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::DPairRegClass);
160 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
163 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
164 if (TM.getSubtarget<ARMSubtarget>().isTargetMachO())
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->isTargetMachO()) {
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->isTargetMachO()) {
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->isTargetMachO()) {
737 // Non-MachO 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::VRSHRs: return "ARMISD::VRSHRs";
1085 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1086 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1087 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1088 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1089 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1090 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1091 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1092 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1093 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1094 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1095 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1096 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1097 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1098 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1099 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1100 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1101 case ARMISD::VDUP: return "ARMISD::VDUP";
1102 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1103 case ARMISD::VEXT: return "ARMISD::VEXT";
1104 case ARMISD::VREV64: return "ARMISD::VREV64";
1105 case ARMISD::VREV32: return "ARMISD::VREV32";
1106 case ARMISD::VREV16: return "ARMISD::VREV16";
1107 case ARMISD::VZIP: return "ARMISD::VZIP";
1108 case ARMISD::VUZP: return "ARMISD::VUZP";
1109 case ARMISD::VTRN: return "ARMISD::VTRN";
1110 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1111 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1112 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1113 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1114 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1115 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1116 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1117 case ARMISD::FMAX: return "ARMISD::FMAX";
1118 case ARMISD::FMIN: return "ARMISD::FMIN";
1119 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1120 case ARMISD::VMINNM: return "ARMISD::VMIN";
1121 case ARMISD::BFI: return "ARMISD::BFI";
1122 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1123 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1124 case ARMISD::VBSL: return "ARMISD::VBSL";
1125 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1126 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1127 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1128 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1129 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1130 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1131 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1132 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1133 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1134 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1135 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1136 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1137 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1138 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1139 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1140 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1141 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1142 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1143 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1144 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1148 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1149 if (!VT.isVector()) return getPointerTy();
1150 return VT.changeVectorElementTypeToInteger();
1153 /// getRegClassFor - Return the register class that should be used for the
1154 /// specified value type.
1155 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1156 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1157 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1158 // load / store 4 to 8 consecutive D registers.
1159 if (Subtarget->hasNEON()) {
1160 if (VT == MVT::v4i64)
1161 return &ARM::QQPRRegClass;
1162 if (VT == MVT::v8i64)
1163 return &ARM::QQQQPRRegClass;
1165 return TargetLowering::getRegClassFor(VT);
1168 // Create a fast isel object.
1170 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1171 const TargetLibraryInfo *libInfo) const {
1172 return ARM::createFastISel(funcInfo, libInfo);
1175 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1176 /// be used for loads / stores from the global.
1177 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1178 return (Subtarget->isThumb1Only() ? 127 : 4095);
1181 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1182 unsigned NumVals = N->getNumValues();
1184 return Sched::RegPressure;
1186 for (unsigned i = 0; i != NumVals; ++i) {
1187 EVT VT = N->getValueType(i);
1188 if (VT == MVT::Glue || VT == MVT::Other)
1190 if (VT.isFloatingPoint() || VT.isVector())
1194 if (!N->isMachineOpcode())
1195 return Sched::RegPressure;
1197 // Load are scheduled for latency even if there instruction itinerary
1198 // is not available.
1199 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1200 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1202 if (MCID.getNumDefs() == 0)
1203 return Sched::RegPressure;
1204 if (!Itins->isEmpty() &&
1205 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1208 return Sched::RegPressure;
1211 //===----------------------------------------------------------------------===//
1213 //===----------------------------------------------------------------------===//
1215 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1216 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1218 default: llvm_unreachable("Unknown condition code!");
1219 case ISD::SETNE: return ARMCC::NE;
1220 case ISD::SETEQ: return ARMCC::EQ;
1221 case ISD::SETGT: return ARMCC::GT;
1222 case ISD::SETGE: return ARMCC::GE;
1223 case ISD::SETLT: return ARMCC::LT;
1224 case ISD::SETLE: return ARMCC::LE;
1225 case ISD::SETUGT: return ARMCC::HI;
1226 case ISD::SETUGE: return ARMCC::HS;
1227 case ISD::SETULT: return ARMCC::LO;
1228 case ISD::SETULE: return ARMCC::LS;
1232 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1233 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1234 ARMCC::CondCodes &CondCode2) {
1235 CondCode2 = ARMCC::AL;
1237 default: llvm_unreachable("Unknown FP condition!");
1239 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1241 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1243 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1244 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1245 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1246 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1247 case ISD::SETO: CondCode = ARMCC::VC; break;
1248 case ISD::SETUO: CondCode = ARMCC::VS; break;
1249 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1250 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1251 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1253 case ISD::SETULT: CondCode = ARMCC::LT; break;
1255 case ISD::SETULE: CondCode = ARMCC::LE; break;
1257 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1261 //===----------------------------------------------------------------------===//
1262 // Calling Convention Implementation
1263 //===----------------------------------------------------------------------===//
1265 #include "ARMGenCallingConv.inc"
1267 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1268 /// given CallingConvention value.
1269 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1271 bool isVarArg) const {
1274 llvm_unreachable("Unsupported calling convention");
1275 case CallingConv::Fast:
1276 if (Subtarget->hasVFP2() && !isVarArg) {
1277 if (!Subtarget->isAAPCS_ABI())
1278 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1279 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1280 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1283 case CallingConv::C: {
1284 // Use target triple & subtarget features to do actual dispatch.
1285 if (!Subtarget->isAAPCS_ABI())
1286 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1287 else if (Subtarget->hasVFP2() &&
1288 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1290 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1291 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1293 case CallingConv::ARM_AAPCS_VFP:
1295 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1297 case CallingConv::ARM_AAPCS:
1298 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1299 case CallingConv::ARM_APCS:
1300 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1301 case CallingConv::GHC:
1302 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1306 /// LowerCallResult - Lower the result values of a call into the
1307 /// appropriate copies out of appropriate physical registers.
1309 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1310 CallingConv::ID CallConv, bool isVarArg,
1311 const SmallVectorImpl<ISD::InputArg> &Ins,
1312 SDLoc dl, SelectionDAG &DAG,
1313 SmallVectorImpl<SDValue> &InVals,
1314 bool isThisReturn, SDValue ThisVal) const {
1316 // Assign locations to each value returned by this call.
1317 SmallVector<CCValAssign, 16> RVLocs;
1318 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1319 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1320 CCInfo.AnalyzeCallResult(Ins,
1321 CCAssignFnForNode(CallConv, /* Return*/ true,
1324 // Copy all of the result registers out of their specified physreg.
1325 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1326 CCValAssign VA = RVLocs[i];
1328 // Pass 'this' value directly from the argument to return value, to avoid
1329 // reg unit interference
1330 if (i == 0 && isThisReturn) {
1331 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1332 "unexpected return calling convention register assignment");
1333 InVals.push_back(ThisVal);
1338 if (VA.needsCustom()) {
1339 // Handle f64 or half of a v2f64.
1340 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1342 Chain = Lo.getValue(1);
1343 InFlag = Lo.getValue(2);
1344 VA = RVLocs[++i]; // skip ahead to next loc
1345 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1347 Chain = Hi.getValue(1);
1348 InFlag = Hi.getValue(2);
1349 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1351 if (VA.getLocVT() == MVT::v2f64) {
1352 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1353 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1354 DAG.getConstant(0, MVT::i32));
1356 VA = RVLocs[++i]; // skip ahead to next loc
1357 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1358 Chain = Lo.getValue(1);
1359 InFlag = Lo.getValue(2);
1360 VA = RVLocs[++i]; // skip ahead to next loc
1361 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1362 Chain = Hi.getValue(1);
1363 InFlag = Hi.getValue(2);
1364 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1365 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1366 DAG.getConstant(1, MVT::i32));
1369 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1371 Chain = Val.getValue(1);
1372 InFlag = Val.getValue(2);
1375 switch (VA.getLocInfo()) {
1376 default: llvm_unreachable("Unknown loc info!");
1377 case CCValAssign::Full: break;
1378 case CCValAssign::BCvt:
1379 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1383 InVals.push_back(Val);
1389 /// LowerMemOpCallTo - Store the argument to the stack.
1391 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1392 SDValue StackPtr, SDValue Arg,
1393 SDLoc dl, SelectionDAG &DAG,
1394 const CCValAssign &VA,
1395 ISD::ArgFlagsTy Flags) const {
1396 unsigned LocMemOffset = VA.getLocMemOffset();
1397 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1398 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1399 return DAG.getStore(Chain, dl, Arg, PtrOff,
1400 MachinePointerInfo::getStack(LocMemOffset),
1404 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1405 SDValue Chain, SDValue &Arg,
1406 RegsToPassVector &RegsToPass,
1407 CCValAssign &VA, CCValAssign &NextVA,
1409 SmallVectorImpl<SDValue> &MemOpChains,
1410 ISD::ArgFlagsTy Flags) const {
1412 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1413 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1414 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1416 if (NextVA.isRegLoc())
1417 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1419 assert(NextVA.isMemLoc());
1420 if (StackPtr.getNode() == 0)
1421 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1423 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1429 /// LowerCall - Lowering a call into a callseq_start <-
1430 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1433 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1434 SmallVectorImpl<SDValue> &InVals) const {
1435 SelectionDAG &DAG = CLI.DAG;
1437 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1438 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1439 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1440 SDValue Chain = CLI.Chain;
1441 SDValue Callee = CLI.Callee;
1442 bool &isTailCall = CLI.IsTailCall;
1443 CallingConv::ID CallConv = CLI.CallConv;
1444 bool doesNotRet = CLI.DoesNotReturn;
1445 bool isVarArg = CLI.IsVarArg;
1447 MachineFunction &MF = DAG.getMachineFunction();
1448 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1449 bool isThisReturn = false;
1450 bool isSibCall = false;
1451 // Disable tail calls if they're not supported.
1452 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1455 // Check if it's really possible to do a tail call.
1456 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1457 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1458 Outs, OutVals, Ins, DAG);
1459 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1460 // detected sibcalls.
1467 // Analyze operands of the call, assigning locations to each operand.
1468 SmallVector<CCValAssign, 16> ArgLocs;
1469 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1470 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1471 CCInfo.AnalyzeCallOperands(Outs,
1472 CCAssignFnForNode(CallConv, /* Return*/ false,
1475 // Get a count of how many bytes are to be pushed on the stack.
1476 unsigned NumBytes = CCInfo.getNextStackOffset();
1478 // For tail calls, memory operands are available in our caller's stack.
1482 // Adjust the stack pointer for the new arguments...
1483 // These operations are automatically eliminated by the prolog/epilog pass
1485 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1488 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1490 RegsToPassVector RegsToPass;
1491 SmallVector<SDValue, 8> MemOpChains;
1493 // Walk the register/memloc assignments, inserting copies/loads. In the case
1494 // of tail call optimization, arguments are handled later.
1495 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1497 ++i, ++realArgIdx) {
1498 CCValAssign &VA = ArgLocs[i];
1499 SDValue Arg = OutVals[realArgIdx];
1500 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1501 bool isByVal = Flags.isByVal();
1503 // Promote the value if needed.
1504 switch (VA.getLocInfo()) {
1505 default: llvm_unreachable("Unknown loc info!");
1506 case CCValAssign::Full: break;
1507 case CCValAssign::SExt:
1508 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1510 case CCValAssign::ZExt:
1511 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1513 case CCValAssign::AExt:
1514 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1516 case CCValAssign::BCvt:
1517 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1521 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1522 if (VA.needsCustom()) {
1523 if (VA.getLocVT() == MVT::v2f64) {
1524 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1525 DAG.getConstant(0, MVT::i32));
1526 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1527 DAG.getConstant(1, MVT::i32));
1529 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1530 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1532 VA = ArgLocs[++i]; // skip ahead to next loc
1533 if (VA.isRegLoc()) {
1534 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1535 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1537 assert(VA.isMemLoc());
1539 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1540 dl, DAG, VA, Flags));
1543 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1544 StackPtr, MemOpChains, Flags);
1546 } else if (VA.isRegLoc()) {
1547 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1548 assert(VA.getLocVT() == MVT::i32 &&
1549 "unexpected calling convention register assignment");
1550 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1551 "unexpected use of 'returned'");
1552 isThisReturn = true;
1554 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1555 } else if (isByVal) {
1556 assert(VA.isMemLoc());
1557 unsigned offset = 0;
1559 // True if this byval aggregate will be split between registers
1561 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1562 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1564 if (CurByValIdx < ByValArgsCount) {
1566 unsigned RegBegin, RegEnd;
1567 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1569 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1571 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1572 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1573 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1574 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1575 MachinePointerInfo(),
1576 false, false, false,
1577 DAG.InferPtrAlignment(AddArg));
1578 MemOpChains.push_back(Load.getValue(1));
1579 RegsToPass.push_back(std::make_pair(j, Load));
1582 // If parameter size outsides register area, "offset" value
1583 // helps us to calculate stack slot for remained part properly.
1584 offset = RegEnd - RegBegin;
1586 CCInfo.nextInRegsParam();
1589 if (Flags.getByValSize() > 4*offset) {
1590 unsigned LocMemOffset = VA.getLocMemOffset();
1591 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1592 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1594 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1595 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1596 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1598 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1600 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1601 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1602 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1603 Ops, array_lengthof(Ops)));
1605 } else if (!isSibCall) {
1606 assert(VA.isMemLoc());
1608 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1609 dl, DAG, VA, Flags));
1613 if (!MemOpChains.empty())
1614 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1615 &MemOpChains[0], MemOpChains.size());
1617 // Build a sequence of copy-to-reg nodes chained together with token chain
1618 // and flag operands which copy the outgoing args into the appropriate regs.
1620 // Tail call byval lowering might overwrite argument registers so in case of
1621 // tail call optimization the copies to registers are lowered later.
1623 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1624 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1625 RegsToPass[i].second, InFlag);
1626 InFlag = Chain.getValue(1);
1629 // For tail calls lower the arguments to the 'real' stack slot.
1631 // Force all the incoming stack arguments to be loaded from the stack
1632 // before any new outgoing arguments are stored to the stack, because the
1633 // outgoing stack slots may alias the incoming argument stack slots, and
1634 // the alias isn't otherwise explicit. This is slightly more conservative
1635 // than necessary, because it means that each store effectively depends
1636 // on every argument instead of just those arguments it would clobber.
1638 // Do not flag preceding copytoreg stuff together with the following stuff.
1640 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1641 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1642 RegsToPass[i].second, InFlag);
1643 InFlag = Chain.getValue(1);
1648 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1649 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1650 // node so that legalize doesn't hack it.
1651 bool isDirect = false;
1652 bool isARMFunc = false;
1653 bool isLocalARMFunc = false;
1654 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1656 if (EnableARMLongCalls) {
1657 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1658 && "long-calls with non-static relocation model!");
1659 // Handle a global address or an external symbol. If it's not one of
1660 // those, the target's already in a register, so we don't need to do
1662 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1663 const GlobalValue *GV = G->getGlobal();
1664 // Create a constant pool entry for the callee address
1665 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1666 ARMConstantPoolValue *CPV =
1667 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1669 // Get the address of the callee into a register
1670 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1671 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1672 Callee = DAG.getLoad(getPointerTy(), dl,
1673 DAG.getEntryNode(), CPAddr,
1674 MachinePointerInfo::getConstantPool(),
1675 false, false, false, 0);
1676 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1677 const char *Sym = S->getSymbol();
1679 // Create a constant pool entry for the callee address
1680 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1681 ARMConstantPoolValue *CPV =
1682 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1683 ARMPCLabelIndex, 0);
1684 // Get the address of the callee into a register
1685 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1686 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1687 Callee = DAG.getLoad(getPointerTy(), dl,
1688 DAG.getEntryNode(), CPAddr,
1689 MachinePointerInfo::getConstantPool(),
1690 false, false, false, 0);
1692 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1693 const GlobalValue *GV = G->getGlobal();
1695 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1696 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1697 getTargetMachine().getRelocationModel() != Reloc::Static;
1698 isARMFunc = !Subtarget->isThumb() || isStub;
1699 // ARM call to a local ARM function is predicable.
1700 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1701 // tBX takes a register source operand.
1702 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1703 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1704 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1705 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1707 // On ELF targets for PIC code, direct calls should go through the PLT
1708 unsigned OpFlags = 0;
1709 if (Subtarget->isTargetELF() &&
1710 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1711 OpFlags = ARMII::MO_PLT;
1712 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1714 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1716 bool isStub = Subtarget->isTargetMachO() &&
1717 getTargetMachine().getRelocationModel() != Reloc::Static;
1718 isARMFunc = !Subtarget->isThumb() || isStub;
1719 // tBX takes a register source operand.
1720 const char *Sym = S->getSymbol();
1721 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1722 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1723 ARMConstantPoolValue *CPV =
1724 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1725 ARMPCLabelIndex, 4);
1726 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1727 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1728 Callee = DAG.getLoad(getPointerTy(), dl,
1729 DAG.getEntryNode(), CPAddr,
1730 MachinePointerInfo::getConstantPool(),
1731 false, false, false, 0);
1732 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1733 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1734 getPointerTy(), Callee, PICLabel);
1736 unsigned OpFlags = 0;
1737 // On ELF targets for PIC code, direct calls should go through the PLT
1738 if (Subtarget->isTargetELF() &&
1739 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1740 OpFlags = ARMII::MO_PLT;
1741 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1745 // FIXME: handle tail calls differently.
1747 bool HasMinSizeAttr = Subtarget->isMinSize();
1748 if (Subtarget->isThumb()) {
1749 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1750 CallOpc = ARMISD::CALL_NOLINK;
1752 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1754 if (!isDirect && !Subtarget->hasV5TOps())
1755 CallOpc = ARMISD::CALL_NOLINK;
1756 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1757 // Emit regular call when code size is the priority
1759 // "mov lr, pc; b _foo" to avoid confusing the RSP
1760 CallOpc = ARMISD::CALL_NOLINK;
1762 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1765 std::vector<SDValue> Ops;
1766 Ops.push_back(Chain);
1767 Ops.push_back(Callee);
1769 // Add argument registers to the end of the list so that they are known live
1771 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1772 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1773 RegsToPass[i].second.getValueType()));
1775 // Add a register mask operand representing the call-preserved registers.
1777 const uint32_t *Mask;
1778 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1779 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1781 // For 'this' returns, use the R0-preserving mask if applicable
1782 Mask = ARI->getThisReturnPreservedMask(CallConv);
1784 // Set isThisReturn to false if the calling convention is not one that
1785 // allows 'returned' to be modeled in this way, so LowerCallResult does
1786 // not try to pass 'this' straight through
1787 isThisReturn = false;
1788 Mask = ARI->getCallPreservedMask(CallConv);
1791 Mask = ARI->getCallPreservedMask(CallConv);
1793 assert(Mask && "Missing call preserved mask for calling convention");
1794 Ops.push_back(DAG.getRegisterMask(Mask));
1797 if (InFlag.getNode())
1798 Ops.push_back(InFlag);
1800 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1802 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1804 // Returns a chain and a flag for retval copy to use.
1805 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1806 InFlag = Chain.getValue(1);
1808 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1809 DAG.getIntPtrConstant(0, true), InFlag, dl);
1811 InFlag = Chain.getValue(1);
1813 // Handle result values, copying them out of physregs into vregs that we
1815 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1816 InVals, isThisReturn,
1817 isThisReturn ? OutVals[0] : SDValue());
1820 /// HandleByVal - Every parameter *after* a byval parameter is passed
1821 /// on the stack. Remember the next parameter register to allocate,
1822 /// and then confiscate the rest of the parameter registers to insure
1825 ARMTargetLowering::HandleByVal(
1826 CCState *State, unsigned &size, unsigned Align) const {
1827 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1828 assert((State->getCallOrPrologue() == Prologue ||
1829 State->getCallOrPrologue() == Call) &&
1830 "unhandled ParmContext");
1832 // For in-prologue parameters handling, we also introduce stack offset
1833 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
1834 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
1835 // NSAA should be evaluted (NSAA means "next stacked argument address").
1836 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
1837 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
1838 unsigned NSAAOffset = State->getNextStackOffset();
1839 if (State->getCallOrPrologue() != Call) {
1840 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
1842 State->getInRegsParamInfo(i, RB, RE);
1843 assert(NSAAOffset >= (RE-RB)*4 &&
1844 "Stack offset for byval regs doesn't introduced anymore?");
1845 NSAAOffset -= (RE-RB)*4;
1848 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1849 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1850 unsigned AlignInRegs = Align / 4;
1851 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1852 for (unsigned i = 0; i < Waste; ++i)
1853 reg = State->AllocateReg(GPRArgRegs, 4);
1856 unsigned excess = 4 * (ARM::R4 - reg);
1858 // Special case when NSAA != SP and parameter size greater than size of
1859 // all remained GPR regs. In that case we can't split parameter, we must
1860 // send it to stack. We also must set NCRN to R4, so waste all
1861 // remained registers.
1862 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1863 while (State->AllocateReg(GPRArgRegs, 4))
1868 // First register for byval parameter is the first register that wasn't
1869 // allocated before this method call, so it would be "reg".
1870 // If parameter is small enough to be saved in range [reg, r4), then
1871 // the end (first after last) register would be reg + param-size-in-regs,
1872 // else parameter would be splitted between registers and stack,
1873 // end register would be r4 in this case.
1874 unsigned ByValRegBegin = reg;
1875 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1876 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1877 // Note, first register is allocated in the beginning of function already,
1878 // allocate remained amount of registers we need.
1879 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1880 State->AllocateReg(GPRArgRegs, 4);
1881 // At a call site, a byval parameter that is split between
1882 // registers and memory needs its size truncated here. In a
1883 // function prologue, such byval parameters are reassembled in
1884 // memory, and are not truncated.
1885 if (State->getCallOrPrologue() == Call) {
1886 // Make remained size equal to 0 in case, when
1887 // the whole structure may be stored into registers.
1897 /// MatchingStackOffset - Return true if the given stack call argument is
1898 /// already available in the same position (relatively) of the caller's
1899 /// incoming argument stack.
1901 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1902 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1903 const TargetInstrInfo *TII) {
1904 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1906 if (Arg.getOpcode() == ISD::CopyFromReg) {
1907 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1908 if (!TargetRegisterInfo::isVirtualRegister(VR))
1910 MachineInstr *Def = MRI->getVRegDef(VR);
1913 if (!Flags.isByVal()) {
1914 if (!TII->isLoadFromStackSlot(Def, FI))
1919 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1920 if (Flags.isByVal())
1921 // ByVal argument is passed in as a pointer but it's now being
1922 // dereferenced. e.g.
1923 // define @foo(%struct.X* %A) {
1924 // tail call @bar(%struct.X* byval %A)
1927 SDValue Ptr = Ld->getBasePtr();
1928 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1931 FI = FINode->getIndex();
1935 assert(FI != INT_MAX);
1936 if (!MFI->isFixedObjectIndex(FI))
1938 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1941 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1942 /// for tail call optimization. Targets which want to do tail call
1943 /// optimization should implement this function.
1945 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1946 CallingConv::ID CalleeCC,
1948 bool isCalleeStructRet,
1949 bool isCallerStructRet,
1950 const SmallVectorImpl<ISD::OutputArg> &Outs,
1951 const SmallVectorImpl<SDValue> &OutVals,
1952 const SmallVectorImpl<ISD::InputArg> &Ins,
1953 SelectionDAG& DAG) const {
1954 const Function *CallerF = DAG.getMachineFunction().getFunction();
1955 CallingConv::ID CallerCC = CallerF->getCallingConv();
1956 bool CCMatch = CallerCC == CalleeCC;
1958 // Look for obvious safe cases to perform tail call optimization that do not
1959 // require ABI changes. This is what gcc calls sibcall.
1961 // Do not sibcall optimize vararg calls unless the call site is not passing
1963 if (isVarArg && !Outs.empty())
1966 // Exception-handling functions need a special set of instructions to indicate
1967 // a return to the hardware. Tail-calling another function would probably
1969 if (CallerF->hasFnAttribute("interrupt"))
1972 // Also avoid sibcall optimization if either caller or callee uses struct
1973 // return semantics.
1974 if (isCalleeStructRet || isCallerStructRet)
1977 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1978 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1979 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1980 // support in the assembler and linker to be used. This would need to be
1981 // fixed to fully support tail calls in Thumb1.
1983 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1984 // LR. This means if we need to reload LR, it takes an extra instructions,
1985 // which outweighs the value of the tail call; but here we don't know yet
1986 // whether LR is going to be used. Probably the right approach is to
1987 // generate the tail call here and turn it back into CALL/RET in
1988 // emitEpilogue if LR is used.
1990 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1991 // but we need to make sure there are enough registers; the only valid
1992 // registers are the 4 used for parameters. We don't currently do this
1994 if (Subtarget->isThumb1Only())
1997 // If the calling conventions do not match, then we'd better make sure the
1998 // results are returned in the same way as what the caller expects.
2000 SmallVector<CCValAssign, 16> RVLocs1;
2001 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2002 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
2003 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2005 SmallVector<CCValAssign, 16> RVLocs2;
2006 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2007 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
2008 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2010 if (RVLocs1.size() != RVLocs2.size())
2012 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2013 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2015 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2017 if (RVLocs1[i].isRegLoc()) {
2018 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2021 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2027 // If Caller's vararg or byval argument has been split between registers and
2028 // stack, do not perform tail call, since part of the argument is in caller's
2030 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2031 getInfo<ARMFunctionInfo>();
2032 if (AFI_Caller->getArgRegsSaveSize())
2035 // If the callee takes no arguments then go on to check the results of the
2037 if (!Outs.empty()) {
2038 // Check if stack adjustment is needed. For now, do not do this if any
2039 // argument is passed on the stack.
2040 SmallVector<CCValAssign, 16> ArgLocs;
2041 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2042 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2043 CCInfo.AnalyzeCallOperands(Outs,
2044 CCAssignFnForNode(CalleeCC, false, isVarArg));
2045 if (CCInfo.getNextStackOffset()) {
2046 MachineFunction &MF = DAG.getMachineFunction();
2048 // Check if the arguments are already laid out in the right way as
2049 // the caller's fixed stack objects.
2050 MachineFrameInfo *MFI = MF.getFrameInfo();
2051 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2052 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2053 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2055 ++i, ++realArgIdx) {
2056 CCValAssign &VA = ArgLocs[i];
2057 EVT RegVT = VA.getLocVT();
2058 SDValue Arg = OutVals[realArgIdx];
2059 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2060 if (VA.getLocInfo() == CCValAssign::Indirect)
2062 if (VA.needsCustom()) {
2063 // f64 and vector types are split into multiple registers or
2064 // register/stack-slot combinations. The types will not match
2065 // the registers; give up on memory f64 refs until we figure
2066 // out what to do about this.
2069 if (!ArgLocs[++i].isRegLoc())
2071 if (RegVT == MVT::v2f64) {
2072 if (!ArgLocs[++i].isRegLoc())
2074 if (!ArgLocs[++i].isRegLoc())
2077 } else if (!VA.isRegLoc()) {
2078 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2090 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2091 MachineFunction &MF, bool isVarArg,
2092 const SmallVectorImpl<ISD::OutputArg> &Outs,
2093 LLVMContext &Context) const {
2094 SmallVector<CCValAssign, 16> RVLocs;
2095 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2096 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2100 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2101 SDLoc DL, SelectionDAG &DAG) {
2102 const MachineFunction &MF = DAG.getMachineFunction();
2103 const Function *F = MF.getFunction();
2105 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2107 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2108 // version of the "preferred return address". These offsets affect the return
2109 // instruction if this is a return from PL1 without hypervisor extensions.
2110 // IRQ/FIQ: +4 "subs pc, lr, #4"
2111 // SWI: 0 "subs pc, lr, #0"
2112 // ABORT: +4 "subs pc, lr, #4"
2113 // UNDEF: +4/+2 "subs pc, lr, #0"
2114 // UNDEF varies depending on where the exception came from ARM or Thumb
2115 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2118 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2121 else if (IntKind == "SWI" || IntKind == "UNDEF")
2124 report_fatal_error("Unsupported interrupt attribute. If present, value "
2125 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2127 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2129 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other,
2130 RetOps.data(), RetOps.size());
2134 ARMTargetLowering::LowerReturn(SDValue Chain,
2135 CallingConv::ID CallConv, bool isVarArg,
2136 const SmallVectorImpl<ISD::OutputArg> &Outs,
2137 const SmallVectorImpl<SDValue> &OutVals,
2138 SDLoc dl, SelectionDAG &DAG) const {
2140 // CCValAssign - represent the assignment of the return value to a location.
2141 SmallVector<CCValAssign, 16> RVLocs;
2143 // CCState - Info about the registers and stack slots.
2144 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2145 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2147 // Analyze outgoing return values.
2148 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2152 SmallVector<SDValue, 4> RetOps;
2153 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2155 // Copy the result values into the output registers.
2156 for (unsigned i = 0, realRVLocIdx = 0;
2158 ++i, ++realRVLocIdx) {
2159 CCValAssign &VA = RVLocs[i];
2160 assert(VA.isRegLoc() && "Can only return in registers!");
2162 SDValue Arg = OutVals[realRVLocIdx];
2164 switch (VA.getLocInfo()) {
2165 default: llvm_unreachable("Unknown loc info!");
2166 case CCValAssign::Full: break;
2167 case CCValAssign::BCvt:
2168 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2172 if (VA.needsCustom()) {
2173 if (VA.getLocVT() == MVT::v2f64) {
2174 // Extract the first half and return it in two registers.
2175 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2176 DAG.getConstant(0, MVT::i32));
2177 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2178 DAG.getVTList(MVT::i32, MVT::i32), Half);
2180 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
2181 Flag = Chain.getValue(1);
2182 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2183 VA = RVLocs[++i]; // skip ahead to next loc
2184 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2185 HalfGPRs.getValue(1), Flag);
2186 Flag = Chain.getValue(1);
2187 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2188 VA = RVLocs[++i]; // skip ahead to next loc
2190 // Extract the 2nd half and fall through to handle it as an f64 value.
2191 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2192 DAG.getConstant(1, MVT::i32));
2194 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2196 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2197 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2198 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2199 Flag = Chain.getValue(1);
2200 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2201 VA = RVLocs[++i]; // skip ahead to next loc
2202 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2205 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2207 // Guarantee that all emitted copies are
2208 // stuck together, avoiding something bad.
2209 Flag = Chain.getValue(1);
2210 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2213 // Update chain and glue.
2216 RetOps.push_back(Flag);
2218 // CPUs which aren't M-class use a special sequence to return from
2219 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2220 // though we use "subs pc, lr, #N").
2222 // M-class CPUs actually use a normal return sequence with a special
2223 // (hardware-provided) value in LR, so the normal code path works.
2224 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2225 !Subtarget->isMClass()) {
2226 if (Subtarget->isThumb1Only())
2227 report_fatal_error("interrupt attribute is not supported in Thumb1");
2228 return LowerInterruptReturn(RetOps, dl, DAG);
2231 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2232 RetOps.data(), RetOps.size());
2235 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2236 if (N->getNumValues() != 1)
2238 if (!N->hasNUsesOfValue(1, 0))
2241 SDValue TCChain = Chain;
2242 SDNode *Copy = *N->use_begin();
2243 if (Copy->getOpcode() == ISD::CopyToReg) {
2244 // If the copy has a glue operand, we conservatively assume it isn't safe to
2245 // perform a tail call.
2246 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2248 TCChain = Copy->getOperand(0);
2249 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2250 SDNode *VMov = Copy;
2251 // f64 returned in a pair of GPRs.
2252 SmallPtrSet<SDNode*, 2> Copies;
2253 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2255 if (UI->getOpcode() != ISD::CopyToReg)
2259 if (Copies.size() > 2)
2262 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2264 SDValue UseChain = UI->getOperand(0);
2265 if (Copies.count(UseChain.getNode()))
2272 } else if (Copy->getOpcode() == ISD::BITCAST) {
2273 // f32 returned in a single GPR.
2274 if (!Copy->hasOneUse())
2276 Copy = *Copy->use_begin();
2277 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2279 TCChain = Copy->getOperand(0);
2284 bool HasRet = false;
2285 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2287 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2288 UI->getOpcode() != ARMISD::INTRET_FLAG)
2300 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2301 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2304 if (!CI->isTailCall())
2307 return !Subtarget->isThumb1Only();
2310 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2311 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2312 // one of the above mentioned nodes. It has to be wrapped because otherwise
2313 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2314 // be used to form addressing mode. These wrapped nodes will be selected
2316 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2317 EVT PtrVT = Op.getValueType();
2318 // FIXME there is no actual debug info here
2320 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2322 if (CP->isMachineConstantPoolEntry())
2323 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2324 CP->getAlignment());
2326 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2327 CP->getAlignment());
2328 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2331 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2332 return MachineJumpTableInfo::EK_Inline;
2335 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2336 SelectionDAG &DAG) const {
2337 MachineFunction &MF = DAG.getMachineFunction();
2338 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2339 unsigned ARMPCLabelIndex = 0;
2341 EVT PtrVT = getPointerTy();
2342 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2343 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2345 if (RelocM == Reloc::Static) {
2346 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2348 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2349 ARMPCLabelIndex = AFI->createPICLabelUId();
2350 ARMConstantPoolValue *CPV =
2351 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2352 ARMCP::CPBlockAddress, PCAdj);
2353 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2355 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2356 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2357 MachinePointerInfo::getConstantPool(),
2358 false, false, false, 0);
2359 if (RelocM == Reloc::Static)
2361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2362 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2365 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2367 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2368 SelectionDAG &DAG) const {
2370 EVT PtrVT = getPointerTy();
2371 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2372 MachineFunction &MF = DAG.getMachineFunction();
2373 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2374 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2375 ARMConstantPoolValue *CPV =
2376 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2377 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2378 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2379 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2380 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2381 MachinePointerInfo::getConstantPool(),
2382 false, false, false, 0);
2383 SDValue Chain = Argument.getValue(1);
2385 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2386 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2388 // call __tls_get_addr.
2391 Entry.Node = Argument;
2392 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2393 Args.push_back(Entry);
2394 // FIXME: is there useful debug info available here?
2395 TargetLowering::CallLoweringInfo CLI(Chain,
2396 (Type *) Type::getInt32Ty(*DAG.getContext()),
2397 false, false, false, false,
2398 0, CallingConv::C, /*isTailCall=*/false,
2399 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2400 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2401 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2402 return CallResult.first;
2405 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2406 // "local exec" model.
2408 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2410 TLSModel::Model model) const {
2411 const GlobalValue *GV = GA->getGlobal();
2414 SDValue Chain = DAG.getEntryNode();
2415 EVT PtrVT = getPointerTy();
2416 // Get the Thread Pointer
2417 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2419 if (model == TLSModel::InitialExec) {
2420 MachineFunction &MF = DAG.getMachineFunction();
2421 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2422 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2423 // Initial exec model.
2424 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2425 ARMConstantPoolValue *CPV =
2426 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2427 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2429 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2430 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2431 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2432 MachinePointerInfo::getConstantPool(),
2433 false, false, false, 0);
2434 Chain = Offset.getValue(1);
2436 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2437 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2439 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2440 MachinePointerInfo::getConstantPool(),
2441 false, false, false, 0);
2444 assert(model == TLSModel::LocalExec);
2445 ARMConstantPoolValue *CPV =
2446 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2447 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2448 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2449 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2450 MachinePointerInfo::getConstantPool(),
2451 false, false, false, 0);
2454 // The address of the thread local variable is the add of the thread
2455 // pointer with the offset of the variable.
2456 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2460 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2461 // TODO: implement the "local dynamic" model
2462 assert(Subtarget->isTargetELF() &&
2463 "TLS not implemented for non-ELF targets");
2464 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2466 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2469 case TLSModel::GeneralDynamic:
2470 case TLSModel::LocalDynamic:
2471 return LowerToTLSGeneralDynamicModel(GA, DAG);
2472 case TLSModel::InitialExec:
2473 case TLSModel::LocalExec:
2474 return LowerToTLSExecModels(GA, DAG, model);
2476 llvm_unreachable("bogus TLS model");
2479 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2480 SelectionDAG &DAG) const {
2481 EVT PtrVT = getPointerTy();
2483 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2484 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2485 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2486 ARMConstantPoolValue *CPV =
2487 ARMConstantPoolConstant::Create(GV,
2488 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2489 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2490 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2491 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2493 MachinePointerInfo::getConstantPool(),
2494 false, false, false, 0);
2495 SDValue Chain = Result.getValue(1);
2496 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2497 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2499 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2500 MachinePointerInfo::getGOT(),
2501 false, false, false, 0);
2505 // If we have T2 ops, we can materialize the address directly via movt/movw
2506 // pair. This is always cheaper.
2507 if (Subtarget->useMovt()) {
2509 // FIXME: Once remat is capable of dealing with instructions with register
2510 // operands, expand this into two nodes.
2511 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2512 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2514 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2515 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2516 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2517 MachinePointerInfo::getConstantPool(),
2518 false, false, false, 0);
2522 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2523 SelectionDAG &DAG) const {
2524 EVT PtrVT = getPointerTy();
2526 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2527 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2529 if (Subtarget->useMovt())
2532 // FIXME: Once remat is capable of dealing with instructions with register
2533 // operands, expand this into multiple nodes
2535 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2537 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2538 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2540 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2541 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2542 MachinePointerInfo::getGOT(), false, false, false, 0);
2546 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2547 SelectionDAG &DAG) const {
2548 assert(Subtarget->isTargetELF() &&
2549 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2550 MachineFunction &MF = DAG.getMachineFunction();
2551 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2552 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2553 EVT PtrVT = getPointerTy();
2555 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2556 ARMConstantPoolValue *CPV =
2557 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2558 ARMPCLabelIndex, PCAdj);
2559 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2560 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2561 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2562 MachinePointerInfo::getConstantPool(),
2563 false, false, false, 0);
2564 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2565 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2569 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2571 SDValue Val = DAG.getConstant(0, MVT::i32);
2572 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2573 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2574 Op.getOperand(1), Val);
2578 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2580 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2581 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2585 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2586 const ARMSubtarget *Subtarget) const {
2587 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2590 default: return SDValue(); // Don't custom lower most intrinsics.
2591 case Intrinsic::arm_thread_pointer: {
2592 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2593 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2595 case Intrinsic::eh_sjlj_lsda: {
2596 MachineFunction &MF = DAG.getMachineFunction();
2597 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2598 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2599 EVT PtrVT = getPointerTy();
2600 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2602 unsigned PCAdj = (RelocM != Reloc::PIC_)
2603 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2604 ARMConstantPoolValue *CPV =
2605 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2606 ARMCP::CPLSDA, PCAdj);
2607 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2608 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2610 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2611 MachinePointerInfo::getConstantPool(),
2612 false, false, false, 0);
2614 if (RelocM == Reloc::PIC_) {
2615 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2616 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2620 case Intrinsic::arm_neon_vmulls:
2621 case Intrinsic::arm_neon_vmullu: {
2622 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2623 ? ARMISD::VMULLs : ARMISD::VMULLu;
2624 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2625 Op.getOperand(1), Op.getOperand(2));
2630 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2631 const ARMSubtarget *Subtarget) {
2632 // FIXME: handle "fence singlethread" more efficiently.
2634 if (!Subtarget->hasDataBarrier()) {
2635 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2636 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2638 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2639 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2640 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2641 DAG.getConstant(0, MVT::i32));
2644 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2645 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2646 unsigned Domain = ARM_MB::ISH;
2647 if (Subtarget->isMClass()) {
2648 // Only a full system barrier exists in the M-class architectures.
2649 Domain = ARM_MB::SY;
2650 } else if (Subtarget->isSwift() && Ord == Release) {
2651 // Swift happens to implement ISHST barriers in a way that's compatible with
2652 // Release semantics but weaker than ISH so we'd be fools not to use
2653 // it. Beware: other processors probably don't!
2654 Domain = ARM_MB::ISHST;
2657 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2658 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2659 DAG.getConstant(Domain, MVT::i32));
2662 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2663 const ARMSubtarget *Subtarget) {
2664 // ARM pre v5TE and Thumb1 does not have preload instructions.
2665 if (!(Subtarget->isThumb2() ||
2666 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2667 // Just preserve the chain.
2668 return Op.getOperand(0);
2671 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2673 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2674 // ARMv7 with MP extension has PLDW.
2675 return Op.getOperand(0);
2677 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2678 if (Subtarget->isThumb()) {
2680 isRead = ~isRead & 1;
2681 isData = ~isData & 1;
2684 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2685 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2686 DAG.getConstant(isData, MVT::i32));
2689 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2690 MachineFunction &MF = DAG.getMachineFunction();
2691 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2693 // vastart just stores the address of the VarArgsFrameIndex slot into the
2694 // memory location argument.
2696 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2697 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2698 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2699 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2700 MachinePointerInfo(SV), false, false, 0);
2704 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2705 SDValue &Root, SelectionDAG &DAG,
2707 MachineFunction &MF = DAG.getMachineFunction();
2708 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2710 const TargetRegisterClass *RC;
2711 if (AFI->isThumb1OnlyFunction())
2712 RC = &ARM::tGPRRegClass;
2714 RC = &ARM::GPRRegClass;
2716 // Transform the arguments stored in physical registers into virtual ones.
2717 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2718 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2721 if (NextVA.isMemLoc()) {
2722 MachineFrameInfo *MFI = MF.getFrameInfo();
2723 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2725 // Create load node to retrieve arguments from the stack.
2726 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2727 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2728 MachinePointerInfo::getFixedStack(FI),
2729 false, false, false, 0);
2731 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2732 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2735 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2739 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2740 unsigned InRegsParamRecordIdx,
2742 unsigned &ArgRegsSize,
2743 unsigned &ArgRegsSaveSize)
2746 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2747 unsigned RBegin, REnd;
2748 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2749 NumGPRs = REnd - RBegin;
2751 unsigned int firstUnalloced;
2752 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2753 sizeof(GPRArgRegs) /
2754 sizeof(GPRArgRegs[0]));
2755 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2758 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2759 ArgRegsSize = NumGPRs * 4;
2761 // If parameter is split between stack and GPRs...
2762 if (NumGPRs && Align == 8 &&
2763 (ArgRegsSize < ArgSize ||
2764 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2765 // Add padding for part of param recovered from GPRs, so
2766 // its last byte must be at address K*8 - 1.
2767 // We need to do it, since remained (stack) part of parameter has
2768 // stack alignment, and we need to "attach" "GPRs head" without gaps
2771 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2772 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2774 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2776 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
2777 (ArgRegsSize + AFI->getArgRegsSaveSize());
2778 ArgRegsSaveSize = ArgRegsSize + Padding;
2780 // We don't need to extend regs save size for byval parameters if they
2781 // are passed via GPRs only.
2782 ArgRegsSaveSize = ArgRegsSize;
2785 // The remaining GPRs hold either the beginning of variable-argument
2786 // data, or the beginning of an aggregate passed by value (usually
2787 // byval). Either way, we allocate stack slots adjacent to the data
2788 // provided by our caller, and store the unallocated registers there.
2789 // If this is a variadic function, the va_list pointer will begin with
2790 // these values; otherwise, this reassembles a (byval) structure that
2791 // was split between registers and memory.
2792 // Return: The frame index registers were stored into.
2794 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2795 SDLoc dl, SDValue &Chain,
2796 const Value *OrigArg,
2797 unsigned InRegsParamRecordIdx,
2798 unsigned OffsetFromOrigArg,
2801 bool ForceMutable) const {
2803 // Currently, two use-cases possible:
2804 // Case #1. Non-var-args function, and we meet first byval parameter.
2805 // Setup first unallocated register as first byval register;
2806 // eat all remained registers
2807 // (these two actions are performed by HandleByVal method).
2808 // Then, here, we initialize stack frame with
2809 // "store-reg" instructions.
2810 // Case #2. Var-args function, that doesn't contain byval parameters.
2811 // The same: eat all remained unallocated registers,
2812 // initialize stack frame.
2814 MachineFunction &MF = DAG.getMachineFunction();
2815 MachineFrameInfo *MFI = MF.getFrameInfo();
2816 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2817 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2818 unsigned RBegin, REnd;
2819 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2820 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2821 firstRegToSaveIndex = RBegin - ARM::R0;
2822 lastRegToSaveIndex = REnd - ARM::R0;
2824 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2825 (GPRArgRegs, array_lengthof(GPRArgRegs));
2826 lastRegToSaveIndex = 4;
2829 unsigned ArgRegsSize, ArgRegsSaveSize;
2830 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2831 ArgRegsSize, ArgRegsSaveSize);
2833 // Store any by-val regs to their spots on the stack so that they may be
2834 // loaded by deferencing the result of formal parameter pointer or va_next.
2835 // Note: once stack area for byval/varargs registers
2836 // was initialized, it can't be initialized again.
2837 if (ArgRegsSaveSize) {
2839 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2842 assert(AFI->getStoredByValParamsPadding() == 0 &&
2843 "The only parameter may be padded.");
2844 AFI->setStoredByValParamsPadding(Padding);
2847 int FrameIndex = MFI->CreateFixedObject(
2849 Padding + ArgOffset,
2851 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2853 SmallVector<SDValue, 4> MemOps;
2854 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2855 ++firstRegToSaveIndex, ++i) {
2856 const TargetRegisterClass *RC;
2857 if (AFI->isThumb1OnlyFunction())
2858 RC = &ARM::tGPRRegClass;
2860 RC = &ARM::GPRRegClass;
2862 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2863 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2865 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2866 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2868 MemOps.push_back(Store);
2869 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2870 DAG.getConstant(4, getPointerTy()));
2873 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2875 if (!MemOps.empty())
2876 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2877 &MemOps[0], MemOps.size());
2880 // This will point to the next argument passed via stack.
2881 return MFI->CreateFixedObject(
2882 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
2885 // Setup stack frame, the va_list pointer will start from.
2887 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2888 SDLoc dl, SDValue &Chain,
2890 bool ForceMutable) const {
2891 MachineFunction &MF = DAG.getMachineFunction();
2892 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2894 // Try to store any remaining integer argument regs
2895 // to their spots on the stack so that they may be loaded by deferencing
2896 // the result of va_next.
2897 // If there is no regs to be stored, just point address after last
2898 // argument passed via stack.
2900 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
2901 0, ArgOffset, 0, ForceMutable);
2903 AFI->setVarArgsFrameIndex(FrameIndex);
2907 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2908 CallingConv::ID CallConv, bool isVarArg,
2909 const SmallVectorImpl<ISD::InputArg>
2911 SDLoc dl, SelectionDAG &DAG,
2912 SmallVectorImpl<SDValue> &InVals)
2914 MachineFunction &MF = DAG.getMachineFunction();
2915 MachineFrameInfo *MFI = MF.getFrameInfo();
2917 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2919 // Assign locations to all of the incoming arguments.
2920 SmallVector<CCValAssign, 16> ArgLocs;
2921 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2922 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2923 CCInfo.AnalyzeFormalArguments(Ins,
2924 CCAssignFnForNode(CallConv, /* Return*/ false,
2927 SmallVector<SDValue, 16> ArgValues;
2928 int lastInsIndex = -1;
2930 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2931 unsigned CurArgIdx = 0;
2933 // Initially ArgRegsSaveSize is zero.
2934 // Then we increase this value each time we meet byval parameter.
2935 // We also increase this value in case of varargs function.
2936 AFI->setArgRegsSaveSize(0);
2938 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2939 CCValAssign &VA = ArgLocs[i];
2940 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2941 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2942 // Arguments stored in registers.
2943 if (VA.isRegLoc()) {
2944 EVT RegVT = VA.getLocVT();
2946 if (VA.needsCustom()) {
2947 // f64 and vector types are split up into multiple registers or
2948 // combinations of registers and stack slots.
2949 if (VA.getLocVT() == MVT::v2f64) {
2950 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2952 VA = ArgLocs[++i]; // skip ahead to next loc
2954 if (VA.isMemLoc()) {
2955 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2956 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2957 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2958 MachinePointerInfo::getFixedStack(FI),
2959 false, false, false, 0);
2961 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2964 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2965 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2966 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2967 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2968 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2970 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2973 const TargetRegisterClass *RC;
2975 if (RegVT == MVT::f32)
2976 RC = &ARM::SPRRegClass;
2977 else if (RegVT == MVT::f64)
2978 RC = &ARM::DPRRegClass;
2979 else if (RegVT == MVT::v2f64)
2980 RC = &ARM::QPRRegClass;
2981 else if (RegVT == MVT::i32)
2982 RC = AFI->isThumb1OnlyFunction() ?
2983 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2984 (const TargetRegisterClass*)&ARM::GPRRegClass;
2986 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2988 // Transform the arguments in physical registers into virtual ones.
2989 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2990 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2993 // If this is an 8 or 16-bit value, it is really passed promoted
2994 // to 32 bits. Insert an assert[sz]ext to capture this, then
2995 // truncate to the right size.
2996 switch (VA.getLocInfo()) {
2997 default: llvm_unreachable("Unknown loc info!");
2998 case CCValAssign::Full: break;
2999 case CCValAssign::BCvt:
3000 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3002 case CCValAssign::SExt:
3003 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3004 DAG.getValueType(VA.getValVT()));
3005 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3007 case CCValAssign::ZExt:
3008 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3009 DAG.getValueType(VA.getValVT()));
3010 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3014 InVals.push_back(ArgValue);
3016 } else { // VA.isRegLoc()
3019 assert(VA.isMemLoc());
3020 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3022 int index = ArgLocs[i].getValNo();
3024 // Some Ins[] entries become multiple ArgLoc[] entries.
3025 // Process them only once.
3026 if (index != lastInsIndex)
3028 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3029 // FIXME: For now, all byval parameter objects are marked mutable.
3030 // This can be changed with more analysis.
3031 // In case of tail call optimization mark all arguments mutable.
3032 // Since they could be overwritten by lowering of arguments in case of
3034 if (Flags.isByVal()) {
3035 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3036 int FrameIndex = StoreByValRegs(
3037 CCInfo, DAG, dl, Chain, CurOrigArg,
3039 Ins[VA.getValNo()].PartOffset,
3040 VA.getLocMemOffset(),
3041 Flags.getByValSize(),
3042 true /*force mutable frames*/);
3043 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3044 CCInfo.nextInRegsParam();
3046 unsigned FIOffset = VA.getLocMemOffset() +
3047 AFI->getStoredByValParamsPadding();
3048 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3051 // Create load nodes to retrieve arguments from the stack.
3052 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3053 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3054 MachinePointerInfo::getFixedStack(FI),
3055 false, false, false, 0));
3057 lastInsIndex = index;
3064 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3065 CCInfo.getNextStackOffset());
3070 /// isFloatingPointZero - Return true if this is +0.0.
3071 static bool isFloatingPointZero(SDValue Op) {
3072 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3073 return CFP->getValueAPF().isPosZero();
3074 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3075 // Maybe this has already been legalized into the constant pool?
3076 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3077 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3078 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3079 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3080 return CFP->getValueAPF().isPosZero();
3086 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3087 /// the given operands.
3089 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3090 SDValue &ARMcc, SelectionDAG &DAG,
3092 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3093 unsigned C = RHSC->getZExtValue();
3094 if (!isLegalICmpImmediate(C)) {
3095 // Constant does not fit, try adjusting it by one?
3100 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3101 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3102 RHS = DAG.getConstant(C-1, MVT::i32);
3107 if (C != 0 && isLegalICmpImmediate(C-1)) {
3108 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3109 RHS = DAG.getConstant(C-1, MVT::i32);
3114 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3115 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3116 RHS = DAG.getConstant(C+1, MVT::i32);
3121 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3122 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3123 RHS = DAG.getConstant(C+1, MVT::i32);
3130 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3131 ARMISD::NodeType CompareType;
3134 CompareType = ARMISD::CMP;
3139 CompareType = ARMISD::CMPZ;
3142 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3143 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3146 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3148 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3151 if (!isFloatingPointZero(RHS))
3152 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3154 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3155 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3158 /// duplicateCmp - Glue values can have only one use, so this function
3159 /// duplicates a comparison node.
3161 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3162 unsigned Opc = Cmp.getOpcode();
3164 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3165 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3167 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3168 Cmp = Cmp.getOperand(0);
3169 Opc = Cmp.getOpcode();
3170 if (Opc == ARMISD::CMPFP)
3171 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3173 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3174 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3176 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3179 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3180 SDValue Cond = Op.getOperand(0);
3181 SDValue SelectTrue = Op.getOperand(1);
3182 SDValue SelectFalse = Op.getOperand(2);
3187 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3188 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3190 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3191 const ConstantSDNode *CMOVTrue =
3192 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3193 const ConstantSDNode *CMOVFalse =
3194 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3196 if (CMOVTrue && CMOVFalse) {
3197 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3198 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3202 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3204 False = SelectFalse;
3205 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3210 if (True.getNode() && False.getNode()) {
3211 EVT VT = Op.getValueType();
3212 SDValue ARMcc = Cond.getOperand(2);
3213 SDValue CCR = Cond.getOperand(3);
3214 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3215 assert(True.getValueType() == VT);
3216 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3221 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3222 // undefined bits before doing a full-word comparison with zero.
3223 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3224 DAG.getConstant(1, Cond.getValueType()));
3226 return DAG.getSelectCC(dl, Cond,
3227 DAG.getConstant(0, Cond.getValueType()),
3228 SelectTrue, SelectFalse, ISD::SETNE);
3231 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3232 if (CC == ISD::SETNE)
3234 return ISD::getSetCCInverse(CC, true);
3237 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3238 bool &swpCmpOps, bool &swpVselOps) {
3239 // Start by selecting the GE condition code for opcodes that return true for
3241 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3243 CondCode = ARMCC::GE;
3245 // and GT for opcodes that return false for 'equality'.
3246 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3248 CondCode = ARMCC::GT;
3250 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3251 // to swap the compare operands.
3252 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3256 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3257 // If we have an unordered opcode, we need to swap the operands to the VSEL
3258 // instruction (effectively negating the condition).
3260 // This also has the effect of swapping which one of 'less' or 'greater'
3261 // returns true, so we also swap the compare operands. It also switches
3262 // whether we return true for 'equality', so we compensate by picking the
3263 // opposite condition code to our original choice.
3264 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3265 CC == ISD::SETUGT) {
3266 swpCmpOps = !swpCmpOps;
3267 swpVselOps = !swpVselOps;
3268 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3271 // 'ordered' is 'anything but unordered', so use the VS condition code and
3272 // swap the VSEL operands.
3273 if (CC == ISD::SETO) {
3274 CondCode = ARMCC::VS;
3278 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3279 // code and swap the VSEL operands.
3280 if (CC == ISD::SETUNE) {
3281 CondCode = ARMCC::EQ;
3286 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3287 EVT VT = Op.getValueType();
3288 SDValue LHS = Op.getOperand(0);
3289 SDValue RHS = Op.getOperand(1);
3290 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3291 SDValue TrueVal = Op.getOperand(2);
3292 SDValue FalseVal = Op.getOperand(3);
3295 if (LHS.getValueType() == MVT::i32) {
3296 // Try to generate VSEL on ARMv8.
3297 // The VSEL instruction can't use all the usual ARM condition
3298 // codes: it only has two bits to select the condition code, so it's
3299 // constrained to use only GE, GT, VS and EQ.
3301 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3302 // swap the operands of the previous compare instruction (effectively
3303 // inverting the compare condition, swapping 'less' and 'greater') and
3304 // sometimes need to swap the operands to the VSEL (which inverts the
3305 // condition in the sense of firing whenever the previous condition didn't)
3306 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3307 TrueVal.getValueType() == MVT::f64)) {
3308 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3309 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3310 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3311 CC = getInverseCCForVSEL(CC);
3312 std::swap(TrueVal, FalseVal);
3317 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3318 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3319 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3323 ARMCC::CondCodes CondCode, CondCode2;
3324 FPCCToARMCC(CC, CondCode, CondCode2);
3326 // Try to generate VSEL on ARMv8.
3327 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3328 TrueVal.getValueType() == MVT::f64)) {
3329 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3330 // same operands, as follows:
3331 // c = fcmp [ogt, olt, ugt, ult] a, b
3333 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3334 // handled differently than the original code sequence.
3335 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3337 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3338 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3339 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3340 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3343 bool swpCmpOps = false;
3344 bool swpVselOps = false;
3345 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3347 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3348 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3350 std::swap(LHS, RHS);
3352 std::swap(TrueVal, FalseVal);
3356 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3357 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3358 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3359 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3361 if (CondCode2 != ARMCC::AL) {
3362 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3363 // FIXME: Needs another CMP because flag can have but one use.
3364 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3365 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3366 Result, TrueVal, ARMcc2, CCR, Cmp2);
3371 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3372 /// to morph to an integer compare sequence.
3373 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3374 const ARMSubtarget *Subtarget) {
3375 SDNode *N = Op.getNode();
3376 if (!N->hasOneUse())
3377 // Otherwise it requires moving the value from fp to integer registers.
3379 if (!N->getNumValues())
3381 EVT VT = Op.getValueType();
3382 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3383 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3384 // vmrs are very slow, e.g. cortex-a8.
3387 if (isFloatingPointZero(Op)) {
3391 return ISD::isNormalLoad(N);
3394 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3395 if (isFloatingPointZero(Op))
3396 return DAG.getConstant(0, MVT::i32);
3398 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3399 return DAG.getLoad(MVT::i32, SDLoc(Op),
3400 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3401 Ld->isVolatile(), Ld->isNonTemporal(),
3402 Ld->isInvariant(), Ld->getAlignment());
3404 llvm_unreachable("Unknown VFP cmp argument!");
3407 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3408 SDValue &RetVal1, SDValue &RetVal2) {
3409 if (isFloatingPointZero(Op)) {
3410 RetVal1 = DAG.getConstant(0, MVT::i32);
3411 RetVal2 = DAG.getConstant(0, MVT::i32);
3415 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3416 SDValue Ptr = Ld->getBasePtr();
3417 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3418 Ld->getChain(), Ptr,
3419 Ld->getPointerInfo(),
3420 Ld->isVolatile(), Ld->isNonTemporal(),
3421 Ld->isInvariant(), Ld->getAlignment());
3423 EVT PtrType = Ptr.getValueType();
3424 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3425 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3426 PtrType, Ptr, DAG.getConstant(4, PtrType));
3427 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3428 Ld->getChain(), NewPtr,
3429 Ld->getPointerInfo().getWithOffset(4),
3430 Ld->isVolatile(), Ld->isNonTemporal(),
3431 Ld->isInvariant(), NewAlign);
3435 llvm_unreachable("Unknown VFP cmp argument!");
3438 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3439 /// f32 and even f64 comparisons to integer ones.
3441 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3442 SDValue Chain = Op.getOperand(0);
3443 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3444 SDValue LHS = Op.getOperand(2);
3445 SDValue RHS = Op.getOperand(3);
3446 SDValue Dest = Op.getOperand(4);
3449 bool LHSSeenZero = false;
3450 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3451 bool RHSSeenZero = false;
3452 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3453 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3454 // If unsafe fp math optimization is enabled and there are no other uses of
3455 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3456 // to an integer comparison.
3457 if (CC == ISD::SETOEQ)
3459 else if (CC == ISD::SETUNE)
3462 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3464 if (LHS.getValueType() == MVT::f32) {
3465 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3466 bitcastf32Toi32(LHS, DAG), Mask);
3467 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3468 bitcastf32Toi32(RHS, DAG), Mask);
3469 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3470 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3471 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3472 Chain, Dest, ARMcc, CCR, Cmp);
3477 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3478 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3479 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3480 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3481 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3482 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3483 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3484 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3485 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3491 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3492 SDValue Chain = Op.getOperand(0);
3493 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3494 SDValue LHS = Op.getOperand(2);
3495 SDValue RHS = Op.getOperand(3);
3496 SDValue Dest = Op.getOperand(4);
3499 if (LHS.getValueType() == MVT::i32) {
3501 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3502 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3503 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3504 Chain, Dest, ARMcc, CCR, Cmp);
3507 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3509 if (getTargetMachine().Options.UnsafeFPMath &&
3510 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3511 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3512 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3513 if (Result.getNode())
3517 ARMCC::CondCodes CondCode, CondCode2;
3518 FPCCToARMCC(CC, CondCode, CondCode2);
3520 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3521 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3522 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3523 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3524 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3525 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3526 if (CondCode2 != ARMCC::AL) {
3527 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3528 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3529 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3534 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3535 SDValue Chain = Op.getOperand(0);
3536 SDValue Table = Op.getOperand(1);
3537 SDValue Index = Op.getOperand(2);
3540 EVT PTy = getPointerTy();
3541 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3542 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3543 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3544 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3545 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3546 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3547 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3548 if (Subtarget->isThumb2()) {
3549 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3550 // which does another jump to the destination. This also makes it easier
3551 // to translate it to TBB / TBH later.
3552 // FIXME: This might not work if the function is extremely large.
3553 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3554 Addr, Op.getOperand(2), JTI, UId);
3556 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3557 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3558 MachinePointerInfo::getJumpTable(),
3559 false, false, false, 0);
3560 Chain = Addr.getValue(1);
3561 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3562 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3564 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3565 MachinePointerInfo::getJumpTable(),
3566 false, false, false, 0);
3567 Chain = Addr.getValue(1);
3568 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3572 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3573 EVT VT = Op.getValueType();
3576 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3577 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3579 return DAG.UnrollVectorOp(Op.getNode());
3582 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3583 "Invalid type for custom lowering!");
3584 if (VT != MVT::v4i16)
3585 return DAG.UnrollVectorOp(Op.getNode());
3587 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3588 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3591 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3592 EVT VT = Op.getValueType();
3594 return LowerVectorFP_TO_INT(Op, DAG);
3599 switch (Op.getOpcode()) {
3600 default: llvm_unreachable("Invalid opcode!");
3601 case ISD::FP_TO_SINT:
3602 Opc = ARMISD::FTOSI;
3604 case ISD::FP_TO_UINT:
3605 Opc = ARMISD::FTOUI;
3608 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3609 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3612 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3613 EVT VT = Op.getValueType();
3616 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3617 if (VT.getVectorElementType() == MVT::f32)
3619 return DAG.UnrollVectorOp(Op.getNode());
3622 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3623 "Invalid type for custom lowering!");
3624 if (VT != MVT::v4f32)
3625 return DAG.UnrollVectorOp(Op.getNode());
3629 switch (Op.getOpcode()) {
3630 default: llvm_unreachable("Invalid opcode!");
3631 case ISD::SINT_TO_FP:
3632 CastOpc = ISD::SIGN_EXTEND;
3633 Opc = ISD::SINT_TO_FP;
3635 case ISD::UINT_TO_FP:
3636 CastOpc = ISD::ZERO_EXTEND;
3637 Opc = ISD::UINT_TO_FP;
3641 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3642 return DAG.getNode(Opc, dl, VT, Op);
3645 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3646 EVT VT = Op.getValueType();
3648 return LowerVectorINT_TO_FP(Op, DAG);
3653 switch (Op.getOpcode()) {
3654 default: llvm_unreachable("Invalid opcode!");
3655 case ISD::SINT_TO_FP:
3656 Opc = ARMISD::SITOF;
3658 case ISD::UINT_TO_FP:
3659 Opc = ARMISD::UITOF;
3663 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3664 return DAG.getNode(Opc, dl, VT, Op);
3667 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3668 // Implement fcopysign with a fabs and a conditional fneg.
3669 SDValue Tmp0 = Op.getOperand(0);
3670 SDValue Tmp1 = Op.getOperand(1);
3672 EVT VT = Op.getValueType();
3673 EVT SrcVT = Tmp1.getValueType();
3674 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3675 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3676 bool UseNEON = !InGPR && Subtarget->hasNEON();
3679 // Use VBSL to copy the sign bit.
3680 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3681 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3682 DAG.getTargetConstant(EncodedVal, MVT::i32));
3683 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3685 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3686 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3687 DAG.getConstant(32, MVT::i32));
3688 else /*if (VT == MVT::f32)*/
3689 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3690 if (SrcVT == MVT::f32) {
3691 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3693 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3694 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3695 DAG.getConstant(32, MVT::i32));
3696 } else if (VT == MVT::f32)
3697 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3698 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3699 DAG.getConstant(32, MVT::i32));
3700 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3701 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3703 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3705 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3706 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3707 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3709 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3710 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3711 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3712 if (VT == MVT::f32) {
3713 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3714 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3715 DAG.getConstant(0, MVT::i32));
3717 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3723 // Bitcast operand 1 to i32.
3724 if (SrcVT == MVT::f64)
3725 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3726 &Tmp1, 1).getValue(1);
3727 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3729 // Or in the signbit with integer operations.
3730 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3731 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3732 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3733 if (VT == MVT::f32) {
3734 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3735 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3736 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3737 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3740 // f64: Or the high part with signbit and then combine two parts.
3741 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3743 SDValue Lo = Tmp0.getValue(0);
3744 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3745 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3746 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3749 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3750 MachineFunction &MF = DAG.getMachineFunction();
3751 MachineFrameInfo *MFI = MF.getFrameInfo();
3752 MFI->setReturnAddressIsTaken(true);
3754 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3757 EVT VT = Op.getValueType();
3759 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3761 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3762 SDValue Offset = DAG.getConstant(4, MVT::i32);
3763 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3764 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3765 MachinePointerInfo(), false, false, false, 0);
3768 // Return LR, which contains the return address. Mark it an implicit live-in.
3769 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3770 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3773 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3774 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3775 MFI->setFrameAddressIsTaken(true);
3777 EVT VT = Op.getValueType();
3778 SDLoc dl(Op); // FIXME probably not meaningful
3779 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3780 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetMachO())
3781 ? ARM::R7 : ARM::R11;
3782 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3784 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3785 MachinePointerInfo(),
3786 false, false, false, 0);
3790 /// ExpandBITCAST - If the target supports VFP, this function is called to
3791 /// expand a bit convert where either the source or destination type is i64 to
3792 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3793 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3794 /// vectors), since the legalizer won't know what to do with that.
3795 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3796 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3798 SDValue Op = N->getOperand(0);
3800 // This function is only supposed to be called for i64 types, either as the
3801 // source or destination of the bit convert.
3802 EVT SrcVT = Op.getValueType();
3803 EVT DstVT = N->getValueType(0);
3804 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3805 "ExpandBITCAST called for non-i64 type");
3807 // Turn i64->f64 into VMOVDRR.
3808 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3809 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3810 DAG.getConstant(0, MVT::i32));
3811 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3812 DAG.getConstant(1, MVT::i32));
3813 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3814 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3817 // Turn f64->i64 into VMOVRRD.
3818 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3819 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3820 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3821 // Merge the pieces into a single i64 value.
3822 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3828 /// getZeroVector - Returns a vector of specified type with all zero elements.
3829 /// Zero vectors are used to represent vector negation and in those cases
3830 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3831 /// not support i64 elements, so sometimes the zero vectors will need to be
3832 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3834 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3835 assert(VT.isVector() && "Expected a vector type");
3836 // The canonical modified immediate encoding of a zero vector is....0!
3837 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3838 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3839 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3840 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3843 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3844 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3845 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3846 SelectionDAG &DAG) const {
3847 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3848 EVT VT = Op.getValueType();
3849 unsigned VTBits = VT.getSizeInBits();
3851 SDValue ShOpLo = Op.getOperand(0);
3852 SDValue ShOpHi = Op.getOperand(1);
3853 SDValue ShAmt = Op.getOperand(2);
3855 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3857 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3859 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3860 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3861 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3862 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3863 DAG.getConstant(VTBits, MVT::i32));
3864 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3865 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3866 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3868 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3869 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3871 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3872 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3875 SDValue Ops[2] = { Lo, Hi };
3876 return DAG.getMergeValues(Ops, 2, dl);
3879 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3880 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3881 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3882 SelectionDAG &DAG) const {
3883 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3884 EVT VT = Op.getValueType();
3885 unsigned VTBits = VT.getSizeInBits();
3887 SDValue ShOpLo = Op.getOperand(0);
3888 SDValue ShOpHi = Op.getOperand(1);
3889 SDValue ShAmt = Op.getOperand(2);
3892 assert(Op.getOpcode() == ISD::SHL_PARTS);
3893 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3894 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3895 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3896 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3897 DAG.getConstant(VTBits, MVT::i32));
3898 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3899 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3901 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3902 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3903 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3905 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3906 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3909 SDValue Ops[2] = { Lo, Hi };
3910 return DAG.getMergeValues(Ops, 2, dl);
3913 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3914 SelectionDAG &DAG) const {
3915 // The rounding mode is in bits 23:22 of the FPSCR.
3916 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3917 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3918 // so that the shift + and get folded into a bitfield extract.
3920 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3921 DAG.getConstant(Intrinsic::arm_get_fpscr,
3923 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3924 DAG.getConstant(1U << 22, MVT::i32));
3925 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3926 DAG.getConstant(22, MVT::i32));
3927 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3928 DAG.getConstant(3, MVT::i32));
3931 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3932 const ARMSubtarget *ST) {
3933 EVT VT = N->getValueType(0);
3936 if (!ST->hasV6T2Ops())
3939 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3940 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3943 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3944 /// for each 16-bit element from operand, repeated. The basic idea is to
3945 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3947 /// Trace for v4i16:
3948 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3949 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3950 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3951 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3952 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3953 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3954 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3955 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3956 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3957 EVT VT = N->getValueType(0);
3960 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3961 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3962 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3963 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3964 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3965 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3968 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3969 /// bit-count for each 16-bit element from the operand. We need slightly
3970 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3971 /// 64/128-bit registers.
3973 /// Trace for v4i16:
3974 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3975 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3976 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3977 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3978 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3979 EVT VT = N->getValueType(0);
3982 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3983 if (VT.is64BitVector()) {
3984 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3985 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3986 DAG.getIntPtrConstant(0));
3988 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3989 BitCounts, DAG.getIntPtrConstant(0));
3990 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3994 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3995 /// bit-count for each 32-bit element from the operand. The idea here is
3996 /// to split the vector into 16-bit elements, leverage the 16-bit count
3997 /// routine, and then combine the results.
3999 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4000 /// input = [v0 v1 ] (vi: 32-bit elements)
4001 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4002 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4003 /// vrev: N0 = [k1 k0 k3 k2 ]
4005 /// N1 =+[k1 k0 k3 k2 ]
4007 /// N2 =+[k1 k3 k0 k2 ]
4009 /// Extended =+[k1 k3 k0 k2 ]
4011 /// Extracted=+[k1 k3 ]
4013 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4014 EVT VT = N->getValueType(0);
4017 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4019 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4020 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4021 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4022 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4023 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4025 if (VT.is64BitVector()) {
4026 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4027 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4028 DAG.getIntPtrConstant(0));
4030 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4031 DAG.getIntPtrConstant(0));
4032 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4036 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4037 const ARMSubtarget *ST) {
4038 EVT VT = N->getValueType(0);
4040 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4041 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4042 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4043 "Unexpected type for custom ctpop lowering");
4045 if (VT.getVectorElementType() == MVT::i32)
4046 return lowerCTPOP32BitElements(N, DAG);
4048 return lowerCTPOP16BitElements(N, DAG);
4051 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4052 const ARMSubtarget *ST) {
4053 EVT VT = N->getValueType(0);
4059 // Lower vector shifts on NEON to use VSHL.
4060 assert(ST->hasNEON() && "unexpected vector shift");
4062 // Left shifts translate directly to the vshiftu intrinsic.
4063 if (N->getOpcode() == ISD::SHL)
4064 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4065 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4066 N->getOperand(0), N->getOperand(1));
4068 assert((N->getOpcode() == ISD::SRA ||
4069 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4071 // NEON uses the same intrinsics for both left and right shifts. For
4072 // right shifts, the shift amounts are negative, so negate the vector of
4074 EVT ShiftVT = N->getOperand(1).getValueType();
4075 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4076 getZeroVector(ShiftVT, DAG, dl),
4078 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4079 Intrinsic::arm_neon_vshifts :
4080 Intrinsic::arm_neon_vshiftu);
4081 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4082 DAG.getConstant(vshiftInt, MVT::i32),
4083 N->getOperand(0), NegatedCount);
4086 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4087 const ARMSubtarget *ST) {
4088 EVT VT = N->getValueType(0);
4091 // We can get here for a node like i32 = ISD::SHL i32, i64
4095 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4096 "Unknown shift to lower!");
4098 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4099 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4100 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4103 // If we are in thumb mode, we don't have RRX.
4104 if (ST->isThumb1Only()) return SDValue();
4106 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4107 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4108 DAG.getConstant(0, MVT::i32));
4109 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4110 DAG.getConstant(1, MVT::i32));
4112 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4113 // captures the result into a carry flag.
4114 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4115 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
4117 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4118 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4120 // Merge the pieces into a single i64 value.
4121 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4124 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4125 SDValue TmpOp0, TmpOp1;
4126 bool Invert = false;
4130 SDValue Op0 = Op.getOperand(0);
4131 SDValue Op1 = Op.getOperand(1);
4132 SDValue CC = Op.getOperand(2);
4133 EVT VT = Op.getValueType();
4134 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4137 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4138 switch (SetCCOpcode) {
4139 default: llvm_unreachable("Illegal FP comparison");
4141 case ISD::SETNE: Invert = true; // Fallthrough
4143 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4145 case ISD::SETLT: Swap = true; // Fallthrough
4147 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4149 case ISD::SETLE: Swap = true; // Fallthrough
4151 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4152 case ISD::SETUGE: Swap = true; // Fallthrough
4153 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4154 case ISD::SETUGT: Swap = true; // Fallthrough
4155 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4156 case ISD::SETUEQ: Invert = true; // Fallthrough
4158 // Expand this to (OLT | OGT).
4162 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4163 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4165 case ISD::SETUO: Invert = true; // Fallthrough
4167 // Expand this to (OLT | OGE).
4171 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4172 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4176 // Integer comparisons.
4177 switch (SetCCOpcode) {
4178 default: llvm_unreachable("Illegal integer comparison");
4179 case ISD::SETNE: Invert = true;
4180 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4181 case ISD::SETLT: Swap = true;
4182 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4183 case ISD::SETLE: Swap = true;
4184 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4185 case ISD::SETULT: Swap = true;
4186 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4187 case ISD::SETULE: Swap = true;
4188 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4191 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4192 if (Opc == ARMISD::VCEQ) {
4195 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4197 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4200 // Ignore bitconvert.
4201 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4202 AndOp = AndOp.getOperand(0);
4204 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4206 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4207 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4214 std::swap(Op0, Op1);
4216 // If one of the operands is a constant vector zero, attempt to fold the
4217 // comparison to a specialized compare-against-zero form.
4219 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4221 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4222 if (Opc == ARMISD::VCGE)
4223 Opc = ARMISD::VCLEZ;
4224 else if (Opc == ARMISD::VCGT)
4225 Opc = ARMISD::VCLTZ;
4230 if (SingleOp.getNode()) {
4233 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4235 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4237 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4239 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4241 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4243 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4246 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4250 Result = DAG.getNOT(dl, Result, VT);
4255 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4256 /// valid vector constant for a NEON instruction with a "modified immediate"
4257 /// operand (e.g., VMOV). If so, return the encoded value.
4258 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4259 unsigned SplatBitSize, SelectionDAG &DAG,
4260 EVT &VT, bool is128Bits, NEONModImmType type) {
4261 unsigned OpCmode, Imm;
4263 // SplatBitSize is set to the smallest size that splats the vector, so a
4264 // zero vector will always have SplatBitSize == 8. However, NEON modified
4265 // immediate instructions others than VMOV do not support the 8-bit encoding
4266 // of a zero vector, and the default encoding of zero is supposed to be the
4271 switch (SplatBitSize) {
4273 if (type != VMOVModImm)
4275 // Any 1-byte value is OK. Op=0, Cmode=1110.
4276 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4279 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4283 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4284 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4285 if ((SplatBits & ~0xff) == 0) {
4286 // Value = 0x00nn: Op=x, Cmode=100x.
4291 if ((SplatBits & ~0xff00) == 0) {
4292 // Value = 0xnn00: Op=x, Cmode=101x.
4294 Imm = SplatBits >> 8;
4300 // NEON's 32-bit VMOV supports splat values where:
4301 // * only one byte is nonzero, or
4302 // * the least significant byte is 0xff and the second byte is nonzero, or
4303 // * the least significant 2 bytes are 0xff and the third is nonzero.
4304 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4305 if ((SplatBits & ~0xff) == 0) {
4306 // Value = 0x000000nn: Op=x, Cmode=000x.
4311 if ((SplatBits & ~0xff00) == 0) {
4312 // Value = 0x0000nn00: Op=x, Cmode=001x.
4314 Imm = SplatBits >> 8;
4317 if ((SplatBits & ~0xff0000) == 0) {
4318 // Value = 0x00nn0000: Op=x, Cmode=010x.
4320 Imm = SplatBits >> 16;
4323 if ((SplatBits & ~0xff000000) == 0) {
4324 // Value = 0xnn000000: Op=x, Cmode=011x.
4326 Imm = SplatBits >> 24;
4330 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4331 if (type == OtherModImm) return SDValue();
4333 if ((SplatBits & ~0xffff) == 0 &&
4334 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4335 // Value = 0x0000nnff: Op=x, Cmode=1100.
4337 Imm = SplatBits >> 8;
4342 if ((SplatBits & ~0xffffff) == 0 &&
4343 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4344 // Value = 0x00nnffff: Op=x, Cmode=1101.
4346 Imm = SplatBits >> 16;
4347 SplatBits |= 0xffff;
4351 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4352 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4353 // VMOV.I32. A (very) minor optimization would be to replicate the value
4354 // and fall through here to test for a valid 64-bit splat. But, then the
4355 // caller would also need to check and handle the change in size.
4359 if (type != VMOVModImm)
4361 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4362 uint64_t BitMask = 0xff;
4364 unsigned ImmMask = 1;
4366 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4367 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4370 } else if ((SplatBits & BitMask) != 0) {
4376 // Op=1, Cmode=1110.
4379 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4384 llvm_unreachable("unexpected size for isNEONModifiedImm");
4387 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4388 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4391 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4392 const ARMSubtarget *ST) const {
4396 bool IsDouble = Op.getValueType() == MVT::f64;
4397 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4399 // Try splatting with a VMOV.f32...
4400 APFloat FPVal = CFP->getValueAPF();
4401 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4404 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4405 // We have code in place to select a valid ConstantFP already, no need to
4410 // It's a float and we are trying to use NEON operations where
4411 // possible. Lower it to a splat followed by an extract.
4413 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4414 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4416 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4417 DAG.getConstant(0, MVT::i32));
4420 // The rest of our options are NEON only, make sure that's allowed before
4422 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4426 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4428 // It wouldn't really be worth bothering for doubles except for one very
4429 // important value, which does happen to match: 0.0. So make sure we don't do
4431 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4434 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4435 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4437 if (NewVal != SDValue()) {
4439 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4442 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4444 // It's a float: cast and extract a vector element.
4445 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4447 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4448 DAG.getConstant(0, MVT::i32));
4451 // Finally, try a VMVN.i32
4452 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4454 if (NewVal != SDValue()) {
4456 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4459 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4461 // It's a float: cast and extract a vector element.
4462 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4464 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4465 DAG.getConstant(0, MVT::i32));
4471 // check if an VEXT instruction can handle the shuffle mask when the
4472 // vector sources of the shuffle are the same.
4473 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4474 unsigned NumElts = VT.getVectorNumElements();
4476 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4482 // If this is a VEXT shuffle, the immediate value is the index of the first
4483 // element. The other shuffle indices must be the successive elements after
4485 unsigned ExpectedElt = Imm;
4486 for (unsigned i = 1; i < NumElts; ++i) {
4487 // Increment the expected index. If it wraps around, just follow it
4488 // back to index zero and keep going.
4490 if (ExpectedElt == NumElts)
4493 if (M[i] < 0) continue; // ignore UNDEF indices
4494 if (ExpectedElt != static_cast<unsigned>(M[i]))
4502 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4503 bool &ReverseVEXT, unsigned &Imm) {
4504 unsigned NumElts = VT.getVectorNumElements();
4505 ReverseVEXT = false;
4507 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4513 // If this is a VEXT shuffle, the immediate value is the index of the first
4514 // element. The other shuffle indices must be the successive elements after
4516 unsigned ExpectedElt = Imm;
4517 for (unsigned i = 1; i < NumElts; ++i) {
4518 // Increment the expected index. If it wraps around, it may still be
4519 // a VEXT but the source vectors must be swapped.
4521 if (ExpectedElt == NumElts * 2) {
4526 if (M[i] < 0) continue; // ignore UNDEF indices
4527 if (ExpectedElt != static_cast<unsigned>(M[i]))
4531 // Adjust the index value if the source operands will be swapped.
4538 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4539 /// instruction with the specified blocksize. (The order of the elements
4540 /// within each block of the vector is reversed.)
4541 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4542 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4543 "Only possible block sizes for VREV are: 16, 32, 64");
4545 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4549 unsigned NumElts = VT.getVectorNumElements();
4550 unsigned BlockElts = M[0] + 1;
4551 // If the first shuffle index is UNDEF, be optimistic.
4553 BlockElts = BlockSize / EltSz;
4555 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4558 for (unsigned i = 0; i < NumElts; ++i) {
4559 if (M[i] < 0) continue; // ignore UNDEF indices
4560 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4567 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4568 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4569 // range, then 0 is placed into the resulting vector. So pretty much any mask
4570 // of 8 elements can work here.
4571 return VT == MVT::v8i8 && M.size() == 8;
4574 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4575 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4579 unsigned NumElts = VT.getVectorNumElements();
4580 WhichResult = (M[0] == 0 ? 0 : 1);
4581 for (unsigned i = 0; i < NumElts; i += 2) {
4582 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4583 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4589 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4590 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4591 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4592 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4593 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4597 unsigned NumElts = VT.getVectorNumElements();
4598 WhichResult = (M[0] == 0 ? 0 : 1);
4599 for (unsigned i = 0; i < NumElts; i += 2) {
4600 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4601 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4607 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4608 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4612 unsigned NumElts = VT.getVectorNumElements();
4613 WhichResult = (M[0] == 0 ? 0 : 1);
4614 for (unsigned i = 0; i != NumElts; ++i) {
4615 if (M[i] < 0) continue; // ignore UNDEF indices
4616 if ((unsigned) M[i] != 2 * i + WhichResult)
4620 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4621 if (VT.is64BitVector() && EltSz == 32)
4627 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4628 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4629 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4630 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4631 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4635 unsigned Half = VT.getVectorNumElements() / 2;
4636 WhichResult = (M[0] == 0 ? 0 : 1);
4637 for (unsigned j = 0; j != 2; ++j) {
4638 unsigned Idx = WhichResult;
4639 for (unsigned i = 0; i != Half; ++i) {
4640 int MIdx = M[i + j * Half];
4641 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4647 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4648 if (VT.is64BitVector() && EltSz == 32)
4654 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4655 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4659 unsigned NumElts = VT.getVectorNumElements();
4660 WhichResult = (M[0] == 0 ? 0 : 1);
4661 unsigned Idx = WhichResult * NumElts / 2;
4662 for (unsigned i = 0; i != NumElts; i += 2) {
4663 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4664 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4669 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4670 if (VT.is64BitVector() && EltSz == 32)
4676 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4677 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4678 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4679 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4680 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4684 unsigned NumElts = VT.getVectorNumElements();
4685 WhichResult = (M[0] == 0 ? 0 : 1);
4686 unsigned Idx = WhichResult * NumElts / 2;
4687 for (unsigned i = 0; i != NumElts; i += 2) {
4688 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4689 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4694 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4695 if (VT.is64BitVector() && EltSz == 32)
4701 /// \return true if this is a reverse operation on an vector.
4702 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4703 unsigned NumElts = VT.getVectorNumElements();
4704 // Make sure the mask has the right size.
4705 if (NumElts != M.size())
4708 // Look for <15, ..., 3, -1, 1, 0>.
4709 for (unsigned i = 0; i != NumElts; ++i)
4710 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4716 // If N is an integer constant that can be moved into a register in one
4717 // instruction, return an SDValue of such a constant (will become a MOV
4718 // instruction). Otherwise return null.
4719 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4720 const ARMSubtarget *ST, SDLoc dl) {
4722 if (!isa<ConstantSDNode>(N))
4724 Val = cast<ConstantSDNode>(N)->getZExtValue();
4726 if (ST->isThumb1Only()) {
4727 if (Val <= 255 || ~Val <= 255)
4728 return DAG.getConstant(Val, MVT::i32);
4730 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4731 return DAG.getConstant(Val, MVT::i32);
4736 // If this is a case we can't handle, return null and let the default
4737 // expansion code take care of it.
4738 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4739 const ARMSubtarget *ST) const {
4740 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4742 EVT VT = Op.getValueType();
4744 APInt SplatBits, SplatUndef;
4745 unsigned SplatBitSize;
4747 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4748 if (SplatBitSize <= 64) {
4749 // Check if an immediate VMOV works.
4751 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4752 SplatUndef.getZExtValue(), SplatBitSize,
4753 DAG, VmovVT, VT.is128BitVector(),
4755 if (Val.getNode()) {
4756 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4757 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4760 // Try an immediate VMVN.
4761 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4762 Val = isNEONModifiedImm(NegatedImm,
4763 SplatUndef.getZExtValue(), SplatBitSize,
4764 DAG, VmovVT, VT.is128BitVector(),
4766 if (Val.getNode()) {
4767 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4768 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4771 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4772 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4773 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4775 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4776 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4782 // Scan through the operands to see if only one value is used.
4784 // As an optimisation, even if more than one value is used it may be more
4785 // profitable to splat with one value then change some lanes.
4787 // Heuristically we decide to do this if the vector has a "dominant" value,
4788 // defined as splatted to more than half of the lanes.
4789 unsigned NumElts = VT.getVectorNumElements();
4790 bool isOnlyLowElement = true;
4791 bool usesOnlyOneValue = true;
4792 bool hasDominantValue = false;
4793 bool isConstant = true;
4795 // Map of the number of times a particular SDValue appears in the
4797 DenseMap<SDValue, unsigned> ValueCounts;
4799 for (unsigned i = 0; i < NumElts; ++i) {
4800 SDValue V = Op.getOperand(i);
4801 if (V.getOpcode() == ISD::UNDEF)
4804 isOnlyLowElement = false;
4805 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4808 ValueCounts.insert(std::make_pair(V, 0));
4809 unsigned &Count = ValueCounts[V];
4811 // Is this value dominant? (takes up more than half of the lanes)
4812 if (++Count > (NumElts / 2)) {
4813 hasDominantValue = true;
4817 if (ValueCounts.size() != 1)
4818 usesOnlyOneValue = false;
4819 if (!Value.getNode() && ValueCounts.size() > 0)
4820 Value = ValueCounts.begin()->first;
4822 if (ValueCounts.size() == 0)
4823 return DAG.getUNDEF(VT);
4825 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4826 // Keep going if we are hitting this case.
4827 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4828 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4830 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4832 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4833 // i32 and try again.
4834 if (hasDominantValue && EltSize <= 32) {
4838 // If we are VDUPing a value that comes directly from a vector, that will
4839 // cause an unnecessary move to and from a GPR, where instead we could
4840 // just use VDUPLANE. We can only do this if the lane being extracted
4841 // is at a constant index, as the VDUP from lane instructions only have
4842 // constant-index forms.
4843 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4844 isa<ConstantSDNode>(Value->getOperand(1))) {
4845 // We need to create a new undef vector to use for the VDUPLANE if the
4846 // size of the vector from which we get the value is different than the
4847 // size of the vector that we need to create. We will insert the element
4848 // such that the register coalescer will remove unnecessary copies.
4849 if (VT != Value->getOperand(0).getValueType()) {
4850 ConstantSDNode *constIndex;
4851 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4852 assert(constIndex && "The index is not a constant!");
4853 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4854 VT.getVectorNumElements();
4855 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4856 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4857 Value, DAG.getConstant(index, MVT::i32)),
4858 DAG.getConstant(index, MVT::i32));
4860 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4861 Value->getOperand(0), Value->getOperand(1));
4863 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4865 if (!usesOnlyOneValue) {
4866 // The dominant value was splatted as 'N', but we now have to insert
4867 // all differing elements.
4868 for (unsigned I = 0; I < NumElts; ++I) {
4869 if (Op.getOperand(I) == Value)
4871 SmallVector<SDValue, 3> Ops;
4873 Ops.push_back(Op.getOperand(I));
4874 Ops.push_back(DAG.getConstant(I, MVT::i32));
4875 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4880 if (VT.getVectorElementType().isFloatingPoint()) {
4881 SmallVector<SDValue, 8> Ops;
4882 for (unsigned i = 0; i < NumElts; ++i)
4883 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4885 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4886 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4887 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4889 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4891 if (usesOnlyOneValue) {
4892 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4893 if (isConstant && Val.getNode())
4894 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4898 // If all elements are constants and the case above didn't get hit, fall back
4899 // to the default expansion, which will generate a load from the constant
4904 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4906 SDValue shuffle = ReconstructShuffle(Op, DAG);
4907 if (shuffle != SDValue())
4911 // Vectors with 32- or 64-bit elements can be built by directly assigning
4912 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4913 // will be legalized.
4914 if (EltSize >= 32) {
4915 // Do the expansion with floating-point types, since that is what the VFP
4916 // registers are defined to use, and since i64 is not legal.
4917 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4918 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4919 SmallVector<SDValue, 8> Ops;
4920 for (unsigned i = 0; i < NumElts; ++i)
4921 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4922 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4923 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4926 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4927 // know the default expansion would otherwise fall back on something even
4928 // worse. For a vector with one or two non-undef values, that's
4929 // scalar_to_vector for the elements followed by a shuffle (provided the
4930 // shuffle is valid for the target) and materialization element by element
4931 // on the stack followed by a load for everything else.
4932 if (!isConstant && !usesOnlyOneValue) {
4933 SDValue Vec = DAG.getUNDEF(VT);
4934 for (unsigned i = 0 ; i < NumElts; ++i) {
4935 SDValue V = Op.getOperand(i);
4936 if (V.getOpcode() == ISD::UNDEF)
4938 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
4939 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
4947 // Gather data to see if the operation can be modelled as a
4948 // shuffle in combination with VEXTs.
4949 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4950 SelectionDAG &DAG) const {
4952 EVT VT = Op.getValueType();
4953 unsigned NumElts = VT.getVectorNumElements();
4955 SmallVector<SDValue, 2> SourceVecs;
4956 SmallVector<unsigned, 2> MinElts;
4957 SmallVector<unsigned, 2> MaxElts;
4959 for (unsigned i = 0; i < NumElts; ++i) {
4960 SDValue V = Op.getOperand(i);
4961 if (V.getOpcode() == ISD::UNDEF)
4963 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4964 // A shuffle can only come from building a vector from various
4965 // elements of other vectors.
4967 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4968 VT.getVectorElementType()) {
4969 // This code doesn't know how to handle shuffles where the vector
4970 // element types do not match (this happens because type legalization
4971 // promotes the return type of EXTRACT_VECTOR_ELT).
4972 // FIXME: It might be appropriate to extend this code to handle
4973 // mismatched types.
4977 // Record this extraction against the appropriate vector if possible...
4978 SDValue SourceVec = V.getOperand(0);
4979 // If the element number isn't a constant, we can't effectively
4980 // analyze what's going on.
4981 if (!isa<ConstantSDNode>(V.getOperand(1)))
4983 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4984 bool FoundSource = false;
4985 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4986 if (SourceVecs[j] == SourceVec) {
4987 if (MinElts[j] > EltNo)
4989 if (MaxElts[j] < EltNo)
4996 // Or record a new source if not...
4998 SourceVecs.push_back(SourceVec);
4999 MinElts.push_back(EltNo);
5000 MaxElts.push_back(EltNo);
5004 // Currently only do something sane when at most two source vectors
5006 if (SourceVecs.size() > 2)
5009 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5010 int VEXTOffsets[2] = {0, 0};
5012 // This loop extracts the usage patterns of the source vectors
5013 // and prepares appropriate SDValues for a shuffle if possible.
5014 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5015 if (SourceVecs[i].getValueType() == VT) {
5016 // No VEXT necessary
5017 ShuffleSrcs[i] = SourceVecs[i];
5020 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5021 // It probably isn't worth padding out a smaller vector just to
5022 // break it down again in a shuffle.
5026 // Since only 64-bit and 128-bit vectors are legal on ARM and
5027 // we've eliminated the other cases...
5028 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5029 "unexpected vector sizes in ReconstructShuffle");
5031 if (MaxElts[i] - MinElts[i] >= NumElts) {
5032 // Span too large for a VEXT to cope
5036 if (MinElts[i] >= NumElts) {
5037 // The extraction can just take the second half
5038 VEXTOffsets[i] = NumElts;
5039 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5041 DAG.getIntPtrConstant(NumElts));
5042 } else if (MaxElts[i] < NumElts) {
5043 // The extraction can just take the first half
5045 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5047 DAG.getIntPtrConstant(0));
5049 // An actual VEXT is needed
5050 VEXTOffsets[i] = MinElts[i];
5051 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5053 DAG.getIntPtrConstant(0));
5054 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5056 DAG.getIntPtrConstant(NumElts));
5057 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5058 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5062 SmallVector<int, 8> Mask;
5064 for (unsigned i = 0; i < NumElts; ++i) {
5065 SDValue Entry = Op.getOperand(i);
5066 if (Entry.getOpcode() == ISD::UNDEF) {
5071 SDValue ExtractVec = Entry.getOperand(0);
5072 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5073 .getOperand(1))->getSExtValue();
5074 if (ExtractVec == SourceVecs[0]) {
5075 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5077 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5081 // Final check before we try to produce nonsense...
5082 if (isShuffleMaskLegal(Mask, VT))
5083 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5089 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5090 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5091 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5092 /// are assumed to be legal.
5094 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5096 if (VT.getVectorNumElements() == 4 &&
5097 (VT.is128BitVector() || VT.is64BitVector())) {
5098 unsigned PFIndexes[4];
5099 for (unsigned i = 0; i != 4; ++i) {
5103 PFIndexes[i] = M[i];
5106 // Compute the index in the perfect shuffle table.
5107 unsigned PFTableIndex =
5108 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5109 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5110 unsigned Cost = (PFEntry >> 30);
5117 unsigned Imm, WhichResult;
5119 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5120 return (EltSize >= 32 ||
5121 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5122 isVREVMask(M, VT, 64) ||
5123 isVREVMask(M, VT, 32) ||
5124 isVREVMask(M, VT, 16) ||
5125 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5126 isVTBLMask(M, VT) ||
5127 isVTRNMask(M, VT, WhichResult) ||
5128 isVUZPMask(M, VT, WhichResult) ||
5129 isVZIPMask(M, VT, WhichResult) ||
5130 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5131 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5132 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5133 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5136 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5137 /// the specified operations to build the shuffle.
5138 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5139 SDValue RHS, SelectionDAG &DAG,
5141 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5142 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5143 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5146 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5155 OP_VUZPL, // VUZP, left result
5156 OP_VUZPR, // VUZP, right result
5157 OP_VZIPL, // VZIP, left result
5158 OP_VZIPR, // VZIP, right result
5159 OP_VTRNL, // VTRN, left result
5160 OP_VTRNR // VTRN, right result
5163 if (OpNum == OP_COPY) {
5164 if (LHSID == (1*9+2)*9+3) return LHS;
5165 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5169 SDValue OpLHS, OpRHS;
5170 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5171 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5172 EVT VT = OpLHS.getValueType();
5175 default: llvm_unreachable("Unknown shuffle opcode!");
5177 // VREV divides the vector in half and swaps within the half.
5178 if (VT.getVectorElementType() == MVT::i32 ||
5179 VT.getVectorElementType() == MVT::f32)
5180 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5181 // vrev <4 x i16> -> VREV32
5182 if (VT.getVectorElementType() == MVT::i16)
5183 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5184 // vrev <4 x i8> -> VREV16
5185 assert(VT.getVectorElementType() == MVT::i8);
5186 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5191 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5192 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5196 return DAG.getNode(ARMISD::VEXT, dl, VT,
5198 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5201 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5202 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5205 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5206 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5209 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5210 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5214 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5215 ArrayRef<int> ShuffleMask,
5216 SelectionDAG &DAG) {
5217 // Check to see if we can use the VTBL instruction.
5218 SDValue V1 = Op.getOperand(0);
5219 SDValue V2 = Op.getOperand(1);
5222 SmallVector<SDValue, 8> VTBLMask;
5223 for (ArrayRef<int>::iterator
5224 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5225 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5227 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5228 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5229 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5232 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5233 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
5237 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5238 SelectionDAG &DAG) {
5240 SDValue OpLHS = Op.getOperand(0);
5241 EVT VT = OpLHS.getValueType();
5243 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5244 "Expect an v8i16/v16i8 type");
5245 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5246 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5247 // extract the first 8 bytes into the top double word and the last 8 bytes
5248 // into the bottom double word. The v8i16 case is similar.
5249 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5250 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5251 DAG.getConstant(ExtractNum, MVT::i32));
5254 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5255 SDValue V1 = Op.getOperand(0);
5256 SDValue V2 = Op.getOperand(1);
5258 EVT VT = Op.getValueType();
5259 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5261 // Convert shuffles that are directly supported on NEON to target-specific
5262 // DAG nodes, instead of keeping them as shuffles and matching them again
5263 // during code selection. This is more efficient and avoids the possibility
5264 // of inconsistencies between legalization and selection.
5265 // FIXME: floating-point vectors should be canonicalized to integer vectors
5266 // of the same time so that they get CSEd properly.
5267 ArrayRef<int> ShuffleMask = SVN->getMask();
5269 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5270 if (EltSize <= 32) {
5271 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5272 int Lane = SVN->getSplatIndex();
5273 // If this is undef splat, generate it via "just" vdup, if possible.
5274 if (Lane == -1) Lane = 0;
5276 // Test if V1 is a SCALAR_TO_VECTOR.
5277 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5278 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5280 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5281 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5283 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5284 !isa<ConstantSDNode>(V1.getOperand(0))) {
5285 bool IsScalarToVector = true;
5286 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5287 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5288 IsScalarToVector = false;
5291 if (IsScalarToVector)
5292 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5294 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5295 DAG.getConstant(Lane, MVT::i32));
5300 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5303 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5304 DAG.getConstant(Imm, MVT::i32));
5307 if (isVREVMask(ShuffleMask, VT, 64))
5308 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5309 if (isVREVMask(ShuffleMask, VT, 32))
5310 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5311 if (isVREVMask(ShuffleMask, VT, 16))
5312 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5314 if (V2->getOpcode() == ISD::UNDEF &&
5315 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5316 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5317 DAG.getConstant(Imm, MVT::i32));
5320 // Check for Neon shuffles that modify both input vectors in place.
5321 // If both results are used, i.e., if there are two shuffles with the same
5322 // source operands and with masks corresponding to both results of one of
5323 // these operations, DAG memoization will ensure that a single node is
5324 // used for both shuffles.
5325 unsigned WhichResult;
5326 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5327 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5328 V1, V2).getValue(WhichResult);
5329 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5330 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5331 V1, V2).getValue(WhichResult);
5332 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5333 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5334 V1, V2).getValue(WhichResult);
5336 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5337 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5338 V1, V1).getValue(WhichResult);
5339 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5340 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5341 V1, V1).getValue(WhichResult);
5342 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5343 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5344 V1, V1).getValue(WhichResult);
5347 // If the shuffle is not directly supported and it has 4 elements, use
5348 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5349 unsigned NumElts = VT.getVectorNumElements();
5351 unsigned PFIndexes[4];
5352 for (unsigned i = 0; i != 4; ++i) {
5353 if (ShuffleMask[i] < 0)
5356 PFIndexes[i] = ShuffleMask[i];
5359 // Compute the index in the perfect shuffle table.
5360 unsigned PFTableIndex =
5361 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5362 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5363 unsigned Cost = (PFEntry >> 30);
5366 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5369 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5370 if (EltSize >= 32) {
5371 // Do the expansion with floating-point types, since that is what the VFP
5372 // registers are defined to use, and since i64 is not legal.
5373 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5374 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5375 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5376 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5377 SmallVector<SDValue, 8> Ops;
5378 for (unsigned i = 0; i < NumElts; ++i) {
5379 if (ShuffleMask[i] < 0)
5380 Ops.push_back(DAG.getUNDEF(EltVT));
5382 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5383 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5384 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5387 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
5388 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5391 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5392 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5394 if (VT == MVT::v8i8) {
5395 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5396 if (NewOp.getNode())
5403 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5404 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5405 SDValue Lane = Op.getOperand(2);
5406 if (!isa<ConstantSDNode>(Lane))
5412 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5413 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5414 SDValue Lane = Op.getOperand(1);
5415 if (!isa<ConstantSDNode>(Lane))
5418 SDValue Vec = Op.getOperand(0);
5419 if (Op.getValueType() == MVT::i32 &&
5420 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5422 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5428 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5429 // The only time a CONCAT_VECTORS operation can have legal types is when
5430 // two 64-bit vectors are concatenated to a 128-bit vector.
5431 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5432 "unexpected CONCAT_VECTORS");
5434 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5435 SDValue Op0 = Op.getOperand(0);
5436 SDValue Op1 = Op.getOperand(1);
5437 if (Op0.getOpcode() != ISD::UNDEF)
5438 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5439 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5440 DAG.getIntPtrConstant(0));
5441 if (Op1.getOpcode() != ISD::UNDEF)
5442 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5443 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5444 DAG.getIntPtrConstant(1));
5445 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5448 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5449 /// element has been zero/sign-extended, depending on the isSigned parameter,
5450 /// from an integer type half its size.
5451 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5453 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5454 EVT VT = N->getValueType(0);
5455 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5456 SDNode *BVN = N->getOperand(0).getNode();
5457 if (BVN->getValueType(0) != MVT::v4i32 ||
5458 BVN->getOpcode() != ISD::BUILD_VECTOR)
5460 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5461 unsigned HiElt = 1 - LoElt;
5462 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5463 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5464 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5465 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5466 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5469 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5470 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5473 if (Hi0->isNullValue() && Hi1->isNullValue())
5479 if (N->getOpcode() != ISD::BUILD_VECTOR)
5482 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5483 SDNode *Elt = N->getOperand(i).getNode();
5484 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5485 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5486 unsigned HalfSize = EltSize / 2;
5488 if (!isIntN(HalfSize, C->getSExtValue()))
5491 if (!isUIntN(HalfSize, C->getZExtValue()))
5502 /// isSignExtended - Check if a node is a vector value that is sign-extended
5503 /// or a constant BUILD_VECTOR with sign-extended elements.
5504 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5505 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5507 if (isExtendedBUILD_VECTOR(N, DAG, true))
5512 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5513 /// or a constant BUILD_VECTOR with zero-extended elements.
5514 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5515 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5517 if (isExtendedBUILD_VECTOR(N, DAG, false))
5522 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5523 if (OrigVT.getSizeInBits() >= 64)
5526 assert(OrigVT.isSimple() && "Expecting a simple value type");
5528 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5529 switch (OrigSimpleTy) {
5530 default: llvm_unreachable("Unexpected Vector Type");
5539 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5540 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5541 /// We insert the required extension here to get the vector to fill a D register.
5542 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5545 unsigned ExtOpcode) {
5546 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5547 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5548 // 64-bits we need to insert a new extension so that it will be 64-bits.
5549 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5550 if (OrigTy.getSizeInBits() >= 64)
5553 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5554 EVT NewVT = getExtensionTo64Bits(OrigTy);
5556 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5559 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5560 /// does not do any sign/zero extension. If the original vector is less
5561 /// than 64 bits, an appropriate extension will be added after the load to
5562 /// reach a total size of 64 bits. We have to add the extension separately
5563 /// because ARM does not have a sign/zero extending load for vectors.
5564 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5565 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5567 // The load already has the right type.
5568 if (ExtendedTy == LD->getMemoryVT())
5569 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5570 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5571 LD->isNonTemporal(), LD->isInvariant(),
5572 LD->getAlignment());
5574 // We need to create a zextload/sextload. We cannot just create a load
5575 // followed by a zext/zext node because LowerMUL is also run during normal
5576 // operation legalization where we can't create illegal types.
5577 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5578 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5579 LD->getMemoryVT(), LD->isVolatile(),
5580 LD->isNonTemporal(), LD->getAlignment());
5583 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5584 /// extending load, or BUILD_VECTOR with extended elements, return the
5585 /// unextended value. The unextended vector should be 64 bits so that it can
5586 /// be used as an operand to a VMULL instruction. If the original vector size
5587 /// before extension is less than 64 bits we add a an extension to resize
5588 /// the vector to 64 bits.
5589 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5590 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5591 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5592 N->getOperand(0)->getValueType(0),
5596 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5597 return SkipLoadExtensionForVMULL(LD, DAG);
5599 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5600 // have been legalized as a BITCAST from v4i32.
5601 if (N->getOpcode() == ISD::BITCAST) {
5602 SDNode *BVN = N->getOperand(0).getNode();
5603 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5604 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5605 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5606 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5607 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5609 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5610 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5611 EVT VT = N->getValueType(0);
5612 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5613 unsigned NumElts = VT.getVectorNumElements();
5614 MVT TruncVT = MVT::getIntegerVT(EltSize);
5615 SmallVector<SDValue, 8> Ops;
5616 for (unsigned i = 0; i != NumElts; ++i) {
5617 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5618 const APInt &CInt = C->getAPIntValue();
5619 // Element types smaller than 32 bits are not legal, so use i32 elements.
5620 // The values are implicitly truncated so sext vs. zext doesn't matter.
5621 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5623 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5624 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5627 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5628 unsigned Opcode = N->getOpcode();
5629 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5630 SDNode *N0 = N->getOperand(0).getNode();
5631 SDNode *N1 = N->getOperand(1).getNode();
5632 return N0->hasOneUse() && N1->hasOneUse() &&
5633 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5638 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5639 unsigned Opcode = N->getOpcode();
5640 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5641 SDNode *N0 = N->getOperand(0).getNode();
5642 SDNode *N1 = N->getOperand(1).getNode();
5643 return N0->hasOneUse() && N1->hasOneUse() &&
5644 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5649 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5650 // Multiplications are only custom-lowered for 128-bit vectors so that
5651 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5652 EVT VT = Op.getValueType();
5653 assert(VT.is128BitVector() && VT.isInteger() &&
5654 "unexpected type for custom-lowering ISD::MUL");
5655 SDNode *N0 = Op.getOperand(0).getNode();
5656 SDNode *N1 = Op.getOperand(1).getNode();
5657 unsigned NewOpc = 0;
5659 bool isN0SExt = isSignExtended(N0, DAG);
5660 bool isN1SExt = isSignExtended(N1, DAG);
5661 if (isN0SExt && isN1SExt)
5662 NewOpc = ARMISD::VMULLs;
5664 bool isN0ZExt = isZeroExtended(N0, DAG);
5665 bool isN1ZExt = isZeroExtended(N1, DAG);
5666 if (isN0ZExt && isN1ZExt)
5667 NewOpc = ARMISD::VMULLu;
5668 else if (isN1SExt || isN1ZExt) {
5669 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5670 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5671 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5672 NewOpc = ARMISD::VMULLs;
5674 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5675 NewOpc = ARMISD::VMULLu;
5677 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5679 NewOpc = ARMISD::VMULLu;
5685 if (VT == MVT::v2i64)
5686 // Fall through to expand this. It is not legal.
5689 // Other vector multiplications are legal.
5694 // Legalize to a VMULL instruction.
5697 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5699 Op0 = SkipExtensionForVMULL(N0, DAG);
5700 assert(Op0.getValueType().is64BitVector() &&
5701 Op1.getValueType().is64BitVector() &&
5702 "unexpected types for extended operands to VMULL");
5703 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5706 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5707 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5714 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5715 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5716 EVT Op1VT = Op1.getValueType();
5717 return DAG.getNode(N0->getOpcode(), DL, VT,
5718 DAG.getNode(NewOpc, DL, VT,
5719 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5720 DAG.getNode(NewOpc, DL, VT,
5721 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5725 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5727 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5728 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5729 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5730 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5731 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5732 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5733 // Get reciprocal estimate.
5734 // float4 recip = vrecpeq_f32(yf);
5735 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5736 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5737 // Because char has a smaller range than uchar, we can actually get away
5738 // without any newton steps. This requires that we use a weird bias
5739 // of 0xb000, however (again, this has been exhaustively tested).
5740 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5741 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5742 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5743 Y = DAG.getConstant(0xb000, MVT::i32);
5744 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5745 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5746 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5747 // Convert back to short.
5748 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5749 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5754 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5756 // Convert to float.
5757 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5758 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5759 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5760 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5761 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5762 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5764 // Use reciprocal estimate and one refinement step.
5765 // float4 recip = vrecpeq_f32(yf);
5766 // recip *= vrecpsq_f32(yf, recip);
5767 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5768 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5769 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5770 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5772 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5773 // Because short has a smaller range than ushort, we can actually get away
5774 // with only a single newton step. This requires that we use a weird bias
5775 // of 89, however (again, this has been exhaustively tested).
5776 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5777 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5778 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5779 N1 = DAG.getConstant(0x89, MVT::i32);
5780 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5781 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5782 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5783 // Convert back to integer and return.
5784 // return vmovn_s32(vcvt_s32_f32(result));
5785 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5786 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5790 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5791 EVT VT = Op.getValueType();
5792 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5793 "unexpected type for custom-lowering ISD::SDIV");
5796 SDValue N0 = Op.getOperand(0);
5797 SDValue N1 = Op.getOperand(1);
5800 if (VT == MVT::v8i8) {
5801 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5802 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5804 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5805 DAG.getIntPtrConstant(4));
5806 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5807 DAG.getIntPtrConstant(4));
5808 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5809 DAG.getIntPtrConstant(0));
5810 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5811 DAG.getIntPtrConstant(0));
5813 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5814 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5816 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5817 N0 = LowerCONCAT_VECTORS(N0, DAG);
5819 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5822 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5825 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5826 EVT VT = Op.getValueType();
5827 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5828 "unexpected type for custom-lowering ISD::UDIV");
5831 SDValue N0 = Op.getOperand(0);
5832 SDValue N1 = Op.getOperand(1);
5835 if (VT == MVT::v8i8) {
5836 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5837 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5839 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5840 DAG.getIntPtrConstant(4));
5841 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5842 DAG.getIntPtrConstant(4));
5843 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5844 DAG.getIntPtrConstant(0));
5845 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5846 DAG.getIntPtrConstant(0));
5848 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5849 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5851 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5852 N0 = LowerCONCAT_VECTORS(N0, DAG);
5854 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5855 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5860 // v4i16 sdiv ... Convert to float.
5861 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5862 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5863 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5864 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5865 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5866 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5868 // Use reciprocal estimate and two refinement steps.
5869 // float4 recip = vrecpeq_f32(yf);
5870 // recip *= vrecpsq_f32(yf, recip);
5871 // recip *= vrecpsq_f32(yf, recip);
5872 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5873 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5874 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5875 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5877 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5878 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5879 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5881 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5882 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5883 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5884 // and that it will never cause us to return an answer too large).
5885 // float4 result = as_float4(as_int4(xf*recip) + 2);
5886 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5887 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5888 N1 = DAG.getConstant(2, MVT::i32);
5889 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5890 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5891 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5892 // Convert back to integer and return.
5893 // return vmovn_u32(vcvt_s32_f32(result));
5894 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5895 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5899 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5900 EVT VT = Op.getNode()->getValueType(0);
5901 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5904 bool ExtraOp = false;
5905 switch (Op.getOpcode()) {
5906 default: llvm_unreachable("Invalid code");
5907 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5908 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5909 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5910 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5914 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5916 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
5917 Op.getOperand(1), Op.getOperand(2));
5920 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
5921 assert(Subtarget->isTargetDarwin());
5923 // For iOS, we want to call an alternative entry point: __sincos_stret,
5924 // return values are passed via sret.
5926 SDValue Arg = Op.getOperand(0);
5927 EVT ArgVT = Arg.getValueType();
5928 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
5930 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5931 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5933 // Pair of floats / doubles used to pass the result.
5934 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
5936 // Create stack object for sret.
5937 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
5938 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
5939 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
5940 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
5946 Entry.Ty = RetTy->getPointerTo();
5947 Entry.isSExt = false;
5948 Entry.isZExt = false;
5949 Entry.isSRet = true;
5950 Args.push_back(Entry);
5954 Entry.isSExt = false;
5955 Entry.isZExt = false;
5956 Args.push_back(Entry);
5958 const char *LibcallName = (ArgVT == MVT::f64)
5959 ? "__sincos_stret" : "__sincosf_stret";
5960 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
5963 CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
5964 false, false, false, false, 0,
5965 CallingConv::C, /*isTaillCall=*/false,
5966 /*doesNotRet=*/false, /*isReturnValueUsed*/false,
5967 Callee, Args, DAG, dl);
5968 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
5970 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
5971 MachinePointerInfo(), false, false, false, 0);
5973 // Address of cos field.
5974 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
5975 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
5976 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
5977 MachinePointerInfo(), false, false, false, 0);
5979 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
5980 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
5981 LoadSin.getValue(0), LoadCos.getValue(0));
5984 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5985 // Monotonic load/store is legal for all targets
5986 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5989 // Acquire/Release load/store is not legal for targets without a
5990 // dmb or equivalent available.
5995 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5996 SelectionDAG &DAG) {
5998 assert (Node->getValueType(0) == MVT::i64 &&
5999 "Only know how to expand i64 atomics");
6000 AtomicSDNode *AN = cast<AtomicSDNode>(Node);
6002 SmallVector<SDValue, 6> Ops;
6003 Ops.push_back(Node->getOperand(0)); // Chain
6004 Ops.push_back(Node->getOperand(1)); // Ptr
6005 for(unsigned i=2; i<Node->getNumOperands(); i++) {
6007 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6008 Node->getOperand(i), DAG.getIntPtrConstant(0)));
6010 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6011 Node->getOperand(i), DAG.getIntPtrConstant(1)));
6013 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6015 DAG.getAtomic(Node->getOpcode(), dl, MVT::i64, Tys, Ops.data(), Ops.size(),
6016 cast<MemSDNode>(Node)->getMemOperand(), AN->getOrdering(),
6017 AN->getSynchScope());
6018 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
6019 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6020 Results.push_back(Result.getValue(2));
6023 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6024 SmallVectorImpl<SDValue> &Results,
6026 const ARMSubtarget *Subtarget) {
6028 SDValue Cycles32, OutChain;
6030 if (Subtarget->hasPerfMon()) {
6031 // Under Power Management extensions, the cycle-count is:
6032 // mrc p15, #0, <Rt>, c9, c13, #0
6033 SDValue Ops[] = { N->getOperand(0), // Chain
6034 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6035 DAG.getConstant(15, MVT::i32),
6036 DAG.getConstant(0, MVT::i32),
6037 DAG.getConstant(9, MVT::i32),
6038 DAG.getConstant(13, MVT::i32),
6039 DAG.getConstant(0, MVT::i32)
6042 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6043 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
6044 array_lengthof(Ops));
6045 OutChain = Cycles32.getValue(1);
6047 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6048 // there are older ARM CPUs that have implementation-specific ways of
6049 // obtaining this information (FIXME!).
6050 Cycles32 = DAG.getConstant(0, MVT::i32);
6051 OutChain = DAG.getEntryNode();
6055 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6056 Cycles32, DAG.getConstant(0, MVT::i32));
6057 Results.push_back(Cycles64);
6058 Results.push_back(OutChain);
6061 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6062 switch (Op.getOpcode()) {
6063 default: llvm_unreachable("Don't know how to custom lower this!");
6064 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6065 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6066 case ISD::GlobalAddress:
6067 return Subtarget->isTargetMachO() ? LowerGlobalAddressDarwin(Op, DAG) :
6068 LowerGlobalAddressELF(Op, DAG);
6069 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6070 case ISD::SELECT: return LowerSELECT(Op, DAG);
6071 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6072 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6073 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6074 case ISD::VASTART: return LowerVASTART(Op, DAG);
6075 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6076 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6077 case ISD::SINT_TO_FP:
6078 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6079 case ISD::FP_TO_SINT:
6080 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6081 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6082 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6083 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6084 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6085 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6086 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6087 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6089 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6092 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6093 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6094 case ISD::SRL_PARTS:
6095 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6096 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6097 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6098 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6099 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6100 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6101 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6102 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6103 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6104 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6105 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6106 case ISD::MUL: return LowerMUL(Op, DAG);
6107 case ISD::SDIV: return LowerSDIV(Op, DAG);
6108 case ISD::UDIV: return LowerUDIV(Op, DAG);
6112 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6113 case ISD::ATOMIC_LOAD:
6114 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6115 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6117 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6121 /// ReplaceNodeResults - Replace the results of node with an illegal result
6122 /// type with new values built out of custom code.
6123 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6124 SmallVectorImpl<SDValue>&Results,
6125 SelectionDAG &DAG) const {
6127 switch (N->getOpcode()) {
6129 llvm_unreachable("Don't know how to custom expand this!");
6131 Res = ExpandBITCAST(N, DAG);
6135 Res = Expand64BitShift(N, DAG, Subtarget);
6137 case ISD::READCYCLECOUNTER:
6138 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6140 case ISD::ATOMIC_STORE:
6141 case ISD::ATOMIC_LOAD:
6142 case ISD::ATOMIC_LOAD_ADD:
6143 case ISD::ATOMIC_LOAD_AND:
6144 case ISD::ATOMIC_LOAD_NAND:
6145 case ISD::ATOMIC_LOAD_OR:
6146 case ISD::ATOMIC_LOAD_SUB:
6147 case ISD::ATOMIC_LOAD_XOR:
6148 case ISD::ATOMIC_SWAP:
6149 case ISD::ATOMIC_CMP_SWAP:
6150 case ISD::ATOMIC_LOAD_MIN:
6151 case ISD::ATOMIC_LOAD_UMIN:
6152 case ISD::ATOMIC_LOAD_MAX:
6153 case ISD::ATOMIC_LOAD_UMAX:
6154 ReplaceATOMIC_OP_64(N, Results, DAG);
6158 Results.push_back(Res);
6161 //===----------------------------------------------------------------------===//
6162 // ARM Scheduler Hooks
6163 //===----------------------------------------------------------------------===//
6166 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
6167 MachineBasicBlock *BB,
6168 unsigned Size) const {
6169 unsigned dest = MI->getOperand(0).getReg();
6170 unsigned ptr = MI->getOperand(1).getReg();
6171 unsigned oldval = MI->getOperand(2).getReg();
6172 unsigned newval = MI->getOperand(3).getReg();
6173 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6174 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
6175 DebugLoc dl = MI->getDebugLoc();
6176 bool isThumb2 = Subtarget->isThumb2();
6178 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6179 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
6180 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6181 (const TargetRegisterClass*)&ARM::GPRRegClass);
6184 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6185 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
6186 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
6189 unsigned ldrOpc, strOpc;
6190 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6192 MachineFunction *MF = BB->getParent();
6193 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6194 MachineFunction::iterator It = BB;
6195 ++It; // insert the new blocks after the current block
6197 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6198 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
6199 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6200 MF->insert(It, loop1MBB);
6201 MF->insert(It, loop2MBB);
6202 MF->insert(It, exitMBB);
6204 // Transfer the remainder of BB and its successor edges to exitMBB.
6205 exitMBB->splice(exitMBB->begin(), BB,
6206 llvm::next(MachineBasicBlock::iterator(MI)),
6208 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6212 // fallthrough --> loop1MBB
6213 BB->addSuccessor(loop1MBB);
6216 // ldrex dest, [ptr]
6220 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6221 if (ldrOpc == ARM::t2LDREX)
6223 AddDefaultPred(MIB);
6224 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6225 .addReg(dest).addReg(oldval));
6226 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6227 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6228 BB->addSuccessor(loop2MBB);
6229 BB->addSuccessor(exitMBB);
6232 // strex scratch, newval, [ptr]
6236 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
6237 if (strOpc == ARM::t2STREX)
6239 AddDefaultPred(MIB);
6240 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6241 .addReg(scratch).addImm(0));
6242 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6243 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6244 BB->addSuccessor(loop1MBB);
6245 BB->addSuccessor(exitMBB);
6251 MI->eraseFromParent(); // The instruction is gone now.
6257 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
6258 unsigned Size, unsigned BinOpcode) const {
6259 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
6260 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6262 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6263 MachineFunction *MF = BB->getParent();
6264 MachineFunction::iterator It = BB;
6267 unsigned dest = MI->getOperand(0).getReg();
6268 unsigned ptr = MI->getOperand(1).getReg();
6269 unsigned incr = MI->getOperand(2).getReg();
6270 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6271 DebugLoc dl = MI->getDebugLoc();
6272 bool isThumb2 = Subtarget->isThumb2();
6274 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6276 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6277 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6278 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6281 unsigned ldrOpc, strOpc;
6282 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6284 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6285 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6286 MF->insert(It, loopMBB);
6287 MF->insert(It, exitMBB);
6289 // Transfer the remainder of BB and its successor edges to exitMBB.
6290 exitMBB->splice(exitMBB->begin(), BB,
6291 llvm::next(MachineBasicBlock::iterator(MI)),
6293 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6295 const TargetRegisterClass *TRC = isThumb2 ?
6296 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6297 (const TargetRegisterClass*)&ARM::GPRRegClass;
6298 unsigned scratch = MRI.createVirtualRegister(TRC);
6299 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
6303 // fallthrough --> loopMBB
6304 BB->addSuccessor(loopMBB);
6308 // <binop> scratch2, dest, incr
6309 // strex scratch, scratch2, ptr
6312 // fallthrough --> exitMBB
6314 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6315 if (ldrOpc == ARM::t2LDREX)
6317 AddDefaultPred(MIB);
6319 // operand order needs to go the other way for NAND
6320 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
6321 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6322 addReg(incr).addReg(dest)).addReg(0);
6324 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
6325 addReg(dest).addReg(incr)).addReg(0);
6328 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6329 if (strOpc == ARM::t2STREX)
6331 AddDefaultPred(MIB);
6332 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6333 .addReg(scratch).addImm(0));
6334 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6335 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6337 BB->addSuccessor(loopMBB);
6338 BB->addSuccessor(exitMBB);
6344 MI->eraseFromParent(); // The instruction is gone now.
6350 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
6351 MachineBasicBlock *BB,
6354 ARMCC::CondCodes Cond) const {
6355 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6357 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6358 MachineFunction *MF = BB->getParent();
6359 MachineFunction::iterator It = BB;
6362 unsigned dest = MI->getOperand(0).getReg();
6363 unsigned ptr = MI->getOperand(1).getReg();
6364 unsigned incr = MI->getOperand(2).getReg();
6365 unsigned oldval = dest;
6366 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6367 DebugLoc dl = MI->getDebugLoc();
6368 bool isThumb2 = Subtarget->isThumb2();
6370 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6372 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
6373 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6374 MRI.constrainRegClass(incr, &ARM::rGPRRegClass);
6377 unsigned ldrOpc, strOpc, extendOpc;
6378 getExclusiveOperation(Size, Ord, isThumb2, ldrOpc, strOpc);
6380 default: llvm_unreachable("unsupported size for AtomicBinaryMinMax!");
6382 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
6385 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
6392 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6393 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6394 MF->insert(It, loopMBB);
6395 MF->insert(It, exitMBB);
6397 // Transfer the remainder of BB and its successor edges to exitMBB.
6398 exitMBB->splice(exitMBB->begin(), BB,
6399 llvm::next(MachineBasicBlock::iterator(MI)),
6401 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6403 const TargetRegisterClass *TRC = isThumb2 ?
6404 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6405 (const TargetRegisterClass*)&ARM::GPRRegClass;
6406 unsigned scratch = MRI.createVirtualRegister(TRC);
6407 unsigned scratch2 = MRI.createVirtualRegister(TRC);
6411 // fallthrough --> loopMBB
6412 BB->addSuccessor(loopMBB);
6416 // (sign extend dest, if required)
6418 // cmov.cond scratch2, incr, dest
6419 // strex scratch, scratch2, ptr
6422 // fallthrough --> exitMBB
6424 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
6425 if (ldrOpc == ARM::t2LDREX)
6427 AddDefaultPred(MIB);
6429 // Sign extend the value, if necessary.
6430 if (signExtend && extendOpc) {
6431 oldval = MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass
6432 : &ARM::GPRnopcRegClass);
6434 MRI.constrainRegClass(dest, &ARM::GPRnopcRegClass);
6435 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
6440 // Build compare and cmov instructions.
6441 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6442 .addReg(oldval).addReg(incr));
6443 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
6444 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
6446 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
6447 if (strOpc == ARM::t2STREX)
6449 AddDefaultPred(MIB);
6450 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6451 .addReg(scratch).addImm(0));
6452 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6453 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6455 BB->addSuccessor(loopMBB);
6456 BB->addSuccessor(exitMBB);
6462 MI->eraseFromParent(); // The instruction is gone now.
6468 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6469 unsigned Op1, unsigned Op2,
6470 bool NeedsCarry, bool IsCmpxchg,
6471 bool IsMinMax, ARMCC::CondCodes CC) const {
6472 // This also handles ATOMIC_SWAP and ATOMIC_STORE, indicated by Op1==0.
6473 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6475 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6476 MachineFunction *MF = BB->getParent();
6477 MachineFunction::iterator It = BB;
6480 bool isStore = (MI->getOpcode() == ARM::ATOMIC_STORE_I64);
6481 unsigned offset = (isStore ? -2 : 0);
6482 unsigned destlo = MI->getOperand(0).getReg();
6483 unsigned desthi = MI->getOperand(1).getReg();
6484 unsigned ptr = MI->getOperand(offset+2).getReg();
6485 unsigned vallo = MI->getOperand(offset+3).getReg();
6486 unsigned valhi = MI->getOperand(offset+4).getReg();
6487 unsigned OrdIdx = offset + (IsCmpxchg ? 7 : 5);
6488 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(OrdIdx).getImm());
6489 DebugLoc dl = MI->getDebugLoc();
6490 bool isThumb2 = Subtarget->isThumb2();
6492 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6494 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6495 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6496 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6497 MRI.constrainRegClass(vallo, &ARM::rGPRRegClass);
6498 MRI.constrainRegClass(valhi, &ARM::rGPRRegClass);
6501 unsigned ldrOpc, strOpc;
6502 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6504 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6505 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6506 if (IsCmpxchg || IsMinMax)
6507 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6509 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6510 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6512 MF->insert(It, loopMBB);
6513 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6514 if (IsCmpxchg) MF->insert(It, cont2BB);
6515 MF->insert(It, exitMBB);
6517 // Transfer the remainder of BB and its successor edges to exitMBB.
6518 exitMBB->splice(exitMBB->begin(), BB,
6519 llvm::next(MachineBasicBlock::iterator(MI)),
6521 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6523 const TargetRegisterClass *TRC = isThumb2 ?
6524 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6525 (const TargetRegisterClass*)&ARM::GPRRegClass;
6526 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6530 // fallthrough --> loopMBB
6531 BB->addSuccessor(loopMBB);
6534 // ldrexd r2, r3, ptr
6535 // <binopa> r0, r2, incr
6536 // <binopb> r1, r3, incr
6537 // strexd storesuccess, r0, r1, ptr
6538 // cmp storesuccess, #0
6540 // fallthrough --> exitMBB
6546 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6547 .addReg(destlo, RegState::Define)
6548 .addReg(desthi, RegState::Define)
6551 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6552 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
6553 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6554 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6555 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6556 .addReg(GPRPair0, 0, ARM::gsub_0);
6557 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6558 .addReg(GPRPair0, 0, ARM::gsub_1);
6562 unsigned StoreLo, StoreHi;
6565 for (unsigned i = 0; i < 2; i++) {
6566 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6568 .addReg(i == 0 ? destlo : desthi)
6569 .addReg(i == 0 ? vallo : valhi));
6570 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6571 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6572 BB->addSuccessor(exitMBB);
6573 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6574 BB = (i == 0 ? contBB : cont2BB);
6577 // Copy to physregs for strexd
6578 StoreLo = MI->getOperand(5).getReg();
6579 StoreHi = MI->getOperand(6).getReg();
6581 // Perform binary operation
6582 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6583 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6584 .addReg(destlo).addReg(vallo))
6585 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6586 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6587 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6588 .addReg(desthi).addReg(valhi))
6589 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6594 // Copy to physregs for strexd
6599 // Compare and branch to exit block.
6600 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6601 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6602 BB->addSuccessor(exitMBB);
6603 BB->addSuccessor(contBB);
6611 MRI.constrainRegClass(StoreLo, &ARM::rGPRRegClass);
6612 MRI.constrainRegClass(StoreHi, &ARM::rGPRRegClass);
6613 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6614 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6616 // Marshal a pair...
6617 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6618 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6619 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6620 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6621 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6624 .addImm(ARM::gsub_0);
6625 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6628 .addImm(ARM::gsub_1);
6631 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
6632 .addReg(StorePair).addReg(ptr));
6635 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6636 .addReg(storesuccess).addImm(0));
6637 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6638 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6640 BB->addSuccessor(loopMBB);
6641 BB->addSuccessor(exitMBB);
6647 MI->eraseFromParent(); // The instruction is gone now.
6653 ARMTargetLowering::EmitAtomicLoad64(MachineInstr *MI, MachineBasicBlock *BB) const {
6655 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6657 unsigned destlo = MI->getOperand(0).getReg();
6658 unsigned desthi = MI->getOperand(1).getReg();
6659 unsigned ptr = MI->getOperand(2).getReg();
6660 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
6661 DebugLoc dl = MI->getDebugLoc();
6662 bool isThumb2 = Subtarget->isThumb2();
6664 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6666 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6667 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6668 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6670 unsigned ldrOpc, strOpc;
6671 getExclusiveOperation(8, Ord, isThumb2, ldrOpc, strOpc);
6673 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(ldrOpc));
6676 MIB.addReg(destlo, RegState::Define)
6677 .addReg(desthi, RegState::Define)
6681 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6682 MIB.addReg(GPRPair0, RegState::Define).addReg(ptr);
6684 // Copy GPRPair0 into dest. (This copy will normally be coalesced.)
6685 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), destlo)
6686 .addReg(GPRPair0, 0, ARM::gsub_0);
6687 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), desthi)
6688 .addReg(GPRPair0, 0, ARM::gsub_1);
6690 AddDefaultPred(MIB);
6692 MI->eraseFromParent(); // The instruction is gone now.
6697 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6698 /// registers the function context.
6699 void ARMTargetLowering::
6700 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6701 MachineBasicBlock *DispatchBB, int FI) const {
6702 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6703 DebugLoc dl = MI->getDebugLoc();
6704 MachineFunction *MF = MBB->getParent();
6705 MachineRegisterInfo *MRI = &MF->getRegInfo();
6706 MachineConstantPool *MCP = MF->getConstantPool();
6707 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6708 const Function *F = MF->getFunction();
6710 bool isThumb = Subtarget->isThumb();
6711 bool isThumb2 = Subtarget->isThumb2();
6713 unsigned PCLabelId = AFI->createPICLabelUId();
6714 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6715 ARMConstantPoolValue *CPV =
6716 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6717 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6719 const TargetRegisterClass *TRC = isThumb ?
6720 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6721 (const TargetRegisterClass*)&ARM::GPRRegClass;
6723 // Grab constant pool and fixed stack memory operands.
6724 MachineMemOperand *CPMMO =
6725 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6726 MachineMemOperand::MOLoad, 4, 4);
6728 MachineMemOperand *FIMMOSt =
6729 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6730 MachineMemOperand::MOStore, 4, 4);
6732 // Load the address of the dispatch MBB into the jump buffer.
6734 // Incoming value: jbuf
6735 // ldr.n r5, LCPI1_1
6738 // str r5, [$jbuf, #+4] ; &jbuf[1]
6739 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6740 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6741 .addConstantPoolIndex(CPI)
6742 .addMemOperand(CPMMO));
6743 // Set the low bit because of thumb mode.
6744 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6746 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6747 .addReg(NewVReg1, RegState::Kill)
6749 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6750 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6751 .addReg(NewVReg2, RegState::Kill)
6753 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6754 .addReg(NewVReg3, RegState::Kill)
6756 .addImm(36) // &jbuf[1] :: pc
6757 .addMemOperand(FIMMOSt));
6758 } else if (isThumb) {
6759 // Incoming value: jbuf
6760 // ldr.n r1, LCPI1_4
6764 // add r2, $jbuf, #+4 ; &jbuf[1]
6766 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6767 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6768 .addConstantPoolIndex(CPI)
6769 .addMemOperand(CPMMO));
6770 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6771 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6772 .addReg(NewVReg1, RegState::Kill)
6774 // Set the low bit because of thumb mode.
6775 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6776 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6777 .addReg(ARM::CPSR, RegState::Define)
6779 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6780 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6781 .addReg(ARM::CPSR, RegState::Define)
6782 .addReg(NewVReg2, RegState::Kill)
6783 .addReg(NewVReg3, RegState::Kill));
6784 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6785 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6787 .addImm(36)); // &jbuf[1] :: pc
6788 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6789 .addReg(NewVReg4, RegState::Kill)
6790 .addReg(NewVReg5, RegState::Kill)
6792 .addMemOperand(FIMMOSt));
6794 // Incoming value: jbuf
6797 // str r1, [$jbuf, #+4] ; &jbuf[1]
6798 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6799 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6800 .addConstantPoolIndex(CPI)
6802 .addMemOperand(CPMMO));
6803 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6804 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6805 .addReg(NewVReg1, RegState::Kill)
6806 .addImm(PCLabelId));
6807 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6808 .addReg(NewVReg2, RegState::Kill)
6810 .addImm(36) // &jbuf[1] :: pc
6811 .addMemOperand(FIMMOSt));
6815 MachineBasicBlock *ARMTargetLowering::
6816 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6817 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6818 DebugLoc dl = MI->getDebugLoc();
6819 MachineFunction *MF = MBB->getParent();
6820 MachineRegisterInfo *MRI = &MF->getRegInfo();
6821 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6822 MachineFrameInfo *MFI = MF->getFrameInfo();
6823 int FI = MFI->getFunctionContextIndex();
6825 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6826 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6827 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6829 // Get a mapping of the call site numbers to all of the landing pads they're
6831 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6832 unsigned MaxCSNum = 0;
6833 MachineModuleInfo &MMI = MF->getMMI();
6834 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6836 if (!BB->isLandingPad()) continue;
6838 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6840 for (MachineBasicBlock::iterator
6841 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6842 if (!II->isEHLabel()) continue;
6844 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6845 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6847 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6848 for (SmallVectorImpl<unsigned>::iterator
6849 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6850 CSI != CSE; ++CSI) {
6851 CallSiteNumToLPad[*CSI].push_back(BB);
6852 MaxCSNum = std::max(MaxCSNum, *CSI);
6858 // Get an ordered list of the machine basic blocks for the jump table.
6859 std::vector<MachineBasicBlock*> LPadList;
6860 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6861 LPadList.reserve(CallSiteNumToLPad.size());
6862 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6863 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6864 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6865 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6866 LPadList.push_back(*II);
6867 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6871 assert(!LPadList.empty() &&
6872 "No landing pad destinations for the dispatch jump table!");
6874 // Create the jump table and associated information.
6875 MachineJumpTableInfo *JTI =
6876 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6877 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6878 unsigned UId = AFI->createJumpTableUId();
6879 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6881 // Create the MBBs for the dispatch code.
6883 // Shove the dispatch's address into the return slot in the function context.
6884 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6885 DispatchBB->setIsLandingPad();
6887 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6888 unsigned trap_opcode;
6889 if (Subtarget->isThumb())
6890 trap_opcode = ARM::tTRAP;
6892 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6894 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6895 DispatchBB->addSuccessor(TrapBB);
6897 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6898 DispatchBB->addSuccessor(DispContBB);
6901 MF->insert(MF->end(), DispatchBB);
6902 MF->insert(MF->end(), DispContBB);
6903 MF->insert(MF->end(), TrapBB);
6905 // Insert code into the entry block that creates and registers the function
6907 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6909 MachineMemOperand *FIMMOLd =
6910 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6911 MachineMemOperand::MOLoad |
6912 MachineMemOperand::MOVolatile, 4, 4);
6914 MachineInstrBuilder MIB;
6915 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6917 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6918 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6920 // Add a register mask with no preserved registers. This results in all
6921 // registers being marked as clobbered.
6922 MIB.addRegMask(RI.getNoPreservedMask());
6924 unsigned NumLPads = LPadList.size();
6925 if (Subtarget->isThumb2()) {
6926 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6927 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6930 .addMemOperand(FIMMOLd));
6932 if (NumLPads < 256) {
6933 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6935 .addImm(LPadList.size()));
6937 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6938 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6939 .addImm(NumLPads & 0xFFFF));
6941 unsigned VReg2 = VReg1;
6942 if ((NumLPads & 0xFFFF0000) != 0) {
6943 VReg2 = MRI->createVirtualRegister(TRC);
6944 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6946 .addImm(NumLPads >> 16));
6949 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6954 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6959 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6960 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6961 .addJumpTableIndex(MJTI)
6964 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6967 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6968 .addReg(NewVReg3, RegState::Kill)
6970 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6972 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6973 .addReg(NewVReg4, RegState::Kill)
6975 .addJumpTableIndex(MJTI)
6977 } else if (Subtarget->isThumb()) {
6978 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6979 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6982 .addMemOperand(FIMMOLd));
6984 if (NumLPads < 256) {
6985 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6989 MachineConstantPool *ConstantPool = MF->getConstantPool();
6990 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6991 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6993 // MachineConstantPool wants an explicit alignment.
6994 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6996 Align = getDataLayout()->getTypeAllocSize(C->getType());
6997 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6999 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7000 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7001 .addReg(VReg1, RegState::Define)
7002 .addConstantPoolIndex(Idx));
7003 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7008 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7013 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7014 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7015 .addReg(ARM::CPSR, RegState::Define)
7019 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7020 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7021 .addJumpTableIndex(MJTI)
7024 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7025 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7026 .addReg(ARM::CPSR, RegState::Define)
7027 .addReg(NewVReg2, RegState::Kill)
7030 MachineMemOperand *JTMMOLd =
7031 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7032 MachineMemOperand::MOLoad, 4, 4);
7034 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7035 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7036 .addReg(NewVReg4, RegState::Kill)
7038 .addMemOperand(JTMMOLd));
7040 unsigned NewVReg6 = NewVReg5;
7041 if (RelocM == Reloc::PIC_) {
7042 NewVReg6 = MRI->createVirtualRegister(TRC);
7043 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7044 .addReg(ARM::CPSR, RegState::Define)
7045 .addReg(NewVReg5, RegState::Kill)
7049 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7050 .addReg(NewVReg6, RegState::Kill)
7051 .addJumpTableIndex(MJTI)
7054 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7055 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7058 .addMemOperand(FIMMOLd));
7060 if (NumLPads < 256) {
7061 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7064 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7065 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7066 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7067 .addImm(NumLPads & 0xFFFF));
7069 unsigned VReg2 = VReg1;
7070 if ((NumLPads & 0xFFFF0000) != 0) {
7071 VReg2 = MRI->createVirtualRegister(TRC);
7072 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7074 .addImm(NumLPads >> 16));
7077 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7081 MachineConstantPool *ConstantPool = MF->getConstantPool();
7082 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7083 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7085 // MachineConstantPool wants an explicit alignment.
7086 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7088 Align = getDataLayout()->getTypeAllocSize(C->getType());
7089 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7091 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7092 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7093 .addReg(VReg1, RegState::Define)
7094 .addConstantPoolIndex(Idx)
7096 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7098 .addReg(VReg1, RegState::Kill));
7101 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7106 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7108 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7110 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7111 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7112 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7113 .addJumpTableIndex(MJTI)
7116 MachineMemOperand *JTMMOLd =
7117 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7118 MachineMemOperand::MOLoad, 4, 4);
7119 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7121 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7122 .addReg(NewVReg3, RegState::Kill)
7125 .addMemOperand(JTMMOLd));
7127 if (RelocM == Reloc::PIC_) {
7128 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7129 .addReg(NewVReg5, RegState::Kill)
7131 .addJumpTableIndex(MJTI)
7134 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7135 .addReg(NewVReg5, RegState::Kill)
7136 .addJumpTableIndex(MJTI)
7141 // Add the jump table entries as successors to the MBB.
7142 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7143 for (std::vector<MachineBasicBlock*>::iterator
7144 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7145 MachineBasicBlock *CurMBB = *I;
7146 if (SeenMBBs.insert(CurMBB))
7147 DispContBB->addSuccessor(CurMBB);
7150 // N.B. the order the invoke BBs are processed in doesn't matter here.
7151 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
7152 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7153 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
7154 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
7155 MachineBasicBlock *BB = *I;
7157 // Remove the landing pad successor from the invoke block and replace it
7158 // with the new dispatch block.
7159 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7161 while (!Successors.empty()) {
7162 MachineBasicBlock *SMBB = Successors.pop_back_val();
7163 if (SMBB->isLandingPad()) {
7164 BB->removeSuccessor(SMBB);
7165 MBBLPads.push_back(SMBB);
7169 BB->addSuccessor(DispatchBB);
7171 // Find the invoke call and mark all of the callee-saved registers as
7172 // 'implicit defined' so that they're spilled. This prevents code from
7173 // moving instructions to before the EH block, where they will never be
7175 for (MachineBasicBlock::reverse_iterator
7176 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7177 if (!II->isCall()) continue;
7179 DenseMap<unsigned, bool> DefRegs;
7180 for (MachineInstr::mop_iterator
7181 OI = II->operands_begin(), OE = II->operands_end();
7183 if (!OI->isReg()) continue;
7184 DefRegs[OI->getReg()] = true;
7187 MachineInstrBuilder MIB(*MF, &*II);
7189 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7190 unsigned Reg = SavedRegs[i];
7191 if (Subtarget->isThumb2() &&
7192 !ARM::tGPRRegClass.contains(Reg) &&
7193 !ARM::hGPRRegClass.contains(Reg))
7195 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7197 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7200 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7207 // Mark all former landing pads as non-landing pads. The dispatch is the only
7209 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7210 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7211 (*I)->setIsLandingPad(false);
7213 // The instruction is gone now.
7214 MI->eraseFromParent();
7220 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7221 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7222 E = MBB->succ_end(); I != E; ++I)
7225 llvm_unreachable("Expecting a BB with two successors!");
7228 /// Return the load opcode for a given load size. If load size >= 8,
7229 /// neon opcode will be returned.
7230 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7232 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7233 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7235 return LdSize == 4 ? ARM::tLDRi
7236 : LdSize == 2 ? ARM::tLDRHi
7237 : LdSize == 1 ? ARM::tLDRBi : 0;
7239 return LdSize == 4 ? ARM::t2LDR_POST
7240 : LdSize == 2 ? ARM::t2LDRH_POST
7241 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7242 return LdSize == 4 ? ARM::LDR_POST_IMM
7243 : LdSize == 2 ? ARM::LDRH_POST
7244 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7247 /// Return the store opcode for a given store size. If store size >= 8,
7248 /// neon opcode will be returned.
7249 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7251 return StSize == 16 ? ARM::VST1q32wb_fixed
7252 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7254 return StSize == 4 ? ARM::tSTRi
7255 : StSize == 2 ? ARM::tSTRHi
7256 : StSize == 1 ? ARM::tSTRBi : 0;
7258 return StSize == 4 ? ARM::t2STR_POST
7259 : StSize == 2 ? ARM::t2STRH_POST
7260 : StSize == 1 ? ARM::t2STRB_POST : 0;
7261 return StSize == 4 ? ARM::STR_POST_IMM
7262 : StSize == 2 ? ARM::STRH_POST
7263 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7266 /// Emit a post-increment load operation with given size. The instructions
7267 /// will be added to BB at Pos.
7268 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7269 const TargetInstrInfo *TII, DebugLoc dl,
7270 unsigned LdSize, unsigned Data, unsigned AddrIn,
7271 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7272 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7273 assert(LdOpc != 0 && "Should have a load opcode");
7275 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7276 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7278 } else if (IsThumb1) {
7279 // load + update AddrIn
7280 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7281 .addReg(AddrIn).addImm(0));
7282 MachineInstrBuilder MIB =
7283 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7284 MIB = AddDefaultT1CC(MIB);
7285 MIB.addReg(AddrIn).addImm(LdSize);
7286 AddDefaultPred(MIB);
7287 } else if (IsThumb2) {
7288 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7289 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7292 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7293 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7294 .addReg(0).addImm(LdSize));
7298 /// Emit a post-increment store operation with given size. The instructions
7299 /// will be added to BB at Pos.
7300 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7301 const TargetInstrInfo *TII, DebugLoc dl,
7302 unsigned StSize, unsigned Data, unsigned AddrIn,
7303 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7304 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7305 assert(StOpc != 0 && "Should have a store opcode");
7307 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7308 .addReg(AddrIn).addImm(0).addReg(Data));
7309 } else if (IsThumb1) {
7310 // store + update AddrIn
7311 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7312 .addReg(AddrIn).addImm(0));
7313 MachineInstrBuilder MIB =
7314 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7315 MIB = AddDefaultT1CC(MIB);
7316 MIB.addReg(AddrIn).addImm(StSize);
7317 AddDefaultPred(MIB);
7318 } else if (IsThumb2) {
7319 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7320 .addReg(Data).addReg(AddrIn).addImm(StSize));
7322 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7323 .addReg(Data).addReg(AddrIn).addReg(0)
7329 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7330 MachineBasicBlock *BB) const {
7331 // This pseudo instruction has 3 operands: dst, src, size
7332 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7333 // Otherwise, we will generate unrolled scalar copies.
7334 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7335 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7336 MachineFunction::iterator It = BB;
7339 unsigned dest = MI->getOperand(0).getReg();
7340 unsigned src = MI->getOperand(1).getReg();
7341 unsigned SizeVal = MI->getOperand(2).getImm();
7342 unsigned Align = MI->getOperand(3).getImm();
7343 DebugLoc dl = MI->getDebugLoc();
7345 MachineFunction *MF = BB->getParent();
7346 MachineRegisterInfo &MRI = MF->getRegInfo();
7347 unsigned UnitSize = 0;
7348 const TargetRegisterClass *TRC = 0;
7349 const TargetRegisterClass *VecTRC = 0;
7351 bool IsThumb1 = Subtarget->isThumb1Only();
7352 bool IsThumb2 = Subtarget->isThumb2();
7356 } else if (Align & 2) {
7359 // Check whether we can use NEON instructions.
7360 if (!MF->getFunction()->getAttributes().
7361 hasAttribute(AttributeSet::FunctionIndex,
7362 Attribute::NoImplicitFloat) &&
7363 Subtarget->hasNEON()) {
7364 if ((Align % 16 == 0) && SizeVal >= 16)
7366 else if ((Align % 8 == 0) && SizeVal >= 8)
7369 // Can't use NEON instructions.
7374 // Select the correct opcode and register class for unit size load/store
7375 bool IsNeon = UnitSize >= 8;
7376 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
7377 : (const TargetRegisterClass *)&ARM::GPRRegClass;
7379 VecTRC = UnitSize == 16
7380 ? (const TargetRegisterClass *)&ARM::DPairRegClass
7382 ? (const TargetRegisterClass *)&ARM::DPRRegClass
7385 unsigned BytesLeft = SizeVal % UnitSize;
7386 unsigned LoopSize = SizeVal - BytesLeft;
7388 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7389 // Use LDR and STR to copy.
7390 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7391 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7392 unsigned srcIn = src;
7393 unsigned destIn = dest;
7394 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7395 unsigned srcOut = MRI.createVirtualRegister(TRC);
7396 unsigned destOut = MRI.createVirtualRegister(TRC);
7397 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7398 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7399 IsThumb1, IsThumb2);
7400 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7401 IsThumb1, IsThumb2);
7406 // Handle the leftover bytes with LDRB and STRB.
7407 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7408 // [destOut] = STRB_POST(scratch, destIn, 1)
7409 for (unsigned i = 0; i < BytesLeft; i++) {
7410 unsigned srcOut = MRI.createVirtualRegister(TRC);
7411 unsigned destOut = MRI.createVirtualRegister(TRC);
7412 unsigned scratch = MRI.createVirtualRegister(TRC);
7413 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7414 IsThumb1, IsThumb2);
7415 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7416 IsThumb1, IsThumb2);
7420 MI->eraseFromParent(); // The instruction is gone now.
7424 // Expand the pseudo op to a loop.
7427 // movw varEnd, # --> with thumb2
7429 // ldrcp varEnd, idx --> without thumb2
7430 // fallthrough --> loopMBB
7432 // PHI varPhi, varEnd, varLoop
7433 // PHI srcPhi, src, srcLoop
7434 // PHI destPhi, dst, destLoop
7435 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7436 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7437 // subs varLoop, varPhi, #UnitSize
7439 // fallthrough --> exitMBB
7441 // epilogue to handle left-over bytes
7442 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7443 // [destOut] = STRB_POST(scratch, destLoop, 1)
7444 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7445 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7446 MF->insert(It, loopMBB);
7447 MF->insert(It, exitMBB);
7449 // Transfer the remainder of BB and its successor edges to exitMBB.
7450 exitMBB->splice(exitMBB->begin(), BB,
7451 llvm::next(MachineBasicBlock::iterator(MI)),
7453 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7455 // Load an immediate to varEnd.
7456 unsigned varEnd = MRI.createVirtualRegister(TRC);
7458 unsigned Vtmp = varEnd;
7459 if ((LoopSize & 0xFFFF0000) != 0)
7460 Vtmp = MRI.createVirtualRegister(TRC);
7461 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7462 .addImm(LoopSize & 0xFFFF));
7464 if ((LoopSize & 0xFFFF0000) != 0)
7465 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7466 .addReg(Vtmp).addImm(LoopSize >> 16));
7468 MachineConstantPool *ConstantPool = MF->getConstantPool();
7469 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7470 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7472 // MachineConstantPool wants an explicit alignment.
7473 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7475 Align = getDataLayout()->getTypeAllocSize(C->getType());
7476 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7479 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7480 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7482 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7483 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7485 BB->addSuccessor(loopMBB);
7487 // Generate the loop body:
7488 // varPhi = PHI(varLoop, varEnd)
7489 // srcPhi = PHI(srcLoop, src)
7490 // destPhi = PHI(destLoop, dst)
7491 MachineBasicBlock *entryBB = BB;
7493 unsigned varLoop = MRI.createVirtualRegister(TRC);
7494 unsigned varPhi = MRI.createVirtualRegister(TRC);
7495 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7496 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7497 unsigned destLoop = MRI.createVirtualRegister(TRC);
7498 unsigned destPhi = MRI.createVirtualRegister(TRC);
7500 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7501 .addReg(varLoop).addMBB(loopMBB)
7502 .addReg(varEnd).addMBB(entryBB);
7503 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7504 .addReg(srcLoop).addMBB(loopMBB)
7505 .addReg(src).addMBB(entryBB);
7506 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7507 .addReg(destLoop).addMBB(loopMBB)
7508 .addReg(dest).addMBB(entryBB);
7510 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7511 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7512 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7513 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7514 IsThumb1, IsThumb2);
7515 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7516 IsThumb1, IsThumb2);
7518 // Decrement loop variable by UnitSize.
7520 MachineInstrBuilder MIB =
7521 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7522 MIB = AddDefaultT1CC(MIB);
7523 MIB.addReg(varPhi).addImm(UnitSize);
7524 AddDefaultPred(MIB);
7526 MachineInstrBuilder MIB =
7527 BuildMI(*BB, BB->end(), dl,
7528 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7529 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7530 MIB->getOperand(5).setReg(ARM::CPSR);
7531 MIB->getOperand(5).setIsDef(true);
7533 BuildMI(*BB, BB->end(), dl,
7534 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7535 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7537 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7538 BB->addSuccessor(loopMBB);
7539 BB->addSuccessor(exitMBB);
7541 // Add epilogue to handle BytesLeft.
7543 MachineInstr *StartOfExit = exitMBB->begin();
7545 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7546 // [destOut] = STRB_POST(scratch, destLoop, 1)
7547 unsigned srcIn = srcLoop;
7548 unsigned destIn = destLoop;
7549 for (unsigned i = 0; i < BytesLeft; i++) {
7550 unsigned srcOut = MRI.createVirtualRegister(TRC);
7551 unsigned destOut = MRI.createVirtualRegister(TRC);
7552 unsigned scratch = MRI.createVirtualRegister(TRC);
7553 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7554 IsThumb1, IsThumb2);
7555 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7556 IsThumb1, IsThumb2);
7561 MI->eraseFromParent(); // The instruction is gone now.
7566 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7567 MachineBasicBlock *BB) const {
7568 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7569 DebugLoc dl = MI->getDebugLoc();
7570 bool isThumb2 = Subtarget->isThumb2();
7571 switch (MI->getOpcode()) {
7574 llvm_unreachable("Unexpected instr type to insert");
7576 // The Thumb2 pre-indexed stores have the same MI operands, they just
7577 // define them differently in the .td files from the isel patterns, so
7578 // they need pseudos.
7579 case ARM::t2STR_preidx:
7580 MI->setDesc(TII->get(ARM::t2STR_PRE));
7582 case ARM::t2STRB_preidx:
7583 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7585 case ARM::t2STRH_preidx:
7586 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7589 case ARM::STRi_preidx:
7590 case ARM::STRBi_preidx: {
7591 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7592 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7593 // Decode the offset.
7594 unsigned Offset = MI->getOperand(4).getImm();
7595 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7596 Offset = ARM_AM::getAM2Offset(Offset);
7600 MachineMemOperand *MMO = *MI->memoperands_begin();
7601 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7602 .addOperand(MI->getOperand(0)) // Rn_wb
7603 .addOperand(MI->getOperand(1)) // Rt
7604 .addOperand(MI->getOperand(2)) // Rn
7605 .addImm(Offset) // offset (skip GPR==zero_reg)
7606 .addOperand(MI->getOperand(5)) // pred
7607 .addOperand(MI->getOperand(6))
7608 .addMemOperand(MMO);
7609 MI->eraseFromParent();
7612 case ARM::STRr_preidx:
7613 case ARM::STRBr_preidx:
7614 case ARM::STRH_preidx: {
7616 switch (MI->getOpcode()) {
7617 default: llvm_unreachable("unexpected opcode!");
7618 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7619 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7620 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7622 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7623 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7624 MIB.addOperand(MI->getOperand(i));
7625 MI->eraseFromParent();
7628 case ARM::ATOMIC_LOAD_ADD_I8:
7629 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7630 case ARM::ATOMIC_LOAD_ADD_I16:
7631 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7632 case ARM::ATOMIC_LOAD_ADD_I32:
7633 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7635 case ARM::ATOMIC_LOAD_AND_I8:
7636 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7637 case ARM::ATOMIC_LOAD_AND_I16:
7638 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7639 case ARM::ATOMIC_LOAD_AND_I32:
7640 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7642 case ARM::ATOMIC_LOAD_OR_I8:
7643 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7644 case ARM::ATOMIC_LOAD_OR_I16:
7645 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7646 case ARM::ATOMIC_LOAD_OR_I32:
7647 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7649 case ARM::ATOMIC_LOAD_XOR_I8:
7650 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7651 case ARM::ATOMIC_LOAD_XOR_I16:
7652 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7653 case ARM::ATOMIC_LOAD_XOR_I32:
7654 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7656 case ARM::ATOMIC_LOAD_NAND_I8:
7657 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7658 case ARM::ATOMIC_LOAD_NAND_I16:
7659 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7660 case ARM::ATOMIC_LOAD_NAND_I32:
7661 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7663 case ARM::ATOMIC_LOAD_SUB_I8:
7664 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7665 case ARM::ATOMIC_LOAD_SUB_I16:
7666 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7667 case ARM::ATOMIC_LOAD_SUB_I32:
7668 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7670 case ARM::ATOMIC_LOAD_MIN_I8:
7671 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7672 case ARM::ATOMIC_LOAD_MIN_I16:
7673 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7674 case ARM::ATOMIC_LOAD_MIN_I32:
7675 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7677 case ARM::ATOMIC_LOAD_MAX_I8:
7678 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7679 case ARM::ATOMIC_LOAD_MAX_I16:
7680 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7681 case ARM::ATOMIC_LOAD_MAX_I32:
7682 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7684 case ARM::ATOMIC_LOAD_UMIN_I8:
7685 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7686 case ARM::ATOMIC_LOAD_UMIN_I16:
7687 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7688 case ARM::ATOMIC_LOAD_UMIN_I32:
7689 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7691 case ARM::ATOMIC_LOAD_UMAX_I8:
7692 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7693 case ARM::ATOMIC_LOAD_UMAX_I16:
7694 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7695 case ARM::ATOMIC_LOAD_UMAX_I32:
7696 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7698 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7699 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7700 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7702 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7703 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7704 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7706 case ARM::ATOMIC_LOAD_I64:
7707 return EmitAtomicLoad64(MI, BB);
7709 case ARM::ATOMIC_LOAD_ADD_I64:
7710 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7711 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7712 /*NeedsCarry*/ true);
7713 case ARM::ATOMIC_LOAD_SUB_I64:
7714 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7715 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7716 /*NeedsCarry*/ true);
7717 case ARM::ATOMIC_LOAD_OR_I64:
7718 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7719 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7720 case ARM::ATOMIC_LOAD_XOR_I64:
7721 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7722 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7723 case ARM::ATOMIC_LOAD_AND_I64:
7724 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7725 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7726 case ARM::ATOMIC_STORE_I64:
7727 case ARM::ATOMIC_SWAP_I64:
7728 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7729 case ARM::ATOMIC_CMP_SWAP_I64:
7730 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7731 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7732 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7733 case ARM::ATOMIC_LOAD_MIN_I64:
7734 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7735 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7736 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7737 /*IsMinMax*/ true, ARMCC::LT);
7738 case ARM::ATOMIC_LOAD_MAX_I64:
7739 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7740 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7741 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7742 /*IsMinMax*/ true, ARMCC::GE);
7743 case ARM::ATOMIC_LOAD_UMIN_I64:
7744 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7745 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7746 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7747 /*IsMinMax*/ true, ARMCC::LO);
7748 case ARM::ATOMIC_LOAD_UMAX_I64:
7749 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7750 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7751 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7752 /*IsMinMax*/ true, ARMCC::HS);
7754 case ARM::tMOVCCr_pseudo: {
7755 // To "insert" a SELECT_CC instruction, we actually have to insert the
7756 // diamond control-flow pattern. The incoming instruction knows the
7757 // destination vreg to set, the condition code register to branch on, the
7758 // true/false values to select between, and a branch opcode to use.
7759 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7760 MachineFunction::iterator It = BB;
7766 // cmpTY ccX, r1, r2
7768 // fallthrough --> copy0MBB
7769 MachineBasicBlock *thisMBB = BB;
7770 MachineFunction *F = BB->getParent();
7771 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7772 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7773 F->insert(It, copy0MBB);
7774 F->insert(It, sinkMBB);
7776 // Transfer the remainder of BB and its successor edges to sinkMBB.
7777 sinkMBB->splice(sinkMBB->begin(), BB,
7778 llvm::next(MachineBasicBlock::iterator(MI)),
7780 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7782 BB->addSuccessor(copy0MBB);
7783 BB->addSuccessor(sinkMBB);
7785 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7786 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7789 // %FalseValue = ...
7790 // # fallthrough to sinkMBB
7793 // Update machine-CFG edges
7794 BB->addSuccessor(sinkMBB);
7797 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7800 BuildMI(*BB, BB->begin(), dl,
7801 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7802 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7803 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7805 MI->eraseFromParent(); // The pseudo instruction is gone now.
7810 case ARM::BCCZi64: {
7811 // If there is an unconditional branch to the other successor, remove it.
7812 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7814 // Compare both parts that make up the double comparison separately for
7816 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7818 unsigned LHS1 = MI->getOperand(1).getReg();
7819 unsigned LHS2 = MI->getOperand(2).getReg();
7821 AddDefaultPred(BuildMI(BB, dl,
7822 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7823 .addReg(LHS1).addImm(0));
7824 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7825 .addReg(LHS2).addImm(0)
7826 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7828 unsigned RHS1 = MI->getOperand(3).getReg();
7829 unsigned RHS2 = MI->getOperand(4).getReg();
7830 AddDefaultPred(BuildMI(BB, dl,
7831 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7832 .addReg(LHS1).addReg(RHS1));
7833 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7834 .addReg(LHS2).addReg(RHS2)
7835 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7838 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7839 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7840 if (MI->getOperand(0).getImm() == ARMCC::NE)
7841 std::swap(destMBB, exitMBB);
7843 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7844 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7846 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7848 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7850 MI->eraseFromParent(); // The pseudo instruction is gone now.
7854 case ARM::Int_eh_sjlj_setjmp:
7855 case ARM::Int_eh_sjlj_setjmp_nofp:
7856 case ARM::tInt_eh_sjlj_setjmp:
7857 case ARM::t2Int_eh_sjlj_setjmp:
7858 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7859 EmitSjLjDispatchBlock(MI, BB);
7864 // To insert an ABS instruction, we have to insert the
7865 // diamond control-flow pattern. The incoming instruction knows the
7866 // source vreg to test against 0, the destination vreg to set,
7867 // the condition code register to branch on, the
7868 // true/false values to select between, and a branch opcode to use.
7873 // BCC (branch to SinkBB if V0 >= 0)
7874 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7875 // SinkBB: V1 = PHI(V2, V3)
7876 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7877 MachineFunction::iterator BBI = BB;
7879 MachineFunction *Fn = BB->getParent();
7880 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7881 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7882 Fn->insert(BBI, RSBBB);
7883 Fn->insert(BBI, SinkBB);
7885 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7886 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7887 bool isThumb2 = Subtarget->isThumb2();
7888 MachineRegisterInfo &MRI = Fn->getRegInfo();
7889 // In Thumb mode S must not be specified if source register is the SP or
7890 // PC and if destination register is the SP, so restrict register class
7891 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7892 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7893 (const TargetRegisterClass*)&ARM::GPRRegClass);
7895 // Transfer the remainder of BB and its successor edges to sinkMBB.
7896 SinkBB->splice(SinkBB->begin(), BB,
7897 llvm::next(MachineBasicBlock::iterator(MI)),
7899 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7901 BB->addSuccessor(RSBBB);
7902 BB->addSuccessor(SinkBB);
7904 // fall through to SinkMBB
7905 RSBBB->addSuccessor(SinkBB);
7907 // insert a cmp at the end of BB
7908 AddDefaultPred(BuildMI(BB, dl,
7909 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7910 .addReg(ABSSrcReg).addImm(0));
7912 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7914 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7915 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7917 // insert rsbri in RSBBB
7918 // Note: BCC and rsbri will be converted into predicated rsbmi
7919 // by if-conversion pass
7920 BuildMI(*RSBBB, RSBBB->begin(), dl,
7921 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7922 .addReg(ABSSrcReg, RegState::Kill)
7923 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7925 // insert PHI in SinkBB,
7926 // reuse ABSDstReg to not change uses of ABS instruction
7927 BuildMI(*SinkBB, SinkBB->begin(), dl,
7928 TII->get(ARM::PHI), ABSDstReg)
7929 .addReg(NewRsbDstReg).addMBB(RSBBB)
7930 .addReg(ABSSrcReg).addMBB(BB);
7932 // remove ABS instruction
7933 MI->eraseFromParent();
7935 // return last added BB
7938 case ARM::COPY_STRUCT_BYVAL_I32:
7940 return EmitStructByval(MI, BB);
7944 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7945 SDNode *Node) const {
7946 if (!MI->hasPostISelHook()) {
7947 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7948 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7952 const MCInstrDesc *MCID = &MI->getDesc();
7953 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7954 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7955 // operand is still set to noreg. If needed, set the optional operand's
7956 // register to CPSR, and remove the redundant implicit def.
7958 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7960 // Rename pseudo opcodes.
7961 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7963 const ARMBaseInstrInfo *TII =
7964 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7965 MCID = &TII->get(NewOpc);
7967 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7968 "converted opcode should be the same except for cc_out");
7972 // Add the optional cc_out operand
7973 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7975 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7977 // Any ARM instruction that sets the 's' bit should specify an optional
7978 // "cc_out" operand in the last operand position.
7979 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7980 assert(!NewOpc && "Optional cc_out operand required");
7983 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7984 // since we already have an optional CPSR def.
7985 bool definesCPSR = false;
7986 bool deadCPSR = false;
7987 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7989 const MachineOperand &MO = MI->getOperand(i);
7990 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7994 MI->RemoveOperand(i);
7999 assert(!NewOpc && "Optional cc_out operand required");
8002 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8004 assert(!MI->getOperand(ccOutIdx).getReg() &&
8005 "expect uninitialized optional cc_out operand");
8009 // If this instruction was defined with an optional CPSR def and its dag node
8010 // had a live implicit CPSR def, then activate the optional CPSR def.
8011 MachineOperand &MO = MI->getOperand(ccOutIdx);
8012 MO.setReg(ARM::CPSR);
8016 //===----------------------------------------------------------------------===//
8017 // ARM Optimization Hooks
8018 //===----------------------------------------------------------------------===//
8020 // Helper function that checks if N is a null or all ones constant.
8021 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8022 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8025 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8028 // Return true if N is conditionally 0 or all ones.
8029 // Detects these expressions where cc is an i1 value:
8031 // (select cc 0, y) [AllOnes=0]
8032 // (select cc y, 0) [AllOnes=0]
8033 // (zext cc) [AllOnes=0]
8034 // (sext cc) [AllOnes=0/1]
8035 // (select cc -1, y) [AllOnes=1]
8036 // (select cc y, -1) [AllOnes=1]
8038 // Invert is set when N is the null/all ones constant when CC is false.
8039 // OtherOp is set to the alternative value of N.
8040 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8041 SDValue &CC, bool &Invert,
8043 SelectionDAG &DAG) {
8044 switch (N->getOpcode()) {
8045 default: return false;
8047 CC = N->getOperand(0);
8048 SDValue N1 = N->getOperand(1);
8049 SDValue N2 = N->getOperand(2);
8050 if (isZeroOrAllOnes(N1, AllOnes)) {
8055 if (isZeroOrAllOnes(N2, AllOnes)) {
8062 case ISD::ZERO_EXTEND:
8063 // (zext cc) can never be the all ones value.
8067 case ISD::SIGN_EXTEND: {
8068 EVT VT = N->getValueType(0);
8069 CC = N->getOperand(0);
8070 if (CC.getValueType() != MVT::i1)
8074 // When looking for an AllOnes constant, N is an sext, and the 'other'
8076 OtherOp = DAG.getConstant(0, VT);
8077 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8078 // When looking for a 0 constant, N can be zext or sext.
8079 OtherOp = DAG.getConstant(1, VT);
8081 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
8087 // Combine a constant select operand into its use:
8089 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8090 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8091 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8092 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8093 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8095 // The transform is rejected if the select doesn't have a constant operand that
8096 // is null, or all ones when AllOnes is set.
8098 // Also recognize sext/zext from i1:
8100 // (add (zext cc), x) -> (select cc (add x, 1), x)
8101 // (add (sext cc), x) -> (select cc (add x, -1), x)
8103 // These transformations eventually create predicated instructions.
8105 // @param N The node to transform.
8106 // @param Slct The N operand that is a select.
8107 // @param OtherOp The other N operand (x above).
8108 // @param DCI Context.
8109 // @param AllOnes Require the select constant to be all ones instead of null.
8110 // @returns The new node, or SDValue() on failure.
8112 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8113 TargetLowering::DAGCombinerInfo &DCI,
8114 bool AllOnes = false) {
8115 SelectionDAG &DAG = DCI.DAG;
8116 EVT VT = N->getValueType(0);
8117 SDValue NonConstantVal;
8120 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8121 NonConstantVal, DAG))
8124 // Slct is now know to be the desired identity constant when CC is true.
8125 SDValue TrueVal = OtherOp;
8126 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8127 OtherOp, NonConstantVal);
8128 // Unless SwapSelectOps says CC should be false.
8130 std::swap(TrueVal, FalseVal);
8132 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8133 CCOp, TrueVal, FalseVal);
8136 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8138 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8139 TargetLowering::DAGCombinerInfo &DCI) {
8140 SDValue N0 = N->getOperand(0);
8141 SDValue N1 = N->getOperand(1);
8142 if (N0.getNode()->hasOneUse()) {
8143 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8144 if (Result.getNode())
8147 if (N1.getNode()->hasOneUse()) {
8148 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8149 if (Result.getNode())
8155 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8156 // (only after legalization).
8157 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8158 TargetLowering::DAGCombinerInfo &DCI,
8159 const ARMSubtarget *Subtarget) {
8161 // Only perform optimization if after legalize, and if NEON is available. We
8162 // also expected both operands to be BUILD_VECTORs.
8163 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8164 || N0.getOpcode() != ISD::BUILD_VECTOR
8165 || N1.getOpcode() != ISD::BUILD_VECTOR)
8168 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8169 EVT VT = N->getValueType(0);
8170 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8173 // Check that the vector operands are of the right form.
8174 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8175 // operands, where N is the size of the formed vector.
8176 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8177 // index such that we have a pair wise add pattern.
8179 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8180 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8182 SDValue Vec = N0->getOperand(0)->getOperand(0);
8183 SDNode *V = Vec.getNode();
8184 unsigned nextIndex = 0;
8186 // For each operands to the ADD which are BUILD_VECTORs,
8187 // check to see if each of their operands are an EXTRACT_VECTOR with
8188 // the same vector and appropriate index.
8189 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8190 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8191 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8193 SDValue ExtVec0 = N0->getOperand(i);
8194 SDValue ExtVec1 = N1->getOperand(i);
8196 // First operand is the vector, verify its the same.
8197 if (V != ExtVec0->getOperand(0).getNode() ||
8198 V != ExtVec1->getOperand(0).getNode())
8201 // Second is the constant, verify its correct.
8202 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8203 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8205 // For the constant, we want to see all the even or all the odd.
8206 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8207 || C1->getZExtValue() != nextIndex+1)
8216 // Create VPADDL node.
8217 SelectionDAG &DAG = DCI.DAG;
8218 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8220 // Build operand list.
8221 SmallVector<SDValue, 8> Ops;
8222 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
8223 TLI.getPointerTy()));
8225 // Input is the vector.
8228 // Get widened type and narrowed type.
8230 unsigned numElem = VT.getVectorNumElements();
8231 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
8232 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8233 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8234 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8236 llvm_unreachable("Invalid vector element type for padd optimization.");
8239 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
8240 widenType, &Ops[0], Ops.size());
8241 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp);
8244 static SDValue findMUL_LOHI(SDValue V) {
8245 if (V->getOpcode() == ISD::UMUL_LOHI ||
8246 V->getOpcode() == ISD::SMUL_LOHI)
8251 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8252 TargetLowering::DAGCombinerInfo &DCI,
8253 const ARMSubtarget *Subtarget) {
8255 if (Subtarget->isThumb1Only()) return SDValue();
8257 // Only perform the checks after legalize when the pattern is available.
8258 if (DCI.isBeforeLegalize()) return SDValue();
8260 // Look for multiply add opportunities.
8261 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8262 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8263 // a glue link from the first add to the second add.
8264 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8265 // a S/UMLAL instruction.
8268 // \ / \ [no multiline comment]
8274 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8275 SDValue AddcOp0 = AddcNode->getOperand(0);
8276 SDValue AddcOp1 = AddcNode->getOperand(1);
8278 // Check if the two operands are from the same mul_lohi node.
8279 if (AddcOp0.getNode() == AddcOp1.getNode())
8282 assert(AddcNode->getNumValues() == 2 &&
8283 AddcNode->getValueType(0) == MVT::i32 &&
8284 "Expect ADDC with two result values. First: i32");
8286 // Check that we have a glued ADDC node.
8287 if (AddcNode->getValueType(1) != MVT::Glue)
8290 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8291 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8292 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8293 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8294 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8297 // Look for the glued ADDE.
8298 SDNode* AddeNode = AddcNode->getGluedUser();
8299 if (AddeNode == NULL)
8302 // Make sure it is really an ADDE.
8303 if (AddeNode->getOpcode() != ISD::ADDE)
8306 assert(AddeNode->getNumOperands() == 3 &&
8307 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8308 "ADDE node has the wrong inputs");
8310 // Check for the triangle shape.
8311 SDValue AddeOp0 = AddeNode->getOperand(0);
8312 SDValue AddeOp1 = AddeNode->getOperand(1);
8314 // Make sure that the ADDE operands are not coming from the same node.
8315 if (AddeOp0.getNode() == AddeOp1.getNode())
8318 // Find the MUL_LOHI node walking up ADDE's operands.
8319 bool IsLeftOperandMUL = false;
8320 SDValue MULOp = findMUL_LOHI(AddeOp0);
8321 if (MULOp == SDValue())
8322 MULOp = findMUL_LOHI(AddeOp1);
8324 IsLeftOperandMUL = true;
8325 if (MULOp == SDValue())
8328 // Figure out the right opcode.
8329 unsigned Opc = MULOp->getOpcode();
8330 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8332 // Figure out the high and low input values to the MLAL node.
8333 SDValue* HiMul = &MULOp;
8334 SDValue* HiAdd = NULL;
8335 SDValue* LoMul = NULL;
8336 SDValue* LowAdd = NULL;
8338 if (IsLeftOperandMUL)
8344 if (AddcOp0->getOpcode() == Opc) {
8348 if (AddcOp1->getOpcode() == Opc) {
8356 if (LoMul->getNode() != HiMul->getNode())
8359 // Create the merged node.
8360 SelectionDAG &DAG = DCI.DAG;
8362 // Build operand list.
8363 SmallVector<SDValue, 8> Ops;
8364 Ops.push_back(LoMul->getOperand(0));
8365 Ops.push_back(LoMul->getOperand(1));
8366 Ops.push_back(*LowAdd);
8367 Ops.push_back(*HiAdd);
8369 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8370 DAG.getVTList(MVT::i32, MVT::i32),
8371 &Ops[0], Ops.size());
8373 // Replace the ADDs' nodes uses by the MLA node's values.
8374 SDValue HiMLALResult(MLALNode.getNode(), 1);
8375 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8377 SDValue LoMLALResult(MLALNode.getNode(), 0);
8378 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8380 // Return original node to notify the driver to stop replacing.
8381 SDValue resNode(AddcNode, 0);
8385 /// PerformADDCCombine - Target-specific dag combine transform from
8386 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8387 static SDValue PerformADDCCombine(SDNode *N,
8388 TargetLowering::DAGCombinerInfo &DCI,
8389 const ARMSubtarget *Subtarget) {
8391 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8395 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8396 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8397 /// called with the default operands, and if that fails, with commuted
8399 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8400 TargetLowering::DAGCombinerInfo &DCI,
8401 const ARMSubtarget *Subtarget){
8403 // Attempt to create vpaddl for this add.
8404 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8405 if (Result.getNode())
8408 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8409 if (N0.getNode()->hasOneUse()) {
8410 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8411 if (Result.getNode()) return Result;
8416 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8418 static SDValue PerformADDCombine(SDNode *N,
8419 TargetLowering::DAGCombinerInfo &DCI,
8420 const ARMSubtarget *Subtarget) {
8421 SDValue N0 = N->getOperand(0);
8422 SDValue N1 = N->getOperand(1);
8424 // First try with the default operand order.
8425 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8426 if (Result.getNode())
8429 // If that didn't work, try again with the operands commuted.
8430 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8433 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8435 static SDValue PerformSUBCombine(SDNode *N,
8436 TargetLowering::DAGCombinerInfo &DCI) {
8437 SDValue N0 = N->getOperand(0);
8438 SDValue N1 = N->getOperand(1);
8440 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8441 if (N1.getNode()->hasOneUse()) {
8442 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8443 if (Result.getNode()) return Result;
8449 /// PerformVMULCombine
8450 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8451 /// special multiplier accumulator forwarding.
8457 // However, for (A + B) * (A + B),
8464 static SDValue PerformVMULCombine(SDNode *N,
8465 TargetLowering::DAGCombinerInfo &DCI,
8466 const ARMSubtarget *Subtarget) {
8467 if (!Subtarget->hasVMLxForwarding())
8470 SelectionDAG &DAG = DCI.DAG;
8471 SDValue N0 = N->getOperand(0);
8472 SDValue N1 = N->getOperand(1);
8473 unsigned Opcode = N0.getOpcode();
8474 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8475 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8476 Opcode = N1.getOpcode();
8477 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8478 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8486 EVT VT = N->getValueType(0);
8488 SDValue N00 = N0->getOperand(0);
8489 SDValue N01 = N0->getOperand(1);
8490 return DAG.getNode(Opcode, DL, VT,
8491 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8492 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8495 static SDValue PerformMULCombine(SDNode *N,
8496 TargetLowering::DAGCombinerInfo &DCI,
8497 const ARMSubtarget *Subtarget) {
8498 SelectionDAG &DAG = DCI.DAG;
8500 if (Subtarget->isThumb1Only())
8503 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8506 EVT VT = N->getValueType(0);
8507 if (VT.is64BitVector() || VT.is128BitVector())
8508 return PerformVMULCombine(N, DCI, Subtarget);
8512 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8516 int64_t MulAmt = C->getSExtValue();
8517 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8519 ShiftAmt = ShiftAmt & (32 - 1);
8520 SDValue V = N->getOperand(0);
8524 MulAmt >>= ShiftAmt;
8527 if (isPowerOf2_32(MulAmt - 1)) {
8528 // (mul x, 2^N + 1) => (add (shl x, N), x)
8529 Res = DAG.getNode(ISD::ADD, DL, VT,
8531 DAG.getNode(ISD::SHL, DL, VT,
8533 DAG.getConstant(Log2_32(MulAmt - 1),
8535 } else if (isPowerOf2_32(MulAmt + 1)) {
8536 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8537 Res = DAG.getNode(ISD::SUB, DL, VT,
8538 DAG.getNode(ISD::SHL, DL, VT,
8540 DAG.getConstant(Log2_32(MulAmt + 1),
8546 uint64_t MulAmtAbs = -MulAmt;
8547 if (isPowerOf2_32(MulAmtAbs + 1)) {
8548 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8549 Res = DAG.getNode(ISD::SUB, DL, VT,
8551 DAG.getNode(ISD::SHL, DL, VT,
8553 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8555 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8556 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8557 Res = DAG.getNode(ISD::ADD, DL, VT,
8559 DAG.getNode(ISD::SHL, DL, VT,
8561 DAG.getConstant(Log2_32(MulAmtAbs-1),
8563 Res = DAG.getNode(ISD::SUB, DL, VT,
8564 DAG.getConstant(0, MVT::i32),Res);
8571 Res = DAG.getNode(ISD::SHL, DL, VT,
8572 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8574 // Do not add new nodes to DAG combiner worklist.
8575 DCI.CombineTo(N, Res, false);
8579 static SDValue PerformANDCombine(SDNode *N,
8580 TargetLowering::DAGCombinerInfo &DCI,
8581 const ARMSubtarget *Subtarget) {
8583 // Attempt to use immediate-form VBIC
8584 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8586 EVT VT = N->getValueType(0);
8587 SelectionDAG &DAG = DCI.DAG;
8589 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8592 APInt SplatBits, SplatUndef;
8593 unsigned SplatBitSize;
8596 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8597 if (SplatBitSize <= 64) {
8599 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8600 SplatUndef.getZExtValue(), SplatBitSize,
8601 DAG, VbicVT, VT.is128BitVector(),
8603 if (Val.getNode()) {
8605 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8606 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8607 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8612 if (!Subtarget->isThumb1Only()) {
8613 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8614 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8615 if (Result.getNode())
8622 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8623 static SDValue PerformORCombine(SDNode *N,
8624 TargetLowering::DAGCombinerInfo &DCI,
8625 const ARMSubtarget *Subtarget) {
8626 // Attempt to use immediate-form VORR
8627 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8629 EVT VT = N->getValueType(0);
8630 SelectionDAG &DAG = DCI.DAG;
8632 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8635 APInt SplatBits, SplatUndef;
8636 unsigned SplatBitSize;
8638 if (BVN && Subtarget->hasNEON() &&
8639 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8640 if (SplatBitSize <= 64) {
8642 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8643 SplatUndef.getZExtValue(), SplatBitSize,
8644 DAG, VorrVT, VT.is128BitVector(),
8646 if (Val.getNode()) {
8648 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8649 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8650 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8655 if (!Subtarget->isThumb1Only()) {
8656 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8657 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8658 if (Result.getNode())
8662 // The code below optimizes (or (and X, Y), Z).
8663 // The AND operand needs to have a single user to make these optimizations
8665 SDValue N0 = N->getOperand(0);
8666 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8668 SDValue N1 = N->getOperand(1);
8670 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8671 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8672 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8674 unsigned SplatBitSize;
8677 APInt SplatBits0, SplatBits1;
8678 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8679 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8680 // Ensure that the second operand of both ands are constants
8681 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8682 HasAnyUndefs) && !HasAnyUndefs) {
8683 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8684 HasAnyUndefs) && !HasAnyUndefs) {
8685 // Ensure that the bit width of the constants are the same and that
8686 // the splat arguments are logical inverses as per the pattern we
8687 // are trying to simplify.
8688 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8689 SplatBits0 == ~SplatBits1) {
8690 // Canonicalize the vector type to make instruction selection
8692 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8693 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8697 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8703 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8706 // BFI is only available on V6T2+
8707 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8711 // 1) or (and A, mask), val => ARMbfi A, val, mask
8712 // iff (val & mask) == val
8714 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8715 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8716 // && mask == ~mask2
8717 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8718 // && ~mask == mask2
8719 // (i.e., copy a bitfield value into another bitfield of the same width)
8724 SDValue N00 = N0.getOperand(0);
8726 // The value and the mask need to be constants so we can verify this is
8727 // actually a bitfield set. If the mask is 0xffff, we can do better
8728 // via a movt instruction, so don't use BFI in that case.
8729 SDValue MaskOp = N0.getOperand(1);
8730 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8733 unsigned Mask = MaskC->getZExtValue();
8737 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8738 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8740 unsigned Val = N1C->getZExtValue();
8741 if ((Val & ~Mask) != Val)
8744 if (ARM::isBitFieldInvertedMask(Mask)) {
8745 Val >>= countTrailingZeros(~Mask);
8747 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8748 DAG.getConstant(Val, MVT::i32),
8749 DAG.getConstant(Mask, MVT::i32));
8751 // Do not add new nodes to DAG combiner worklist.
8752 DCI.CombineTo(N, Res, false);
8755 } else if (N1.getOpcode() == ISD::AND) {
8756 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8757 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8760 unsigned Mask2 = N11C->getZExtValue();
8762 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8764 if (ARM::isBitFieldInvertedMask(Mask) &&
8766 // The pack halfword instruction works better for masks that fit it,
8767 // so use that when it's available.
8768 if (Subtarget->hasT2ExtractPack() &&
8769 (Mask == 0xffff || Mask == 0xffff0000))
8772 unsigned amt = countTrailingZeros(Mask2);
8773 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8774 DAG.getConstant(amt, MVT::i32));
8775 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8776 DAG.getConstant(Mask, MVT::i32));
8777 // Do not add new nodes to DAG combiner worklist.
8778 DCI.CombineTo(N, Res, false);
8780 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8782 // The pack halfword instruction works better for masks that fit it,
8783 // so use that when it's available.
8784 if (Subtarget->hasT2ExtractPack() &&
8785 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8788 unsigned lsb = countTrailingZeros(Mask);
8789 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8790 DAG.getConstant(lsb, MVT::i32));
8791 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8792 DAG.getConstant(Mask2, MVT::i32));
8793 // Do not add new nodes to DAG combiner worklist.
8794 DCI.CombineTo(N, Res, false);
8799 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8800 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8801 ARM::isBitFieldInvertedMask(~Mask)) {
8802 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8803 // where lsb(mask) == #shamt and masked bits of B are known zero.
8804 SDValue ShAmt = N00.getOperand(1);
8805 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8806 unsigned LSB = countTrailingZeros(Mask);
8810 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8811 DAG.getConstant(~Mask, MVT::i32));
8813 // Do not add new nodes to DAG combiner worklist.
8814 DCI.CombineTo(N, Res, false);
8820 static SDValue PerformXORCombine(SDNode *N,
8821 TargetLowering::DAGCombinerInfo &DCI,
8822 const ARMSubtarget *Subtarget) {
8823 EVT VT = N->getValueType(0);
8824 SelectionDAG &DAG = DCI.DAG;
8826 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8829 if (!Subtarget->isThumb1Only()) {
8830 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8831 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8832 if (Result.getNode())
8839 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8840 /// the bits being cleared by the AND are not demanded by the BFI.
8841 static SDValue PerformBFICombine(SDNode *N,
8842 TargetLowering::DAGCombinerInfo &DCI) {
8843 SDValue N1 = N->getOperand(1);
8844 if (N1.getOpcode() == ISD::AND) {
8845 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8848 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8849 unsigned LSB = countTrailingZeros(~InvMask);
8850 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8851 unsigned Mask = (1 << Width)-1;
8852 unsigned Mask2 = N11C->getZExtValue();
8853 if ((Mask & (~Mask2)) == 0)
8854 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8855 N->getOperand(0), N1.getOperand(0),
8861 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8862 /// ARMISD::VMOVRRD.
8863 static SDValue PerformVMOVRRDCombine(SDNode *N,
8864 TargetLowering::DAGCombinerInfo &DCI) {
8865 // vmovrrd(vmovdrr x, y) -> x,y
8866 SDValue InDouble = N->getOperand(0);
8867 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8868 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8870 // vmovrrd(load f64) -> (load i32), (load i32)
8871 SDNode *InNode = InDouble.getNode();
8872 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8873 InNode->getValueType(0) == MVT::f64 &&
8874 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8875 !cast<LoadSDNode>(InNode)->isVolatile()) {
8876 // TODO: Should this be done for non-FrameIndex operands?
8877 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8879 SelectionDAG &DAG = DCI.DAG;
8881 SDValue BasePtr = LD->getBasePtr();
8882 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8883 LD->getPointerInfo(), LD->isVolatile(),
8884 LD->isNonTemporal(), LD->isInvariant(),
8885 LD->getAlignment());
8887 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8888 DAG.getConstant(4, MVT::i32));
8889 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8890 LD->getPointerInfo(), LD->isVolatile(),
8891 LD->isNonTemporal(), LD->isInvariant(),
8892 std::min(4U, LD->getAlignment() / 2));
8894 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8895 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8896 DCI.RemoveFromWorklist(LD);
8904 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8905 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8906 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8907 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8908 SDValue Op0 = N->getOperand(0);
8909 SDValue Op1 = N->getOperand(1);
8910 if (Op0.getOpcode() == ISD::BITCAST)
8911 Op0 = Op0.getOperand(0);
8912 if (Op1.getOpcode() == ISD::BITCAST)
8913 Op1 = Op1.getOperand(0);
8914 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8915 Op0.getNode() == Op1.getNode() &&
8916 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8917 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8918 N->getValueType(0), Op0.getOperand(0));
8922 /// PerformSTORECombine - Target-specific dag combine xforms for
8924 static SDValue PerformSTORECombine(SDNode *N,
8925 TargetLowering::DAGCombinerInfo &DCI) {
8926 StoreSDNode *St = cast<StoreSDNode>(N);
8927 if (St->isVolatile())
8930 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8931 // pack all of the elements in one place. Next, store to memory in fewer
8933 SDValue StVal = St->getValue();
8934 EVT VT = StVal.getValueType();
8935 if (St->isTruncatingStore() && VT.isVector()) {
8936 SelectionDAG &DAG = DCI.DAG;
8937 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8938 EVT StVT = St->getMemoryVT();
8939 unsigned NumElems = VT.getVectorNumElements();
8940 assert(StVT != VT && "Cannot truncate to the same type");
8941 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8942 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8944 // From, To sizes and ElemCount must be pow of two
8945 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8947 // We are going to use the original vector elt for storing.
8948 // Accumulated smaller vector elements must be a multiple of the store size.
8949 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8951 unsigned SizeRatio = FromEltSz / ToEltSz;
8952 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8954 // Create a type on which we perform the shuffle.
8955 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8956 NumElems*SizeRatio);
8957 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8960 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8961 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8962 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8964 // Can't shuffle using an illegal type.
8965 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8967 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8968 DAG.getUNDEF(WideVec.getValueType()),
8970 // At this point all of the data is stored at the bottom of the
8971 // register. We now need to save it to mem.
8973 // Find the largest store unit
8974 MVT StoreType = MVT::i8;
8975 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8976 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8977 MVT Tp = (MVT::SimpleValueType)tp;
8978 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8981 // Didn't find a legal store type.
8982 if (!TLI.isTypeLegal(StoreType))
8985 // Bitcast the original vector into a vector of store-size units
8986 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8987 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8988 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8989 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8990 SmallVector<SDValue, 8> Chains;
8991 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8992 TLI.getPointerTy());
8993 SDValue BasePtr = St->getBasePtr();
8995 // Perform one or more big stores into memory.
8996 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8997 for (unsigned I = 0; I < E; I++) {
8998 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8999 StoreType, ShuffWide,
9000 DAG.getIntPtrConstant(I));
9001 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9002 St->getPointerInfo(), St->isVolatile(),
9003 St->isNonTemporal(), St->getAlignment());
9004 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9006 Chains.push_back(Ch);
9008 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
9012 if (!ISD::isNormalStore(St))
9015 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9016 // ARM stores of arguments in the same cache line.
9017 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9018 StVal.getNode()->hasOneUse()) {
9019 SelectionDAG &DAG = DCI.DAG;
9021 SDValue BasePtr = St->getBasePtr();
9022 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9023 StVal.getNode()->getOperand(0), BasePtr,
9024 St->getPointerInfo(), St->isVolatile(),
9025 St->isNonTemporal(), St->getAlignment());
9027 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9028 DAG.getConstant(4, MVT::i32));
9029 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
9030 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9031 St->isNonTemporal(),
9032 std::min(4U, St->getAlignment() / 2));
9035 if (StVal.getValueType() != MVT::i64 ||
9036 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9039 // Bitcast an i64 store extracted from a vector to f64.
9040 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9041 SelectionDAG &DAG = DCI.DAG;
9043 SDValue IntVec = StVal.getOperand(0);
9044 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9045 IntVec.getValueType().getVectorNumElements());
9046 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9047 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9048 Vec, StVal.getOperand(1));
9050 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9051 // Make the DAGCombiner fold the bitcasts.
9052 DCI.AddToWorklist(Vec.getNode());
9053 DCI.AddToWorklist(ExtElt.getNode());
9054 DCI.AddToWorklist(V.getNode());
9055 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9056 St->getPointerInfo(), St->isVolatile(),
9057 St->isNonTemporal(), St->getAlignment(),
9061 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9062 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9063 /// i64 vector to have f64 elements, since the value can then be loaded
9064 /// directly into a VFP register.
9065 static bool hasNormalLoadOperand(SDNode *N) {
9066 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9067 for (unsigned i = 0; i < NumElts; ++i) {
9068 SDNode *Elt = N->getOperand(i).getNode();
9069 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9075 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9076 /// ISD::BUILD_VECTOR.
9077 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9078 TargetLowering::DAGCombinerInfo &DCI){
9079 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9080 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9081 // into a pair of GPRs, which is fine when the value is used as a scalar,
9082 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9083 SelectionDAG &DAG = DCI.DAG;
9084 if (N->getNumOperands() == 2) {
9085 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9090 // Load i64 elements as f64 values so that type legalization does not split
9091 // them up into i32 values.
9092 EVT VT = N->getValueType(0);
9093 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9096 SmallVector<SDValue, 8> Ops;
9097 unsigned NumElts = VT.getVectorNumElements();
9098 for (unsigned i = 0; i < NumElts; ++i) {
9099 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9101 // Make the DAGCombiner fold the bitcast.
9102 DCI.AddToWorklist(V.getNode());
9104 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9105 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
9106 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9109 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9111 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9112 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9113 // At that time, we may have inserted bitcasts from integer to float.
9114 // If these bitcasts have survived DAGCombine, change the lowering of this
9115 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9116 // force to use floating point types.
9118 // Make sure we can change the type of the vector.
9119 // This is possible iff:
9120 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9121 // 1.1. Vector is used only once.
9122 // 1.2. Use is a bit convert to an integer type.
9123 // 2. The size of its operands are 32-bits (64-bits are not legal).
9124 EVT VT = N->getValueType(0);
9125 EVT EltVT = VT.getVectorElementType();
9127 // Check 1.1. and 2.
9128 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9131 // By construction, the input type must be float.
9132 assert(EltVT == MVT::f32 && "Unexpected type!");
9135 SDNode *Use = *N->use_begin();
9136 if (Use->getOpcode() != ISD::BITCAST ||
9137 Use->getValueType(0).isFloatingPoint())
9140 // Check profitability.
9141 // Model is, if more than half of the relevant operands are bitcast from
9142 // i32, turn the build_vector into a sequence of insert_vector_elt.
9143 // Relevant operands are everything that is not statically
9144 // (i.e., at compile time) bitcasted.
9145 unsigned NumOfBitCastedElts = 0;
9146 unsigned NumElts = VT.getVectorNumElements();
9147 unsigned NumOfRelevantElts = NumElts;
9148 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9149 SDValue Elt = N->getOperand(Idx);
9150 if (Elt->getOpcode() == ISD::BITCAST) {
9151 // Assume only bit cast to i32 will go away.
9152 if (Elt->getOperand(0).getValueType() == MVT::i32)
9153 ++NumOfBitCastedElts;
9154 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9155 // Constants are statically casted, thus do not count them as
9156 // relevant operands.
9157 --NumOfRelevantElts;
9160 // Check if more than half of the elements require a non-free bitcast.
9161 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9164 SelectionDAG &DAG = DCI.DAG;
9165 // Create the new vector type.
9166 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9167 // Check if the type is legal.
9168 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9169 if (!TLI.isTypeLegal(VecVT))
9173 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9174 // => BITCAST INSERT_VECTOR_ELT
9175 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9177 SDValue Vec = DAG.getUNDEF(VecVT);
9179 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9180 SDValue V = N->getOperand(Idx);
9181 if (V.getOpcode() == ISD::UNDEF)
9183 if (V.getOpcode() == ISD::BITCAST &&
9184 V->getOperand(0).getValueType() == MVT::i32)
9185 // Fold obvious case.
9186 V = V.getOperand(0);
9188 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9189 // Make the DAGCombiner fold the bitcasts.
9190 DCI.AddToWorklist(V.getNode());
9192 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
9193 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9195 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9196 // Make the DAGCombiner fold the bitcasts.
9197 DCI.AddToWorklist(Vec.getNode());
9201 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9202 /// ISD::INSERT_VECTOR_ELT.
9203 static SDValue PerformInsertEltCombine(SDNode *N,
9204 TargetLowering::DAGCombinerInfo &DCI) {
9205 // Bitcast an i64 load inserted into a vector to f64.
9206 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9207 EVT VT = N->getValueType(0);
9208 SDNode *Elt = N->getOperand(1).getNode();
9209 if (VT.getVectorElementType() != MVT::i64 ||
9210 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9213 SelectionDAG &DAG = DCI.DAG;
9215 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9216 VT.getVectorNumElements());
9217 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9218 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9219 // Make the DAGCombiner fold the bitcasts.
9220 DCI.AddToWorklist(Vec.getNode());
9221 DCI.AddToWorklist(V.getNode());
9222 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9223 Vec, V, N->getOperand(2));
9224 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9227 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9228 /// ISD::VECTOR_SHUFFLE.
9229 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9230 // The LLVM shufflevector instruction does not require the shuffle mask
9231 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9232 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9233 // operands do not match the mask length, they are extended by concatenating
9234 // them with undef vectors. That is probably the right thing for other
9235 // targets, but for NEON it is better to concatenate two double-register
9236 // size vector operands into a single quad-register size vector. Do that
9237 // transformation here:
9238 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9239 // shuffle(concat(v1, v2), undef)
9240 SDValue Op0 = N->getOperand(0);
9241 SDValue Op1 = N->getOperand(1);
9242 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9243 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9244 Op0.getNumOperands() != 2 ||
9245 Op1.getNumOperands() != 2)
9247 SDValue Concat0Op1 = Op0.getOperand(1);
9248 SDValue Concat1Op1 = Op1.getOperand(1);
9249 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9250 Concat1Op1.getOpcode() != ISD::UNDEF)
9252 // Skip the transformation if any of the types are illegal.
9253 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9254 EVT VT = N->getValueType(0);
9255 if (!TLI.isTypeLegal(VT) ||
9256 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9257 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9260 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9261 Op0.getOperand(0), Op1.getOperand(0));
9262 // Translate the shuffle mask.
9263 SmallVector<int, 16> NewMask;
9264 unsigned NumElts = VT.getVectorNumElements();
9265 unsigned HalfElts = NumElts/2;
9266 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9267 for (unsigned n = 0; n < NumElts; ++n) {
9268 int MaskElt = SVN->getMaskElt(n);
9270 if (MaskElt < (int)HalfElts)
9272 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9273 NewElt = HalfElts + MaskElt - NumElts;
9274 NewMask.push_back(NewElt);
9276 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9277 DAG.getUNDEF(VT), NewMask.data());
9280 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9281 /// NEON load/store intrinsics to merge base address updates.
9282 static SDValue CombineBaseUpdate(SDNode *N,
9283 TargetLowering::DAGCombinerInfo &DCI) {
9284 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9287 SelectionDAG &DAG = DCI.DAG;
9288 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9289 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9290 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9291 SDValue Addr = N->getOperand(AddrOpIdx);
9293 // Search for a use of the address operand that is an increment.
9294 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9295 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9297 if (User->getOpcode() != ISD::ADD ||
9298 UI.getUse().getResNo() != Addr.getResNo())
9301 // Check that the add is independent of the load/store. Otherwise, folding
9302 // it would create a cycle.
9303 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9306 // Find the new opcode for the updating load/store.
9308 bool isLaneOp = false;
9309 unsigned NewOpc = 0;
9310 unsigned NumVecs = 0;
9312 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9314 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9315 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9317 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9319 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9321 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9323 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9324 NumVecs = 2; isLaneOp = true; break;
9325 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9326 NumVecs = 3; isLaneOp = true; break;
9327 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9328 NumVecs = 4; isLaneOp = true; break;
9329 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9330 NumVecs = 1; isLoad = false; break;
9331 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9332 NumVecs = 2; isLoad = false; break;
9333 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9334 NumVecs = 3; isLoad = false; break;
9335 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9336 NumVecs = 4; isLoad = false; break;
9337 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9338 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9339 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9340 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9341 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9342 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9346 switch (N->getOpcode()) {
9347 default: llvm_unreachable("unexpected opcode for Neon base update");
9348 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9349 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9350 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9354 // Find the size of memory referenced by the load/store.
9357 VecTy = N->getValueType(0);
9359 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9360 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9362 NumBytes /= VecTy.getVectorNumElements();
9364 // If the increment is a constant, it must match the memory ref size.
9365 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9366 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9367 uint64_t IncVal = CInc->getZExtValue();
9368 if (IncVal != NumBytes)
9370 } else if (NumBytes >= 3 * 16) {
9371 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9372 // separate instructions that make it harder to use a non-constant update.
9376 // Create the new updating load/store node.
9378 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9380 for (n = 0; n < NumResultVecs; ++n)
9382 Tys[n++] = MVT::i32;
9383 Tys[n] = MVT::Other;
9384 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
9385 SmallVector<SDValue, 8> Ops;
9386 Ops.push_back(N->getOperand(0)); // incoming chain
9387 Ops.push_back(N->getOperand(AddrOpIdx));
9389 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9390 Ops.push_back(N->getOperand(i));
9392 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9393 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9394 Ops.data(), Ops.size(),
9395 MemInt->getMemoryVT(),
9396 MemInt->getMemOperand());
9399 std::vector<SDValue> NewResults;
9400 for (unsigned i = 0; i < NumResultVecs; ++i) {
9401 NewResults.push_back(SDValue(UpdN.getNode(), i));
9403 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9404 DCI.CombineTo(N, NewResults);
9405 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9412 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9413 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9414 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9416 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9417 SelectionDAG &DAG = DCI.DAG;
9418 EVT VT = N->getValueType(0);
9419 // vldN-dup instructions only support 64-bit vectors for N > 1.
9420 if (!VT.is64BitVector())
9423 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9424 SDNode *VLD = N->getOperand(0).getNode();
9425 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9427 unsigned NumVecs = 0;
9428 unsigned NewOpc = 0;
9429 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9430 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9432 NewOpc = ARMISD::VLD2DUP;
9433 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9435 NewOpc = ARMISD::VLD3DUP;
9436 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9438 NewOpc = ARMISD::VLD4DUP;
9443 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9444 // numbers match the load.
9445 unsigned VLDLaneNo =
9446 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9447 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9449 // Ignore uses of the chain result.
9450 if (UI.getUse().getResNo() == NumVecs)
9453 if (User->getOpcode() != ARMISD::VDUPLANE ||
9454 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9458 // Create the vldN-dup node.
9461 for (n = 0; n < NumVecs; ++n)
9463 Tys[n] = MVT::Other;
9464 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
9465 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9466 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9467 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9468 Ops, 2, VLDMemInt->getMemoryVT(),
9469 VLDMemInt->getMemOperand());
9472 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9474 unsigned ResNo = UI.getUse().getResNo();
9475 // Ignore uses of the chain result.
9476 if (ResNo == NumVecs)
9479 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9482 // Now the vldN-lane intrinsic is dead except for its chain result.
9483 // Update uses of the chain.
9484 std::vector<SDValue> VLDDupResults;
9485 for (unsigned n = 0; n < NumVecs; ++n)
9486 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9487 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9488 DCI.CombineTo(VLD, VLDDupResults);
9493 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9494 /// ARMISD::VDUPLANE.
9495 static SDValue PerformVDUPLANECombine(SDNode *N,
9496 TargetLowering::DAGCombinerInfo &DCI) {
9497 SDValue Op = N->getOperand(0);
9499 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9500 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9501 if (CombineVLDDUP(N, DCI))
9502 return SDValue(N, 0);
9504 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9505 // redundant. Ignore bit_converts for now; element sizes are checked below.
9506 while (Op.getOpcode() == ISD::BITCAST)
9507 Op = Op.getOperand(0);
9508 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9511 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9512 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9513 // The canonical VMOV for a zero vector uses a 32-bit element size.
9514 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9516 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9518 EVT VT = N->getValueType(0);
9519 if (EltSize > VT.getVectorElementType().getSizeInBits())
9522 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9525 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9526 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9527 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9531 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9533 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9538 APFloat APF = C->getValueAPF();
9539 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9540 != APFloat::opOK || !isExact)
9543 c0 = (I == 0) ? cN : c0;
9544 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9551 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9552 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9553 /// when the VMUL has a constant operand that is a power of 2.
9555 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9556 /// vmul.f32 d16, d17, d16
9557 /// vcvt.s32.f32 d16, d16
9559 /// vcvt.s32.f32 d16, d16, #3
9560 static SDValue PerformVCVTCombine(SDNode *N,
9561 TargetLowering::DAGCombinerInfo &DCI,
9562 const ARMSubtarget *Subtarget) {
9563 SelectionDAG &DAG = DCI.DAG;
9564 SDValue Op = N->getOperand(0);
9566 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9567 Op.getOpcode() != ISD::FMUL)
9571 SDValue N0 = Op->getOperand(0);
9572 SDValue ConstVec = Op->getOperand(1);
9573 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9575 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9576 !isConstVecPow2(ConstVec, isSigned, C))
9579 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9580 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9581 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9582 // These instructions only exist converting from f32 to i32. We can handle
9583 // smaller integers by generating an extra truncate, but larger ones would
9588 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9589 Intrinsic::arm_neon_vcvtfp2fxu;
9590 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9591 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9592 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9593 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9594 DAG.getConstant(Log2_64(C), MVT::i32));
9596 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9597 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9602 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9603 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9604 /// when the VDIV has a constant operand that is a power of 2.
9606 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9607 /// vcvt.f32.s32 d16, d16
9608 /// vdiv.f32 d16, d17, d16
9610 /// vcvt.f32.s32 d16, d16, #3
9611 static SDValue PerformVDIVCombine(SDNode *N,
9612 TargetLowering::DAGCombinerInfo &DCI,
9613 const ARMSubtarget *Subtarget) {
9614 SelectionDAG &DAG = DCI.DAG;
9615 SDValue Op = N->getOperand(0);
9616 unsigned OpOpcode = Op.getNode()->getOpcode();
9618 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9619 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9623 SDValue ConstVec = N->getOperand(1);
9624 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9626 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9627 !isConstVecPow2(ConstVec, isSigned, C))
9630 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9631 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9632 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9633 // These instructions only exist converting from i32 to f32. We can handle
9634 // smaller integers by generating an extra extend, but larger ones would
9639 SDValue ConvInput = Op.getOperand(0);
9640 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9641 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9642 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9643 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9646 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9647 Intrinsic::arm_neon_vcvtfxu2fp;
9648 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9650 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9651 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9654 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9655 /// operand of a vector shift operation, where all the elements of the
9656 /// build_vector must have the same constant integer value.
9657 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9658 // Ignore bit_converts.
9659 while (Op.getOpcode() == ISD::BITCAST)
9660 Op = Op.getOperand(0);
9661 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9662 APInt SplatBits, SplatUndef;
9663 unsigned SplatBitSize;
9665 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9666 HasAnyUndefs, ElementBits) ||
9667 SplatBitSize > ElementBits)
9669 Cnt = SplatBits.getSExtValue();
9673 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9674 /// operand of a vector shift left operation. That value must be in the range:
9675 /// 0 <= Value < ElementBits for a left shift; or
9676 /// 0 <= Value <= ElementBits for a long left shift.
9677 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9678 assert(VT.isVector() && "vector shift count is not a vector type");
9679 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9680 if (! getVShiftImm(Op, ElementBits, Cnt))
9682 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9685 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9686 /// operand of a vector shift right operation. For a shift opcode, the value
9687 /// is positive, but for an intrinsic the value count must be negative. The
9688 /// absolute value must be in the range:
9689 /// 1 <= |Value| <= ElementBits for a right shift; or
9690 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9691 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9693 assert(VT.isVector() && "vector shift count is not a vector type");
9694 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9695 if (! getVShiftImm(Op, ElementBits, Cnt))
9699 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9702 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9703 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9704 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9707 // Don't do anything for most intrinsics.
9710 // Vector shifts: check for immediate versions and lower them.
9711 // Note: This is done during DAG combining instead of DAG legalizing because
9712 // the build_vectors for 64-bit vector element shift counts are generally
9713 // not legal, and it is hard to see their values after they get legalized to
9714 // loads from a constant pool.
9715 case Intrinsic::arm_neon_vshifts:
9716 case Intrinsic::arm_neon_vshiftu:
9717 case Intrinsic::arm_neon_vshiftls:
9718 case Intrinsic::arm_neon_vshiftlu:
9719 case Intrinsic::arm_neon_vrshifts:
9720 case Intrinsic::arm_neon_vrshiftu:
9721 case Intrinsic::arm_neon_vrshiftn:
9722 case Intrinsic::arm_neon_vqshifts:
9723 case Intrinsic::arm_neon_vqshiftu:
9724 case Intrinsic::arm_neon_vqshiftsu:
9725 case Intrinsic::arm_neon_vqshiftns:
9726 case Intrinsic::arm_neon_vqshiftnu:
9727 case Intrinsic::arm_neon_vqshiftnsu:
9728 case Intrinsic::arm_neon_vqrshiftns:
9729 case Intrinsic::arm_neon_vqrshiftnu:
9730 case Intrinsic::arm_neon_vqrshiftnsu: {
9731 EVT VT = N->getOperand(1).getValueType();
9733 unsigned VShiftOpc = 0;
9736 case Intrinsic::arm_neon_vshifts:
9737 case Intrinsic::arm_neon_vshiftu:
9738 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9739 VShiftOpc = ARMISD::VSHL;
9742 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9743 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9744 ARMISD::VSHRs : ARMISD::VSHRu);
9749 case Intrinsic::arm_neon_vshiftls:
9750 case Intrinsic::arm_neon_vshiftlu:
9751 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9753 llvm_unreachable("invalid shift count for vshll intrinsic");
9755 case Intrinsic::arm_neon_vrshifts:
9756 case Intrinsic::arm_neon_vrshiftu:
9757 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9761 case Intrinsic::arm_neon_vqshifts:
9762 case Intrinsic::arm_neon_vqshiftu:
9763 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9767 case Intrinsic::arm_neon_vqshiftsu:
9768 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9770 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9772 case Intrinsic::arm_neon_vrshiftn:
9773 case Intrinsic::arm_neon_vqshiftns:
9774 case Intrinsic::arm_neon_vqshiftnu:
9775 case Intrinsic::arm_neon_vqshiftnsu:
9776 case Intrinsic::arm_neon_vqrshiftns:
9777 case Intrinsic::arm_neon_vqrshiftnu:
9778 case Intrinsic::arm_neon_vqrshiftnsu:
9779 // Narrowing shifts require an immediate right shift.
9780 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9782 llvm_unreachable("invalid shift count for narrowing vector shift "
9786 llvm_unreachable("unhandled vector shift");
9790 case Intrinsic::arm_neon_vshifts:
9791 case Intrinsic::arm_neon_vshiftu:
9792 // Opcode already set above.
9794 case Intrinsic::arm_neon_vshiftls:
9795 case Intrinsic::arm_neon_vshiftlu:
9796 if (Cnt == VT.getVectorElementType().getSizeInBits())
9797 VShiftOpc = ARMISD::VSHLLi;
9799 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9800 ARMISD::VSHLLs : ARMISD::VSHLLu);
9802 case Intrinsic::arm_neon_vrshifts:
9803 VShiftOpc = ARMISD::VRSHRs; break;
9804 case Intrinsic::arm_neon_vrshiftu:
9805 VShiftOpc = ARMISD::VRSHRu; break;
9806 case Intrinsic::arm_neon_vrshiftn:
9807 VShiftOpc = ARMISD::VRSHRN; break;
9808 case Intrinsic::arm_neon_vqshifts:
9809 VShiftOpc = ARMISD::VQSHLs; break;
9810 case Intrinsic::arm_neon_vqshiftu:
9811 VShiftOpc = ARMISD::VQSHLu; break;
9812 case Intrinsic::arm_neon_vqshiftsu:
9813 VShiftOpc = ARMISD::VQSHLsu; break;
9814 case Intrinsic::arm_neon_vqshiftns:
9815 VShiftOpc = ARMISD::VQSHRNs; break;
9816 case Intrinsic::arm_neon_vqshiftnu:
9817 VShiftOpc = ARMISD::VQSHRNu; break;
9818 case Intrinsic::arm_neon_vqshiftnsu:
9819 VShiftOpc = ARMISD::VQSHRNsu; break;
9820 case Intrinsic::arm_neon_vqrshiftns:
9821 VShiftOpc = ARMISD::VQRSHRNs; break;
9822 case Intrinsic::arm_neon_vqrshiftnu:
9823 VShiftOpc = ARMISD::VQRSHRNu; break;
9824 case Intrinsic::arm_neon_vqrshiftnsu:
9825 VShiftOpc = ARMISD::VQRSHRNsu; break;
9828 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9829 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9832 case Intrinsic::arm_neon_vshiftins: {
9833 EVT VT = N->getOperand(1).getValueType();
9835 unsigned VShiftOpc = 0;
9837 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9838 VShiftOpc = ARMISD::VSLI;
9839 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9840 VShiftOpc = ARMISD::VSRI;
9842 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9845 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9846 N->getOperand(1), N->getOperand(2),
9847 DAG.getConstant(Cnt, MVT::i32));
9850 case Intrinsic::arm_neon_vqrshifts:
9851 case Intrinsic::arm_neon_vqrshiftu:
9852 // No immediate versions of these to check for.
9859 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9860 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9861 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9862 /// vector element shift counts are generally not legal, and it is hard to see
9863 /// their values after they get legalized to loads from a constant pool.
9864 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9865 const ARMSubtarget *ST) {
9866 EVT VT = N->getValueType(0);
9867 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9868 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9869 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9870 SDValue N1 = N->getOperand(1);
9871 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9872 SDValue N0 = N->getOperand(0);
9873 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9874 DAG.MaskedValueIsZero(N0.getOperand(0),
9875 APInt::getHighBitsSet(32, 16)))
9876 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9880 // Nothing to be done for scalar shifts.
9881 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9882 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9885 assert(ST->hasNEON() && "unexpected vector shift");
9888 switch (N->getOpcode()) {
9889 default: llvm_unreachable("unexpected shift opcode");
9892 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9893 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9894 DAG.getConstant(Cnt, MVT::i32));
9899 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9900 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9901 ARMISD::VSHRs : ARMISD::VSHRu);
9902 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9903 DAG.getConstant(Cnt, MVT::i32));
9909 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9910 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9911 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9912 const ARMSubtarget *ST) {
9913 SDValue N0 = N->getOperand(0);
9915 // Check for sign- and zero-extensions of vector extract operations of 8-
9916 // and 16-bit vector elements. NEON supports these directly. They are
9917 // handled during DAG combining because type legalization will promote them
9918 // to 32-bit types and it is messy to recognize the operations after that.
9919 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9920 SDValue Vec = N0.getOperand(0);
9921 SDValue Lane = N0.getOperand(1);
9922 EVT VT = N->getValueType(0);
9923 EVT EltVT = N0.getValueType();
9924 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9926 if (VT == MVT::i32 &&
9927 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9928 TLI.isTypeLegal(Vec.getValueType()) &&
9929 isa<ConstantSDNode>(Lane)) {
9932 switch (N->getOpcode()) {
9933 default: llvm_unreachable("unexpected opcode");
9934 case ISD::SIGN_EXTEND:
9935 Opc = ARMISD::VGETLANEs;
9937 case ISD::ZERO_EXTEND:
9938 case ISD::ANY_EXTEND:
9939 Opc = ARMISD::VGETLANEu;
9942 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9949 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9950 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9951 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9952 const ARMSubtarget *ST) {
9953 // If the target supports NEON, try to use vmax/vmin instructions for f32
9954 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9955 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9956 // a NaN; only do the transformation when it matches that behavior.
9958 // For now only do this when using NEON for FP operations; if using VFP, it
9959 // is not obvious that the benefit outweighs the cost of switching to the
9961 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9962 N->getValueType(0) != MVT::f32)
9965 SDValue CondLHS = N->getOperand(0);
9966 SDValue CondRHS = N->getOperand(1);
9967 SDValue LHS = N->getOperand(2);
9968 SDValue RHS = N->getOperand(3);
9969 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9971 unsigned Opcode = 0;
9973 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9974 IsReversed = false; // x CC y ? x : y
9975 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9976 IsReversed = true ; // x CC y ? y : x
9990 // If LHS is NaN, an ordered comparison will be false and the result will
9991 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9992 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9993 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9994 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9996 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9997 // will return -0, so vmin can only be used for unsafe math or if one of
9998 // the operands is known to be nonzero.
9999 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
10000 !DAG.getTarget().Options.UnsafeFPMath &&
10001 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10003 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
10012 // If LHS is NaN, an ordered comparison will be false and the result will
10013 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
10014 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10015 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
10016 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10018 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
10019 // will return +0, so vmax can only be used for unsafe math or if one of
10020 // the operands is known to be nonzero.
10021 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
10022 !DAG.getTarget().Options.UnsafeFPMath &&
10023 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10025 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
10031 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
10034 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10036 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10037 SDValue Cmp = N->getOperand(4);
10038 if (Cmp.getOpcode() != ARMISD::CMPZ)
10039 // Only looking at EQ and NE cases.
10042 EVT VT = N->getValueType(0);
10044 SDValue LHS = Cmp.getOperand(0);
10045 SDValue RHS = Cmp.getOperand(1);
10046 SDValue FalseVal = N->getOperand(0);
10047 SDValue TrueVal = N->getOperand(1);
10048 SDValue ARMcc = N->getOperand(2);
10049 ARMCC::CondCodes CC =
10050 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10068 /// FIXME: Turn this into a target neutral optimization?
10070 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10071 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10072 N->getOperand(3), Cmp);
10073 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10075 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10076 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10077 N->getOperand(3), NewCmp);
10080 if (Res.getNode()) {
10081 APInt KnownZero, KnownOne;
10082 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
10083 // Capture demanded bits information that would be otherwise lost.
10084 if (KnownZero == 0xfffffffe)
10085 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10086 DAG.getValueType(MVT::i1));
10087 else if (KnownZero == 0xffffff00)
10088 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10089 DAG.getValueType(MVT::i8));
10090 else if (KnownZero == 0xffff0000)
10091 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10092 DAG.getValueType(MVT::i16));
10098 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10099 DAGCombinerInfo &DCI) const {
10100 switch (N->getOpcode()) {
10102 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10103 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10104 case ISD::SUB: return PerformSUBCombine(N, DCI);
10105 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10106 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10107 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10108 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10109 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10110 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
10111 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10112 case ISD::STORE: return PerformSTORECombine(N, DCI);
10113 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
10114 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10115 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10116 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10117 case ISD::FP_TO_SINT:
10118 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10119 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10120 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10123 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10124 case ISD::SIGN_EXTEND:
10125 case ISD::ZERO_EXTEND:
10126 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10127 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10128 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10129 case ARMISD::VLD2DUP:
10130 case ARMISD::VLD3DUP:
10131 case ARMISD::VLD4DUP:
10132 return CombineBaseUpdate(N, DCI);
10133 case ARMISD::BUILD_VECTOR:
10134 return PerformARMBUILD_VECTORCombine(N, DCI);
10135 case ISD::INTRINSIC_VOID:
10136 case ISD::INTRINSIC_W_CHAIN:
10137 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10138 case Intrinsic::arm_neon_vld1:
10139 case Intrinsic::arm_neon_vld2:
10140 case Intrinsic::arm_neon_vld3:
10141 case Intrinsic::arm_neon_vld4:
10142 case Intrinsic::arm_neon_vld2lane:
10143 case Intrinsic::arm_neon_vld3lane:
10144 case Intrinsic::arm_neon_vld4lane:
10145 case Intrinsic::arm_neon_vst1:
10146 case Intrinsic::arm_neon_vst2:
10147 case Intrinsic::arm_neon_vst3:
10148 case Intrinsic::arm_neon_vst4:
10149 case Intrinsic::arm_neon_vst2lane:
10150 case Intrinsic::arm_neon_vst3lane:
10151 case Intrinsic::arm_neon_vst4lane:
10152 return CombineBaseUpdate(N, DCI);
10160 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10162 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10165 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
10166 bool *Fast) const {
10167 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10168 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10170 switch (VT.getSimpleVT().SimpleTy) {
10176 // Unaligned access can use (for example) LRDB, LRDH, LDR
10177 if (AllowsUnaligned) {
10179 *Fast = Subtarget->hasV7Ops();
10186 // For any little-endian targets with neon, we can support unaligned ld/st
10187 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10188 // A big-endian target may also explicitly support unaligned accesses
10189 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10199 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10200 unsigned AlignCheck) {
10201 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10202 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10205 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10206 unsigned DstAlign, unsigned SrcAlign,
10207 bool IsMemset, bool ZeroMemset,
10209 MachineFunction &MF) const {
10210 const Function *F = MF.getFunction();
10212 // See if we can use NEON instructions for this...
10213 if ((!IsMemset || ZeroMemset) &&
10214 Subtarget->hasNEON() &&
10215 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10216 Attribute::NoImplicitFloat)) {
10219 (memOpAlign(SrcAlign, DstAlign, 16) ||
10220 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
10222 } else if (Size >= 8 &&
10223 (memOpAlign(SrcAlign, DstAlign, 8) ||
10224 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
10229 // Lowering to i32/i16 if the size permits.
10232 else if (Size >= 2)
10235 // Let the target-independent logic figure it out.
10239 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10240 if (Val.getOpcode() != ISD::LOAD)
10243 EVT VT1 = Val.getValueType();
10244 if (!VT1.isSimple() || !VT1.isInteger() ||
10245 !VT2.isSimple() || !VT2.isInteger())
10248 switch (VT1.getSimpleVT().SimpleTy) {
10253 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10260 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10261 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10264 if (!isTypeLegal(EVT::getEVT(Ty1)))
10267 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10269 // Assuming the caller doesn't have a zeroext or signext return parameter,
10270 // truncation all the way down to i1 is valid.
10275 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10279 unsigned Scale = 1;
10280 switch (VT.getSimpleVT().SimpleTy) {
10281 default: return false;
10296 if ((V & (Scale - 1)) != 0)
10299 return V == (V & ((1LL << 5) - 1));
10302 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10303 const ARMSubtarget *Subtarget) {
10304 bool isNeg = false;
10310 switch (VT.getSimpleVT().SimpleTy) {
10311 default: return false;
10316 // + imm12 or - imm8
10318 return V == (V & ((1LL << 8) - 1));
10319 return V == (V & ((1LL << 12) - 1));
10322 // Same as ARM mode. FIXME: NEON?
10323 if (!Subtarget->hasVFP2())
10328 return V == (V & ((1LL << 8) - 1));
10332 /// isLegalAddressImmediate - Return true if the integer value can be used
10333 /// as the offset of the target addressing mode for load / store of the
10335 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10336 const ARMSubtarget *Subtarget) {
10340 if (!VT.isSimple())
10343 if (Subtarget->isThumb1Only())
10344 return isLegalT1AddressImmediate(V, VT);
10345 else if (Subtarget->isThumb2())
10346 return isLegalT2AddressImmediate(V, VT, Subtarget);
10351 switch (VT.getSimpleVT().SimpleTy) {
10352 default: return false;
10357 return V == (V & ((1LL << 12) - 1));
10360 return V == (V & ((1LL << 8) - 1));
10363 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10368 return V == (V & ((1LL << 8) - 1));
10372 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10374 int Scale = AM.Scale;
10378 switch (VT.getSimpleVT().SimpleTy) {
10379 default: return false;
10387 Scale = Scale & ~1;
10388 return Scale == 2 || Scale == 4 || Scale == 8;
10391 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10395 // Note, we allow "void" uses (basically, uses that aren't loads or
10396 // stores), because arm allows folding a scale into many arithmetic
10397 // operations. This should be made more precise and revisited later.
10399 // Allow r << imm, but the imm has to be a multiple of two.
10400 if (Scale & 1) return false;
10401 return isPowerOf2_32(Scale);
10405 /// isLegalAddressingMode - Return true if the addressing mode represented
10406 /// by AM is legal for this target, for a load/store of the specified type.
10407 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10409 EVT VT = getValueType(Ty, true);
10410 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10413 // Can never fold addr of global into load/store.
10417 switch (AM.Scale) {
10418 case 0: // no scale reg, must be "r+i" or "r", or "i".
10421 if (Subtarget->isThumb1Only())
10425 // ARM doesn't support any R+R*scale+imm addr modes.
10429 if (!VT.isSimple())
10432 if (Subtarget->isThumb2())
10433 return isLegalT2ScaledAddressingMode(AM, VT);
10435 int Scale = AM.Scale;
10436 switch (VT.getSimpleVT().SimpleTy) {
10437 default: return false;
10441 if (Scale < 0) Scale = -Scale;
10445 return isPowerOf2_32(Scale & ~1);
10449 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10454 // Note, we allow "void" uses (basically, uses that aren't loads or
10455 // stores), because arm allows folding a scale into many arithmetic
10456 // operations. This should be made more precise and revisited later.
10458 // Allow r << imm, but the imm has to be a multiple of two.
10459 if (Scale & 1) return false;
10460 return isPowerOf2_32(Scale);
10466 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10467 /// icmp immediate, that is the target has icmp instructions which can compare
10468 /// a register against the immediate without having to materialize the
10469 /// immediate into a register.
10470 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10471 // Thumb2 and ARM modes can use cmn for negative immediates.
10472 if (!Subtarget->isThumb())
10473 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10474 if (Subtarget->isThumb2())
10475 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10476 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10477 return Imm >= 0 && Imm <= 255;
10480 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10481 /// *or sub* immediate, that is the target has add or sub instructions which can
10482 /// add a register with the immediate without having to materialize the
10483 /// immediate into a register.
10484 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10485 // Same encoding for add/sub, just flip the sign.
10486 int64_t AbsImm = llvm::abs64(Imm);
10487 if (!Subtarget->isThumb())
10488 return ARM_AM::getSOImmVal(AbsImm) != -1;
10489 if (Subtarget->isThumb2())
10490 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10491 // Thumb1 only has 8-bit unsigned immediate.
10492 return AbsImm >= 0 && AbsImm <= 255;
10495 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10496 bool isSEXTLoad, SDValue &Base,
10497 SDValue &Offset, bool &isInc,
10498 SelectionDAG &DAG) {
10499 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10502 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10503 // AddressingMode 3
10504 Base = Ptr->getOperand(0);
10505 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10506 int RHSC = (int)RHS->getZExtValue();
10507 if (RHSC < 0 && RHSC > -256) {
10508 assert(Ptr->getOpcode() == ISD::ADD);
10510 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10514 isInc = (Ptr->getOpcode() == ISD::ADD);
10515 Offset = Ptr->getOperand(1);
10517 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10518 // AddressingMode 2
10519 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10520 int RHSC = (int)RHS->getZExtValue();
10521 if (RHSC < 0 && RHSC > -0x1000) {
10522 assert(Ptr->getOpcode() == ISD::ADD);
10524 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10525 Base = Ptr->getOperand(0);
10530 if (Ptr->getOpcode() == ISD::ADD) {
10532 ARM_AM::ShiftOpc ShOpcVal=
10533 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10534 if (ShOpcVal != ARM_AM::no_shift) {
10535 Base = Ptr->getOperand(1);
10536 Offset = Ptr->getOperand(0);
10538 Base = Ptr->getOperand(0);
10539 Offset = Ptr->getOperand(1);
10544 isInc = (Ptr->getOpcode() == ISD::ADD);
10545 Base = Ptr->getOperand(0);
10546 Offset = Ptr->getOperand(1);
10550 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10554 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10555 bool isSEXTLoad, SDValue &Base,
10556 SDValue &Offset, bool &isInc,
10557 SelectionDAG &DAG) {
10558 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10561 Base = Ptr->getOperand(0);
10562 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10563 int RHSC = (int)RHS->getZExtValue();
10564 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10565 assert(Ptr->getOpcode() == ISD::ADD);
10567 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10569 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10570 isInc = Ptr->getOpcode() == ISD::ADD;
10571 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10579 /// getPreIndexedAddressParts - returns true by value, base pointer and
10580 /// offset pointer and addressing mode by reference if the node's address
10581 /// can be legally represented as pre-indexed load / store address.
10583 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10585 ISD::MemIndexedMode &AM,
10586 SelectionDAG &DAG) const {
10587 if (Subtarget->isThumb1Only())
10592 bool isSEXTLoad = false;
10593 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10594 Ptr = LD->getBasePtr();
10595 VT = LD->getMemoryVT();
10596 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10597 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10598 Ptr = ST->getBasePtr();
10599 VT = ST->getMemoryVT();
10604 bool isLegal = false;
10605 if (Subtarget->isThumb2())
10606 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10607 Offset, isInc, DAG);
10609 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10610 Offset, isInc, DAG);
10614 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10618 /// getPostIndexedAddressParts - returns true by value, base pointer and
10619 /// offset pointer and addressing mode by reference if this node can be
10620 /// combined with a load / store to form a post-indexed load / store.
10621 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10624 ISD::MemIndexedMode &AM,
10625 SelectionDAG &DAG) const {
10626 if (Subtarget->isThumb1Only())
10631 bool isSEXTLoad = false;
10632 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10633 VT = LD->getMemoryVT();
10634 Ptr = LD->getBasePtr();
10635 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10636 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10637 VT = ST->getMemoryVT();
10638 Ptr = ST->getBasePtr();
10643 bool isLegal = false;
10644 if (Subtarget->isThumb2())
10645 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10648 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10654 // Swap base ptr and offset to catch more post-index load / store when
10655 // it's legal. In Thumb2 mode, offset must be an immediate.
10656 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10657 !Subtarget->isThumb2())
10658 std::swap(Base, Offset);
10660 // Post-indexed load / store update the base pointer.
10665 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10669 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10672 const SelectionDAG &DAG,
10673 unsigned Depth) const {
10674 unsigned BitWidth = KnownOne.getBitWidth();
10675 KnownZero = KnownOne = APInt(BitWidth, 0);
10676 switch (Op.getOpcode()) {
10682 // These nodes' second result is a boolean
10683 if (Op.getResNo() == 0)
10685 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10687 case ARMISD::CMOV: {
10688 // Bits are known zero/one if known on the LHS and RHS.
10689 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10690 if (KnownZero == 0 && KnownOne == 0) return;
10692 APInt KnownZeroRHS, KnownOneRHS;
10693 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10694 KnownZero &= KnownZeroRHS;
10695 KnownOne &= KnownOneRHS;
10701 //===----------------------------------------------------------------------===//
10702 // ARM Inline Assembly Support
10703 //===----------------------------------------------------------------------===//
10705 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10706 // Looking for "rev" which is V6+.
10707 if (!Subtarget->hasV6Ops())
10710 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10711 std::string AsmStr = IA->getAsmString();
10712 SmallVector<StringRef, 4> AsmPieces;
10713 SplitString(AsmStr, AsmPieces, ";\n");
10715 switch (AsmPieces.size()) {
10716 default: return false;
10718 AsmStr = AsmPieces[0];
10720 SplitString(AsmStr, AsmPieces, " \t,");
10723 if (AsmPieces.size() == 3 &&
10724 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10725 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10726 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10727 if (Ty && Ty->getBitWidth() == 32)
10728 return IntrinsicLowering::LowerToByteSwap(CI);
10736 /// getConstraintType - Given a constraint letter, return the type of
10737 /// constraint it is for this target.
10738 ARMTargetLowering::ConstraintType
10739 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10740 if (Constraint.size() == 1) {
10741 switch (Constraint[0]) {
10743 case 'l': return C_RegisterClass;
10744 case 'w': return C_RegisterClass;
10745 case 'h': return C_RegisterClass;
10746 case 'x': return C_RegisterClass;
10747 case 't': return C_RegisterClass;
10748 case 'j': return C_Other; // Constant for movw.
10749 // An address with a single base register. Due to the way we
10750 // currently handle addresses it is the same as an 'r' memory constraint.
10751 case 'Q': return C_Memory;
10753 } else if (Constraint.size() == 2) {
10754 switch (Constraint[0]) {
10756 // All 'U+' constraints are addresses.
10757 case 'U': return C_Memory;
10760 return TargetLowering::getConstraintType(Constraint);
10763 /// Examine constraint type and operand type and determine a weight value.
10764 /// This object must already have been set up with the operand type
10765 /// and the current alternative constraint selected.
10766 TargetLowering::ConstraintWeight
10767 ARMTargetLowering::getSingleConstraintMatchWeight(
10768 AsmOperandInfo &info, const char *constraint) const {
10769 ConstraintWeight weight = CW_Invalid;
10770 Value *CallOperandVal = info.CallOperandVal;
10771 // If we don't have a value, we can't do a match,
10772 // but allow it at the lowest weight.
10773 if (CallOperandVal == NULL)
10775 Type *type = CallOperandVal->getType();
10776 // Look at the constraint type.
10777 switch (*constraint) {
10779 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10782 if (type->isIntegerTy()) {
10783 if (Subtarget->isThumb())
10784 weight = CW_SpecificReg;
10786 weight = CW_Register;
10790 if (type->isFloatingPointTy())
10791 weight = CW_Register;
10797 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10799 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10801 if (Constraint.size() == 1) {
10802 // GCC ARM Constraint Letters
10803 switch (Constraint[0]) {
10804 case 'l': // Low regs or general regs.
10805 if (Subtarget->isThumb())
10806 return RCPair(0U, &ARM::tGPRRegClass);
10807 return RCPair(0U, &ARM::GPRRegClass);
10808 case 'h': // High regs or no regs.
10809 if (Subtarget->isThumb())
10810 return RCPair(0U, &ARM::hGPRRegClass);
10813 return RCPair(0U, &ARM::GPRRegClass);
10815 if (VT == MVT::Other)
10817 if (VT == MVT::f32)
10818 return RCPair(0U, &ARM::SPRRegClass);
10819 if (VT.getSizeInBits() == 64)
10820 return RCPair(0U, &ARM::DPRRegClass);
10821 if (VT.getSizeInBits() == 128)
10822 return RCPair(0U, &ARM::QPRRegClass);
10825 if (VT == MVT::Other)
10827 if (VT == MVT::f32)
10828 return RCPair(0U, &ARM::SPR_8RegClass);
10829 if (VT.getSizeInBits() == 64)
10830 return RCPair(0U, &ARM::DPR_8RegClass);
10831 if (VT.getSizeInBits() == 128)
10832 return RCPair(0U, &ARM::QPR_8RegClass);
10835 if (VT == MVT::f32)
10836 return RCPair(0U, &ARM::SPRRegClass);
10840 if (StringRef("{cc}").equals_lower(Constraint))
10841 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10843 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10846 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10847 /// vector. If it is invalid, don't add anything to Ops.
10848 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10849 std::string &Constraint,
10850 std::vector<SDValue>&Ops,
10851 SelectionDAG &DAG) const {
10852 SDValue Result(0, 0);
10854 // Currently only support length 1 constraints.
10855 if (Constraint.length() != 1) return;
10857 char ConstraintLetter = Constraint[0];
10858 switch (ConstraintLetter) {
10861 case 'I': case 'J': case 'K': case 'L':
10862 case 'M': case 'N': case 'O':
10863 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10867 int64_t CVal64 = C->getSExtValue();
10868 int CVal = (int) CVal64;
10869 // None of these constraints allow values larger than 32 bits. Check
10870 // that the value fits in an int.
10871 if (CVal != CVal64)
10874 switch (ConstraintLetter) {
10876 // Constant suitable for movw, must be between 0 and
10878 if (Subtarget->hasV6T2Ops())
10879 if (CVal >= 0 && CVal <= 65535)
10883 if (Subtarget->isThumb1Only()) {
10884 // This must be a constant between 0 and 255, for ADD
10886 if (CVal >= 0 && CVal <= 255)
10888 } else if (Subtarget->isThumb2()) {
10889 // A constant that can be used as an immediate value in a
10890 // data-processing instruction.
10891 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10894 // A constant that can be used as an immediate value in a
10895 // data-processing instruction.
10896 if (ARM_AM::getSOImmVal(CVal) != -1)
10902 if (Subtarget->isThumb()) { // FIXME thumb2
10903 // This must be a constant between -255 and -1, for negated ADD
10904 // immediates. This can be used in GCC with an "n" modifier that
10905 // prints the negated value, for use with SUB instructions. It is
10906 // not useful otherwise but is implemented for compatibility.
10907 if (CVal >= -255 && CVal <= -1)
10910 // This must be a constant between -4095 and 4095. It is not clear
10911 // what this constraint is intended for. Implemented for
10912 // compatibility with GCC.
10913 if (CVal >= -4095 && CVal <= 4095)
10919 if (Subtarget->isThumb1Only()) {
10920 // A 32-bit value where only one byte has a nonzero value. Exclude
10921 // zero to match GCC. This constraint is used by GCC internally for
10922 // constants that can be loaded with a move/shift combination.
10923 // It is not useful otherwise but is implemented for compatibility.
10924 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10926 } else if (Subtarget->isThumb2()) {
10927 // A constant whose bitwise inverse can be used as an immediate
10928 // value in a data-processing instruction. This can be used in GCC
10929 // with a "B" modifier that prints the inverted value, for use with
10930 // BIC and MVN instructions. It is not useful otherwise but is
10931 // implemented for compatibility.
10932 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10935 // A constant whose bitwise inverse can be used as an immediate
10936 // value in a data-processing instruction. This can be used in GCC
10937 // with a "B" modifier that prints the inverted value, for use with
10938 // BIC and MVN instructions. It is not useful otherwise but is
10939 // implemented for compatibility.
10940 if (ARM_AM::getSOImmVal(~CVal) != -1)
10946 if (Subtarget->isThumb1Only()) {
10947 // This must be a constant between -7 and 7,
10948 // for 3-operand ADD/SUB immediate instructions.
10949 if (CVal >= -7 && CVal < 7)
10951 } else if (Subtarget->isThumb2()) {
10952 // A constant whose negation can be used as an immediate value in a
10953 // data-processing instruction. This can be used in GCC with an "n"
10954 // modifier that prints the negated value, for use with SUB
10955 // instructions. It is not useful otherwise but is implemented for
10957 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10960 // A constant whose negation can be used as an immediate value in a
10961 // data-processing instruction. This can be used in GCC with an "n"
10962 // modifier that prints the negated value, for use with SUB
10963 // instructions. It is not useful otherwise but is implemented for
10965 if (ARM_AM::getSOImmVal(-CVal) != -1)
10971 if (Subtarget->isThumb()) { // FIXME thumb2
10972 // This must be a multiple of 4 between 0 and 1020, for
10973 // ADD sp + immediate.
10974 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10977 // A power of two or a constant between 0 and 32. This is used in
10978 // GCC for the shift amount on shifted register operands, but it is
10979 // useful in general for any shift amounts.
10980 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10986 if (Subtarget->isThumb()) { // FIXME thumb2
10987 // This must be a constant between 0 and 31, for shift amounts.
10988 if (CVal >= 0 && CVal <= 31)
10994 if (Subtarget->isThumb()) { // FIXME thumb2
10995 // This must be a multiple of 4 between -508 and 508, for
10996 // ADD/SUB sp = sp + immediate.
10997 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11002 Result = DAG.getTargetConstant(CVal, Op.getValueType());
11006 if (Result.getNode()) {
11007 Ops.push_back(Result);
11010 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11013 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11014 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
11015 unsigned Opcode = Op->getOpcode();
11016 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11017 "Invalid opcode for Div/Rem lowering");
11018 bool isSigned = (Opcode == ISD::SDIVREM);
11019 EVT VT = Op->getValueType(0);
11020 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11023 switch (VT.getSimpleVT().SimpleTy) {
11024 default: llvm_unreachable("Unexpected request for libcall!");
11025 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11026 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11027 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11028 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11031 SDValue InChain = DAG.getEntryNode();
11033 TargetLowering::ArgListTy Args;
11034 TargetLowering::ArgListEntry Entry;
11035 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
11036 EVT ArgVT = Op->getOperand(i).getValueType();
11037 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
11038 Entry.Node = Op->getOperand(i);
11040 Entry.isSExt = isSigned;
11041 Entry.isZExt = !isSigned;
11042 Args.push_back(Entry);
11045 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11048 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
11052 CallLoweringInfo CLI(InChain, RetTy, isSigned, !isSigned, false, true,
11053 0, getLibcallCallingConv(LC), /*isTailCall=*/false,
11054 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
11055 Callee, Args, DAG, dl);
11056 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11058 return CallInfo.first;
11062 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11063 // The ARM target isn't yet aware of offsets.
11067 bool ARM::isBitFieldInvertedMask(unsigned v) {
11068 if (v == 0xffffffff)
11071 // there can be 1's on either or both "outsides", all the "inside"
11072 // bits must be 0's
11073 unsigned TO = CountTrailingOnes_32(v);
11074 unsigned LO = CountLeadingOnes_32(v);
11075 v = (v >> TO) << TO;
11076 v = (v << LO) >> LO;
11080 /// isFPImmLegal - Returns true if the target can instruction select the
11081 /// specified FP immediate natively. If false, the legalizer will
11082 /// materialize the FP immediate as a load from a constant pool.
11083 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11084 if (!Subtarget->hasVFP3())
11086 if (VT == MVT::f32)
11087 return ARM_AM::getFP32Imm(Imm) != -1;
11088 if (VT == MVT::f64)
11089 return ARM_AM::getFP64Imm(Imm) != -1;
11093 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11094 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11095 /// specified in the intrinsic calls.
11096 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11098 unsigned Intrinsic) const {
11099 switch (Intrinsic) {
11100 case Intrinsic::arm_neon_vld1:
11101 case Intrinsic::arm_neon_vld2:
11102 case Intrinsic::arm_neon_vld3:
11103 case Intrinsic::arm_neon_vld4:
11104 case Intrinsic::arm_neon_vld2lane:
11105 case Intrinsic::arm_neon_vld3lane:
11106 case Intrinsic::arm_neon_vld4lane: {
11107 Info.opc = ISD::INTRINSIC_W_CHAIN;
11108 // Conservatively set memVT to the entire set of vectors loaded.
11109 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11110 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11111 Info.ptrVal = I.getArgOperand(0);
11113 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11114 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11115 Info.vol = false; // volatile loads with NEON intrinsics not supported
11116 Info.readMem = true;
11117 Info.writeMem = false;
11120 case Intrinsic::arm_neon_vst1:
11121 case Intrinsic::arm_neon_vst2:
11122 case Intrinsic::arm_neon_vst3:
11123 case Intrinsic::arm_neon_vst4:
11124 case Intrinsic::arm_neon_vst2lane:
11125 case Intrinsic::arm_neon_vst3lane:
11126 case Intrinsic::arm_neon_vst4lane: {
11127 Info.opc = ISD::INTRINSIC_VOID;
11128 // Conservatively set memVT to the entire set of vectors stored.
11129 unsigned NumElts = 0;
11130 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11131 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11132 if (!ArgTy->isVectorTy())
11134 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11136 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11137 Info.ptrVal = I.getArgOperand(0);
11139 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11140 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11141 Info.vol = false; // volatile stores with NEON intrinsics not supported
11142 Info.readMem = false;
11143 Info.writeMem = true;
11146 case Intrinsic::arm_ldrex: {
11147 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11148 Info.opc = ISD::INTRINSIC_W_CHAIN;
11149 Info.memVT = MVT::getVT(PtrTy->getElementType());
11150 Info.ptrVal = I.getArgOperand(0);
11152 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11154 Info.readMem = true;
11155 Info.writeMem = false;
11158 case Intrinsic::arm_strex: {
11159 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11160 Info.opc = ISD::INTRINSIC_W_CHAIN;
11161 Info.memVT = MVT::getVT(PtrTy->getElementType());
11162 Info.ptrVal = I.getArgOperand(1);
11164 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11166 Info.readMem = false;
11167 Info.writeMem = true;
11170 case Intrinsic::arm_strexd: {
11171 Info.opc = ISD::INTRINSIC_W_CHAIN;
11172 Info.memVT = MVT::i64;
11173 Info.ptrVal = I.getArgOperand(2);
11177 Info.readMem = false;
11178 Info.writeMem = true;
11181 case Intrinsic::arm_ldrexd: {
11182 Info.opc = ISD::INTRINSIC_W_CHAIN;
11183 Info.memVT = MVT::i64;
11184 Info.ptrVal = I.getArgOperand(0);
11188 Info.readMem = true;
11189 Info.writeMem = false;
11199 /// \brief Returns true if it is beneficial to convert a load of a constant
11200 /// to just the constant itself.
11201 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11203 assert(Ty->isIntegerTy());
11205 unsigned Bits = Ty->getPrimitiveSizeInBits();
11206 if (Bits == 0 || Bits > 32)