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"
54 STATISTIC(NumTailCalls, "Number of tail calls");
55 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
56 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
58 // This option should go away when tail calls fully work.
60 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
61 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
65 EnableARMLongCalls("arm-long-calls", cl::Hidden,
66 cl::desc("Generate calls via indirect call instructions"),
70 ARMInterworking("arm-interworking", cl::Hidden,
71 cl::desc("Enable / disable ARM interworking (for debugging only)"),
75 class ARMCCState : public CCState {
77 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
78 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs,
79 LLVMContext &C, ParmContext PC)
80 : CCState(CC, isVarArg, MF, TM, locs, C) {
81 assert(((PC == Call) || (PC == Prologue)) &&
82 "ARMCCState users must specify whether their context is call"
83 "or prologue generation.");
89 // The APCS parameter registers.
90 static const uint16_t GPRArgRegs[] = {
91 ARM::R0, ARM::R1, ARM::R2, ARM::R3
94 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
95 MVT PromotedBitwiseVT) {
96 if (VT != PromotedLdStVT) {
97 setOperationAction(ISD::LOAD, VT, Promote);
98 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
100 setOperationAction(ISD::STORE, VT, Promote);
101 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
104 MVT ElemTy = VT.getVectorElementType();
105 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
106 setOperationAction(ISD::SETCC, VT, Custom);
107 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
108 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
109 if (ElemTy == MVT::i32) {
110 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
112 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
113 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
115 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
117 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
118 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
120 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
121 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
122 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
123 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
124 setOperationAction(ISD::SELECT, VT, Expand);
125 setOperationAction(ISD::SELECT_CC, VT, Expand);
126 setOperationAction(ISD::VSELECT, VT, Expand);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
128 if (VT.isInteger()) {
129 setOperationAction(ISD::SHL, VT, Custom);
130 setOperationAction(ISD::SRA, VT, Custom);
131 setOperationAction(ISD::SRL, VT, Custom);
134 // Promote all bit-wise operations.
135 if (VT.isInteger() && VT != PromotedBitwiseVT) {
136 setOperationAction(ISD::AND, VT, Promote);
137 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
138 setOperationAction(ISD::OR, VT, Promote);
139 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
140 setOperationAction(ISD::XOR, VT, Promote);
141 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
144 // Neon does not support vector divide/remainder operations.
145 setOperationAction(ISD::SDIV, VT, Expand);
146 setOperationAction(ISD::UDIV, VT, Expand);
147 setOperationAction(ISD::FDIV, VT, Expand);
148 setOperationAction(ISD::SREM, VT, Expand);
149 setOperationAction(ISD::UREM, VT, Expand);
150 setOperationAction(ISD::FREM, VT, Expand);
153 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
154 addRegisterClass(VT, &ARM::DPRRegClass);
155 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
158 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
159 addRegisterClass(VT, &ARM::QPRRegClass);
160 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
163 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
164 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
165 return new TargetLoweringObjectFileMachO();
167 return new ARMElfTargetObjectFile();
170 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
171 : TargetLowering(TM, createTLOF(TM)) {
172 Subtarget = &TM.getSubtarget<ARMSubtarget>();
173 RegInfo = TM.getRegisterInfo();
174 Itins = TM.getInstrItineraryData();
176 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
178 if (Subtarget->isTargetIOS()) {
179 // Uses VFP for Thumb libfuncs if available.
180 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
181 Subtarget->hasARMOps()) {
182 // Single-precision floating-point arithmetic.
183 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
184 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
185 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
186 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
188 // Double-precision floating-point arithmetic.
189 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
190 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
191 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
192 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
194 // Single-precision comparisons.
195 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
196 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
197 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
198 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
199 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
200 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
201 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
202 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
204 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
213 // Double-precision comparisons.
214 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
215 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
216 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
217 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
218 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
219 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
220 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
221 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
223 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
232 // Floating-point to integer conversions.
233 // i64 conversions are done via library routines even when generating VFP
234 // instructions, so use the same ones.
235 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
237 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
238 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
240 // Conversions between floating types.
241 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
242 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
244 // Integer to floating-point conversions.
245 // i64 conversions are done via library routines even when generating VFP
246 // instructions, so use the same ones.
247 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
248 // e.g., __floatunsidf vs. __floatunssidfvfp.
249 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
251 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
252 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
256 // These libcalls are not available in 32-bit.
257 setLibcallName(RTLIB::SHL_I128, 0);
258 setLibcallName(RTLIB::SRL_I128, 0);
259 setLibcallName(RTLIB::SRA_I128, 0);
261 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
262 // Double-precision floating-point arithmetic helper functions
263 // RTABI chapter 4.1.2, Table 2
264 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
265 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
266 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
267 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
268 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
270 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
273 // Double-precision floating-point comparison helper functions
274 // RTABI chapter 4.1.2, Table 3
275 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
277 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
278 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
279 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
280 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
282 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
284 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
286 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
287 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
289 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
290 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
291 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
300 // Single-precision floating-point arithmetic helper functions
301 // RTABI chapter 4.1.2, Table 4
302 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
303 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
304 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
305 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
306 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
308 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
311 // Single-precision floating-point comparison helper functions
312 // RTABI chapter 4.1.2, Table 5
313 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
315 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
316 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
317 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
318 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
320 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
322 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
324 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
325 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
327 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
328 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
329 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
338 // Floating-point to integer conversions.
339 // RTABI chapter 4.1.2, Table 6
340 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
342 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
343 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
346 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
347 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
357 // Conversions between floating types.
358 // RTABI chapter 4.1.2, Table 7
359 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
360 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
361 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
362 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
364 // Integer to floating-point conversions.
365 // RTABI chapter 4.1.2, Table 8
366 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
367 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
368 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
369 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
370 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
371 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
372 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
373 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
374 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
383 // Long long helper functions
384 // RTABI chapter 4.2, Table 9
385 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
386 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
387 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
388 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
389 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
396 // Integer division functions
397 // RTABI chapter 4.3.1
398 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
400 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
401 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
402 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
404 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
405 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
406 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
416 // RTABI chapter 4.3.4
417 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
418 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
419 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
420 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
421 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
422 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
425 // Use divmod compiler-rt calls for iOS 5.0 and later.
426 if (Subtarget->getTargetTriple().isiOS() &&
427 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
428 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
429 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
432 if (Subtarget->isThumb1Only())
433 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
435 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
436 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
437 !Subtarget->isThumb1Only()) {
438 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
439 if (!Subtarget->isFPOnlySP())
440 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
442 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
445 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
447 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
448 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
449 setTruncStoreAction((MVT::SimpleValueType)VT,
450 (MVT::SimpleValueType)InnerVT, Expand);
451 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
452 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
453 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
456 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
457 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
459 if (Subtarget->hasNEON()) {
460 addDRTypeForNEON(MVT::v2f32);
461 addDRTypeForNEON(MVT::v8i8);
462 addDRTypeForNEON(MVT::v4i16);
463 addDRTypeForNEON(MVT::v2i32);
464 addDRTypeForNEON(MVT::v1i64);
466 addQRTypeForNEON(MVT::v4f32);
467 addQRTypeForNEON(MVT::v2f64);
468 addQRTypeForNEON(MVT::v16i8);
469 addQRTypeForNEON(MVT::v8i16);
470 addQRTypeForNEON(MVT::v4i32);
471 addQRTypeForNEON(MVT::v2i64);
473 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
474 // neither Neon nor VFP support any arithmetic operations on it.
475 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
476 // supported for v4f32.
477 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
478 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
479 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
480 // FIXME: Code duplication: FDIV and FREM are expanded always, see
481 // ARMTargetLowering::addTypeForNEON method for details.
482 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
483 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
484 // FIXME: Create unittest.
485 // In another words, find a way when "copysign" appears in DAG with vector
487 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
488 // FIXME: Code duplication: SETCC has custom operation action, see
489 // ARMTargetLowering::addTypeForNEON method for details.
490 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
491 // FIXME: Create unittest for FNEG and for FABS.
492 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
493 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
495 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
496 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
497 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
498 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
499 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
500 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
501 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
502 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
503 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
504 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
505 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
506 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
507 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
508 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
509 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
510 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
512 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
513 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
514 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
515 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
516 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
518 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
520 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
521 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
522 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
523 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
524 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
525 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
526 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
528 // Mark v2f32 intrinsics.
529 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
530 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
531 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
532 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
533 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
534 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
535 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
536 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
537 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
538 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
539 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
540 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
541 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
542 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
543 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
545 // Neon does not support some operations on v1i64 and v2i64 types.
546 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
547 // Custom handling for some quad-vector types to detect VMULL.
548 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
549 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
550 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
551 // Custom handling for some vector types to avoid expensive expansions
552 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
553 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
554 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
555 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
556 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
557 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
558 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
559 // a destination type that is wider than the source, and nor does
560 // it have a FP_TO_[SU]INT instruction with a narrower destination than
562 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
563 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
564 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
565 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
567 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
568 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
570 // NEON does not have single instruction CTPOP for vectors with element
571 // types wider than 8-bits. However, custom lowering can leverage the
572 // v8i8/v16i8 vcnt instruction.
573 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
574 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
575 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
576 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
578 // NEON only has FMA instructions as of VFP4.
579 if (!Subtarget->hasVFP4()) {
580 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
581 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
584 setTargetDAGCombine(ISD::INTRINSIC_VOID);
585 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
586 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
587 setTargetDAGCombine(ISD::SHL);
588 setTargetDAGCombine(ISD::SRL);
589 setTargetDAGCombine(ISD::SRA);
590 setTargetDAGCombine(ISD::SIGN_EXTEND);
591 setTargetDAGCombine(ISD::ZERO_EXTEND);
592 setTargetDAGCombine(ISD::ANY_EXTEND);
593 setTargetDAGCombine(ISD::SELECT_CC);
594 setTargetDAGCombine(ISD::BUILD_VECTOR);
595 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
596 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
597 setTargetDAGCombine(ISD::STORE);
598 setTargetDAGCombine(ISD::FP_TO_SINT);
599 setTargetDAGCombine(ISD::FP_TO_UINT);
600 setTargetDAGCombine(ISD::FDIV);
602 // It is legal to extload from v4i8 to v4i16 or v4i32.
603 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
604 MVT::v4i16, MVT::v2i16,
606 for (unsigned i = 0; i < 6; ++i) {
607 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
608 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
609 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
613 // ARM and Thumb2 support UMLAL/SMLAL.
614 if (!Subtarget->isThumb1Only())
615 setTargetDAGCombine(ISD::ADDC);
618 computeRegisterProperties();
620 // ARM does not have f32 extending load.
621 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
623 // ARM does not have i1 sign extending load.
624 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
626 // ARM supports all 4 flavors of integer indexed load / store.
627 if (!Subtarget->isThumb1Only()) {
628 for (unsigned im = (unsigned)ISD::PRE_INC;
629 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
630 setIndexedLoadAction(im, MVT::i1, Legal);
631 setIndexedLoadAction(im, MVT::i8, Legal);
632 setIndexedLoadAction(im, MVT::i16, Legal);
633 setIndexedLoadAction(im, MVT::i32, Legal);
634 setIndexedStoreAction(im, MVT::i1, Legal);
635 setIndexedStoreAction(im, MVT::i8, Legal);
636 setIndexedStoreAction(im, MVT::i16, Legal);
637 setIndexedStoreAction(im, MVT::i32, Legal);
641 // i64 operation support.
642 setOperationAction(ISD::MUL, MVT::i64, Expand);
643 setOperationAction(ISD::MULHU, MVT::i32, Expand);
644 if (Subtarget->isThumb1Only()) {
645 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
646 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
648 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
649 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
650 setOperationAction(ISD::MULHS, MVT::i32, Expand);
652 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
653 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
654 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
655 setOperationAction(ISD::SRL, MVT::i64, Custom);
656 setOperationAction(ISD::SRA, MVT::i64, Custom);
658 if (!Subtarget->isThumb1Only()) {
659 // FIXME: We should do this for Thumb1 as well.
660 setOperationAction(ISD::ADDC, MVT::i32, Custom);
661 setOperationAction(ISD::ADDE, MVT::i32, Custom);
662 setOperationAction(ISD::SUBC, MVT::i32, Custom);
663 setOperationAction(ISD::SUBE, MVT::i32, Custom);
666 // ARM does not have ROTL.
667 setOperationAction(ISD::ROTL, MVT::i32, Expand);
668 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
669 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
670 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
671 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
673 // These just redirect to CTTZ and CTLZ on ARM.
674 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
675 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
677 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
679 // Only ARMv6 has BSWAP.
680 if (!Subtarget->hasV6Ops())
681 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
683 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
684 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
685 // These are expanded into libcalls if the cpu doesn't have HW divider.
686 setOperationAction(ISD::SDIV, MVT::i32, Expand);
687 setOperationAction(ISD::UDIV, MVT::i32, Expand);
690 // FIXME: Also set divmod for SREM on EABI
691 setOperationAction(ISD::SREM, MVT::i32, Expand);
692 setOperationAction(ISD::UREM, MVT::i32, Expand);
693 // Register based DivRem for AEABI (RTABI 4.2)
694 if (Subtarget->isTargetAEABI()) {
695 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
696 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
697 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
698 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
699 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
700 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
701 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
702 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
704 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
705 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
706 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
707 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
708 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
709 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
710 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
711 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
713 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
714 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
716 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
717 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
720 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
721 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
722 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
723 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
724 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
726 setOperationAction(ISD::TRAP, MVT::Other, Legal);
728 // Use the default implementation.
729 setOperationAction(ISD::VASTART, MVT::Other, Custom);
730 setOperationAction(ISD::VAARG, MVT::Other, Expand);
731 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
732 setOperationAction(ISD::VAEND, MVT::Other, Expand);
733 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
734 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
736 if (!Subtarget->isTargetDarwin()) {
737 // Non-Darwin platforms may return values in these registers via the
738 // personality function.
739 setExceptionPointerRegister(ARM::R0);
740 setExceptionSelectorRegister(ARM::R1);
743 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
744 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
745 // the default expansion.
746 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
747 // ATOMIC_FENCE needs custom lowering; the other 32-bit ones are legal and
749 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
750 // Custom lowering for 64-bit ops
751 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
752 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
753 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
754 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
755 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
756 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
757 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
758 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
759 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
760 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
761 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
762 // On v8, we have particularly efficient implementations of atomic fences
763 // if they can be combined with nearby atomic loads and stores.
764 if (!Subtarget->hasV8Ops()) {
765 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
766 setInsertFencesForAtomic(true);
768 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
770 // If there's anything we can use as a barrier, go through custom lowering
772 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
773 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
775 // Set them all for expansion, which will force libcalls.
776 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
777 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
778 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
779 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
780 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
781 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
782 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
783 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
784 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
785 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
786 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
787 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
788 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
789 // Unordered/Monotonic case.
790 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
791 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
794 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
796 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
797 if (!Subtarget->hasV6Ops()) {
798 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
799 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
801 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
803 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
804 !Subtarget->isThumb1Only()) {
805 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
806 // iff target supports vfp2.
807 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
808 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
811 // We want to custom lower some of our intrinsics.
812 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
813 if (Subtarget->isTargetDarwin()) {
814 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
815 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
816 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
819 setOperationAction(ISD::SETCC, MVT::i32, Expand);
820 setOperationAction(ISD::SETCC, MVT::f32, Expand);
821 setOperationAction(ISD::SETCC, MVT::f64, Expand);
822 setOperationAction(ISD::SELECT, MVT::i32, Custom);
823 setOperationAction(ISD::SELECT, MVT::f32, Custom);
824 setOperationAction(ISD::SELECT, MVT::f64, Custom);
825 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
826 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
827 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
829 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
830 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
831 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
832 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
833 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
835 // We don't support sin/cos/fmod/copysign/pow
836 setOperationAction(ISD::FSIN, MVT::f64, Expand);
837 setOperationAction(ISD::FSIN, MVT::f32, Expand);
838 setOperationAction(ISD::FCOS, MVT::f32, Expand);
839 setOperationAction(ISD::FCOS, MVT::f64, Expand);
840 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
841 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
842 setOperationAction(ISD::FREM, MVT::f64, Expand);
843 setOperationAction(ISD::FREM, MVT::f32, Expand);
844 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
845 !Subtarget->isThumb1Only()) {
846 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
847 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
849 setOperationAction(ISD::FPOW, MVT::f64, Expand);
850 setOperationAction(ISD::FPOW, MVT::f32, Expand);
852 if (!Subtarget->hasVFP4()) {
853 setOperationAction(ISD::FMA, MVT::f64, Expand);
854 setOperationAction(ISD::FMA, MVT::f32, Expand);
857 // Various VFP goodness
858 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
859 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
860 if (Subtarget->hasVFP2()) {
861 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
862 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
863 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
864 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
866 // Special handling for half-precision FP.
867 if (!Subtarget->hasFP16()) {
868 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
869 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
873 // Combine sin / cos into one node or libcall if possible.
874 if (Subtarget->hasSinCos()) {
875 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
876 setLibcallName(RTLIB::SINCOS_F64, "sincos");
877 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
878 // For iOS, we don't want to the normal expansion of a libcall to
879 // sincos. We want to issue a libcall to __sincos_stret.
880 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
881 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
885 // We have target-specific dag combine patterns for the following nodes:
886 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
887 setTargetDAGCombine(ISD::ADD);
888 setTargetDAGCombine(ISD::SUB);
889 setTargetDAGCombine(ISD::MUL);
890 setTargetDAGCombine(ISD::AND);
891 setTargetDAGCombine(ISD::OR);
892 setTargetDAGCombine(ISD::XOR);
894 if (Subtarget->hasV6Ops())
895 setTargetDAGCombine(ISD::SRL);
897 setStackPointerRegisterToSaveRestore(ARM::SP);
899 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
900 !Subtarget->hasVFP2())
901 setSchedulingPreference(Sched::RegPressure);
903 setSchedulingPreference(Sched::Hybrid);
905 //// temporary - rewrite interface to use type
906 MaxStoresPerMemset = 8;
907 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
908 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
909 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
910 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
911 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
913 // On ARM arguments smaller than 4 bytes are extended, so all arguments
914 // are at least 4 bytes aligned.
915 setMinStackArgumentAlignment(4);
917 // Prefer likely predicted branches to selects on out-of-order cores.
918 PredictableSelectIsExpensive = Subtarget->isLikeA9();
920 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
923 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
924 bool isThumb2, unsigned &LdrOpc,
926 static const unsigned LoadBares[4][2] = {{ARM::LDREXB, ARM::t2LDREXB},
927 {ARM::LDREXH, ARM::t2LDREXH},
928 {ARM::LDREX, ARM::t2LDREX},
929 {ARM::LDREXD, ARM::t2LDREXD}};
930 static const unsigned LoadAcqs[4][2] = {{ARM::LDAEXB, ARM::t2LDAEXB},
931 {ARM::LDAEXH, ARM::t2LDAEXH},
932 {ARM::LDAEX, ARM::t2LDAEX},
933 {ARM::LDAEXD, ARM::t2LDAEXD}};
934 static const unsigned StoreBares[4][2] = {{ARM::STREXB, ARM::t2STREXB},
935 {ARM::STREXH, ARM::t2STREXH},
936 {ARM::STREX, ARM::t2STREX},
937 {ARM::STREXD, ARM::t2STREXD}};
938 static const unsigned StoreRels[4][2] = {{ARM::STLEXB, ARM::t2STLEXB},
939 {ARM::STLEXH, ARM::t2STLEXH},
940 {ARM::STLEX, ARM::t2STLEX},
941 {ARM::STLEXD, ARM::t2STLEXD}};
943 const unsigned (*LoadOps)[2], (*StoreOps)[2];
944 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
949 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
950 StoreOps = StoreRels;
952 StoreOps = StoreBares;
954 assert(isPowerOf2_32(Size) && Size <= 8 &&
955 "unsupported size for atomic binary op!");
957 LdrOpc = LoadOps[Log2_32(Size)][isThumb2];
958 StrOpc = StoreOps[Log2_32(Size)][isThumb2];
961 // FIXME: It might make sense to define the representative register class as the
962 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
963 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
964 // SPR's representative would be DPR_VFP2. This should work well if register
965 // pressure tracking were modified such that a register use would increment the
966 // pressure of the register class's representative and all of it's super
967 // classes' representatives transitively. We have not implemented this because
968 // of the difficulty prior to coalescing of modeling operand register classes
969 // due to the common occurrence of cross class copies and subregister insertions
971 std::pair<const TargetRegisterClass*, uint8_t>
972 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
973 const TargetRegisterClass *RRC = 0;
975 switch (VT.SimpleTy) {
977 return TargetLowering::findRepresentativeClass(VT);
978 // Use DPR as representative register class for all floating point
979 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
980 // the cost is 1 for both f32 and f64.
981 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
982 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
983 RRC = &ARM::DPRRegClass;
984 // When NEON is used for SP, only half of the register file is available
985 // because operations that define both SP and DP results will be constrained
986 // to the VFP2 class (D0-D15). We currently model this constraint prior to
987 // coalescing by double-counting the SP regs. See the FIXME above.
988 if (Subtarget->useNEONForSinglePrecisionFP())
991 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
992 case MVT::v4f32: case MVT::v2f64:
993 RRC = &ARM::DPRRegClass;
997 RRC = &ARM::DPRRegClass;
1001 RRC = &ARM::DPRRegClass;
1005 return std::make_pair(RRC, Cost);
1008 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1011 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1012 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1013 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1014 case ARMISD::CALL: return "ARMISD::CALL";
1015 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1016 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1017 case ARMISD::tCALL: return "ARMISD::tCALL";
1018 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1019 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1020 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1021 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1022 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1023 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1024 case ARMISD::CMP: return "ARMISD::CMP";
1025 case ARMISD::CMN: return "ARMISD::CMN";
1026 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1027 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1028 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1029 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1030 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1032 case ARMISD::CMOV: return "ARMISD::CMOV";
1034 case ARMISD::RBIT: return "ARMISD::RBIT";
1036 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1037 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1038 case ARMISD::SITOF: return "ARMISD::SITOF";
1039 case ARMISD::UITOF: return "ARMISD::UITOF";
1041 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1042 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1043 case ARMISD::RRX: return "ARMISD::RRX";
1045 case ARMISD::ADDC: return "ARMISD::ADDC";
1046 case ARMISD::ADDE: return "ARMISD::ADDE";
1047 case ARMISD::SUBC: return "ARMISD::SUBC";
1048 case ARMISD::SUBE: return "ARMISD::SUBE";
1050 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1051 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1053 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1054 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1056 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1058 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1060 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1062 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1064 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1066 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1067 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1068 case ARMISD::VCGE: return "ARMISD::VCGE";
1069 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1070 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1071 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1072 case ARMISD::VCGT: return "ARMISD::VCGT";
1073 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1074 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1075 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1076 case ARMISD::VTST: return "ARMISD::VTST";
1078 case ARMISD::VSHL: return "ARMISD::VSHL";
1079 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1080 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1081 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1082 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1083 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1084 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1085 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1086 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1087 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1088 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1089 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1090 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1091 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1092 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1093 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1094 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1095 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1096 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1097 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1098 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1099 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1100 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1101 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1102 case ARMISD::VDUP: return "ARMISD::VDUP";
1103 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1104 case ARMISD::VEXT: return "ARMISD::VEXT";
1105 case ARMISD::VREV64: return "ARMISD::VREV64";
1106 case ARMISD::VREV32: return "ARMISD::VREV32";
1107 case ARMISD::VREV16: return "ARMISD::VREV16";
1108 case ARMISD::VZIP: return "ARMISD::VZIP";
1109 case ARMISD::VUZP: return "ARMISD::VUZP";
1110 case ARMISD::VTRN: return "ARMISD::VTRN";
1111 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1112 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1113 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1114 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1115 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1116 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1117 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1118 case ARMISD::FMAX: return "ARMISD::FMAX";
1119 case ARMISD::FMIN: return "ARMISD::FMIN";
1120 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1121 case ARMISD::VMINNM: return "ARMISD::VMIN";
1122 case ARMISD::BFI: return "ARMISD::BFI";
1123 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1124 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1125 case ARMISD::VBSL: return "ARMISD::VBSL";
1126 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1127 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1128 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1129 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1130 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1131 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1132 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1133 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1134 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1135 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1136 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1137 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1138 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1139 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1140 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1141 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1142 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1143 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1144 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1145 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1149 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1150 if (!VT.isVector()) return getPointerTy();
1151 return VT.changeVectorElementTypeToInteger();
1154 /// getRegClassFor - Return the register class that should be used for the
1155 /// specified value type.
1156 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1157 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1158 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1159 // load / store 4 to 8 consecutive D registers.
1160 if (Subtarget->hasNEON()) {
1161 if (VT == MVT::v4i64)
1162 return &ARM::QQPRRegClass;
1163 if (VT == MVT::v8i64)
1164 return &ARM::QQQQPRRegClass;
1166 return TargetLowering::getRegClassFor(VT);
1169 // Create a fast isel object.
1171 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1172 const TargetLibraryInfo *libInfo) const {
1173 return ARM::createFastISel(funcInfo, libInfo);
1176 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1177 /// be used for loads / stores from the global.
1178 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1179 return (Subtarget->isThumb1Only() ? 127 : 4095);
1182 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1183 unsigned NumVals = N->getNumValues();
1185 return Sched::RegPressure;
1187 for (unsigned i = 0; i != NumVals; ++i) {
1188 EVT VT = N->getValueType(i);
1189 if (VT == MVT::Glue || VT == MVT::Other)
1191 if (VT.isFloatingPoint() || VT.isVector())
1195 if (!N->isMachineOpcode())
1196 return Sched::RegPressure;
1198 // Load are scheduled for latency even if there instruction itinerary
1199 // is not available.
1200 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1201 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1203 if (MCID.getNumDefs() == 0)
1204 return Sched::RegPressure;
1205 if (!Itins->isEmpty() &&
1206 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1209 return Sched::RegPressure;
1212 //===----------------------------------------------------------------------===//
1214 //===----------------------------------------------------------------------===//
1216 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1217 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1219 default: llvm_unreachable("Unknown condition code!");
1220 case ISD::SETNE: return ARMCC::NE;
1221 case ISD::SETEQ: return ARMCC::EQ;
1222 case ISD::SETGT: return ARMCC::GT;
1223 case ISD::SETGE: return ARMCC::GE;
1224 case ISD::SETLT: return ARMCC::LT;
1225 case ISD::SETLE: return ARMCC::LE;
1226 case ISD::SETUGT: return ARMCC::HI;
1227 case ISD::SETUGE: return ARMCC::HS;
1228 case ISD::SETULT: return ARMCC::LO;
1229 case ISD::SETULE: return ARMCC::LS;
1233 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1234 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1235 ARMCC::CondCodes &CondCode2) {
1236 CondCode2 = ARMCC::AL;
1238 default: llvm_unreachable("Unknown FP condition!");
1240 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1242 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1244 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1245 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1246 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1247 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1248 case ISD::SETO: CondCode = ARMCC::VC; break;
1249 case ISD::SETUO: CondCode = ARMCC::VS; break;
1250 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1251 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1252 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1254 case ISD::SETULT: CondCode = ARMCC::LT; break;
1256 case ISD::SETULE: CondCode = ARMCC::LE; break;
1258 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1262 //===----------------------------------------------------------------------===//
1263 // Calling Convention Implementation
1264 //===----------------------------------------------------------------------===//
1266 #include "ARMGenCallingConv.inc"
1268 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1269 /// given CallingConvention value.
1270 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1272 bool isVarArg) const {
1275 llvm_unreachable("Unsupported calling convention");
1276 case CallingConv::Fast:
1277 if (Subtarget->hasVFP2() && !isVarArg) {
1278 if (!Subtarget->isAAPCS_ABI())
1279 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1280 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1281 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1284 case CallingConv::C: {
1285 // Use target triple & subtarget features to do actual dispatch.
1286 if (!Subtarget->isAAPCS_ABI())
1287 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1288 else if (Subtarget->hasVFP2() &&
1289 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1291 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1292 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1294 case CallingConv::ARM_AAPCS_VFP:
1296 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1298 case CallingConv::ARM_AAPCS:
1299 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1300 case CallingConv::ARM_APCS:
1301 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1302 case CallingConv::GHC:
1303 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1307 /// LowerCallResult - Lower the result values of a call into the
1308 /// appropriate copies out of appropriate physical registers.
1310 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1311 CallingConv::ID CallConv, bool isVarArg,
1312 const SmallVectorImpl<ISD::InputArg> &Ins,
1313 SDLoc dl, SelectionDAG &DAG,
1314 SmallVectorImpl<SDValue> &InVals,
1315 bool isThisReturn, SDValue ThisVal) const {
1317 // Assign locations to each value returned by this call.
1318 SmallVector<CCValAssign, 16> RVLocs;
1319 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1320 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1321 CCInfo.AnalyzeCallResult(Ins,
1322 CCAssignFnForNode(CallConv, /* Return*/ true,
1325 // Copy all of the result registers out of their specified physreg.
1326 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1327 CCValAssign VA = RVLocs[i];
1329 // Pass 'this' value directly from the argument to return value, to avoid
1330 // reg unit interference
1331 if (i == 0 && isThisReturn) {
1332 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1333 "unexpected return calling convention register assignment");
1334 InVals.push_back(ThisVal);
1339 if (VA.needsCustom()) {
1340 // Handle f64 or half of a v2f64.
1341 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1343 Chain = Lo.getValue(1);
1344 InFlag = Lo.getValue(2);
1345 VA = RVLocs[++i]; // skip ahead to next loc
1346 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1348 Chain = Hi.getValue(1);
1349 InFlag = Hi.getValue(2);
1350 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1352 if (VA.getLocVT() == MVT::v2f64) {
1353 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1354 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1355 DAG.getConstant(0, MVT::i32));
1357 VA = RVLocs[++i]; // skip ahead to next loc
1358 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1359 Chain = Lo.getValue(1);
1360 InFlag = Lo.getValue(2);
1361 VA = RVLocs[++i]; // skip ahead to next loc
1362 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1363 Chain = Hi.getValue(1);
1364 InFlag = Hi.getValue(2);
1365 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1366 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1367 DAG.getConstant(1, MVT::i32));
1370 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1372 Chain = Val.getValue(1);
1373 InFlag = Val.getValue(2);
1376 switch (VA.getLocInfo()) {
1377 default: llvm_unreachable("Unknown loc info!");
1378 case CCValAssign::Full: break;
1379 case CCValAssign::BCvt:
1380 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1384 InVals.push_back(Val);
1390 /// LowerMemOpCallTo - Store the argument to the stack.
1392 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1393 SDValue StackPtr, SDValue Arg,
1394 SDLoc dl, SelectionDAG &DAG,
1395 const CCValAssign &VA,
1396 ISD::ArgFlagsTy Flags) const {
1397 unsigned LocMemOffset = VA.getLocMemOffset();
1398 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1399 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1400 return DAG.getStore(Chain, dl, Arg, PtrOff,
1401 MachinePointerInfo::getStack(LocMemOffset),
1405 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1406 SDValue Chain, SDValue &Arg,
1407 RegsToPassVector &RegsToPass,
1408 CCValAssign &VA, CCValAssign &NextVA,
1410 SmallVectorImpl<SDValue> &MemOpChains,
1411 ISD::ArgFlagsTy Flags) const {
1413 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1414 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1415 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1417 if (NextVA.isRegLoc())
1418 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1420 assert(NextVA.isMemLoc());
1421 if (StackPtr.getNode() == 0)
1422 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1424 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1430 /// LowerCall - Lowering a call into a callseq_start <-
1431 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1434 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1435 SmallVectorImpl<SDValue> &InVals) const {
1436 SelectionDAG &DAG = CLI.DAG;
1438 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1439 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1440 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1441 SDValue Chain = CLI.Chain;
1442 SDValue Callee = CLI.Callee;
1443 bool &isTailCall = CLI.IsTailCall;
1444 CallingConv::ID CallConv = CLI.CallConv;
1445 bool doesNotRet = CLI.DoesNotReturn;
1446 bool isVarArg = CLI.IsVarArg;
1448 MachineFunction &MF = DAG.getMachineFunction();
1449 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1450 bool isThisReturn = false;
1451 bool isSibCall = false;
1452 // Disable tail calls if they're not supported.
1453 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1456 // Check if it's really possible to do a tail call.
1457 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1458 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1459 Outs, OutVals, Ins, DAG);
1460 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1461 // detected sibcalls.
1468 // Analyze operands of the call, assigning locations to each operand.
1469 SmallVector<CCValAssign, 16> ArgLocs;
1470 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1471 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1472 CCInfo.AnalyzeCallOperands(Outs,
1473 CCAssignFnForNode(CallConv, /* Return*/ false,
1476 // Get a count of how many bytes are to be pushed on the stack.
1477 unsigned NumBytes = CCInfo.getNextStackOffset();
1479 // For tail calls, memory operands are available in our caller's stack.
1483 // Adjust the stack pointer for the new arguments...
1484 // These operations are automatically eliminated by the prolog/epilog pass
1486 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1489 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1491 RegsToPassVector RegsToPass;
1492 SmallVector<SDValue, 8> MemOpChains;
1494 // Walk the register/memloc assignments, inserting copies/loads. In the case
1495 // of tail call optimization, arguments are handled later.
1496 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1498 ++i, ++realArgIdx) {
1499 CCValAssign &VA = ArgLocs[i];
1500 SDValue Arg = OutVals[realArgIdx];
1501 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1502 bool isByVal = Flags.isByVal();
1504 // Promote the value if needed.
1505 switch (VA.getLocInfo()) {
1506 default: llvm_unreachable("Unknown loc info!");
1507 case CCValAssign::Full: break;
1508 case CCValAssign::SExt:
1509 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1511 case CCValAssign::ZExt:
1512 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1514 case CCValAssign::AExt:
1515 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1517 case CCValAssign::BCvt:
1518 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1522 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1523 if (VA.needsCustom()) {
1524 if (VA.getLocVT() == MVT::v2f64) {
1525 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1526 DAG.getConstant(0, MVT::i32));
1527 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1528 DAG.getConstant(1, MVT::i32));
1530 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1531 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1533 VA = ArgLocs[++i]; // skip ahead to next loc
1534 if (VA.isRegLoc()) {
1535 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1536 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1538 assert(VA.isMemLoc());
1540 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1541 dl, DAG, VA, Flags));
1544 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1545 StackPtr, MemOpChains, Flags);
1547 } else if (VA.isRegLoc()) {
1548 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1549 assert(VA.getLocVT() == MVT::i32 &&
1550 "unexpected calling convention register assignment");
1551 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1552 "unexpected use of 'returned'");
1553 isThisReturn = true;
1555 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1556 } else if (isByVal) {
1557 assert(VA.isMemLoc());
1558 unsigned offset = 0;
1560 // True if this byval aggregate will be split between registers
1562 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1563 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1565 if (CurByValIdx < ByValArgsCount) {
1567 unsigned RegBegin, RegEnd;
1568 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1570 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1572 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1573 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1574 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1575 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1576 MachinePointerInfo(),
1577 false, false, false,
1578 DAG.InferPtrAlignment(AddArg));
1579 MemOpChains.push_back(Load.getValue(1));
1580 RegsToPass.push_back(std::make_pair(j, Load));
1583 // If parameter size outsides register area, "offset" value
1584 // helps us to calculate stack slot for remained part properly.
1585 offset = RegEnd - RegBegin;
1587 CCInfo.nextInRegsParam();
1590 if (Flags.getByValSize() > 4*offset) {
1591 unsigned LocMemOffset = VA.getLocMemOffset();
1592 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1593 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1595 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1596 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1597 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1599 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1601 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1602 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1603 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1604 Ops, array_lengthof(Ops)));
1606 } else if (!isSibCall) {
1607 assert(VA.isMemLoc());
1609 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1610 dl, DAG, VA, Flags));
1614 if (!MemOpChains.empty())
1615 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1616 &MemOpChains[0], MemOpChains.size());
1618 // Build a sequence of copy-to-reg nodes chained together with token chain
1619 // and flag operands which copy the outgoing args into the appropriate regs.
1621 // Tail call byval lowering might overwrite argument registers so in case of
1622 // tail call optimization the copies to registers are lowered later.
1624 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1625 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1626 RegsToPass[i].second, InFlag);
1627 InFlag = Chain.getValue(1);
1630 // For tail calls lower the arguments to the 'real' stack slot.
1632 // Force all the incoming stack arguments to be loaded from the stack
1633 // before any new outgoing arguments are stored to the stack, because the
1634 // outgoing stack slots may alias the incoming argument stack slots, and
1635 // the alias isn't otherwise explicit. This is slightly more conservative
1636 // than necessary, because it means that each store effectively depends
1637 // on every argument instead of just those arguments it would clobber.
1639 // Do not flag preceding copytoreg stuff together with the following stuff.
1641 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1642 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1643 RegsToPass[i].second, InFlag);
1644 InFlag = Chain.getValue(1);
1649 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1650 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1651 // node so that legalize doesn't hack it.
1652 bool isDirect = false;
1653 bool isARMFunc = false;
1654 bool isLocalARMFunc = false;
1655 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1657 if (EnableARMLongCalls) {
1658 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1659 && "long-calls with non-static relocation model!");
1660 // Handle a global address or an external symbol. If it's not one of
1661 // those, the target's already in a register, so we don't need to do
1663 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1664 const GlobalValue *GV = G->getGlobal();
1665 // Create a constant pool entry for the callee address
1666 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1667 ARMConstantPoolValue *CPV =
1668 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1670 // Get the address of the callee into a register
1671 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1672 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1673 Callee = DAG.getLoad(getPointerTy(), dl,
1674 DAG.getEntryNode(), CPAddr,
1675 MachinePointerInfo::getConstantPool(),
1676 false, false, false, 0);
1677 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1678 const char *Sym = S->getSymbol();
1680 // Create a constant pool entry for the callee address
1681 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1682 ARMConstantPoolValue *CPV =
1683 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1684 ARMPCLabelIndex, 0);
1685 // Get the address of the callee into a register
1686 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1687 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1688 Callee = DAG.getLoad(getPointerTy(), dl,
1689 DAG.getEntryNode(), CPAddr,
1690 MachinePointerInfo::getConstantPool(),
1691 false, false, false, 0);
1693 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1694 const GlobalValue *GV = G->getGlobal();
1696 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1697 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1698 getTargetMachine().getRelocationModel() != Reloc::Static;
1699 isARMFunc = !Subtarget->isThumb() || isStub;
1700 // ARM call to a local ARM function is predicable.
1701 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1702 // tBX takes a register source operand.
1703 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1704 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1705 ARMConstantPoolValue *CPV =
1706 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1707 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1708 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1709 Callee = DAG.getLoad(getPointerTy(), dl,
1710 DAG.getEntryNode(), CPAddr,
1711 MachinePointerInfo::getConstantPool(),
1712 false, false, false, 0);
1713 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1714 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1715 getPointerTy(), Callee, PICLabel);
1717 // On ELF targets for PIC code, direct calls should go through the PLT
1718 unsigned OpFlags = 0;
1719 if (Subtarget->isTargetELF() &&
1720 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1721 OpFlags = ARMII::MO_PLT;
1722 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1724 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1726 bool isStub = Subtarget->isTargetDarwin() &&
1727 getTargetMachine().getRelocationModel() != Reloc::Static;
1728 isARMFunc = !Subtarget->isThumb() || isStub;
1729 // tBX takes a register source operand.
1730 const char *Sym = S->getSymbol();
1731 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1732 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1733 ARMConstantPoolValue *CPV =
1734 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1735 ARMPCLabelIndex, 4);
1736 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1737 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1738 Callee = DAG.getLoad(getPointerTy(), dl,
1739 DAG.getEntryNode(), CPAddr,
1740 MachinePointerInfo::getConstantPool(),
1741 false, false, false, 0);
1742 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1743 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1744 getPointerTy(), Callee, PICLabel);
1746 unsigned OpFlags = 0;
1747 // On ELF targets for PIC code, direct calls should go through the PLT
1748 if (Subtarget->isTargetELF() &&
1749 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1750 OpFlags = ARMII::MO_PLT;
1751 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1755 // FIXME: handle tail calls differently.
1757 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1758 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1759 if (Subtarget->isThumb()) {
1760 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1761 CallOpc = ARMISD::CALL_NOLINK;
1763 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1765 if (!isDirect && !Subtarget->hasV5TOps())
1766 CallOpc = ARMISD::CALL_NOLINK;
1767 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1768 // Emit regular call when code size is the priority
1770 // "mov lr, pc; b _foo" to avoid confusing the RSP
1771 CallOpc = ARMISD::CALL_NOLINK;
1773 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1776 std::vector<SDValue> Ops;
1777 Ops.push_back(Chain);
1778 Ops.push_back(Callee);
1780 // Add argument registers to the end of the list so that they are known live
1782 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1783 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1784 RegsToPass[i].second.getValueType()));
1786 // Add a register mask operand representing the call-preserved registers.
1788 const uint32_t *Mask;
1789 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1790 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1792 // For 'this' returns, use the R0-preserving mask if applicable
1793 Mask = ARI->getThisReturnPreservedMask(CallConv);
1795 // Set isThisReturn to false if the calling convention is not one that
1796 // allows 'returned' to be modeled in this way, so LowerCallResult does
1797 // not try to pass 'this' straight through
1798 isThisReturn = false;
1799 Mask = ARI->getCallPreservedMask(CallConv);
1802 Mask = ARI->getCallPreservedMask(CallConv);
1804 assert(Mask && "Missing call preserved mask for calling convention");
1805 Ops.push_back(DAG.getRegisterMask(Mask));
1808 if (InFlag.getNode())
1809 Ops.push_back(InFlag);
1811 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1813 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1815 // Returns a chain and a flag for retval copy to use.
1816 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1817 InFlag = Chain.getValue(1);
1819 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1820 DAG.getIntPtrConstant(0, true), InFlag, dl);
1822 InFlag = Chain.getValue(1);
1824 // Handle result values, copying them out of physregs into vregs that we
1826 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1827 InVals, isThisReturn,
1828 isThisReturn ? OutVals[0] : SDValue());
1831 /// HandleByVal - Every parameter *after* a byval parameter is passed
1832 /// on the stack. Remember the next parameter register to allocate,
1833 /// and then confiscate the rest of the parameter registers to insure
1836 ARMTargetLowering::HandleByVal(
1837 CCState *State, unsigned &size, unsigned Align) const {
1838 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1839 assert((State->getCallOrPrologue() == Prologue ||
1840 State->getCallOrPrologue() == Call) &&
1841 "unhandled ParmContext");
1843 // For in-prologue parameters handling, we also introduce stack offset
1844 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
1845 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
1846 // NSAA should be evaluted (NSAA means "next stacked argument address").
1847 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
1848 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
1849 unsigned NSAAOffset = State->getNextStackOffset();
1850 if (State->getCallOrPrologue() != Call) {
1851 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
1853 State->getInRegsParamInfo(i, RB, RE);
1854 assert(NSAAOffset >= (RE-RB)*4 &&
1855 "Stack offset for byval regs doesn't introduced anymore?");
1856 NSAAOffset -= (RE-RB)*4;
1859 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1860 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1861 unsigned AlignInRegs = Align / 4;
1862 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1863 for (unsigned i = 0; i < Waste; ++i)
1864 reg = State->AllocateReg(GPRArgRegs, 4);
1867 unsigned excess = 4 * (ARM::R4 - reg);
1869 // Special case when NSAA != SP and parameter size greater than size of
1870 // all remained GPR regs. In that case we can't split parameter, we must
1871 // send it to stack. We also must set NCRN to R4, so waste all
1872 // remained registers.
1873 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1874 while (State->AllocateReg(GPRArgRegs, 4))
1879 // First register for byval parameter is the first register that wasn't
1880 // allocated before this method call, so it would be "reg".
1881 // If parameter is small enough to be saved in range [reg, r4), then
1882 // the end (first after last) register would be reg + param-size-in-regs,
1883 // else parameter would be splitted between registers and stack,
1884 // end register would be r4 in this case.
1885 unsigned ByValRegBegin = reg;
1886 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1887 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1888 // Note, first register is allocated in the beginning of function already,
1889 // allocate remained amount of registers we need.
1890 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1891 State->AllocateReg(GPRArgRegs, 4);
1892 // At a call site, a byval parameter that is split between
1893 // registers and memory needs its size truncated here. In a
1894 // function prologue, such byval parameters are reassembled in
1895 // memory, and are not truncated.
1896 if (State->getCallOrPrologue() == Call) {
1897 // Make remained size equal to 0 in case, when
1898 // the whole structure may be stored into registers.
1908 /// MatchingStackOffset - Return true if the given stack call argument is
1909 /// already available in the same position (relatively) of the caller's
1910 /// incoming argument stack.
1912 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1913 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1914 const TargetInstrInfo *TII) {
1915 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1917 if (Arg.getOpcode() == ISD::CopyFromReg) {
1918 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1919 if (!TargetRegisterInfo::isVirtualRegister(VR))
1921 MachineInstr *Def = MRI->getVRegDef(VR);
1924 if (!Flags.isByVal()) {
1925 if (!TII->isLoadFromStackSlot(Def, FI))
1930 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1931 if (Flags.isByVal())
1932 // ByVal argument is passed in as a pointer but it's now being
1933 // dereferenced. e.g.
1934 // define @foo(%struct.X* %A) {
1935 // tail call @bar(%struct.X* byval %A)
1938 SDValue Ptr = Ld->getBasePtr();
1939 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1942 FI = FINode->getIndex();
1946 assert(FI != INT_MAX);
1947 if (!MFI->isFixedObjectIndex(FI))
1949 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1952 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1953 /// for tail call optimization. Targets which want to do tail call
1954 /// optimization should implement this function.
1956 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1957 CallingConv::ID CalleeCC,
1959 bool isCalleeStructRet,
1960 bool isCallerStructRet,
1961 const SmallVectorImpl<ISD::OutputArg> &Outs,
1962 const SmallVectorImpl<SDValue> &OutVals,
1963 const SmallVectorImpl<ISD::InputArg> &Ins,
1964 SelectionDAG& DAG) const {
1965 const Function *CallerF = DAG.getMachineFunction().getFunction();
1966 CallingConv::ID CallerCC = CallerF->getCallingConv();
1967 bool CCMatch = CallerCC == CalleeCC;
1969 // Look for obvious safe cases to perform tail call optimization that do not
1970 // require ABI changes. This is what gcc calls sibcall.
1972 // Do not sibcall optimize vararg calls unless the call site is not passing
1974 if (isVarArg && !Outs.empty())
1977 // Exception-handling functions need a special set of instructions to indicate
1978 // a return to the hardware. Tail-calling another function would probably
1980 if (CallerF->hasFnAttribute("interrupt"))
1983 // Also avoid sibcall optimization if either caller or callee uses struct
1984 // return semantics.
1985 if (isCalleeStructRet || isCallerStructRet)
1988 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1989 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1990 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1991 // support in the assembler and linker to be used. This would need to be
1992 // fixed to fully support tail calls in Thumb1.
1994 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1995 // LR. This means if we need to reload LR, it takes an extra instructions,
1996 // which outweighs the value of the tail call; but here we don't know yet
1997 // whether LR is going to be used. Probably the right approach is to
1998 // generate the tail call here and turn it back into CALL/RET in
1999 // emitEpilogue if LR is used.
2001 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2002 // but we need to make sure there are enough registers; the only valid
2003 // registers are the 4 used for parameters. We don't currently do this
2005 if (Subtarget->isThumb1Only())
2008 // If the calling conventions do not match, then we'd better make sure the
2009 // results are returned in the same way as what the caller expects.
2011 SmallVector<CCValAssign, 16> RVLocs1;
2012 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2013 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
2014 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2016 SmallVector<CCValAssign, 16> RVLocs2;
2017 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2018 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
2019 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2021 if (RVLocs1.size() != RVLocs2.size())
2023 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2024 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2026 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2028 if (RVLocs1[i].isRegLoc()) {
2029 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2032 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2038 // If Caller's vararg or byval argument has been split between registers and
2039 // stack, do not perform tail call, since part of the argument is in caller's
2041 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2042 getInfo<ARMFunctionInfo>();
2043 if (AFI_Caller->getArgRegsSaveSize())
2046 // If the callee takes no arguments then go on to check the results of the
2048 if (!Outs.empty()) {
2049 // Check if stack adjustment is needed. For now, do not do this if any
2050 // argument is passed on the stack.
2051 SmallVector<CCValAssign, 16> ArgLocs;
2052 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2053 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2054 CCInfo.AnalyzeCallOperands(Outs,
2055 CCAssignFnForNode(CalleeCC, false, isVarArg));
2056 if (CCInfo.getNextStackOffset()) {
2057 MachineFunction &MF = DAG.getMachineFunction();
2059 // Check if the arguments are already laid out in the right way as
2060 // the caller's fixed stack objects.
2061 MachineFrameInfo *MFI = MF.getFrameInfo();
2062 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2063 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2064 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2066 ++i, ++realArgIdx) {
2067 CCValAssign &VA = ArgLocs[i];
2068 EVT RegVT = VA.getLocVT();
2069 SDValue Arg = OutVals[realArgIdx];
2070 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2071 if (VA.getLocInfo() == CCValAssign::Indirect)
2073 if (VA.needsCustom()) {
2074 // f64 and vector types are split into multiple registers or
2075 // register/stack-slot combinations. The types will not match
2076 // the registers; give up on memory f64 refs until we figure
2077 // out what to do about this.
2080 if (!ArgLocs[++i].isRegLoc())
2082 if (RegVT == MVT::v2f64) {
2083 if (!ArgLocs[++i].isRegLoc())
2085 if (!ArgLocs[++i].isRegLoc())
2088 } else if (!VA.isRegLoc()) {
2089 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2101 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2102 MachineFunction &MF, bool isVarArg,
2103 const SmallVectorImpl<ISD::OutputArg> &Outs,
2104 LLVMContext &Context) const {
2105 SmallVector<CCValAssign, 16> RVLocs;
2106 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2107 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2111 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2112 SDLoc DL, SelectionDAG &DAG) {
2113 const MachineFunction &MF = DAG.getMachineFunction();
2114 const Function *F = MF.getFunction();
2116 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2118 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2119 // version of the "preferred return address". These offsets affect the return
2120 // instruction if this is a return from PL1 without hypervisor extensions.
2121 // IRQ/FIQ: +4 "subs pc, lr, #4"
2122 // SWI: 0 "subs pc, lr, #0"
2123 // ABORT: +4 "subs pc, lr, #4"
2124 // UNDEF: +4/+2 "subs pc, lr, #0"
2125 // UNDEF varies depending on where the exception came from ARM or Thumb
2126 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2129 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2132 else if (IntKind == "SWI" || IntKind == "UNDEF")
2135 report_fatal_error("Unsupported interrupt attribute. If present, value "
2136 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2138 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2140 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other,
2141 RetOps.data(), RetOps.size());
2145 ARMTargetLowering::LowerReturn(SDValue Chain,
2146 CallingConv::ID CallConv, bool isVarArg,
2147 const SmallVectorImpl<ISD::OutputArg> &Outs,
2148 const SmallVectorImpl<SDValue> &OutVals,
2149 SDLoc dl, SelectionDAG &DAG) const {
2151 // CCValAssign - represent the assignment of the return value to a location.
2152 SmallVector<CCValAssign, 16> RVLocs;
2154 // CCState - Info about the registers and stack slots.
2155 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2156 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2158 // Analyze outgoing return values.
2159 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2163 SmallVector<SDValue, 4> RetOps;
2164 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2166 // Copy the result values into the output registers.
2167 for (unsigned i = 0, realRVLocIdx = 0;
2169 ++i, ++realRVLocIdx) {
2170 CCValAssign &VA = RVLocs[i];
2171 assert(VA.isRegLoc() && "Can only return in registers!");
2173 SDValue Arg = OutVals[realRVLocIdx];
2175 switch (VA.getLocInfo()) {
2176 default: llvm_unreachable("Unknown loc info!");
2177 case CCValAssign::Full: break;
2178 case CCValAssign::BCvt:
2179 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2183 if (VA.needsCustom()) {
2184 if (VA.getLocVT() == MVT::v2f64) {
2185 // Extract the first half and return it in two registers.
2186 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2187 DAG.getConstant(0, MVT::i32));
2188 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2189 DAG.getVTList(MVT::i32, MVT::i32), Half);
2191 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2192 Flag = Chain.getValue(1);
2193 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2194 VA = RVLocs[++i]; // skip ahead to next loc
2195 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2196 HalfGPRs.getValue(1), Flag);
2197 Flag = Chain.getValue(1);
2198 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2199 VA = RVLocs[++i]; // skip ahead to next loc
2201 // Extract the 2nd half and fall through to handle it as an f64 value.
2202 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2203 DAG.getConstant(1, MVT::i32));
2205 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2207 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2208 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2209 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2210 Flag = Chain.getValue(1);
2211 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2212 VA = RVLocs[++i]; // skip ahead to next loc
2213 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2216 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2218 // Guarantee that all emitted copies are
2219 // stuck together, avoiding something bad.
2220 Flag = Chain.getValue(1);
2221 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2224 // Update chain and glue.
2227 RetOps.push_back(Flag);
2229 // CPUs which aren't M-class use a special sequence to return from
2230 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2231 // though we use "subs pc, lr, #N").
2233 // M-class CPUs actually use a normal return sequence with a special
2234 // (hardware-provided) value in LR, so the normal code path works.
2235 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2236 !Subtarget->isMClass()) {
2237 if (Subtarget->isThumb1Only())
2238 report_fatal_error("interrupt attribute is not supported in Thumb1");
2239 return LowerInterruptReturn(RetOps, dl, DAG);
2242 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2243 RetOps.data(), RetOps.size());
2246 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2247 if (N->getNumValues() != 1)
2249 if (!N->hasNUsesOfValue(1, 0))
2252 SDValue TCChain = Chain;
2253 SDNode *Copy = *N->use_begin();
2254 if (Copy->getOpcode() == ISD::CopyToReg) {
2255 // If the copy has a glue operand, we conservatively assume it isn't safe to
2256 // perform a tail call.
2257 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2259 TCChain = Copy->getOperand(0);
2260 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2261 SDNode *VMov = Copy;
2262 // f64 returned in a pair of GPRs.
2263 SmallPtrSet<SDNode*, 2> Copies;
2264 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2266 if (UI->getOpcode() != ISD::CopyToReg)
2270 if (Copies.size() > 2)
2273 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2275 SDValue UseChain = UI->getOperand(0);
2276 if (Copies.count(UseChain.getNode()))
2283 } else if (Copy->getOpcode() == ISD::BITCAST) {
2284 // f32 returned in a single GPR.
2285 if (!Copy->hasOneUse())
2287 Copy = *Copy->use_begin();
2288 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2290 TCChain = Copy->getOperand(0);
2295 bool HasRet = false;
2296 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2298 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2299 UI->getOpcode() != ARMISD::INTRET_FLAG)
2311 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2312 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2315 if (!CI->isTailCall())
2318 return !Subtarget->isThumb1Only();
2321 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2322 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2323 // one of the above mentioned nodes. It has to be wrapped because otherwise
2324 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2325 // be used to form addressing mode. These wrapped nodes will be selected
2327 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2328 EVT PtrVT = Op.getValueType();
2329 // FIXME there is no actual debug info here
2331 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2333 if (CP->isMachineConstantPoolEntry())
2334 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2335 CP->getAlignment());
2337 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2338 CP->getAlignment());
2339 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2342 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2343 return MachineJumpTableInfo::EK_Inline;
2346 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2347 SelectionDAG &DAG) const {
2348 MachineFunction &MF = DAG.getMachineFunction();
2349 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2350 unsigned ARMPCLabelIndex = 0;
2352 EVT PtrVT = getPointerTy();
2353 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2354 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2356 if (RelocM == Reloc::Static) {
2357 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2359 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2360 ARMPCLabelIndex = AFI->createPICLabelUId();
2361 ARMConstantPoolValue *CPV =
2362 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2363 ARMCP::CPBlockAddress, PCAdj);
2364 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2366 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2367 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2368 MachinePointerInfo::getConstantPool(),
2369 false, false, false, 0);
2370 if (RelocM == Reloc::Static)
2372 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2373 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2376 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2378 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2379 SelectionDAG &DAG) const {
2381 EVT PtrVT = getPointerTy();
2382 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2383 MachineFunction &MF = DAG.getMachineFunction();
2384 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2385 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2386 ARMConstantPoolValue *CPV =
2387 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2388 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2389 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2390 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2391 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2392 MachinePointerInfo::getConstantPool(),
2393 false, false, false, 0);
2394 SDValue Chain = Argument.getValue(1);
2396 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2397 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2399 // call __tls_get_addr.
2402 Entry.Node = Argument;
2403 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2404 Args.push_back(Entry);
2405 // FIXME: is there useful debug info available here?
2406 TargetLowering::CallLoweringInfo CLI(Chain,
2407 (Type *) Type::getInt32Ty(*DAG.getContext()),
2408 false, false, false, false,
2409 0, CallingConv::C, /*isTailCall=*/false,
2410 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2411 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2412 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2413 return CallResult.first;
2416 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2417 // "local exec" model.
2419 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2421 TLSModel::Model model) const {
2422 const GlobalValue *GV = GA->getGlobal();
2425 SDValue Chain = DAG.getEntryNode();
2426 EVT PtrVT = getPointerTy();
2427 // Get the Thread Pointer
2428 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2430 if (model == TLSModel::InitialExec) {
2431 MachineFunction &MF = DAG.getMachineFunction();
2432 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2433 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2434 // Initial exec model.
2435 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2436 ARMConstantPoolValue *CPV =
2437 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2438 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2440 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2441 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2442 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2443 MachinePointerInfo::getConstantPool(),
2444 false, false, false, 0);
2445 Chain = Offset.getValue(1);
2447 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2448 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2450 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2451 MachinePointerInfo::getConstantPool(),
2452 false, false, false, 0);
2455 assert(model == TLSModel::LocalExec);
2456 ARMConstantPoolValue *CPV =
2457 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2458 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2459 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2460 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2461 MachinePointerInfo::getConstantPool(),
2462 false, false, false, 0);
2465 // The address of the thread local variable is the add of the thread
2466 // pointer with the offset of the variable.
2467 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2471 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2472 // TODO: implement the "local dynamic" model
2473 assert(Subtarget->isTargetELF() &&
2474 "TLS not implemented for non-ELF targets");
2475 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2477 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2480 case TLSModel::GeneralDynamic:
2481 case TLSModel::LocalDynamic:
2482 return LowerToTLSGeneralDynamicModel(GA, DAG);
2483 case TLSModel::InitialExec:
2484 case TLSModel::LocalExec:
2485 return LowerToTLSExecModels(GA, DAG, model);
2487 llvm_unreachable("bogus TLS model");
2490 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2491 SelectionDAG &DAG) const {
2492 EVT PtrVT = getPointerTy();
2494 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2495 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2496 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2497 ARMConstantPoolValue *CPV =
2498 ARMConstantPoolConstant::Create(GV,
2499 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2500 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2501 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2502 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2504 MachinePointerInfo::getConstantPool(),
2505 false, false, false, 0);
2506 SDValue Chain = Result.getValue(1);
2507 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2508 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2510 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2511 MachinePointerInfo::getGOT(),
2512 false, false, false, 0);
2516 // If we have T2 ops, we can materialize the address directly via movt/movw
2517 // pair. This is always cheaper.
2518 if (Subtarget->useMovt()) {
2520 // FIXME: Once remat is capable of dealing with instructions with register
2521 // operands, expand this into two nodes.
2522 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2523 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2525 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2526 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2527 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2528 MachinePointerInfo::getConstantPool(),
2529 false, false, false, 0);
2533 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2534 SelectionDAG &DAG) const {
2535 EVT PtrVT = getPointerTy();
2537 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2538 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2540 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2541 // update ARMFastISel::ARMMaterializeGV.
2542 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2544 // FIXME: Once remat is capable of dealing with instructions with register
2545 // operands, expand this into two nodes.
2546 if (RelocM == Reloc::Static)
2547 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2548 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2551 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2553 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2554 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2556 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2557 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2558 MachinePointerInfo::getGOT(),
2559 false, false, false, 0);
2563 unsigned ARMPCLabelIndex = 0;
2565 if (RelocM == Reloc::Static) {
2566 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2568 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2569 ARMPCLabelIndex = AFI->createPICLabelUId();
2570 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2571 ARMConstantPoolValue *CPV =
2572 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2574 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2576 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2578 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2579 MachinePointerInfo::getConstantPool(),
2580 false, false, false, 0);
2581 SDValue Chain = Result.getValue(1);
2583 if (RelocM == Reloc::PIC_) {
2584 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2585 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2588 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2589 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2590 false, false, false, 0);
2595 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2596 SelectionDAG &DAG) const {
2597 assert(Subtarget->isTargetELF() &&
2598 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2599 MachineFunction &MF = DAG.getMachineFunction();
2600 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2601 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2602 EVT PtrVT = getPointerTy();
2604 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2605 ARMConstantPoolValue *CPV =
2606 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2607 ARMPCLabelIndex, PCAdj);
2608 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2609 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2610 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2611 MachinePointerInfo::getConstantPool(),
2612 false, false, false, 0);
2613 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2614 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2618 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2620 SDValue Val = DAG.getConstant(0, MVT::i32);
2621 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2622 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2623 Op.getOperand(1), Val);
2627 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2629 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2630 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2634 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2635 const ARMSubtarget *Subtarget) const {
2636 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2639 default: return SDValue(); // Don't custom lower most intrinsics.
2640 case Intrinsic::arm_thread_pointer: {
2641 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2642 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2644 case Intrinsic::eh_sjlj_lsda: {
2645 MachineFunction &MF = DAG.getMachineFunction();
2646 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2647 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2648 EVT PtrVT = getPointerTy();
2649 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2651 unsigned PCAdj = (RelocM != Reloc::PIC_)
2652 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2653 ARMConstantPoolValue *CPV =
2654 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2655 ARMCP::CPLSDA, PCAdj);
2656 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2657 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2659 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2660 MachinePointerInfo::getConstantPool(),
2661 false, false, false, 0);
2663 if (RelocM == Reloc::PIC_) {
2664 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2665 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2669 case Intrinsic::arm_neon_vmulls:
2670 case Intrinsic::arm_neon_vmullu: {
2671 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2672 ? ARMISD::VMULLs : ARMISD::VMULLu;
2673 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2674 Op.getOperand(1), Op.getOperand(2));
2679 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2680 const ARMSubtarget *Subtarget) {
2681 // FIXME: handle "fence singlethread" more efficiently.
2683 if (!Subtarget->hasDataBarrier()) {
2684 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2685 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2687 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2688 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2689 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2690 DAG.getConstant(0, MVT::i32));
2693 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2694 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2695 unsigned Domain = ARM_MB::ISH;
2696 if (Subtarget->isMClass()) {
2697 // Only a full system barrier exists in the M-class architectures.
2698 Domain = ARM_MB::SY;
2699 } else if (Subtarget->isSwift() && Ord == Release) {
2700 // Swift happens to implement ISHST barriers in a way that's compatible with
2701 // Release semantics but weaker than ISH so we'd be fools not to use
2702 // it. Beware: other processors probably don't!
2703 Domain = ARM_MB::ISHST;
2706 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2707 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2708 DAG.getConstant(Domain, MVT::i32));
2711 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2712 const ARMSubtarget *Subtarget) {
2713 // ARM pre v5TE and Thumb1 does not have preload instructions.
2714 if (!(Subtarget->isThumb2() ||
2715 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2716 // Just preserve the chain.
2717 return Op.getOperand(0);
2720 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2722 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2723 // ARMv7 with MP extension has PLDW.
2724 return Op.getOperand(0);
2726 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2727 if (Subtarget->isThumb()) {
2729 isRead = ~isRead & 1;
2730 isData = ~isData & 1;
2733 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2734 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2735 DAG.getConstant(isData, MVT::i32));
2738 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2739 MachineFunction &MF = DAG.getMachineFunction();
2740 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2742 // vastart just stores the address of the VarArgsFrameIndex slot into the
2743 // memory location argument.
2745 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2746 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2747 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2748 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2749 MachinePointerInfo(SV), false, false, 0);
2753 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2754 SDValue &Root, SelectionDAG &DAG,
2756 MachineFunction &MF = DAG.getMachineFunction();
2757 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2759 const TargetRegisterClass *RC;
2760 if (AFI->isThumb1OnlyFunction())
2761 RC = &ARM::tGPRRegClass;
2763 RC = &ARM::GPRRegClass;
2765 // Transform the arguments stored in physical registers into virtual ones.
2766 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2767 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2770 if (NextVA.isMemLoc()) {
2771 MachineFrameInfo *MFI = MF.getFrameInfo();
2772 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2774 // Create load node to retrieve arguments from the stack.
2775 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2776 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2777 MachinePointerInfo::getFixedStack(FI),
2778 false, false, false, 0);
2780 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2781 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2784 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2788 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2789 unsigned InRegsParamRecordIdx,
2791 unsigned &ArgRegsSize,
2792 unsigned &ArgRegsSaveSize)
2795 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2796 unsigned RBegin, REnd;
2797 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2798 NumGPRs = REnd - RBegin;
2800 unsigned int firstUnalloced;
2801 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2802 sizeof(GPRArgRegs) /
2803 sizeof(GPRArgRegs[0]));
2804 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2807 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2808 ArgRegsSize = NumGPRs * 4;
2810 // If parameter is split between stack and GPRs...
2811 if (NumGPRs && Align == 8 &&
2812 (ArgRegsSize < ArgSize ||
2813 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2814 // Add padding for part of param recovered from GPRs, so
2815 // its last byte must be at address K*8 - 1.
2816 // We need to do it, since remained (stack) part of parameter has
2817 // stack alignment, and we need to "attach" "GPRs head" without gaps
2820 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2821 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2823 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2825 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
2826 (ArgRegsSize + AFI->getArgRegsSaveSize());
2827 ArgRegsSaveSize = ArgRegsSize + Padding;
2829 // We don't need to extend regs save size for byval parameters if they
2830 // are passed via GPRs only.
2831 ArgRegsSaveSize = ArgRegsSize;
2834 // The remaining GPRs hold either the beginning of variable-argument
2835 // data, or the beginning of an aggregate passed by value (usually
2836 // byval). Either way, we allocate stack slots adjacent to the data
2837 // provided by our caller, and store the unallocated registers there.
2838 // If this is a variadic function, the va_list pointer will begin with
2839 // these values; otherwise, this reassembles a (byval) structure that
2840 // was split between registers and memory.
2841 // Return: The frame index registers were stored into.
2843 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2844 SDLoc dl, SDValue &Chain,
2845 const Value *OrigArg,
2846 unsigned InRegsParamRecordIdx,
2847 unsigned OffsetFromOrigArg,
2850 bool ForceMutable) const {
2852 // Currently, two use-cases possible:
2853 // Case #1. Non var-args function, and we meet first byval parameter.
2854 // Setup first unallocated register as first byval register;
2855 // eat all remained registers
2856 // (these two actions are performed by HandleByVal method).
2857 // Then, here, we initialize stack frame with
2858 // "store-reg" instructions.
2859 // Case #2. Var-args function, that doesn't contain byval parameters.
2860 // The same: eat all remained unallocated registers,
2861 // initialize stack frame.
2863 MachineFunction &MF = DAG.getMachineFunction();
2864 MachineFrameInfo *MFI = MF.getFrameInfo();
2865 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2866 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2867 unsigned RBegin, REnd;
2868 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2869 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2870 firstRegToSaveIndex = RBegin - ARM::R0;
2871 lastRegToSaveIndex = REnd - ARM::R0;
2873 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2874 (GPRArgRegs, array_lengthof(GPRArgRegs));
2875 lastRegToSaveIndex = 4;
2878 unsigned ArgRegsSize, ArgRegsSaveSize;
2879 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2880 ArgRegsSize, ArgRegsSaveSize);
2882 // Store any by-val regs to their spots on the stack so that they may be
2883 // loaded by deferencing the result of formal parameter pointer or va_next.
2884 // Note: once stack area for byval/varargs registers
2885 // was initialized, it can't be initialized again.
2886 if (ArgRegsSaveSize) {
2888 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2891 assert(AFI->getStoredByValParamsPadding() == 0 &&
2892 "The only parameter may be padded.");
2893 AFI->setStoredByValParamsPadding(Padding);
2896 int FrameIndex = MFI->CreateFixedObject(
2898 Padding + ArgOffset,
2900 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2902 SmallVector<SDValue, 4> MemOps;
2903 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2904 ++firstRegToSaveIndex, ++i) {
2905 const TargetRegisterClass *RC;
2906 if (AFI->isThumb1OnlyFunction())
2907 RC = &ARM::tGPRRegClass;
2909 RC = &ARM::GPRRegClass;
2911 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2912 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2914 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2915 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2917 MemOps.push_back(Store);
2918 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2919 DAG.getConstant(4, getPointerTy()));
2922 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2924 if (!MemOps.empty())
2925 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2926 &MemOps[0], MemOps.size());
2929 // This will point to the next argument passed via stack.
2930 return MFI->CreateFixedObject(
2931 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
2934 // Setup stack frame, the va_list pointer will start from.
2936 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2937 SDLoc dl, SDValue &Chain,
2939 bool ForceMutable) const {
2940 MachineFunction &MF = DAG.getMachineFunction();
2941 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2943 // Try to store any remaining integer argument regs
2944 // to their spots on the stack so that they may be loaded by deferencing
2945 // the result of va_next.
2946 // If there is no regs to be stored, just point address after last
2947 // argument passed via stack.
2949 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2950 0, ArgOffset, 0, ForceMutable);
2952 AFI->setVarArgsFrameIndex(FrameIndex);
2956 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2957 CallingConv::ID CallConv, bool isVarArg,
2958 const SmallVectorImpl<ISD::InputArg>
2960 SDLoc dl, SelectionDAG &DAG,
2961 SmallVectorImpl<SDValue> &InVals)
2963 MachineFunction &MF = DAG.getMachineFunction();
2964 MachineFrameInfo *MFI = MF.getFrameInfo();
2966 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2968 // Assign locations to all of the incoming arguments.
2969 SmallVector<CCValAssign, 16> ArgLocs;
2970 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2971 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2972 CCInfo.AnalyzeFormalArguments(Ins,
2973 CCAssignFnForNode(CallConv, /* Return*/ false,
2976 SmallVector<SDValue, 16> ArgValues;
2977 int lastInsIndex = -1;
2979 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2980 unsigned CurArgIdx = 0;
2982 // Initially ArgRegsSaveSize is zero.
2983 // Then we increase this value each time we meet byval parameter.
2984 // We also increase this value in case of varargs function.
2985 AFI->setArgRegsSaveSize(0);
2987 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2988 CCValAssign &VA = ArgLocs[i];
2989 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2990 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2991 // Arguments stored in registers.
2992 if (VA.isRegLoc()) {
2993 EVT RegVT = VA.getLocVT();
2995 if (VA.needsCustom()) {
2996 // f64 and vector types are split up into multiple registers or
2997 // combinations of registers and stack slots.
2998 if (VA.getLocVT() == MVT::v2f64) {
2999 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3001 VA = ArgLocs[++i]; // skip ahead to next loc
3003 if (VA.isMemLoc()) {
3004 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3005 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3006 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3007 MachinePointerInfo::getFixedStack(FI),
3008 false, false, false, 0);
3010 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3013 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3014 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3015 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3016 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3017 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3019 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3022 const TargetRegisterClass *RC;
3024 if (RegVT == MVT::f32)
3025 RC = &ARM::SPRRegClass;
3026 else if (RegVT == MVT::f64)
3027 RC = &ARM::DPRRegClass;
3028 else if (RegVT == MVT::v2f64)
3029 RC = &ARM::QPRRegClass;
3030 else if (RegVT == MVT::i32)
3031 RC = AFI->isThumb1OnlyFunction() ?
3032 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3033 (const TargetRegisterClass*)&ARM::GPRRegClass;
3035 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3037 // Transform the arguments in physical registers into virtual ones.
3038 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3039 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3042 // If this is an 8 or 16-bit value, it is really passed promoted
3043 // to 32 bits. Insert an assert[sz]ext to capture this, then
3044 // truncate to the right size.
3045 switch (VA.getLocInfo()) {
3046 default: llvm_unreachable("Unknown loc info!");
3047 case CCValAssign::Full: break;
3048 case CCValAssign::BCvt:
3049 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3051 case CCValAssign::SExt:
3052 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3053 DAG.getValueType(VA.getValVT()));
3054 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3056 case CCValAssign::ZExt:
3057 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3058 DAG.getValueType(VA.getValVT()));
3059 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3063 InVals.push_back(ArgValue);
3065 } else { // VA.isRegLoc()
3068 assert(VA.isMemLoc());
3069 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3071 int index = ArgLocs[i].getValNo();
3073 // Some Ins[] entries become multiple ArgLoc[] entries.
3074 // Process them only once.
3075 if (index != lastInsIndex)
3077 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3078 // FIXME: For now, all byval parameter objects are marked mutable.
3079 // This can be changed with more analysis.
3080 // In case of tail call optimization mark all arguments mutable.
3081 // Since they could be overwritten by lowering of arguments in case of
3083 if (Flags.isByVal()) {
3084 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3085 int FrameIndex = StoreByValRegs(
3086 CCInfo, DAG, dl, Chain, CurOrigArg,
3088 Ins[VA.getValNo()].PartOffset,
3089 VA.getLocMemOffset(),
3090 Flags.getByValSize(),
3091 true /*force mutable frames*/);
3092 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3093 CCInfo.nextInRegsParam();
3095 unsigned FIOffset = VA.getLocMemOffset() +
3096 AFI->getStoredByValParamsPadding();
3097 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3100 // Create load nodes to retrieve arguments from the stack.
3101 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3102 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3103 MachinePointerInfo::getFixedStack(FI),
3104 false, false, false, 0));
3106 lastInsIndex = index;
3113 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3114 CCInfo.getNextStackOffset());
3119 /// isFloatingPointZero - Return true if this is +0.0.
3120 static bool isFloatingPointZero(SDValue Op) {
3121 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3122 return CFP->getValueAPF().isPosZero();
3123 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3124 // Maybe this has already been legalized into the constant pool?
3125 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3126 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3127 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3128 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3129 return CFP->getValueAPF().isPosZero();
3135 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3136 /// the given operands.
3138 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3139 SDValue &ARMcc, SelectionDAG &DAG,
3141 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3142 unsigned C = RHSC->getZExtValue();
3143 if (!isLegalICmpImmediate(C)) {
3144 // Constant does not fit, try adjusting it by one?
3149 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3150 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3151 RHS = DAG.getConstant(C-1, MVT::i32);
3156 if (C != 0 && isLegalICmpImmediate(C-1)) {
3157 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3158 RHS = DAG.getConstant(C-1, MVT::i32);
3163 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3164 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3165 RHS = DAG.getConstant(C+1, MVT::i32);
3170 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3171 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3172 RHS = DAG.getConstant(C+1, MVT::i32);
3179 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3180 ARMISD::NodeType CompareType;
3183 CompareType = ARMISD::CMP;
3188 CompareType = ARMISD::CMPZ;
3191 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3192 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3195 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3197 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3200 if (!isFloatingPointZero(RHS))
3201 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3203 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3204 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3207 /// duplicateCmp - Glue values can have only one use, so this function
3208 /// duplicates a comparison node.
3210 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3211 unsigned Opc = Cmp.getOpcode();
3213 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3214 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3216 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3217 Cmp = Cmp.getOperand(0);
3218 Opc = Cmp.getOpcode();
3219 if (Opc == ARMISD::CMPFP)
3220 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3222 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3223 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3225 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3228 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3229 SDValue Cond = Op.getOperand(0);
3230 SDValue SelectTrue = Op.getOperand(1);
3231 SDValue SelectFalse = Op.getOperand(2);
3236 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3237 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3239 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3240 const ConstantSDNode *CMOVTrue =
3241 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3242 const ConstantSDNode *CMOVFalse =
3243 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3245 if (CMOVTrue && CMOVFalse) {
3246 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3247 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3251 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3253 False = SelectFalse;
3254 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3259 if (True.getNode() && False.getNode()) {
3260 EVT VT = Op.getValueType();
3261 SDValue ARMcc = Cond.getOperand(2);
3262 SDValue CCR = Cond.getOperand(3);
3263 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3264 assert(True.getValueType() == VT);
3265 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3270 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3271 // undefined bits before doing a full-word comparison with zero.
3272 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3273 DAG.getConstant(1, Cond.getValueType()));
3275 return DAG.getSelectCC(dl, Cond,
3276 DAG.getConstant(0, Cond.getValueType()),
3277 SelectTrue, SelectFalse, ISD::SETNE);
3280 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3281 if (CC == ISD::SETNE)
3283 return ISD::getSetCCSwappedOperands(CC);
3286 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3287 bool &swpCmpOps, bool &swpVselOps) {
3288 // Start by selecting the GE condition code for opcodes that return true for
3290 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3292 CondCode = ARMCC::GE;
3294 // and GT for opcodes that return false for 'equality'.
3295 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3297 CondCode = ARMCC::GT;
3299 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3300 // to swap the compare operands.
3301 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3305 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3306 // If we have an unordered opcode, we need to swap the operands to the VSEL
3307 // instruction (effectively negating the condition).
3309 // This also has the effect of swapping which one of 'less' or 'greater'
3310 // returns true, so we also swap the compare operands. It also switches
3311 // whether we return true for 'equality', so we compensate by picking the
3312 // opposite condition code to our original choice.
3313 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3314 CC == ISD::SETUGT) {
3315 swpCmpOps = !swpCmpOps;
3316 swpVselOps = !swpVselOps;
3317 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3320 // 'ordered' is 'anything but unordered', so use the VS condition code and
3321 // swap the VSEL operands.
3322 if (CC == ISD::SETO) {
3323 CondCode = ARMCC::VS;
3327 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3328 // code and swap the VSEL operands.
3329 if (CC == ISD::SETUNE) {
3330 CondCode = ARMCC::EQ;
3335 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3336 EVT VT = Op.getValueType();
3337 SDValue LHS = Op.getOperand(0);
3338 SDValue RHS = Op.getOperand(1);
3339 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3340 SDValue TrueVal = Op.getOperand(2);
3341 SDValue FalseVal = Op.getOperand(3);
3344 if (LHS.getValueType() == MVT::i32) {
3345 // Try to generate VSEL on ARMv8.
3346 // The VSEL instruction can't use all the usual ARM condition
3347 // codes: it only has two bits to select the condition code, so it's
3348 // constrained to use only GE, GT, VS and EQ.
3350 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3351 // swap the operands of the previous compare instruction (effectively
3352 // inverting the compare condition, swapping 'less' and 'greater') and
3353 // sometimes need to swap the operands to the VSEL (which inverts the
3354 // condition in the sense of firing whenever the previous condition didn't)
3355 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3356 TrueVal.getValueType() == MVT::f64)) {
3357 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3358 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3359 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3360 CC = getInverseCCForVSEL(CC);
3361 std::swap(TrueVal, FalseVal);
3366 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3367 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3368 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3372 ARMCC::CondCodes CondCode, CondCode2;
3373 FPCCToARMCC(CC, CondCode, CondCode2);
3375 // Try to generate VSEL on ARMv8.
3376 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3377 TrueVal.getValueType() == MVT::f64)) {
3378 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3379 // same operands, as follows:
3380 // c = fcmp [ogt, olt, ugt, ult] a, b
3382 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3383 // handled differently than the original code sequence.
3384 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3386 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3387 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3388 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3389 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3392 bool swpCmpOps = false;
3393 bool swpVselOps = false;
3394 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3396 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3397 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3399 std::swap(LHS, RHS);
3401 std::swap(TrueVal, FalseVal);
3405 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3406 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3407 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3408 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3410 if (CondCode2 != ARMCC::AL) {
3411 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3412 // FIXME: Needs another CMP because flag can have but one use.
3413 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3414 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3415 Result, TrueVal, ARMcc2, CCR, Cmp2);
3420 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3421 /// to morph to an integer compare sequence.
3422 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3423 const ARMSubtarget *Subtarget) {
3424 SDNode *N = Op.getNode();
3425 if (!N->hasOneUse())
3426 // Otherwise it requires moving the value from fp to integer registers.
3428 if (!N->getNumValues())
3430 EVT VT = Op.getValueType();
3431 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3432 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3433 // vmrs are very slow, e.g. cortex-a8.
3436 if (isFloatingPointZero(Op)) {
3440 return ISD::isNormalLoad(N);
3443 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3444 if (isFloatingPointZero(Op))
3445 return DAG.getConstant(0, MVT::i32);
3447 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3448 return DAG.getLoad(MVT::i32, SDLoc(Op),
3449 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3450 Ld->isVolatile(), Ld->isNonTemporal(),
3451 Ld->isInvariant(), Ld->getAlignment());
3453 llvm_unreachable("Unknown VFP cmp argument!");
3456 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3457 SDValue &RetVal1, SDValue &RetVal2) {
3458 if (isFloatingPointZero(Op)) {
3459 RetVal1 = DAG.getConstant(0, MVT::i32);
3460 RetVal2 = DAG.getConstant(0, MVT::i32);
3464 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3465 SDValue Ptr = Ld->getBasePtr();
3466 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3467 Ld->getChain(), Ptr,
3468 Ld->getPointerInfo(),
3469 Ld->isVolatile(), Ld->isNonTemporal(),
3470 Ld->isInvariant(), Ld->getAlignment());
3472 EVT PtrType = Ptr.getValueType();
3473 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3474 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3475 PtrType, Ptr, DAG.getConstant(4, PtrType));
3476 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3477 Ld->getChain(), NewPtr,
3478 Ld->getPointerInfo().getWithOffset(4),
3479 Ld->isVolatile(), Ld->isNonTemporal(),
3480 Ld->isInvariant(), NewAlign);
3484 llvm_unreachable("Unknown VFP cmp argument!");
3487 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3488 /// f32 and even f64 comparisons to integer ones.
3490 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3491 SDValue Chain = Op.getOperand(0);
3492 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3493 SDValue LHS = Op.getOperand(2);
3494 SDValue RHS = Op.getOperand(3);
3495 SDValue Dest = Op.getOperand(4);
3498 bool LHSSeenZero = false;
3499 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3500 bool RHSSeenZero = false;
3501 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3502 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3503 // If unsafe fp math optimization is enabled and there are no other uses of
3504 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3505 // to an integer comparison.
3506 if (CC == ISD::SETOEQ)
3508 else if (CC == ISD::SETUNE)
3511 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3513 if (LHS.getValueType() == MVT::f32) {
3514 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3515 bitcastf32Toi32(LHS, DAG), Mask);
3516 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3517 bitcastf32Toi32(RHS, DAG), Mask);
3518 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3519 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3520 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3521 Chain, Dest, ARMcc, CCR, Cmp);
3526 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3527 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3528 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3529 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3530 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3531 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3532 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3533 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3534 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3540 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3541 SDValue Chain = Op.getOperand(0);
3542 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3543 SDValue LHS = Op.getOperand(2);
3544 SDValue RHS = Op.getOperand(3);
3545 SDValue Dest = Op.getOperand(4);
3548 if (LHS.getValueType() == MVT::i32) {
3550 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3551 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3552 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3553 Chain, Dest, ARMcc, CCR, Cmp);
3556 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3558 if (getTargetMachine().Options.UnsafeFPMath &&
3559 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3560 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3561 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3562 if (Result.getNode())
3566 ARMCC::CondCodes CondCode, CondCode2;
3567 FPCCToARMCC(CC, CondCode, CondCode2);
3569 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3570 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3571 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3572 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3573 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3574 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3575 if (CondCode2 != ARMCC::AL) {
3576 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3577 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3578 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3583 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3584 SDValue Chain = Op.getOperand(0);
3585 SDValue Table = Op.getOperand(1);
3586 SDValue Index = Op.getOperand(2);
3589 EVT PTy = getPointerTy();
3590 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3591 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3592 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3593 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3594 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3595 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3596 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3597 if (Subtarget->isThumb2()) {
3598 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3599 // which does another jump to the destination. This also makes it easier
3600 // to translate it to TBB / TBH later.
3601 // FIXME: This might not work if the function is extremely large.
3602 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3603 Addr, Op.getOperand(2), JTI, UId);
3605 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3606 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3607 MachinePointerInfo::getJumpTable(),
3608 false, false, false, 0);
3609 Chain = Addr.getValue(1);
3610 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3611 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3613 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3614 MachinePointerInfo::getJumpTable(),
3615 false, false, false, 0);
3616 Chain = Addr.getValue(1);
3617 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3621 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3622 EVT VT = Op.getValueType();
3625 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3626 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3628 return DAG.UnrollVectorOp(Op.getNode());
3631 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3632 "Invalid type for custom lowering!");
3633 if (VT != MVT::v4i16)
3634 return DAG.UnrollVectorOp(Op.getNode());
3636 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3637 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3640 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3641 EVT VT = Op.getValueType();
3643 return LowerVectorFP_TO_INT(Op, DAG);
3648 switch (Op.getOpcode()) {
3649 default: llvm_unreachable("Invalid opcode!");
3650 case ISD::FP_TO_SINT:
3651 Opc = ARMISD::FTOSI;
3653 case ISD::FP_TO_UINT:
3654 Opc = ARMISD::FTOUI;
3657 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3658 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3661 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3662 EVT VT = Op.getValueType();
3665 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3666 if (VT.getVectorElementType() == MVT::f32)
3668 return DAG.UnrollVectorOp(Op.getNode());
3671 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3672 "Invalid type for custom lowering!");
3673 if (VT != MVT::v4f32)
3674 return DAG.UnrollVectorOp(Op.getNode());
3678 switch (Op.getOpcode()) {
3679 default: llvm_unreachable("Invalid opcode!");
3680 case ISD::SINT_TO_FP:
3681 CastOpc = ISD::SIGN_EXTEND;
3682 Opc = ISD::SINT_TO_FP;
3684 case ISD::UINT_TO_FP:
3685 CastOpc = ISD::ZERO_EXTEND;
3686 Opc = ISD::UINT_TO_FP;
3690 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3691 return DAG.getNode(Opc, dl, VT, Op);
3694 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3695 EVT VT = Op.getValueType();
3697 return LowerVectorINT_TO_FP(Op, DAG);
3702 switch (Op.getOpcode()) {
3703 default: llvm_unreachable("Invalid opcode!");
3704 case ISD::SINT_TO_FP:
3705 Opc = ARMISD::SITOF;
3707 case ISD::UINT_TO_FP:
3708 Opc = ARMISD::UITOF;
3712 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3713 return DAG.getNode(Opc, dl, VT, Op);
3716 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3717 // Implement fcopysign with a fabs and a conditional fneg.
3718 SDValue Tmp0 = Op.getOperand(0);
3719 SDValue Tmp1 = Op.getOperand(1);
3721 EVT VT = Op.getValueType();
3722 EVT SrcVT = Tmp1.getValueType();
3723 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3724 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3725 bool UseNEON = !InGPR && Subtarget->hasNEON();
3728 // Use VBSL to copy the sign bit.
3729 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3730 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3731 DAG.getTargetConstant(EncodedVal, MVT::i32));
3732 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3734 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3735 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3736 DAG.getConstant(32, MVT::i32));
3737 else /*if (VT == MVT::f32)*/
3738 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3739 if (SrcVT == MVT::f32) {
3740 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3742 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3743 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3744 DAG.getConstant(32, MVT::i32));
3745 } else if (VT == MVT::f32)
3746 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3747 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3748 DAG.getConstant(32, MVT::i32));
3749 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3750 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3752 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3754 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3755 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3756 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3758 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3759 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3760 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3761 if (VT == MVT::f32) {
3762 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3763 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3764 DAG.getConstant(0, MVT::i32));
3766 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3772 // Bitcast operand 1 to i32.
3773 if (SrcVT == MVT::f64)
3774 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3775 &Tmp1, 1).getValue(1);
3776 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3778 // Or in the signbit with integer operations.
3779 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3780 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3781 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3782 if (VT == MVT::f32) {
3783 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3784 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3785 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3786 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3789 // f64: Or the high part with signbit and then combine two parts.
3790 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3792 SDValue Lo = Tmp0.getValue(0);
3793 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3794 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3795 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3798 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3799 MachineFunction &MF = DAG.getMachineFunction();
3800 MachineFrameInfo *MFI = MF.getFrameInfo();
3801 MFI->setReturnAddressIsTaken(true);
3803 EVT VT = Op.getValueType();
3805 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3807 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3808 SDValue Offset = DAG.getConstant(4, MVT::i32);
3809 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3810 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3811 MachinePointerInfo(), false, false, false, 0);
3814 // Return LR, which contains the return address. Mark it an implicit live-in.
3815 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3816 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3819 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3820 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3821 MFI->setFrameAddressIsTaken(true);
3823 EVT VT = Op.getValueType();
3824 SDLoc dl(Op); // FIXME probably not meaningful
3825 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3826 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3827 ? ARM::R7 : ARM::R11;
3828 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3830 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3831 MachinePointerInfo(),
3832 false, false, false, 0);
3836 /// ExpandBITCAST - If the target supports VFP, this function is called to
3837 /// expand a bit convert where either the source or destination type is i64 to
3838 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3839 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3840 /// vectors), since the legalizer won't know what to do with that.
3841 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3842 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3844 SDValue Op = N->getOperand(0);
3846 // This function is only supposed to be called for i64 types, either as the
3847 // source or destination of the bit convert.
3848 EVT SrcVT = Op.getValueType();
3849 EVT DstVT = N->getValueType(0);
3850 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3851 "ExpandBITCAST called for non-i64 type");
3853 // Turn i64->f64 into VMOVDRR.
3854 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3855 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3856 DAG.getConstant(0, MVT::i32));
3857 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3858 DAG.getConstant(1, MVT::i32));
3859 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3860 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3863 // Turn f64->i64 into VMOVRRD.
3864 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3865 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3866 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3867 // Merge the pieces into a single i64 value.
3868 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3874 /// getZeroVector - Returns a vector of specified type with all zero elements.
3875 /// Zero vectors are used to represent vector negation and in those cases
3876 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3877 /// not support i64 elements, so sometimes the zero vectors will need to be
3878 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3880 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3881 assert(VT.isVector() && "Expected a vector type");
3882 // The canonical modified immediate encoding of a zero vector is....0!
3883 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3884 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3885 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3886 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3889 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3890 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3891 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3892 SelectionDAG &DAG) const {
3893 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3894 EVT VT = Op.getValueType();
3895 unsigned VTBits = VT.getSizeInBits();
3897 SDValue ShOpLo = Op.getOperand(0);
3898 SDValue ShOpHi = Op.getOperand(1);
3899 SDValue ShAmt = Op.getOperand(2);
3901 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3903 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3905 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3906 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3907 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3908 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3909 DAG.getConstant(VTBits, MVT::i32));
3910 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3911 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3912 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3914 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3915 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3917 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3918 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3921 SDValue Ops[2] = { Lo, Hi };
3922 return DAG.getMergeValues(Ops, 2, dl);
3925 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3926 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3927 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3928 SelectionDAG &DAG) const {
3929 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3930 EVT VT = Op.getValueType();
3931 unsigned VTBits = VT.getSizeInBits();
3933 SDValue ShOpLo = Op.getOperand(0);
3934 SDValue ShOpHi = Op.getOperand(1);
3935 SDValue ShAmt = Op.getOperand(2);
3938 assert(Op.getOpcode() == ISD::SHL_PARTS);
3939 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3940 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3941 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3942 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3943 DAG.getConstant(VTBits, MVT::i32));
3944 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3945 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3947 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3948 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3949 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3951 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3952 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3955 SDValue Ops[2] = { Lo, Hi };
3956 return DAG.getMergeValues(Ops, 2, dl);
3959 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3960 SelectionDAG &DAG) const {
3961 // The rounding mode is in bits 23:22 of the FPSCR.
3962 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3963 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3964 // so that the shift + and get folded into a bitfield extract.
3966 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3967 DAG.getConstant(Intrinsic::arm_get_fpscr,
3969 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3970 DAG.getConstant(1U << 22, MVT::i32));
3971 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3972 DAG.getConstant(22, MVT::i32));
3973 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3974 DAG.getConstant(3, MVT::i32));
3977 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3978 const ARMSubtarget *ST) {
3979 EVT VT = N->getValueType(0);
3982 if (!ST->hasV6T2Ops())
3985 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3986 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3989 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3990 /// for each 16-bit element from operand, repeated. The basic idea is to
3991 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3993 /// Trace for v4i16:
3994 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3995 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3996 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3997 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3998 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3999 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4000 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4001 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4002 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4003 EVT VT = N->getValueType(0);
4006 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4007 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4008 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4009 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4010 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4011 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4014 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4015 /// bit-count for each 16-bit element from the operand. We need slightly
4016 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4017 /// 64/128-bit registers.
4019 /// Trace for v4i16:
4020 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4021 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4022 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4023 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4024 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4025 EVT VT = N->getValueType(0);
4028 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4029 if (VT.is64BitVector()) {
4030 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4031 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4032 DAG.getIntPtrConstant(0));
4034 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4035 BitCounts, DAG.getIntPtrConstant(0));
4036 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4040 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4041 /// bit-count for each 32-bit element from the operand. The idea here is
4042 /// to split the vector into 16-bit elements, leverage the 16-bit count
4043 /// routine, and then combine the results.
4045 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4046 /// input = [v0 v1 ] (vi: 32-bit elements)
4047 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4048 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4049 /// vrev: N0 = [k1 k0 k3 k2 ]
4051 /// N1 =+[k1 k0 k3 k2 ]
4053 /// N2 =+[k1 k3 k0 k2 ]
4055 /// Extended =+[k1 k3 k0 k2 ]
4057 /// Extracted=+[k1 k3 ]
4059 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4060 EVT VT = N->getValueType(0);
4063 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4065 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4066 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4067 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4068 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4069 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4071 if (VT.is64BitVector()) {
4072 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4073 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4074 DAG.getIntPtrConstant(0));
4076 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4077 DAG.getIntPtrConstant(0));
4078 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4082 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4083 const ARMSubtarget *ST) {
4084 EVT VT = N->getValueType(0);
4086 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4087 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4088 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4089 "Unexpected type for custom ctpop lowering");
4091 if (VT.getVectorElementType() == MVT::i32)
4092 return lowerCTPOP32BitElements(N, DAG);
4094 return lowerCTPOP16BitElements(N, DAG);
4097 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4098 const ARMSubtarget *ST) {
4099 EVT VT = N->getValueType(0);
4105 // Lower vector shifts on NEON to use VSHL.
4106 assert(ST->hasNEON() && "unexpected vector shift");
4108 // Left shifts translate directly to the vshiftu intrinsic.
4109 if (N->getOpcode() == ISD::SHL)
4110 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4111 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4112 N->getOperand(0), N->getOperand(1));
4114 assert((N->getOpcode() == ISD::SRA ||
4115 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4117 // NEON uses the same intrinsics for both left and right shifts. For
4118 // right shifts, the shift amounts are negative, so negate the vector of
4120 EVT ShiftVT = N->getOperand(1).getValueType();
4121 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4122 getZeroVector(ShiftVT, DAG, dl),
4124 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4125 Intrinsic::arm_neon_vshifts :
4126 Intrinsic::arm_neon_vshiftu);
4127 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4128 DAG.getConstant(vshiftInt, MVT::i32),
4129 N->getOperand(0), NegatedCount);
4132 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4133 const ARMSubtarget *ST) {
4134 EVT VT = N->getValueType(0);
4137 // We can get here for a node like i32 = ISD::SHL i32, i64
4141 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4142 "Unknown shift to lower!");
4144 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4145 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4146 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4149 // If we are in thumb mode, we don't have RRX.
4150 if (ST->isThumb1Only()) return SDValue();
4152 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4153 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4154 DAG.getConstant(0, MVT::i32));
4155 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4156 DAG.getConstant(1, MVT::i32));
4158 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4159 // captures the result into a carry flag.
4160 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4161 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4163 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4164 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4166 // Merge the pieces into a single i64 value.
4167 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4170 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4171 SDValue TmpOp0, TmpOp1;
4172 bool Invert = false;
4176 SDValue Op0 = Op.getOperand(0);
4177 SDValue Op1 = Op.getOperand(1);
4178 SDValue CC = Op.getOperand(2);
4179 EVT VT = Op.getValueType();
4180 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4183 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4184 switch (SetCCOpcode) {
4185 default: llvm_unreachable("Illegal FP comparison");
4187 case ISD::SETNE: Invert = true; // Fallthrough
4189 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4191 case ISD::SETLT: Swap = true; // Fallthrough
4193 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4195 case ISD::SETLE: Swap = true; // Fallthrough
4197 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4198 case ISD::SETUGE: Swap = true; // Fallthrough
4199 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4200 case ISD::SETUGT: Swap = true; // Fallthrough
4201 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4202 case ISD::SETUEQ: Invert = true; // Fallthrough
4204 // Expand this to (OLT | OGT).
4208 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4209 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4211 case ISD::SETUO: Invert = true; // Fallthrough
4213 // Expand this to (OLT | OGE).
4217 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4218 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4222 // Integer comparisons.
4223 switch (SetCCOpcode) {
4224 default: llvm_unreachable("Illegal integer comparison");
4225 case ISD::SETNE: Invert = true;
4226 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4227 case ISD::SETLT: Swap = true;
4228 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4229 case ISD::SETLE: Swap = true;
4230 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4231 case ISD::SETULT: Swap = true;
4232 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4233 case ISD::SETULE: Swap = true;
4234 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4237 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4238 if (Opc == ARMISD::VCEQ) {
4241 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4243 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4246 // Ignore bitconvert.
4247 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4248 AndOp = AndOp.getOperand(0);
4250 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4252 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4253 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4260 std::swap(Op0, Op1);
4262 // If one of the operands is a constant vector zero, attempt to fold the
4263 // comparison to a specialized compare-against-zero form.
4265 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4267 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4268 if (Opc == ARMISD::VCGE)
4269 Opc = ARMISD::VCLEZ;
4270 else if (Opc == ARMISD::VCGT)
4271 Opc = ARMISD::VCLTZ;
4276 if (SingleOp.getNode()) {
4279 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4281 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4283 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4285 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4287 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4289 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4292 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4296 Result = DAG.getNOT(dl, Result, VT);
4301 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4302 /// valid vector constant for a NEON instruction with a "modified immediate"
4303 /// operand (e.g., VMOV). If so, return the encoded value.
4304 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4305 unsigned SplatBitSize, SelectionDAG &DAG,
4306 EVT &VT, bool is128Bits, NEONModImmType type) {
4307 unsigned OpCmode, Imm;
4309 // SplatBitSize is set to the smallest size that splats the vector, so a
4310 // zero vector will always have SplatBitSize == 8. However, NEON modified
4311 // immediate instructions others than VMOV do not support the 8-bit encoding
4312 // of a zero vector, and the default encoding of zero is supposed to be the
4317 switch (SplatBitSize) {
4319 if (type != VMOVModImm)
4321 // Any 1-byte value is OK. Op=0, Cmode=1110.
4322 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4325 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4329 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4330 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4331 if ((SplatBits & ~0xff) == 0) {
4332 // Value = 0x00nn: Op=x, Cmode=100x.
4337 if ((SplatBits & ~0xff00) == 0) {
4338 // Value = 0xnn00: Op=x, Cmode=101x.
4340 Imm = SplatBits >> 8;
4346 // NEON's 32-bit VMOV supports splat values where:
4347 // * only one byte is nonzero, or
4348 // * the least significant byte is 0xff and the second byte is nonzero, or
4349 // * the least significant 2 bytes are 0xff and the third is nonzero.
4350 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4351 if ((SplatBits & ~0xff) == 0) {
4352 // Value = 0x000000nn: Op=x, Cmode=000x.
4357 if ((SplatBits & ~0xff00) == 0) {
4358 // Value = 0x0000nn00: Op=x, Cmode=001x.
4360 Imm = SplatBits >> 8;
4363 if ((SplatBits & ~0xff0000) == 0) {
4364 // Value = 0x00nn0000: Op=x, Cmode=010x.
4366 Imm = SplatBits >> 16;
4369 if ((SplatBits & ~0xff000000) == 0) {
4370 // Value = 0xnn000000: Op=x, Cmode=011x.
4372 Imm = SplatBits >> 24;
4376 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4377 if (type == OtherModImm) return SDValue();
4379 if ((SplatBits & ~0xffff) == 0 &&
4380 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4381 // Value = 0x0000nnff: Op=x, Cmode=1100.
4383 Imm = SplatBits >> 8;
4388 if ((SplatBits & ~0xffffff) == 0 &&
4389 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4390 // Value = 0x00nnffff: Op=x, Cmode=1101.
4392 Imm = SplatBits >> 16;
4393 SplatBits |= 0xffff;
4397 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4398 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4399 // VMOV.I32. A (very) minor optimization would be to replicate the value
4400 // and fall through here to test for a valid 64-bit splat. But, then the
4401 // caller would also need to check and handle the change in size.
4405 if (type != VMOVModImm)
4407 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4408 uint64_t BitMask = 0xff;
4410 unsigned ImmMask = 1;
4412 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4413 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4416 } else if ((SplatBits & BitMask) != 0) {
4422 // Op=1, Cmode=1110.
4425 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4430 llvm_unreachable("unexpected size for isNEONModifiedImm");
4433 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4434 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4437 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4438 const ARMSubtarget *ST) const {
4442 bool IsDouble = Op.getValueType() == MVT::f64;
4443 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4445 // Try splatting with a VMOV.f32...
4446 APFloat FPVal = CFP->getValueAPF();
4447 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4450 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4451 // We have code in place to select a valid ConstantFP already, no need to
4456 // It's a float and we are trying to use NEON operations where
4457 // possible. Lower it to a splat followed by an extract.
4459 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4460 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4462 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4463 DAG.getConstant(0, MVT::i32));
4466 // The rest of our options are NEON only, make sure that's allowed before
4468 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4472 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4474 // It wouldn't really be worth bothering for doubles except for one very
4475 // important value, which does happen to match: 0.0. So make sure we don't do
4477 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4480 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4481 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4483 if (NewVal != SDValue()) {
4485 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4488 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4490 // It's a float: cast and extract a vector element.
4491 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4493 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4494 DAG.getConstant(0, MVT::i32));
4497 // Finally, try a VMVN.i32
4498 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4500 if (NewVal != SDValue()) {
4502 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4505 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4507 // It's a float: cast and extract a vector element.
4508 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4510 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4511 DAG.getConstant(0, MVT::i32));
4517 // check if an VEXT instruction can handle the shuffle mask when the
4518 // vector sources of the shuffle are the same.
4519 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4520 unsigned NumElts = VT.getVectorNumElements();
4522 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4528 // If this is a VEXT shuffle, the immediate value is the index of the first
4529 // element. The other shuffle indices must be the successive elements after
4531 unsigned ExpectedElt = Imm;
4532 for (unsigned i = 1; i < NumElts; ++i) {
4533 // Increment the expected index. If it wraps around, just follow it
4534 // back to index zero and keep going.
4536 if (ExpectedElt == NumElts)
4539 if (M[i] < 0) continue; // ignore UNDEF indices
4540 if (ExpectedElt != static_cast<unsigned>(M[i]))
4548 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4549 bool &ReverseVEXT, unsigned &Imm) {
4550 unsigned NumElts = VT.getVectorNumElements();
4551 ReverseVEXT = false;
4553 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4559 // If this is a VEXT shuffle, the immediate value is the index of the first
4560 // element. The other shuffle indices must be the successive elements after
4562 unsigned ExpectedElt = Imm;
4563 for (unsigned i = 1; i < NumElts; ++i) {
4564 // Increment the expected index. If it wraps around, it may still be
4565 // a VEXT but the source vectors must be swapped.
4567 if (ExpectedElt == NumElts * 2) {
4572 if (M[i] < 0) continue; // ignore UNDEF indices
4573 if (ExpectedElt != static_cast<unsigned>(M[i]))
4577 // Adjust the index value if the source operands will be swapped.
4584 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4585 /// instruction with the specified blocksize. (The order of the elements
4586 /// within each block of the vector is reversed.)
4587 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4588 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4589 "Only possible block sizes for VREV are: 16, 32, 64");
4591 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4595 unsigned NumElts = VT.getVectorNumElements();
4596 unsigned BlockElts = M[0] + 1;
4597 // If the first shuffle index is UNDEF, be optimistic.
4599 BlockElts = BlockSize / EltSz;
4601 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4604 for (unsigned i = 0; i < NumElts; ++i) {
4605 if (M[i] < 0) continue; // ignore UNDEF indices
4606 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4613 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4614 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4615 // range, then 0 is placed into the resulting vector. So pretty much any mask
4616 // of 8 elements can work here.
4617 return VT == MVT::v8i8 && M.size() == 8;
4620 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4621 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4625 unsigned NumElts = VT.getVectorNumElements();
4626 WhichResult = (M[0] == 0 ? 0 : 1);
4627 for (unsigned i = 0; i < NumElts; i += 2) {
4628 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4629 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4635 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4636 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4637 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4638 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4639 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4643 unsigned NumElts = VT.getVectorNumElements();
4644 WhichResult = (M[0] == 0 ? 0 : 1);
4645 for (unsigned i = 0; i < NumElts; i += 2) {
4646 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4647 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4653 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4654 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4658 unsigned NumElts = VT.getVectorNumElements();
4659 WhichResult = (M[0] == 0 ? 0 : 1);
4660 for (unsigned i = 0; i != NumElts; ++i) {
4661 if (M[i] < 0) continue; // ignore UNDEF indices
4662 if ((unsigned) M[i] != 2 * i + WhichResult)
4666 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4667 if (VT.is64BitVector() && EltSz == 32)
4673 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4674 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4675 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4676 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4677 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4681 unsigned Half = VT.getVectorNumElements() / 2;
4682 WhichResult = (M[0] == 0 ? 0 : 1);
4683 for (unsigned j = 0; j != 2; ++j) {
4684 unsigned Idx = WhichResult;
4685 for (unsigned i = 0; i != Half; ++i) {
4686 int MIdx = M[i + j * Half];
4687 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4693 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4694 if (VT.is64BitVector() && EltSz == 32)
4700 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4701 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4705 unsigned NumElts = VT.getVectorNumElements();
4706 WhichResult = (M[0] == 0 ? 0 : 1);
4707 unsigned Idx = WhichResult * NumElts / 2;
4708 for (unsigned i = 0; i != NumElts; i += 2) {
4709 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4710 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4715 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4716 if (VT.is64BitVector() && EltSz == 32)
4722 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4723 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4724 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4725 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4726 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4730 unsigned NumElts = VT.getVectorNumElements();
4731 WhichResult = (M[0] == 0 ? 0 : 1);
4732 unsigned Idx = WhichResult * NumElts / 2;
4733 for (unsigned i = 0; i != NumElts; i += 2) {
4734 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4735 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4740 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4741 if (VT.is64BitVector() && EltSz == 32)
4747 /// \return true if this is a reverse operation on an vector.
4748 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4749 unsigned NumElts = VT.getVectorNumElements();
4750 // Make sure the mask has the right size.
4751 if (NumElts != M.size())
4754 // Look for <15, ..., 3, -1, 1, 0>.
4755 for (unsigned i = 0; i != NumElts; ++i)
4756 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4762 // If N is an integer constant that can be moved into a register in one
4763 // instruction, return an SDValue of such a constant (will become a MOV
4764 // instruction). Otherwise return null.
4765 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4766 const ARMSubtarget *ST, SDLoc dl) {
4768 if (!isa<ConstantSDNode>(N))
4770 Val = cast<ConstantSDNode>(N)->getZExtValue();
4772 if (ST->isThumb1Only()) {
4773 if (Val <= 255 || ~Val <= 255)
4774 return DAG.getConstant(Val, MVT::i32);
4776 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4777 return DAG.getConstant(Val, MVT::i32);
4782 // If this is a case we can't handle, return null and let the default
4783 // expansion code take care of it.
4784 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4785 const ARMSubtarget *ST) const {
4786 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4788 EVT VT = Op.getValueType();
4790 APInt SplatBits, SplatUndef;
4791 unsigned SplatBitSize;
4793 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4794 if (SplatBitSize <= 64) {
4795 // Check if an immediate VMOV works.
4797 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4798 SplatUndef.getZExtValue(), SplatBitSize,
4799 DAG, VmovVT, VT.is128BitVector(),
4801 if (Val.getNode()) {
4802 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4803 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4806 // Try an immediate VMVN.
4807 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4808 Val = isNEONModifiedImm(NegatedImm,
4809 SplatUndef.getZExtValue(), SplatBitSize,
4810 DAG, VmovVT, VT.is128BitVector(),
4812 if (Val.getNode()) {
4813 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4814 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4817 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4818 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4819 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4821 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4822 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4828 // Scan through the operands to see if only one value is used.
4830 // As an optimisation, even if more than one value is used it may be more
4831 // profitable to splat with one value then change some lanes.
4833 // Heuristically we decide to do this if the vector has a "dominant" value,
4834 // defined as splatted to more than half of the lanes.
4835 unsigned NumElts = VT.getVectorNumElements();
4836 bool isOnlyLowElement = true;
4837 bool usesOnlyOneValue = true;
4838 bool hasDominantValue = false;
4839 bool isConstant = true;
4841 // Map of the number of times a particular SDValue appears in the
4843 DenseMap<SDValue, unsigned> ValueCounts;
4845 for (unsigned i = 0; i < NumElts; ++i) {
4846 SDValue V = Op.getOperand(i);
4847 if (V.getOpcode() == ISD::UNDEF)
4850 isOnlyLowElement = false;
4851 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4854 ValueCounts.insert(std::make_pair(V, 0));
4855 unsigned &Count = ValueCounts[V];
4857 // Is this value dominant? (takes up more than half of the lanes)
4858 if (++Count > (NumElts / 2)) {
4859 hasDominantValue = true;
4863 if (ValueCounts.size() != 1)
4864 usesOnlyOneValue = false;
4865 if (!Value.getNode() && ValueCounts.size() > 0)
4866 Value = ValueCounts.begin()->first;
4868 if (ValueCounts.size() == 0)
4869 return DAG.getUNDEF(VT);
4871 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4872 // Keep going if we are hitting this case.
4873 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4874 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4876 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4878 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4879 // i32 and try again.
4880 if (hasDominantValue && EltSize <= 32) {
4884 // If we are VDUPing a value that comes directly from a vector, that will
4885 // cause an unnecessary move to and from a GPR, where instead we could
4886 // just use VDUPLANE. We can only do this if the lane being extracted
4887 // is at a constant index, as the VDUP from lane instructions only have
4888 // constant-index forms.
4889 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4890 isa<ConstantSDNode>(Value->getOperand(1))) {
4891 // We need to create a new undef vector to use for the VDUPLANE if the
4892 // size of the vector from which we get the value is different than the
4893 // size of the vector that we need to create. We will insert the element
4894 // such that the register coalescer will remove unnecessary copies.
4895 if (VT != Value->getOperand(0).getValueType()) {
4896 ConstantSDNode *constIndex;
4897 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4898 assert(constIndex && "The index is not a constant!");
4899 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4900 VT.getVectorNumElements();
4901 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4902 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4903 Value, DAG.getConstant(index, MVT::i32)),
4904 DAG.getConstant(index, MVT::i32));
4906 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4907 Value->getOperand(0), Value->getOperand(1));
4909 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4911 if (!usesOnlyOneValue) {
4912 // The dominant value was splatted as 'N', but we now have to insert
4913 // all differing elements.
4914 for (unsigned I = 0; I < NumElts; ++I) {
4915 if (Op.getOperand(I) == Value)
4917 SmallVector<SDValue, 3> Ops;
4919 Ops.push_back(Op.getOperand(I));
4920 Ops.push_back(DAG.getConstant(I, MVT::i32));
4921 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4926 if (VT.getVectorElementType().isFloatingPoint()) {
4927 SmallVector<SDValue, 8> Ops;
4928 for (unsigned i = 0; i < NumElts; ++i)
4929 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4931 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4932 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4933 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4935 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4937 if (usesOnlyOneValue) {
4938 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4939 if (isConstant && Val.getNode())
4940 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4944 // If all elements are constants and the case above didn't get hit, fall back
4945 // to the default expansion, which will generate a load from the constant
4950 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4952 SDValue shuffle = ReconstructShuffle(Op, DAG);
4953 if (shuffle != SDValue())
4957 // Vectors with 32- or 64-bit elements can be built by directly assigning
4958 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4959 // will be legalized.
4960 if (EltSize >= 32) {
4961 // Do the expansion with floating-point types, since that is what the VFP
4962 // registers are defined to use, and since i64 is not legal.
4963 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4964 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4965 SmallVector<SDValue, 8> Ops;
4966 for (unsigned i = 0; i < NumElts; ++i)
4967 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4968 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4969 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4972 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4973 // know the default expansion would otherwise fall back on something even
4974 // worse. For a vector with one or two non-undef values, that's
4975 // scalar_to_vector for the elements followed by a shuffle (provided the
4976 // shuffle is valid for the target) and materialization element by element
4977 // on the stack followed by a load for everything else.
4978 if (!isConstant && !usesOnlyOneValue) {
4979 SDValue Vec = DAG.getUNDEF(VT);
4980 for (unsigned i = 0 ; i < NumElts; ++i) {
4981 SDValue V = Op.getOperand(i);
4982 if (V.getOpcode() == ISD::UNDEF)
4984 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4985 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4993 // Gather data to see if the operation can be modelled as a
4994 // shuffle in combination with VEXTs.
4995 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4996 SelectionDAG &DAG) const {
4998 EVT VT = Op.getValueType();
4999 unsigned NumElts = VT.getVectorNumElements();
5001 SmallVector<SDValue, 2> SourceVecs;
5002 SmallVector<unsigned, 2> MinElts;
5003 SmallVector<unsigned, 2> MaxElts;
5005 for (unsigned i = 0; i < NumElts; ++i) {
5006 SDValue V = Op.getOperand(i);
5007 if (V.getOpcode() == ISD::UNDEF)
5009 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5010 // A shuffle can only come from building a vector from various
5011 // elements of other vectors.
5013 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5014 VT.getVectorElementType()) {
5015 // This code doesn't know how to handle shuffles where the vector
5016 // element types do not match (this happens because type legalization
5017 // promotes the return type of EXTRACT_VECTOR_ELT).
5018 // FIXME: It might be appropriate to extend this code to handle
5019 // mismatched types.
5023 // Record this extraction against the appropriate vector if possible...
5024 SDValue SourceVec = V.getOperand(0);
5025 // If the element number isn't a constant, we can't effectively
5026 // analyze what's going on.
5027 if (!isa<ConstantSDNode>(V.getOperand(1)))
5029 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5030 bool FoundSource = false;
5031 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5032 if (SourceVecs[j] == SourceVec) {
5033 if (MinElts[j] > EltNo)
5035 if (MaxElts[j] < EltNo)
5042 // Or record a new source if not...
5044 SourceVecs.push_back(SourceVec);
5045 MinElts.push_back(EltNo);
5046 MaxElts.push_back(EltNo);
5050 // Currently only do something sane when at most two source vectors
5052 if (SourceVecs.size() > 2)
5055 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5056 int VEXTOffsets[2] = {0, 0};
5058 // This loop extracts the usage patterns of the source vectors
5059 // and prepares appropriate SDValues for a shuffle if possible.
5060 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5061 if (SourceVecs[i].getValueType() == VT) {
5062 // No VEXT necessary
5063 ShuffleSrcs[i] = SourceVecs[i];
5066 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5067 // It probably isn't worth padding out a smaller vector just to
5068 // break it down again in a shuffle.
5072 // Since only 64-bit and 128-bit vectors are legal on ARM and
5073 // we've eliminated the other cases...
5074 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5075 "unexpected vector sizes in ReconstructShuffle");
5077 if (MaxElts[i] - MinElts[i] >= NumElts) {
5078 // Span too large for a VEXT to cope
5082 if (MinElts[i] >= NumElts) {
5083 // The extraction can just take the second half
5084 VEXTOffsets[i] = NumElts;
5085 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5087 DAG.getIntPtrConstant(NumElts));
5088 } else if (MaxElts[i] < NumElts) {
5089 // The extraction can just take the first half
5091 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5093 DAG.getIntPtrConstant(0));
5095 // An actual VEXT is needed
5096 VEXTOffsets[i] = MinElts[i];
5097 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5099 DAG.getIntPtrConstant(0));
5100 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5102 DAG.getIntPtrConstant(NumElts));
5103 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5104 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5108 SmallVector<int, 8> Mask;
5110 for (unsigned i = 0; i < NumElts; ++i) {
5111 SDValue Entry = Op.getOperand(i);
5112 if (Entry.getOpcode() == ISD::UNDEF) {
5117 SDValue ExtractVec = Entry.getOperand(0);
5118 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5119 .getOperand(1))->getSExtValue();
5120 if (ExtractVec == SourceVecs[0]) {
5121 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5123 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5127 // Final check before we try to produce nonsense...
5128 if (isShuffleMaskLegal(Mask, VT))
5129 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5135 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5136 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5137 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5138 /// are assumed to be legal.
5140 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5142 if (VT.getVectorNumElements() == 4 &&
5143 (VT.is128BitVector() || VT.is64BitVector())) {
5144 unsigned PFIndexes[4];
5145 for (unsigned i = 0; i != 4; ++i) {
5149 PFIndexes[i] = M[i];
5152 // Compute the index in the perfect shuffle table.
5153 unsigned PFTableIndex =
5154 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5155 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5156 unsigned Cost = (PFEntry >> 30);
5163 unsigned Imm, WhichResult;
5165 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5166 return (EltSize >= 32 ||
5167 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5168 isVREVMask(M, VT, 64) ||
5169 isVREVMask(M, VT, 32) ||
5170 isVREVMask(M, VT, 16) ||
5171 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5172 isVTBLMask(M, VT) ||
5173 isVTRNMask(M, VT, WhichResult) ||
5174 isVUZPMask(M, VT, WhichResult) ||
5175 isVZIPMask(M, VT, WhichResult) ||
5176 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5177 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5178 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5179 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5182 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5183 /// the specified operations to build the shuffle.
5184 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5185 SDValue RHS, SelectionDAG &DAG,
5187 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5188 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5189 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5192 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5201 OP_VUZPL, // VUZP, left result
5202 OP_VUZPR, // VUZP, right result
5203 OP_VZIPL, // VZIP, left result
5204 OP_VZIPR, // VZIP, right result
5205 OP_VTRNL, // VTRN, left result
5206 OP_VTRNR // VTRN, right result
5209 if (OpNum == OP_COPY) {
5210 if (LHSID == (1*9+2)*9+3) return LHS;
5211 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5215 SDValue OpLHS, OpRHS;
5216 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5217 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5218 EVT VT = OpLHS.getValueType();
5221 default: llvm_unreachable("Unknown shuffle opcode!");
5223 // VREV divides the vector in half and swaps within the half.
5224 if (VT.getVectorElementType() == MVT::i32 ||
5225 VT.getVectorElementType() == MVT::f32)
5226 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5227 // vrev <4 x i16> -> VREV32
5228 if (VT.getVectorElementType() == MVT::i16)
5229 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5230 // vrev <4 x i8> -> VREV16
5231 assert(VT.getVectorElementType() == MVT::i8);
5232 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5237 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5238 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5242 return DAG.getNode(ARMISD::VEXT, dl, VT,
5244 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5247 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5248 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5251 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5252 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5255 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5256 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5260 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5261 ArrayRef<int> ShuffleMask,
5262 SelectionDAG &DAG) {
5263 // Check to see if we can use the VTBL instruction.
5264 SDValue V1 = Op.getOperand(0);
5265 SDValue V2 = Op.getOperand(1);
5268 SmallVector<SDValue, 8> VTBLMask;
5269 for (ArrayRef<int>::iterator
5270 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5271 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5273 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5274 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5275 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5278 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5279 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5283 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5284 SelectionDAG &DAG) {
5286 SDValue OpLHS = Op.getOperand(0);
5287 EVT VT = OpLHS.getValueType();
5289 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5290 "Expect an v8i16/v16i8 type");
5291 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5292 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5293 // extract the first 8 bytes into the top double word and the last 8 bytes
5294 // into the bottom double word. The v8i16 case is similar.
5295 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5296 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5297 DAG.getConstant(ExtractNum, MVT::i32));
5300 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5301 SDValue V1 = Op.getOperand(0);
5302 SDValue V2 = Op.getOperand(1);
5304 EVT VT = Op.getValueType();
5305 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5307 // Convert shuffles that are directly supported on NEON to target-specific
5308 // DAG nodes, instead of keeping them as shuffles and matching them again
5309 // during code selection. This is more efficient and avoids the possibility
5310 // of inconsistencies between legalization and selection.
5311 // FIXME: floating-point vectors should be canonicalized to integer vectors
5312 // of the same time so that they get CSEd properly.
5313 ArrayRef<int> ShuffleMask = SVN->getMask();
5315 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5316 if (EltSize <= 32) {
5317 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5318 int Lane = SVN->getSplatIndex();
5319 // If this is undef splat, generate it via "just" vdup, if possible.
5320 if (Lane == -1) Lane = 0;
5322 // Test if V1 is a SCALAR_TO_VECTOR.
5323 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5324 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5326 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5327 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5329 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5330 !isa<ConstantSDNode>(V1.getOperand(0))) {
5331 bool IsScalarToVector = true;
5332 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5333 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5334 IsScalarToVector = false;
5337 if (IsScalarToVector)
5338 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5340 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5341 DAG.getConstant(Lane, MVT::i32));
5346 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5349 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5350 DAG.getConstant(Imm, MVT::i32));
5353 if (isVREVMask(ShuffleMask, VT, 64))
5354 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5355 if (isVREVMask(ShuffleMask, VT, 32))
5356 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5357 if (isVREVMask(ShuffleMask, VT, 16))
5358 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5360 if (V2->getOpcode() == ISD::UNDEF &&
5361 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5362 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5363 DAG.getConstant(Imm, MVT::i32));
5366 // Check for Neon shuffles that modify both input vectors in place.
5367 // If both results are used, i.e., if there are two shuffles with the same
5368 // source operands and with masks corresponding to both results of one of
5369 // these operations, DAG memoization will ensure that a single node is
5370 // used for both shuffles.
5371 unsigned WhichResult;
5372 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5373 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5374 V1, V2).getValue(WhichResult);
5375 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5376 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5377 V1, V2).getValue(WhichResult);
5378 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5379 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5380 V1, V2).getValue(WhichResult);
5382 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5383 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5384 V1, V1).getValue(WhichResult);
5385 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5386 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5387 V1, V1).getValue(WhichResult);
5388 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5389 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5390 V1, V1).getValue(WhichResult);
5393 // If the shuffle is not directly supported and it has 4 elements, use
5394 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5395 unsigned NumElts = VT.getVectorNumElements();
5397 unsigned PFIndexes[4];
5398 for (unsigned i = 0; i != 4; ++i) {
5399 if (ShuffleMask[i] < 0)
5402 PFIndexes[i] = ShuffleMask[i];
5405 // Compute the index in the perfect shuffle table.
5406 unsigned PFTableIndex =
5407 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5408 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5409 unsigned Cost = (PFEntry >> 30);
5412 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5415 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5416 if (EltSize >= 32) {
5417 // Do the expansion with floating-point types, since that is what the VFP
5418 // registers are defined to use, and since i64 is not legal.
5419 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5420 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5421 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5422 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5423 SmallVector<SDValue, 8> Ops;
5424 for (unsigned i = 0; i < NumElts; ++i) {
5425 if (ShuffleMask[i] < 0)
5426 Ops.push_back(DAG.getUNDEF(EltVT));
5428 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5429 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5430 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5433 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5434 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5437 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5438 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5440 if (VT == MVT::v8i8) {
5441 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5442 if (NewOp.getNode())
5449 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5450 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5451 SDValue Lane = Op.getOperand(2);
5452 if (!isa<ConstantSDNode>(Lane))
5458 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5459 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5460 SDValue Lane = Op.getOperand(1);
5461 if (!isa<ConstantSDNode>(Lane))
5464 SDValue Vec = Op.getOperand(0);
5465 if (Op.getValueType() == MVT::i32 &&
5466 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5468 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5474 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5475 // The only time a CONCAT_VECTORS operation can have legal types is when
5476 // two 64-bit vectors are concatenated to a 128-bit vector.
5477 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5478 "unexpected CONCAT_VECTORS");
5480 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5481 SDValue Op0 = Op.getOperand(0);
5482 SDValue Op1 = Op.getOperand(1);
5483 if (Op0.getOpcode() != ISD::UNDEF)
5484 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5485 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5486 DAG.getIntPtrConstant(0));
5487 if (Op1.getOpcode() != ISD::UNDEF)
5488 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5489 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5490 DAG.getIntPtrConstant(1));
5491 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5494 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5495 /// element has been zero/sign-extended, depending on the isSigned parameter,
5496 /// from an integer type half its size.
5497 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5499 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5500 EVT VT = N->getValueType(0);
5501 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5502 SDNode *BVN = N->getOperand(0).getNode();
5503 if (BVN->getValueType(0) != MVT::v4i32 ||
5504 BVN->getOpcode() != ISD::BUILD_VECTOR)
5506 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5507 unsigned HiElt = 1 - LoElt;
5508 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5509 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5510 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5511 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5512 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5515 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5516 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5519 if (Hi0->isNullValue() && Hi1->isNullValue())
5525 if (N->getOpcode() != ISD::BUILD_VECTOR)
5528 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5529 SDNode *Elt = N->getOperand(i).getNode();
5530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5531 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5532 unsigned HalfSize = EltSize / 2;
5534 if (!isIntN(HalfSize, C->getSExtValue()))
5537 if (!isUIntN(HalfSize, C->getZExtValue()))
5548 /// isSignExtended - Check if a node is a vector value that is sign-extended
5549 /// or a constant BUILD_VECTOR with sign-extended elements.
5550 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5551 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5553 if (isExtendedBUILD_VECTOR(N, DAG, true))
5558 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5559 /// or a constant BUILD_VECTOR with zero-extended elements.
5560 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5561 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5563 if (isExtendedBUILD_VECTOR(N, DAG, false))
5568 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5569 if (OrigVT.getSizeInBits() >= 64)
5572 assert(OrigVT.isSimple() && "Expecting a simple value type");
5574 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5575 switch (OrigSimpleTy) {
5576 default: llvm_unreachable("Unexpected Vector Type");
5585 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5586 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5587 /// We insert the required extension here to get the vector to fill a D register.
5588 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5591 unsigned ExtOpcode) {
5592 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5593 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5594 // 64-bits we need to insert a new extension so that it will be 64-bits.
5595 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5596 if (OrigTy.getSizeInBits() >= 64)
5599 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5600 EVT NewVT = getExtensionTo64Bits(OrigTy);
5602 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5605 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5606 /// does not do any sign/zero extension. If the original vector is less
5607 /// than 64 bits, an appropriate extension will be added after the load to
5608 /// reach a total size of 64 bits. We have to add the extension separately
5609 /// because ARM does not have a sign/zero extending load for vectors.
5610 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5611 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5613 // The load already has the right type.
5614 if (ExtendedTy == LD->getMemoryVT())
5615 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5616 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5617 LD->isNonTemporal(), LD->isInvariant(),
5618 LD->getAlignment());
5620 // We need to create a zextload/sextload. We cannot just create a load
5621 // followed by a zext/zext node because LowerMUL is also run during normal
5622 // operation legalization where we can't create illegal types.
5623 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5624 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5625 LD->getMemoryVT(), LD->isVolatile(),
5626 LD->isNonTemporal(), LD->getAlignment());
5629 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5630 /// extending load, or BUILD_VECTOR with extended elements, return the
5631 /// unextended value. The unextended vector should be 64 bits so that it can
5632 /// be used as an operand to a VMULL instruction. If the original vector size
5633 /// before extension is less than 64 bits we add a an extension to resize
5634 /// the vector to 64 bits.
5635 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5636 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5637 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5638 N->getOperand(0)->getValueType(0),
5642 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5643 return SkipLoadExtensionForVMULL(LD, DAG);
5645 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5646 // have been legalized as a BITCAST from v4i32.
5647 if (N->getOpcode() == ISD::BITCAST) {
5648 SDNode *BVN = N->getOperand(0).getNode();
5649 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5650 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5651 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5652 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5653 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5655 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5656 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5657 EVT VT = N->getValueType(0);
5658 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5659 unsigned NumElts = VT.getVectorNumElements();
5660 MVT TruncVT = MVT::getIntegerVT(EltSize);
5661 SmallVector<SDValue, 8> Ops;
5662 for (unsigned i = 0; i != NumElts; ++i) {
5663 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5664 const APInt &CInt = C->getAPIntValue();
5665 // Element types smaller than 32 bits are not legal, so use i32 elements.
5666 // The values are implicitly truncated so sext vs. zext doesn't matter.
5667 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5669 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5670 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5673 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5674 unsigned Opcode = N->getOpcode();
5675 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5676 SDNode *N0 = N->getOperand(0).getNode();
5677 SDNode *N1 = N->getOperand(1).getNode();
5678 return N0->hasOneUse() && N1->hasOneUse() &&
5679 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5684 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5685 unsigned Opcode = N->getOpcode();
5686 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5687 SDNode *N0 = N->getOperand(0).getNode();
5688 SDNode *N1 = N->getOperand(1).getNode();
5689 return N0->hasOneUse() && N1->hasOneUse() &&
5690 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5695 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5696 // Multiplications are only custom-lowered for 128-bit vectors so that
5697 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5698 EVT VT = Op.getValueType();
5699 assert(VT.is128BitVector() && VT.isInteger() &&
5700 "unexpected type for custom-lowering ISD::MUL");
5701 SDNode *N0 = Op.getOperand(0).getNode();
5702 SDNode *N1 = Op.getOperand(1).getNode();
5703 unsigned NewOpc = 0;
5705 bool isN0SExt = isSignExtended(N0, DAG);
5706 bool isN1SExt = isSignExtended(N1, DAG);
5707 if (isN0SExt && isN1SExt)
5708 NewOpc = ARMISD::VMULLs;
5710 bool isN0ZExt = isZeroExtended(N0, DAG);
5711 bool isN1ZExt = isZeroExtended(N1, DAG);
5712 if (isN0ZExt && isN1ZExt)
5713 NewOpc = ARMISD::VMULLu;
5714 else if (isN1SExt || isN1ZExt) {
5715 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5716 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5717 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5718 NewOpc = ARMISD::VMULLs;
5720 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5721 NewOpc = ARMISD::VMULLu;
5723 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5725 NewOpc = ARMISD::VMULLu;
5731 if (VT == MVT::v2i64)
5732 // Fall through to expand this. It is not legal.
5735 // Other vector multiplications are legal.
5740 // Legalize to a VMULL instruction.
5743 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5745 Op0 = SkipExtensionForVMULL(N0, DAG);
5746 assert(Op0.getValueType().is64BitVector() &&
5747 Op1.getValueType().is64BitVector() &&
5748 "unexpected types for extended operands to VMULL");
5749 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5752 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5753 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5760 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5761 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5762 EVT Op1VT = Op1.getValueType();
5763 return DAG.getNode(N0->getOpcode(), DL, VT,
5764 DAG.getNode(NewOpc, DL, VT,
5765 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5766 DAG.getNode(NewOpc, DL, VT,
5767 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5771 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5773 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5774 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5775 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5776 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5777 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5778 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5779 // Get reciprocal estimate.
5780 // float4 recip = vrecpeq_f32(yf);
5781 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5782 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5783 // Because char has a smaller range than uchar, we can actually get away
5784 // without any newton steps. This requires that we use a weird bias
5785 // of 0xb000, however (again, this has been exhaustively tested).
5786 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5787 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5788 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5789 Y = DAG.getConstant(0xb000, MVT::i32);
5790 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5791 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5792 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5793 // Convert back to short.
5794 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5795 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5800 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5802 // Convert to float.
5803 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5804 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5805 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5806 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5807 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5808 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5810 // Use reciprocal estimate and one refinement step.
5811 // float4 recip = vrecpeq_f32(yf);
5812 // recip *= vrecpsq_f32(yf, recip);
5813 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5814 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5815 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5816 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5818 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5819 // Because short has a smaller range than ushort, we can actually get away
5820 // with only a single newton step. This requires that we use a weird bias
5821 // of 89, however (again, this has been exhaustively tested).
5822 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5823 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5824 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5825 N1 = DAG.getConstant(0x89, MVT::i32);
5826 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5827 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5828 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5829 // Convert back to integer and return.
5830 // return vmovn_s32(vcvt_s32_f32(result));
5831 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5832 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5836 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5837 EVT VT = Op.getValueType();
5838 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5839 "unexpected type for custom-lowering ISD::SDIV");
5842 SDValue N0 = Op.getOperand(0);
5843 SDValue N1 = Op.getOperand(1);
5846 if (VT == MVT::v8i8) {
5847 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5848 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5850 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5851 DAG.getIntPtrConstant(4));
5852 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5853 DAG.getIntPtrConstant(4));
5854 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5855 DAG.getIntPtrConstant(0));
5856 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5857 DAG.getIntPtrConstant(0));
5859 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5860 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5862 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5863 N0 = LowerCONCAT_VECTORS(N0, DAG);
5865 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5868 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5871 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5872 EVT VT = Op.getValueType();
5873 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5874 "unexpected type for custom-lowering ISD::UDIV");
5877 SDValue N0 = Op.getOperand(0);
5878 SDValue N1 = Op.getOperand(1);
5881 if (VT == MVT::v8i8) {
5882 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5883 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5885 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5886 DAG.getIntPtrConstant(4));
5887 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5888 DAG.getIntPtrConstant(4));
5889 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5890 DAG.getIntPtrConstant(0));
5891 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5892 DAG.getIntPtrConstant(0));
5894 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5895 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5897 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5898 N0 = LowerCONCAT_VECTORS(N0, DAG);
5900 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5901 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5906 // v4i16 sdiv ... Convert to float.
5907 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5908 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5909 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5910 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5911 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5912 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5914 // Use reciprocal estimate and two refinement steps.
5915 // float4 recip = vrecpeq_f32(yf);
5916 // recip *= vrecpsq_f32(yf, recip);
5917 // recip *= vrecpsq_f32(yf, recip);
5918 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5919 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5920 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5921 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5923 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5924 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5925 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5927 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5928 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5929 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5930 // and that it will never cause us to return an answer too large).
5931 // float4 result = as_float4(as_int4(xf*recip) + 2);
5932 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5933 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5934 N1 = DAG.getConstant(2, MVT::i32);
5935 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5936 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5937 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5938 // Convert back to integer and return.
5939 // return vmovn_u32(vcvt_s32_f32(result));
5940 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5941 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5945 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5946 EVT VT = Op.getNode()->getValueType(0);
5947 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5950 bool ExtraOp = false;
5951 switch (Op.getOpcode()) {
5952 default: llvm_unreachable("Invalid code");
5953 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5954 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5955 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5956 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5960 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5962 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5963 Op.getOperand(1), Op.getOperand(2));
5966 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
5967 assert(Subtarget->isTargetDarwin());
5969 // For iOS, we want to call an alternative entry point: __sincos_stret,
5970 // return values are passed via sret.
5972 SDValue Arg = Op.getOperand(0);
5973 EVT ArgVT = Arg.getValueType();
5974 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
5976 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5977 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5979 // Pair of floats / doubles used to pass the result.
5980 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
5982 // Create stack object for sret.
5983 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
5984 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
5985 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
5986 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
5992 Entry.Ty = RetTy->getPointerTo();
5993 Entry.isSExt = false;
5994 Entry.isZExt = false;
5995 Entry.isSRet = true;
5996 Args.push_back(Entry);
6000 Entry.isSExt = false;
6001 Entry.isZExt = false;
6002 Args.push_back(Entry);
6004 const char *LibcallName = (ArgVT == MVT::f64)
6005 ? "__sincos_stret" : "__sincosf_stret";
6006 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6009 CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
6010 false, false, false, false, 0,
6011 CallingConv::C, /*isTaillCall=*/false,
6012 /*doesNotRet=*/false, /*isReturnValueUsed*/false,
6013 Callee, Args, DAG, dl);
6014 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6016 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6017 MachinePointerInfo(), false, false, false, 0);
6019 // Address of cos field.
6020 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6021 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6022 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6023 MachinePointerInfo(), false, false, false, 0);
6025 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6026 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6027 LoadSin.getValue(0), LoadCos.getValue(0));
6030 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6031 // Monotonic load/store is legal for all targets
6032 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6035 // Aquire/Release load/store is not legal for targets without a
6036 // dmb or equivalent available.
6041 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
6042 SelectionDAG &DAG) {
6044 assert (Node->getValueType(0) == MVT::i64 &&
6045 "Only know how to expand i64 atomics");
6046 AtomicSDNode *AN = cast<AtomicSDNode>(Node);
6048 SmallVector<SDValue, 6> Ops;
6049 Ops.push_back(Node->getOperand(0)); // Chain
6050 Ops.push_back(Node->getOperand(1)); // Ptr
6051 for(unsigned i=2; i<Node->getNumOperands(); i++) {
6053 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6054 Node->getOperand(i), DAG.getIntPtrConstant(0)));
6056 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6057 Node->getOperand(i), DAG.getIntPtrConstant(1)));
6059 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6061 DAG.getAtomic(Node->getOpcode(), dl, MVT::i64, Tys, Ops.data(), Ops.size(),
6062 cast<MemSDNode>(Node)->getMemOperand(), AN->getOrdering(),
6063 AN->getSynchScope());
6064 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
6065 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6066 Results.push_back(Result.getValue(2));
6069 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6070 SmallVectorImpl<SDValue> &Results,
6072 const ARMSubtarget *Subtarget) {
6074 SDValue Cycles32, OutChain;
6076 if (Subtarget->hasPerfMon()) {
6077 // Under Power Management extensions, the cycle-count is:
6078 // mrc p15, #0, <Rt>, c9, c13, #0
6079 SDValue Ops[] = { N->getOperand(0), // Chain
6080 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6081 DAG.getConstant(15, MVT::i32),
6082 DAG.getConstant(0, MVT::i32),
6083 DAG.getConstant(9, MVT::i32),
6084 DAG.getConstant(13, MVT::i32),
6085 DAG.getConstant(0, MVT::i32)
6088 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6089 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6090 array_lengthof(Ops));
6091 OutChain = Cycles32.getValue(1);
6093 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6094 // there are older ARM CPUs that have implementation-specific ways of
6095 // obtaining this information (FIXME!).
6096 Cycles32 = DAG.getConstant(0, MVT::i32);
6097 OutChain = DAG.getEntryNode();
6101 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6102 Cycles32, DAG.getConstant(0, MVT::i32));
6103 Results.push_back(Cycles64);
6104 Results.push_back(OutChain);
6107 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6108 switch (Op.getOpcode()) {
6109 default: llvm_unreachable("Don't know how to custom lower this!");
6110 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6111 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6112 case ISD::GlobalAddress:
6113 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
6114 LowerGlobalAddressELF(Op, DAG);
6115 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6116 case ISD::SELECT: return LowerSELECT(Op, DAG);
6117 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6118 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6119 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6120 case ISD::VASTART: return LowerVASTART(Op, DAG);
6121 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6122 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6123 case ISD::SINT_TO_FP:
6124 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6125 case ISD::FP_TO_SINT:
6126 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6127 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6128 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6129 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6130 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6131 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6132 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6133 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6135 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6138 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6139 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6140 case ISD::SRL_PARTS:
6141 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6142 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6143 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6144 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6145 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6146 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6147 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6148 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6149 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6150 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6151 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6152 case ISD::MUL: return LowerMUL(Op, DAG);
6153 case ISD::SDIV: return LowerSDIV(Op, DAG);
6154 case ISD::UDIV: return LowerUDIV(Op, DAG);
6158 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6159 case ISD::ATOMIC_LOAD:
6160 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6161 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6163 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6167 /// ReplaceNodeResults - Replace the results of node with an illegal result
6168 /// type with new values built out of custom code.
6169 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6170 SmallVectorImpl<SDValue>&Results,
6171 SelectionDAG &DAG) const {
6173 switch (N->getOpcode()) {
6175 llvm_unreachable("Don't know how to custom expand this!");
6177 Res = ExpandBITCAST(N, DAG);
6181 Res = Expand64BitShift(N, DAG, Subtarget);
6183 case ISD::READCYCLECOUNTER:
6184 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6186 case ISD::ATOMIC_STORE:
6187 case ISD::ATOMIC_LOAD:
6188 case ISD::ATOMIC_LOAD_ADD:
6189 case ISD::ATOMIC_LOAD_AND:
6190 case ISD::ATOMIC_LOAD_NAND:
6191 case ISD::ATOMIC_LOAD_OR:
6192 case ISD::ATOMIC_LOAD_SUB:
6193 case ISD::ATOMIC_LOAD_XOR:
6194 case ISD::ATOMIC_SWAP:
6195 case ISD::ATOMIC_CMP_SWAP:
6196 case ISD::ATOMIC_LOAD_MIN:
6197 case ISD::ATOMIC_LOAD_UMIN:
6198 case ISD::ATOMIC_LOAD_MAX:
6199 case ISD::ATOMIC_LOAD_UMAX:
6200 ReplaceATOMIC_OP_64(N, Results, DAG);
6204 Results.push_back(Res);
6207 //===----------------------------------------------------------------------===//
6208 // ARM Scheduler Hooks
6209 //===----------------------------------------------------------------------===//
6212 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6213 MachineBasicBlock *BB,
6214 unsigned Size) const {
6215 unsigned dest = MI->getOperand(0).getReg();
6216 unsigned ptr = MI->getOperand(1).getReg();
6217 unsigned oldval = MI->getOperand(2).getReg();
6218 unsigned newval = MI->getOperand(3).getReg();
6219 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6220 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
6221 DebugLoc dl = MI->getDebugLoc();
6222 bool isThumb2 = Subtarget->isThumb2();
6224 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6225 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6226 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6227 (const TargetRegisterClass*)&ARM::GPRRegClass);
6230 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6231 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6232 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6235 unsigned ldrOpc, strOpc;
6236 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6238 MachineFunction *MF = BB->getParent();
6239 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6240 MachineFunction::iterator It = BB;
6241 ++It; // insert the new blocks after the current block
6243 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6244 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6245 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6246 MF->insert(It, loop1MBB);
6247 MF->insert(It, loop2MBB);
6248 MF->insert(It, exitMBB);
6250 // Transfer the remainder of BB and its successor edges to exitMBB.
6251 exitMBB->splice(exitMBB->begin(), BB,
6252 llvm::next(MachineBasicBlock::iterator(MI)),
6254 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6258 // fallthrough --> loop1MBB
6259 BB->addSuccessor(loop1MBB);
6262 // ldrex dest, [ptr]
6266 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6267 if (ldrOpc == ARM::t2LDREX)
6269 AddDefaultPred(MIB);
6270 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6271 .addReg(dest).addReg(oldval));
6272 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6273 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6274 BB->addSuccessor(loop2MBB);
6275 BB->addSuccessor(exitMBB);
6278 // strex scratch, newval, [ptr]
6282 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6283 if (strOpc == ARM::t2STREX)
6285 AddDefaultPred(MIB);
6286 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6287 .addReg(scratch).addImm(0));
6288 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6289 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6290 BB->addSuccessor(loop1MBB);
6291 BB->addSuccessor(exitMBB);
6297 MI->eraseFromParent(); // The instruction is gone now.
6303 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6304 unsigned Size, unsigned BinOpcode) const {
6305 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6308 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6309 MachineFunction *MF = BB->getParent();
6310 MachineFunction::iterator It = BB;
6313 unsigned dest = MI->getOperand(0).getReg();
6314 unsigned ptr = MI->getOperand(1).getReg();
6315 unsigned incr = MI->getOperand(2).getReg();
6316 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6317 DebugLoc dl = MI->getDebugLoc();
6318 bool isThumb2 = Subtarget->isThumb2();
6320 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6322 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6323 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6324 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6327 unsigned ldrOpc, strOpc;
6328 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6330 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6331 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6332 MF->insert(It, loopMBB);
6333 MF->insert(It, exitMBB);
6335 // Transfer the remainder of BB and its successor edges to exitMBB.
6336 exitMBB->splice(exitMBB->begin(), BB,
6337 llvm::next(MachineBasicBlock::iterator(MI)),
6339 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6341 const TargetRegisterClass *TRC = isThumb2 ?
6342 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6343 (const TargetRegisterClass*)&ARM::GPRRegClass;
6344 unsigned scratch = MRI.createVirtualRegister(TRC);
6345 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6349 // fallthrough --> loopMBB
6350 BB->addSuccessor(loopMBB);
6354 // <binop> scratch2, dest, incr
6355 // strex scratch, scratch2, ptr
6358 // fallthrough --> exitMBB
6360 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6361 if (ldrOpc == ARM::t2LDREX)
6363 AddDefaultPred(MIB);
6365 // operand order needs to go the other way for NAND
6366 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6367 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6368 addReg(incr).addReg(dest)).addReg(0);
6370 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6371 addReg(dest).addReg(incr)).addReg(0);
6374 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6375 if (strOpc == ARM::t2STREX)
6377 AddDefaultPred(MIB);
6378 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6379 .addReg(scratch).addImm(0));
6380 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6381 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6383 BB->addSuccessor(loopMBB);
6384 BB->addSuccessor(exitMBB);
6390 MI->eraseFromParent(); // The instruction is gone now.
6396 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6397 MachineBasicBlock *BB,
6400 ARMCC::CondCodes Cond) const {
6401 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6403 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6404 MachineFunction *MF = BB->getParent();
6405 MachineFunction::iterator It = BB;
6408 unsigned dest = MI->getOperand(0).getReg();
6409 unsigned ptr = MI->getOperand(1).getReg();
6410 unsigned incr = MI->getOperand(2).getReg();
6411 unsigned oldval = dest;
6412 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6413 DebugLoc dl = MI->getDebugLoc();
6414 bool isThumb2 = Subtarget->isThumb2();
6416 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6418 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6419 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6420 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6423 unsigned ldrOpc, strOpc, extendOpc;
6424 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6426 default: llvm_unreachable("unsupported size for AtomicBinaryMinMax!");
6428 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6431 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6438 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6439 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6440 MF->insert(It, loopMBB);
6441 MF->insert(It, exitMBB);
6443 // Transfer the remainder of BB and its successor edges to exitMBB.
6444 exitMBB->splice(exitMBB->begin(), BB,
6445 llvm::next(MachineBasicBlock::iterator(MI)),
6447 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6449 const TargetRegisterClass *TRC = isThumb2 ?
6450 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6451 (const TargetRegisterClass*)&ARM::GPRRegClass;
6452 unsigned scratch = MRI.createVirtualRegister(TRC);
6453 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6457 // fallthrough --> loopMBB
6458 BB->addSuccessor(loopMBB);
6462 // (sign extend dest, if required)
6464 // cmov.cond scratch2, incr, dest
6465 // strex scratch, scratch2, ptr
6468 // fallthrough --> exitMBB
6470 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6471 if (ldrOpc == ARM::t2LDREX)
6473 AddDefaultPred(MIB);
6475 // Sign extend the value, if necessary.
6476 if (signExtend && extendOpc) {
6477 oldval = MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass
6478 : &ARM::GPRnopcRegClass);
6480 MRI.constrainRegClass(dest, &ARM::GPRnopcRegClass);
6481 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6486 // Build compare and cmov instructions.
6487 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6488 .addReg(oldval).addReg(incr));
6489 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6490 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6492 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6493 if (strOpc == ARM::t2STREX)
6495 AddDefaultPred(MIB);
6496 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6497 .addReg(scratch).addImm(0));
6498 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6499 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6501 BB->addSuccessor(loopMBB);
6502 BB->addSuccessor(exitMBB);
6508 MI->eraseFromParent(); // The instruction is gone now.
6514 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6515 unsigned Op1, unsigned Op2,
6516 bool NeedsCarry, bool IsCmpxchg,
6517 bool IsMinMax, ARMCC::CondCodes CC) const {
6518 // This also handles ATOMIC_SWAP and ATOMIC_STORE, indicated by Op1==0.
6519 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6521 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6522 MachineFunction *MF = BB->getParent();
6523 MachineFunction::iterator It = BB;
6526 bool isStore = (MI->getOpcode() == ARM::ATOMIC_STORE_I64);
6527 unsigned offset = (isStore ? -2 : 0);
6528 unsigned destlo = MI->getOperand(0).getReg();
6529 unsigned desthi = MI->getOperand(1).getReg();
6530 unsigned ptr = MI->getOperand(offset+2).getReg();
6531 unsigned vallo = MI->getOperand(offset+3).getReg();
6532 unsigned valhi = MI->getOperand(offset+4).getReg();
6533 unsigned OrdIdx = offset + (IsCmpxchg ? 7 : 5);
6534 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(OrdIdx).getImm());
6535 DebugLoc dl = MI->getDebugLoc();
6536 bool isThumb2 = Subtarget->isThumb2();
6538 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6540 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6541 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6542 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6543 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6544 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6547 unsigned ldrOpc, strOpc;
6548 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6550 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6551 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6552 if (IsCmpxchg || IsMinMax)
6553 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6555 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6556 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6558 MF->insert(It, loopMBB);
6559 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6560 if (IsCmpxchg) MF->insert(It, cont2BB);
6561 MF->insert(It, exitMBB);
6563 // Transfer the remainder of BB and its successor edges to exitMBB.
6564 exitMBB->splice(exitMBB->begin(), BB,
6565 llvm::next(MachineBasicBlock::iterator(MI)),
6567 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6569 const TargetRegisterClass *TRC = isThumb2 ?
6570 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6571 (const TargetRegisterClass*)&ARM::GPRRegClass;
6572 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6576 // fallthrough --> loopMBB
6577 BB->addSuccessor(loopMBB);
6580 // ldrexd r2, r3, ptr
6581 // <binopa> r0, r2, incr
6582 // <binopb> r1, r3, incr
6583 // strexd storesuccess, r0, r1, ptr
6584 // cmp storesuccess, #0
6586 // fallthrough --> exitMBB
6592 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6593 .addReg(destlo, RegState::Define)
6594 .addReg(desthi, RegState::Define)
6597 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6598 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6599 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6600 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6601 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6602 .addReg(GPRPair0, 0, ARM::gsub_0);
6603 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6604 .addReg(GPRPair0, 0, ARM::gsub_1);
6608 unsigned StoreLo, StoreHi;
6611 for (unsigned i = 0; i < 2; i++) {
6612 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6614 .addReg(i == 0 ? destlo : desthi)
6615 .addReg(i == 0 ? vallo : valhi));
6616 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6617 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6618 BB->addSuccessor(exitMBB);
6619 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6620 BB = (i == 0 ? contBB : cont2BB);
6623 // Copy to physregs for strexd
6624 StoreLo = MI->getOperand(5).getReg();
6625 StoreHi = MI->getOperand(6).getReg();
6627 // Perform binary operation
6628 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6629 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6630 .addReg(destlo).addReg(vallo))
6631 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6632 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6633 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6634 .addReg(desthi).addReg(valhi))
6635 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6640 // Copy to physregs for strexd
6645 // Compare and branch to exit block.
6646 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6647 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6648 BB->addSuccessor(exitMBB);
6649 BB->addSuccessor(contBB);
6657 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6658 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6659 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6660 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6662 // Marshal a pair...
6663 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6664 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6665 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6666 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6667 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6670 .addImm(ARM::gsub_0);
6671 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6674 .addImm(ARM::gsub_1);
6677 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6678 .addReg(StorePair).addReg(ptr));
6681 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6682 .addReg(storesuccess).addImm(0));
6683 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6684 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6686 BB->addSuccessor(loopMBB);
6687 BB->addSuccessor(exitMBB);
6693 MI->eraseFromParent(); // The instruction is gone now.
6699 ARMTargetLowering::EmitAtomicLoad64(MachineInstr *MI, MachineBasicBlock *BB) const {
6701 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6703 unsigned destlo = MI->getOperand(0).getReg();
6704 unsigned desthi = MI->getOperand(1).getReg();
6705 unsigned ptr = MI->getOperand(2).getReg();
6706 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6707 DebugLoc dl = MI->getDebugLoc();
6708 bool isThumb2 = Subtarget->isThumb2();
6710 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6712 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6713 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6714 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6716 unsigned ldrOpc, strOpc;
6717 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6719 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(ldrOpc));
6722 MIB.addReg(destlo, RegState::Define)
6723 .addReg(desthi, RegState::Define)
6727 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6728 MIB.addReg(GPRPair0, RegState::Define).addReg(ptr);
6730 // Copy GPRPair0 into dest. (This copy will normally be coalesced.)
6731 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), destlo)
6732 .addReg(GPRPair0, 0, ARM::gsub_0);
6733 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), desthi)
6734 .addReg(GPRPair0, 0, ARM::gsub_1);
6736 AddDefaultPred(MIB);
6738 MI->eraseFromParent(); // The instruction is gone now.
6743 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6744 /// registers the function context.
6745 void ARMTargetLowering::
6746 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6747 MachineBasicBlock *DispatchBB, int FI) const {
6748 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6749 DebugLoc dl = MI->getDebugLoc();
6750 MachineFunction *MF = MBB->getParent();
6751 MachineRegisterInfo *MRI = &MF->getRegInfo();
6752 MachineConstantPool *MCP = MF->getConstantPool();
6753 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6754 const Function *F = MF->getFunction();
6756 bool isThumb = Subtarget->isThumb();
6757 bool isThumb2 = Subtarget->isThumb2();
6759 unsigned PCLabelId = AFI->createPICLabelUId();
6760 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6761 ARMConstantPoolValue *CPV =
6762 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6763 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6765 const TargetRegisterClass *TRC = isThumb ?
6766 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6767 (const TargetRegisterClass*)&ARM::GPRRegClass;
6769 // Grab constant pool and fixed stack memory operands.
6770 MachineMemOperand *CPMMO =
6771 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6772 MachineMemOperand::MOLoad, 4, 4);
6774 MachineMemOperand *FIMMOSt =
6775 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6776 MachineMemOperand::MOStore, 4, 4);
6778 // Load the address of the dispatch MBB into the jump buffer.
6780 // Incoming value: jbuf
6781 // ldr.n r5, LCPI1_1
6784 // str r5, [$jbuf, #+4] ; &jbuf[1]
6785 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6786 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6787 .addConstantPoolIndex(CPI)
6788 .addMemOperand(CPMMO));
6789 // Set the low bit because of thumb mode.
6790 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6792 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6793 .addReg(NewVReg1, RegState::Kill)
6795 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6796 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6797 .addReg(NewVReg2, RegState::Kill)
6799 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6800 .addReg(NewVReg3, RegState::Kill)
6802 .addImm(36) // &jbuf[1] :: pc
6803 .addMemOperand(FIMMOSt));
6804 } else if (isThumb) {
6805 // Incoming value: jbuf
6806 // ldr.n r1, LCPI1_4
6810 // add r2, $jbuf, #+4 ; &jbuf[1]
6812 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6813 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6814 .addConstantPoolIndex(CPI)
6815 .addMemOperand(CPMMO));
6816 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6817 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6818 .addReg(NewVReg1, RegState::Kill)
6820 // Set the low bit because of thumb mode.
6821 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6822 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6823 .addReg(ARM::CPSR, RegState::Define)
6825 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6826 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6827 .addReg(ARM::CPSR, RegState::Define)
6828 .addReg(NewVReg2, RegState::Kill)
6829 .addReg(NewVReg3, RegState::Kill));
6830 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6831 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6833 .addImm(36)); // &jbuf[1] :: pc
6834 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6835 .addReg(NewVReg4, RegState::Kill)
6836 .addReg(NewVReg5, RegState::Kill)
6838 .addMemOperand(FIMMOSt));
6840 // Incoming value: jbuf
6843 // str r1, [$jbuf, #+4] ; &jbuf[1]
6844 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6845 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6846 .addConstantPoolIndex(CPI)
6848 .addMemOperand(CPMMO));
6849 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6850 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6851 .addReg(NewVReg1, RegState::Kill)
6852 .addImm(PCLabelId));
6853 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6854 .addReg(NewVReg2, RegState::Kill)
6856 .addImm(36) // &jbuf[1] :: pc
6857 .addMemOperand(FIMMOSt));
6861 MachineBasicBlock *ARMTargetLowering::
6862 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6863 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6864 DebugLoc dl = MI->getDebugLoc();
6865 MachineFunction *MF = MBB->getParent();
6866 MachineRegisterInfo *MRI = &MF->getRegInfo();
6867 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6868 MachineFrameInfo *MFI = MF->getFrameInfo();
6869 int FI = MFI->getFunctionContextIndex();
6871 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6872 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6873 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6875 // Get a mapping of the call site numbers to all of the landing pads they're
6877 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6878 unsigned MaxCSNum = 0;
6879 MachineModuleInfo &MMI = MF->getMMI();
6880 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6882 if (!BB->isLandingPad()) continue;
6884 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6886 for (MachineBasicBlock::iterator
6887 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6888 if (!II->isEHLabel()) continue;
6890 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6891 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6893 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6894 for (SmallVectorImpl<unsigned>::iterator
6895 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6896 CSI != CSE; ++CSI) {
6897 CallSiteNumToLPad[*CSI].push_back(BB);
6898 MaxCSNum = std::max(MaxCSNum, *CSI);
6904 // Get an ordered list of the machine basic blocks for the jump table.
6905 std::vector<MachineBasicBlock*> LPadList;
6906 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6907 LPadList.reserve(CallSiteNumToLPad.size());
6908 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6909 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6910 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6911 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6912 LPadList.push_back(*II);
6913 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6917 assert(!LPadList.empty() &&
6918 "No landing pad destinations for the dispatch jump table!");
6920 // Create the jump table and associated information.
6921 MachineJumpTableInfo *JTI =
6922 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6923 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6924 unsigned UId = AFI->createJumpTableUId();
6925 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6927 // Create the MBBs for the dispatch code.
6929 // Shove the dispatch's address into the return slot in the function context.
6930 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6931 DispatchBB->setIsLandingPad();
6933 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6934 unsigned trap_opcode;
6935 if (Subtarget->isThumb())
6936 trap_opcode = ARM::tTRAP;
6938 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6940 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6941 DispatchBB->addSuccessor(TrapBB);
6943 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6944 DispatchBB->addSuccessor(DispContBB);
6947 MF->insert(MF->end(), DispatchBB);
6948 MF->insert(MF->end(), DispContBB);
6949 MF->insert(MF->end(), TrapBB);
6951 // Insert code into the entry block that creates and registers the function
6953 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6955 MachineMemOperand *FIMMOLd =
6956 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6957 MachineMemOperand::MOLoad |
6958 MachineMemOperand::MOVolatile, 4, 4);
6960 MachineInstrBuilder MIB;
6961 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6963 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6964 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6966 // Add a register mask with no preserved registers. This results in all
6967 // registers being marked as clobbered.
6968 MIB.addRegMask(RI.getNoPreservedMask());
6970 unsigned NumLPads = LPadList.size();
6971 if (Subtarget->isThumb2()) {
6972 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6973 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6976 .addMemOperand(FIMMOLd));
6978 if (NumLPads < 256) {
6979 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6981 .addImm(LPadList.size()));
6983 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6984 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6985 .addImm(NumLPads & 0xFFFF));
6987 unsigned VReg2 = VReg1;
6988 if ((NumLPads & 0xFFFF0000) != 0) {
6989 VReg2 = MRI->createVirtualRegister(TRC);
6990 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6992 .addImm(NumLPads >> 16));
6995 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7000 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7005 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7006 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7007 .addJumpTableIndex(MJTI)
7010 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7013 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7014 .addReg(NewVReg3, RegState::Kill)
7016 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7018 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7019 .addReg(NewVReg4, RegState::Kill)
7021 .addJumpTableIndex(MJTI)
7023 } else if (Subtarget->isThumb()) {
7024 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7025 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7028 .addMemOperand(FIMMOLd));
7030 if (NumLPads < 256) {
7031 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7035 MachineConstantPool *ConstantPool = MF->getConstantPool();
7036 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7037 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7039 // MachineConstantPool wants an explicit alignment.
7040 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7042 Align = getDataLayout()->getTypeAllocSize(C->getType());
7043 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7045 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7046 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7047 .addReg(VReg1, RegState::Define)
7048 .addConstantPoolIndex(Idx));
7049 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7054 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7059 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7060 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7061 .addReg(ARM::CPSR, RegState::Define)
7065 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7066 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7067 .addJumpTableIndex(MJTI)
7070 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7071 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7072 .addReg(ARM::CPSR, RegState::Define)
7073 .addReg(NewVReg2, RegState::Kill)
7076 MachineMemOperand *JTMMOLd =
7077 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7078 MachineMemOperand::MOLoad, 4, 4);
7080 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7081 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7082 .addReg(NewVReg4, RegState::Kill)
7084 .addMemOperand(JTMMOLd));
7086 unsigned NewVReg6 = NewVReg5;
7087 if (RelocM == Reloc::PIC_) {
7088 NewVReg6 = MRI->createVirtualRegister(TRC);
7089 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7090 .addReg(ARM::CPSR, RegState::Define)
7091 .addReg(NewVReg5, RegState::Kill)
7095 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7096 .addReg(NewVReg6, RegState::Kill)
7097 .addJumpTableIndex(MJTI)
7100 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7101 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7104 .addMemOperand(FIMMOLd));
7106 if (NumLPads < 256) {
7107 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7110 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7111 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7112 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7113 .addImm(NumLPads & 0xFFFF));
7115 unsigned VReg2 = VReg1;
7116 if ((NumLPads & 0xFFFF0000) != 0) {
7117 VReg2 = MRI->createVirtualRegister(TRC);
7118 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7120 .addImm(NumLPads >> 16));
7123 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7127 MachineConstantPool *ConstantPool = MF->getConstantPool();
7128 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7129 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7131 // MachineConstantPool wants an explicit alignment.
7132 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7134 Align = getDataLayout()->getTypeAllocSize(C->getType());
7135 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7137 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7138 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7139 .addReg(VReg1, RegState::Define)
7140 .addConstantPoolIndex(Idx)
7142 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7144 .addReg(VReg1, RegState::Kill));
7147 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7152 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7154 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7156 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7157 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7158 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7159 .addJumpTableIndex(MJTI)
7162 MachineMemOperand *JTMMOLd =
7163 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7164 MachineMemOperand::MOLoad, 4, 4);
7165 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7167 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7168 .addReg(NewVReg3, RegState::Kill)
7171 .addMemOperand(JTMMOLd));
7173 if (RelocM == Reloc::PIC_) {
7174 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7175 .addReg(NewVReg5, RegState::Kill)
7177 .addJumpTableIndex(MJTI)
7180 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7181 .addReg(NewVReg5, RegState::Kill)
7182 .addJumpTableIndex(MJTI)
7187 // Add the jump table entries as successors to the MBB.
7188 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7189 for (std::vector<MachineBasicBlock*>::iterator
7190 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7191 MachineBasicBlock *CurMBB = *I;
7192 if (SeenMBBs.insert(CurMBB))
7193 DispContBB->addSuccessor(CurMBB);
7196 // N.B. the order the invoke BBs are processed in doesn't matter here.
7197 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7198 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7199 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7200 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7201 MachineBasicBlock *BB = *I;
7203 // Remove the landing pad successor from the invoke block and replace it
7204 // with the new dispatch block.
7205 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7207 while (!Successors.empty()) {
7208 MachineBasicBlock *SMBB = Successors.pop_back_val();
7209 if (SMBB->isLandingPad()) {
7210 BB->removeSuccessor(SMBB);
7211 MBBLPads.push_back(SMBB);
7215 BB->addSuccessor(DispatchBB);
7217 // Find the invoke call and mark all of the callee-saved registers as
7218 // 'implicit defined' so that they're spilled. This prevents code from
7219 // moving instructions to before the EH block, where they will never be
7221 for (MachineBasicBlock::reverse_iterator
7222 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7223 if (!II->isCall()) continue;
7225 DenseMap<unsigned, bool> DefRegs;
7226 for (MachineInstr::mop_iterator
7227 OI = II->operands_begin(), OE = II->operands_end();
7229 if (!OI->isReg()) continue;
7230 DefRegs[OI->getReg()] = true;
7233 MachineInstrBuilder MIB(*MF, &*II);
7235 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7236 unsigned Reg = SavedRegs[i];
7237 if (Subtarget->isThumb2() &&
7238 !ARM::tGPRRegClass.contains(Reg) &&
7239 !ARM::hGPRRegClass.contains(Reg))
7241 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7243 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7246 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7253 // Mark all former landing pads as non-landing pads. The dispatch is the only
7255 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7256 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7257 (*I)->setIsLandingPad(false);
7259 // The instruction is gone now.
7260 MI->eraseFromParent();
7266 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7267 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7268 E = MBB->succ_end(); I != E; ++I)
7271 llvm_unreachable("Expecting a BB with two successors!");
7274 /// Return the load opcode for a given load size. If load size >= 8,
7275 /// neon opcode will be returned.
7276 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7278 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7279 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7281 return LdSize == 4 ? ARM::tLDRi
7282 : LdSize == 2 ? ARM::tLDRHi
7283 : LdSize == 1 ? ARM::tLDRBi : 0;
7285 return LdSize == 4 ? ARM::t2LDR_POST
7286 : LdSize == 2 ? ARM::t2LDRH_POST
7287 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7288 return LdSize == 4 ? ARM::LDR_POST_IMM
7289 : LdSize == 2 ? ARM::LDRH_POST
7290 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7293 /// Return the store opcode for a given store size. If store size >= 8,
7294 /// neon opcode will be returned.
7295 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7297 return StSize == 16 ? ARM::VST1q32wb_fixed
7298 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7300 return StSize == 4 ? ARM::tSTRi
7301 : StSize == 2 ? ARM::tSTRHi
7302 : StSize == 1 ? ARM::tSTRBi : 0;
7304 return StSize == 4 ? ARM::t2STR_POST
7305 : StSize == 2 ? ARM::t2STRH_POST
7306 : StSize == 1 ? ARM::t2STRB_POST : 0;
7307 return StSize == 4 ? ARM::STR_POST_IMM
7308 : StSize == 2 ? ARM::STRH_POST
7309 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7312 /// Emit a post-increment load operation with given size. The instructions
7313 /// will be added to BB at Pos.
7314 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7315 const TargetInstrInfo *TII, DebugLoc dl,
7316 unsigned LdSize, unsigned Data, unsigned AddrIn,
7317 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7318 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7319 assert(LdOpc != 0 && "Should have a load opcode");
7321 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7322 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7324 } else if (IsThumb1) {
7325 // load + update AddrIn
7326 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7327 .addReg(AddrIn).addImm(0));
7328 MachineInstrBuilder MIB =
7329 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7330 MIB = AddDefaultT1CC(MIB);
7331 MIB.addReg(AddrIn).addImm(LdSize);
7332 AddDefaultPred(MIB);
7333 } else if (IsThumb2) {
7334 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7335 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7338 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7339 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7340 .addReg(0).addImm(LdSize));
7344 /// Emit a post-increment store operation with given size. The instructions
7345 /// will be added to BB at Pos.
7346 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7347 const TargetInstrInfo *TII, DebugLoc dl,
7348 unsigned StSize, unsigned Data, unsigned AddrIn,
7349 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7350 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7351 assert(StOpc != 0 && "Should have a store opcode");
7353 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7354 .addReg(AddrIn).addImm(0).addReg(Data));
7355 } else if (IsThumb1) {
7356 // store + update AddrIn
7357 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7358 .addReg(AddrIn).addImm(0));
7359 MachineInstrBuilder MIB =
7360 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7361 MIB = AddDefaultT1CC(MIB);
7362 MIB.addReg(AddrIn).addImm(StSize);
7363 AddDefaultPred(MIB);
7364 } else if (IsThumb2) {
7365 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7366 .addReg(Data).addReg(AddrIn).addImm(StSize));
7368 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7369 .addReg(Data).addReg(AddrIn).addReg(0)
7375 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7376 MachineBasicBlock *BB) const {
7377 // This pseudo instruction has 3 operands: dst, src, size
7378 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7379 // Otherwise, we will generate unrolled scalar copies.
7380 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7381 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7382 MachineFunction::iterator It = BB;
7385 unsigned dest = MI->getOperand(0).getReg();
7386 unsigned src = MI->getOperand(1).getReg();
7387 unsigned SizeVal = MI->getOperand(2).getImm();
7388 unsigned Align = MI->getOperand(3).getImm();
7389 DebugLoc dl = MI->getDebugLoc();
7391 MachineFunction *MF = BB->getParent();
7392 MachineRegisterInfo &MRI = MF->getRegInfo();
7393 unsigned UnitSize = 0;
7394 const TargetRegisterClass *TRC = 0;
7395 const TargetRegisterClass *VecTRC = 0;
7397 bool IsThumb1 = Subtarget->isThumb1Only();
7398 bool IsThumb2 = Subtarget->isThumb2();
7402 } else if (Align & 2) {
7405 // Check whether we can use NEON instructions.
7406 if (!MF->getFunction()->getAttributes().
7407 hasAttribute(AttributeSet::FunctionIndex,
7408 Attribute::NoImplicitFloat) &&
7409 Subtarget->hasNEON()) {
7410 if ((Align % 16 == 0) && SizeVal >= 16)
7412 else if ((Align % 8 == 0) && SizeVal >= 8)
7415 // Can't use NEON instructions.
7420 // Select the correct opcode and register class for unit size load/store
7421 bool IsNeon = UnitSize >= 8;
7422 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
7423 : (const TargetRegisterClass *)&ARM::GPRRegClass;
7425 VecTRC = UnitSize == 16
7426 ? (const TargetRegisterClass *)&ARM::DPairRegClass
7428 ? (const TargetRegisterClass *)&ARM::DPRRegClass
7431 unsigned BytesLeft = SizeVal % UnitSize;
7432 unsigned LoopSize = SizeVal - BytesLeft;
7434 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7435 // Use LDR and STR to copy.
7436 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7437 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7438 unsigned srcIn = src;
7439 unsigned destIn = dest;
7440 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7441 unsigned srcOut = MRI.createVirtualRegister(TRC);
7442 unsigned destOut = MRI.createVirtualRegister(TRC);
7443 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7444 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7445 IsThumb1, IsThumb2);
7446 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7447 IsThumb1, IsThumb2);
7452 // Handle the leftover bytes with LDRB and STRB.
7453 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7454 // [destOut] = STRB_POST(scratch, destIn, 1)
7455 for (unsigned i = 0; i < BytesLeft; i++) {
7456 unsigned srcOut = MRI.createVirtualRegister(TRC);
7457 unsigned destOut = MRI.createVirtualRegister(TRC);
7458 unsigned scratch = MRI.createVirtualRegister(TRC);
7459 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7460 IsThumb1, IsThumb2);
7461 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7462 IsThumb1, IsThumb2);
7466 MI->eraseFromParent(); // The instruction is gone now.
7470 // Expand the pseudo op to a loop.
7473 // movw varEnd, # --> with thumb2
7475 // ldrcp varEnd, idx --> without thumb2
7476 // fallthrough --> loopMBB
7478 // PHI varPhi, varEnd, varLoop
7479 // PHI srcPhi, src, srcLoop
7480 // PHI destPhi, dst, destLoop
7481 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7482 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7483 // subs varLoop, varPhi, #UnitSize
7485 // fallthrough --> exitMBB
7487 // epilogue to handle left-over bytes
7488 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7489 // [destOut] = STRB_POST(scratch, destLoop, 1)
7490 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7491 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7492 MF->insert(It, loopMBB);
7493 MF->insert(It, exitMBB);
7495 // Transfer the remainder of BB and its successor edges to exitMBB.
7496 exitMBB->splice(exitMBB->begin(), BB,
7497 llvm::next(MachineBasicBlock::iterator(MI)),
7499 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7501 // Load an immediate to varEnd.
7502 unsigned varEnd = MRI.createVirtualRegister(TRC);
7504 unsigned Vtmp = varEnd;
7505 if ((LoopSize & 0xFFFF0000) != 0)
7506 Vtmp = MRI.createVirtualRegister(TRC);
7507 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7508 .addImm(LoopSize & 0xFFFF));
7510 if ((LoopSize & 0xFFFF0000) != 0)
7511 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7512 .addReg(Vtmp).addImm(LoopSize >> 16));
7514 MachineConstantPool *ConstantPool = MF->getConstantPool();
7515 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7516 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7518 // MachineConstantPool wants an explicit alignment.
7519 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7521 Align = getDataLayout()->getTypeAllocSize(C->getType());
7522 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7525 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7526 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7528 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7529 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7531 BB->addSuccessor(loopMBB);
7533 // Generate the loop body:
7534 // varPhi = PHI(varLoop, varEnd)
7535 // srcPhi = PHI(srcLoop, src)
7536 // destPhi = PHI(destLoop, dst)
7537 MachineBasicBlock *entryBB = BB;
7539 unsigned varLoop = MRI.createVirtualRegister(TRC);
7540 unsigned varPhi = MRI.createVirtualRegister(TRC);
7541 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7542 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7543 unsigned destLoop = MRI.createVirtualRegister(TRC);
7544 unsigned destPhi = MRI.createVirtualRegister(TRC);
7546 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7547 .addReg(varLoop).addMBB(loopMBB)
7548 .addReg(varEnd).addMBB(entryBB);
7549 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7550 .addReg(srcLoop).addMBB(loopMBB)
7551 .addReg(src).addMBB(entryBB);
7552 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7553 .addReg(destLoop).addMBB(loopMBB)
7554 .addReg(dest).addMBB(entryBB);
7556 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7557 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7558 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7559 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7560 IsThumb1, IsThumb2);
7561 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7562 IsThumb1, IsThumb2);
7564 // Decrement loop variable by UnitSize.
7566 MachineInstrBuilder MIB =
7567 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7568 MIB = AddDefaultT1CC(MIB);
7569 MIB.addReg(varPhi).addImm(UnitSize);
7570 AddDefaultPred(MIB);
7572 MachineInstrBuilder MIB =
7573 BuildMI(*BB, BB->end(), dl,
7574 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7575 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7576 MIB->getOperand(5).setReg(ARM::CPSR);
7577 MIB->getOperand(5).setIsDef(true);
7579 BuildMI(*BB, BB->end(), dl,
7580 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7581 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7583 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7584 BB->addSuccessor(loopMBB);
7585 BB->addSuccessor(exitMBB);
7587 // Add epilogue to handle BytesLeft.
7589 MachineInstr *StartOfExit = exitMBB->begin();
7591 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7592 // [destOut] = STRB_POST(scratch, destLoop, 1)
7593 unsigned srcIn = srcLoop;
7594 unsigned destIn = destLoop;
7595 for (unsigned i = 0; i < BytesLeft; i++) {
7596 unsigned srcOut = MRI.createVirtualRegister(TRC);
7597 unsigned destOut = MRI.createVirtualRegister(TRC);
7598 unsigned scratch = MRI.createVirtualRegister(TRC);
7599 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7600 IsThumb1, IsThumb2);
7601 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7602 IsThumb1, IsThumb2);
7607 MI->eraseFromParent(); // The instruction is gone now.
7612 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7613 MachineBasicBlock *BB) const {
7614 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7615 DebugLoc dl = MI->getDebugLoc();
7616 bool isThumb2 = Subtarget->isThumb2();
7617 switch (MI->getOpcode()) {
7620 llvm_unreachable("Unexpected instr type to insert");
7622 // The Thumb2 pre-indexed stores have the same MI operands, they just
7623 // define them differently in the .td files from the isel patterns, so
7624 // they need pseudos.
7625 case ARM::t2STR_preidx:
7626 MI->setDesc(TII->get(ARM::t2STR_PRE));
7628 case ARM::t2STRB_preidx:
7629 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7631 case ARM::t2STRH_preidx:
7632 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7635 case ARM::STRi_preidx:
7636 case ARM::STRBi_preidx: {
7637 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7638 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7639 // Decode the offset.
7640 unsigned Offset = MI->getOperand(4).getImm();
7641 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7642 Offset = ARM_AM::getAM2Offset(Offset);
7646 MachineMemOperand *MMO = *MI->memoperands_begin();
7647 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7648 .addOperand(MI->getOperand(0)) // Rn_wb
7649 .addOperand(MI->getOperand(1)) // Rt
7650 .addOperand(MI->getOperand(2)) // Rn
7651 .addImm(Offset) // offset (skip GPR==zero_reg)
7652 .addOperand(MI->getOperand(5)) // pred
7653 .addOperand(MI->getOperand(6))
7654 .addMemOperand(MMO);
7655 MI->eraseFromParent();
7658 case ARM::STRr_preidx:
7659 case ARM::STRBr_preidx:
7660 case ARM::STRH_preidx: {
7662 switch (MI->getOpcode()) {
7663 default: llvm_unreachable("unexpected opcode!");
7664 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7665 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7666 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7668 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7669 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7670 MIB.addOperand(MI->getOperand(i));
7671 MI->eraseFromParent();
7674 case ARM::ATOMIC_LOAD_ADD_I8:
7675 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7676 case ARM::ATOMIC_LOAD_ADD_I16:
7677 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7678 case ARM::ATOMIC_LOAD_ADD_I32:
7679 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7681 case ARM::ATOMIC_LOAD_AND_I8:
7682 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7683 case ARM::ATOMIC_LOAD_AND_I16:
7684 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7685 case ARM::ATOMIC_LOAD_AND_I32:
7686 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7688 case ARM::ATOMIC_LOAD_OR_I8:
7689 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7690 case ARM::ATOMIC_LOAD_OR_I16:
7691 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7692 case ARM::ATOMIC_LOAD_OR_I32:
7693 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7695 case ARM::ATOMIC_LOAD_XOR_I8:
7696 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7697 case ARM::ATOMIC_LOAD_XOR_I16:
7698 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7699 case ARM::ATOMIC_LOAD_XOR_I32:
7700 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7702 case ARM::ATOMIC_LOAD_NAND_I8:
7703 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7704 case ARM::ATOMIC_LOAD_NAND_I16:
7705 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7706 case ARM::ATOMIC_LOAD_NAND_I32:
7707 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7709 case ARM::ATOMIC_LOAD_SUB_I8:
7710 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7711 case ARM::ATOMIC_LOAD_SUB_I16:
7712 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7713 case ARM::ATOMIC_LOAD_SUB_I32:
7714 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7716 case ARM::ATOMIC_LOAD_MIN_I8:
7717 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7718 case ARM::ATOMIC_LOAD_MIN_I16:
7719 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7720 case ARM::ATOMIC_LOAD_MIN_I32:
7721 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7723 case ARM::ATOMIC_LOAD_MAX_I8:
7724 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7725 case ARM::ATOMIC_LOAD_MAX_I16:
7726 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7727 case ARM::ATOMIC_LOAD_MAX_I32:
7728 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7730 case ARM::ATOMIC_LOAD_UMIN_I8:
7731 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7732 case ARM::ATOMIC_LOAD_UMIN_I16:
7733 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7734 case ARM::ATOMIC_LOAD_UMIN_I32:
7735 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7737 case ARM::ATOMIC_LOAD_UMAX_I8:
7738 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7739 case ARM::ATOMIC_LOAD_UMAX_I16:
7740 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7741 case ARM::ATOMIC_LOAD_UMAX_I32:
7742 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7744 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7745 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7746 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7748 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7749 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7750 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7752 case ARM::ATOMIC_LOAD_I64:
7753 return EmitAtomicLoad64(MI, BB);
7755 case ARM::ATOMIC_LOAD_ADD_I64:
7756 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7757 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7758 /*NeedsCarry*/ true);
7759 case ARM::ATOMIC_LOAD_SUB_I64:
7760 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7761 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7762 /*NeedsCarry*/ true);
7763 case ARM::ATOMIC_LOAD_OR_I64:
7764 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7765 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7766 case ARM::ATOMIC_LOAD_XOR_I64:
7767 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7768 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7769 case ARM::ATOMIC_LOAD_AND_I64:
7770 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7771 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7772 case ARM::ATOMIC_STORE_I64:
7773 case ARM::ATOMIC_SWAP_I64:
7774 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7775 case ARM::ATOMIC_CMP_SWAP_I64:
7776 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7777 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7778 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7779 case ARM::ATOMIC_LOAD_MIN_I64:
7780 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7781 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7782 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7783 /*IsMinMax*/ true, ARMCC::LT);
7784 case ARM::ATOMIC_LOAD_MAX_I64:
7785 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7786 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7787 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7788 /*IsMinMax*/ true, ARMCC::GE);
7789 case ARM::ATOMIC_LOAD_UMIN_I64:
7790 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7791 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7792 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7793 /*IsMinMax*/ true, ARMCC::LO);
7794 case ARM::ATOMIC_LOAD_UMAX_I64:
7795 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7796 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7797 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7798 /*IsMinMax*/ true, ARMCC::HS);
7800 case ARM::tMOVCCr_pseudo: {
7801 // To "insert" a SELECT_CC instruction, we actually have to insert the
7802 // diamond control-flow pattern. The incoming instruction knows the
7803 // destination vreg to set, the condition code register to branch on, the
7804 // true/false values to select between, and a branch opcode to use.
7805 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7806 MachineFunction::iterator It = BB;
7812 // cmpTY ccX, r1, r2
7814 // fallthrough --> copy0MBB
7815 MachineBasicBlock *thisMBB = BB;
7816 MachineFunction *F = BB->getParent();
7817 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7818 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7819 F->insert(It, copy0MBB);
7820 F->insert(It, sinkMBB);
7822 // Transfer the remainder of BB and its successor edges to sinkMBB.
7823 sinkMBB->splice(sinkMBB->begin(), BB,
7824 llvm::next(MachineBasicBlock::iterator(MI)),
7826 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7828 BB->addSuccessor(copy0MBB);
7829 BB->addSuccessor(sinkMBB);
7831 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7832 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7835 // %FalseValue = ...
7836 // # fallthrough to sinkMBB
7839 // Update machine-CFG edges
7840 BB->addSuccessor(sinkMBB);
7843 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7846 BuildMI(*BB, BB->begin(), dl,
7847 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7848 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7849 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7851 MI->eraseFromParent(); // The pseudo instruction is gone now.
7856 case ARM::BCCZi64: {
7857 // If there is an unconditional branch to the other successor, remove it.
7858 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7860 // Compare both parts that make up the double comparison separately for
7862 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7864 unsigned LHS1 = MI->getOperand(1).getReg();
7865 unsigned LHS2 = MI->getOperand(2).getReg();
7867 AddDefaultPred(BuildMI(BB, dl,
7868 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7869 .addReg(LHS1).addImm(0));
7870 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7871 .addReg(LHS2).addImm(0)
7872 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7874 unsigned RHS1 = MI->getOperand(3).getReg();
7875 unsigned RHS2 = MI->getOperand(4).getReg();
7876 AddDefaultPred(BuildMI(BB, dl,
7877 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7878 .addReg(LHS1).addReg(RHS1));
7879 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7880 .addReg(LHS2).addReg(RHS2)
7881 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7884 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7885 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7886 if (MI->getOperand(0).getImm() == ARMCC::NE)
7887 std::swap(destMBB, exitMBB);
7889 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7890 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7892 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7894 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7896 MI->eraseFromParent(); // The pseudo instruction is gone now.
7900 case ARM::Int_eh_sjlj_setjmp:
7901 case ARM::Int_eh_sjlj_setjmp_nofp:
7902 case ARM::tInt_eh_sjlj_setjmp:
7903 case ARM::t2Int_eh_sjlj_setjmp:
7904 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7905 EmitSjLjDispatchBlock(MI, BB);
7910 // To insert an ABS instruction, we have to insert the
7911 // diamond control-flow pattern. The incoming instruction knows the
7912 // source vreg to test against 0, the destination vreg to set,
7913 // the condition code register to branch on, the
7914 // true/false values to select between, and a branch opcode to use.
7919 // BCC (branch to SinkBB if V0 >= 0)
7920 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7921 // SinkBB: V1 = PHI(V2, V3)
7922 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7923 MachineFunction::iterator BBI = BB;
7925 MachineFunction *Fn = BB->getParent();
7926 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7927 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7928 Fn->insert(BBI, RSBBB);
7929 Fn->insert(BBI, SinkBB);
7931 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7932 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7933 bool isThumb2 = Subtarget->isThumb2();
7934 MachineRegisterInfo &MRI = Fn->getRegInfo();
7935 // In Thumb mode S must not be specified if source register is the SP or
7936 // PC and if destination register is the SP, so restrict register class
7937 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7938 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7939 (const TargetRegisterClass*)&ARM::GPRRegClass);
7941 // Transfer the remainder of BB and its successor edges to sinkMBB.
7942 SinkBB->splice(SinkBB->begin(), BB,
7943 llvm::next(MachineBasicBlock::iterator(MI)),
7945 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7947 BB->addSuccessor(RSBBB);
7948 BB->addSuccessor(SinkBB);
7950 // fall through to SinkMBB
7951 RSBBB->addSuccessor(SinkBB);
7953 // insert a cmp at the end of BB
7954 AddDefaultPred(BuildMI(BB, dl,
7955 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7956 .addReg(ABSSrcReg).addImm(0));
7958 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7960 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7961 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7963 // insert rsbri in RSBBB
7964 // Note: BCC and rsbri will be converted into predicated rsbmi
7965 // by if-conversion pass
7966 BuildMI(*RSBBB, RSBBB->begin(), dl,
7967 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7968 .addReg(ABSSrcReg, RegState::Kill)
7969 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7971 // insert PHI in SinkBB,
7972 // reuse ABSDstReg to not change uses of ABS instruction
7973 BuildMI(*SinkBB, SinkBB->begin(), dl,
7974 TII->get(ARM::PHI), ABSDstReg)
7975 .addReg(NewRsbDstReg).addMBB(RSBBB)
7976 .addReg(ABSSrcReg).addMBB(BB);
7978 // remove ABS instruction
7979 MI->eraseFromParent();
7981 // return last added BB
7984 case ARM::COPY_STRUCT_BYVAL_I32:
7986 return EmitStructByval(MI, BB);
7990 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7991 SDNode *Node) const {
7992 if (!MI->hasPostISelHook()) {
7993 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7994 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7998 const MCInstrDesc *MCID = &MI->getDesc();
7999 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8000 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8001 // operand is still set to noreg. If needed, set the optional operand's
8002 // register to CPSR, and remove the redundant implicit def.
8004 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8006 // Rename pseudo opcodes.
8007 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8009 const ARMBaseInstrInfo *TII =
8010 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
8011 MCID = &TII->get(NewOpc);
8013 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8014 "converted opcode should be the same except for cc_out");
8018 // Add the optional cc_out operand
8019 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8021 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8023 // Any ARM instruction that sets the 's' bit should specify an optional
8024 // "cc_out" operand in the last operand position.
8025 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8026 assert(!NewOpc && "Optional cc_out operand required");
8029 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8030 // since we already have an optional CPSR def.
8031 bool definesCPSR = false;
8032 bool deadCPSR = false;
8033 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8035 const MachineOperand &MO = MI->getOperand(i);
8036 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8040 MI->RemoveOperand(i);
8045 assert(!NewOpc && "Optional cc_out operand required");
8048 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8050 assert(!MI->getOperand(ccOutIdx).getReg() &&
8051 "expect uninitialized optional cc_out operand");
8055 // If this instruction was defined with an optional CPSR def and its dag node
8056 // had a live implicit CPSR def, then activate the optional CPSR def.
8057 MachineOperand &MO = MI->getOperand(ccOutIdx);
8058 MO.setReg(ARM::CPSR);
8062 //===----------------------------------------------------------------------===//
8063 // ARM Optimization Hooks
8064 //===----------------------------------------------------------------------===//
8066 // Helper function that checks if N is a null or all ones constant.
8067 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8068 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8071 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8074 // Return true if N is conditionally 0 or all ones.
8075 // Detects these expressions where cc is an i1 value:
8077 // (select cc 0, y) [AllOnes=0]
8078 // (select cc y, 0) [AllOnes=0]
8079 // (zext cc) [AllOnes=0]
8080 // (sext cc) [AllOnes=0/1]
8081 // (select cc -1, y) [AllOnes=1]
8082 // (select cc y, -1) [AllOnes=1]
8084 // Invert is set when N is the null/all ones constant when CC is false.
8085 // OtherOp is set to the alternative value of N.
8086 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8087 SDValue &CC, bool &Invert,
8089 SelectionDAG &DAG) {
8090 switch (N->getOpcode()) {
8091 default: return false;
8093 CC = N->getOperand(0);
8094 SDValue N1 = N->getOperand(1);
8095 SDValue N2 = N->getOperand(2);
8096 if (isZeroOrAllOnes(N1, AllOnes)) {
8101 if (isZeroOrAllOnes(N2, AllOnes)) {
8108 case ISD::ZERO_EXTEND:
8109 // (zext cc) can never be the all ones value.
8113 case ISD::SIGN_EXTEND: {
8114 EVT VT = N->getValueType(0);
8115 CC = N->getOperand(0);
8116 if (CC.getValueType() != MVT::i1)
8120 // When looking for an AllOnes constant, N is an sext, and the 'other'
8122 OtherOp = DAG.getConstant(0, VT);
8123 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8124 // When looking for a 0 constant, N can be zext or sext.
8125 OtherOp = DAG.getConstant(1, VT);
8127 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
8133 // Combine a constant select operand into its use:
8135 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8136 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8137 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8138 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8139 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8141 // The transform is rejected if the select doesn't have a constant operand that
8142 // is null, or all ones when AllOnes is set.
8144 // Also recognize sext/zext from i1:
8146 // (add (zext cc), x) -> (select cc (add x, 1), x)
8147 // (add (sext cc), x) -> (select cc (add x, -1), x)
8149 // These transformations eventually create predicated instructions.
8151 // @param N The node to transform.
8152 // @param Slct The N operand that is a select.
8153 // @param OtherOp The other N operand (x above).
8154 // @param DCI Context.
8155 // @param AllOnes Require the select constant to be all ones instead of null.
8156 // @returns The new node, or SDValue() on failure.
8158 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8159 TargetLowering::DAGCombinerInfo &DCI,
8160 bool AllOnes = false) {
8161 SelectionDAG &DAG = DCI.DAG;
8162 EVT VT = N->getValueType(0);
8163 SDValue NonConstantVal;
8166 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8167 NonConstantVal, DAG))
8170 // Slct is now know to be the desired identity constant when CC is true.
8171 SDValue TrueVal = OtherOp;
8172 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8173 OtherOp, NonConstantVal);
8174 // Unless SwapSelectOps says CC should be false.
8176 std::swap(TrueVal, FalseVal);
8178 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8179 CCOp, TrueVal, FalseVal);
8182 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8184 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8185 TargetLowering::DAGCombinerInfo &DCI) {
8186 SDValue N0 = N->getOperand(0);
8187 SDValue N1 = N->getOperand(1);
8188 if (N0.getNode()->hasOneUse()) {
8189 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8190 if (Result.getNode())
8193 if (N1.getNode()->hasOneUse()) {
8194 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8195 if (Result.getNode())
8201 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8202 // (only after legalization).
8203 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8204 TargetLowering::DAGCombinerInfo &DCI,
8205 const ARMSubtarget *Subtarget) {
8207 // Only perform optimization if after legalize, and if NEON is available. We
8208 // also expected both operands to be BUILD_VECTORs.
8209 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8210 || N0.getOpcode() != ISD::BUILD_VECTOR
8211 || N1.getOpcode() != ISD::BUILD_VECTOR)
8214 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8215 EVT VT = N->getValueType(0);
8216 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8219 // Check that the vector operands are of the right form.
8220 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8221 // operands, where N is the size of the formed vector.
8222 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8223 // index such that we have a pair wise add pattern.
8225 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8226 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8228 SDValue Vec = N0->getOperand(0)->getOperand(0);
8229 SDNode *V = Vec.getNode();
8230 unsigned nextIndex = 0;
8232 // For each operands to the ADD which are BUILD_VECTORs,
8233 // check to see if each of their operands are an EXTRACT_VECTOR with
8234 // the same vector and appropriate index.
8235 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8236 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8237 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8239 SDValue ExtVec0 = N0->getOperand(i);
8240 SDValue ExtVec1 = N1->getOperand(i);
8242 // First operand is the vector, verify its the same.
8243 if (V != ExtVec0->getOperand(0).getNode() ||
8244 V != ExtVec1->getOperand(0).getNode())
8247 // Second is the constant, verify its correct.
8248 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8249 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8251 // For the constant, we want to see all the even or all the odd.
8252 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8253 || C1->getZExtValue() != nextIndex+1)
8262 // Create VPADDL node.
8263 SelectionDAG &DAG = DCI.DAG;
8264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8266 // Build operand list.
8267 SmallVector<SDValue, 8> Ops;
8268 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8269 TLI.getPointerTy()));
8271 // Input is the vector.
8274 // Get widened type and narrowed type.
8276 unsigned numElem = VT.getVectorNumElements();
8277 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8278 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8279 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8280 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8282 llvm_unreachable("Invalid vector element type for padd optimization.");
8285 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8286 widenType, &Ops[0], Ops.size());
8287 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8290 static SDValue findMUL_LOHI(SDValue V) {
8291 if (V->getOpcode() == ISD::UMUL_LOHI ||
8292 V->getOpcode() == ISD::SMUL_LOHI)
8297 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8298 TargetLowering::DAGCombinerInfo &DCI,
8299 const ARMSubtarget *Subtarget) {
8301 if (Subtarget->isThumb1Only()) return SDValue();
8303 // Only perform the checks after legalize when the pattern is available.
8304 if (DCI.isBeforeLegalize()) return SDValue();
8306 // Look for multiply add opportunities.
8307 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8308 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8309 // a glue link from the first add to the second add.
8310 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8311 // a S/UMLAL instruction.
8314 // \ / \ [no multiline comment]
8320 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8321 SDValue AddcOp0 = AddcNode->getOperand(0);
8322 SDValue AddcOp1 = AddcNode->getOperand(1);
8324 // Check if the two operands are from the same mul_lohi node.
8325 if (AddcOp0.getNode() == AddcOp1.getNode())
8328 assert(AddcNode->getNumValues() == 2 &&
8329 AddcNode->getValueType(0) == MVT::i32 &&
8330 "Expect ADDC with two result values. First: i32");
8332 // Check that we have a glued ADDC node.
8333 if (AddcNode->getValueType(1) != MVT::Glue)
8336 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8337 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8338 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8339 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8340 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8343 // Look for the glued ADDE.
8344 SDNode* AddeNode = AddcNode->getGluedUser();
8345 if (AddeNode == NULL)
8348 // Make sure it is really an ADDE.
8349 if (AddeNode->getOpcode() != ISD::ADDE)
8352 assert(AddeNode->getNumOperands() == 3 &&
8353 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8354 "ADDE node has the wrong inputs");
8356 // Check for the triangle shape.
8357 SDValue AddeOp0 = AddeNode->getOperand(0);
8358 SDValue AddeOp1 = AddeNode->getOperand(1);
8360 // Make sure that the ADDE operands are not coming from the same node.
8361 if (AddeOp0.getNode() == AddeOp1.getNode())
8364 // Find the MUL_LOHI node walking up ADDE's operands.
8365 bool IsLeftOperandMUL = false;
8366 SDValue MULOp = findMUL_LOHI(AddeOp0);
8367 if (MULOp == SDValue())
8368 MULOp = findMUL_LOHI(AddeOp1);
8370 IsLeftOperandMUL = true;
8371 if (MULOp == SDValue())
8374 // Figure out the right opcode.
8375 unsigned Opc = MULOp->getOpcode();
8376 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8378 // Figure out the high and low input values to the MLAL node.
8379 SDValue* HiMul = &MULOp;
8380 SDValue* HiAdd = NULL;
8381 SDValue* LoMul = NULL;
8382 SDValue* LowAdd = NULL;
8384 if (IsLeftOperandMUL)
8390 if (AddcOp0->getOpcode() == Opc) {
8394 if (AddcOp1->getOpcode() == Opc) {
8402 if (LoMul->getNode() != HiMul->getNode())
8405 // Create the merged node.
8406 SelectionDAG &DAG = DCI.DAG;
8408 // Build operand list.
8409 SmallVector<SDValue, 8> Ops;
8410 Ops.push_back(LoMul->getOperand(0));
8411 Ops.push_back(LoMul->getOperand(1));
8412 Ops.push_back(*LowAdd);
8413 Ops.push_back(*HiAdd);
8415 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8416 DAG.getVTList(MVT::i32, MVT::i32),
8417 &Ops[0], Ops.size());
8419 // Replace the ADDs' nodes uses by the MLA node's values.
8420 SDValue HiMLALResult(MLALNode.getNode(), 1);
8421 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8423 SDValue LoMLALResult(MLALNode.getNode(), 0);
8424 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8426 // Return original node to notify the driver to stop replacing.
8427 SDValue resNode(AddcNode, 0);
8431 /// PerformADDCCombine - Target-specific dag combine transform from
8432 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8433 static SDValue PerformADDCCombine(SDNode *N,
8434 TargetLowering::DAGCombinerInfo &DCI,
8435 const ARMSubtarget *Subtarget) {
8437 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8441 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8442 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8443 /// called with the default operands, and if that fails, with commuted
8445 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8446 TargetLowering::DAGCombinerInfo &DCI,
8447 const ARMSubtarget *Subtarget){
8449 // Attempt to create vpaddl for this add.
8450 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8451 if (Result.getNode())
8454 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8455 if (N0.getNode()->hasOneUse()) {
8456 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8457 if (Result.getNode()) return Result;
8462 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8464 static SDValue PerformADDCombine(SDNode *N,
8465 TargetLowering::DAGCombinerInfo &DCI,
8466 const ARMSubtarget *Subtarget) {
8467 SDValue N0 = N->getOperand(0);
8468 SDValue N1 = N->getOperand(1);
8470 // First try with the default operand order.
8471 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8472 if (Result.getNode())
8475 // If that didn't work, try again with the operands commuted.
8476 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8479 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8481 static SDValue PerformSUBCombine(SDNode *N,
8482 TargetLowering::DAGCombinerInfo &DCI) {
8483 SDValue N0 = N->getOperand(0);
8484 SDValue N1 = N->getOperand(1);
8486 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8487 if (N1.getNode()->hasOneUse()) {
8488 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8489 if (Result.getNode()) return Result;
8495 /// PerformVMULCombine
8496 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8497 /// special multiplier accumulator forwarding.
8503 // However, for (A + B) * (A + B),
8510 static SDValue PerformVMULCombine(SDNode *N,
8511 TargetLowering::DAGCombinerInfo &DCI,
8512 const ARMSubtarget *Subtarget) {
8513 if (!Subtarget->hasVMLxForwarding())
8516 SelectionDAG &DAG = DCI.DAG;
8517 SDValue N0 = N->getOperand(0);
8518 SDValue N1 = N->getOperand(1);
8519 unsigned Opcode = N0.getOpcode();
8520 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8521 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8522 Opcode = N1.getOpcode();
8523 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8524 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8532 EVT VT = N->getValueType(0);
8534 SDValue N00 = N0->getOperand(0);
8535 SDValue N01 = N0->getOperand(1);
8536 return DAG.getNode(Opcode, DL, VT,
8537 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8538 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8541 static SDValue PerformMULCombine(SDNode *N,
8542 TargetLowering::DAGCombinerInfo &DCI,
8543 const ARMSubtarget *Subtarget) {
8544 SelectionDAG &DAG = DCI.DAG;
8546 if (Subtarget->isThumb1Only())
8549 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8552 EVT VT = N->getValueType(0);
8553 if (VT.is64BitVector() || VT.is128BitVector())
8554 return PerformVMULCombine(N, DCI, Subtarget);
8558 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8562 int64_t MulAmt = C->getSExtValue();
8563 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8565 ShiftAmt = ShiftAmt & (32 - 1);
8566 SDValue V = N->getOperand(0);
8570 MulAmt >>= ShiftAmt;
8573 if (isPowerOf2_32(MulAmt - 1)) {
8574 // (mul x, 2^N + 1) => (add (shl x, N), x)
8575 Res = DAG.getNode(ISD::ADD, DL, VT,
8577 DAG.getNode(ISD::SHL, DL, VT,
8579 DAG.getConstant(Log2_32(MulAmt - 1),
8581 } else if (isPowerOf2_32(MulAmt + 1)) {
8582 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8583 Res = DAG.getNode(ISD::SUB, DL, VT,
8584 DAG.getNode(ISD::SHL, DL, VT,
8586 DAG.getConstant(Log2_32(MulAmt + 1),
8592 uint64_t MulAmtAbs = -MulAmt;
8593 if (isPowerOf2_32(MulAmtAbs + 1)) {
8594 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8595 Res = DAG.getNode(ISD::SUB, DL, VT,
8597 DAG.getNode(ISD::SHL, DL, VT,
8599 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8601 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8602 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8603 Res = DAG.getNode(ISD::ADD, DL, VT,
8605 DAG.getNode(ISD::SHL, DL, VT,
8607 DAG.getConstant(Log2_32(MulAmtAbs-1),
8609 Res = DAG.getNode(ISD::SUB, DL, VT,
8610 DAG.getConstant(0, MVT::i32),Res);
8617 Res = DAG.getNode(ISD::SHL, DL, VT,
8618 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8620 // Do not add new nodes to DAG combiner worklist.
8621 DCI.CombineTo(N, Res, false);
8625 static SDValue PerformANDCombine(SDNode *N,
8626 TargetLowering::DAGCombinerInfo &DCI,
8627 const ARMSubtarget *Subtarget) {
8629 // Attempt to use immediate-form VBIC
8630 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8632 EVT VT = N->getValueType(0);
8633 SelectionDAG &DAG = DCI.DAG;
8635 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8638 APInt SplatBits, SplatUndef;
8639 unsigned SplatBitSize;
8642 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8643 if (SplatBitSize <= 64) {
8645 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8646 SplatUndef.getZExtValue(), SplatBitSize,
8647 DAG, VbicVT, VT.is128BitVector(),
8649 if (Val.getNode()) {
8651 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8652 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8653 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8658 if (!Subtarget->isThumb1Only()) {
8659 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8660 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8661 if (Result.getNode())
8668 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8669 static SDValue PerformORCombine(SDNode *N,
8670 TargetLowering::DAGCombinerInfo &DCI,
8671 const ARMSubtarget *Subtarget) {
8672 // Attempt to use immediate-form VORR
8673 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8675 EVT VT = N->getValueType(0);
8676 SelectionDAG &DAG = DCI.DAG;
8678 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8681 APInt SplatBits, SplatUndef;
8682 unsigned SplatBitSize;
8684 if (BVN && Subtarget->hasNEON() &&
8685 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8686 if (SplatBitSize <= 64) {
8688 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8689 SplatUndef.getZExtValue(), SplatBitSize,
8690 DAG, VorrVT, VT.is128BitVector(),
8692 if (Val.getNode()) {
8694 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8695 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8696 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8701 if (!Subtarget->isThumb1Only()) {
8702 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8703 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8704 if (Result.getNode())
8708 // The code below optimizes (or (and X, Y), Z).
8709 // The AND operand needs to have a single user to make these optimizations
8711 SDValue N0 = N->getOperand(0);
8712 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8714 SDValue N1 = N->getOperand(1);
8716 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8717 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8718 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8720 unsigned SplatBitSize;
8723 APInt SplatBits0, SplatBits1;
8724 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8725 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8726 // Ensure that the second operand of both ands are constants
8727 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8728 HasAnyUndefs) && !HasAnyUndefs) {
8729 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8730 HasAnyUndefs) && !HasAnyUndefs) {
8731 // Ensure that the bit width of the constants are the same and that
8732 // the splat arguments are logical inverses as per the pattern we
8733 // are trying to simplify.
8734 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8735 SplatBits0 == ~SplatBits1) {
8736 // Canonicalize the vector type to make instruction selection
8738 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8739 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8743 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8749 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8752 // BFI is only available on V6T2+
8753 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8757 // 1) or (and A, mask), val => ARMbfi A, val, mask
8758 // iff (val & mask) == val
8760 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8761 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8762 // && mask == ~mask2
8763 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8764 // && ~mask == mask2
8765 // (i.e., copy a bitfield value into another bitfield of the same width)
8770 SDValue N00 = N0.getOperand(0);
8772 // The value and the mask need to be constants so we can verify this is
8773 // actually a bitfield set. If the mask is 0xffff, we can do better
8774 // via a movt instruction, so don't use BFI in that case.
8775 SDValue MaskOp = N0.getOperand(1);
8776 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8779 unsigned Mask = MaskC->getZExtValue();
8783 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8784 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8786 unsigned Val = N1C->getZExtValue();
8787 if ((Val & ~Mask) != Val)
8790 if (ARM::isBitFieldInvertedMask(Mask)) {
8791 Val >>= countTrailingZeros(~Mask);
8793 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8794 DAG.getConstant(Val, MVT::i32),
8795 DAG.getConstant(Mask, MVT::i32));
8797 // Do not add new nodes to DAG combiner worklist.
8798 DCI.CombineTo(N, Res, false);
8801 } else if (N1.getOpcode() == ISD::AND) {
8802 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8803 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8806 unsigned Mask2 = N11C->getZExtValue();
8808 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8810 if (ARM::isBitFieldInvertedMask(Mask) &&
8812 // The pack halfword instruction works better for masks that fit it,
8813 // so use that when it's available.
8814 if (Subtarget->hasT2ExtractPack() &&
8815 (Mask == 0xffff || Mask == 0xffff0000))
8818 unsigned amt = countTrailingZeros(Mask2);
8819 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8820 DAG.getConstant(amt, MVT::i32));
8821 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8822 DAG.getConstant(Mask, MVT::i32));
8823 // Do not add new nodes to DAG combiner worklist.
8824 DCI.CombineTo(N, Res, false);
8826 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8828 // The pack halfword instruction works better for masks that fit it,
8829 // so use that when it's available.
8830 if (Subtarget->hasT2ExtractPack() &&
8831 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8834 unsigned lsb = countTrailingZeros(Mask);
8835 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8836 DAG.getConstant(lsb, MVT::i32));
8837 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8838 DAG.getConstant(Mask2, MVT::i32));
8839 // Do not add new nodes to DAG combiner worklist.
8840 DCI.CombineTo(N, Res, false);
8845 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8846 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8847 ARM::isBitFieldInvertedMask(~Mask)) {
8848 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8849 // where lsb(mask) == #shamt and masked bits of B are known zero.
8850 SDValue ShAmt = N00.getOperand(1);
8851 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8852 unsigned LSB = countTrailingZeros(Mask);
8856 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8857 DAG.getConstant(~Mask, MVT::i32));
8859 // Do not add new nodes to DAG combiner worklist.
8860 DCI.CombineTo(N, Res, false);
8866 static SDValue PerformXORCombine(SDNode *N,
8867 TargetLowering::DAGCombinerInfo &DCI,
8868 const ARMSubtarget *Subtarget) {
8869 EVT VT = N->getValueType(0);
8870 SelectionDAG &DAG = DCI.DAG;
8872 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8875 if (!Subtarget->isThumb1Only()) {
8876 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8877 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8878 if (Result.getNode())
8885 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8886 /// the bits being cleared by the AND are not demanded by the BFI.
8887 static SDValue PerformBFICombine(SDNode *N,
8888 TargetLowering::DAGCombinerInfo &DCI) {
8889 SDValue N1 = N->getOperand(1);
8890 if (N1.getOpcode() == ISD::AND) {
8891 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8894 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8895 unsigned LSB = countTrailingZeros(~InvMask);
8896 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8897 unsigned Mask = (1 << Width)-1;
8898 unsigned Mask2 = N11C->getZExtValue();
8899 if ((Mask & (~Mask2)) == 0)
8900 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8901 N->getOperand(0), N1.getOperand(0),
8907 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8908 /// ARMISD::VMOVRRD.
8909 static SDValue PerformVMOVRRDCombine(SDNode *N,
8910 TargetLowering::DAGCombinerInfo &DCI) {
8911 // vmovrrd(vmovdrr x, y) -> x,y
8912 SDValue InDouble = N->getOperand(0);
8913 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8914 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8916 // vmovrrd(load f64) -> (load i32), (load i32)
8917 SDNode *InNode = InDouble.getNode();
8918 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8919 InNode->getValueType(0) == MVT::f64 &&
8920 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8921 !cast<LoadSDNode>(InNode)->isVolatile()) {
8922 // TODO: Should this be done for non-FrameIndex operands?
8923 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8925 SelectionDAG &DAG = DCI.DAG;
8927 SDValue BasePtr = LD->getBasePtr();
8928 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8929 LD->getPointerInfo(), LD->isVolatile(),
8930 LD->isNonTemporal(), LD->isInvariant(),
8931 LD->getAlignment());
8933 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8934 DAG.getConstant(4, MVT::i32));
8935 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8936 LD->getPointerInfo(), LD->isVolatile(),
8937 LD->isNonTemporal(), LD->isInvariant(),
8938 std::min(4U, LD->getAlignment() / 2));
8940 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8941 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8942 DCI.RemoveFromWorklist(LD);
8950 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8951 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8952 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8953 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8954 SDValue Op0 = N->getOperand(0);
8955 SDValue Op1 = N->getOperand(1);
8956 if (Op0.getOpcode() == ISD::BITCAST)
8957 Op0 = Op0.getOperand(0);
8958 if (Op1.getOpcode() == ISD::BITCAST)
8959 Op1 = Op1.getOperand(0);
8960 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8961 Op0.getNode() == Op1.getNode() &&
8962 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8963 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8964 N->getValueType(0), Op0.getOperand(0));
8968 /// PerformSTORECombine - Target-specific dag combine xforms for
8970 static SDValue PerformSTORECombine(SDNode *N,
8971 TargetLowering::DAGCombinerInfo &DCI) {
8972 StoreSDNode *St = cast<StoreSDNode>(N);
8973 if (St->isVolatile())
8976 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8977 // pack all of the elements in one place. Next, store to memory in fewer
8979 SDValue StVal = St->getValue();
8980 EVT VT = StVal.getValueType();
8981 if (St->isTruncatingStore() && VT.isVector()) {
8982 SelectionDAG &DAG = DCI.DAG;
8983 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8984 EVT StVT = St->getMemoryVT();
8985 unsigned NumElems = VT.getVectorNumElements();
8986 assert(StVT != VT && "Cannot truncate to the same type");
8987 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8988 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8990 // From, To sizes and ElemCount must be pow of two
8991 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8993 // We are going to use the original vector elt for storing.
8994 // Accumulated smaller vector elements must be a multiple of the store size.
8995 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8997 unsigned SizeRatio = FromEltSz / ToEltSz;
8998 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9000 // Create a type on which we perform the shuffle.
9001 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9002 NumElems*SizeRatio);
9003 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9006 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9007 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9008 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
9010 // Can't shuffle using an illegal type.
9011 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9013 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9014 DAG.getUNDEF(WideVec.getValueType()),
9016 // At this point all of the data is stored at the bottom of the
9017 // register. We now need to save it to mem.
9019 // Find the largest store unit
9020 MVT StoreType = MVT::i8;
9021 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
9022 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
9023 MVT Tp = (MVT::SimpleValueType)tp;
9024 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9027 // Didn't find a legal store type.
9028 if (!TLI.isTypeLegal(StoreType))
9031 // Bitcast the original vector into a vector of store-size units
9032 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9033 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9034 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9035 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9036 SmallVector<SDValue, 8> Chains;
9037 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9038 TLI.getPointerTy());
9039 SDValue BasePtr = St->getBasePtr();
9041 // Perform one or more big stores into memory.
9042 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9043 for (unsigned I = 0; I < E; I++) {
9044 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9045 StoreType, ShuffWide,
9046 DAG.getIntPtrConstant(I));
9047 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9048 St->getPointerInfo(), St->isVolatile(),
9049 St->isNonTemporal(), St->getAlignment());
9050 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9052 Chains.push_back(Ch);
9054 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
9058 if (!ISD::isNormalStore(St))
9061 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9062 // ARM stores of arguments in the same cache line.
9063 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9064 StVal.getNode()->hasOneUse()) {
9065 SelectionDAG &DAG = DCI.DAG;
9067 SDValue BasePtr = St->getBasePtr();
9068 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9069 StVal.getNode()->getOperand(0), BasePtr,
9070 St->getPointerInfo(), St->isVolatile(),
9071 St->isNonTemporal(), St->getAlignment());
9073 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9074 DAG.getConstant(4, MVT::i32));
9075 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
9076 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9077 St->isNonTemporal(),
9078 std::min(4U, St->getAlignment() / 2));
9081 if (StVal.getValueType() != MVT::i64 ||
9082 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9085 // Bitcast an i64 store extracted from a vector to f64.
9086 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9087 SelectionDAG &DAG = DCI.DAG;
9089 SDValue IntVec = StVal.getOperand(0);
9090 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9091 IntVec.getValueType().getVectorNumElements());
9092 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9093 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9094 Vec, StVal.getOperand(1));
9096 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9097 // Make the DAGCombiner fold the bitcasts.
9098 DCI.AddToWorklist(Vec.getNode());
9099 DCI.AddToWorklist(ExtElt.getNode());
9100 DCI.AddToWorklist(V.getNode());
9101 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9102 St->getPointerInfo(), St->isVolatile(),
9103 St->isNonTemporal(), St->getAlignment(),
9107 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9108 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9109 /// i64 vector to have f64 elements, since the value can then be loaded
9110 /// directly into a VFP register.
9111 static bool hasNormalLoadOperand(SDNode *N) {
9112 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9113 for (unsigned i = 0; i < NumElts; ++i) {
9114 SDNode *Elt = N->getOperand(i).getNode();
9115 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9121 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9122 /// ISD::BUILD_VECTOR.
9123 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9124 TargetLowering::DAGCombinerInfo &DCI){
9125 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9126 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9127 // into a pair of GPRs, which is fine when the value is used as a scalar,
9128 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9129 SelectionDAG &DAG = DCI.DAG;
9130 if (N->getNumOperands() == 2) {
9131 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9136 // Load i64 elements as f64 values so that type legalization does not split
9137 // them up into i32 values.
9138 EVT VT = N->getValueType(0);
9139 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9142 SmallVector<SDValue, 8> Ops;
9143 unsigned NumElts = VT.getVectorNumElements();
9144 for (unsigned i = 0; i < NumElts; ++i) {
9145 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9147 // Make the DAGCombiner fold the bitcast.
9148 DCI.AddToWorklist(V.getNode());
9150 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9151 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
9152 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9155 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9157 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9158 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9159 // At that time, we may have inserted bitcasts from integer to float.
9160 // If these bitcasts have survived DAGCombine, change the lowering of this
9161 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9162 // force to use floating point types.
9164 // Make sure we can change the type of the vector.
9165 // This is possible iff:
9166 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9167 // 1.1. Vector is used only once.
9168 // 1.2. Use is a bit convert to an integer type.
9169 // 2. The size of its operands are 32-bits (64-bits are not legal).
9170 EVT VT = N->getValueType(0);
9171 EVT EltVT = VT.getVectorElementType();
9173 // Check 1.1. and 2.
9174 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9177 // By construction, the input type must be float.
9178 assert(EltVT == MVT::f32 && "Unexpected type!");
9181 SDNode *Use = *N->use_begin();
9182 if (Use->getOpcode() != ISD::BITCAST ||
9183 Use->getValueType(0).isFloatingPoint())
9186 // Check profitability.
9187 // Model is, if more than half of the relevant operands are bitcast from
9188 // i32, turn the build_vector into a sequence of insert_vector_elt.
9189 // Relevant operands are everything that is not statically
9190 // (i.e., at compile time) bitcasted.
9191 unsigned NumOfBitCastedElts = 0;
9192 unsigned NumElts = VT.getVectorNumElements();
9193 unsigned NumOfRelevantElts = NumElts;
9194 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9195 SDValue Elt = N->getOperand(Idx);
9196 if (Elt->getOpcode() == ISD::BITCAST) {
9197 // Assume only bit cast to i32 will go away.
9198 if (Elt->getOperand(0).getValueType() == MVT::i32)
9199 ++NumOfBitCastedElts;
9200 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9201 // Constants are statically casted, thus do not count them as
9202 // relevant operands.
9203 --NumOfRelevantElts;
9206 // Check if more than half of the elements require a non-free bitcast.
9207 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9210 SelectionDAG &DAG = DCI.DAG;
9211 // Create the new vector type.
9212 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9213 // Check if the type is legal.
9214 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9215 if (!TLI.isTypeLegal(VecVT))
9219 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9220 // => BITCAST INSERT_VECTOR_ELT
9221 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9223 SDValue Vec = DAG.getUNDEF(VecVT);
9225 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9226 SDValue V = N->getOperand(Idx);
9227 if (V.getOpcode() == ISD::UNDEF)
9229 if (V.getOpcode() == ISD::BITCAST &&
9230 V->getOperand(0).getValueType() == MVT::i32)
9231 // Fold obvious case.
9232 V = V.getOperand(0);
9234 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9235 // Make the DAGCombiner fold the bitcasts.
9236 DCI.AddToWorklist(V.getNode());
9238 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9239 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9241 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9242 // Make the DAGCombiner fold the bitcasts.
9243 DCI.AddToWorklist(Vec.getNode());
9247 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9248 /// ISD::INSERT_VECTOR_ELT.
9249 static SDValue PerformInsertEltCombine(SDNode *N,
9250 TargetLowering::DAGCombinerInfo &DCI) {
9251 // Bitcast an i64 load inserted into a vector to f64.
9252 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9253 EVT VT = N->getValueType(0);
9254 SDNode *Elt = N->getOperand(1).getNode();
9255 if (VT.getVectorElementType() != MVT::i64 ||
9256 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9259 SelectionDAG &DAG = DCI.DAG;
9261 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9262 VT.getVectorNumElements());
9263 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9264 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9265 // Make the DAGCombiner fold the bitcasts.
9266 DCI.AddToWorklist(Vec.getNode());
9267 DCI.AddToWorklist(V.getNode());
9268 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9269 Vec, V, N->getOperand(2));
9270 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9273 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9274 /// ISD::VECTOR_SHUFFLE.
9275 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9276 // The LLVM shufflevector instruction does not require the shuffle mask
9277 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9278 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9279 // operands do not match the mask length, they are extended by concatenating
9280 // them with undef vectors. That is probably the right thing for other
9281 // targets, but for NEON it is better to concatenate two double-register
9282 // size vector operands into a single quad-register size vector. Do that
9283 // transformation here:
9284 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9285 // shuffle(concat(v1, v2), undef)
9286 SDValue Op0 = N->getOperand(0);
9287 SDValue Op1 = N->getOperand(1);
9288 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9289 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9290 Op0.getNumOperands() != 2 ||
9291 Op1.getNumOperands() != 2)
9293 SDValue Concat0Op1 = Op0.getOperand(1);
9294 SDValue Concat1Op1 = Op1.getOperand(1);
9295 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9296 Concat1Op1.getOpcode() != ISD::UNDEF)
9298 // Skip the transformation if any of the types are illegal.
9299 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9300 EVT VT = N->getValueType(0);
9301 if (!TLI.isTypeLegal(VT) ||
9302 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9303 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9306 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9307 Op0.getOperand(0), Op1.getOperand(0));
9308 // Translate the shuffle mask.
9309 SmallVector<int, 16> NewMask;
9310 unsigned NumElts = VT.getVectorNumElements();
9311 unsigned HalfElts = NumElts/2;
9312 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9313 for (unsigned n = 0; n < NumElts; ++n) {
9314 int MaskElt = SVN->getMaskElt(n);
9316 if (MaskElt < (int)HalfElts)
9318 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9319 NewElt = HalfElts + MaskElt - NumElts;
9320 NewMask.push_back(NewElt);
9322 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9323 DAG.getUNDEF(VT), NewMask.data());
9326 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9327 /// NEON load/store intrinsics to merge base address updates.
9328 static SDValue CombineBaseUpdate(SDNode *N,
9329 TargetLowering::DAGCombinerInfo &DCI) {
9330 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9333 SelectionDAG &DAG = DCI.DAG;
9334 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9335 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9336 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9337 SDValue Addr = N->getOperand(AddrOpIdx);
9339 // Search for a use of the address operand that is an increment.
9340 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9341 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9343 if (User->getOpcode() != ISD::ADD ||
9344 UI.getUse().getResNo() != Addr.getResNo())
9347 // Check that the add is independent of the load/store. Otherwise, folding
9348 // it would create a cycle.
9349 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9352 // Find the new opcode for the updating load/store.
9354 bool isLaneOp = false;
9355 unsigned NewOpc = 0;
9356 unsigned NumVecs = 0;
9358 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9360 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9361 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9363 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9365 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9367 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9369 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9370 NumVecs = 2; isLaneOp = true; break;
9371 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9372 NumVecs = 3; isLaneOp = true; break;
9373 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9374 NumVecs = 4; isLaneOp = true; break;
9375 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9376 NumVecs = 1; isLoad = false; break;
9377 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9378 NumVecs = 2; isLoad = false; break;
9379 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9380 NumVecs = 3; isLoad = false; break;
9381 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9382 NumVecs = 4; isLoad = false; break;
9383 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9384 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9385 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9386 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9387 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9388 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9392 switch (N->getOpcode()) {
9393 default: llvm_unreachable("unexpected opcode for Neon base update");
9394 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9395 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9396 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9400 // Find the size of memory referenced by the load/store.
9403 VecTy = N->getValueType(0);
9405 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9406 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9408 NumBytes /= VecTy.getVectorNumElements();
9410 // If the increment is a constant, it must match the memory ref size.
9411 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9412 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9413 uint64_t IncVal = CInc->getZExtValue();
9414 if (IncVal != NumBytes)
9416 } else if (NumBytes >= 3 * 16) {
9417 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9418 // separate instructions that make it harder to use a non-constant update.
9422 // Create the new updating load/store node.
9424 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9426 for (n = 0; n < NumResultVecs; ++n)
9428 Tys[n++] = MVT::i32;
9429 Tys[n] = MVT::Other;
9430 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9431 SmallVector<SDValue, 8> Ops;
9432 Ops.push_back(N->getOperand(0)); // incoming chain
9433 Ops.push_back(N->getOperand(AddrOpIdx));
9435 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9436 Ops.push_back(N->getOperand(i));
9438 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9439 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9440 Ops.data(), Ops.size(),
9441 MemInt->getMemoryVT(),
9442 MemInt->getMemOperand());
9445 std::vector<SDValue> NewResults;
9446 for (unsigned i = 0; i < NumResultVecs; ++i) {
9447 NewResults.push_back(SDValue(UpdN.getNode(), i));
9449 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9450 DCI.CombineTo(N, NewResults);
9451 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9458 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9459 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9460 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9462 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9463 SelectionDAG &DAG = DCI.DAG;
9464 EVT VT = N->getValueType(0);
9465 // vldN-dup instructions only support 64-bit vectors for N > 1.
9466 if (!VT.is64BitVector())
9469 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9470 SDNode *VLD = N->getOperand(0).getNode();
9471 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9473 unsigned NumVecs = 0;
9474 unsigned NewOpc = 0;
9475 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9476 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9478 NewOpc = ARMISD::VLD2DUP;
9479 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9481 NewOpc = ARMISD::VLD3DUP;
9482 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9484 NewOpc = ARMISD::VLD4DUP;
9489 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9490 // numbers match the load.
9491 unsigned VLDLaneNo =
9492 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9493 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9495 // Ignore uses of the chain result.
9496 if (UI.getUse().getResNo() == NumVecs)
9499 if (User->getOpcode() != ARMISD::VDUPLANE ||
9500 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9504 // Create the vldN-dup node.
9507 for (n = 0; n < NumVecs; ++n)
9509 Tys[n] = MVT::Other;
9510 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9511 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9512 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9513 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9514 Ops, 2, VLDMemInt->getMemoryVT(),
9515 VLDMemInt->getMemOperand());
9518 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9520 unsigned ResNo = UI.getUse().getResNo();
9521 // Ignore uses of the chain result.
9522 if (ResNo == NumVecs)
9525 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9528 // Now the vldN-lane intrinsic is dead except for its chain result.
9529 // Update uses of the chain.
9530 std::vector<SDValue> VLDDupResults;
9531 for (unsigned n = 0; n < NumVecs; ++n)
9532 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9533 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9534 DCI.CombineTo(VLD, VLDDupResults);
9539 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9540 /// ARMISD::VDUPLANE.
9541 static SDValue PerformVDUPLANECombine(SDNode *N,
9542 TargetLowering::DAGCombinerInfo &DCI) {
9543 SDValue Op = N->getOperand(0);
9545 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9546 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9547 if (CombineVLDDUP(N, DCI))
9548 return SDValue(N, 0);
9550 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9551 // redundant. Ignore bit_converts for now; element sizes are checked below.
9552 while (Op.getOpcode() == ISD::BITCAST)
9553 Op = Op.getOperand(0);
9554 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9557 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9558 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9559 // The canonical VMOV for a zero vector uses a 32-bit element size.
9560 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9562 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9564 EVT VT = N->getValueType(0);
9565 if (EltSize > VT.getVectorElementType().getSizeInBits())
9568 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9571 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9572 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9573 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9577 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9579 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9584 APFloat APF = C->getValueAPF();
9585 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9586 != APFloat::opOK || !isExact)
9589 c0 = (I == 0) ? cN : c0;
9590 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9597 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9598 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9599 /// when the VMUL has a constant operand that is a power of 2.
9601 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9602 /// vmul.f32 d16, d17, d16
9603 /// vcvt.s32.f32 d16, d16
9605 /// vcvt.s32.f32 d16, d16, #3
9606 static SDValue PerformVCVTCombine(SDNode *N,
9607 TargetLowering::DAGCombinerInfo &DCI,
9608 const ARMSubtarget *Subtarget) {
9609 SelectionDAG &DAG = DCI.DAG;
9610 SDValue Op = N->getOperand(0);
9612 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9613 Op.getOpcode() != ISD::FMUL)
9617 SDValue N0 = Op->getOperand(0);
9618 SDValue ConstVec = Op->getOperand(1);
9619 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9621 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9622 !isConstVecPow2(ConstVec, isSigned, C))
9625 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9626 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9627 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9628 // These instructions only exist converting from f32 to i32. We can handle
9629 // smaller integers by generating an extra truncate, but larger ones would
9634 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9635 Intrinsic::arm_neon_vcvtfp2fxu;
9636 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9637 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9638 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9639 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9640 DAG.getConstant(Log2_64(C), MVT::i32));
9642 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9643 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9648 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9649 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9650 /// when the VDIV has a constant operand that is a power of 2.
9652 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9653 /// vcvt.f32.s32 d16, d16
9654 /// vdiv.f32 d16, d17, d16
9656 /// vcvt.f32.s32 d16, d16, #3
9657 static SDValue PerformVDIVCombine(SDNode *N,
9658 TargetLowering::DAGCombinerInfo &DCI,
9659 const ARMSubtarget *Subtarget) {
9660 SelectionDAG &DAG = DCI.DAG;
9661 SDValue Op = N->getOperand(0);
9662 unsigned OpOpcode = Op.getNode()->getOpcode();
9664 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9665 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9669 SDValue ConstVec = N->getOperand(1);
9670 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9672 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9673 !isConstVecPow2(ConstVec, isSigned, C))
9676 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9677 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9678 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9679 // These instructions only exist converting from i32 to f32. We can handle
9680 // smaller integers by generating an extra extend, but larger ones would
9685 SDValue ConvInput = Op.getOperand(0);
9686 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9687 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9688 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9689 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9692 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9693 Intrinsic::arm_neon_vcvtfxu2fp;
9694 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9696 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9697 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9700 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9701 /// operand of a vector shift operation, where all the elements of the
9702 /// build_vector must have the same constant integer value.
9703 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9704 // Ignore bit_converts.
9705 while (Op.getOpcode() == ISD::BITCAST)
9706 Op = Op.getOperand(0);
9707 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9708 APInt SplatBits, SplatUndef;
9709 unsigned SplatBitSize;
9711 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9712 HasAnyUndefs, ElementBits) ||
9713 SplatBitSize > ElementBits)
9715 Cnt = SplatBits.getSExtValue();
9719 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9720 /// operand of a vector shift left operation. That value must be in the range:
9721 /// 0 <= Value < ElementBits for a left shift; or
9722 /// 0 <= Value <= ElementBits for a long left shift.
9723 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9724 assert(VT.isVector() && "vector shift count is not a vector type");
9725 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9726 if (! getVShiftImm(Op, ElementBits, Cnt))
9728 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9731 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9732 /// operand of a vector shift right operation. For a shift opcode, the value
9733 /// is positive, but for an intrinsic the value count must be negative. The
9734 /// absolute value must be in the range:
9735 /// 1 <= |Value| <= ElementBits for a right shift; or
9736 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9737 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9739 assert(VT.isVector() && "vector shift count is not a vector type");
9740 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9741 if (! getVShiftImm(Op, ElementBits, Cnt))
9745 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9748 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9749 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9750 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9753 // Don't do anything for most intrinsics.
9756 // Vector shifts: check for immediate versions and lower them.
9757 // Note: This is done during DAG combining instead of DAG legalizing because
9758 // the build_vectors for 64-bit vector element shift counts are generally
9759 // not legal, and it is hard to see their values after they get legalized to
9760 // loads from a constant pool.
9761 case Intrinsic::arm_neon_vshifts:
9762 case Intrinsic::arm_neon_vshiftu:
9763 case Intrinsic::arm_neon_vshiftls:
9764 case Intrinsic::arm_neon_vshiftlu:
9765 case Intrinsic::arm_neon_vshiftn:
9766 case Intrinsic::arm_neon_vrshifts:
9767 case Intrinsic::arm_neon_vrshiftu:
9768 case Intrinsic::arm_neon_vrshiftn:
9769 case Intrinsic::arm_neon_vqshifts:
9770 case Intrinsic::arm_neon_vqshiftu:
9771 case Intrinsic::arm_neon_vqshiftsu:
9772 case Intrinsic::arm_neon_vqshiftns:
9773 case Intrinsic::arm_neon_vqshiftnu:
9774 case Intrinsic::arm_neon_vqshiftnsu:
9775 case Intrinsic::arm_neon_vqrshiftns:
9776 case Intrinsic::arm_neon_vqrshiftnu:
9777 case Intrinsic::arm_neon_vqrshiftnsu: {
9778 EVT VT = N->getOperand(1).getValueType();
9780 unsigned VShiftOpc = 0;
9783 case Intrinsic::arm_neon_vshifts:
9784 case Intrinsic::arm_neon_vshiftu:
9785 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9786 VShiftOpc = ARMISD::VSHL;
9789 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9790 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9791 ARMISD::VSHRs : ARMISD::VSHRu);
9796 case Intrinsic::arm_neon_vshiftls:
9797 case Intrinsic::arm_neon_vshiftlu:
9798 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9800 llvm_unreachable("invalid shift count for vshll intrinsic");
9802 case Intrinsic::arm_neon_vrshifts:
9803 case Intrinsic::arm_neon_vrshiftu:
9804 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9808 case Intrinsic::arm_neon_vqshifts:
9809 case Intrinsic::arm_neon_vqshiftu:
9810 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9814 case Intrinsic::arm_neon_vqshiftsu:
9815 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9817 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9819 case Intrinsic::arm_neon_vshiftn:
9820 case Intrinsic::arm_neon_vrshiftn:
9821 case Intrinsic::arm_neon_vqshiftns:
9822 case Intrinsic::arm_neon_vqshiftnu:
9823 case Intrinsic::arm_neon_vqshiftnsu:
9824 case Intrinsic::arm_neon_vqrshiftns:
9825 case Intrinsic::arm_neon_vqrshiftnu:
9826 case Intrinsic::arm_neon_vqrshiftnsu:
9827 // Narrowing shifts require an immediate right shift.
9828 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9830 llvm_unreachable("invalid shift count for narrowing vector shift "
9834 llvm_unreachable("unhandled vector shift");
9838 case Intrinsic::arm_neon_vshifts:
9839 case Intrinsic::arm_neon_vshiftu:
9840 // Opcode already set above.
9842 case Intrinsic::arm_neon_vshiftls:
9843 case Intrinsic::arm_neon_vshiftlu:
9844 if (Cnt == VT.getVectorElementType().getSizeInBits())
9845 VShiftOpc = ARMISD::VSHLLi;
9847 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9848 ARMISD::VSHLLs : ARMISD::VSHLLu);
9850 case Intrinsic::arm_neon_vshiftn:
9851 VShiftOpc = ARMISD::VSHRN; break;
9852 case Intrinsic::arm_neon_vrshifts:
9853 VShiftOpc = ARMISD::VRSHRs; break;
9854 case Intrinsic::arm_neon_vrshiftu:
9855 VShiftOpc = ARMISD::VRSHRu; break;
9856 case Intrinsic::arm_neon_vrshiftn:
9857 VShiftOpc = ARMISD::VRSHRN; break;
9858 case Intrinsic::arm_neon_vqshifts:
9859 VShiftOpc = ARMISD::VQSHLs; break;
9860 case Intrinsic::arm_neon_vqshiftu:
9861 VShiftOpc = ARMISD::VQSHLu; break;
9862 case Intrinsic::arm_neon_vqshiftsu:
9863 VShiftOpc = ARMISD::VQSHLsu; break;
9864 case Intrinsic::arm_neon_vqshiftns:
9865 VShiftOpc = ARMISD::VQSHRNs; break;
9866 case Intrinsic::arm_neon_vqshiftnu:
9867 VShiftOpc = ARMISD::VQSHRNu; break;
9868 case Intrinsic::arm_neon_vqshiftnsu:
9869 VShiftOpc = ARMISD::VQSHRNsu; break;
9870 case Intrinsic::arm_neon_vqrshiftns:
9871 VShiftOpc = ARMISD::VQRSHRNs; break;
9872 case Intrinsic::arm_neon_vqrshiftnu:
9873 VShiftOpc = ARMISD::VQRSHRNu; break;
9874 case Intrinsic::arm_neon_vqrshiftnsu:
9875 VShiftOpc = ARMISD::VQRSHRNsu; break;
9878 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9879 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9882 case Intrinsic::arm_neon_vshiftins: {
9883 EVT VT = N->getOperand(1).getValueType();
9885 unsigned VShiftOpc = 0;
9887 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9888 VShiftOpc = ARMISD::VSLI;
9889 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9890 VShiftOpc = ARMISD::VSRI;
9892 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9895 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9896 N->getOperand(1), N->getOperand(2),
9897 DAG.getConstant(Cnt, MVT::i32));
9900 case Intrinsic::arm_neon_vqrshifts:
9901 case Intrinsic::arm_neon_vqrshiftu:
9902 // No immediate versions of these to check for.
9909 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9910 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9911 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9912 /// vector element shift counts are generally not legal, and it is hard to see
9913 /// their values after they get legalized to loads from a constant pool.
9914 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9915 const ARMSubtarget *ST) {
9916 EVT VT = N->getValueType(0);
9917 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9918 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9919 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9920 SDValue N1 = N->getOperand(1);
9921 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9922 SDValue N0 = N->getOperand(0);
9923 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9924 DAG.MaskedValueIsZero(N0.getOperand(0),
9925 APInt::getHighBitsSet(32, 16)))
9926 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9930 // Nothing to be done for scalar shifts.
9931 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9932 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9935 assert(ST->hasNEON() && "unexpected vector shift");
9938 switch (N->getOpcode()) {
9939 default: llvm_unreachable("unexpected shift opcode");
9942 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9943 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9944 DAG.getConstant(Cnt, MVT::i32));
9949 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9950 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9951 ARMISD::VSHRs : ARMISD::VSHRu);
9952 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9953 DAG.getConstant(Cnt, MVT::i32));
9959 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9960 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9961 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9962 const ARMSubtarget *ST) {
9963 SDValue N0 = N->getOperand(0);
9965 // Check for sign- and zero-extensions of vector extract operations of 8-
9966 // and 16-bit vector elements. NEON supports these directly. They are
9967 // handled during DAG combining because type legalization will promote them
9968 // to 32-bit types and it is messy to recognize the operations after that.
9969 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9970 SDValue Vec = N0.getOperand(0);
9971 SDValue Lane = N0.getOperand(1);
9972 EVT VT = N->getValueType(0);
9973 EVT EltVT = N0.getValueType();
9974 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9976 if (VT == MVT::i32 &&
9977 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9978 TLI.isTypeLegal(Vec.getValueType()) &&
9979 isa<ConstantSDNode>(Lane)) {
9982 switch (N->getOpcode()) {
9983 default: llvm_unreachable("unexpected opcode");
9984 case ISD::SIGN_EXTEND:
9985 Opc = ARMISD::VGETLANEs;
9987 case ISD::ZERO_EXTEND:
9988 case ISD::ANY_EXTEND:
9989 Opc = ARMISD::VGETLANEu;
9992 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9999 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
10000 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
10001 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
10002 const ARMSubtarget *ST) {
10003 // If the target supports NEON, try to use vmax/vmin instructions for f32
10004 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
10005 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
10006 // a NaN; only do the transformation when it matches that behavior.
10008 // For now only do this when using NEON for FP operations; if using VFP, it
10009 // is not obvious that the benefit outweighs the cost of switching to the
10011 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
10012 N->getValueType(0) != MVT::f32)
10015 SDValue CondLHS = N->getOperand(0);
10016 SDValue CondRHS = N->getOperand(1);
10017 SDValue LHS = N->getOperand(2);
10018 SDValue RHS = N->getOperand(3);
10019 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
10021 unsigned Opcode = 0;
10023 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
10024 IsReversed = false; // x CC y ? x : y
10025 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
10026 IsReversed = true ; // x CC y ? y : x
10040 // If LHS is NaN, an ordered comparison will be false and the result will
10041 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
10042 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10043 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
10044 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10046 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
10047 // will return -0, so vmin can only be used for unsafe math or if one of
10048 // the operands is known to be nonzero.
10049 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
10050 !DAG.getTarget().Options.UnsafeFPMath &&
10051 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10053 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
10062 // If LHS is NaN, an ordered comparison will be false and the result will
10063 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
10064 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10065 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
10066 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10068 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
10069 // will return +0, so vmax can only be used for unsafe math or if one of
10070 // the operands is known to be nonzero.
10071 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
10072 !DAG.getTarget().Options.UnsafeFPMath &&
10073 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10075 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
10081 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
10084 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10086 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10087 SDValue Cmp = N->getOperand(4);
10088 if (Cmp.getOpcode() != ARMISD::CMPZ)
10089 // Only looking at EQ and NE cases.
10092 EVT VT = N->getValueType(0);
10094 SDValue LHS = Cmp.getOperand(0);
10095 SDValue RHS = Cmp.getOperand(1);
10096 SDValue FalseVal = N->getOperand(0);
10097 SDValue TrueVal = N->getOperand(1);
10098 SDValue ARMcc = N->getOperand(2);
10099 ARMCC::CondCodes CC =
10100 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10118 /// FIXME: Turn this into a target neutral optimization?
10120 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10121 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10122 N->getOperand(3), Cmp);
10123 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10125 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10126 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10127 N->getOperand(3), NewCmp);
10130 if (Res.getNode()) {
10131 APInt KnownZero, KnownOne;
10132 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
10133 // Capture demanded bits information that would be otherwise lost.
10134 if (KnownZero == 0xfffffffe)
10135 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10136 DAG.getValueType(MVT::i1));
10137 else if (KnownZero == 0xffffff00)
10138 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10139 DAG.getValueType(MVT::i8));
10140 else if (KnownZero == 0xffff0000)
10141 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10142 DAG.getValueType(MVT::i16));
10148 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10149 DAGCombinerInfo &DCI) const {
10150 switch (N->getOpcode()) {
10152 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10153 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10154 case ISD::SUB: return PerformSUBCombine(N, DCI);
10155 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10156 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10157 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10158 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10159 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10160 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
10161 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10162 case ISD::STORE: return PerformSTORECombine(N, DCI);
10163 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
10164 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10165 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10166 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10167 case ISD::FP_TO_SINT:
10168 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10169 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10170 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10173 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10174 case ISD::SIGN_EXTEND:
10175 case ISD::ZERO_EXTEND:
10176 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10177 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10178 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10179 case ARMISD::VLD2DUP:
10180 case ARMISD::VLD3DUP:
10181 case ARMISD::VLD4DUP:
10182 return CombineBaseUpdate(N, DCI);
10183 case ARMISD::BUILD_VECTOR:
10184 return PerformARMBUILD_VECTORCombine(N, DCI);
10185 case ISD::INTRINSIC_VOID:
10186 case ISD::INTRINSIC_W_CHAIN:
10187 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10188 case Intrinsic::arm_neon_vld1:
10189 case Intrinsic::arm_neon_vld2:
10190 case Intrinsic::arm_neon_vld3:
10191 case Intrinsic::arm_neon_vld4:
10192 case Intrinsic::arm_neon_vld2lane:
10193 case Intrinsic::arm_neon_vld3lane:
10194 case Intrinsic::arm_neon_vld4lane:
10195 case Intrinsic::arm_neon_vst1:
10196 case Intrinsic::arm_neon_vst2:
10197 case Intrinsic::arm_neon_vst3:
10198 case Intrinsic::arm_neon_vst4:
10199 case Intrinsic::arm_neon_vst2lane:
10200 case Intrinsic::arm_neon_vst3lane:
10201 case Intrinsic::arm_neon_vst4lane:
10202 return CombineBaseUpdate(N, DCI);
10210 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10212 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10215 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
10216 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10217 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10219 switch (VT.getSimpleVT().SimpleTy) {
10225 // Unaligned access can use (for example) LRDB, LRDH, LDR
10226 if (AllowsUnaligned) {
10228 *Fast = Subtarget->hasV7Ops();
10235 // For any little-endian targets with neon, we can support unaligned ld/st
10236 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10237 // A big-endian target may also explictly support unaligned accesses
10238 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10248 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10249 unsigned AlignCheck) {
10250 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10251 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10254 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10255 unsigned DstAlign, unsigned SrcAlign,
10256 bool IsMemset, bool ZeroMemset,
10258 MachineFunction &MF) const {
10259 const Function *F = MF.getFunction();
10261 // See if we can use NEON instructions for this...
10262 if ((!IsMemset || ZeroMemset) &&
10263 Subtarget->hasNEON() &&
10264 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10265 Attribute::NoImplicitFloat)) {
10268 (memOpAlign(SrcAlign, DstAlign, 16) ||
10269 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
10271 } else if (Size >= 8 &&
10272 (memOpAlign(SrcAlign, DstAlign, 8) ||
10273 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
10278 // Lowering to i32/i16 if the size permits.
10281 else if (Size >= 2)
10284 // Let the target-independent logic figure it out.
10288 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10289 if (Val.getOpcode() != ISD::LOAD)
10292 EVT VT1 = Val.getValueType();
10293 if (!VT1.isSimple() || !VT1.isInteger() ||
10294 !VT2.isSimple() || !VT2.isInteger())
10297 switch (VT1.getSimpleVT().SimpleTy) {
10302 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10309 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10310 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10313 if (!isTypeLegal(EVT::getEVT(Ty1)))
10316 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10318 // Assuming the caller doesn't have a zeroext or signext return parameter,
10319 // truncation all the way down to i1 is valid.
10324 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10328 unsigned Scale = 1;
10329 switch (VT.getSimpleVT().SimpleTy) {
10330 default: return false;
10345 if ((V & (Scale - 1)) != 0)
10348 return V == (V & ((1LL << 5) - 1));
10351 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10352 const ARMSubtarget *Subtarget) {
10353 bool isNeg = false;
10359 switch (VT.getSimpleVT().SimpleTy) {
10360 default: return false;
10365 // + imm12 or - imm8
10367 return V == (V & ((1LL << 8) - 1));
10368 return V == (V & ((1LL << 12) - 1));
10371 // Same as ARM mode. FIXME: NEON?
10372 if (!Subtarget->hasVFP2())
10377 return V == (V & ((1LL << 8) - 1));
10381 /// isLegalAddressImmediate - Return true if the integer value can be used
10382 /// as the offset of the target addressing mode for load / store of the
10384 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10385 const ARMSubtarget *Subtarget) {
10389 if (!VT.isSimple())
10392 if (Subtarget->isThumb1Only())
10393 return isLegalT1AddressImmediate(V, VT);
10394 else if (Subtarget->isThumb2())
10395 return isLegalT2AddressImmediate(V, VT, Subtarget);
10400 switch (VT.getSimpleVT().SimpleTy) {
10401 default: return false;
10406 return V == (V & ((1LL << 12) - 1));
10409 return V == (V & ((1LL << 8) - 1));
10412 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10417 return V == (V & ((1LL << 8) - 1));
10421 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10423 int Scale = AM.Scale;
10427 switch (VT.getSimpleVT().SimpleTy) {
10428 default: return false;
10436 Scale = Scale & ~1;
10437 return Scale == 2 || Scale == 4 || Scale == 8;
10440 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10444 // Note, we allow "void" uses (basically, uses that aren't loads or
10445 // stores), because arm allows folding a scale into many arithmetic
10446 // operations. This should be made more precise and revisited later.
10448 // Allow r << imm, but the imm has to be a multiple of two.
10449 if (Scale & 1) return false;
10450 return isPowerOf2_32(Scale);
10454 /// isLegalAddressingMode - Return true if the addressing mode represented
10455 /// by AM is legal for this target, for a load/store of the specified type.
10456 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10458 EVT VT = getValueType(Ty, true);
10459 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10462 // Can never fold addr of global into load/store.
10466 switch (AM.Scale) {
10467 case 0: // no scale reg, must be "r+i" or "r", or "i".
10470 if (Subtarget->isThumb1Only())
10474 // ARM doesn't support any R+R*scale+imm addr modes.
10478 if (!VT.isSimple())
10481 if (Subtarget->isThumb2())
10482 return isLegalT2ScaledAddressingMode(AM, VT);
10484 int Scale = AM.Scale;
10485 switch (VT.getSimpleVT().SimpleTy) {
10486 default: return false;
10490 if (Scale < 0) Scale = -Scale;
10494 return isPowerOf2_32(Scale & ~1);
10498 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10503 // Note, we allow "void" uses (basically, uses that aren't loads or
10504 // stores), because arm allows folding a scale into many arithmetic
10505 // operations. This should be made more precise and revisited later.
10507 // Allow r << imm, but the imm has to be a multiple of two.
10508 if (Scale & 1) return false;
10509 return isPowerOf2_32(Scale);
10515 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10516 /// icmp immediate, that is the target has icmp instructions which can compare
10517 /// a register against the immediate without having to materialize the
10518 /// immediate into a register.
10519 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10520 // Thumb2 and ARM modes can use cmn for negative immediates.
10521 if (!Subtarget->isThumb())
10522 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10523 if (Subtarget->isThumb2())
10524 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10525 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10526 return Imm >= 0 && Imm <= 255;
10529 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10530 /// *or sub* immediate, that is the target has add or sub instructions which can
10531 /// add a register with the immediate without having to materialize the
10532 /// immediate into a register.
10533 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10534 // Same encoding for add/sub, just flip the sign.
10535 int64_t AbsImm = llvm::abs64(Imm);
10536 if (!Subtarget->isThumb())
10537 return ARM_AM::getSOImmVal(AbsImm) != -1;
10538 if (Subtarget->isThumb2())
10539 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10540 // Thumb1 only has 8-bit unsigned immediate.
10541 return AbsImm >= 0 && AbsImm <= 255;
10544 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10545 bool isSEXTLoad, SDValue &Base,
10546 SDValue &Offset, bool &isInc,
10547 SelectionDAG &DAG) {
10548 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10551 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10552 // AddressingMode 3
10553 Base = Ptr->getOperand(0);
10554 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10555 int RHSC = (int)RHS->getZExtValue();
10556 if (RHSC < 0 && RHSC > -256) {
10557 assert(Ptr->getOpcode() == ISD::ADD);
10559 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10563 isInc = (Ptr->getOpcode() == ISD::ADD);
10564 Offset = Ptr->getOperand(1);
10566 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10567 // AddressingMode 2
10568 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10569 int RHSC = (int)RHS->getZExtValue();
10570 if (RHSC < 0 && RHSC > -0x1000) {
10571 assert(Ptr->getOpcode() == ISD::ADD);
10573 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10574 Base = Ptr->getOperand(0);
10579 if (Ptr->getOpcode() == ISD::ADD) {
10581 ARM_AM::ShiftOpc ShOpcVal=
10582 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10583 if (ShOpcVal != ARM_AM::no_shift) {
10584 Base = Ptr->getOperand(1);
10585 Offset = Ptr->getOperand(0);
10587 Base = Ptr->getOperand(0);
10588 Offset = Ptr->getOperand(1);
10593 isInc = (Ptr->getOpcode() == ISD::ADD);
10594 Base = Ptr->getOperand(0);
10595 Offset = Ptr->getOperand(1);
10599 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10603 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10604 bool isSEXTLoad, SDValue &Base,
10605 SDValue &Offset, bool &isInc,
10606 SelectionDAG &DAG) {
10607 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10610 Base = Ptr->getOperand(0);
10611 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10612 int RHSC = (int)RHS->getZExtValue();
10613 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10614 assert(Ptr->getOpcode() == ISD::ADD);
10616 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10618 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10619 isInc = Ptr->getOpcode() == ISD::ADD;
10620 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10628 /// getPreIndexedAddressParts - returns true by value, base pointer and
10629 /// offset pointer and addressing mode by reference if the node's address
10630 /// can be legally represented as pre-indexed load / store address.
10632 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10634 ISD::MemIndexedMode &AM,
10635 SelectionDAG &DAG) const {
10636 if (Subtarget->isThumb1Only())
10641 bool isSEXTLoad = false;
10642 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10643 Ptr = LD->getBasePtr();
10644 VT = LD->getMemoryVT();
10645 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10646 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10647 Ptr = ST->getBasePtr();
10648 VT = ST->getMemoryVT();
10653 bool isLegal = false;
10654 if (Subtarget->isThumb2())
10655 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10656 Offset, isInc, DAG);
10658 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10659 Offset, isInc, DAG);
10663 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10667 /// getPostIndexedAddressParts - returns true by value, base pointer and
10668 /// offset pointer and addressing mode by reference if this node can be
10669 /// combined with a load / store to form a post-indexed load / store.
10670 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10673 ISD::MemIndexedMode &AM,
10674 SelectionDAG &DAG) const {
10675 if (Subtarget->isThumb1Only())
10680 bool isSEXTLoad = false;
10681 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10682 VT = LD->getMemoryVT();
10683 Ptr = LD->getBasePtr();
10684 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10685 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10686 VT = ST->getMemoryVT();
10687 Ptr = ST->getBasePtr();
10692 bool isLegal = false;
10693 if (Subtarget->isThumb2())
10694 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10697 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10703 // Swap base ptr and offset to catch more post-index load / store when
10704 // it's legal. In Thumb2 mode, offset must be an immediate.
10705 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10706 !Subtarget->isThumb2())
10707 std::swap(Base, Offset);
10709 // Post-indexed load / store update the base pointer.
10714 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10718 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10721 const SelectionDAG &DAG,
10722 unsigned Depth) const {
10723 unsigned BitWidth = KnownOne.getBitWidth();
10724 KnownZero = KnownOne = APInt(BitWidth, 0);
10725 switch (Op.getOpcode()) {
10731 // These nodes' second result is a boolean
10732 if (Op.getResNo() == 0)
10734 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10736 case ARMISD::CMOV: {
10737 // Bits are known zero/one if known on the LHS and RHS.
10738 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10739 if (KnownZero == 0 && KnownOne == 0) return;
10741 APInt KnownZeroRHS, KnownOneRHS;
10742 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10743 KnownZero &= KnownZeroRHS;
10744 KnownOne &= KnownOneRHS;
10750 //===----------------------------------------------------------------------===//
10751 // ARM Inline Assembly Support
10752 //===----------------------------------------------------------------------===//
10754 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10755 // Looking for "rev" which is V6+.
10756 if (!Subtarget->hasV6Ops())
10759 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10760 std::string AsmStr = IA->getAsmString();
10761 SmallVector<StringRef, 4> AsmPieces;
10762 SplitString(AsmStr, AsmPieces, ";\n");
10764 switch (AsmPieces.size()) {
10765 default: return false;
10767 AsmStr = AsmPieces[0];
10769 SplitString(AsmStr, AsmPieces, " \t,");
10772 if (AsmPieces.size() == 3 &&
10773 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10774 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10775 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10776 if (Ty && Ty->getBitWidth() == 32)
10777 return IntrinsicLowering::LowerToByteSwap(CI);
10785 /// getConstraintType - Given a constraint letter, return the type of
10786 /// constraint it is for this target.
10787 ARMTargetLowering::ConstraintType
10788 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10789 if (Constraint.size() == 1) {
10790 switch (Constraint[0]) {
10792 case 'l': return C_RegisterClass;
10793 case 'w': return C_RegisterClass;
10794 case 'h': return C_RegisterClass;
10795 case 'x': return C_RegisterClass;
10796 case 't': return C_RegisterClass;
10797 case 'j': return C_Other; // Constant for movw.
10798 // An address with a single base register. Due to the way we
10799 // currently handle addresses it is the same as an 'r' memory constraint.
10800 case 'Q': return C_Memory;
10802 } else if (Constraint.size() == 2) {
10803 switch (Constraint[0]) {
10805 // All 'U+' constraints are addresses.
10806 case 'U': return C_Memory;
10809 return TargetLowering::getConstraintType(Constraint);
10812 /// Examine constraint type and operand type and determine a weight value.
10813 /// This object must already have been set up with the operand type
10814 /// and the current alternative constraint selected.
10815 TargetLowering::ConstraintWeight
10816 ARMTargetLowering::getSingleConstraintMatchWeight(
10817 AsmOperandInfo &info, const char *constraint) const {
10818 ConstraintWeight weight = CW_Invalid;
10819 Value *CallOperandVal = info.CallOperandVal;
10820 // If we don't have a value, we can't do a match,
10821 // but allow it at the lowest weight.
10822 if (CallOperandVal == NULL)
10824 Type *type = CallOperandVal->getType();
10825 // Look at the constraint type.
10826 switch (*constraint) {
10828 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10831 if (type->isIntegerTy()) {
10832 if (Subtarget->isThumb())
10833 weight = CW_SpecificReg;
10835 weight = CW_Register;
10839 if (type->isFloatingPointTy())
10840 weight = CW_Register;
10846 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10848 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10850 if (Constraint.size() == 1) {
10851 // GCC ARM Constraint Letters
10852 switch (Constraint[0]) {
10853 case 'l': // Low regs or general regs.
10854 if (Subtarget->isThumb())
10855 return RCPair(0U, &ARM::tGPRRegClass);
10856 return RCPair(0U, &ARM::GPRRegClass);
10857 case 'h': // High regs or no regs.
10858 if (Subtarget->isThumb())
10859 return RCPair(0U, &ARM::hGPRRegClass);
10862 return RCPair(0U, &ARM::GPRRegClass);
10864 if (VT == MVT::Other)
10866 if (VT == MVT::f32)
10867 return RCPair(0U, &ARM::SPRRegClass);
10868 if (VT.getSizeInBits() == 64)
10869 return RCPair(0U, &ARM::DPRRegClass);
10870 if (VT.getSizeInBits() == 128)
10871 return RCPair(0U, &ARM::QPRRegClass);
10874 if (VT == MVT::Other)
10876 if (VT == MVT::f32)
10877 return RCPair(0U, &ARM::SPR_8RegClass);
10878 if (VT.getSizeInBits() == 64)
10879 return RCPair(0U, &ARM::DPR_8RegClass);
10880 if (VT.getSizeInBits() == 128)
10881 return RCPair(0U, &ARM::QPR_8RegClass);
10884 if (VT == MVT::f32)
10885 return RCPair(0U, &ARM::SPRRegClass);
10889 if (StringRef("{cc}").equals_lower(Constraint))
10890 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10892 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10895 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10896 /// vector. If it is invalid, don't add anything to Ops.
10897 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10898 std::string &Constraint,
10899 std::vector<SDValue>&Ops,
10900 SelectionDAG &DAG) const {
10901 SDValue Result(0, 0);
10903 // Currently only support length 1 constraints.
10904 if (Constraint.length() != 1) return;
10906 char ConstraintLetter = Constraint[0];
10907 switch (ConstraintLetter) {
10910 case 'I': case 'J': case 'K': case 'L':
10911 case 'M': case 'N': case 'O':
10912 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10916 int64_t CVal64 = C->getSExtValue();
10917 int CVal = (int) CVal64;
10918 // None of these constraints allow values larger than 32 bits. Check
10919 // that the value fits in an int.
10920 if (CVal != CVal64)
10923 switch (ConstraintLetter) {
10925 // Constant suitable for movw, must be between 0 and
10927 if (Subtarget->hasV6T2Ops())
10928 if (CVal >= 0 && CVal <= 65535)
10932 if (Subtarget->isThumb1Only()) {
10933 // This must be a constant between 0 and 255, for ADD
10935 if (CVal >= 0 && CVal <= 255)
10937 } else if (Subtarget->isThumb2()) {
10938 // A constant that can be used as an immediate value in a
10939 // data-processing instruction.
10940 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10943 // A constant that can be used as an immediate value in a
10944 // data-processing instruction.
10945 if (ARM_AM::getSOImmVal(CVal) != -1)
10951 if (Subtarget->isThumb()) { // FIXME thumb2
10952 // This must be a constant between -255 and -1, for negated ADD
10953 // immediates. This can be used in GCC with an "n" modifier that
10954 // prints the negated value, for use with SUB instructions. It is
10955 // not useful otherwise but is implemented for compatibility.
10956 if (CVal >= -255 && CVal <= -1)
10959 // This must be a constant between -4095 and 4095. It is not clear
10960 // what this constraint is intended for. Implemented for
10961 // compatibility with GCC.
10962 if (CVal >= -4095 && CVal <= 4095)
10968 if (Subtarget->isThumb1Only()) {
10969 // A 32-bit value where only one byte has a nonzero value. Exclude
10970 // zero to match GCC. This constraint is used by GCC internally for
10971 // constants that can be loaded with a move/shift combination.
10972 // It is not useful otherwise but is implemented for compatibility.
10973 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10975 } else if (Subtarget->isThumb2()) {
10976 // A constant whose bitwise inverse can be used as an immediate
10977 // value in a data-processing instruction. This can be used in GCC
10978 // with a "B" modifier that prints the inverted value, for use with
10979 // BIC and MVN instructions. It is not useful otherwise but is
10980 // implemented for compatibility.
10981 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10984 // A constant whose bitwise inverse can be used as an immediate
10985 // value in a data-processing instruction. This can be used in GCC
10986 // with a "B" modifier that prints the inverted value, for use with
10987 // BIC and MVN instructions. It is not useful otherwise but is
10988 // implemented for compatibility.
10989 if (ARM_AM::getSOImmVal(~CVal) != -1)
10995 if (Subtarget->isThumb1Only()) {
10996 // This must be a constant between -7 and 7,
10997 // for 3-operand ADD/SUB immediate instructions.
10998 if (CVal >= -7 && CVal < 7)
11000 } else if (Subtarget->isThumb2()) {
11001 // A constant whose negation can be used as an immediate value in a
11002 // data-processing instruction. This can be used in GCC with an "n"
11003 // modifier that prints the negated value, for use with SUB
11004 // instructions. It is not useful otherwise but is implemented for
11006 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11009 // A constant whose negation can be used as an immediate value in a
11010 // data-processing instruction. This can be used in GCC with an "n"
11011 // modifier that prints the negated value, for use with SUB
11012 // instructions. It is not useful otherwise but is implemented for
11014 if (ARM_AM::getSOImmVal(-CVal) != -1)
11020 if (Subtarget->isThumb()) { // FIXME thumb2
11021 // This must be a multiple of 4 between 0 and 1020, for
11022 // ADD sp + immediate.
11023 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11026 // A power of two or a constant between 0 and 32. This is used in
11027 // GCC for the shift amount on shifted register operands, but it is
11028 // useful in general for any shift amounts.
11029 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11035 if (Subtarget->isThumb()) { // FIXME thumb2
11036 // This must be a constant between 0 and 31, for shift amounts.
11037 if (CVal >= 0 && CVal <= 31)
11043 if (Subtarget->isThumb()) { // FIXME thumb2
11044 // This must be a multiple of 4 between -508 and 508, for
11045 // ADD/SUB sp = sp + immediate.
11046 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11051 Result = DAG.getTargetConstant(CVal, Op.getValueType());
11055 if (Result.getNode()) {
11056 Ops.push_back(Result);
11059 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11062 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11063 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
11064 unsigned Opcode = Op->getOpcode();
11065 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11066 "Invalid opcode for Div/Rem lowering");
11067 bool isSigned = (Opcode == ISD::SDIVREM);
11068 EVT VT = Op->getValueType(0);
11069 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11072 switch (VT.getSimpleVT().SimpleTy) {
11073 default: llvm_unreachable("Unexpected request for libcall!");
11074 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11075 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11076 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11077 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11080 SDValue InChain = DAG.getEntryNode();
11082 TargetLowering::ArgListTy Args;
11083 TargetLowering::ArgListEntry Entry;
11084 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
11085 EVT ArgVT = Op->getOperand(i).getValueType();
11086 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
11087 Entry.Node = Op->getOperand(i);
11089 Entry.isSExt = isSigned;
11090 Entry.isZExt = !isSigned;
11091 Args.push_back(Entry);
11094 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11097 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
11101 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
11102 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
11103 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
11104 Callee, Args, DAG, dl);
11105 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11107 return CallInfo.first;
11111 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11112 // The ARM target isn't yet aware of offsets.
11116 bool ARM::isBitFieldInvertedMask(unsigned v) {
11117 if (v == 0xffffffff)
11120 // there can be 1's on either or both "outsides", all the "inside"
11121 // bits must be 0's
11122 unsigned TO = CountTrailingOnes_32(v);
11123 unsigned LO = CountLeadingOnes_32(v);
11124 v = (v >> TO) << TO;
11125 v = (v << LO) >> LO;
11129 /// isFPImmLegal - Returns true if the target can instruction select the
11130 /// specified FP immediate natively. If false, the legalizer will
11131 /// materialize the FP immediate as a load from a constant pool.
11132 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11133 if (!Subtarget->hasVFP3())
11135 if (VT == MVT::f32)
11136 return ARM_AM::getFP32Imm(Imm) != -1;
11137 if (VT == MVT::f64)
11138 return ARM_AM::getFP64Imm(Imm) != -1;
11142 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11143 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11144 /// specified in the intrinsic calls.
11145 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11147 unsigned Intrinsic) const {
11148 switch (Intrinsic) {
11149 case Intrinsic::arm_neon_vld1:
11150 case Intrinsic::arm_neon_vld2:
11151 case Intrinsic::arm_neon_vld3:
11152 case Intrinsic::arm_neon_vld4:
11153 case Intrinsic::arm_neon_vld2lane:
11154 case Intrinsic::arm_neon_vld3lane:
11155 case Intrinsic::arm_neon_vld4lane: {
11156 Info.opc = ISD::INTRINSIC_W_CHAIN;
11157 // Conservatively set memVT to the entire set of vectors loaded.
11158 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11159 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11160 Info.ptrVal = I.getArgOperand(0);
11162 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11163 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11164 Info.vol = false; // volatile loads with NEON intrinsics not supported
11165 Info.readMem = true;
11166 Info.writeMem = false;
11169 case Intrinsic::arm_neon_vst1:
11170 case Intrinsic::arm_neon_vst2:
11171 case Intrinsic::arm_neon_vst3:
11172 case Intrinsic::arm_neon_vst4:
11173 case Intrinsic::arm_neon_vst2lane:
11174 case Intrinsic::arm_neon_vst3lane:
11175 case Intrinsic::arm_neon_vst4lane: {
11176 Info.opc = ISD::INTRINSIC_VOID;
11177 // Conservatively set memVT to the entire set of vectors stored.
11178 unsigned NumElts = 0;
11179 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11180 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11181 if (!ArgTy->isVectorTy())
11183 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11185 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11186 Info.ptrVal = I.getArgOperand(0);
11188 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11189 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11190 Info.vol = false; // volatile stores with NEON intrinsics not supported
11191 Info.readMem = false;
11192 Info.writeMem = true;
11195 case Intrinsic::arm_ldrex: {
11196 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11197 Info.opc = ISD::INTRINSIC_W_CHAIN;
11198 Info.memVT = MVT::getVT(PtrTy->getElementType());
11199 Info.ptrVal = I.getArgOperand(0);
11201 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11203 Info.readMem = true;
11204 Info.writeMem = false;
11207 case Intrinsic::arm_strex: {
11208 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11209 Info.opc = ISD::INTRINSIC_W_CHAIN;
11210 Info.memVT = MVT::getVT(PtrTy->getElementType());
11211 Info.ptrVal = I.getArgOperand(1);
11213 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11215 Info.readMem = false;
11216 Info.writeMem = true;
11219 case Intrinsic::arm_strexd: {
11220 Info.opc = ISD::INTRINSIC_W_CHAIN;
11221 Info.memVT = MVT::i64;
11222 Info.ptrVal = I.getArgOperand(2);
11226 Info.readMem = false;
11227 Info.writeMem = true;
11230 case Intrinsic::arm_ldrexd: {
11231 Info.opc = ISD::INTRINSIC_W_CHAIN;
11232 Info.memVT = MVT::i64;
11233 Info.ptrVal = I.getArgOperand(0);
11237 Info.readMem = true;
11238 Info.writeMem = false;