155 SDValue Op1 = Op.getOperand(1);
4156 SDValue CC = Op.getOperand(2);
4157 EVT VT = Op.getValueType();
4158 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4161 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4162 switch (SetCCOpcode) {
4163 default: llvm_unreachable("Illegal FP comparison");
4165 case ISD::SETNE: Invert = true; // Fallthrough
4167 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4169 case ISD::SETLT: Swap = true; // Fallthrough
4171 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4173 case ISD::SETLE: Swap = true; // Fallthrough
4175 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4176 case ISD::SETUGE: Swap = true; // Fallthrough
4177 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4178 case ISD::SETUGT: Swap = true; // Fallthrough
4179 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4180 case ISD::SETUEQ: Invert = true; // Fallthrough
4182 // Expand this to (OLT | OGT).
4186 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4187 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4189 case ISD::SETUO: Invert = true; // Fallthrough
4191 // Expand this to (OLT | OGE).
4195 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4196 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4200 // Integer comparisons.
4201 switch (SetCCOpcode) {
4202 default: llvm_unreachable("Illegal integer comparison");
4203 case ISD::SETNE: Invert = true;
4204 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4205 case ISD::SETLT: Swap = true;
4206 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4207 case ISD::SETLE: Swap = true;
4208 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4209 case ISD::SETULT: Swap = true;
4210 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4211 case ISD::SETULE: Swap = true;
4212 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4215 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4216 if (Opc == ARMISD::VCEQ) {
4219 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4221 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4224 // Ignore bitconvert.
4225 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4226 AndOp = AndOp.getOperand(0);
4228 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4230 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4231 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4238 std::swap(Op0, Op1);
4240 // If one of the operands is a constant vector zero, attempt to fold the
4241 // comparison to a specialized compare-against-zero form.
4243 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4245 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4246 if (Opc == ARMISD::VCGE)
4247 Opc = ARMISD::VCLEZ;
4248 else if (Opc == ARMISD::VCGT)
4249 Opc = ARMISD::VCLTZ;
4254 if (SingleOp.getNode()) {
4257 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4259 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4261 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4263 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4265 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4267 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4270 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4274 Result = DAG.getNOT(dl, Result, VT);
4279 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4280 /// valid vector constant for a NEON instruction with a "modified immediate"
4281 /// operand (e.g., VMOV). If so, return the encoded value.
4282 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4283 unsigned SplatBitSize, SelectionDAG &DAG,
4284 EVT &VT, bool is128Bits, NEONModImmType type) {
4285 unsigned OpCmode, Imm;
4287 // SplatBitSize is set to the smallest size that splats the vector, so a
4288 // zero vector will always have SplatBitSize == 8. However, NEON modified
4289 // immediate instructions others than VMOV do not support the 8-bit encoding
4290 // of a zero vector, and the default encoding of zero is supposed to be the
4295 switch (SplatBitSize) {
4297 if (type != VMOVModImm)
4299 // Any 1-byte value is OK. Op=0, Cmode=1110.
4300 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4303 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4307 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4308 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4309 if ((SplatBits & ~0xff) == 0) {
4310 // Value = 0x00nn: Op=x, Cmode=100x.
4315 if ((SplatBits & ~0xff00) == 0) {
4316 // Value = 0xnn00: Op=x, Cmode=101x.
4318 Imm = SplatBits >> 8;
4324 // NEON's 32-bit VMOV supports splat values where:
4325 // * only one byte is nonzero, or
4326 // * the least significant byte is 0xff and the second byte is nonzero, or
4327 // * the least significant 2 bytes are 0xff and the third is nonzero.
4328 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4329 if ((SplatBits & ~0xff) == 0) {
4330 // Value = 0x000000nn: Op=x, Cmode=000x.
4335 if ((SplatBits & ~0xff00) == 0) {
4336 // Value = 0x0000nn00: Op=x, Cmode=001x.
4338 Imm = SplatBits >> 8;
4341 if ((SplatBits & ~0xff0000) == 0) {
4342 // Value = 0x00nn0000: Op=x, Cmode=010x.
4344 Imm = SplatBits >> 16;
4347 if ((SplatBits & ~0xff000000) == 0) {
4348 // Value = 0xnn000000: Op=x, Cmode=011x.
4350 Imm = SplatBits >> 24;
4354 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4355 if (type == OtherModImm) return SDValue();
4357 if ((SplatBits & ~0xffff) == 0 &&
4358 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4359 // Value = 0x0000nnff: Op=x, Cmode=1100.
4361 Imm = SplatBits >> 8;
4366 if ((SplatBits & ~0xffffff) == 0 &&
4367 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4368 // Value = 0x00nnffff: Op=x, Cmode=1101.
4370 Imm = SplatBits >> 16;
4371 SplatBits |= 0xffff;
4375 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4376 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4377 // VMOV.I32. A (very) minor optimization would be to replicate the value
4378 // and fall through here to test for a valid 64-bit splat. But, then the
4379 // caller would also need to check and handle the change in size.
4383 if (type != VMOVModImm)
4385 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4386 uint64_t BitMask = 0xff;
4388 unsigned ImmMask = 1;
4390 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4391 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4394 } else if ((SplatBits & BitMask) != 0) {
4400 // Op=1, Cmode=1110.
4403 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4408 llvm_unreachable("unexpected size for isNEONModifiedImm");
4411 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4412 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4415 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4416 const ARMSubtarget *ST) const {
4420 bool IsDouble = Op.getValueType() == MVT::f64;
4421 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4423 // Try splatting with a VMOV.f32...
4424 APFloat FPVal = CFP->getValueAPF();
4425 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4428 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4429 // We have code in place to select a valid ConstantFP already, no need to
4434 // It's a float and we are trying to use NEON operations where
4435 // possible. Lower it to a splat followed by an extract.
4437 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4438 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4440 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4441 DAG.getConstant(0, MVT::i32));
4444 // The rest of our options are NEON only, make sure that's allowed before
4446 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4450 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4452 // It wouldn't really be worth bothering for doubles except for one very
4453 // important value, which does happen to match: 0.0. So make sure we don't do
4455 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4458 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4459 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4461 if (NewVal != SDValue()) {
4463 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4466 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4468 // It's a float: cast and extract a vector element.
4469 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4471 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4472 DAG.getConstant(0, MVT::i32));
4475 // Finally, try a VMVN.i32
4476 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4478 if (NewVal != SDValue()) {
4480 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4483 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4485 // It's a float: cast and extract a vector element.
4486 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4488 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4489 DAG.getConstant(0, MVT::i32));
4495 // check if an VEXT instruction can handle the shuffle mask when the
4496 // vector sources of the shuffle are the same.
4497 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4498 unsigned NumElts = VT.getVectorNumElements();
4500 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4506 // If this is a VEXT shuffle, the immediate value is the index of the first
4507 // element. The other shuffle indices must be the successive elements after
4509 unsigned ExpectedElt = Imm;
4510 for (unsigned i = 1; i < NumElts; ++i) {
4511 // Increment the expected index. If it wraps around, just follow it
4512 // back to index zero and keep going.
4514 if (ExpectedElt == NumElts)
4517 if (M[i] < 0) continue; // ignore UNDEF indices
4518 if (ExpectedElt != static_cast<unsigned>(M[i]))
4526 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4527 bool &ReverseVEXT, unsigned &Imm) {
4528 unsigned NumElts = VT.getVectorNumElements();
4529 ReverseVEXT = false;
4531 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4537 // If this is a VEXT shuffle, the immediate value is the index of the first
4538 // element. The other shuffle indices must be the successive elements after
4540 unsigned ExpectedElt = Imm;
4541 for (unsigned i = 1; i < NumElts; ++i) {
4542 // Increment the expected index. If it wraps around, it may still be
4543 // a VEXT but the source vectors must be swapped.
4545 if (ExpectedElt == NumElts * 2) {
4550 if (M[i] < 0) continue; // ignore UNDEF indices
4551 if (ExpectedElt != static_cast<unsigned>(M[i]))
4555 // Adjust the index value if the source operands will be swapped.
4562 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4563 /// instruction with the specified blocksize. (The order of the elements
4564 /// within each block of the vector is reversed.)
4565 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4566 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4567 "Only possible block sizes for VREV are: 16, 32, 64");
4569 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4573 unsigned NumElts = VT.getVectorNumElements();
4574 unsigned BlockElts = M[0] + 1;
4575 // If the first shuffle index is UNDEF, be optimistic.
4577 BlockElts = BlockSize / EltSz;
4579 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4582 for (unsigned i = 0; i < NumElts; ++i) {
4583 if (M[i] < 0) continue; // ignore UNDEF indices
4584 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4591 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4592 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4593 // range, then 0 is placed into the resulting vector. So pretty much any mask
4594 // of 8 elements can work here.
4595 return VT == MVT::v8i8 && M.size() == 8;
4598 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4599 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4603 unsigned NumElts = VT.getVectorNumElements();
4604 WhichResult = (M[0] == 0 ? 0 : 1);
4605 for (unsigned i = 0; i < NumElts; i += 2) {
4606 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4607 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4613 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4614 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4615 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4616 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4617 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4621 unsigned NumElts = VT.getVectorNumElements();
4622 WhichResult = (M[0] == 0 ? 0 : 1);
4623 for (unsigned i = 0; i < NumElts; i += 2) {
4624 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4625 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4631 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4632 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4636 unsigned NumElts = VT.getVectorNumElements();
4637 WhichResult = (M[0] == 0 ? 0 : 1);
4638 for (unsigned i = 0; i != NumElts; ++i) {
4639 if (M[i] < 0) continue; // ignore UNDEF indices
4640 if ((unsigned) M[i] != 2 * i + WhichResult)
4644 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4645 if (VT.is64BitVector() && EltSz == 32)
4651 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4652 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4653 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4654 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4655 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4659 unsigned Half = VT.getVectorNumElements() / 2;
4660 WhichResult = (M[0] == 0 ? 0 : 1);
4661 for (unsigned j = 0; j != 2; ++j) {
4662 unsigned Idx = WhichResult;
4663 for (unsigned i = 0; i != Half; ++i) {
4664 int MIdx = M[i + j * Half];
4665 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4671 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4672 if (VT.is64BitVector() && EltSz == 32)
4678 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4679 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4683 unsigned NumElts = VT.getVectorNumElements();
4684 WhichResult = (M[0] == 0 ? 0 : 1);
4685 unsigned Idx = WhichResult * NumElts / 2;
4686 for (unsigned i = 0; i != NumElts; i += 2) {
4687 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4688 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4693 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4694 if (VT.is64BitVector() && EltSz == 32)
4700 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4701 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4702 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4703 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4704 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4708 unsigned NumElts = VT.getVectorNumElements();
4709 WhichResult = (M[0] == 0 ? 0 : 1);
4710 unsigned Idx = WhichResult * NumElts / 2;
4711 for (unsigned i = 0; i != NumElts; i += 2) {
4712 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4713 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4718 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4719 if (VT.is64BitVector() && EltSz == 32)
4725 /// \return true if this is a reverse operation on an vector.
4726 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4727 unsigned NumElts = VT.getVectorNumElements();
4728 // Make sure the mask has the right size.
4729 if (NumElts != M.size())
4732 // Look for <15, ..., 3, -1, 1, 0>.
4733 for (unsigned i = 0; i != NumElts; ++i)
4734 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4740 // If N is an integer constant that can be moved into a register in one
4741 // instruction, return an SDValue of such a constant (will become a MOV
4742 // instruction). Otherwise return null.
4743 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4744 const ARMSubtarget *ST, SDLoc dl) {
4746 if (!isa<ConstantSDNode>(N))
4748 Val = cast<ConstantSDNode>(N)->getZExtValue();
4750 if (ST->isThumb1Only()) {
4751 if (Val <= 255 || ~Val <= 255)
4752 return DAG.getConstant(Val, MVT::i32);
4754 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4755 return DAG.getConstant(Val, MVT::i32);
4760 // If this is a case we can't handle, return null and let the default
4761 // expansion code take care of it.
4762 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4763 const ARMSubtarget *ST) const {
4764 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4766 EVT VT = Op.getValueType();
4768 APInt SplatBits, SplatUndef;
4769 unsigned SplatBitSize;
4771 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4772 if (SplatBitSize <= 64) {
4773 // Check if an immediate VMOV works.
4775 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4776 SplatUndef.getZExtValue(), SplatBitSize,
4777 DAG, VmovVT, VT.is128BitVector(),
4779 if (Val.getNode()) {
4780 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4781 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4784 // Try an immediate VMVN.
4785 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4786 Val = isNEONModifiedImm(NegatedImm,
4787 SplatUndef.getZExtValue(), SplatBitSize,
4788 DAG, VmovVT, VT.is128BitVector(),
4790 if (Val.getNode()) {
4791 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4792 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4795 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4796 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4797 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4799 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4800 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4806 // Scan through the operands to see if only one value is used.
4808 // As an optimisation, even if more than one value is used it may be more
4809 // profitable to splat with one value then change some lanes.
4811 // Heuristically we decide to do this if the vector has a "dominant" value,
4812 // defined as splatted to more than half of the lanes.
4813 unsigned NumElts = VT.getVectorNumElements();
4814 bool isOnlyLowElement = true;
4815 bool usesOnlyOneValue = true;
4816 bool hasDominantValue = false;
4817 bool isConstant = true;
4819 // Map of the number of times a particular SDValue appears in the
4821 DenseMap<SDValue, unsigned> ValueCounts;
4823 for (unsigned i = 0; i < NumElts; ++i) {
4824 SDValue V = Op.getOperand(i);
4825 if (V.getOpcode() == ISD::UNDEF)
4828 isOnlyLowElement = false;
4829 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4832 ValueCounts.insert(std::make_pair(V, 0));
4833 unsigned &Count = ValueCounts[V];
4835 // Is this value dominant? (takes up more than half of the lanes)
4836 if (++Count > (NumElts / 2)) {
4837 hasDominantValue = true;
4841 if (ValueCounts.size() != 1)
4842 usesOnlyOneValue = false;
4843 if (!Value.getNode() && ValueCounts.size() > 0)
4844 Value = ValueCounts.begin()->first;
4846 if (ValueCounts.size() == 0)
4847 return DAG.getUNDEF(VT);
4849 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4850 // Keep going if we are hitting this case.
4851 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4852 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4854 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4856 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4857 // i32 and try again.
4858 if (hasDominantValue && EltSize <= 32) {
4862 // If we are VDUPing a value that comes directly from a vector, that will
4863 // cause an unnecessary move to and from a GPR, where instead we could
4864 // just use VDUPLANE. We can only do this if the lane being extracted
4865 // is at a constant index, as the VDUP from lane instructions only have
4866 // constant-index forms.
4867 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4868 isa<ConstantSDNode>(Value->getOperand(1))) {
4869 // We need to create a new undef vector to use for the VDUPLANE if the
4870 // size of the vector from which we get the value is different than the
4871 // size of the vector that we need to create. We will insert the element
4872 // such that the register coalescer will remove unnecessary copies.
4873 if (VT != Value->getOperand(0).getValueType()) {
4874 ConstantSDNode *constIndex;
4875 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4876 assert(constIndex && "The index is not a constant!");
4877 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4878 VT.getVectorNumElements();
4879 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4880 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4881 Value, DAG.getConstant(index, MVT::i32)),
4882 DAG.getConstant(index, MVT::i32));
4884 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4885 Value->getOperand(0), Value->getOperand(1));
4887 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4889 if (!usesOnlyOneValue) {
4890 // The dominant value was splatted as 'N', but we now have to insert
4891 // all differing elements.
4892 for (unsigned I = 0; I < NumElts; ++I) {
4893 if (Op.getOperand(I) == Value)
4895 SmallVector<SDValue, 3> Ops;
4897 Ops.push_back(Op.getOperand(I));
4898 Ops.push_back(DAG.getConstant(I, MVT::i32));
4899 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4904 if (VT.getVectorElementType().isFloatingPoint()) {
4905 SmallVector<SDValue, 8> Ops;
4906 for (unsigned i = 0; i < NumElts; ++i)
4907 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4909 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4910 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4911 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4913 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4915 if (usesOnlyOneValue) {
4916 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4917 if (isConstant && Val.getNode())
4918 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4922 // If all elements are constants and the case above didn't get hit, fall back
4923 // to the default expansion, which will generate a load from the constant
4928 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4930 SDValue shuffle = ReconstructShuffle(Op, DAG);
4931 if (shuffle != SDValue())
4935 // Vectors with 32- or 64-bit elements can be built by directly assigning
4936 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4937 // will be legalized.
4938 if (EltSize >= 32) {
4939 // Do the expansion with floating-point types, since that is what the VFP
4940 // registers are defined to use, and since i64 is not legal.
4941 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4942 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4943 SmallVector<SDValue, 8> Ops;
4944 for (unsigned i = 0; i < NumElts; ++i)
4945 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4946 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4947 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4950 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4951 // know the default expansion would otherwise fall back on something even
4952 // worse. For a vector with one or two non-undef values, that's
4953 // scalar_to_vector for the elements followed by a shuffle (provided the
4954 // shuffle is valid for the target) and materialization element by element
4955 // on the stack followed by a load for everything else.
4956 if (!isConstant && !usesOnlyOneValue) {
4957 SDValue Vec = DAG.getUNDEF(VT);
4958 for (unsigned i = 0 ; i < NumElts; ++i) {
4959 SDValue V = Op.getOperand(i);
4960 if (V.getOpcode() == ISD::UNDEF)
4962 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4963 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4971 // Gather data to see if the operation can be modelled as a
4972 // shuffle in combination with VEXTs.
4973 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4974 SelectionDAG &DAG) const {
4976 EVT VT = Op.getValueType();
4977 unsigned NumElts = VT.getVectorNumElements();
4979 SmallVector<SDValue, 2> SourceVecs;
4980 SmallVector<unsigned, 2> MinElts;
4981 SmallVector<unsigned, 2> MaxElts;
4983 for (unsigned i = 0; i < NumElts; ++i) {
4984 SDValue V = Op.getOperand(i);
4985 if (V.getOpcode() == ISD::UNDEF)
4987 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4988 // A shuffle can only come from building a vector from various
4989 // elements of other vectors.
4991 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4992 VT.getVectorElementType()) {
4993 // This code doesn't know how to handle shuffles where the vector
4994 // element types do not match (this happens because type legalization
4995 // promotes the return type of EXTRACT_VECTOR_ELT).
4996 // FIXME: It might be appropriate to extend this code to handle
4997 // mismatched types.
5001 // Record this extraction against the appropriate vector if possible...
5002 SDValue SourceVec = V.getOperand(0);
5003 // If the element number isn't a constant, we can't effectively
5004 // analyze what's going on.
5005 if (!isa<ConstantSDNode>(V.getOperand(1)))
5007 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5008 bool FoundSource = false;
5009 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5010 if (SourceVecs[j] == SourceVec) {
5011 if (MinElts[j] > EltNo)
5013 if (MaxElts[j] < EltNo)
5020 // Or record a new source if not...
5022 SourceVecs.push_back(SourceVec);
5023 MinElts.push_back(EltNo);
5024 MaxElts.push_back(EltNo);
5028 // Currently only do something sane when at most two source vectors
5030 if (SourceVecs.size() > 2)
5033 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5034 int VEXTOffsets[2] = {0, 0};
5036 // This loop extracts the usage patterns of the source vectors
5037 // and prepares appropriate SDValues for a shuffle if possible.
5038 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5039 if (SourceVecs[i].getValueType() == VT) {
5040 // No VEXT necessary
5041 ShuffleSrcs[i] = SourceVecs[i];
5044 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5045 // It probably isn't worth padding out a smaller vector just to
5046 // break it down again in a shuffle.
5050 // Since only 64-bit and 128-bit vectors are legal on ARM and
5051 // we've eliminated the other cases...
5052 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5053 "unexpected vector sizes in ReconstructShuffle");
5055 if (MaxElts[i] - MinElts[i] >= NumElts) {
5056 // Span too large for a VEXT to cope
5060 if (MinElts[i] >= NumElts) {
5061 // The extraction can just take the second half
5062 VEXTOffsets[i] = NumElts;
5063 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5065 DAG.getIntPtrConstant(NumElts));
5066 } else if (MaxElts[i] < NumElts) {
5067 // The extraction can just take the first half
5069 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5071 DAG.getIntPtrConstant(0));
5073 // An actual VEXT is needed
5074 VEXTOffsets[i] = MinElts[i];
5075 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5077 DAG.getIntPtrConstant(0));
5078 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5080 DAG.getIntPtrConstant(NumElts));
5081 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5082 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5086 SmallVector<int, 8> Mask;
5088 for (unsigned i = 0; i < NumElts; ++i) {
5089 SDValue Entry = Op.getOperand(i);
5090 if (Entry.getOpcode() == ISD::UNDEF) {
5095 SDValue ExtractVec = Entry.getOperand(0);
5096 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5097 .getOperand(1))->getSExtValue();
5098 if (ExtractVec == SourceVecs[0]) {
5099 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5101 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5105 // Final check before we try to produce nonsense...
5106 if (isShuffleMaskLegal(Mask, VT))
5107 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5113 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5114 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5115 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5116 /// are assumed to be legal.
5118 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5120 if (VT.getVectorNumElements() == 4 &&
5121 (VT.is128BitVector() || VT.is64BitVector())) {
5122 unsigned PFIndexes[4];
5123 for (unsigned i = 0; i != 4; ++i) {
5127 PFIndexes[i] = M[i];
5130 // Compute the index in the perfect shuffle table.
5131 unsigned PFTableIndex =
5132 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5133 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5134 unsigned Cost = (PFEntry >> 30);
5141 unsigned Imm, WhichResult;
5143 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5144 return (EltSize >= 32 ||
5145 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5146 isVREVMask(M, VT, 64) ||
5147 isVREVMask(M, VT, 32) ||
5148 isVREVMask(M, VT, 16) ||
5149 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5150 isVTBLMask(M, VT) ||
5151 isVTRNMask(M, VT, WhichResult) ||
5152 isVUZPMask(M, VT, WhichResult) ||
5153 isVZIPMask(M, VT, WhichResult) ||
5154 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5155 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5156 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5157 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5160 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5161 /// the specified operations to build the shuffle.
5162 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5163 SDValue RHS, SelectionDAG &DAG,
5165 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5166 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5167 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5170 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5179 OP_VUZPL, // VUZP, left result
5180 OP_VUZPR, // VUZP, right result
5181 OP_VZIPL, // VZIP, left result
5182 OP_VZIPR, // VZIP, right result
5183 OP_VTRNL, // VTRN, left result
5184 OP_VTRNR // VTRN, right result
5187 if (OpNum == OP_COPY) {
5188 if (LHSID == (1*9+2)*9+3) return LHS;
5189 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5193 SDValue OpLHS, OpRHS;
5194 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5195 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5196 EVT VT = OpLHS.getValueType();
5199 default: llvm_unreachable("Unknown shuffle opcode!");
5201 // VREV divides the vector in half and swaps within the half.
5202 if (VT.getVectorElementType() == MVT::i32 ||
5203 VT.getVectorElementType() == MVT::f32)
5204 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5205 // vrev <4 x i16> -> VREV32
5206 if (VT.getVectorElementType() == MVT::i16)
5207 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5208 // vrev <4 x i8> -> VREV16
5209 assert(VT.getVectorElementType() == MVT::i8);
5210 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5215 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5216 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5220 return DAG.getNode(ARMISD::VEXT, dl, VT,
5222 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5225 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5226 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5229 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5230 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5233 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5234 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5238 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5239 ArrayRef<int> ShuffleMask,
5240 SelectionDAG &DAG) {
5241 // Check to see if we can use the VTBL instruction.
5242 SDValue V1 = Op.getOperand(0);
5243 SDValue V2 = Op.getOperand(1);
5246 SmallVector<SDValue, 8> VTBLMask;
5247 for (ArrayRef<int>::iterator
5248 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5249 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5251 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5252 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5253 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5256 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5257 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5261 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5262 SelectionDAG &DAG) {
5264 SDValue OpLHS = Op.getOperand(0);
5265 EVT VT = OpLHS.getValueType();
5267 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5268 "Expect an v8i16/v16i8 type");
5269 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5270 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5271 // extract the first 8 bytes into the top double word and the last 8 bytes
5272 // into the bottom double word. The v8i16 case is similar.
5273 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5274 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5275 DAG.getConstant(ExtractNum, MVT::i32));
5278 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5279 SDValue V1 = Op.getOperand(0);
5280 SDValue V2 = Op.getOperand(1);
5282 EVT VT = Op.getValueType();
5283 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5285 // Convert shuffles that are directly supported on NEON to target-specific
5286 // DAG nodes, instead of keeping them as shuffles and matching them again
5287 // during code selection. This is more efficient and avoids the possibility
5288 // of inconsistencies between legalization and selection.
5289 // FIXME: floating-point vectors should be canonicalized to integer vectors
5290 // of the same time so that they get CSEd properly.
5291 ArrayRef<int> ShuffleMask = SVN->getMask();
5293 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5294 if (EltSize <= 32) {
5295 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5296 int Lane = SVN->getSplatIndex();
5297 // If this is undef splat, generate it via "just" vdup, if possible.
5298 if (Lane == -1) Lane = 0;
5300 // Test if V1 is a SCALAR_TO_VECTOR.
5301 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5302 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5304 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5305 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5307 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5308 !isa<ConstantSDNode>(V1.getOperand(0))) {
5309 bool IsScalarToVector = true;
5310 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5311 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5312 IsScalarToVector = false;
5315 if (IsScalarToVector)
5316 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5318 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5319 DAG.getConstant(Lane, MVT::i32));
5324 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5327 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5328 DAG.getConstant(Imm, MVT::i32));
5331 if (isVREVMask(ShuffleMask, VT, 64))
5332 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5333 if (isVREVMask(ShuffleMask, VT, 32))
5334 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5335 if (isVREVMask(ShuffleMask, VT, 16))
5336 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5338 if (V2->getOpcode() == ISD::UNDEF &&
5339 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5340 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5341 DAG.getConstant(Imm, MVT::i32));
5344 // Check for Neon shuffles that modify both input vectors in place.
5345 // If both results are used, i.e., if there are two shuffles with the same
5346 // source operands and with masks corresponding to both results of one of
5347 // these operations, DAG memoization will ensure that a single node is
5348 // used for both shuffles.
5349 unsigned WhichResult;
5350 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5351 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5352 V1, V2).getValue(WhichResult);
5353 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5354 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5355 V1, V2).getValue(WhichResult);
5356 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5357 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5358 V1, V2).getValue(WhichResult);
5360 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5361 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5362 V1, V1).getValue(WhichResult);
5363 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5364 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5365 V1, V1).getValue(WhichResult);
5366 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5367 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5368 V1, V1).getValue(WhichResult);
5371 // If the shuffle is not directly supported and it has 4 elements, use
5372 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5373 unsigned NumElts = VT.getVectorNumElements();
5375 unsigned PFIndexes[4];
5376 for (unsigned i = 0; i != 4; ++i) {
5377 if (ShuffleMask[i] < 0)
5380 PFIndexes[i] = ShuffleMask[i];
5383 // Compute the index in the perfect shuffle table.
5384 unsigned PFTableIndex =
5385 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5386 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5387 unsigned Cost = (PFEntry >> 30);
5390 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5393 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5394 if (EltSize >= 32) {
5395 // Do the expansion with floating-point types, since that is what the VFP
5396 // registers are defined to use, and since i64 is not legal.
5397 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5398 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5399 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5400 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5401 SmallVector<SDValue, 8> Ops;
5402 for (unsigned i = 0; i < NumElts; ++i) {
5403 if (ShuffleMask[i] < 0)
5404 Ops.push_back(DAG.getUNDEF(EltVT));
5406 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5407 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5408 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5411 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5412 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5415 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5416 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5418 if (VT == MVT::v8i8) {
5419 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5420 if (NewOp.getNode())
5427 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5428 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5429 SDValue Lane = Op.getOperand(2);
5430 if (!isa<ConstantSDNode>(Lane))
5436 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5437 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5438 SDValue Lane = Op.getOperand(1);
5439 if (!isa<ConstantSDNode>(Lane))
5442 SDValue Vec = Op.getOperand(0);
5443 if (Op.getValueType() == MVT::i32 &&
5444 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5446 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5452 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5453 // The only time a CONCAT_VECTORS operation can have legal types is when
5454 // two 64-bit vectors are concatenated to a 128-bit vector.
5455 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5456 "unexpected CONCAT_VECTORS");
5458 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5459 SDValue Op0 = Op.getOperand(0);
5460 SDValue Op1 = Op.getOperand(1);
5461 if (Op0.getOpcode() != ISD::UNDEF)
5462 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5463 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5464 DAG.getIntPtrConstant(0));
5465 if (Op1.getOpcode() != ISD::UNDEF)
5466 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5467 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5468 DAG.getIntPtrConstant(1));
5469 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5472 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5473 /// element has been zero/sign-extended, depending on the isSigned parameter,
5474 /// from an integer type half its size.
5475 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5477 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5478 EVT VT = N->getValueType(0);
5479 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5480 SDNode *BVN = N->getOperand(0).getNode();
5481 if (BVN->getValueType(0) != MVT::v4i32 ||
5482 BVN->getOpcode() != ISD::BUILD_VECTOR)
5484 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5485 unsigned HiElt = 1 - LoElt;
5486 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5487 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5488 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5489 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5490 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5493 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5494 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5497 if (Hi0->isNullValue() && Hi1->isNullValue())
5503 if (N->getOpcode() != ISD::BUILD_VECTOR)
5506 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5507 SDNode *Elt = N->getOperand(i).getNode();
5508 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5509 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5510 unsigned HalfSize = EltSize / 2;
5512 if (!isIntN(HalfSize, C->getSExtValue()))
5515 if (!isUIntN(HalfSize, C->getZExtValue()))
5526 /// isSignExtended - Check if a node is a vector value that is sign-extended
5527 /// or a constant BUILD_VECTOR with sign-extended elements.
5528 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5529 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5531 if (isExtendedBUILD_VECTOR(N, DAG, true))
5536 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5537 /// or a constant BUILD_VECTOR with zero-extended elements.
5538 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5539 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5541 if (isExtendedBUILD_VECTOR(N, DAG, false))
5546 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5547 if (OrigVT.getSizeInBits() >= 64)
5550 assert(OrigVT.isSimple() && "Expecting a simple value type");
5552 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5553 switch (OrigSimpleTy) {
5554 default: llvm_unreachable("Unexpected Vector Type");
5563 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5564 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5565 /// We insert the required extension here to get the vector to fill a D register.
5566 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5569 unsigned ExtOpcode) {
5570 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5571 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5572 // 64-bits we need to insert a new extension so that it will be 64-bits.
5573 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5574 if (OrigTy.getSizeInBits() >= 64)
5577 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5578 EVT NewVT = getExtensionTo64Bits(OrigTy);
5580 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5583 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5584 /// does not do any sign/zero extension. If the original vector is less
5585 /// than 64 bits, an appropriate extension will be added after the load to
5586 /// reach a total size of 64 bits. We have to add the extension separately
5587 /// because ARM does not have a sign/zero extending load for vectors.
5588 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5589 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5591 // The load already has the right type.
5592 if (ExtendedTy == LD->getMemoryVT())
5593 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5594 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5595 LD->isNonTemporal(), LD->isInvariant(),
5596 LD->getAlignment());
5598 // We need to create a zextload/sextload. We cannot just create a load
5599 // followed by a zext/zext node because LowerMUL is also run during normal
5600 // operation legalization where we can't create illegal types.
5601 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5602 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5603 LD->getMemoryVT(), LD->isVolatile(),
5604 LD->isNonTemporal(), LD->getAlignment());
5607 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5608 /// extending load, or BUILD_VECTOR with extended elements, return the
5609 /// unextended value. The unextended vector should be 64 bits so that it can
5610 /// be used as an operand to a VMULL instruction. If the original vector size
5611 /// before extension is less than 64 bits we add a an extension to resize
5612 /// the vector to 64 bits.
5613 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5614 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5615 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5616 N->getOperand(0)->getValueType(0),
5620 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5621 return SkipLoadExtensionForVMULL(LD, DAG);
5623 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5624 // have been legalized as a BITCAST from v4i32.
5625 if (N->getOpcode() == ISD::BITCAST) {
5626 SDNode *BVN = N->getOperand(0).getNode();
5627 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5628 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5629 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5630 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5631 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5633 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5634 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5635 EVT VT = N->getValueType(0);
5636 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5637 unsigned NumElts = VT.getVectorNumElements();
5638 MVT TruncVT = MVT::getIntegerVT(EltSize);
5639 SmallVector<SDValue, 8> Ops;
5640 for (unsigned i = 0; i != NumElts; ++i) {
5641 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5642 const APInt &CInt = C->getAPIntValue();
5643 // Element types smaller than 32 bits are not legal, so use i32 elements.
5644 // The values are implicitly truncated so sext vs. zext doesn't matter.
5645 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5647 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5648 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5651 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5652 unsigned Opcode = N->getOpcode();
5653 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5654 SDNode *N0 = N->getOperand(0).getNode();
5655 SDNode *N1 = N->getOperand(1).getNode();
5656 return N0->hasOneUse() && N1->hasOneUse() &&
5657 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5662 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5663 unsigned Opcode = N->getOpcode();
5664 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5665 SDNode *N0 = N->getOperand(0).getNode();
5666 SDNode *N1 = N->getOperand(1).getNode();
5667 return N0->hasOneUse() && N1->hasOneUse() &&
5668 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5673 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5674 // Multiplications are only custom-lowered for 128-bit vectors so that
5675 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5676 EVT VT = Op.getValueType();
5677 assert(VT.is128BitVector() && VT.isInteger() &&
5678 "unexpected type for custom-lowering ISD::MUL");
5679 SDNode *N0 = Op.getOperand(0).getNode();
5680 SDNode *N1 = Op.getOperand(1).getNode();
5681 unsigned NewOpc = 0;
5683 bool isN0SExt = isSignExtended(N0, DAG);
5684 bool isN1SExt = isSignExtended(N1, DAG);
5685 if (isN0SExt && isN1SExt)
5686 NewOpc = ARMISD::VMULLs;
5688 bool isN0ZExt = isZeroExtended(N0, DAG);
5689 bool isN1ZExt = isZeroExtended(N1, DAG);
5690 if (isN0ZExt && isN1ZExt)
5691 NewOpc = ARMISD::VMULLu;
5692 else if (isN1SExt || isN1ZExt) {
5693 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5694 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5695 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5696 NewOpc = ARMISD::VMULLs;
5698 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5699 NewOpc = ARMISD::VMULLu;
5701 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5703 NewOpc = ARMISD::VMULLu;
5709 if (VT == MVT::v2i64)
5710 // Fall through to expand this. It is not legal.
5713 // Other vector multiplications are legal.
5718 // Legalize to a VMULL instruction.
5721 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5723 Op0 = SkipExtensionForVMULL(N0, DAG);
5724 assert(Op0.getValueType().is64BitVector() &&
5725 Op1.getValueType().is64BitVector() &&
5726 "unexpected types for extended operands to VMULL");
5727 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5730 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5731 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5738 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5739 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5740 EVT Op1VT = Op1.getValueType();
5741 return DAG.getNode(N0->getOpcode(), DL, VT,
5742 DAG.getNode(NewOpc, DL, VT,
5743 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5744 DAG.getNode(NewOpc, DL, VT,
5745 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5749 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5751 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5752 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5753 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5754 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5755 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5756 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5757 // Get reciprocal estimate.
5758 // float4 recip = vrecpeq_f32(yf);
5759 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5760 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5761 // Because char has a smaller range than uchar, we can actually get away
5762 // without any newton steps. This requires that we use a weird bias
5763 // of 0xb000, however (again, this has been exhaustively tested).
5764 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5765 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5766 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5767 Y = DAG.getConstant(0xb000, MVT::i32);
5768 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5769 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5770 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5771 // Convert back to short.
5772 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5773 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5778 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5780 // Convert to float.
5781 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5782 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5783 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5784 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5785 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5786 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5788 // Use reciprocal estimate and one refinement step.
5789 // float4 recip = vrecpeq_f32(yf);
5790 // recip *= vrecpsq_f32(yf, recip);
5791 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5792 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5793 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5794 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5796 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5797 // Because short has a smaller range than ushort, we can actually get away
5798 // with only a single newton step. This requires that we use a weird bias
5799 // of 89, however (again, this has been exhaustively tested).
5800 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5801 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5802 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5803 N1 = DAG.getConstant(0x89, MVT::i32);
5804 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5805 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5806 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5807 // Convert back to integer and return.
5808 // return vmovn_s32(vcvt_s32_f32(result));
5809 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5810 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5814 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5815 EVT VT = Op.getValueType();
5816 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5817 "unexpected type for custom-lowering ISD::SDIV");
5820 SDValue N0 = Op.getOperand(0);
5821 SDValue N1 = Op.getOperand(1);
5824 if (VT == MVT::v8i8) {
5825 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5826 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5828 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5829 DAG.getIntPtrConstant(4));
5830 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5831 DAG.getIntPtrConstant(4));
5832 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5833 DAG.getIntPtrConstant(0));
5834 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5835 DAG.getIntPtrConstant(0));
5837 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5838 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5840 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5841 N0 = LowerCONCAT_VECTORS(N0, DAG);
5843 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5846 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5849 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5850 EVT VT = Op.getValueType();
5851 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5852 "unexpected type for custom-lowering ISD::UDIV");
5855 SDValue N0 = Op.getOperand(0);
5856 SDValue N1 = Op.getOperand(1);
5859 if (VT == MVT::v8i8) {
5860 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5861 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5863 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5864 DAG.getIntPtrConstant(4));
5865 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5866 DAG.getIntPtrConstant(4));
5867 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5868 DAG.getIntPtrConstant(0));
5869 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5870 DAG.getIntPtrConstant(0));
5872 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5873 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5875 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5876 N0 = LowerCONCAT_VECTORS(N0, DAG);
5878 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5879 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5884 // v4i16 sdiv ... Convert to float.
5885 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5886 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5887 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5888 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5889 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5890 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5892 // Use reciprocal estimate and two refinement steps.
5893 // float4 recip = vrecpeq_f32(yf);
5894 // recip *= vrecpsq_f32(yf, recip);
5895 // recip *= vrecpsq_f32(yf, recip);
5896 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5897 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5898 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5899 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5901 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5902 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5903 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5905 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5906 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5907 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5908 // and that it will never cause us to return an answer too large).
5909 // float4 result = as_float4(as_int4(xf*recip) + 2);
5910 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5911 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5912 N1 = DAG.getConstant(2, MVT::i32);
5913 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5914 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5915 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5916 // Convert back to integer and return.
5917 // return vmovn_u32(vcvt_s32_f32(result));
5918 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5919 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5923 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5924 EVT VT = Op.getNode()->getValueType(0);
5925 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5928 bool ExtraOp = false;
5929 switch (Op.getOpcode()) {
5930 default: llvm_unreachable("Invalid code");
5931 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5932 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5933 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5934 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5938 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5940 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5941 Op.getOperand(1), Op.getOperand(2));
5944 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5945 // Monotonic load/store is legal for all targets
5946 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5949 // Aquire/Release load/store is not legal for targets without a
5950 // dmb or equivalent available.
5955 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5956 SelectionDAG &DAG) {
5958 assert (Node->getValueType(0) == MVT::i64 &&
5959 "Only know how to expand i64 atomics");
5960 AtomicSDNode *AN = cast<AtomicSDNode>(Node);
5962 SmallVector<SDValue, 6> Ops;
5963 Ops.push_back(Node->getOperand(0)); // Chain
5964 Ops.push_back(Node->getOperand(1)); // Ptr
5965 for(unsigned i=2; i<Node->getNumOperands(); i++) {
5967 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5968 Node->getOperand(i), DAG.getIntPtrConstant(0)));
5970 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5971 Node->getOperand(i), DAG.getIntPtrConstant(1)));
5973 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5975 DAG.getAtomic(Node->getOpcode(), dl, MVT::i64, Tys, Ops.data(), Ops.size(),
5976 cast<MemSDNode>(Node)->getMemOperand(), AN->getOrdering(),
5977 AN->getSynchScope());
5978 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5979 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5980 Results.push_back(Result.getValue(2));
5983 static void ReplaceREADCYCLECOUNTER(SDNode *N,
5984 SmallVectorImpl<SDValue> &Results,
5986 const ARMSubtarget *Subtarget) {
5988 SDValue Cycles32, OutChain;
5990 if (Subtarget->hasPerfMon()) {
5991 // Under Power Management extensions, the cycle-count is:
5992 // mrc p15, #0, <Rt>, c9, c13, #0
5993 SDValue Ops[] = { N->getOperand(0), // Chain
5994 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
5995 DAG.getConstant(15, MVT::i32),
5996 DAG.getConstant(0, MVT::i32),
5997 DAG.getConstant(9, MVT::i32),
5998 DAG.getConstant(13, MVT::i32),
5999 DAG.getConstant(0, MVT::i32)
6002 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6003 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6004 array_lengthof(Ops));
6005 OutChain = Cycles32.getValue(1);
6007 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6008 // there are older ARM CPUs that have implementation-specific ways of
6009 // obtaining this information (FIXME!).
6010 Cycles32 = DAG.getConstant(0, MVT::i32);
6011 OutChain = DAG.getEntryNode();
6015 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6016 Cycles32, DAG.getConstant(0, MVT::i32));
6017 Results.push_back(Cycles64);
6018 Results.push_back(OutChain);
6021 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6022 switch (Op.getOpcode()) {
6023 default: llvm_unreachable("Don't know how to custom lower this!");
6024 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6025 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6026 case ISD::GlobalAddress:
6027 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
6028 LowerGlobalAddressELF(Op, DAG);
6029 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6030 case ISD::SELECT: return LowerSELECT(Op, DAG);
6031 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6032 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6033 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6034 case ISD::VASTART: return LowerVASTART(Op, DAG);
6035 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6036 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6037 case ISD::SINT_TO_FP:
6038 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6039 case ISD::FP_TO_SINT:
6040 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6041 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6042 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6043 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6044 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6045 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6046 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6047 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6049 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6052 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6053 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6054 case ISD::SRL_PARTS:
6055 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6056 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6057 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6058 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6059 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6060 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6061 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6062 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6063 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6064 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6065 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6066 case ISD::MUL: return LowerMUL(Op, DAG);
6067 case ISD::SDIV: return LowerSDIV(Op, DAG);
6068 case ISD::UDIV: return LowerUDIV(Op, DAG);
6072 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6073 case ISD::ATOMIC_LOAD:
6074 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6076 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6080 /// ReplaceNodeResults - Replace the results of node with an illegal result
6081 /// type with new values built out of custom code.
6082 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6083 SmallVectorImpl<SDValue>&Results,
6084 SelectionDAG &DAG) const {
6086 switch (N->getOpcode()) {
6088 llvm_unreachable("Don't know how to custom expand this!");
6090 Res = ExpandBITCAST(N, DAG);
6092 case ISD::SIGN_EXTEND:
6093 case ISD::ZERO_EXTEND:
6094 Res = ExpandVectorExtension(N, DAG);
6098 Res = Expand64BitShift(N, DAG, Subtarget);
6100 case ISD::READCYCLECOUNTER:
6101 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6103 case ISD::ATOMIC_STORE:
6104 case ISD::ATOMIC_LOAD:
6105 case ISD::ATOMIC_LOAD_ADD:
6106 case ISD::ATOMIC_LOAD_AND:
6107 case ISD::ATOMIC_LOAD_NAND:
6108 case ISD::ATOMIC_LOAD_OR:
6109 case ISD::ATOMIC_LOAD_SUB:
6110 case ISD::ATOMIC_LOAD_XOR:
6111 case ISD::ATOMIC_SWAP:
6112 case ISD::ATOMIC_CMP_SWAP:
6113 case ISD::ATOMIC_LOAD_MIN:
6114 case ISD::ATOMIC_LOAD_UMIN:
6115 case ISD::ATOMIC_LOAD_MAX:
6116 case ISD::ATOMIC_LOAD_UMAX:
6117 ReplaceATOMIC_OP_64(N, Results, DAG);
6121 Results.push_back(Res);
6124 //===----------------------------------------------------------------------===//
6125 // ARM Scheduler Hooks
6126 //===----------------------------------------------------------------------===//
6129 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6130 MachineBasicBlock *BB,
6131 unsigned Size) const {
6132 unsigned dest = MI->getOperand(0).getReg();
6133 unsigned ptr = MI->getOperand(1).getReg();
6134 unsigned oldval = MI->getOperand(2).getReg();
6135 unsigned newval = MI->getOperand(3).getReg();
6136 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6137 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
6138 DebugLoc dl = MI->getDebugLoc();
6139 bool isThumb2 = Subtarget->isThumb2();
6141 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6142 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6143 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6144 (const TargetRegisterClass*)&ARM::GPRRegClass);
6147 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6148 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6149 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6152 unsigned ldrOpc, strOpc;
6153 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6155 MachineFunction *MF = BB->getParent();
6156 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6157 MachineFunction::iterator It = BB;
6158 ++It; // insert the new blocks after the current block
6160 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6161 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6162 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6163 MF->insert(It, loop1MBB);
6164 MF->insert(It, loop2MBB);
6165 MF->insert(It, exitMBB);
6167 // Transfer the remainder of BB and its successor edges to exitMBB.
6168 exitMBB->splice(exitMBB->begin(), BB,
6169 llvm::next(MachineBasicBlock::iterator(MI)),
6171 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6175 // fallthrough --> loop1MBB
6176 BB->addSuccessor(loop1MBB);
6179 // ldrex dest, [ptr]
6183 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6184 if (ldrOpc == ARM::t2LDREX)
6186 AddDefaultPred(MIB);
6187 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6188 .addReg(dest).addReg(oldval));
6189 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6190 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6191 BB->addSuccessor(loop2MBB);
6192 BB->addSuccessor(exitMBB);
6195 // strex scratch, newval, [ptr]
6199 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6200 if (strOpc == ARM::t2STREX)
6202 AddDefaultPred(MIB);
6203 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6204 .addReg(scratch).addImm(0));
6205 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6206 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6207 BB->addSuccessor(loop1MBB);
6208 BB->addSuccessor(exitMBB);
6214 MI->eraseFromParent(); // The instruction is gone now.
6220 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6221 unsigned Size, unsigned BinOpcode) const {
6222 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6223 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6225 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6226 MachineFunction *MF = BB->getParent();
6227 MachineFunction::iterator It = BB;
6230 unsigned dest = MI->getOperand(0).getReg();
6231 unsigned ptr = MI->getOperand(1).getReg();
6232 unsigned incr = MI->getOperand(2).getReg();
6233 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6234 DebugLoc dl = MI->getDebugLoc();
6235 bool isThumb2 = Subtarget->isThumb2();
6237 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6239 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6240 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6241 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6244 unsigned ldrOpc, strOpc;
6245 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6247 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6248 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6249 MF->insert(It, loopMBB);
6250 MF->insert(It, exitMBB);
6252 // Transfer the remainder of BB and its successor edges to exitMBB.
6253 exitMBB->splice(exitMBB->begin(), BB,
6254 llvm::next(MachineBasicBlock::iterator(MI)),
6256 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6258 const TargetRegisterClass *TRC = isThumb2 ?
6259 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6260 (const TargetRegisterClass*)&ARM::GPRRegClass;
6261 unsigned scratch = MRI.createVirtualRegister(TRC);
6262 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6266 // fallthrough --> loopMBB
6267 BB->addSuccessor(loopMBB);
6271 // <binop> scratch2, dest, incr
6272 // strex scratch, scratch2, ptr
6275 // fallthrough --> exitMBB
6277 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6278 if (ldrOpc == ARM::t2LDREX)
6280 AddDefaultPred(MIB);
6282 // operand order needs to go the other way for NAND
6283 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6284 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6285 addReg(incr).addReg(dest)).addReg(0);
6287 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6288 addReg(dest).addReg(incr)).addReg(0);
6291 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6292 if (strOpc == ARM::t2STREX)
6294 AddDefaultPred(MIB);
6295 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6296 .addReg(scratch).addImm(0));
6297 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6298 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6300 BB->addSuccessor(loopMBB);
6301 BB->addSuccessor(exitMBB);
6307 MI->eraseFromParent(); // The instruction is gone now.
6313 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6314 MachineBasicBlock *BB,
6317 ARMCC::CondCodes Cond) const {
6318 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6320 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6321 MachineFunction *MF = BB->getParent();
6322 MachineFunction::iterator It = BB;
6325 unsigned dest = MI->getOperand(0).getReg();
6326 unsigned ptr = MI->getOperand(1).getReg();
6327 unsigned incr = MI->getOperand(2).getReg();
6328 unsigned oldval = dest;
6329 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6330 DebugLoc dl = MI->getDebugLoc();
6331 bool isThumb2 = Subtarget->isThumb2();
6333 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6335 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6336 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6337 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6340 unsigned ldrOpc, strOpc, extendOpc;
6341 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6343 default: llvm_unreachable("unsupported size for AtomicBinaryMinMax!");
6345 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6348 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6355 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6356 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6357 MF->insert(It, loopMBB);
6358 MF->insert(It, exitMBB);
6360 // Transfer the remainder of BB and its successor edges to exitMBB.
6361 exitMBB->splice(exitMBB->begin(), BB,
6362 llvm::next(MachineBasicBlock::iterator(MI)),
6364 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6366 const TargetRegisterClass *TRC = isThumb2 ?
6367 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6368 (const TargetRegisterClass*)&ARM::GPRRegClass;
6369 unsigned scratch = MRI.createVirtualRegister(TRC);
6370 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6374 // fallthrough --> loopMBB
6375 BB->addSuccessor(loopMBB);
6379 // (sign extend dest, if required)
6381 // cmov.cond scratch2, incr, dest
6382 // strex scratch, scratch2, ptr
6385 // fallthrough --> exitMBB
6387 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6388 if (ldrOpc == ARM::t2LDREX)
6390 AddDefaultPred(MIB);
6392 // Sign extend the value, if necessary.
6393 if (signExtend && extendOpc) {
6394 oldval = MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass
6395 : &ARM::GPRnopcRegClass);
6397 MRI.constrainRegClass(dest, &ARM::GPRnopcRegClass);
6398 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6403 // Build compare and cmov instructions.
6404 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6405 .addReg(oldval).addReg(incr));
6406 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6407 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6409 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6410 if (strOpc == ARM::t2STREX)
6412 AddDefaultPred(MIB);
6413 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6414 .addReg(scratch).addImm(0));
6415 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6416 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6418 BB->addSuccessor(loopMBB);
6419 BB->addSuccessor(exitMBB);
6425 MI->eraseFromParent(); // The instruction is gone now.
6431 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6432 unsigned Op1, unsigned Op2,
6433 bool NeedsCarry, bool IsCmpxchg,
6434 bool IsMinMax, ARMCC::CondCodes CC) const {
6435 // This also handles ATOMIC_SWAP and ATOMIC_STORE, indicated by Op1==0.
6436 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6438 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6439 MachineFunction *MF = BB->getParent();
6440 MachineFunction::iterator It = BB;
6443 bool isStore = (MI->getOpcode() == ARM::ATOMIC_STORE_I64);
6444 unsigned offset = (isStore ? -2 : 0);
6445 unsigned destlo = MI->getOperand(0).getReg();
6446 unsigned desthi = MI->getOperand(1).getReg();
6447 unsigned ptr = MI->getOperand(offset+2).getReg();
6448 unsigned vallo = MI->getOperand(offset+3).getReg();
6449 unsigned valhi = MI->getOperand(offset+4).getReg();
6450 unsigned OrdIdx = offset + (IsCmpxchg ? 7 : 5);
6451 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(OrdIdx).getImm());
6452 DebugLoc dl = MI->getDebugLoc();
6453 bool isThumb2 = Subtarget->isThumb2();
6455 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6457 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6458 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6459 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6460 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6461 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6464 unsigned ldrOpc, strOpc;
6465 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6467 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6468 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6469 if (IsCmpxchg || IsMinMax)
6470 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6472 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6473 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6475 MF->insert(It, loopMBB);
6476 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6477 if (IsCmpxchg) MF->insert(It, cont2BB);
6478 MF->insert(It, exitMBB);
6480 // Transfer the remainder of BB and its successor edges to exitMBB.
6481 exitMBB->splice(exitMBB->begin(), BB,
6482 llvm::next(MachineBasicBlock::iterator(MI)),
6484 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6486 const TargetRegisterClass *TRC = isThumb2 ?
6487 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6488 (const TargetRegisterClass*)&ARM::GPRRegClass;
6489 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6493 // fallthrough --> loopMBB
6494 BB->addSuccessor(loopMBB);
6497 // ldrexd r2, r3, ptr
6498 // <binopa> r0, r2, incr
6499 // <binopb> r1, r3, incr
6500 // strexd storesuccess, r0, r1, ptr
6501 // cmp storesuccess, #0
6503 // fallthrough --> exitMBB
6509 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6510 .addReg(destlo, RegState::Define)
6511 .addReg(desthi, RegState::Define)
6514 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6515 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6516 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6517 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6518 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6519 .addReg(GPRPair0, 0, ARM::gsub_0);
6520 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6521 .addReg(GPRPair0, 0, ARM::gsub_1);
6525 unsigned StoreLo, StoreHi;
6528 for (unsigned i = 0; i < 2; i++) {
6529 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6531 .addReg(i == 0 ? destlo : desthi)
6532 .addReg(i == 0 ? vallo : valhi));
6533 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6534 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6535 BB->addSuccessor(exitMBB);
6536 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6537 BB = (i == 0 ? contBB : cont2BB);
6540 // Copy to physregs for strexd
6541 StoreLo = MI->getOperand(5).getReg();
6542 StoreHi = MI->getOperand(6).getReg();
6544 // Perform binary operation
6545 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6546 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6547 .addReg(destlo).addReg(vallo))
6548 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6549 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6550 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6551 .addReg(desthi).addReg(valhi))
6552 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6557 // Copy to physregs for strexd
6562 // Compare and branch to exit block.
6563 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6564 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6565 BB->addSuccessor(exitMBB);
6566 BB->addSuccessor(contBB);
6574 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6575 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6576 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6577 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6579 // Marshal a pair...
6580 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6581 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6582 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6583 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6584 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6587 .addImm(ARM::gsub_0);
6588 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6591 .addImm(ARM::gsub_1);
6594 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6595 .addReg(StorePair).addReg(ptr));
6598 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6599 .addReg(storesuccess).addImm(0));
6600 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6601 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6603 BB->addSuccessor(loopMBB);
6604 BB->addSuccessor(exitMBB);
6610 MI->eraseFromParent(); // The instruction is gone now.
6616 ARMTargetLowering::EmitAtomicLoad64(MachineInstr *MI, MachineBasicBlock *BB) const {
6618 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6620 unsigned destlo = MI->getOperand(0).getReg();
6621 unsigned desthi = MI->getOperand(1).getReg();
6622 unsigned ptr = MI->getOperand(2).getReg();
6623 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6624 DebugLoc dl = MI->getDebugLoc();
6625 bool isThumb2 = Subtarget->isThumb2();
6627 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6629 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6630 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6631 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6633 unsigned ldrOpc, strOpc;
6634 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6636 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(ldrOpc));
6639 MIB.addReg(destlo, RegState::Define)
6640 .addReg(desthi, RegState::Define)
6644 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6645 MIB.addReg(GPRPair0, RegState::Define).addReg(ptr);
6647 // Copy GPRPair0 into dest. (This copy will normally be coalesced.)
6648 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), destlo)
6649 .addReg(GPRPair0, 0, ARM::gsub_0);
6650 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), desthi)
6651 .addReg(GPRPair0, 0, ARM::gsub_1);
6653 AddDefaultPred(MIB);
6655 MI->eraseFromParent(); // The instruction is gone now.
6660 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6661 /// registers the function context.
6662 void ARMTargetLowering::
6663 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6664 MachineBasicBlock *DispatchBB, int FI) const {
6665 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6666 DebugLoc dl = MI->getDebugLoc();
6667 MachineFunction *MF = MBB->getParent();
6668 MachineRegisterInfo *MRI = &MF->getRegInfo();
6669 MachineConstantPool *MCP = MF->getConstantPool();
6670 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6671 const Function *F = MF->getFunction();
6673 bool isThumb = Subtarget->isThumb();
6674 bool isThumb2 = Subtarget->isThumb2();
6676 unsigned PCLabelId = AFI->createPICLabelUId();
6677 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6678 ARMConstantPoolValue *CPV =
6679 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6680 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6682 const TargetRegisterClass *TRC = isThumb ?
6683 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6684 (const TargetRegisterClass*)&ARM::GPRRegClass;
6686 // Grab constant pool and fixed stack memory operands.
6687 MachineMemOperand *CPMMO =
6688 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6689 MachineMemOperand::MOLoad, 4, 4);
6691 MachineMemOperand *FIMMOSt =
6692 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6693 MachineMemOperand::MOStore, 4, 4);
6695 // Load the address of the dispatch MBB into the jump buffer.
6697 // Incoming value: jbuf
6698 // ldr.n r5, LCPI1_1
6701 // str r5, [$jbuf, #+4] ; &jbuf[1]
6702 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6703 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6704 .addConstantPoolIndex(CPI)
6705 .addMemOperand(CPMMO));
6706 // Set the low bit because of thumb mode.
6707 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6709 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6710 .addReg(NewVReg1, RegState::Kill)
6712 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6713 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6714 .addReg(NewVReg2, RegState::Kill)
6716 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6717 .addReg(NewVReg3, RegState::Kill)
6719 .addImm(36) // &jbuf[1] :: pc
6720 .addMemOperand(FIMMOSt));
6721 } else if (isThumb) {
6722 // Incoming value: jbuf
6723 // ldr.n r1, LCPI1_4
6727 // add r2, $jbuf, #+4 ; &jbuf[1]
6729 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6730 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6731 .addConstantPoolIndex(CPI)
6732 .addMemOperand(CPMMO));
6733 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6734 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6735 .addReg(NewVReg1, RegState::Kill)
6737 // Set the low bit because of thumb mode.
6738 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6739 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6740 .addReg(ARM::CPSR, RegState::Define)
6742 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6743 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6744 .addReg(ARM::CPSR, RegState::Define)
6745 .addReg(NewVReg2, RegState::Kill)
6746 .addReg(NewVReg3, RegState::Kill));
6747 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6748 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6750 .addImm(36)); // &jbuf[1] :: pc
6751 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6752 .addReg(NewVReg4, RegState::Kill)
6753 .addReg(NewVReg5, RegState::Kill)
6755 .addMemOperand(FIMMOSt));
6757 // Incoming value: jbuf
6760 // str r1, [$jbuf, #+4] ; &jbuf[1]
6761 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6762 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6763 .addConstantPoolIndex(CPI)
6765 .addMemOperand(CPMMO));
6766 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6767 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6768 .addReg(NewVReg1, RegState::Kill)
6769 .addImm(PCLabelId));
6770 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6771 .addReg(NewVReg2, RegState::Kill)
6773 .addImm(36) // &jbuf[1] :: pc
6774 .addMemOperand(FIMMOSt));
6778 MachineBasicBlock *ARMTargetLowering::
6779 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6780 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6781 DebugLoc dl = MI->getDebugLoc();
6782 MachineFunction *MF = MBB->getParent();
6783 MachineRegisterInfo *MRI = &MF->getRegInfo();
6784 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6785 MachineFrameInfo *MFI = MF->getFrameInfo();
6786 int FI = MFI->getFunctionContextIndex();
6788 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6789 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6790 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6792 // Get a mapping of the call site numbers to all of the landing pads they're
6794 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6795 unsigned MaxCSNum = 0;
6796 MachineModuleInfo &MMI = MF->getMMI();
6797 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6799 if (!BB->isLandingPad()) continue;
6801 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6803 for (MachineBasicBlock::iterator
6804 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6805 if (!II->isEHLabel()) continue;
6807 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6808 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6810 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6811 for (SmallVectorImpl<unsigned>::iterator
6812 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6813 CSI != CSE; ++CSI) {
6814 CallSiteNumToLPad[*CSI].push_back(BB);
6815 MaxCSNum = std::max(MaxCSNum, *CSI);
6821 // Get an ordered list of the machine basic blocks for the jump table.
6822 std::vector<MachineBasicBlock*> LPadList;
6823 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6824 LPadList.reserve(CallSiteNumToLPad.size());
6825 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6826 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6827 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6828 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6829 LPadList.push_back(*II);
6830 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6834 assert(!LPadList.empty() &&
6835 "No landing pad destinations for the dispatch jump table!");
6837 // Create the jump table and associated information.
6838 MachineJumpTableInfo *JTI =
6839 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6840 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6841 unsigned UId = AFI->createJumpTableUId();
6842 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6844 // Create the MBBs for the dispatch code.
6846 // Shove the dispatch's address into the return slot in the function context.
6847 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6848 DispatchBB->setIsLandingPad();
6850 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6851 unsigned trap_opcode;
6852 if (Subtarget->isThumb())
6853 trap_opcode = ARM::tTRAP;
6855 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6857 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6858 DispatchBB->addSuccessor(TrapBB);
6860 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6861 DispatchBB->addSuccessor(DispContBB);
6864 MF->insert(MF->end(), DispatchBB);
6865 MF->insert(MF->end(), DispContBB);
6866 MF->insert(MF->end(), TrapBB);
6868 // Insert code into the entry block that creates and registers the function
6870 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6872 MachineMemOperand *FIMMOLd =
6873 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6874 MachineMemOperand::MOLoad |
6875 MachineMemOperand::MOVolatile, 4, 4);
6877 MachineInstrBuilder MIB;
6878 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6880 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6881 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6883 // Add a register mask with no preserved registers. This results in all
6884 // registers being marked as clobbered.
6885 MIB.addRegMask(RI.getNoPreservedMask());
6887 unsigned NumLPads = LPadList.size();
6888 if (Subtarget->isThumb2()) {
6889 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6890 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6893 .addMemOperand(FIMMOLd));
6895 if (NumLPads < 256) {
6896 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6898 .addImm(LPadList.size()));
6900 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6901 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6902 .addImm(NumLPads & 0xFFFF));
6904 unsigned VReg2 = VReg1;
6905 if ((NumLPads & 0xFFFF0000) != 0) {
6906 VReg2 = MRI->createVirtualRegister(TRC);
6907 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6909 .addImm(NumLPads >> 16));
6912 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6917 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6922 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6923 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6924 .addJumpTableIndex(MJTI)
6927 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6930 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6931 .addReg(NewVReg3, RegState::Kill)
6933 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6935 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6936 .addReg(NewVReg4, RegState::Kill)
6938 .addJumpTableIndex(MJTI)
6940 } else if (Subtarget->isThumb()) {
6941 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6942 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6945 .addMemOperand(FIMMOLd));
6947 if (NumLPads < 256) {
6948 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6952 MachineConstantPool *ConstantPool = MF->getConstantPool();
6953 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6954 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6956 // MachineConstantPool wants an explicit alignment.
6957 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6959 Align = getDataLayout()->getTypeAllocSize(C->getType());
6960 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6962 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6963 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6964 .addReg(VReg1, RegState::Define)
6965 .addConstantPoolIndex(Idx));
6966 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6971 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6976 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6977 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6978 .addReg(ARM::CPSR, RegState::Define)
6982 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6983 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6984 .addJumpTableIndex(MJTI)
6987 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6988 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6989 .addReg(ARM::CPSR, RegState::Define)
6990 .addReg(NewVReg2, RegState::Kill)
6993 MachineMemOperand *JTMMOLd =
6994 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6995 MachineMemOperand::MOLoad, 4, 4);
6997 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6998 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6999 .addReg(NewVReg4, RegState::Kill)
7001 .addMemOperand(JTMMOLd));
7003 unsigned NewVReg6 = NewVReg5;
7004 if (RelocM == Reloc::PIC_) {
7005 NewVReg6 = MRI->createVirtualRegister(TRC);
7006 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7007 .addReg(ARM::CPSR, RegState::Define)
7008 .addReg(NewVReg5, RegState::Kill)
7012 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7013 .addReg(NewVReg6, RegState::Kill)
7014 .addJumpTableIndex(MJTI)
7017 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7018 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7021 .addMemOperand(FIMMOLd));
7023 if (NumLPads < 256) {
7024 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7027 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7028 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7029 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7030 .addImm(NumLPads & 0xFFFF));
7032 unsigned VReg2 = VReg1;
7033 if ((NumLPads & 0xFFFF0000) != 0) {
7034 VReg2 = MRI->createVirtualRegister(TRC);
7035 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7037 .addImm(NumLPads >> 16));
7040 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7044 MachineConstantPool *ConstantPool = MF->getConstantPool();
7045 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7046 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7048 // MachineConstantPool wants an explicit alignment.
7049 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7051 Align = getDataLayout()->getTypeAllocSize(C->getType());
7052 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7054 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7055 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7056 .addReg(VReg1, RegState::Define)
7057 .addConstantPoolIndex(Idx)
7059 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7061 .addReg(VReg1, RegState::Kill));
7064 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7069 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7071 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7073 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7074 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7075 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7076 .addJumpTableIndex(MJTI)
7079 MachineMemOperand *JTMMOLd =
7080 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7081 MachineMemOperand::MOLoad, 4, 4);
7082 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7084 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7085 .addReg(NewVReg3, RegState::Kill)
7088 .addMemOperand(JTMMOLd));
7090 if (RelocM == Reloc::PIC_) {
7091 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7092 .addReg(NewVReg5, RegState::Kill)
7094 .addJumpTableIndex(MJTI)
7097 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7098 .addReg(NewVReg5, RegState::Kill)
7099 .addJumpTableIndex(MJTI)
7104 // Add the jump table entries as successors to the MBB.
7105 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7106 for (std::vector<MachineBasicBlock*>::iterator
7107 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7108 MachineBasicBlock *CurMBB = *I;
7109 if (SeenMBBs.insert(CurMBB))
7110 DispContBB->addSuccessor(CurMBB);
7113 // N.B. the order the invoke BBs are processed in doesn't matter here.
7114 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7115 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7116 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7117 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7118 MachineBasicBlock *BB = *I;
7120 // Remove the landing pad successor from the invoke block and replace it
7121 // with the new dispatch block.
7122 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7124 while (!Successors.empty()) {
7125 MachineBasicBlock *SMBB = Successors.pop_back_val();
7126 if (SMBB->isLandingPad()) {
7127 BB->removeSuccessor(SMBB);
7128 MBBLPads.push_back(SMBB);
7132 BB->addSuccessor(DispatchBB);
7134 // Find the invoke call and mark all of the callee-saved registers as
7135 // 'implicit defined' so that they're spilled. This prevents code from
7136 // moving instructions to before the EH block, where they will never be
7138 for (MachineBasicBlock::reverse_iterator
7139 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7140 if (!II->isCall()) continue;
7142 DenseMap<unsigned, bool> DefRegs;
7143 for (MachineInstr::mop_iterator
7144 OI = II->operands_begin(), OE = II->operands_end();
7146 if (!OI->isReg()) continue;
7147 DefRegs[OI->getReg()] = true;
7150 MachineInstrBuilder MIB(*MF, &*II);
7152 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7153 unsigned Reg = SavedRegs[i];
7154 if (Subtarget->isThumb2() &&
7155 !ARM::tGPRRegClass.contains(Reg) &&
7156 !ARM::hGPRRegClass.contains(Reg))
7158 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7160 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7163 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7170 // Mark all former landing pads as non-landing pads. The dispatch is the only
7172 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7173 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7174 (*I)->setIsLandingPad(false);
7176 // The instruction is gone now.
7177 MI->eraseFromParent();
7183 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7184 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7185 E = MBB->succ_end(); I != E; ++I)
7188 llvm_unreachable("Expecting a BB with two successors!");
7191 MachineBasicBlock *ARMTargetLowering::
7192 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
7193 // This pseudo instruction has 3 operands: dst, src, size
7194 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7195 // Otherwise, we will generate unrolled scalar copies.
7196 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7197 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7198 MachineFunction::iterator It = BB;
7201 unsigned dest = MI->getOperand(0).getReg();
7202 unsigned src = MI->getOperand(1).getReg();
7203 unsigned SizeVal = MI->getOperand(2).getImm();
7204 unsigned Align = MI->getOperand(3).getImm();
7205 DebugLoc dl = MI->getDebugLoc();
7207 bool isThumb2 = Subtarget->isThumb2();
7208 MachineFunction *MF = BB->getParent();
7209 MachineRegisterInfo &MRI = MF->getRegInfo();
7210 unsigned ldrOpc, strOpc, UnitSize = 0;
7212 const TargetRegisterClass *TRC = isThumb2 ?
7213 (const TargetRegisterClass*)&ARM::tGPRRegClass :
7214 (const TargetRegisterClass*)&ARM::GPRRegClass;
7215 const TargetRegisterClass *TRC_Vec = 0;
7218 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7219 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7221 } else if (Align & 2) {
7222 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
7223 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
7226 // Check whether we can use NEON instructions.
7227 if (!MF->getFunction()->getAttributes().
7228 hasAttribute(AttributeSet::FunctionIndex,
7229 Attribute::NoImplicitFloat) &&
7230 Subtarget->hasNEON()) {
7231 if ((Align % 16 == 0) && SizeVal >= 16) {
7232 ldrOpc = ARM::VLD1q32wb_fixed;
7233 strOpc = ARM::VST1q32wb_fixed;
7235 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
7237 else if ((Align % 8 == 0) && SizeVal >= 8) {
7238 ldrOpc = ARM::VLD1d32wb_fixed;
7239 strOpc = ARM::VST1d32wb_fixed;
7241 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
7244 // Can't use NEON instructions.
7245 if (UnitSize == 0) {
7246 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
7247 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
7252 unsigned BytesLeft = SizeVal % UnitSize;
7253 unsigned LoopSize = SizeVal - BytesLeft;
7255 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7256 // Use LDR and STR to copy.
7257 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7258 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7259 unsigned srcIn = src;
7260 unsigned destIn = dest;
7261 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7262 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7263 unsigned srcOut = MRI.createVirtualRegister(TRC);
7264 unsigned destOut = MRI.createVirtualRegister(TRC);
7265 if (UnitSize >= 8) {
7266 AddDefaultPred(BuildMI(*BB, MI, dl,
7267 TII->get(ldrOpc), scratch)
7268 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
7270 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7271 .addReg(destIn).addImm(0).addReg(scratch));
7272 } else if (isThumb2) {
7273 AddDefaultPred(BuildMI(*BB, MI, dl,
7274 TII->get(ldrOpc), scratch)
7275 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
7277 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7278 .addReg(scratch).addReg(destIn)
7281 AddDefaultPred(BuildMI(*BB, MI, dl,
7282 TII->get(ldrOpc), scratch)
7283 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
7286 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7287 .addReg(scratch).addReg(destIn)
7288 .addReg(0).addImm(UnitSize));
7294 // Handle the leftover bytes with LDRB and STRB.
7295 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7296 // [destOut] = STRB_POST(scratch, destIn, 1)
7297 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7298 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7299 for (unsigned i = 0; i < BytesLeft; i++) {
7300 unsigned scratch = MRI.createVirtualRegister(TRC);
7301 unsigned srcOut = MRI.createVirtualRegister(TRC);
7302 unsigned destOut = MRI.createVirtualRegister(TRC);
7304 AddDefaultPred(BuildMI(*BB, MI, dl,
7305 TII->get(ldrOpc),scratch)
7306 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7308 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7309 .addReg(scratch).addReg(destIn)
7310 .addReg(0).addImm(1));
7312 AddDefaultPred(BuildMI(*BB, MI, dl,
7313 TII->get(ldrOpc),scratch)
7314 .addReg(srcOut, RegState::Define).addReg(srcIn)
7315 .addReg(0).addImm(1));
7317 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
7318 .addReg(scratch).addReg(destIn)
7319 .addReg(0).addImm(1));
7324 MI->eraseFromParent(); // The instruction is gone now.
7328 // Expand the pseudo op to a loop.
7331 // movw varEnd, # --> with thumb2
7333 // ldrcp varEnd, idx --> without thumb2
7334 // fallthrough --> loopMBB
7336 // PHI varPhi, varEnd, varLoop
7337 // PHI srcPhi, src, srcLoop
7338 // PHI destPhi, dst, destLoop
7339 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7340 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7341 // subs varLoop, varPhi, #UnitSize
7343 // fallthrough --> exitMBB
7345 // epilogue to handle left-over bytes
7346 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7347 // [destOut] = STRB_POST(scratch, destLoop, 1)
7348 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7349 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7350 MF->insert(It, loopMBB);
7351 MF->insert(It, exitMBB);
7353 // Transfer the remainder of BB and its successor edges to exitMBB.
7354 exitMBB->splice(exitMBB->begin(), BB,
7355 llvm::next(MachineBasicBlock::iterator(MI)),
7357 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7359 // Load an immediate to varEnd.
7360 unsigned varEnd = MRI.createVirtualRegister(TRC);
7362 unsigned VReg1 = varEnd;
7363 if ((LoopSize & 0xFFFF0000) != 0)
7364 VReg1 = MRI.createVirtualRegister(TRC);
7365 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
7366 .addImm(LoopSize & 0xFFFF));
7368 if ((LoopSize & 0xFFFF0000) != 0)
7369 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7371 .addImm(LoopSize >> 16));
7373 MachineConstantPool *ConstantPool = MF->getConstantPool();
7374 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7375 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7377 // MachineConstantPool wants an explicit alignment.
7378 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7380 Align = getDataLayout()->getTypeAllocSize(C->getType());
7381 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7383 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
7384 .addReg(varEnd, RegState::Define)
7385 .addConstantPoolIndex(Idx)
7388 BB->addSuccessor(loopMBB);
7390 // Generate the loop body:
7391 // varPhi = PHI(varLoop, varEnd)
7392 // srcPhi = PHI(srcLoop, src)
7393 // destPhi = PHI(destLoop, dst)
7394 MachineBasicBlock *entryBB = BB;
7396 unsigned varLoop = MRI.createVirtualRegister(TRC);
7397 unsigned varPhi = MRI.createVirtualRegister(TRC);
7398 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7399 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7400 unsigned destLoop = MRI.createVirtualRegister(TRC);
7401 unsigned destPhi = MRI.createVirtualRegister(TRC);
7403 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7404 .addReg(varLoop).addMBB(loopMBB)
7405 .addReg(varEnd).addMBB(entryBB);
7406 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7407 .addReg(srcLoop).addMBB(loopMBB)
7408 .addReg(src).addMBB(entryBB);
7409 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7410 .addReg(destLoop).addMBB(loopMBB)
7411 .addReg(dest).addMBB(entryBB);
7413 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7414 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7415 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
7416 if (UnitSize >= 8) {
7417 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7418 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
7420 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7421 .addReg(destPhi).addImm(0).addReg(scratch));
7422 } else if (isThumb2) {
7423 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7424 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
7426 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7427 .addReg(scratch).addReg(destPhi)
7430 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
7431 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
7434 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
7435 .addReg(scratch).addReg(destPhi)
7436 .addReg(0).addImm(UnitSize));
7439 // Decrement loop variable by UnitSize.
7440 MachineInstrBuilder MIB = BuildMI(BB, dl,
7441 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7442 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7443 MIB->getOperand(5).setReg(ARM::CPSR);
7444 MIB->getOperand(5).setIsDef(true);
7446 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7447 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7449 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7450 BB->addSuccessor(loopMBB);
7451 BB->addSuccessor(exitMBB);
7453 // Add epilogue to handle BytesLeft.
7455 MachineInstr *StartOfExit = exitMBB->begin();
7456 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
7457 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
7459 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7460 // [destOut] = STRB_POST(scratch, destLoop, 1)
7461 unsigned srcIn = srcLoop;
7462 unsigned destIn = destLoop;
7463 for (unsigned i = 0; i < BytesLeft; i++) {
7464 unsigned scratch = MRI.createVirtualRegister(TRC);
7465 unsigned srcOut = MRI.createVirtualRegister(TRC);
7466 unsigned destOut = MRI.createVirtualRegister(TRC);
7468 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7469 TII->get(ldrOpc),scratch)
7470 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
7472 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7473 .addReg(scratch).addReg(destIn)
7476 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
7477 TII->get(ldrOpc),scratch)
7478 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
7480 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
7481 .addReg(scratch).addReg(destIn)
7482 .addReg(0).addImm(1));
7488 MI->eraseFromParent(); // The instruction is gone now.
7493 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7494 MachineBasicBlock *BB) const {
7495 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7496 DebugLoc dl = MI->getDebugLoc();
7497 bool isThumb2 = Subtarget->isThumb2();
7498 switch (MI->getOpcode()) {
7501 llvm_unreachable("Unexpected instr type to insert");
7503 // The Thumb2 pre-indexed stores have the same MI operands, they just
7504 // define them differently in the .td files from the isel patterns, so
7505 // they need pseudos.
7506 case ARM::t2STR_preidx:
7507 MI->setDesc(TII->get(ARM::t2STR_PRE));
7509 case ARM::t2STRB_preidx:
7510 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7512 case ARM::t2STRH_preidx:
7513 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7516 case ARM::STRi_preidx:
7517 case ARM::STRBi_preidx: {
7518 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7519 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7520 // Decode the offset.
7521 unsigned Offset = MI->getOperand(4).getImm();
7522 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7523 Offset = ARM_AM::getAM2Offset(Offset);
7527 MachineMemOperand *MMO = *MI->memoperands_begin();
7528 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7529 .addOperand(MI->getOperand(0)) // Rn_wb
7530 .addOperand(MI->getOperand(1)) // Rt
7531 .addOperand(MI->getOperand(2)) // Rn
7532 .addImm(Offset) // offset (skip GPR==zero_reg)
7533 .addOperand(MI->getOperand(5)) // pred
7534 .addOperand(MI->getOperand(6))
7535 .addMemOperand(MMO);
7536 MI->eraseFromParent();
7539 case ARM::STRr_preidx:
7540 case ARM::STRBr_preidx:
7541 case ARM::STRH_preidx: {
7543 switch (MI->getOpcode()) {
7544 default: llvm_unreachable("unexpected opcode!");
7545 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7546 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7547 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7549 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7550 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7551 MIB.addOperand(MI->getOperand(i));
7552 MI->eraseFromParent();
7555 case ARM::ATOMIC_LOAD_ADD_I8:
7556 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7557 case ARM::ATOMIC_LOAD_ADD_I16:
7558 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7559 case ARM::ATOMIC_LOAD_ADD_I32:
7560 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7562 case ARM::ATOMIC_LOAD_AND_I8:
7563 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7564 case ARM::ATOMIC_LOAD_AND_I16:
7565 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7566 case ARM::ATOMIC_LOAD_AND_I32:
7567 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7569 case ARM::ATOMIC_LOAD_OR_I8:
7570 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7571 case ARM::ATOMIC_LOAD_OR_I16:
7572 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7573 case ARM::ATOMIC_LOAD_OR_I32:
7574 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7576 case ARM::ATOMIC_LOAD_XOR_I8:
7577 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7578 case ARM::ATOMIC_LOAD_XOR_I16:
7579 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7580 case ARM::ATOMIC_LOAD_XOR_I32:
7581 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7583 case ARM::ATOMIC_LOAD_NAND_I8:
7584 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7585 case ARM::ATOMIC_LOAD_NAND_I16:
7586 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7587 case ARM::ATOMIC_LOAD_NAND_I32:
7588 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7590 case ARM::ATOMIC_LOAD_SUB_I8:
7591 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7592 case ARM::ATOMIC_LOAD_SUB_I16:
7593 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7594 case ARM::ATOMIC_LOAD_SUB_I32:
7595 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7597 case ARM::ATOMIC_LOAD_MIN_I8:
7598 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7599 case ARM::ATOMIC_LOAD_MIN_I16:
7600 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7601 case ARM::ATOMIC_LOAD_MIN_I32:
7602 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7604 case ARM::ATOMIC_LOAD_MAX_I8:
7605 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7606 case ARM::ATOMIC_LOAD_MAX_I16:
7607 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7608 case ARM::ATOMIC_LOAD_MAX_I32:
7609 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7611 case ARM::ATOMIC_LOAD_UMIN_I8:
7612 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7613 case ARM::ATOMIC_LOAD_UMIN_I16:
7614 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7615 case ARM::ATOMIC_LOAD_UMIN_I32:
7616 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7618 case ARM::ATOMIC_LOAD_UMAX_I8:
7619 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7620 case ARM::ATOMIC_LOAD_UMAX_I16:
7621 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7622 case ARM::ATOMIC_LOAD_UMAX_I32:
7623 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7625 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7626 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7627 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7629 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7630 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7631 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7633 case ARM::ATOMIC_LOAD_I64:
7634 return EmitAtomicLoad64(MI, BB);
7636 case ARM::ATOMIC_LOAD_ADD_I64:
7637 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7638 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7639 /*NeedsCarry*/ true);
7640 case ARM::ATOMIC_LOAD_SUB_I64:
7641 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7642 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7643 /*NeedsCarry*/ true);
7644 case ARM::ATOMIC_LOAD_OR_I64:
7645 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7646 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7647 case ARM::ATOMIC_LOAD_XOR_I64:
7648 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7649 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7650 case ARM::ATOMIC_LOAD_AND_I64:
7651 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7652 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7653 case ARM::ATOMIC_STORE_I64:
7654 case ARM::ATOMIC_SWAP_I64:
7655 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7656 case ARM::ATOMIC_CMP_SWAP_I64:
7657 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7658 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7659 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7660 case ARM::ATOMIC_LOAD_MIN_I64:
7661 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7662 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7663 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7664 /*IsMinMax*/ true, ARMCC::LT);
7665 case ARM::ATOMIC_LOAD_MAX_I64:
7666 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7667 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7668 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7669 /*IsMinMax*/ true, ARMCC::GE);
7670 case ARM::ATOMIC_LOAD_UMIN_I64:
7671 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7672 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7673 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7674 /*IsMinMax*/ true, ARMCC::LO);
7675 case ARM::ATOMIC_LOAD_UMAX_I64:
7676 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7677 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7678 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7679 /*IsMinMax*/ true, ARMCC::HS);
7681 case ARM::tMOVCCr_pseudo: {
7682 // To "insert" a SELECT_CC instruction, we actually have to insert the
7683 // diamond control-flow pattern. The incoming instruction knows the
7684 // destination vreg to set, the condition code register to branch on, the
7685 // true/false values to select between, and a branch opcode to use.
7686 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7687 MachineFunction::iterator It = BB;
7693 // cmpTY ccX, r1, r2
7695 // fallthrough --> copy0MBB
7696 MachineBasicBlock *thisMBB = BB;
7697 MachineFunction *F = BB->getParent();
7698 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7699 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7700 F->insert(It, copy0MBB);
7701 F->insert(It, sinkMBB);
7703 // Transfer the remainder of BB and its successor edges to sinkMBB.
7704 sinkMBB->splice(sinkMBB->begin(), BB,
7705 llvm::next(MachineBasicBlock::iterator(MI)),
7707 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7709 BB->addSuccessor(copy0MBB);
7710 BB->addSuccessor(sinkMBB);
7712 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7713 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7716 // %FalseValue = ...
7717 // # fallthrough to sinkMBB
7720 // Update machine-CFG edges
7721 BB->addSuccessor(sinkMBB);
7724 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7727 BuildMI(*BB, BB->begin(), dl,
7728 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7729 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7730 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7732 MI->eraseFromParent(); // The pseudo instruction is gone now.
7737 case ARM::BCCZi64: {
7738 // If there is an unconditional branch to the other successor, remove it.
7739 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7741 // Compare both parts that make up the double comparison separately for
7743 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7745 unsigned LHS1 = MI->getOperand(1).getReg();
7746 unsigned LHS2 = MI->getOperand(2).getReg();
7748 AddDefaultPred(BuildMI(BB, dl,
7749 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7750 .addReg(LHS1).addImm(0));
7751 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7752 .addReg(LHS2).addImm(0)
7753 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7755 unsigned RHS1 = MI->getOperand(3).getReg();
7756 unsigned RHS2 = MI->getOperand(4).getReg();
7757 AddDefaultPred(BuildMI(BB, dl,
7758 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7759 .addReg(LHS1).addReg(RHS1));
7760 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7761 .addReg(LHS2).addReg(RHS2)
7762 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7765 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7766 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7767 if (MI->getOperand(0).getImm() == ARMCC::NE)
7768 std::swap(destMBB, exitMBB);
7770 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7771 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7773 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7775 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7777 MI->eraseFromParent(); // The pseudo instruction is gone now.
7781 case ARM::Int_eh_sjlj_setjmp:
7782 case ARM::Int_eh_sjlj_setjmp_nofp:
7783 case ARM::tInt_eh_sjlj_setjmp:
7784 case ARM::t2Int_eh_sjlj_setjmp:
7785 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7786 EmitSjLjDispatchBlock(MI, BB);
7791 // To insert an ABS instruction, we have to insert the
7792 // diamond control-flow pattern. The incoming instruction knows the
7793 // source vreg to test against 0, the destination vreg to set,
7794 // the condition code register to branch on, the
7795 // true/false values to select between, and a branch opcode to use.
7800 // BCC (branch to SinkBB if V0 >= 0)
7801 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7802 // SinkBB: V1 = PHI(V2, V3)
7803 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7804 MachineFunction::iterator BBI = BB;
7806 MachineFunction *Fn = BB->getParent();
7807 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7808 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7809 Fn->insert(BBI, RSBBB);
7810 Fn->insert(BBI, SinkBB);
7812 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7813 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7814 bool isThumb2 = Subtarget->isThumb2();
7815 MachineRegisterInfo &MRI = Fn->getRegInfo();
7816 // In Thumb mode S must not be specified if source register is the SP or
7817 // PC and if destination register is the SP, so restrict register class
7818 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7819 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7820 (const TargetRegisterClass*)&ARM::GPRRegClass);
7822 // Transfer the remainder of BB and its successor edges to sinkMBB.
7823 SinkBB->splice(SinkBB->begin(), BB,
7824 llvm::next(MachineBasicBlock::iterator(MI)),
7826 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7828 BB->addSuccessor(RSBBB);
7829 BB->addSuccessor(SinkBB);
7831 // fall through to SinkMBB
7832 RSBBB->addSuccessor(SinkBB);
7834 // insert a cmp at the end of BB
7835 AddDefaultPred(BuildMI(BB, dl,
7836 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7837 .addReg(ABSSrcReg).addImm(0));
7839 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7841 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7842 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7844 // insert rsbri in RSBBB
7845 // Note: BCC and rsbri will be converted into predicated rsbmi
7846 // by if-conversion pass
7847 BuildMI(*RSBBB, RSBBB->begin(), dl,
7848 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7849 .addReg(ABSSrcReg, RegState::Kill)
7850 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7852 // insert PHI in SinkBB,
7853 // reuse ABSDstReg to not change uses of ABS instruction
7854 BuildMI(*SinkBB, SinkBB->begin(), dl,
7855 TII->get(ARM::PHI), ABSDstReg)
7856 .addReg(NewRsbDstReg).addMBB(RSBBB)
7857 .addReg(ABSSrcReg).addMBB(BB);
7859 // remove ABS instruction
7860 MI->eraseFromParent();
7862 // return last added BB
7865 case ARM::COPY_STRUCT_BYVAL_I32:
7867 return EmitStructByval(MI, BB);
7871 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7872 SDNode *Node) const {
7873 if (!MI->hasPostISelHook()) {
7874 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7875 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7879 const MCInstrDesc *MCID = &MI->getDesc();
7880 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7881 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7882 // operand is still set to noreg. If needed, set the optional operand's
7883 // register to CPSR, and remove the redundant implicit def.
7885 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7887 // Rename pseudo opcodes.
7888 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7890 const ARMBaseInstrInfo *TII =
7891 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7892 MCID = &TII->get(NewOpc);
7894 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7895 "converted opcode should be the same except for cc_out");
7899 // Add the optional cc_out operand
7900 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7902 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7904 // Any ARM instruction that sets the 's' bit should specify an optional
7905 // "cc_out" operand in the last operand position.
7906 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7907 assert(!NewOpc && "Optional cc_out operand required");
7910 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7911 // since we already have an optional CPSR def.
7912 bool definesCPSR = false;
7913 bool deadCPSR = false;
7914 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7916 const MachineOperand &MO = MI->getOperand(i);
7917 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7921 MI->RemoveOperand(i);
7926 assert(!NewOpc && "Optional cc_out operand required");
7929 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7931 assert(!MI->getOperand(ccOutIdx).getReg() &&
7932 "expect uninitialized optional cc_out operand");
7936 // If this instruction was defined with an optional CPSR def and its dag node
7937 // had a live implicit CPSR def, then activate the optional CPSR def.
7938 MachineOperand &MO = MI->getOperand(ccOutIdx);
7939 MO.setReg(ARM::CPSR);
7943 //===----------------------------------------------------------------------===//
7944 // ARM Optimization Hooks
7945 //===----------------------------------------------------------------------===//
7947 // Helper function that checks if N is a null or all ones constant.
7948 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7949 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7952 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7955 // Return true if N is conditionally 0 or all ones.
7956 // Detects these expressions where cc is an i1 value:
7958 // (select cc 0, y) [AllOnes=0]
7959 // (select cc y, 0) [AllOnes=0]
7960 // (zext cc) [AllOnes=0]
7961 // (sext cc) [AllOnes=0/1]
7962 // (select cc -1, y) [AllOnes=1]
7963 // (select cc y, -1) [AllOnes=1]
7965 // Invert is set when N is the null/all ones constant when CC is false.
7966 // OtherOp is set to the alternative value of N.
7967 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7968 SDValue &CC, bool &Invert,
7970 SelectionDAG &DAG) {
7971 switch (N->getOpcode()) {
7972 default: return false;
7974 CC = N->getOperand(0);
7975 SDValue N1 = N->getOperand(1);
7976 SDValue N2 = N->getOperand(2);
7977 if (isZeroOrAllOnes(N1, AllOnes)) {
7982 if (isZeroOrAllOnes(N2, AllOnes)) {
7989 case ISD::ZERO_EXTEND:
7990 // (zext cc) can never be the all ones value.
7994 case ISD::SIGN_EXTEND: {
7995 EVT VT = N->getValueType(0);
7996 CC = N->getOperand(0);
7997 if (CC.getValueType() != MVT::i1)
8001 // When looking for an AllOnes constant, N is an sext, and the 'other'
8003 OtherOp = DAG.getConstant(0, VT);
8004 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8005 // When looking for a 0 constant, N can be zext or sext.
8006 OtherOp = DAG.getConstant(1, VT);
8008 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
8014 // Combine a constant select operand into its use:
8016 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8017 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8018 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8019 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8020 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8022 // The transform is rejected if the select doesn't have a constant operand that
8023 // is null, or all ones when AllOnes is set.
8025 // Also recognize sext/zext from i1:
8027 // (add (zext cc), x) -> (select cc (add x, 1), x)
8028 // (add (sext cc), x) -> (select cc (add x, -1), x)
8030 // These transformations eventually create predicated instructions.
8032 // @param N The node to transform.
8033 // @param Slct The N operand that is a select.
8034 // @param OtherOp The other N operand (x above).
8035 // @param DCI Context.
8036 // @param AllOnes Require the select constant to be all ones instead of null.
8037 // @returns The new node, or SDValue() on failure.
8039 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8040 TargetLowering::DAGCombinerInfo &DCI,
8041 bool AllOnes = false) {
8042 SelectionDAG &DAG = DCI.DAG;
8043 EVT VT = N->getValueType(0);
8044 SDValue NonConstantVal;
8047 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8048 NonConstantVal, DAG))
8051 // Slct is now know to be the desired identity constant when CC is true.
8052 SDValue TrueVal = OtherOp;
8053 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8054 OtherOp, NonConstantVal);
8055 // Unless SwapSelectOps says CC should be false.
8057 std::swap(TrueVal, FalseVal);
8059 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8060 CCOp, TrueVal, FalseVal);
8063 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8065 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8066 TargetLowering::DAGCombinerInfo &DCI) {
8067 SDValue N0 = N->getOperand(0);
8068 SDValue N1 = N->getOperand(1);
8069 if (N0.getNode()->hasOneUse()) {
8070 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8071 if (Result.getNode())
8074 if (N1.getNode()->hasOneUse()) {
8075 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8076 if (Result.getNode())
8082 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8083 // (only after legalization).
8084 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8085 TargetLowering::DAGCombinerInfo &DCI,
8086 const ARMSubtarget *Subtarget) {
8088 // Only perform optimization if after legalize, and if NEON is available. We
8089 // also expected both operands to be BUILD_VECTORs.
8090 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8091 || N0.getOpcode() != ISD::BUILD_VECTOR
8092 || N1.getOpcode() != ISD::BUILD_VECTOR)
8095 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8096 EVT VT = N->getValueType(0);
8097 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8100 // Check that the vector operands are of the right form.
8101 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8102 // operands, where N is the size of the formed vector.
8103 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8104 // index such that we have a pair wise add pattern.
8106 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8107 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8109 SDValue Vec = N0->getOperand(0)->getOperand(0);
8110 SDNode *V = Vec.getNode();
8111 unsigned nextIndex = 0;
8113 // For each operands to the ADD which are BUILD_VECTORs,
8114 // check to see if each of their operands are an EXTRACT_VECTOR with
8115 // the same vector and appropriate index.
8116 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8117 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8118 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8120 SDValue ExtVec0 = N0->getOperand(i);
8121 SDValue ExtVec1 = N1->getOperand(i);
8123 // First operand is the vector, verify its the same.
8124 if (V != ExtVec0->getOperand(0).getNode() ||
8125 V != ExtVec1->getOperand(0).getNode())
8128 // Second is the constant, verify its correct.
8129 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8130 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8132 // For the constant, we want to see all the even or all the odd.
8133 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8134 || C1->getZExtValue() != nextIndex+1)
8143 // Create VPADDL node.
8144 SelectionDAG &DAG = DCI.DAG;
8145 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8147 // Build operand list.
8148 SmallVector<SDValue, 8> Ops;
8149 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8150 TLI.getPointerTy()));
8152 // Input is the vector.
8155 // Get widened type and narrowed type.
8157 unsigned numElem = VT.getVectorNumElements();
8158 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8159 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8160 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8161 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8163 llvm_unreachable("Invalid vector element type for padd optimization.");
8166 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8167 widenType, &Ops[0], Ops.size());
8168 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8171 static SDValue findMUL_LOHI(SDValue V) {
8172 if (V->getOpcode() == ISD::UMUL_LOHI ||
8173 V->getOpcode() == ISD::SMUL_LOHI)
8178 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8179 TargetLowering::DAGCombinerInfo &DCI,
8180 const ARMSubtarget *Subtarget) {
8182 if (Subtarget->isThumb1Only()) return SDValue();
8184 // Only perform the checks after legalize when the pattern is available.
8185 if (DCI.isBeforeLegalize()) return SDValue();
8187 // Look for multiply add opportunities.
8188 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8189 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8190 // a glue link from the first add to the second add.
8191 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8192 // a S/UMLAL instruction.
8195 // \ / \ [no multiline comment]
8201 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8202 SDValue AddcOp0 = AddcNode->getOperand(0);
8203 SDValue AddcOp1 = AddcNode->getOperand(1);
8205 // Check if the two operands are from the same mul_lohi node.
8206 if (AddcOp0.getNode() == AddcOp1.getNode())
8209 assert(AddcNode->getNumValues() == 2 &&
8210 AddcNode->getValueType(0) == MVT::i32 &&
8211 "Expect ADDC with two result values. First: i32");
8213 // Check that we have a glued ADDC node.
8214 if (AddcNode->getValueType(1) != MVT::Glue)
8217 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8218 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8219 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8220 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8221 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8224 // Look for the glued ADDE.
8225 SDNode* AddeNode = AddcNode->getGluedUser();
8226 if (AddeNode == NULL)
8229 // Make sure it is really an ADDE.
8230 if (AddeNode->getOpcode() != ISD::ADDE)
8233 assert(AddeNode->getNumOperands() == 3 &&
8234 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8235 "ADDE node has the wrong inputs");
8237 // Check for the triangle shape.
8238 SDValue AddeOp0 = AddeNode->getOperand(0);
8239 SDValue AddeOp1 = AddeNode->getOperand(1);
8241 // Make sure that the ADDE operands are not coming from the same node.
8242 if (AddeOp0.getNode() == AddeOp1.getNode())
8245 // Find the MUL_LOHI node walking up ADDE's operands.
8246 bool IsLeftOperandMUL = false;
8247 SDValue MULOp = findMUL_LOHI(AddeOp0);
8248 if (MULOp == SDValue())
8249 MULOp = findMUL_LOHI(AddeOp1);
8251 IsLeftOperandMUL = true;
8252 if (MULOp == SDValue())
8255 // Figure out the right opcode.
8256 unsigned Opc = MULOp->getOpcode();
8257 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8259 // Figure out the high and low input values to the MLAL node.
8260 SDValue* HiMul = &MULOp;
8261 SDValue* HiAdd = NULL;
8262 SDValue* LoMul = NULL;
8263 SDValue* LowAdd = NULL;
8265 if (IsLeftOperandMUL)
8271 if (AddcOp0->getOpcode() == Opc) {
8275 if (AddcOp1->getOpcode() == Opc) {
8283 if (LoMul->getNode() != HiMul->getNode())
8286 // Create the merged node.
8287 SelectionDAG &DAG = DCI.DAG;
8289 // Build operand list.
8290 SmallVector<SDValue, 8> Ops;
8291 Ops.push_back(LoMul->getOperand(0));
8292 Ops.push_back(LoMul->getOperand(1));
8293 Ops.push_back(*LowAdd);
8294 Ops.push_back(*HiAdd);
8296 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8297 DAG.getVTList(MVT::i32, MVT::i32),
8298 &Ops[0], Ops.size());
8300 // Replace the ADDs' nodes uses by the MLA node's values.
8301 SDValue HiMLALResult(MLALNode.getNode(), 1);
8302 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8304 SDValue LoMLALResult(MLALNode.getNode(), 0);
8305 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8307 // Return original node to notify the driver to stop replacing.
8308 SDValue resNode(AddcNode, 0);
8312 /// PerformADDCCombine - Target-specific dag combine transform from
8313 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8314 static SDValue PerformADDCCombine(SDNode *N,
8315 TargetLowering::DAGCombinerInfo &DCI,
8316 const ARMSubtarget *Subtarget) {
8318 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8322 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8323 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8324 /// called with the default operands, and if that fails, with commuted
8326 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8327 TargetLowering::DAGCombinerInfo &DCI,
8328 const ARMSubtarget *Subtarget){
8330 // Attempt to create vpaddl for this add.
8331 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8332 if (Result.getNode())
8335 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8336 if (N0.getNode()->hasOneUse()) {
8337 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8338 if (Result.getNode()) return Result;
8343 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8345 static SDValue PerformADDCombine(SDNode *N,
8346 TargetLowering::DAGCombinerInfo &DCI,
8347 const ARMSubtarget *Subtarget) {
8348 SDValue N0 = N->getOperand(0);
8349 SDValue N1 = N->getOperand(1);
8351 // First try with the default operand order.
8352 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8353 if (Result.getNode())
8356 // If that didn't work, try again with the operands commuted.
8357 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8360 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8362 static SDValue PerformSUBCombine(SDNode *N,
8363 TargetLowering::DAGCombinerInfo &DCI) {
8364 SDValue N0 = N->getOperand(0);
8365 SDValue N1 = N->getOperand(1);
8367 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8368 if (N1.getNode()->hasOneUse()) {
8369 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8370 if (Result.getNode()) return Result;
8376 /// PerformVMULCombine
8377 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8378 /// special multiplier accumulator forwarding.
8384 // However, for (A + B) * (A + B),
8391 static SDValue PerformVMULCombine(SDNode *N,
8392 TargetLowering::DAGCombinerInfo &DCI,
8393 const ARMSubtarget *Subtarget) {
8394 if (!Subtarget->hasVMLxForwarding())
8397 SelectionDAG &DAG = DCI.DAG;
8398 SDValue N0 = N->getOperand(0);
8399 SDValue N1 = N->getOperand(1);
8400 unsigned Opcode = N0.getOpcode();
8401 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8402 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8403 Opcode = N1.getOpcode();
8404 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8405 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8413 EVT VT = N->getValueType(0);
8415 SDValue N00 = N0->getOperand(0);
8416 SDValue N01 = N0->getOperand(1);
8417 return DAG.getNode(Opcode, DL, VT,
8418 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8419 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8422 static SDValue PerformMULCombine(SDNode *N,
8423 TargetLowering::DAGCombinerInfo &DCI,
8424 const ARMSubtarget *Subtarget) {
8425 SelectionDAG &DAG = DCI.DAG;
8427 if (Subtarget->isThumb1Only())
8430 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8433 EVT VT = N->getValueType(0);
8434 if (VT.is64BitVector() || VT.is128BitVector())
8435 return PerformVMULCombine(N, DCI, Subtarget);
8439 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8443 int64_t MulAmt = C->getSExtValue();
8444 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8446 ShiftAmt = ShiftAmt & (32 - 1);
8447 SDValue V = N->getOperand(0);
8451 MulAmt >>= ShiftAmt;
8454 if (isPowerOf2_32(MulAmt - 1)) {
8455 // (mul x, 2^N + 1) => (add (shl x, N), x)
8456 Res = DAG.getNode(ISD::ADD, DL, VT,
8458 DAG.getNode(ISD::SHL, DL, VT,
8460 DAG.getConstant(Log2_32(MulAmt - 1),
8462 } else if (isPowerOf2_32(MulAmt + 1)) {
8463 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8464 Res = DAG.getNode(ISD::SUB, DL, VT,
8465 DAG.getNode(ISD::SHL, DL, VT,
8467 DAG.getConstant(Log2_32(MulAmt + 1),
8473 uint64_t MulAmtAbs = -MulAmt;
8474 if (isPowerOf2_32(MulAmtAbs + 1)) {
8475 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8476 Res = DAG.getNode(ISD::SUB, DL, VT,
8478 DAG.getNode(ISD::SHL, DL, VT,
8480 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8482 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8483 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8484 Res = DAG.getNode(ISD::ADD, DL, VT,
8486 DAG.getNode(ISD::SHL, DL, VT,
8488 DAG.getConstant(Log2_32(MulAmtAbs-1),
8490 Res = DAG.getNode(ISD::SUB, DL, VT,
8491 DAG.getConstant(0, MVT::i32),Res);
8498 Res = DAG.getNode(ISD::SHL, DL, VT,
8499 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8501 // Do not add new nodes to DAG combiner worklist.
8502 DCI.CombineTo(N, Res, false);
8506 static SDValue PerformANDCombine(SDNode *N,
8507 TargetLowering::DAGCombinerInfo &DCI,
8508 const ARMSubtarget *Subtarget) {
8510 // Attempt to use immediate-form VBIC
8511 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8513 EVT VT = N->getValueType(0);
8514 SelectionDAG &DAG = DCI.DAG;
8516 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8519 APInt SplatBits, SplatUndef;
8520 unsigned SplatBitSize;
8523 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8524 if (SplatBitSize <= 64) {
8526 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8527 SplatUndef.getZExtValue(), SplatBitSize,
8528 DAG, VbicVT, VT.is128BitVector(),
8530 if (Val.getNode()) {
8532 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8533 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8534 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8539 if (!Subtarget->isThumb1Only()) {
8540 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8541 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8542 if (Result.getNode())
8549 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8550 static SDValue PerformORCombine(SDNode *N,
8551 TargetLowering::DAGCombinerInfo &DCI,
8552 const ARMSubtarget *Subtarget) {
8553 // Attempt to use immediate-form VORR
8554 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8556 EVT VT = N->getValueType(0);
8557 SelectionDAG &DAG = DCI.DAG;
8559 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8562 APInt SplatBits, SplatUndef;
8563 unsigned SplatBitSize;
8565 if (BVN && Subtarget->hasNEON() &&
8566 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8567 if (SplatBitSize <= 64) {
8569 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8570 SplatUndef.getZExtValue(), SplatBitSize,
8571 DAG, VorrVT, VT.is128BitVector(),
8573 if (Val.getNode()) {
8575 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8576 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8577 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8582 if (!Subtarget->isThumb1Only()) {
8583 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8584 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8585 if (Result.getNode())
8589 // The code below optimizes (or (and X, Y), Z).
8590 // The AND operand needs to have a single user to make these optimizations
8592 SDValue N0 = N->getOperand(0);
8593 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8595 SDValue N1 = N->getOperand(1);
8597 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8598 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8599 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8601 unsigned SplatBitSize;
8604 APInt SplatBits0, SplatBits1;
8605 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8606 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8607 // Ensure that the second operand of both ands are constants
8608 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8609 HasAnyUndefs) && !HasAnyUndefs) {
8610 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8611 HasAnyUndefs) && !HasAnyUndefs) {
8612 // Ensure that the bit width of the constants are the same and that
8613 // the splat arguments are logical inverses as per the pattern we
8614 // are trying to simplify.
8615 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8616 SplatBits0 == ~SplatBits1) {
8617 // Canonicalize the vector type to make instruction selection
8619 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8620 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8624 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8630 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8633 // BFI is only available on V6T2+
8634 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8638 // 1) or (and A, mask), val => ARMbfi A, val, mask
8639 // iff (val & mask) == val
8641 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8642 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8643 // && mask == ~mask2
8644 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8645 // && ~mask == mask2
8646 // (i.e., copy a bitfield value into another bitfield of the same width)
8651 SDValue N00 = N0.getOperand(0);
8653 // The value and the mask need to be constants so we can verify this is
8654 // actually a bitfield set. If the mask is 0xffff, we can do better
8655 // via a movt instruction, so don't use BFI in that case.
8656 SDValue MaskOp = N0.getOperand(1);
8657 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8660 unsigned Mask = MaskC->getZExtValue();
8664 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8665 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8667 unsigned Val = N1C->getZExtValue();
8668 if ((Val & ~Mask) != Val)
8671 if (ARM::isBitFieldInvertedMask(Mask)) {
8672 Val >>= countTrailingZeros(~Mask);
8674 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8675 DAG.getConstant(Val, MVT::i32),
8676 DAG.getConstant(Mask, MVT::i32));
8678 // Do not add new nodes to DAG combiner worklist.
8679 DCI.CombineTo(N, Res, false);
8682 } else if (N1.getOpcode() == ISD::AND) {
8683 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8684 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8687 unsigned Mask2 = N11C->getZExtValue();
8689 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8691 if (ARM::isBitFieldInvertedMask(Mask) &&
8693 // The pack halfword instruction works better for masks that fit it,
8694 // so use that when it's available.
8695 if (Subtarget->hasT2ExtractPack() &&
8696 (Mask == 0xffff || Mask == 0xffff0000))
8699 unsigned amt = countTrailingZeros(Mask2);
8700 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8701 DAG.getConstant(amt, MVT::i32));
8702 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8703 DAG.getConstant(Mask, MVT::i32));
8704 // Do not add new nodes to DAG combiner worklist.
8705 DCI.CombineTo(N, Res, false);
8707 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8709 // The pack halfword instruction works better for masks that fit it,
8710 // so use that when it's available.
8711 if (Subtarget->hasT2ExtractPack() &&
8712 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8715 unsigned lsb = countTrailingZeros(Mask);
8716 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8717 DAG.getConstant(lsb, MVT::i32));
8718 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8719 DAG.getConstant(Mask2, MVT::i32));
8720 // Do not add new nodes to DAG combiner worklist.
8721 DCI.CombineTo(N, Res, false);
8726 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8727 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8728 ARM::isBitFieldInvertedMask(~Mask)) {
8729 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8730 // where lsb(mask) == #shamt and masked bits of B are known zero.
8731 SDValue ShAmt = N00.getOperand(1);
8732 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8733 unsigned LSB = countTrailingZeros(Mask);
8737 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8738 DAG.getConstant(~Mask, MVT::i32));
8740 // Do not add new nodes to DAG combiner worklist.
8741 DCI.CombineTo(N, Res, false);
8747 static SDValue PerformXORCombine(SDNode *N,
8748 TargetLowering::DAGCombinerInfo &DCI,
8749 const ARMSubtarget *Subtarget) {
8750 EVT VT = N->getValueType(0);
8751 SelectionDAG &DAG = DCI.DAG;
8753 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8756 if (!Subtarget->isThumb1Only()) {
8757 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8758 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8759 if (Result.getNode())
8766 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8767 /// the bits being cleared by the AND are not demanded by the BFI.
8768 static SDValue PerformBFICombine(SDNode *N,
8769 TargetLowering::DAGCombinerInfo &DCI) {
8770 SDValue N1 = N->getOperand(1);
8771 if (N1.getOpcode() == ISD::AND) {
8772 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8775 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8776 unsigned LSB = countTrailingZeros(~InvMask);
8777 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8778 unsigned Mask = (1 << Width)-1;
8779 unsigned Mask2 = N11C->getZExtValue();
8780 if ((Mask & (~Mask2)) == 0)
8781 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8782 N->getOperand(0), N1.getOperand(0),
8788 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8789 /// ARMISD::VMOVRRD.
8790 static SDValue PerformVMOVRRDCombine(SDNode *N,
8791 TargetLowering::DAGCombinerInfo &DCI) {
8792 // vmovrrd(vmovdrr x, y) -> x,y
8793 SDValue InDouble = N->getOperand(0);
8794 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8795 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8797 // vmovrrd(load f64) -> (load i32), (load i32)
8798 SDNode *InNode = InDouble.getNode();
8799 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8800 InNode->getValueType(0) == MVT::f64 &&
8801 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8802 !cast<LoadSDNode>(InNode)->isVolatile()) {
8803 // TODO: Should this be done for non-FrameIndex operands?
8804 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8806 SelectionDAG &DAG = DCI.DAG;
8808 SDValue BasePtr = LD->getBasePtr();
8809 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8810 LD->getPointerInfo(), LD->isVolatile(),
8811 LD->isNonTemporal(), LD->isInvariant(),
8812 LD->getAlignment());
8814 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8815 DAG.getConstant(4, MVT::i32));
8816 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8817 LD->getPointerInfo(), LD->isVolatile(),
8818 LD->isNonTemporal(), LD->isInvariant(),
8819 std::min(4U, LD->getAlignment() / 2));
8821 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8822 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8823 DCI.RemoveFromWorklist(LD);
8831 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8832 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8833 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8834 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8835 SDValue Op0 = N->getOperand(0);
8836 SDValue Op1 = N->getOperand(1);
8837 if (Op0.getOpcode() == ISD::BITCAST)
8838 Op0 = Op0.getOperand(0);
8839 if (Op1.getOpcode() == ISD::BITCAST)
8840 Op1 = Op1.getOperand(0);
8841 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8842 Op0.getNode() == Op1.getNode() &&
8843 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8844 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8845 N->getValueType(0), Op0.getOperand(0));
8849 /// PerformSTORECombine - Target-specific dag combine xforms for
8851 static SDValue PerformSTORECombine(SDNode *N,
8852 TargetLowering::DAGCombinerInfo &DCI) {
8853 StoreSDNode *St = cast<StoreSDNode>(N);
8854 if (St->isVolatile())
8857 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8858 // pack all of the elements in one place. Next, store to memory in fewer
8860 SDValue StVal = St->getValue();
8861 EVT VT = StVal.getValueType();
8862 if (St->isTruncatingStore() && VT.isVector()) {
8863 SelectionDAG &DAG = DCI.DAG;
8864 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8865 EVT StVT = St->getMemoryVT();
8866 unsigned NumElems = VT.getVectorNumElements();
8867 assert(StVT != VT && "Cannot truncate to the same type");
8868 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8869 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8871 // From, To sizes and ElemCount must be pow of two
8872 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8874 // We are going to use the original vector elt for storing.
8875 // Accumulated smaller vector elements must be a multiple of the store size.
8876 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8878 unsigned SizeRatio = FromEltSz / ToEltSz;
8879 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8881 // Create a type on which we perform the shuffle.
8882 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8883 NumElems*SizeRatio);
8884 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8887 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8888 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8889 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8891 // Can't shuffle using an illegal type.
8892 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8894 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8895 DAG.getUNDEF(WideVec.getValueType()),
8897 // At this point all of the data is stored at the bottom of the
8898 // register. We now need to save it to mem.
8900 // Find the largest store unit
8901 MVT StoreType = MVT::i8;
8902 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8903 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8904 MVT Tp = (MVT::SimpleValueType)tp;
8905 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8908 // Didn't find a legal store type.
8909 if (!TLI.isTypeLegal(StoreType))
8912 // Bitcast the original vector into a vector of store-size units
8913 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8914 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8915 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8916 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8917 SmallVector<SDValue, 8> Chains;
8918 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8919 TLI.getPointerTy());
8920 SDValue BasePtr = St->getBasePtr();
8922 // Perform one or more big stores into memory.
8923 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8924 for (unsigned I = 0; I < E; I++) {
8925 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8926 StoreType, ShuffWide,
8927 DAG.getIntPtrConstant(I));
8928 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8929 St->getPointerInfo(), St->isVolatile(),
8930 St->isNonTemporal(), St->getAlignment());
8931 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8933 Chains.push_back(Ch);
8935 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8939 if (!ISD::isNormalStore(St))
8942 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8943 // ARM stores of arguments in the same cache line.
8944 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8945 StVal.getNode()->hasOneUse()) {
8946 SelectionDAG &DAG = DCI.DAG;
8948 SDValue BasePtr = St->getBasePtr();
8949 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8950 StVal.getNode()->getOperand(0), BasePtr,
8951 St->getPointerInfo(), St->isVolatile(),
8952 St->isNonTemporal(), St->getAlignment());
8954 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8955 DAG.getConstant(4, MVT::i32));
8956 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8957 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8958 St->isNonTemporal(),
8959 std::min(4U, St->getAlignment() / 2));
8962 if (StVal.getValueType() != MVT::i64 ||
8963 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8966 // Bitcast an i64 store extracted from a vector to f64.
8967 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8968 SelectionDAG &DAG = DCI.DAG;
8970 SDValue IntVec = StVal.getOperand(0);
8971 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8972 IntVec.getValueType().getVectorNumElements());
8973 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8974 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8975 Vec, StVal.getOperand(1));
8977 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8978 // Make the DAGCombiner fold the bitcasts.
8979 DCI.AddToWorklist(Vec.getNode());
8980 DCI.AddToWorklist(ExtElt.getNode());
8981 DCI.AddToWorklist(V.getNode());
8982 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8983 St->getPointerInfo(), St->isVolatile(),
8984 St->isNonTemporal(), St->getAlignment(),
8988 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8989 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8990 /// i64 vector to have f64 elements, since the value can then be loaded
8991 /// directly into a VFP register.
8992 static bool hasNormalLoadOperand(SDNode *N) {
8993 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8994 for (unsigned i = 0; i < NumElts; ++i) {
8995 SDNode *Elt = N->getOperand(i).getNode();
8996 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9002 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9003 /// ISD::BUILD_VECTOR.
9004 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9005 TargetLowering::DAGCombinerInfo &DCI){
9006 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9007 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9008 // into a pair of GPRs, which is fine when the value is used as a scalar,
9009 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9010 SelectionDAG &DAG = DCI.DAG;
9011 if (N->getNumOperands() == 2) {
9012 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9017 // Load i64 elements as f64 values so that type legalization does not split
9018 // them up into i32 values.
9019 EVT VT = N->getValueType(0);
9020 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9023 SmallVector<SDValue, 8> Ops;
9024 unsigned NumElts = VT.getVectorNumElements();
9025 for (unsigned i = 0; i < NumElts; ++i) {
9026 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9028 // Make the DAGCombiner fold the bitcast.
9029 DCI.AddToWorklist(V.getNode());
9031 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9032 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
9033 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9036 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9038 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9039 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9040 // At that time, we may have inserted bitcasts from integer to float.
9041 // If these bitcasts have survived DAGCombine, change the lowering of this
9042 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9043 // force to use floating point types.
9045 // Make sure we can change the type of the vector.
9046 // This is possible iff:
9047 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9048 // 1.1. Vector is used only once.
9049 // 1.2. Use is a bit convert to an integer type.
9050 // 2. The size of its operands are 32-bits (64-bits are not legal).
9051 EVT VT = N->getValueType(0);
9052 EVT EltVT = VT.getVectorElementType();
9054 // Check 1.1. and 2.
9055 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9058 // By construction, the input type must be float.
9059 assert(EltVT == MVT::f32 && "Unexpected type!");
9062 SDNode *Use = *N->use_begin();
9063 if (Use->getOpcode() != ISD::BITCAST ||
9064 Use->getValueType(0).isFloatingPoint())
9067 // Check profitability.
9068 // Model is, if more than half of the relevant operands are bitcast from
9069 // i32, turn the build_vector into a sequence of insert_vector_elt.
9070 // Relevant operands are everything that is not statically
9071 // (i.e., at compile time) bitcasted.
9072 unsigned NumOfBitCastedElts = 0;
9073 unsigned NumElts = VT.getVectorNumElements();
9074 unsigned NumOfRelevantElts = NumElts;
9075 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9076 SDValue Elt = N->getOperand(Idx);
9077 if (Elt->getOpcode() == ISD::BITCAST) {
9078 // Assume only bit cast to i32 will go away.
9079 if (Elt->getOperand(0).getValueType() == MVT::i32)
9080 ++NumOfBitCastedElts;
9081 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9082 // Constants are statically casted, thus do not count them as
9083 // relevant operands.
9084 --NumOfRelevantElts;
9087 // Check if more than half of the elements require a non-free bitcast.
9088 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9091 SelectionDAG &DAG = DCI.DAG;
9092 // Create the new vector type.
9093 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9094 // Check if the type is legal.
9095 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9096 if (!TLI.isTypeLegal(VecVT))
9100 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9101 // => BITCAST INSERT_VECTOR_ELT
9102 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9104 SDValue Vec = DAG.getUNDEF(VecVT);
9106 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9107 SDValue V = N->getOperand(Idx);
9108 if (V.getOpcode() == ISD::UNDEF)
9110 if (V.getOpcode() == ISD::BITCAST &&
9111 V->getOperand(0).getValueType() == MVT::i32)
9112 // Fold obvious case.
9113 V = V.getOperand(0);
9115 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9116 // Make the DAGCombiner fold the bitcasts.
9117 DCI.AddToWorklist(V.getNode());
9119 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9120 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9122 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9123 // Make the DAGCombiner fold the bitcasts.
9124 DCI.AddToWorklist(Vec.getNode());
9128 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9129 /// ISD::INSERT_VECTOR_ELT.
9130 static SDValue PerformInsertEltCombine(SDNode *N,
9131 TargetLowering::DAGCombinerInfo &DCI) {
9132 // Bitcast an i64 load inserted into a vector to f64.
9133 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9134 EVT VT = N->getValueType(0);
9135 SDNode *Elt = N->getOperand(1).getNode();
9136 if (VT.getVectorElementType() != MVT::i64 ||
9137 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9140 SelectionDAG &DAG = DCI.DAG;
9142 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9143 VT.getVectorNumElements());
9144 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9145 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9146 // Make the DAGCombiner fold the bitcasts.
9147 DCI.AddToWorklist(Vec.getNode());
9148 DCI.AddToWorklist(V.getNode());
9149 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9150 Vec, V, N->getOperand(2));
9151 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9154 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9155 /// ISD::VECTOR_SHUFFLE.
9156 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9157 // The LLVM shufflevector instruction does not require the shuffle mask
9158 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9159 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9160 // operands do not match the mask length, they are extended by concatenating
9161 // them with undef vectors. That is probably the right thing for other
9162 // targets, but for NEON it is better to concatenate two double-register
9163 // size vector operands into a single quad-register size vector. Do that
9164 // transformation here:
9165 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9166 // shuffle(concat(v1, v2), undef)
9167 SDValue Op0 = N->getOperand(0);
9168 SDValue Op1 = N->getOperand(1);
9169 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9170 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9171 Op0.getNumOperands() != 2 ||
9172 Op1.getNumOperands() != 2)
9174 SDValue Concat0Op1 = Op0.getOperand(1);
9175 SDValue Concat1Op1 = Op1.getOperand(1);
9176 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9177 Concat1Op1.getOpcode() != ISD::UNDEF)
9179 // Skip the transformation if any of the types are illegal.
9180 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9181 EVT VT = N->getValueType(0);
9182 if (!TLI.isTypeLegal(VT) ||
9183 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9184 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9187 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9188 Op0.getOperand(0), Op1.getOperand(0));
9189 // Translate the shuffle mask.
9190 SmallVector<int, 16> NewMask;
9191 unsigned NumElts = VT.getVectorNumElements();
9192 unsigned HalfElts = NumElts/2;
9193 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9194 for (unsigned n = 0; n < NumElts; ++n) {
9195 int MaskElt = SVN->getMaskElt(n);
9197 if (MaskElt < (int)HalfElts)
9199 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9200 NewElt = HalfElts + MaskElt - NumElts;
9201 NewMask.push_back(NewElt);
9203 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9204 DAG.getUNDEF(VT), NewMask.data());
9207 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9208 /// NEON load/store intrinsics to merge base address updates.
9209 static SDValue CombineBaseUpdate(SDNode *N,
9210 TargetLowering::DAGCombinerInfo &DCI) {
9211 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9214 SelectionDAG &DAG = DCI.DAG;
9215 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9216 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9217 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9218 SDValue Addr = N->getOperand(AddrOpIdx);
9220 // Search for a use of the address operand that is an increment.
9221 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9222 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9224 if (User->getOpcode() != ISD::ADD ||
9225 UI.getUse().getResNo() != Addr.getResNo())
9228 // Check that the add is independent of the load/store. Otherwise, folding
9229 // it would create a cycle.
9230 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9233 // Find the new opcode for the updating load/store.
9235 bool isLaneOp = false;
9236 unsigned NewOpc = 0;
9237 unsigned NumVecs = 0;
9239 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9241 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9242 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9244 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9246 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9248 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9250 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9251 NumVecs = 2; isLaneOp = true; break;
9252 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9253 NumVecs = 3; isLaneOp = true; break;
9254 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9255 NumVecs = 4; isLaneOp = true; break;
9256 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9257 NumVecs = 1; isLoad = false; break;
9258 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9259 NumVecs = 2; isLoad = false; break;
9260 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9261 NumVecs = 3; isLoad = false; break;
9262 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9263 NumVecs = 4; isLoad = false; break;
9264 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9265 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9266 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9267 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9268 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9269 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9273 switch (N->getOpcode()) {
9274 default: llvm_unreachable("unexpected opcode for Neon base update");
9275 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9276 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9277 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9281 // Find the size of memory referenced by the load/store.
9284 VecTy = N->getValueType(0);
9286 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9287 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9289 NumBytes /= VecTy.getVectorNumElements();
9291 // If the increment is a constant, it must match the memory ref size.
9292 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9293 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9294 uint64_t IncVal = CInc->getZExtValue();
9295 if (IncVal != NumBytes)
9297 } else if (NumBytes >= 3 * 16) {
9298 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9299 // separate instructions that make it harder to use a non-constant update.
9303 // Create the new updating load/store node.
9305 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9307 for (n = 0; n < NumResultVecs; ++n)
9309 Tys[n++] = MVT::i32;
9310 Tys[n] = MVT::Other;
9311 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9312 SmallVector<SDValue, 8> Ops;
9313 Ops.push_back(N->getOperand(0)); // incoming chain
9314 Ops.push_back(N->getOperand(AddrOpIdx));
9316 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9317 Ops.push_back(N->getOperand(i));
9319 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9320 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9321 Ops.data(), Ops.size(),
9322 MemInt->getMemoryVT(),
9323 MemInt->getMemOperand());
9326 std::vector<SDValue> NewResults;
9327 for (unsigned i = 0; i < NumResultVecs; ++i) {
9328 NewResults.push_back(SDValue(UpdN.getNode(), i));
9330 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9331 DCI.CombineTo(N, NewResults);
9332 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9339 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9340 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9341 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9343 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9344 SelectionDAG &DAG = DCI.DAG;
9345 EVT VT = N->getValueType(0);
9346 // vldN-dup instructions only support 64-bit vectors for N > 1.
9347 if (!VT.is64BitVector())
9350 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9351 SDNode *VLD = N->getOperand(0).getNode();
9352 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9354 unsigned NumVecs = 0;
9355 unsigned NewOpc = 0;
9356 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9357 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9359 NewOpc = ARMISD::VLD2DUP;
9360 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9362 NewOpc = ARMISD::VLD3DUP;
9363 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9365 NewOpc = ARMISD::VLD4DUP;
9370 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9371 // numbers match the load.
9372 unsigned VLDLaneNo =
9373 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9374 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9376 // Ignore uses of the chain result.
9377 if (UI.getUse().getResNo() == NumVecs)
9380 if (User->getOpcode() != ARMISD::VDUPLANE ||
9381 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9385 // Create the vldN-dup node.
9388 for (n = 0; n < NumVecs; ++n)
9390 Tys[n] = MVT::Other;
9391 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9392 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9393 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9394 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9395 Ops, 2, VLDMemInt->getMemoryVT(),
9396 VLDMemInt->getMemOperand());
9399 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9401 unsigned ResNo = UI.getUse().getResNo();
9402 // Ignore uses of the chain result.
9403 if (ResNo == NumVecs)
9406 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9409 // Now the vldN-lane intrinsic is dead except for its chain result.
9410 // Update uses of the chain.
9411 std::vector<SDValue> VLDDupResults;
9412 for (unsigned n = 0; n < NumVecs; ++n)
9413 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9414 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9415 DCI.CombineTo(VLD, VLDDupResults);
9420 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9421 /// ARMISD::VDUPLANE.
9422 static SDValue PerformVDUPLANECombine(SDNode *N,
9423 TargetLowering::DAGCombinerInfo &DCI) {
9424 SDValue Op = N->getOperand(0);
9426 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9427 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9428 if (CombineVLDDUP(N, DCI))
9429 return SDValue(N, 0);
9431 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9432 // redundant. Ignore bit_converts for now; element sizes are checked below.
9433 while (Op.getOpcode() == ISD::BITCAST)
9434 Op = Op.getOperand(0);
9435 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9438 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9439 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9440 // The canonical VMOV for a zero vector uses a 32-bit element size.
9441 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9443 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9445 EVT VT = N->getValueType(0);
9446 if (EltSize > VT.getVectorElementType().getSizeInBits())
9449 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9452 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9453 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9454 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9458 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9460 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9465 APFloat APF = C->getValueAPF();
9466 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9467 != APFloat::opOK || !isExact)
9470 c0 = (I == 0) ? cN : c0;
9471 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9478 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9479 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9480 /// when the VMUL has a constant operand that is a power of 2.
9482 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9483 /// vmul.f32 d16, d17, d16
9484 /// vcvt.s32.f32 d16, d16
9486 /// vcvt.s32.f32 d16, d16, #3
9487 static SDValue PerformVCVTCombine(SDNode *N,
9488 TargetLowering::DAGCombinerInfo &DCI,
9489 const ARMSubtarget *Subtarget) {
9490 SelectionDAG &DAG = DCI.DAG;
9491 SDValue Op = N->getOperand(0);
9493 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9494 Op.getOpcode() != ISD::FMUL)
9498 SDValue N0 = Op->getOperand(0);
9499 SDValue ConstVec = Op->getOperand(1);
9500 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9502 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9503 !isConstVecPow2(ConstVec, isSigned, C))
9506 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9507 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9508 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9509 // These instructions only exist converting from f32 to i32. We can handle
9510 // smaller integers by generating an extra truncate, but larger ones would
9515 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9516 Intrinsic::arm_neon_vcvtfp2fxu;
9517 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9518 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9519 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9520 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9521 DAG.getConstant(Log2_64(C), MVT::i32));
9523 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9524 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9529 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9530 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9531 /// when the VDIV has a constant operand that is a power of 2.
9533 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9534 /// vcvt.f32.s32 d16, d16
9535 /// vdiv.f32 d16, d17, d16
9537 /// vcvt.f32.s32 d16, d16, #3
9538 static SDValue PerformVDIVCombine(SDNode *N,
9539 TargetLowering::DAGCombinerInfo &DCI,
9540 const ARMSubtarget *Subtarget) {
9541 SelectionDAG &DAG = DCI.DAG;
9542 SDValue Op = N->getOperand(0);
9543 unsigned OpOpcode = Op.getNode()->getOpcode();
9545 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9546 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9550 SDValue ConstVec = N->getOperand(1);
9551 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9553 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9554 !isConstVecPow2(ConstVec, isSigned, C))
9557 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9558 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9559 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9560 // These instructions only exist converting from i32 to f32. We can handle
9561 // smaller integers by generating an extra extend, but larger ones would
9566 SDValue ConvInput = Op.getOperand(0);
9567 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9568 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9569 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9570 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9573 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9574 Intrinsic::arm_neon_vcvtfxu2fp;
9575 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9577 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9578 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9581 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9582 /// operand of a vector shift operation, where all the elements of the
9583 /// build_vector must have the same constant integer value.
9584 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9585 // Ignore bit_converts.
9586 while (Op.getOpcode() == ISD::BITCAST)
9587 Op = Op.getOperand(0);
9588 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9589 APInt SplatBits, SplatUndef;
9590 unsigned SplatBitSize;
9592 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9593 HasAnyUndefs, ElementBits) ||
9594 SplatBitSize > ElementBits)
9596 Cnt = SplatBits.getSExtValue();
9600 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9601 /// operand of a vector shift left operation. That value must be in the range:
9602 /// 0 <= Value < ElementBits for a left shift; or
9603 /// 0 <= Value <= ElementBits for a long left shift.
9604 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9605 assert(VT.isVector() && "vector shift count is not a vector type");
9606 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9607 if (! getVShiftImm(Op, ElementBits, Cnt))
9609 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9612 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9613 /// operand of a vector shift right operation. For a shift opcode, the value
9614 /// is positive, but for an intrinsic the value count must be negative. The
9615 /// absolute value must be in the range:
9616 /// 1 <= |Value| <= ElementBits for a right shift; or
9617 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9618 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9620 assert(VT.isVector() && "vector shift count is not a vector type");
9621 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9622 if (! getVShiftImm(Op, ElementBits, Cnt))
9626 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9629 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9630 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9631 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9634 // Don't do anything for most intrinsics.
9637 // Vector shifts: check for immediate versions and lower them.
9638 // Note: This is done during DAG combining instead of DAG legalizing because
9639 // the build_vectors for 64-bit vector element shift counts are generally
9640 // not legal, and it is hard to see their values after they get legalized to
9641 // loads from a constant pool.
9642 case Intrinsic::arm_neon_vshifts:
9643 case Intrinsic::arm_neon_vshiftu:
9644 case Intrinsic::arm_neon_vshiftls:
9645 case Intrinsic::arm_neon_vshiftlu:
9646 case Intrinsic::arm_neon_vshiftn:
9647 case Intrinsic::arm_neon_vrshifts:
9648 case Intrinsic::arm_neon_vrshiftu:
9649 case Intrinsic::arm_neon_vrshiftn:
9650 case Intrinsic::arm_neon_vqshifts:
9651 case Intrinsic::arm_neon_vqshiftu:
9652 case Intrinsic::arm_neon_vqshiftsu:
9653 case Intrinsic::arm_neon_vqshiftns:
9654 case Intrinsic::arm_neon_vqshiftnu:
9655 case Intrinsic::arm_neon_vqshiftnsu:
9656 case Intrinsic::arm_neon_vqrshiftns:
9657 case Intrinsic::arm_neon_vqrshiftnu:
9658 case Intrinsic::arm_neon_vqrshiftnsu: {
9659 EVT VT = N->getOperand(1).getValueType();
9661 unsigned VShiftOpc = 0;
9664 case Intrinsic::arm_neon_vshifts:
9665 case Intrinsic::arm_neon_vshiftu:
9666 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9667 VShiftOpc = ARMISD::VSHL;
9670 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9671 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9672 ARMISD::VSHRs : ARMISD::VSHRu);
9677 case Intrinsic::arm_neon_vshiftls:
9678 case Intrinsic::arm_neon_vshiftlu:
9679 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9681 llvm_unreachable("invalid shift count for vshll intrinsic");
9683 case Intrinsic::arm_neon_vrshifts:
9684 case Intrinsic::arm_neon_vrshiftu:
9685 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9689 case Intrinsic::arm_neon_vqshifts:
9690 case Intrinsic::arm_neon_vqshiftu:
9691 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9695 case Intrinsic::arm_neon_vqshiftsu:
9696 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9698 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9700 case Intrinsic::arm_neon_vshiftn:
9701 case Intrinsic::arm_neon_vrshiftn:
9702 case Intrinsic::arm_neon_vqshiftns:
9703 case Intrinsic::arm_neon_vqshiftnu:
9704 case Intrinsic::arm_neon_vqshiftnsu:
9705 case Intrinsic::arm_neon_vqrshiftns:
9706 case Intrinsic::arm_neon_vqrshiftnu:
9707 case Intrinsic::arm_neon_vqrshiftnsu:
9708 // Narrowing shifts require an immediate right shift.
9709 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9711 llvm_unreachable("invalid shift count for narrowing vector shift "
9715 llvm_unreachable("unhandled vector shift");
9719 case Intrinsic::arm_neon_vshifts:
9720 case Intrinsic::arm_neon_vshiftu:
9721 // Opcode already set above.
9723 case Intrinsic::arm_neon_vshiftls:
9724 case Intrinsic::arm_neon_vshiftlu:
9725 if (Cnt == VT.getVectorElementType().getSizeInBits())
9726 VShiftOpc = ARMISD::VSHLLi;
9728 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9729 ARMISD::VSHLLs : ARMISD::VSHLLu);
9731 case Intrinsic::arm_neon_vshiftn:
9732 VShiftOpc = ARMISD::VSHRN; break;
9733 case Intrinsic::arm_neon_vrshifts:
9734 VShiftOpc = ARMISD::VRSHRs; break;
9735 case Intrinsic::arm_neon_vrshiftu:
9736 VShiftOpc = ARMISD::VRSHRu; break;
9737 case Intrinsic::arm_neon_vrshiftn:
9738 VShiftOpc = ARMISD::VRSHRN; break;
9739 case Intrinsic::arm_neon_vqshifts:
9740 VShiftOpc = ARMISD::VQSHLs; break;
9741 case Intrinsic::arm_neon_vqshiftu:
9742 VShiftOpc = ARMISD::VQSHLu; break;
9743 case Intrinsic::arm_neon_vqshiftsu:
9744 VShiftOpc = ARMISD::VQSHLsu; break;
9745 case Intrinsic::arm_neon_vqshiftns:
9746 VShiftOpc = ARMISD::VQSHRNs; break;
9747 case Intrinsic::arm_neon_vqshiftnu:
9748 VShiftOpc = ARMISD::VQSHRNu; break;
9749 case Intrinsic::arm_neon_vqshiftnsu:
9750 VShiftOpc = ARMISD::VQSHRNsu; break;
9751 case Intrinsic::arm_neon_vqrshiftns:
9752 VShiftOpc = ARMISD::VQRSHRNs; break;
9753 case Intrinsic::arm_neon_vqrshiftnu:
9754 VShiftOpc = ARMISD::VQRSHRNu; break;
9755 case Intrinsic::arm_neon_vqrshiftnsu:
9756 VShiftOpc = ARMISD::VQRSHRNsu; break;
9759 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9760 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9763 case Intrinsic::arm_neon_vshiftins: {
9764 EVT VT = N->getOperand(1).getValueType();
9766 unsigned VShiftOpc = 0;
9768 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9769 VShiftOpc = ARMISD::VSLI;
9770 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9771 VShiftOpc = ARMISD::VSRI;
9773 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9776 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9777 N->getOperand(1), N->getOperand(2),
9778 DAG.getConstant(Cnt, MVT::i32));
9781 case Intrinsic::arm_neon_vqrshifts:
9782 case Intrinsic::arm_neon_vqrshiftu:
9783 // No immediate versions of these to check for.
9790 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9791 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9792 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9793 /// vector element shift counts are generally not legal, and it is hard to see
9794 /// their values after they get legalized to loads from a constant pool.
9795 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9796 const ARMSubtarget *ST) {
9797 EVT VT = N->getValueType(0);
9798 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9799 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9800 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9801 SDValue N1 = N->getOperand(1);
9802 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9803 SDValue N0 = N->getOperand(0);
9804 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9805 DAG.MaskedValueIsZero(N0.getOperand(0),
9806 APInt::getHighBitsSet(32, 16)))
9807 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9811 // Nothing to be done for scalar shifts.
9812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9813 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9816 assert(ST->hasNEON() && "unexpected vector shift");
9819 switch (N->getOpcode()) {
9820 default: llvm_unreachable("unexpected shift opcode");
9823 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9824 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9825 DAG.getConstant(Cnt, MVT::i32));
9830 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9831 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9832 ARMISD::VSHRs : ARMISD::VSHRu);
9833 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9834 DAG.getConstant(Cnt, MVT::i32));
9840 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9841 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9842 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9843 const ARMSubtarget *ST) {
9844 SDValue N0 = N->getOperand(0);
9846 // Check for sign- and zero-extensions of vector extract operations of 8-
9847 // and 16-bit vector elements. NEON supports these directly. They are
9848 // handled during DAG combining because type legalization will promote them
9849 // to 32-bit types and it is messy to recognize the operations after that.
9850 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9851 SDValue Vec = N0.getOperand(0);
9852 SDValue Lane = N0.getOperand(1);
9853 EVT VT = N->getValueType(0);
9854 EVT EltVT = N0.getValueType();
9855 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9857 if (VT == MVT::i32 &&
9858 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9859 TLI.isTypeLegal(Vec.getValueType()) &&
9860 isa<ConstantSDNode>(Lane)) {
9863 switch (N->getOpcode()) {
9864 default: llvm_unreachable("unexpected opcode");
9865 case ISD::SIGN_EXTEND:
9866 Opc = ARMISD::VGETLANEs;
9868 case ISD::ZERO_EXTEND:
9869 case ISD::ANY_EXTEND:
9870 Opc = ARMISD::VGETLANEu;
9873 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9880 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9881 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9882 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9883 const ARMSubtarget *ST) {
9884 // If the target supports NEON, try to use vmax/vmin instructions for f32
9885 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9886 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9887 // a NaN; only do the transformation when it matches that behavior.
9889 // For now only do this when using NEON for FP operations; if using VFP, it
9890 // is not obvious that the benefit outweighs the cost of switching to the
9892 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9893 N->getValueType(0) != MVT::f32)
9896 SDValue CondLHS = N->getOperand(0);
9897 SDValue CondRHS = N->getOperand(1);
9898 SDValue LHS = N->getOperand(2);
9899 SDValue RHS = N->getOperand(3);
9900 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9902 unsigned Opcode = 0;
9904 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9905 IsReversed = false; // x CC y ? x : y
9906 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9907 IsReversed = true ; // x CC y ? y : x
9921 // If LHS is NaN, an ordered comparison will be false and the result will
9922 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9923 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9924 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9925 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9927 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9928 // will return -0, so vmin can only be used for unsafe math or if one of
9929 // the operands is known to be nonzero.
9930 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9931 !DAG.getTarget().Options.UnsafeFPMath &&
9932 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9934 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9943 // If LHS is NaN, an ordered comparison will be false and the result will
9944 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9945 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9946 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9947 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9949 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9950 // will return +0, so vmax can only be used for unsafe math or if one of
9951 // the operands is known to be nonzero.
9952 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9953 !DAG.getTarget().Options.UnsafeFPMath &&
9954 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9956 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9962 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9965 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9967 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9968 SDValue Cmp = N->getOperand(4);
9969 if (Cmp.getOpcode() != ARMISD::CMPZ)
9970 // Only looking at EQ and NE cases.
9973 EVT VT = N->getValueType(0);
9975 SDValue LHS = Cmp.getOperand(0);
9976 SDValue RHS = Cmp.getOperand(1);
9977 SDValue FalseVal = N->getOperand(0);
9978 SDValue TrueVal = N->getOperand(1);
9979 SDValue ARMcc = N->getOperand(2);
9980 ARMCC::CondCodes CC =
9981 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9999 /// FIXME: Turn this into a target neutral optimization?
10001 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10002 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10003 N->getOperand(3), Cmp);
10004 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10006 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10007 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10008 N->getOperand(3), NewCmp);
10011 if (Res.getNode()) {
10012 APInt KnownZero, KnownOne;
10013 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
10014 // Capture demanded bits information that would be otherwise lost.
10015 if (KnownZero == 0xfffffffe)
10016 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10017 DAG.getValueType(MVT::i1));
10018 else if (KnownZero == 0xffffff00)
10019 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10020 DAG.getValueType(MVT::i8));
10021 else if (KnownZero == 0xffff0000)
10022 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10023 DAG.getValueType(MVT::i16));
10029 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10030 DAGCombinerInfo &DCI) const {
10031 switch (N->getOpcode()) {
10033 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10034 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10035 case ISD::SUB: return PerformSUBCombine(N, DCI);
10036 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10037 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10038 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10039 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10040 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10041 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
10042 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10043 case ISD::STORE: return PerformSTORECombine(N, DCI);
10044 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
10045 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10046 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10047 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10048 case ISD::FP_TO_SINT:
10049 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10050 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10051 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10054 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10055 case ISD::SIGN_EXTEND:
10056 case ISD::ZERO_EXTEND:
10057 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10058 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10059 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10060 case ARMISD::VLD2DUP:
10061 case ARMISD::VLD3DUP:
10062 case ARMISD::VLD4DUP:
10063 return CombineBaseUpdate(N, DCI);
10064 case ARMISD::BUILD_VECTOR:
10065 return PerformARMBUILD_VECTORCombine(N, DCI);
10066 case ISD::INTRINSIC_VOID:
10067 case ISD::INTRINSIC_W_CHAIN:
10068 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10069 case Intrinsic::arm_neon_vld1:
10070 case Intrinsic::arm_neon_vld2:
10071 case Intrinsic::arm_neon_vld3:
10072 case Intrinsic::arm_neon_vld4:
10073 case Intrinsic::arm_neon_vld2lane:
10074 case Intrinsic::arm_neon_vld3lane:
10075 case Intrinsic::arm_neon_vld4lane:
10076 case Intrinsic::arm_neon_vst1:
10077 case Intrinsic::arm_neon_vst2:
10078 case Intrinsic::arm_neon_vst3:
10079 case Intrinsic::arm_neon_vst4:
10080 case Intrinsic::arm_neon_vst2lane:
10081 case Intrinsic::arm_neon_vst3lane:
10082 case Intrinsic::arm_neon_vst4lane:
10083 return CombineBaseUpdate(N, DCI);
10091 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10093 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10096 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
10097 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10098 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10100 switch (VT.getSimpleVT().SimpleTy) {
10106 // Unaligned access can use (for example) LRDB, LRDH, LDR
10107 if (AllowsUnaligned) {
10109 *Fast = Subtarget->hasV7Ops();
10116 // For any little-endian targets with neon, we can support unaligned ld/st
10117 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10118 // A big-endian target may also explictly support unaligned accesses
10119 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10129 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10130 unsigned AlignCheck) {
10131 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10132 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10135 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10136 unsigned DstAlign, unsigned SrcAlign,
10137 bool IsMemset, bool ZeroMemset,
10139 MachineFunction &MF) const {
10140 const Function *F = MF.getFunction();
10142 // See if we can use NEON instructions for this...
10143 if ((!IsMemset || ZeroMemset) &&
10144 Subtarget->hasNEON() &&
10145 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10146 Attribute::NoImplicitFloat)) {
10149 (memOpAlign(SrcAlign, DstAlign, 16) ||
10150 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
10152 } else if (Size >= 8 &&
10153 (memOpAlign(SrcAlign, DstAlign, 8) ||
10154 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
10159 // Lowering to i32/i16 if the size permits.
10162 else if (Size >= 2)
10165 // Let the target-independent logic figure it out.
10169 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10170 if (Val.getOpcode() != ISD::LOAD)
10173 EVT VT1 = Val.getValueType();
10174 if (!VT1.isSimple() || !VT1.isInteger() ||
10175 !VT2.isSimple() || !VT2.isInteger())
10178 switch (VT1.getSimpleVT().SimpleTy) {
10183 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10190 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10191 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10194 if (!isTypeLegal(EVT::getEVT(Ty1)))
10197 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10199 // Assuming the caller doesn't have a zeroext or signext return parameter,
10200 // truncation all the way down to i1 is valid.
10205 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10209 unsigned Scale = 1;
10210 switch (VT.getSimpleVT().SimpleTy) {
10211 default: return false;
10226 if ((V & (Scale - 1)) != 0)
10229 return V == (V & ((1LL << 5) - 1));
10232 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10233 const ARMSubtarget *Subtarget) {
10234 bool isNeg = false;
10240 switch (VT.getSimpleVT().SimpleTy) {
10241 default: return false;
10246 // + imm12 or - imm8
10248 return V == (V & ((1LL << 8) - 1));
10249 return V == (V & ((1LL << 12) - 1));
10252 // Same as ARM mode. FIXME: NEON?
10253 if (!Subtarget->hasVFP2())
10258 return V == (V & ((1LL << 8) - 1));
10262 /// isLegalAddressImmediate - Return true if the integer value can be used
10263 /// as the offset of the target addressing mode for load / store of the
10265 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10266 const ARMSubtarget *Subtarget) {
10270 if (!VT.isSimple())
10273 if (Subtarget->isThumb1Only())
10274 return isLegalT1AddressImmediate(V, VT);
10275 else if (Subtarget->isThumb2())
10276 return isLegalT2AddressImmediate(V, VT, Subtarget);
10281 switch (VT.getSimpleVT().SimpleTy) {
10282 default: return false;
10287 return V == (V & ((1LL << 12) - 1));
10290 return V == (V & ((1LL << 8) - 1));
10293 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10298 return V == (V & ((1LL << 8) - 1));
10302 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10304 int Scale = AM.Scale;
10308 switch (VT.getSimpleVT().SimpleTy) {
10309 default: return false;
10317 Scale = Scale & ~1;
10318 return Scale == 2 || Scale == 4 || Scale == 8;
10321 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10325 // Note, we allow "void" uses (basically, uses that aren't loads or
10326 // stores), because arm allows folding a scale into many arithmetic
10327 // operations. This should be made more precise and revisited later.
10329 // Allow r << imm, but the imm has to be a multiple of two.
10330 if (Scale & 1) return false;
10331 return isPowerOf2_32(Scale);
10335 /// isLegalAddressingMode - Return true if the addressing mode represented
10336 /// by AM is legal for this target, for a load/store of the specified type.
10337 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10339 EVT VT = getValueType(Ty, true);
10340 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10343 // Can never fold addr of global into load/store.
10347 switch (AM.Scale) {
10348 case 0: // no scale reg, must be "r+i" or "r", or "i".
10351 if (Subtarget->isThumb1Only())
10355 // ARM doesn't support any R+R*scale+imm addr modes.
10359 if (!VT.isSimple())
10362 if (Subtarget->isThumb2())
10363 return isLegalT2ScaledAddressingMode(AM, VT);
10365 int Scale = AM.Scale;
10366 switch (VT.getSimpleVT().SimpleTy) {
10367 default: return false;
10371 if (Scale < 0) Scale = -Scale;
10375 return isPowerOf2_32(Scale & ~1);
10379 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10384 // Note, we allow "void" uses (basically, uses that aren't loads or
10385 // stores), because arm allows folding a scale into many arithmetic
10386 // operations. This should be made more precise and revisited later.
10388 // Allow r << imm, but the imm has to be a multiple of two.
10389 if (Scale & 1) return false;
10390 return isPowerOf2_32(Scale);
10396 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10397 /// icmp immediate, that is the target has icmp instructions which can compare
10398 /// a register against the immediate without having to materialize the
10399 /// immediate into a register.
10400 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10401 // Thumb2 and ARM modes can use cmn for negative immediates.
10402 if (!Subtarget->isThumb())
10403 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10404 if (Subtarget->isThumb2())
10405 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10406 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10407 return Imm >= 0 && Imm <= 255;
10410 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10411 /// *or sub* immediate, that is the target has add or sub instructions which can
10412 /// add a register with the immediate without having to materialize the
10413 /// immediate into a register.
10414 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10415 // Same encoding for add/sub, just flip the sign.
10416 int64_t AbsImm = llvm::abs64(Imm);
10417 if (!Subtarget->isThumb())
10418 return ARM_AM::getSOImmVal(AbsImm) != -1;
10419 if (Subtarget->isThumb2())
10420 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10421 // Thumb1 only has 8-bit unsigned immediate.
10422 return AbsImm >= 0 && AbsImm <= 255;
10425 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10426 bool isSEXTLoad, SDValue &Base,
10427 SDValue &Offset, bool &isInc,
10428 SelectionDAG &DAG) {
10429 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10432 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10433 // AddressingMode 3
10434 Base = Ptr->getOperand(0);
10435 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10436 int RHSC = (int)RHS->getZExtValue();
10437 if (RHSC < 0 && RHSC > -256) {
10438 assert(Ptr->getOpcode() == ISD::ADD);
10440 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10444 isInc = (Ptr->getOpcode() == ISD::ADD);
10445 Offset = Ptr->getOperand(1);
10447 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10448 // AddressingMode 2
10449 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10450 int RHSC = (int)RHS->getZExtValue();
10451 if (RHSC < 0 && RHSC > -0x1000) {
10452 assert(Ptr->getOpcode() == ISD::ADD);
10454 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10455 Base = Ptr->getOperand(0);
10460 if (Ptr->getOpcode() == ISD::ADD) {
10462 ARM_AM::ShiftOpc ShOpcVal=
10463 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10464 if (ShOpcVal != ARM_AM::no_shift) {
10465 Base = Ptr->getOperand(1);
10466 Offset = Ptr->getOperand(0);
10468 Base = Ptr->getOperand(0);
10469 Offset = Ptr->getOperand(1);
10474 isInc = (Ptr->getOpcode() == ISD::ADD);
10475 Base = Ptr->getOperand(0);
10476 Offset = Ptr->getOperand(1);
10480 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10484 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10485 bool isSEXTLoad, SDValue &Base,
10486 SDValue &Offset, bool &isInc,
10487 SelectionDAG &DAG) {
10488 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10491 Base = Ptr->getOperand(0);
10492 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10493 int RHSC = (int)RHS->getZExtValue();
10494 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10495 assert(Ptr->getOpcode() == ISD::ADD);
10497 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10499 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10500 isInc = Ptr->getOpcode() == ISD::ADD;
10501 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10509 /// getPreIndexedAddressParts - returns true by value, base pointer and
10510 /// offset pointer and addressing mode by reference if the node's address
10511 /// can be legally represented as pre-indexed load / store address.
10513 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10515 ISD::MemIndexedMode &AM,
10516 SelectionDAG &DAG) const {
10517 if (Subtarget->isThumb1Only())
10522 bool isSEXTLoad = false;
10523 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10524 Ptr = LD->getBasePtr();
10525 VT = LD->getMemoryVT();
10526 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10527 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10528 Ptr = ST->getBasePtr();
10529 VT = ST->getMemoryVT();
10534 bool isLegal = false;
10535 if (Subtarget->isThumb2())
10536 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10537 Offset, isInc, DAG);
10539 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10540 Offset, isInc, DAG);
10544 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10548 /// getPostIndexedAddressParts - returns true by value, base pointer and
10549 /// offset pointer and addressing mode by reference if this node can be
10550 /// combined with a load / store to form a post-indexed load / store.
10551 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10554 ISD::MemIndexedMode &AM,
10555 SelectionDAG &DAG) const {
10556 if (Subtarget->isThumb1Only())
10561 bool isSEXTLoad = false;
10562 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10563 VT = LD->getMemoryVT();
10564 Ptr = LD->getBasePtr();
10565 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10566 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10567 VT = ST->getMemoryVT();
10568 Ptr = ST->getBasePtr();
10573 bool isLegal = false;
10574 if (Subtarget->isThumb2())
10575 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10578 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10584 // Swap base ptr and offset to catch more post-index load / store when
10585 // it's legal. In Thumb2 mode, offset must be an immediate.
10586 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10587 !Subtarget->isThumb2())
10588 std::swap(Base, Offset);
10590 // Post-indexed load / store update the base pointer.
10595 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10599 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10602 const SelectionDAG &DAG,
10603 unsigned Depth) const {
10604 unsigned BitWidth = KnownOne.getBitWidth();
10605 KnownZero = KnownOne = APInt(BitWidth, 0);
10606 switch (Op.getOpcode()) {
10612 // These nodes' second result is a boolean
10613 if (Op.getResNo() == 0)
10615 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10617 case ARMISD::CMOV: {
10618 // Bits are known zero/one if known on the LHS and RHS.
10619 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10620 if (KnownZero == 0 && KnownOne == 0) return;
10622 APInt KnownZeroRHS, KnownOneRHS;
10623 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10624 KnownZero &= KnownZeroRHS;
10625 KnownOne &= KnownOneRHS;
10631 //===----------------------------------------------------------------------===//
10632 // ARM Inline Assembly Support
10633 //===----------------------------------------------------------------------===//
10635 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10636 // Looking for "rev" which is V6+.
10637 if (!Subtarget->hasV6Ops())
10640 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10641 std::string AsmStr = IA->getAsmString();
10642 SmallVector<StringRef, 4> AsmPieces;
10643 SplitString(AsmStr, AsmPieces, ";\n");
10645 switch (AsmPieces.size()) {
10646 default: return false;
10648 AsmStr = AsmPieces[0];
10650 SplitString(AsmStr, AsmPieces, " \t,");
10653 if (AsmPieces.size() == 3 &&
10654 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10655 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10656 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10657 if (Ty && Ty->getBitWidth() == 32)
10658 return IntrinsicLowering::LowerToByteSwap(CI);
10666 /// getConstraintType - Given a constraint letter, return the type of
10667 /// constraint it is for this target.
10668 ARMTargetLowering::ConstraintType
10669 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10670 if (Constraint.size() == 1) {
10671 switch (Constraint[0]) {
10673 case 'l': return C_RegisterClass;
10674 case 'w': return C_RegisterClass;
10675 case 'h': return C_RegisterClass;
10676 case 'x': return C_RegisterClass;
10677 case 't': return C_RegisterClass;
10678 case 'j': return C_Other; // Constant for movw.
10679 // An address with a single base register. Due to the way we
10680 // currently handle addresses it is the same as an 'r' memory constraint.
10681 case 'Q': return C_Memory;
10683 } else if (Constraint.size() == 2) {
10684 switch (Constraint[0]) {
10686 // All 'U+' constraints are addresses.
10687 case 'U': return C_Memory;
10690 return TargetLowering::getConstraintType(Constraint);
10693 /// Examine constraint type and operand type and determine a weight value.
10694 /// This object must already have been set up with the operand type
10695 /// and the current alternative constraint selected.
10696 TargetLowering::ConstraintWeight
10697 ARMTargetLowering::getSingleConstraintMatchWeight(
10698 AsmOperandInfo &info, const char *constraint) const {
10699 ConstraintWeight weight = CW_Invalid;
10700 Value *CallOperandVal = info.CallOperandVal;
10701 // If we don't have a value, we can't do a match,
10702 // but allow it at the lowest weight.
10703 if (CallOperandVal == NULL)
10705 Type *type = CallOperandVal->getType();
10706 // Look at the constraint type.
10707 switch (*constraint) {
10709 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10712 if (type->isIntegerTy()) {
10713 if (Subtarget->isThumb())
10714 weight = CW_SpecificReg;
10716 weight = CW_Register;
10720 if (type->isFloatingPointTy())
10721 weight = CW_Register;
10727 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10729 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10731 if (Constraint.size() == 1) {
10732 // GCC ARM Constraint Letters
10733 switch (Constraint[0]) {
10734 case 'l': // Low regs or general regs.
10735 if (Subtarget->isThumb())
10736 return RCPair(0U, &ARM::tGPRRegClass);
10737 return RCPair(0U, &ARM::GPRRegClass);
10738 case 'h': // High regs or no regs.
10739 if (Subtarget->isThumb())
10740 return RCPair(0U, &ARM::hGPRRegClass);
10743 return RCPair(0U, &ARM::GPRRegClass);
10745 if (VT == MVT::f32)
10746 return RCPair(0U, &ARM::SPRRegClass);
10747 if (VT.getSizeInBits() == 64)
10748 return RCPair(0U, &ARM::DPRRegClass);
10749 if (VT.getSizeInBits() == 128)
10750 return RCPair(0U, &ARM::QPRRegClass);
10753 if (VT == MVT::f32)
10754 return RCPair(0U, &ARM::SPR_8RegClass);
10755 if (VT.getSizeInBits() == 64)
10756 return RCPair(0U, &ARM::DPR_8RegClass);
10757 if (VT.getSizeInBits() == 128)
10758 return RCPair(0U, &ARM::QPR_8RegClass);
10761 if (VT == MVT::f32)
10762 return RCPair(0U, &ARM::SPRRegClass);
10766 if (StringRef("{cc}").equals_lower(Constraint))
10767 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10769 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10772 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10773 /// vector. If it is invalid, don't add anything to Ops.
10774 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10775 std::string &Constraint,
10776 std::vector<SDValue>&Ops,
10777 SelectionDAG &DAG) const {
10778 SDValue Result(0, 0);
10780 // Currently only support length 1 constraints.
10781 if (Constraint.length() != 1) return;
10783 char ConstraintLetter = Constraint[0];
10784 switch (ConstraintLetter) {
10787 case 'I': case 'J': case 'K': case 'L':
10788 case 'M': case 'N': case 'O':
10789 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10793 int64_t CVal64 = C->getSExtValue();
10794 int CVal = (int) CVal64;
10795 // None of these constraints allow values larger than 32 bits. Check
10796 // that the value fits in an int.
10797 if (CVal != CVal64)
10800 switch (ConstraintLetter) {
10802 // Constant suitable for movw, must be between 0 and
10804 if (Subtarget->hasV6T2Ops())
10805 if (CVal >= 0 && CVal <= 65535)
10809 if (Subtarget->isThumb1Only()) {
10810 // This must be a constant between 0 and 255, for ADD
10812 if (CVal >= 0 && CVal <= 255)
10814 } else if (Subtarget->isThumb2()) {
10815 // A constant that can be used as an immediate value in a
10816 // data-processing instruction.
10817 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10820 // A constant that can be used as an immediate value in a
10821 // data-processing instruction.
10822 if (ARM_AM::getSOImmVal(CVal) != -1)
10828 if (Subtarget->isThumb()) { // FIXME thumb2
10829 // This must be a constant between -255 and -1, for negated ADD
10830 // immediates. This can be used in GCC with an "n" modifier that
10831 // prints the negated value, for use with SUB instructions. It is
10832 // not useful otherwise but is implemented for compatibility.
10833 if (CVal >= -255 && CVal <= -1)
10836 // This must be a constant between -4095 and 4095. It is not clear
10837 // what this constraint is intended for. Implemented for
10838 // compatibility with GCC.
10839 if (CVal >= -4095 && CVal <= 4095)
10845 if (Subtarget->isThumb1Only()) {
10846 // A 32-bit value where only one byte has a nonzero value. Exclude
10847 // zero to match GCC. This constraint is used by GCC internally for
10848 // constants that can be loaded with a move/shift combination.
10849 // It is not useful otherwise but is implemented for compatibility.
10850 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10852 } else if (Subtarget->isThumb2()) {
10853 // A constant whose bitwise inverse can be used as an immediate
10854 // value in a data-processing instruction. This can be used in GCC
10855 // with a "B" modifier that prints the inverted value, for use with
10856 // BIC and MVN instructions. It is not useful otherwise but is
10857 // implemented for compatibility.
10858 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10861 // A constant whose bitwise inverse can be used as an immediate
10862 // value in a data-processing instruction. This can be used in GCC
10863 // with a "B" modifier that prints the inverted value, for use with
10864 // BIC and MVN instructions. It is not useful otherwise but is
10865 // implemented for compatibility.
10866 if (ARM_AM::getSOImmVal(~CVal) != -1)
10872 if (Subtarget->isThumb1Only()) {
10873 // This must be a constant between -7 and 7,
10874 // for 3-operand ADD/SUB immediate instructions.
10875 if (CVal >= -7 && CVal < 7)
10877 } else if (Subtarget->isThumb2()) {
10878 // A constant whose negation can be used as an immediate value in a
10879 // data-processing instruction. This can be used in GCC with an "n"
10880 // modifier that prints the negated value, for use with SUB
10881 // instructions. It is not useful otherwise but is implemented for
10883 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10886 // A constant whose negation can be used as an immediate value in a
10887 // data-processing instruction. This can be used in GCC with an "n"
10888 // modifier that prints the negated value, for use with SUB
10889 // instructions. It is not useful otherwise but is implemented for
10891 if (ARM_AM::getSOImmVal(-CVal) != -1)
10897 if (Subtarget->isThumb()) { // FIXME thumb2
10898 // This must be a multiple of 4 between 0 and 1020, for
10899 // ADD sp + immediate.
10900 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10903 // A power of two or a constant between 0 and 32. This is used in
10904 // GCC for the shift amount on shifted register operands, but it is
10905 // useful in general for any shift amounts.
10906 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10912 if (Subtarget->isThumb()) { // FIXME thumb2
10913 // This must be a constant between 0 and 31, for shift amounts.
10914 if (CVal >= 0 && CVal <= 31)
10920 if (Subtarget->isThumb()) { // FIXME thumb2
10921 // This must be a multiple of 4 between -508 and 508, for
10922 // ADD/SUB sp = sp + immediate.
10923 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10928 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10932 if (Result.getNode()) {
10933 Ops.push_back(Result);
10936 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10939 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10940 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10941 unsigned Opcode = Op->getOpcode();
10942 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10943 "Invalid opcode for Div/Rem lowering");
10944 bool isSigned = (Opcode == ISD::SDIVREM);
10945 EVT VT = Op->getValueType(0);
10946 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10949 switch (VT.getSimpleVT().SimpleTy) {
10950 default: llvm_unreachable("Unexpected request for libcall!");
10951 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10952 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10953 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10954 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10957 SDValue InChain = DAG.getEntryNode();
10959 TargetLowering::ArgListTy Args;
10960 TargetLowering::ArgListEntry Entry;
10961 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10962 EVT ArgVT = Op->getOperand(i).getValueType();
10963 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10964 Entry.Node = Op->getOperand(i);
10966 Entry.isSExt = isSigned;
10967 Entry.isZExt = !isSigned;
10968 Args.push_back(Entry);
10971 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10974 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10978 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
10979 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
10980 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
10981 Callee, Args, DAG, dl);
10982 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10984 return CallInfo.first;
10988 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10989 // The ARM target isn't yet aware of offsets.
10993 bool ARM::isBitFieldInvertedMask(unsigned v) {
10994 if (v == 0xffffffff)
10997 // there can be 1's on either or both "outsides", all the "inside"
10998 // bits must be 0's
10999 unsigned TO = CountTrailingOnes_32(v);
11000 unsigned LO = CountLeadingOnes_32(v);
11001 v = (v >> TO) << TO;
11002 v = (v << LO) >> LO;
11006 /// isFPImmLegal - Returns true if the target can instruction select the
11007 /// specified FP immediate natively. If false, the legalizer will
11008 /// materialize the FP immediate as a load from a constant pool.
11009 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11010 if (!Subtarget->hasVFP3())
11012 if (VT == MVT::f32)
11013 return ARM_AM::getFP32Imm(Imm) != -1;
11014 if (VT == MVT::f64)
11015 return ARM_AM::getFP64Imm(Imm) != -1;
11019 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11020 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11021 /// specified in the intrinsic calls.
11022 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11024 unsigned Intrinsic) const {
11025 switch (Intrinsic) {
11026 case Intrinsic::arm_neon_vld1:
11027 case Intrinsic::arm_neon_vld2:
11028 case Intrinsic::arm_neon_vld3:
11029 case Intrinsic::arm_neon_vld4:
11030 case Intrinsic::arm_neon_vld2lane:
11031 case Intrinsic::arm_neon_vld3lane:
11032 case Intrinsic::arm_neon_vld4lane: {
11033 Info.opc = ISD::INTRINSIC_W_CHAIN;
11034 // Conservatively set memVT to the entire set of vectors loaded.
11035 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11036 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11037 Info.ptrVal = I.getArgOperand(0);
11039 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11040 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11041 Info.vol = false; // volatile loads with NEON intrinsics not supported
11042 Info.readMem = true;
11043 Info.writeMem = false;
11046 case Intrinsic::arm_neon_vst1:
11047 case Intrinsic::arm_neon_vst2:
11048 case Intrinsic::arm_neon_vst3:
11049 case Intrinsic::arm_neon_vst4:
11050 case Intrinsic::arm_neon_vst2lane:
11051 case Intrinsic::arm_neon_vst3lane:
11052 case Intrinsic::arm_neon_vst4lane: {
11053 Info.opc = ISD::INTRINSIC_VOID;
11054 // Conservatively set memVT to the entire set of vectors stored.
11055 unsigned NumElts = 0;
11056 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11057 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11058 if (!ArgTy->isVectorTy())
11060 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11062 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11063 Info.ptrVal = I.getArgOperand(0);
11065 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11066 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11067 Info.vol = false; // volatile stores with NEON intrinsics not supported
11068 Info.readMem = false;
11069 Info.writeMem = true;
11072 case Intrinsic::arm_ldrex: {
11073 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11074 Info.opc = ISD::INTRINSIC_W_CHAIN;
11075 Info.memVT = MVT::getVT(PtrTy->getElementType());
11076 Info.ptrVal = I.getArgOperand(0);
11078 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11080 Info.readMem = true;
11081 Info.writeMem = false;
11084 case Intrinsic::arm_strex: {
11085 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11086 Info.opc = ISD::INTRINSIC_W_CHAIN;
11087 Info.memVT = MVT::getVT(PtrTy->getElementType());
11088 Info.ptrVal = I.getArgOperand(1);
11090 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11092 Info.readMem = false;
11093 Info.writeMem = true;
11096 case Intrinsic::arm_strexd: {
11097 Info.opc = ISD::INTRINSIC_W_CHAIN;
11098 Info.memVT = MVT::i64;
11099 Info.ptrVal = I.getArgOperand(2);
11103 Info.readMem = false;
11104 Info.writeMem = true;
11107 case Intrinsic::arm_ldrexd: {
11108 Info.opc = ISD::INTRINSIC_W_CHAIN;
11109 Info.memVT = MVT::i64;
11110 Info.ptrVal = I.getArgOperand(0);
11114 Info.readMem = true;
11115 Info.writeMem = false;