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/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/IntrinsicLowering.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::QPRRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
164 return new TargetLoweringObjectFileMachO();
166 return new ARMElfTargetObjectFile();
169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<ARMSubtarget>();
172 RegInfo = TM.getRegisterInfo();
173 Itins = TM.getInstrItineraryData();
175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
177 if (Subtarget->isTargetDarwin()) {
178 // Uses VFP for Thumb libfuncs if available.
179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
180 // Single-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
186 // Double-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
192 // Single-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
211 // Double-precision comparisons.
212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
230 // Floating-point to integer conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
238 // Conversions between floating types.
239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
242 // Integer to floating-point conversions.
243 // i64 conversions are done via library routines even when generating VFP
244 // instructions, so use the same ones.
245 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
246 // e.g., __floatunsidf vs. __floatunssidfvfp.
247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, 0);
256 setLibcallName(RTLIB::SRL_I128, 0);
257 setLibcallName(RTLIB::SRA_I128, 0);
259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point arithmetic helper functions
299 // RTABI chapter 4.1.2, Table 4
300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
309 // Single-precision floating-point comparison helper functions
310 // RTABI chapter 4.1.2, Table 5
311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
336 // Floating-point to integer conversions.
337 // RTABI chapter 4.1.2, Table 6
338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
355 // Conversions between floating types.
356 // RTABI chapter 4.1.2, Table 7
357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
362 // Integer to floating-point conversions.
363 // RTABI chapter 4.1.2, Table 8
364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 // Long long helper functions
382 // RTABI chapter 4.2, Table 9
383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
394 // Integer division functions
395 // RTABI chapter 4.3.1
396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
414 // RTABI chapter 4.3.4
415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
423 // Use divmod compiler-rt calls for iOS 5.0 and later.
424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430 if (Subtarget->isThumb1Only())
431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
433 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
435 !Subtarget->isThumb1Only()) {
436 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
437 if (!Subtarget->isFPOnlySP())
438 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
447 setTruncStoreAction((MVT::SimpleValueType)VT,
448 (MVT::SimpleValueType)InnerVT, Expand);
449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
456 if (Subtarget->hasNEON()) {
457 addDRTypeForNEON(MVT::v2f32);
458 addDRTypeForNEON(MVT::v8i8);
459 addDRTypeForNEON(MVT::v4i16);
460 addDRTypeForNEON(MVT::v2i32);
461 addDRTypeForNEON(MVT::v1i64);
463 addQRTypeForNEON(MVT::v4f32);
464 addQRTypeForNEON(MVT::v2f64);
465 addQRTypeForNEON(MVT::v16i8);
466 addQRTypeForNEON(MVT::v8i16);
467 addQRTypeForNEON(MVT::v4i32);
468 addQRTypeForNEON(MVT::v2i64);
470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
471 // neither Neon nor VFP support any arithmetic operations on it.
472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
473 // supported for v4f32.
474 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
477 // FIXME: Code duplication: FDIV and FREM are expanded always, see
478 // ARMTargetLowering::addTypeForNEON method for details.
479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
480 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
481 // FIXME: Create unittest.
482 // In another words, find a way when "copysign" appears in DAG with vector
484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
485 // FIXME: Code duplication: SETCC has custom operation action, see
486 // ARMTargetLowering::addTypeForNEON method for details.
487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
488 // FIXME: Create unittest for FNEG and for FABS.
489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
490 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
508 setOperationAction(ISD::FABS, MVT::v4f32, Expand);
509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
525 // Neon does not support some operations on v1i64 and v2i64 types.
526 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
527 // Custom handling for some quad-vector types to detect VMULL.
528 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
529 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
530 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
531 // Custom handling for some vector types to avoid expensive expansions
532 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
533 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
534 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
535 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
536 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
537 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
538 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
539 // a destination type that is wider than the source, and nor does
540 // it have a FP_TO_[SU]INT instruction with a narrower destination than
542 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
543 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
544 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
545 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
547 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
549 setTargetDAGCombine(ISD::INTRINSIC_VOID);
550 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
551 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
552 setTargetDAGCombine(ISD::SHL);
553 setTargetDAGCombine(ISD::SRL);
554 setTargetDAGCombine(ISD::SRA);
555 setTargetDAGCombine(ISD::SIGN_EXTEND);
556 setTargetDAGCombine(ISD::ZERO_EXTEND);
557 setTargetDAGCombine(ISD::ANY_EXTEND);
558 setTargetDAGCombine(ISD::SELECT_CC);
559 setTargetDAGCombine(ISD::BUILD_VECTOR);
560 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
561 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
562 setTargetDAGCombine(ISD::STORE);
563 setTargetDAGCombine(ISD::FP_TO_SINT);
564 setTargetDAGCombine(ISD::FP_TO_UINT);
565 setTargetDAGCombine(ISD::FDIV);
567 // It is legal to extload from v4i8 to v4i16 or v4i32.
568 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
569 MVT::v4i16, MVT::v2i16,
571 for (unsigned i = 0; i < 6; ++i) {
572 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
573 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
574 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
578 // ARM and Thumb2 support UMLAL/SMLAL.
579 if (!Subtarget->isThumb1Only())
580 setTargetDAGCombine(ISD::ADDC);
583 computeRegisterProperties();
585 // ARM does not have f32 extending load.
586 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
588 // ARM does not have i1 sign extending load.
589 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
591 // ARM supports all 4 flavors of integer indexed load / store.
592 if (!Subtarget->isThumb1Only()) {
593 for (unsigned im = (unsigned)ISD::PRE_INC;
594 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
595 setIndexedLoadAction(im, MVT::i1, Legal);
596 setIndexedLoadAction(im, MVT::i8, Legal);
597 setIndexedLoadAction(im, MVT::i16, Legal);
598 setIndexedLoadAction(im, MVT::i32, Legal);
599 setIndexedStoreAction(im, MVT::i1, Legal);
600 setIndexedStoreAction(im, MVT::i8, Legal);
601 setIndexedStoreAction(im, MVT::i16, Legal);
602 setIndexedStoreAction(im, MVT::i32, Legal);
606 // i64 operation support.
607 setOperationAction(ISD::MUL, MVT::i64, Expand);
608 setOperationAction(ISD::MULHU, MVT::i32, Expand);
609 if (Subtarget->isThumb1Only()) {
610 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
611 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
613 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
614 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
615 setOperationAction(ISD::MULHS, MVT::i32, Expand);
617 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
618 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
619 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
620 setOperationAction(ISD::SRL, MVT::i64, Custom);
621 setOperationAction(ISD::SRA, MVT::i64, Custom);
623 if (!Subtarget->isThumb1Only()) {
624 // FIXME: We should do this for Thumb1 as well.
625 setOperationAction(ISD::ADDC, MVT::i32, Custom);
626 setOperationAction(ISD::ADDE, MVT::i32, Custom);
627 setOperationAction(ISD::SUBC, MVT::i32, Custom);
628 setOperationAction(ISD::SUBE, MVT::i32, Custom);
631 // ARM does not have ROTL.
632 setOperationAction(ISD::ROTL, MVT::i32, Expand);
633 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
634 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
635 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
636 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
638 // These just redirect to CTTZ and CTLZ on ARM.
639 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
640 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
642 // Only ARMv6 has BSWAP.
643 if (!Subtarget->hasV6Ops())
644 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
646 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
647 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
648 // These are expanded into libcalls if the cpu doesn't have HW divider.
649 setOperationAction(ISD::SDIV, MVT::i32, Expand);
650 setOperationAction(ISD::UDIV, MVT::i32, Expand);
652 setOperationAction(ISD::SREM, MVT::i32, Expand);
653 setOperationAction(ISD::UREM, MVT::i32, Expand);
654 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
655 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
657 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
658 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
659 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
660 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
661 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
663 setOperationAction(ISD::TRAP, MVT::Other, Legal);
665 // Use the default implementation.
666 setOperationAction(ISD::VASTART, MVT::Other, Custom);
667 setOperationAction(ISD::VAARG, MVT::Other, Expand);
668 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
669 setOperationAction(ISD::VAEND, MVT::Other, Expand);
670 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
671 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
673 if (!Subtarget->isTargetDarwin()) {
674 // Non-Darwin platforms may return values in these registers via the
675 // personality function.
676 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
677 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
678 setExceptionPointerRegister(ARM::R0);
679 setExceptionSelectorRegister(ARM::R1);
682 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
683 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
684 // the default expansion.
685 // FIXME: This should be checking for v6k, not just v6.
686 if (Subtarget->hasDataBarrier() ||
687 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
688 // membarrier needs custom lowering; the rest are legal and handled
690 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
691 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
692 // Custom lowering for 64-bit ops
693 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
694 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
695 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
696 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
697 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
698 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
699 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
700 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
701 setInsertFencesForAtomic(true);
703 // Set them all for expansion, which will force libcalls.
704 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
705 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
706 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
711 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
712 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
713 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
714 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
715 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
716 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
717 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
718 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
719 // Unordered/Monotonic case.
720 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
721 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
722 // Since the libcalls include locking, fold in the fences
723 setShouldFoldAtomicFences(true);
726 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
728 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
729 if (!Subtarget->hasV6Ops()) {
730 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
731 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
733 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
735 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
736 !Subtarget->isThumb1Only()) {
737 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
738 // iff target supports vfp2.
739 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
740 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
743 // We want to custom lower some of our intrinsics.
744 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
745 if (Subtarget->isTargetDarwin()) {
746 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
747 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
748 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
751 setOperationAction(ISD::SETCC, MVT::i32, Expand);
752 setOperationAction(ISD::SETCC, MVT::f32, Expand);
753 setOperationAction(ISD::SETCC, MVT::f64, Expand);
754 setOperationAction(ISD::SELECT, MVT::i32, Custom);
755 setOperationAction(ISD::SELECT, MVT::f32, Custom);
756 setOperationAction(ISD::SELECT, MVT::f64, Custom);
757 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
758 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
759 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
761 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
762 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
763 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
764 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
765 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
767 // We don't support sin/cos/fmod/copysign/pow
768 setOperationAction(ISD::FSIN, MVT::f64, Expand);
769 setOperationAction(ISD::FSIN, MVT::f32, Expand);
770 setOperationAction(ISD::FCOS, MVT::f32, Expand);
771 setOperationAction(ISD::FCOS, MVT::f64, Expand);
772 setOperationAction(ISD::FREM, MVT::f64, Expand);
773 setOperationAction(ISD::FREM, MVT::f32, Expand);
774 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
775 !Subtarget->isThumb1Only()) {
776 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
777 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
779 setOperationAction(ISD::FPOW, MVT::f64, Expand);
780 setOperationAction(ISD::FPOW, MVT::f32, Expand);
782 if (!Subtarget->hasVFP4()) {
783 setOperationAction(ISD::FMA, MVT::f64, Expand);
784 setOperationAction(ISD::FMA, MVT::f32, Expand);
787 // Various VFP goodness
788 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
789 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
790 if (Subtarget->hasVFP2()) {
791 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
792 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
793 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
794 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
796 // Special handling for half-precision FP.
797 if (!Subtarget->hasFP16()) {
798 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
799 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
803 // We have target-specific dag combine patterns for the following nodes:
804 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
805 setTargetDAGCombine(ISD::ADD);
806 setTargetDAGCombine(ISD::SUB);
807 setTargetDAGCombine(ISD::MUL);
808 setTargetDAGCombine(ISD::AND);
809 setTargetDAGCombine(ISD::OR);
810 setTargetDAGCombine(ISD::XOR);
812 if (Subtarget->hasV6Ops())
813 setTargetDAGCombine(ISD::SRL);
815 setStackPointerRegisterToSaveRestore(ARM::SP);
817 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
818 !Subtarget->hasVFP2())
819 setSchedulingPreference(Sched::RegPressure);
821 setSchedulingPreference(Sched::Hybrid);
823 //// temporary - rewrite interface to use type
824 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
825 maxStoresPerMemset = 16;
826 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
828 // On ARM arguments smaller than 4 bytes are extended, so all arguments
829 // are at least 4 bytes aligned.
830 setMinStackArgumentAlignment(4);
832 benefitFromCodePlacementOpt = true;
834 // Prefer likely predicted branches to selects on out-of-order cores.
835 predictableSelectIsExpensive = Subtarget->isLikeA9();
837 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
840 // FIXME: It might make sense to define the representative register class as the
841 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
842 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
843 // SPR's representative would be DPR_VFP2. This should work well if register
844 // pressure tracking were modified such that a register use would increment the
845 // pressure of the register class's representative and all of it's super
846 // classes' representatives transitively. We have not implemented this because
847 // of the difficulty prior to coalescing of modeling operand register classes
848 // due to the common occurrence of cross class copies and subregister insertions
850 std::pair<const TargetRegisterClass*, uint8_t>
851 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
852 const TargetRegisterClass *RRC = 0;
854 switch (VT.getSimpleVT().SimpleTy) {
856 return TargetLowering::findRepresentativeClass(VT);
857 // Use DPR as representative register class for all floating point
858 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
859 // the cost is 1 for both f32 and f64.
860 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
861 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
862 RRC = &ARM::DPRRegClass;
863 // When NEON is used for SP, only half of the register file is available
864 // because operations that define both SP and DP results will be constrained
865 // to the VFP2 class (D0-D15). We currently model this constraint prior to
866 // coalescing by double-counting the SP regs. See the FIXME above.
867 if (Subtarget->useNEONForSinglePrecisionFP())
870 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
871 case MVT::v4f32: case MVT::v2f64:
872 RRC = &ARM::DPRRegClass;
876 RRC = &ARM::DPRRegClass;
880 RRC = &ARM::DPRRegClass;
884 return std::make_pair(RRC, Cost);
887 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
890 case ARMISD::Wrapper: return "ARMISD::Wrapper";
891 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
892 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
893 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
894 case ARMISD::CALL: return "ARMISD::CALL";
895 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
896 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
897 case ARMISD::tCALL: return "ARMISD::tCALL";
898 case ARMISD::BRCOND: return "ARMISD::BRCOND";
899 case ARMISD::BR_JT: return "ARMISD::BR_JT";
900 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
901 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
902 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
903 case ARMISD::CMP: return "ARMISD::CMP";
904 case ARMISD::CMN: return "ARMISD::CMN";
905 case ARMISD::CMPZ: return "ARMISD::CMPZ";
906 case ARMISD::CMPFP: return "ARMISD::CMPFP";
907 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
908 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
909 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
911 case ARMISD::CMOV: return "ARMISD::CMOV";
913 case ARMISD::RBIT: return "ARMISD::RBIT";
915 case ARMISD::FTOSI: return "ARMISD::FTOSI";
916 case ARMISD::FTOUI: return "ARMISD::FTOUI";
917 case ARMISD::SITOF: return "ARMISD::SITOF";
918 case ARMISD::UITOF: return "ARMISD::UITOF";
920 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
921 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
922 case ARMISD::RRX: return "ARMISD::RRX";
924 case ARMISD::ADDC: return "ARMISD::ADDC";
925 case ARMISD::ADDE: return "ARMISD::ADDE";
926 case ARMISD::SUBC: return "ARMISD::SUBC";
927 case ARMISD::SUBE: return "ARMISD::SUBE";
929 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
930 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
932 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
933 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
935 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
937 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
939 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
941 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
942 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
944 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
946 case ARMISD::VCEQ: return "ARMISD::VCEQ";
947 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
948 case ARMISD::VCGE: return "ARMISD::VCGE";
949 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
950 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
951 case ARMISD::VCGEU: return "ARMISD::VCGEU";
952 case ARMISD::VCGT: return "ARMISD::VCGT";
953 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
954 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
955 case ARMISD::VCGTU: return "ARMISD::VCGTU";
956 case ARMISD::VTST: return "ARMISD::VTST";
958 case ARMISD::VSHL: return "ARMISD::VSHL";
959 case ARMISD::VSHRs: return "ARMISD::VSHRs";
960 case ARMISD::VSHRu: return "ARMISD::VSHRu";
961 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
962 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
963 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
964 case ARMISD::VSHRN: return "ARMISD::VSHRN";
965 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
966 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
967 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
968 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
969 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
970 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
971 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
972 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
973 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
974 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
975 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
976 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
977 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
978 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
979 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
980 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
981 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
982 case ARMISD::VDUP: return "ARMISD::VDUP";
983 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
984 case ARMISD::VEXT: return "ARMISD::VEXT";
985 case ARMISD::VREV64: return "ARMISD::VREV64";
986 case ARMISD::VREV32: return "ARMISD::VREV32";
987 case ARMISD::VREV16: return "ARMISD::VREV16";
988 case ARMISD::VZIP: return "ARMISD::VZIP";
989 case ARMISD::VUZP: return "ARMISD::VUZP";
990 case ARMISD::VTRN: return "ARMISD::VTRN";
991 case ARMISD::VTBL1: return "ARMISD::VTBL1";
992 case ARMISD::VTBL2: return "ARMISD::VTBL2";
993 case ARMISD::VMULLs: return "ARMISD::VMULLs";
994 case ARMISD::VMULLu: return "ARMISD::VMULLu";
995 case ARMISD::UMLAL: return "ARMISD::UMLAL";
996 case ARMISD::SMLAL: return "ARMISD::SMLAL";
997 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
998 case ARMISD::FMAX: return "ARMISD::FMAX";
999 case ARMISD::FMIN: return "ARMISD::FMIN";
1000 case ARMISD::BFI: return "ARMISD::BFI";
1001 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1002 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1003 case ARMISD::VBSL: return "ARMISD::VBSL";
1004 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1005 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1006 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1007 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1008 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1009 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1010 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1011 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1012 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1013 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1014 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1015 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1016 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1017 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1018 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1019 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1020 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1021 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1022 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1023 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1027 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1028 if (!VT.isVector()) return getPointerTy();
1029 return VT.changeVectorElementTypeToInteger();
1032 /// getRegClassFor - Return the register class that should be used for the
1033 /// specified value type.
1034 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1035 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1036 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1037 // load / store 4 to 8 consecutive D registers.
1038 if (Subtarget->hasNEON()) {
1039 if (VT == MVT::v4i64)
1040 return &ARM::QQPRRegClass;
1041 if (VT == MVT::v8i64)
1042 return &ARM::QQQQPRRegClass;
1044 return TargetLowering::getRegClassFor(VT);
1047 // Create a fast isel object.
1049 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1050 const TargetLibraryInfo *libInfo) const {
1051 return ARM::createFastISel(funcInfo, libInfo);
1054 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1055 /// be used for loads / stores from the global.
1056 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1057 return (Subtarget->isThumb1Only() ? 127 : 4095);
1060 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1061 unsigned NumVals = N->getNumValues();
1063 return Sched::RegPressure;
1065 for (unsigned i = 0; i != NumVals; ++i) {
1066 EVT VT = N->getValueType(i);
1067 if (VT == MVT::Glue || VT == MVT::Other)
1069 if (VT.isFloatingPoint() || VT.isVector())
1073 if (!N->isMachineOpcode())
1074 return Sched::RegPressure;
1076 // Load are scheduled for latency even if there instruction itinerary
1077 // is not available.
1078 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1079 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1081 if (MCID.getNumDefs() == 0)
1082 return Sched::RegPressure;
1083 if (!Itins->isEmpty() &&
1084 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1087 return Sched::RegPressure;
1090 //===----------------------------------------------------------------------===//
1092 //===----------------------------------------------------------------------===//
1094 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1095 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1097 default: llvm_unreachable("Unknown condition code!");
1098 case ISD::SETNE: return ARMCC::NE;
1099 case ISD::SETEQ: return ARMCC::EQ;
1100 case ISD::SETGT: return ARMCC::GT;
1101 case ISD::SETGE: return ARMCC::GE;
1102 case ISD::SETLT: return ARMCC::LT;
1103 case ISD::SETLE: return ARMCC::LE;
1104 case ISD::SETUGT: return ARMCC::HI;
1105 case ISD::SETUGE: return ARMCC::HS;
1106 case ISD::SETULT: return ARMCC::LO;
1107 case ISD::SETULE: return ARMCC::LS;
1111 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1112 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1113 ARMCC::CondCodes &CondCode2) {
1114 CondCode2 = ARMCC::AL;
1116 default: llvm_unreachable("Unknown FP condition!");
1118 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1120 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1122 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1123 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1124 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1125 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1126 case ISD::SETO: CondCode = ARMCC::VC; break;
1127 case ISD::SETUO: CondCode = ARMCC::VS; break;
1128 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1129 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1130 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1132 case ISD::SETULT: CondCode = ARMCC::LT; break;
1134 case ISD::SETULE: CondCode = ARMCC::LE; break;
1136 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1140 //===----------------------------------------------------------------------===//
1141 // Calling Convention Implementation
1142 //===----------------------------------------------------------------------===//
1144 #include "ARMGenCallingConv.inc"
1146 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1147 /// given CallingConvention value.
1148 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1150 bool isVarArg) const {
1153 llvm_unreachable("Unsupported calling convention");
1154 case CallingConv::Fast:
1155 if (Subtarget->hasVFP2() && !isVarArg) {
1156 if (!Subtarget->isAAPCS_ABI())
1157 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1158 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1159 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1162 case CallingConv::C: {
1163 // Use target triple & subtarget features to do actual dispatch.
1164 if (!Subtarget->isAAPCS_ABI())
1165 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1166 else if (Subtarget->hasVFP2() &&
1167 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1169 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1172 case CallingConv::ARM_AAPCS_VFP:
1174 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1176 case CallingConv::ARM_AAPCS:
1177 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1178 case CallingConv::ARM_APCS:
1179 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1180 case CallingConv::GHC:
1181 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1185 /// LowerCallResult - Lower the result values of a call into the
1186 /// appropriate copies out of appropriate physical registers.
1188 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1189 CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<ISD::InputArg> &Ins,
1191 DebugLoc dl, SelectionDAG &DAG,
1192 SmallVectorImpl<SDValue> &InVals) const {
1194 // Assign locations to each value returned by this call.
1195 SmallVector<CCValAssign, 16> RVLocs;
1196 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1197 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1198 CCInfo.AnalyzeCallResult(Ins,
1199 CCAssignFnForNode(CallConv, /* Return*/ true,
1202 // Copy all of the result registers out of their specified physreg.
1203 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1204 CCValAssign VA = RVLocs[i];
1207 if (VA.needsCustom()) {
1208 // Handle f64 or half of a v2f64.
1209 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1211 Chain = Lo.getValue(1);
1212 InFlag = Lo.getValue(2);
1213 VA = RVLocs[++i]; // skip ahead to next loc
1214 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1216 Chain = Hi.getValue(1);
1217 InFlag = Hi.getValue(2);
1218 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1220 if (VA.getLocVT() == MVT::v2f64) {
1221 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1222 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1223 DAG.getConstant(0, MVT::i32));
1225 VA = RVLocs[++i]; // skip ahead to next loc
1226 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1227 Chain = Lo.getValue(1);
1228 InFlag = Lo.getValue(2);
1229 VA = RVLocs[++i]; // skip ahead to next loc
1230 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1231 Chain = Hi.getValue(1);
1232 InFlag = Hi.getValue(2);
1233 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1234 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1235 DAG.getConstant(1, MVT::i32));
1238 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1240 Chain = Val.getValue(1);
1241 InFlag = Val.getValue(2);
1244 switch (VA.getLocInfo()) {
1245 default: llvm_unreachable("Unknown loc info!");
1246 case CCValAssign::Full: break;
1247 case CCValAssign::BCvt:
1248 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1252 InVals.push_back(Val);
1258 /// LowerMemOpCallTo - Store the argument to the stack.
1260 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1261 SDValue StackPtr, SDValue Arg,
1262 DebugLoc dl, SelectionDAG &DAG,
1263 const CCValAssign &VA,
1264 ISD::ArgFlagsTy Flags) const {
1265 unsigned LocMemOffset = VA.getLocMemOffset();
1266 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1267 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1268 return DAG.getStore(Chain, dl, Arg, PtrOff,
1269 MachinePointerInfo::getStack(LocMemOffset),
1273 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1274 SDValue Chain, SDValue &Arg,
1275 RegsToPassVector &RegsToPass,
1276 CCValAssign &VA, CCValAssign &NextVA,
1278 SmallVector<SDValue, 8> &MemOpChains,
1279 ISD::ArgFlagsTy Flags) const {
1281 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1282 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1283 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1285 if (NextVA.isRegLoc())
1286 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1288 assert(NextVA.isMemLoc());
1289 if (StackPtr.getNode() == 0)
1290 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1292 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1298 /// LowerCall - Lowering a call into a callseq_start <-
1299 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1302 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1303 SmallVectorImpl<SDValue> &InVals) const {
1304 SelectionDAG &DAG = CLI.DAG;
1305 DebugLoc &dl = CLI.DL;
1306 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1307 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1308 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1309 SDValue Chain = CLI.Chain;
1310 SDValue Callee = CLI.Callee;
1311 bool &isTailCall = CLI.IsTailCall;
1312 CallingConv::ID CallConv = CLI.CallConv;
1313 bool doesNotRet = CLI.DoesNotReturn;
1314 bool isVarArg = CLI.IsVarArg;
1316 MachineFunction &MF = DAG.getMachineFunction();
1317 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1318 bool IsSibCall = false;
1319 // Disable tail calls if they're not supported.
1320 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1323 // Check if it's really possible to do a tail call.
1324 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1325 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1326 Outs, OutVals, Ins, DAG);
1327 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1328 // detected sibcalls.
1335 // Analyze operands of the call, assigning locations to each operand.
1336 SmallVector<CCValAssign, 16> ArgLocs;
1337 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1338 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1339 CCInfo.AnalyzeCallOperands(Outs,
1340 CCAssignFnForNode(CallConv, /* Return*/ false,
1343 // Get a count of how many bytes are to be pushed on the stack.
1344 unsigned NumBytes = CCInfo.getNextStackOffset();
1346 // For tail calls, memory operands are available in our caller's stack.
1350 // Adjust the stack pointer for the new arguments...
1351 // These operations are automatically eliminated by the prolog/epilog pass
1353 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1355 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1357 RegsToPassVector RegsToPass;
1358 SmallVector<SDValue, 8> MemOpChains;
1360 // Walk the register/memloc assignments, inserting copies/loads. In the case
1361 // of tail call optimization, arguments are handled later.
1362 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1364 ++i, ++realArgIdx) {
1365 CCValAssign &VA = ArgLocs[i];
1366 SDValue Arg = OutVals[realArgIdx];
1367 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1368 bool isByVal = Flags.isByVal();
1370 // Promote the value if needed.
1371 switch (VA.getLocInfo()) {
1372 default: llvm_unreachable("Unknown loc info!");
1373 case CCValAssign::Full: break;
1374 case CCValAssign::SExt:
1375 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1377 case CCValAssign::ZExt:
1378 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1380 case CCValAssign::AExt:
1381 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1383 case CCValAssign::BCvt:
1384 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1388 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1389 if (VA.needsCustom()) {
1390 if (VA.getLocVT() == MVT::v2f64) {
1391 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1392 DAG.getConstant(0, MVT::i32));
1393 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1394 DAG.getConstant(1, MVT::i32));
1396 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1397 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1399 VA = ArgLocs[++i]; // skip ahead to next loc
1400 if (VA.isRegLoc()) {
1401 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1402 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1404 assert(VA.isMemLoc());
1406 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1407 dl, DAG, VA, Flags));
1410 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1411 StackPtr, MemOpChains, Flags);
1413 } else if (VA.isRegLoc()) {
1414 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1415 } else if (isByVal) {
1416 assert(VA.isMemLoc());
1417 unsigned offset = 0;
1419 // True if this byval aggregate will be split between registers
1421 if (CCInfo.isFirstByValRegValid()) {
1422 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1424 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1425 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1426 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1427 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1428 MachinePointerInfo(),
1429 false, false, false, 0);
1430 MemOpChains.push_back(Load.getValue(1));
1431 RegsToPass.push_back(std::make_pair(j, Load));
1433 offset = ARM::R4 - CCInfo.getFirstByValReg();
1434 CCInfo.clearFirstByValReg();
1437 if (Flags.getByValSize() - 4*offset > 0) {
1438 unsigned LocMemOffset = VA.getLocMemOffset();
1439 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1440 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1442 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1443 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1444 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1446 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1448 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1449 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1450 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1451 Ops, array_lengthof(Ops)));
1453 } else if (!IsSibCall) {
1454 assert(VA.isMemLoc());
1456 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1457 dl, DAG, VA, Flags));
1461 if (!MemOpChains.empty())
1462 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1463 &MemOpChains[0], MemOpChains.size());
1465 // Build a sequence of copy-to-reg nodes chained together with token chain
1466 // and flag operands which copy the outgoing args into the appropriate regs.
1468 // Tail call byval lowering might overwrite argument registers so in case of
1469 // tail call optimization the copies to registers are lowered later.
1471 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1472 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1473 RegsToPass[i].second, InFlag);
1474 InFlag = Chain.getValue(1);
1477 // For tail calls lower the arguments to the 'real' stack slot.
1479 // Force all the incoming stack arguments to be loaded from the stack
1480 // before any new outgoing arguments are stored to the stack, because the
1481 // outgoing stack slots may alias the incoming argument stack slots, and
1482 // the alias isn't otherwise explicit. This is slightly more conservative
1483 // than necessary, because it means that each store effectively depends
1484 // on every argument instead of just those arguments it would clobber.
1486 // Do not flag preceding copytoreg stuff together with the following stuff.
1488 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1489 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1490 RegsToPass[i].second, InFlag);
1491 InFlag = Chain.getValue(1);
1496 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1497 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1498 // node so that legalize doesn't hack it.
1499 bool isDirect = false;
1500 bool isARMFunc = false;
1501 bool isLocalARMFunc = false;
1502 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1504 if (EnableARMLongCalls) {
1505 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1506 && "long-calls with non-static relocation model!");
1507 // Handle a global address or an external symbol. If it's not one of
1508 // those, the target's already in a register, so we don't need to do
1510 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1511 const GlobalValue *GV = G->getGlobal();
1512 // Create a constant pool entry for the callee address
1513 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1514 ARMConstantPoolValue *CPV =
1515 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1517 // Get the address of the callee into a register
1518 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1519 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1520 Callee = DAG.getLoad(getPointerTy(), dl,
1521 DAG.getEntryNode(), CPAddr,
1522 MachinePointerInfo::getConstantPool(),
1523 false, false, false, 0);
1524 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1525 const char *Sym = S->getSymbol();
1527 // Create a constant pool entry for the callee address
1528 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1529 ARMConstantPoolValue *CPV =
1530 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1531 ARMPCLabelIndex, 0);
1532 // Get the address of the callee into a register
1533 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1534 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1535 Callee = DAG.getLoad(getPointerTy(), dl,
1536 DAG.getEntryNode(), CPAddr,
1537 MachinePointerInfo::getConstantPool(),
1538 false, false, false, 0);
1540 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1541 const GlobalValue *GV = G->getGlobal();
1543 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1544 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1545 getTargetMachine().getRelocationModel() != Reloc::Static;
1546 isARMFunc = !Subtarget->isThumb() || isStub;
1547 // ARM call to a local ARM function is predicable.
1548 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1549 // tBX takes a register source operand.
1550 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1551 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1552 ARMConstantPoolValue *CPV =
1553 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1554 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1555 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1556 Callee = DAG.getLoad(getPointerTy(), dl,
1557 DAG.getEntryNode(), CPAddr,
1558 MachinePointerInfo::getConstantPool(),
1559 false, false, false, 0);
1560 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1561 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1562 getPointerTy(), Callee, PICLabel);
1564 // On ELF targets for PIC code, direct calls should go through the PLT
1565 unsigned OpFlags = 0;
1566 if (Subtarget->isTargetELF() &&
1567 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1568 OpFlags = ARMII::MO_PLT;
1569 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1571 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1573 bool isStub = Subtarget->isTargetDarwin() &&
1574 getTargetMachine().getRelocationModel() != Reloc::Static;
1575 isARMFunc = !Subtarget->isThumb() || isStub;
1576 // tBX takes a register source operand.
1577 const char *Sym = S->getSymbol();
1578 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1579 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1580 ARMConstantPoolValue *CPV =
1581 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1582 ARMPCLabelIndex, 4);
1583 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1584 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1585 Callee = DAG.getLoad(getPointerTy(), dl,
1586 DAG.getEntryNode(), CPAddr,
1587 MachinePointerInfo::getConstantPool(),
1588 false, false, false, 0);
1589 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1590 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1591 getPointerTy(), Callee, PICLabel);
1593 unsigned OpFlags = 0;
1594 // On ELF targets for PIC code, direct calls should go through the PLT
1595 if (Subtarget->isTargetELF() &&
1596 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1597 OpFlags = ARMII::MO_PLT;
1598 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1602 // FIXME: handle tail calls differently.
1604 bool HasMinSizeAttr = MF.getFunction()->getFnAttributes().
1605 hasAttribute(Attributes::MinSize);
1606 if (Subtarget->isThumb()) {
1607 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1608 CallOpc = ARMISD::CALL_NOLINK;
1610 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1612 if (!isDirect && !Subtarget->hasV5TOps())
1613 CallOpc = ARMISD::CALL_NOLINK;
1614 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1615 // Emit regular call when code size is the priority
1617 // "mov lr, pc; b _foo" to avoid confusing the RSP
1618 CallOpc = ARMISD::CALL_NOLINK;
1620 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1623 std::vector<SDValue> Ops;
1624 Ops.push_back(Chain);
1625 Ops.push_back(Callee);
1627 // Add argument registers to the end of the list so that they are known live
1629 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1630 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1631 RegsToPass[i].second.getValueType()));
1633 // Add a register mask operand representing the call-preserved registers.
1634 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1635 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1636 assert(Mask && "Missing call preserved mask for calling convention");
1637 Ops.push_back(DAG.getRegisterMask(Mask));
1639 if (InFlag.getNode())
1640 Ops.push_back(InFlag);
1642 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1644 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1646 // Returns a chain and a flag for retval copy to use.
1647 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1648 InFlag = Chain.getValue(1);
1650 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1651 DAG.getIntPtrConstant(0, true), InFlag);
1653 InFlag = Chain.getValue(1);
1655 // Handle result values, copying them out of physregs into vregs that we
1657 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1661 /// HandleByVal - Every parameter *after* a byval parameter is passed
1662 /// on the stack. Remember the next parameter register to allocate,
1663 /// and then confiscate the rest of the parameter registers to insure
1666 ARMTargetLowering::HandleByVal(
1667 CCState *State, unsigned &size, unsigned Align) const {
1668 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1669 assert((State->getCallOrPrologue() == Prologue ||
1670 State->getCallOrPrologue() == Call) &&
1671 "unhandled ParmContext");
1672 if ((!State->isFirstByValRegValid()) &&
1673 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1674 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1675 unsigned AlignInRegs = Align / 4;
1676 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1677 for (unsigned i = 0; i < Waste; ++i)
1678 reg = State->AllocateReg(GPRArgRegs, 4);
1681 State->setFirstByValReg(reg);
1682 // At a call site, a byval parameter that is split between
1683 // registers and memory needs its size truncated here. In a
1684 // function prologue, such byval parameters are reassembled in
1685 // memory, and are not truncated.
1686 if (State->getCallOrPrologue() == Call) {
1687 unsigned excess = 4 * (ARM::R4 - reg);
1688 assert(size >= excess && "expected larger existing stack allocation");
1693 // Confiscate any remaining parameter registers to preclude their
1694 // assignment to subsequent parameters.
1695 while (State->AllocateReg(GPRArgRegs, 4))
1699 /// MatchingStackOffset - Return true if the given stack call argument is
1700 /// already available in the same position (relatively) of the caller's
1701 /// incoming argument stack.
1703 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1704 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1705 const TargetInstrInfo *TII) {
1706 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1708 if (Arg.getOpcode() == ISD::CopyFromReg) {
1709 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1710 if (!TargetRegisterInfo::isVirtualRegister(VR))
1712 MachineInstr *Def = MRI->getVRegDef(VR);
1715 if (!Flags.isByVal()) {
1716 if (!TII->isLoadFromStackSlot(Def, FI))
1721 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1722 if (Flags.isByVal())
1723 // ByVal argument is passed in as a pointer but it's now being
1724 // dereferenced. e.g.
1725 // define @foo(%struct.X* %A) {
1726 // tail call @bar(%struct.X* byval %A)
1729 SDValue Ptr = Ld->getBasePtr();
1730 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1733 FI = FINode->getIndex();
1737 assert(FI != INT_MAX);
1738 if (!MFI->isFixedObjectIndex(FI))
1740 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1743 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1744 /// for tail call optimization. Targets which want to do tail call
1745 /// optimization should implement this function.
1747 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1748 CallingConv::ID CalleeCC,
1750 bool isCalleeStructRet,
1751 bool isCallerStructRet,
1752 const SmallVectorImpl<ISD::OutputArg> &Outs,
1753 const SmallVectorImpl<SDValue> &OutVals,
1754 const SmallVectorImpl<ISD::InputArg> &Ins,
1755 SelectionDAG& DAG) const {
1756 const Function *CallerF = DAG.getMachineFunction().getFunction();
1757 CallingConv::ID CallerCC = CallerF->getCallingConv();
1758 bool CCMatch = CallerCC == CalleeCC;
1760 // Look for obvious safe cases to perform tail call optimization that do not
1761 // require ABI changes. This is what gcc calls sibcall.
1763 // Do not sibcall optimize vararg calls unless the call site is not passing
1765 if (isVarArg && !Outs.empty())
1768 // Also avoid sibcall optimization if either caller or callee uses struct
1769 // return semantics.
1770 if (isCalleeStructRet || isCallerStructRet)
1773 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1774 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1775 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1776 // support in the assembler and linker to be used. This would need to be
1777 // fixed to fully support tail calls in Thumb1.
1779 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1780 // LR. This means if we need to reload LR, it takes an extra instructions,
1781 // which outweighs the value of the tail call; but here we don't know yet
1782 // whether LR is going to be used. Probably the right approach is to
1783 // generate the tail call here and turn it back into CALL/RET in
1784 // emitEpilogue if LR is used.
1786 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1787 // but we need to make sure there are enough registers; the only valid
1788 // registers are the 4 used for parameters. We don't currently do this
1790 if (Subtarget->isThumb1Only())
1793 // If the calling conventions do not match, then we'd better make sure the
1794 // results are returned in the same way as what the caller expects.
1796 SmallVector<CCValAssign, 16> RVLocs1;
1797 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1798 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1799 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1801 SmallVector<CCValAssign, 16> RVLocs2;
1802 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1803 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1804 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1806 if (RVLocs1.size() != RVLocs2.size())
1808 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1809 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1811 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1813 if (RVLocs1[i].isRegLoc()) {
1814 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1817 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1823 // If Caller's vararg or byval argument has been split between registers and
1824 // stack, do not perform tail call, since part of the argument is in caller's
1826 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1827 getInfo<ARMFunctionInfo>();
1828 if (AFI_Caller->getVarArgsRegSaveSize())
1831 // If the callee takes no arguments then go on to check the results of the
1833 if (!Outs.empty()) {
1834 // Check if stack adjustment is needed. For now, do not do this if any
1835 // argument is passed on the stack.
1836 SmallVector<CCValAssign, 16> ArgLocs;
1837 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1838 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1839 CCInfo.AnalyzeCallOperands(Outs,
1840 CCAssignFnForNode(CalleeCC, false, isVarArg));
1841 if (CCInfo.getNextStackOffset()) {
1842 MachineFunction &MF = DAG.getMachineFunction();
1844 // Check if the arguments are already laid out in the right way as
1845 // the caller's fixed stack objects.
1846 MachineFrameInfo *MFI = MF.getFrameInfo();
1847 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1848 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1849 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1851 ++i, ++realArgIdx) {
1852 CCValAssign &VA = ArgLocs[i];
1853 EVT RegVT = VA.getLocVT();
1854 SDValue Arg = OutVals[realArgIdx];
1855 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1856 if (VA.getLocInfo() == CCValAssign::Indirect)
1858 if (VA.needsCustom()) {
1859 // f64 and vector types are split into multiple registers or
1860 // register/stack-slot combinations. The types will not match
1861 // the registers; give up on memory f64 refs until we figure
1862 // out what to do about this.
1865 if (!ArgLocs[++i].isRegLoc())
1867 if (RegVT == MVT::v2f64) {
1868 if (!ArgLocs[++i].isRegLoc())
1870 if (!ArgLocs[++i].isRegLoc())
1873 } else if (!VA.isRegLoc()) {
1874 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1886 ARMTargetLowering::LowerReturn(SDValue Chain,
1887 CallingConv::ID CallConv, bool isVarArg,
1888 const SmallVectorImpl<ISD::OutputArg> &Outs,
1889 const SmallVectorImpl<SDValue> &OutVals,
1890 DebugLoc dl, SelectionDAG &DAG) const {
1892 // CCValAssign - represent the assignment of the return value to a location.
1893 SmallVector<CCValAssign, 16> RVLocs;
1895 // CCState - Info about the registers and stack slots.
1896 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1897 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1899 // Analyze outgoing return values.
1900 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1903 // If this is the first return lowered for this function, add
1904 // the regs to the liveout set for the function.
1905 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1906 for (unsigned i = 0; i != RVLocs.size(); ++i)
1907 if (RVLocs[i].isRegLoc())
1908 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1913 // Copy the result values into the output registers.
1914 for (unsigned i = 0, realRVLocIdx = 0;
1916 ++i, ++realRVLocIdx) {
1917 CCValAssign &VA = RVLocs[i];
1918 assert(VA.isRegLoc() && "Can only return in registers!");
1920 SDValue Arg = OutVals[realRVLocIdx];
1922 switch (VA.getLocInfo()) {
1923 default: llvm_unreachable("Unknown loc info!");
1924 case CCValAssign::Full: break;
1925 case CCValAssign::BCvt:
1926 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1930 if (VA.needsCustom()) {
1931 if (VA.getLocVT() == MVT::v2f64) {
1932 // Extract the first half and return it in two registers.
1933 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1934 DAG.getConstant(0, MVT::i32));
1935 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1936 DAG.getVTList(MVT::i32, MVT::i32), Half);
1938 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1939 Flag = Chain.getValue(1);
1940 VA = RVLocs[++i]; // skip ahead to next loc
1941 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1942 HalfGPRs.getValue(1), Flag);
1943 Flag = Chain.getValue(1);
1944 VA = RVLocs[++i]; // skip ahead to next loc
1946 // Extract the 2nd half and fall through to handle it as an f64 value.
1947 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1948 DAG.getConstant(1, MVT::i32));
1950 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1952 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1953 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1954 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1955 Flag = Chain.getValue(1);
1956 VA = RVLocs[++i]; // skip ahead to next loc
1957 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1960 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1962 // Guarantee that all emitted copies are
1963 // stuck together, avoiding something bad.
1964 Flag = Chain.getValue(1);
1969 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1971 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1976 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1977 if (N->getNumValues() != 1)
1979 if (!N->hasNUsesOfValue(1, 0))
1982 SDValue TCChain = Chain;
1983 SDNode *Copy = *N->use_begin();
1984 if (Copy->getOpcode() == ISD::CopyToReg) {
1985 // If the copy has a glue operand, we conservatively assume it isn't safe to
1986 // perform a tail call.
1987 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1989 TCChain = Copy->getOperand(0);
1990 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1991 SDNode *VMov = Copy;
1992 // f64 returned in a pair of GPRs.
1993 SmallPtrSet<SDNode*, 2> Copies;
1994 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1996 if (UI->getOpcode() != ISD::CopyToReg)
2000 if (Copies.size() > 2)
2003 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2005 SDValue UseChain = UI->getOperand(0);
2006 if (Copies.count(UseChain.getNode()))
2013 } else if (Copy->getOpcode() == ISD::BITCAST) {
2014 // f32 returned in a single GPR.
2015 if (!Copy->hasOneUse())
2017 Copy = *Copy->use_begin();
2018 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2020 Chain = Copy->getOperand(0);
2025 bool HasRet = false;
2026 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2028 if (UI->getOpcode() != ARMISD::RET_FLAG)
2040 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2041 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2044 if (!CI->isTailCall())
2047 return !Subtarget->isThumb1Only();
2050 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2051 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2052 // one of the above mentioned nodes. It has to be wrapped because otherwise
2053 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2054 // be used to form addressing mode. These wrapped nodes will be selected
2056 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2057 EVT PtrVT = Op.getValueType();
2058 // FIXME there is no actual debug info here
2059 DebugLoc dl = Op.getDebugLoc();
2060 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2062 if (CP->isMachineConstantPoolEntry())
2063 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2064 CP->getAlignment());
2066 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2067 CP->getAlignment());
2068 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2071 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2072 return MachineJumpTableInfo::EK_Inline;
2075 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2076 SelectionDAG &DAG) const {
2077 MachineFunction &MF = DAG.getMachineFunction();
2078 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2079 unsigned ARMPCLabelIndex = 0;
2080 DebugLoc DL = Op.getDebugLoc();
2081 EVT PtrVT = getPointerTy();
2082 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2083 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2085 if (RelocM == Reloc::Static) {
2086 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2088 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2089 ARMPCLabelIndex = AFI->createPICLabelUId();
2090 ARMConstantPoolValue *CPV =
2091 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2092 ARMCP::CPBlockAddress, PCAdj);
2093 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2095 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2096 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2097 MachinePointerInfo::getConstantPool(),
2098 false, false, false, 0);
2099 if (RelocM == Reloc::Static)
2101 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2102 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2105 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2107 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2108 SelectionDAG &DAG) const {
2109 DebugLoc dl = GA->getDebugLoc();
2110 EVT PtrVT = getPointerTy();
2111 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2112 MachineFunction &MF = DAG.getMachineFunction();
2113 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2114 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2115 ARMConstantPoolValue *CPV =
2116 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2117 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2118 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2119 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2120 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2121 MachinePointerInfo::getConstantPool(),
2122 false, false, false, 0);
2123 SDValue Chain = Argument.getValue(1);
2125 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2126 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2128 // call __tls_get_addr.
2131 Entry.Node = Argument;
2132 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2133 Args.push_back(Entry);
2134 // FIXME: is there useful debug info available here?
2135 TargetLowering::CallLoweringInfo CLI(Chain,
2136 (Type *) Type::getInt32Ty(*DAG.getContext()),
2137 false, false, false, false,
2138 0, CallingConv::C, /*isTailCall=*/false,
2139 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2140 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2141 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2142 return CallResult.first;
2145 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2146 // "local exec" model.
2148 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2150 TLSModel::Model model) const {
2151 const GlobalValue *GV = GA->getGlobal();
2152 DebugLoc dl = GA->getDebugLoc();
2154 SDValue Chain = DAG.getEntryNode();
2155 EVT PtrVT = getPointerTy();
2156 // Get the Thread Pointer
2157 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2159 if (model == TLSModel::InitialExec) {
2160 MachineFunction &MF = DAG.getMachineFunction();
2161 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2162 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2163 // Initial exec model.
2164 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2165 ARMConstantPoolValue *CPV =
2166 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2167 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2169 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2170 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2171 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2172 MachinePointerInfo::getConstantPool(),
2173 false, false, false, 0);
2174 Chain = Offset.getValue(1);
2176 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2177 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2179 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2180 MachinePointerInfo::getConstantPool(),
2181 false, false, false, 0);
2184 assert(model == TLSModel::LocalExec);
2185 ARMConstantPoolValue *CPV =
2186 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2187 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2188 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2189 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2190 MachinePointerInfo::getConstantPool(),
2191 false, false, false, 0);
2194 // The address of the thread local variable is the add of the thread
2195 // pointer with the offset of the variable.
2196 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2200 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2201 // TODO: implement the "local dynamic" model
2202 assert(Subtarget->isTargetELF() &&
2203 "TLS not implemented for non-ELF targets");
2204 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2206 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2209 case TLSModel::GeneralDynamic:
2210 case TLSModel::LocalDynamic:
2211 return LowerToTLSGeneralDynamicModel(GA, DAG);
2212 case TLSModel::InitialExec:
2213 case TLSModel::LocalExec:
2214 return LowerToTLSExecModels(GA, DAG, model);
2216 llvm_unreachable("bogus TLS model");
2219 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2220 SelectionDAG &DAG) const {
2221 EVT PtrVT = getPointerTy();
2222 DebugLoc dl = Op.getDebugLoc();
2223 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2224 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2225 if (RelocM == Reloc::PIC_) {
2226 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2227 ARMConstantPoolValue *CPV =
2228 ARMConstantPoolConstant::Create(GV,
2229 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2230 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2231 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2232 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2234 MachinePointerInfo::getConstantPool(),
2235 false, false, false, 0);
2236 SDValue Chain = Result.getValue(1);
2237 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2238 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2240 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2241 MachinePointerInfo::getGOT(),
2242 false, false, false, 0);
2246 // If we have T2 ops, we can materialize the address directly via movt/movw
2247 // pair. This is always cheaper.
2248 if (Subtarget->useMovt()) {
2250 // FIXME: Once remat is capable of dealing with instructions with register
2251 // operands, expand this into two nodes.
2252 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2253 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2255 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2256 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2257 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2258 MachinePointerInfo::getConstantPool(),
2259 false, false, false, 0);
2263 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2264 SelectionDAG &DAG) const {
2265 EVT PtrVT = getPointerTy();
2266 DebugLoc dl = Op.getDebugLoc();
2267 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2268 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2269 MachineFunction &MF = DAG.getMachineFunction();
2270 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2272 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2273 // update ARMFastISel::ARMMaterializeGV.
2274 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2276 // FIXME: Once remat is capable of dealing with instructions with register
2277 // operands, expand this into two nodes.
2278 if (RelocM == Reloc::Static)
2279 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2280 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2282 unsigned Wrapper = (RelocM == Reloc::PIC_)
2283 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2284 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2285 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2286 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2287 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2288 MachinePointerInfo::getGOT(),
2289 false, false, false, 0);
2293 unsigned ARMPCLabelIndex = 0;
2295 if (RelocM == Reloc::Static) {
2296 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2298 ARMPCLabelIndex = AFI->createPICLabelUId();
2299 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2300 ARMConstantPoolValue *CPV =
2301 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2303 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2305 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2307 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2308 MachinePointerInfo::getConstantPool(),
2309 false, false, false, 0);
2310 SDValue Chain = Result.getValue(1);
2312 if (RelocM == Reloc::PIC_) {
2313 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2314 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2317 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2318 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2319 false, false, false, 0);
2324 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2325 SelectionDAG &DAG) const {
2326 assert(Subtarget->isTargetELF() &&
2327 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2328 MachineFunction &MF = DAG.getMachineFunction();
2329 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2330 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2331 EVT PtrVT = getPointerTy();
2332 DebugLoc dl = Op.getDebugLoc();
2333 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2334 ARMConstantPoolValue *CPV =
2335 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2336 ARMPCLabelIndex, PCAdj);
2337 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2338 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2339 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2340 MachinePointerInfo::getConstantPool(),
2341 false, false, false, 0);
2342 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2343 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2347 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2348 DebugLoc dl = Op.getDebugLoc();
2349 SDValue Val = DAG.getConstant(0, MVT::i32);
2350 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2351 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2352 Op.getOperand(1), Val);
2356 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2357 DebugLoc dl = Op.getDebugLoc();
2358 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2359 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2363 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2364 const ARMSubtarget *Subtarget) const {
2365 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2366 DebugLoc dl = Op.getDebugLoc();
2368 default: return SDValue(); // Don't custom lower most intrinsics.
2369 case Intrinsic::arm_thread_pointer: {
2370 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2371 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2373 case Intrinsic::eh_sjlj_lsda: {
2374 MachineFunction &MF = DAG.getMachineFunction();
2375 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2376 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2377 EVT PtrVT = getPointerTy();
2378 DebugLoc dl = Op.getDebugLoc();
2379 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2381 unsigned PCAdj = (RelocM != Reloc::PIC_)
2382 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2383 ARMConstantPoolValue *CPV =
2384 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2385 ARMCP::CPLSDA, PCAdj);
2386 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2387 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2389 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2390 MachinePointerInfo::getConstantPool(),
2391 false, false, false, 0);
2393 if (RelocM == Reloc::PIC_) {
2394 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2395 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2399 case Intrinsic::arm_neon_vmulls:
2400 case Intrinsic::arm_neon_vmullu: {
2401 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2402 ? ARMISD::VMULLs : ARMISD::VMULLu;
2403 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2404 Op.getOperand(1), Op.getOperand(2));
2409 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2410 const ARMSubtarget *Subtarget) {
2411 DebugLoc dl = Op.getDebugLoc();
2412 if (!Subtarget->hasDataBarrier()) {
2413 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2414 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2416 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2417 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2418 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2419 DAG.getConstant(0, MVT::i32));
2422 SDValue Op5 = Op.getOperand(5);
2423 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2424 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2425 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2426 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2428 ARM_MB::MemBOpt DMBOpt;
2429 if (isDeviceBarrier)
2430 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2432 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2433 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2434 DAG.getConstant(DMBOpt, MVT::i32));
2438 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2439 const ARMSubtarget *Subtarget) {
2440 // FIXME: handle "fence singlethread" more efficiently.
2441 DebugLoc dl = Op.getDebugLoc();
2442 if (!Subtarget->hasDataBarrier()) {
2443 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2444 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2446 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2447 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2448 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2449 DAG.getConstant(0, MVT::i32));
2452 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2453 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2456 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2457 const ARMSubtarget *Subtarget) {
2458 // ARM pre v5TE and Thumb1 does not have preload instructions.
2459 if (!(Subtarget->isThumb2() ||
2460 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2461 // Just preserve the chain.
2462 return Op.getOperand(0);
2464 DebugLoc dl = Op.getDebugLoc();
2465 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2467 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2468 // ARMv7 with MP extension has PLDW.
2469 return Op.getOperand(0);
2471 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2472 if (Subtarget->isThumb()) {
2474 isRead = ~isRead & 1;
2475 isData = ~isData & 1;
2478 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2479 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2480 DAG.getConstant(isData, MVT::i32));
2483 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2484 MachineFunction &MF = DAG.getMachineFunction();
2485 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2487 // vastart just stores the address of the VarArgsFrameIndex slot into the
2488 // memory location argument.
2489 DebugLoc dl = Op.getDebugLoc();
2490 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2491 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2492 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2493 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2494 MachinePointerInfo(SV), false, false, 0);
2498 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2499 SDValue &Root, SelectionDAG &DAG,
2500 DebugLoc dl) const {
2501 MachineFunction &MF = DAG.getMachineFunction();
2502 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2504 const TargetRegisterClass *RC;
2505 if (AFI->isThumb1OnlyFunction())
2506 RC = &ARM::tGPRRegClass;
2508 RC = &ARM::GPRRegClass;
2510 // Transform the arguments stored in physical registers into virtual ones.
2511 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2512 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2515 if (NextVA.isMemLoc()) {
2516 MachineFrameInfo *MFI = MF.getFrameInfo();
2517 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2519 // Create load node to retrieve arguments from the stack.
2520 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2521 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2522 MachinePointerInfo::getFixedStack(FI),
2523 false, false, false, 0);
2525 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2526 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2529 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2533 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2534 unsigned &VARegSize, unsigned &VARegSaveSize)
2537 if (CCInfo.isFirstByValRegValid())
2538 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2540 unsigned int firstUnalloced;
2541 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2542 sizeof(GPRArgRegs) /
2543 sizeof(GPRArgRegs[0]));
2544 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2547 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2548 VARegSize = NumGPRs * 4;
2549 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2552 // The remaining GPRs hold either the beginning of variable-argument
2553 // data, or the beginning of an aggregate passed by value (usuall
2554 // byval). Either way, we allocate stack slots adjacent to the data
2555 // provided by our caller, and store the unallocated registers there.
2556 // If this is a variadic function, the va_list pointer will begin with
2557 // these values; otherwise, this reassembles a (byval) structure that
2558 // was split between registers and memory.
2560 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2561 DebugLoc dl, SDValue &Chain,
2562 const Value *OrigArg,
2563 unsigned OffsetFromOrigArg,
2565 bool ForceMutable) const {
2566 MachineFunction &MF = DAG.getMachineFunction();
2567 MachineFrameInfo *MFI = MF.getFrameInfo();
2568 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2569 unsigned firstRegToSaveIndex;
2570 if (CCInfo.isFirstByValRegValid())
2571 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2573 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2574 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2577 unsigned VARegSize, VARegSaveSize;
2578 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2579 if (VARegSaveSize) {
2580 // If this function is vararg, store any remaining integer argument regs
2581 // to their spots on the stack so that they may be loaded by deferencing
2582 // the result of va_next.
2583 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2584 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2585 ArgOffset + VARegSaveSize
2588 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2591 SmallVector<SDValue, 4> MemOps;
2592 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2593 const TargetRegisterClass *RC;
2594 if (AFI->isThumb1OnlyFunction())
2595 RC = &ARM::tGPRRegClass;
2597 RC = &ARM::GPRRegClass;
2599 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2600 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2602 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2603 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2605 MemOps.push_back(Store);
2606 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2607 DAG.getConstant(4, getPointerTy()));
2609 if (!MemOps.empty())
2610 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2611 &MemOps[0], MemOps.size());
2613 // This will point to the next argument passed via stack.
2614 AFI->setVarArgsFrameIndex(
2615 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2619 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2620 CallingConv::ID CallConv, bool isVarArg,
2621 const SmallVectorImpl<ISD::InputArg>
2623 DebugLoc dl, SelectionDAG &DAG,
2624 SmallVectorImpl<SDValue> &InVals)
2626 MachineFunction &MF = DAG.getMachineFunction();
2627 MachineFrameInfo *MFI = MF.getFrameInfo();
2629 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2631 // Assign locations to all of the incoming arguments.
2632 SmallVector<CCValAssign, 16> ArgLocs;
2633 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2634 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2635 CCInfo.AnalyzeFormalArguments(Ins,
2636 CCAssignFnForNode(CallConv, /* Return*/ false,
2639 SmallVector<SDValue, 16> ArgValues;
2640 int lastInsIndex = -1;
2642 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2643 unsigned CurArgIdx = 0;
2644 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2645 CCValAssign &VA = ArgLocs[i];
2646 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2647 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2648 // Arguments stored in registers.
2649 if (VA.isRegLoc()) {
2650 EVT RegVT = VA.getLocVT();
2652 if (VA.needsCustom()) {
2653 // f64 and vector types are split up into multiple registers or
2654 // combinations of registers and stack slots.
2655 if (VA.getLocVT() == MVT::v2f64) {
2656 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2658 VA = ArgLocs[++i]; // skip ahead to next loc
2660 if (VA.isMemLoc()) {
2661 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2662 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2663 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2664 MachinePointerInfo::getFixedStack(FI),
2665 false, false, false, 0);
2667 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2670 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2671 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2672 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2673 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2674 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2676 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2679 const TargetRegisterClass *RC;
2681 if (RegVT == MVT::f32)
2682 RC = &ARM::SPRRegClass;
2683 else if (RegVT == MVT::f64)
2684 RC = &ARM::DPRRegClass;
2685 else if (RegVT == MVT::v2f64)
2686 RC = &ARM::QPRRegClass;
2687 else if (RegVT == MVT::i32)
2688 RC = AFI->isThumb1OnlyFunction() ?
2689 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2690 (const TargetRegisterClass*)&ARM::GPRRegClass;
2692 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2694 // Transform the arguments in physical registers into virtual ones.
2695 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2696 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2699 // If this is an 8 or 16-bit value, it is really passed promoted
2700 // to 32 bits. Insert an assert[sz]ext to capture this, then
2701 // truncate to the right size.
2702 switch (VA.getLocInfo()) {
2703 default: llvm_unreachable("Unknown loc info!");
2704 case CCValAssign::Full: break;
2705 case CCValAssign::BCvt:
2706 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2708 case CCValAssign::SExt:
2709 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2710 DAG.getValueType(VA.getValVT()));
2711 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2713 case CCValAssign::ZExt:
2714 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2715 DAG.getValueType(VA.getValVT()));
2716 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2720 InVals.push_back(ArgValue);
2722 } else { // VA.isRegLoc()
2725 assert(VA.isMemLoc());
2726 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2728 int index = ArgLocs[i].getValNo();
2730 // Some Ins[] entries become multiple ArgLoc[] entries.
2731 // Process them only once.
2732 if (index != lastInsIndex)
2734 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2735 // FIXME: For now, all byval parameter objects are marked mutable.
2736 // This can be changed with more analysis.
2737 // In case of tail call optimization mark all arguments mutable.
2738 // Since they could be overwritten by lowering of arguments in case of
2740 if (Flags.isByVal()) {
2741 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2742 if (!AFI->getVarArgsFrameIndex()) {
2743 VarArgStyleRegisters(CCInfo, DAG,
2744 dl, Chain, CurOrigArg,
2745 Ins[VA.getValNo()].PartOffset,
2746 VA.getLocMemOffset(),
2747 true /*force mutable frames*/);
2748 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2749 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2751 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2752 VA.getLocMemOffset(), false);
2753 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2756 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2757 VA.getLocMemOffset(), true);
2759 // Create load nodes to retrieve arguments from the stack.
2760 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2761 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2762 MachinePointerInfo::getFixedStack(FI),
2763 false, false, false, 0));
2765 lastInsIndex = index;
2772 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2773 CCInfo.getNextStackOffset());
2778 /// isFloatingPointZero - Return true if this is +0.0.
2779 static bool isFloatingPointZero(SDValue Op) {
2780 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2781 return CFP->getValueAPF().isPosZero();
2782 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2783 // Maybe this has already been legalized into the constant pool?
2784 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2785 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2786 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2787 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2788 return CFP->getValueAPF().isPosZero();
2794 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2795 /// the given operands.
2797 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2798 SDValue &ARMcc, SelectionDAG &DAG,
2799 DebugLoc dl) const {
2800 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2801 unsigned C = RHSC->getZExtValue();
2802 if (!isLegalICmpImmediate(C)) {
2803 // Constant does not fit, try adjusting it by one?
2808 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2809 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2810 RHS = DAG.getConstant(C-1, MVT::i32);
2815 if (C != 0 && isLegalICmpImmediate(C-1)) {
2816 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2817 RHS = DAG.getConstant(C-1, MVT::i32);
2822 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2823 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2824 RHS = DAG.getConstant(C+1, MVT::i32);
2829 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2830 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2831 RHS = DAG.getConstant(C+1, MVT::i32);
2838 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2839 ARMISD::NodeType CompareType;
2842 CompareType = ARMISD::CMP;
2847 CompareType = ARMISD::CMPZ;
2850 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2851 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2854 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2856 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2857 DebugLoc dl) const {
2859 if (!isFloatingPointZero(RHS))
2860 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2862 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2863 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2866 /// duplicateCmp - Glue values can have only one use, so this function
2867 /// duplicates a comparison node.
2869 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2870 unsigned Opc = Cmp.getOpcode();
2871 DebugLoc DL = Cmp.getDebugLoc();
2872 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2873 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2875 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2876 Cmp = Cmp.getOperand(0);
2877 Opc = Cmp.getOpcode();
2878 if (Opc == ARMISD::CMPFP)
2879 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2881 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2882 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2884 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2887 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2888 SDValue Cond = Op.getOperand(0);
2889 SDValue SelectTrue = Op.getOperand(1);
2890 SDValue SelectFalse = Op.getOperand(2);
2891 DebugLoc dl = Op.getDebugLoc();
2895 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2896 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2898 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2899 const ConstantSDNode *CMOVTrue =
2900 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2901 const ConstantSDNode *CMOVFalse =
2902 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2904 if (CMOVTrue && CMOVFalse) {
2905 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2906 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2910 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2912 False = SelectFalse;
2913 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2918 if (True.getNode() && False.getNode()) {
2919 EVT VT = Op.getValueType();
2920 SDValue ARMcc = Cond.getOperand(2);
2921 SDValue CCR = Cond.getOperand(3);
2922 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2923 assert(True.getValueType() == VT);
2924 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2929 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2930 // undefined bits before doing a full-word comparison with zero.
2931 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2932 DAG.getConstant(1, Cond.getValueType()));
2934 return DAG.getSelectCC(dl, Cond,
2935 DAG.getConstant(0, Cond.getValueType()),
2936 SelectTrue, SelectFalse, ISD::SETNE);
2939 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2940 EVT VT = Op.getValueType();
2941 SDValue LHS = Op.getOperand(0);
2942 SDValue RHS = Op.getOperand(1);
2943 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2944 SDValue TrueVal = Op.getOperand(2);
2945 SDValue FalseVal = Op.getOperand(3);
2946 DebugLoc dl = Op.getDebugLoc();
2948 if (LHS.getValueType() == MVT::i32) {
2950 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2951 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2952 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2955 ARMCC::CondCodes CondCode, CondCode2;
2956 FPCCToARMCC(CC, CondCode, CondCode2);
2958 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2959 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2960 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2961 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2963 if (CondCode2 != ARMCC::AL) {
2964 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2965 // FIXME: Needs another CMP because flag can have but one use.
2966 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2967 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2968 Result, TrueVal, ARMcc2, CCR, Cmp2);
2973 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2974 /// to morph to an integer compare sequence.
2975 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2976 const ARMSubtarget *Subtarget) {
2977 SDNode *N = Op.getNode();
2978 if (!N->hasOneUse())
2979 // Otherwise it requires moving the value from fp to integer registers.
2981 if (!N->getNumValues())
2983 EVT VT = Op.getValueType();
2984 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2985 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2986 // vmrs are very slow, e.g. cortex-a8.
2989 if (isFloatingPointZero(Op)) {
2993 return ISD::isNormalLoad(N);
2996 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2997 if (isFloatingPointZero(Op))
2998 return DAG.getConstant(0, MVT::i32);
3000 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3001 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3002 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3003 Ld->isVolatile(), Ld->isNonTemporal(),
3004 Ld->isInvariant(), Ld->getAlignment());
3006 llvm_unreachable("Unknown VFP cmp argument!");
3009 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3010 SDValue &RetVal1, SDValue &RetVal2) {
3011 if (isFloatingPointZero(Op)) {
3012 RetVal1 = DAG.getConstant(0, MVT::i32);
3013 RetVal2 = DAG.getConstant(0, MVT::i32);
3017 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3018 SDValue Ptr = Ld->getBasePtr();
3019 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3020 Ld->getChain(), Ptr,
3021 Ld->getPointerInfo(),
3022 Ld->isVolatile(), Ld->isNonTemporal(),
3023 Ld->isInvariant(), Ld->getAlignment());
3025 EVT PtrType = Ptr.getValueType();
3026 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3027 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3028 PtrType, Ptr, DAG.getConstant(4, PtrType));
3029 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3030 Ld->getChain(), NewPtr,
3031 Ld->getPointerInfo().getWithOffset(4),
3032 Ld->isVolatile(), Ld->isNonTemporal(),
3033 Ld->isInvariant(), NewAlign);
3037 llvm_unreachable("Unknown VFP cmp argument!");
3040 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3041 /// f32 and even f64 comparisons to integer ones.
3043 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3044 SDValue Chain = Op.getOperand(0);
3045 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3046 SDValue LHS = Op.getOperand(2);
3047 SDValue RHS = Op.getOperand(3);
3048 SDValue Dest = Op.getOperand(4);
3049 DebugLoc dl = Op.getDebugLoc();
3051 bool LHSSeenZero = false;
3052 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3053 bool RHSSeenZero = false;
3054 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3055 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3056 // If unsafe fp math optimization is enabled and there are no other uses of
3057 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3058 // to an integer comparison.
3059 if (CC == ISD::SETOEQ)
3061 else if (CC == ISD::SETUNE)
3064 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3066 if (LHS.getValueType() == MVT::f32) {
3067 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3068 bitcastf32Toi32(LHS, DAG), Mask);
3069 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3070 bitcastf32Toi32(RHS, DAG), Mask);
3071 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3072 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3073 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3074 Chain, Dest, ARMcc, CCR, Cmp);
3079 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3080 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3081 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3082 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3083 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3084 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3085 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3086 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3087 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3093 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3094 SDValue Chain = Op.getOperand(0);
3095 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3096 SDValue LHS = Op.getOperand(2);
3097 SDValue RHS = Op.getOperand(3);
3098 SDValue Dest = Op.getOperand(4);
3099 DebugLoc dl = Op.getDebugLoc();
3101 if (LHS.getValueType() == MVT::i32) {
3103 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3104 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3105 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3106 Chain, Dest, ARMcc, CCR, Cmp);
3109 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3111 if (getTargetMachine().Options.UnsafeFPMath &&
3112 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3113 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3114 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3115 if (Result.getNode())
3119 ARMCC::CondCodes CondCode, CondCode2;
3120 FPCCToARMCC(CC, CondCode, CondCode2);
3122 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3123 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3124 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3125 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3126 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3127 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3128 if (CondCode2 != ARMCC::AL) {
3129 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3130 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3131 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3136 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3137 SDValue Chain = Op.getOperand(0);
3138 SDValue Table = Op.getOperand(1);
3139 SDValue Index = Op.getOperand(2);
3140 DebugLoc dl = Op.getDebugLoc();
3142 EVT PTy = getPointerTy();
3143 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3144 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3145 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3146 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3147 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3148 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3149 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3150 if (Subtarget->isThumb2()) {
3151 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3152 // which does another jump to the destination. This also makes it easier
3153 // to translate it to TBB / TBH later.
3154 // FIXME: This might not work if the function is extremely large.
3155 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3156 Addr, Op.getOperand(2), JTI, UId);
3158 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3159 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3160 MachinePointerInfo::getJumpTable(),
3161 false, false, false, 0);
3162 Chain = Addr.getValue(1);
3163 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3164 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3166 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3167 MachinePointerInfo::getJumpTable(),
3168 false, false, false, 0);
3169 Chain = Addr.getValue(1);
3170 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3174 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3175 EVT VT = Op.getValueType();
3176 DebugLoc dl = Op.getDebugLoc();
3178 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3179 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3181 return DAG.UnrollVectorOp(Op.getNode());
3184 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3185 "Invalid type for custom lowering!");
3186 if (VT != MVT::v4i16)
3187 return DAG.UnrollVectorOp(Op.getNode());
3189 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3190 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3193 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3194 EVT VT = Op.getValueType();
3196 return LowerVectorFP_TO_INT(Op, DAG);
3198 DebugLoc dl = Op.getDebugLoc();
3201 switch (Op.getOpcode()) {
3202 default: llvm_unreachable("Invalid opcode!");
3203 case ISD::FP_TO_SINT:
3204 Opc = ARMISD::FTOSI;
3206 case ISD::FP_TO_UINT:
3207 Opc = ARMISD::FTOUI;
3210 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3211 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3214 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3215 EVT VT = Op.getValueType();
3216 DebugLoc dl = Op.getDebugLoc();
3218 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3219 if (VT.getVectorElementType() == MVT::f32)
3221 return DAG.UnrollVectorOp(Op.getNode());
3224 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3225 "Invalid type for custom lowering!");
3226 if (VT != MVT::v4f32)
3227 return DAG.UnrollVectorOp(Op.getNode());
3231 switch (Op.getOpcode()) {
3232 default: llvm_unreachable("Invalid opcode!");
3233 case ISD::SINT_TO_FP:
3234 CastOpc = ISD::SIGN_EXTEND;
3235 Opc = ISD::SINT_TO_FP;
3237 case ISD::UINT_TO_FP:
3238 CastOpc = ISD::ZERO_EXTEND;
3239 Opc = ISD::UINT_TO_FP;
3243 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3244 return DAG.getNode(Opc, dl, VT, Op);
3247 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3248 EVT VT = Op.getValueType();
3250 return LowerVectorINT_TO_FP(Op, DAG);
3252 DebugLoc dl = Op.getDebugLoc();
3255 switch (Op.getOpcode()) {
3256 default: llvm_unreachable("Invalid opcode!");
3257 case ISD::SINT_TO_FP:
3258 Opc = ARMISD::SITOF;
3260 case ISD::UINT_TO_FP:
3261 Opc = ARMISD::UITOF;
3265 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3266 return DAG.getNode(Opc, dl, VT, Op);
3269 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3270 // Implement fcopysign with a fabs and a conditional fneg.
3271 SDValue Tmp0 = Op.getOperand(0);
3272 SDValue Tmp1 = Op.getOperand(1);
3273 DebugLoc dl = Op.getDebugLoc();
3274 EVT VT = Op.getValueType();
3275 EVT SrcVT = Tmp1.getValueType();
3276 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3277 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3278 bool UseNEON = !InGPR && Subtarget->hasNEON();
3281 // Use VBSL to copy the sign bit.
3282 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3283 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3284 DAG.getTargetConstant(EncodedVal, MVT::i32));
3285 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3287 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3288 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3289 DAG.getConstant(32, MVT::i32));
3290 else /*if (VT == MVT::f32)*/
3291 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3292 if (SrcVT == MVT::f32) {
3293 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3295 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3296 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3297 DAG.getConstant(32, MVT::i32));
3298 } else if (VT == MVT::f32)
3299 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3300 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3301 DAG.getConstant(32, MVT::i32));
3302 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3303 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3305 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3307 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3308 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3309 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3311 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3312 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3313 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3314 if (VT == MVT::f32) {
3315 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3316 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3317 DAG.getConstant(0, MVT::i32));
3319 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3325 // Bitcast operand 1 to i32.
3326 if (SrcVT == MVT::f64)
3327 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3328 &Tmp1, 1).getValue(1);
3329 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3331 // Or in the signbit with integer operations.
3332 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3333 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3334 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3335 if (VT == MVT::f32) {
3336 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3337 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3338 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3339 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3342 // f64: Or the high part with signbit and then combine two parts.
3343 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3345 SDValue Lo = Tmp0.getValue(0);
3346 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3347 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3348 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3351 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3352 MachineFunction &MF = DAG.getMachineFunction();
3353 MachineFrameInfo *MFI = MF.getFrameInfo();
3354 MFI->setReturnAddressIsTaken(true);
3356 EVT VT = Op.getValueType();
3357 DebugLoc dl = Op.getDebugLoc();
3358 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3360 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3361 SDValue Offset = DAG.getConstant(4, MVT::i32);
3362 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3363 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3364 MachinePointerInfo(), false, false, false, 0);
3367 // Return LR, which contains the return address. Mark it an implicit live-in.
3368 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3369 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3372 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3373 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3374 MFI->setFrameAddressIsTaken(true);
3376 EVT VT = Op.getValueType();
3377 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3378 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3379 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3380 ? ARM::R7 : ARM::R11;
3381 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3383 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3384 MachinePointerInfo(),
3385 false, false, false, 0);
3389 /// ExpandBITCAST - If the target supports VFP, this function is called to
3390 /// expand a bit convert where either the source or destination type is i64 to
3391 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3392 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3393 /// vectors), since the legalizer won't know what to do with that.
3394 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3395 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3396 DebugLoc dl = N->getDebugLoc();
3397 SDValue Op = N->getOperand(0);
3399 // This function is only supposed to be called for i64 types, either as the
3400 // source or destination of the bit convert.
3401 EVT SrcVT = Op.getValueType();
3402 EVT DstVT = N->getValueType(0);
3403 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3404 "ExpandBITCAST called for non-i64 type");
3406 // Turn i64->f64 into VMOVDRR.
3407 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3408 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3409 DAG.getConstant(0, MVT::i32));
3410 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3411 DAG.getConstant(1, MVT::i32));
3412 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3413 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3416 // Turn f64->i64 into VMOVRRD.
3417 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3418 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3419 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3420 // Merge the pieces into a single i64 value.
3421 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3427 /// getZeroVector - Returns a vector of specified type with all zero elements.
3428 /// Zero vectors are used to represent vector negation and in those cases
3429 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3430 /// not support i64 elements, so sometimes the zero vectors will need to be
3431 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3433 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3434 assert(VT.isVector() && "Expected a vector type");
3435 // The canonical modified immediate encoding of a zero vector is....0!
3436 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3437 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3438 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3439 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3442 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3443 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3444 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3445 SelectionDAG &DAG) const {
3446 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3447 EVT VT = Op.getValueType();
3448 unsigned VTBits = VT.getSizeInBits();
3449 DebugLoc dl = Op.getDebugLoc();
3450 SDValue ShOpLo = Op.getOperand(0);
3451 SDValue ShOpHi = Op.getOperand(1);
3452 SDValue ShAmt = Op.getOperand(2);
3454 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3456 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3458 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3459 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3460 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3461 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3462 DAG.getConstant(VTBits, MVT::i32));
3463 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3464 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3465 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3467 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3468 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3470 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3471 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3474 SDValue Ops[2] = { Lo, Hi };
3475 return DAG.getMergeValues(Ops, 2, dl);
3478 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3479 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3480 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3481 SelectionDAG &DAG) const {
3482 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3483 EVT VT = Op.getValueType();
3484 unsigned VTBits = VT.getSizeInBits();
3485 DebugLoc dl = Op.getDebugLoc();
3486 SDValue ShOpLo = Op.getOperand(0);
3487 SDValue ShOpHi = Op.getOperand(1);
3488 SDValue ShAmt = Op.getOperand(2);
3491 assert(Op.getOpcode() == ISD::SHL_PARTS);
3492 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3493 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3494 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3495 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3496 DAG.getConstant(VTBits, MVT::i32));
3497 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3498 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3500 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3501 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3502 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3504 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3505 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3508 SDValue Ops[2] = { Lo, Hi };
3509 return DAG.getMergeValues(Ops, 2, dl);
3512 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3513 SelectionDAG &DAG) const {
3514 // The rounding mode is in bits 23:22 of the FPSCR.
3515 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3516 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3517 // so that the shift + and get folded into a bitfield extract.
3518 DebugLoc dl = Op.getDebugLoc();
3519 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3520 DAG.getConstant(Intrinsic::arm_get_fpscr,
3522 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3523 DAG.getConstant(1U << 22, MVT::i32));
3524 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3525 DAG.getConstant(22, MVT::i32));
3526 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3527 DAG.getConstant(3, MVT::i32));
3530 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3531 const ARMSubtarget *ST) {
3532 EVT VT = N->getValueType(0);
3533 DebugLoc dl = N->getDebugLoc();
3535 if (!ST->hasV6T2Ops())
3538 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3539 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3542 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3543 const ARMSubtarget *ST) {
3544 EVT VT = N->getValueType(0);
3545 DebugLoc dl = N->getDebugLoc();
3550 // Lower vector shifts on NEON to use VSHL.
3551 assert(ST->hasNEON() && "unexpected vector shift");
3553 // Left shifts translate directly to the vshiftu intrinsic.
3554 if (N->getOpcode() == ISD::SHL)
3555 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3556 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3557 N->getOperand(0), N->getOperand(1));
3559 assert((N->getOpcode() == ISD::SRA ||
3560 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3562 // NEON uses the same intrinsics for both left and right shifts. For
3563 // right shifts, the shift amounts are negative, so negate the vector of
3565 EVT ShiftVT = N->getOperand(1).getValueType();
3566 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3567 getZeroVector(ShiftVT, DAG, dl),
3569 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3570 Intrinsic::arm_neon_vshifts :
3571 Intrinsic::arm_neon_vshiftu);
3572 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3573 DAG.getConstant(vshiftInt, MVT::i32),
3574 N->getOperand(0), NegatedCount);
3577 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3578 const ARMSubtarget *ST) {
3579 EVT VT = N->getValueType(0);
3580 DebugLoc dl = N->getDebugLoc();
3582 // We can get here for a node like i32 = ISD::SHL i32, i64
3586 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3587 "Unknown shift to lower!");
3589 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3590 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3591 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3594 // If we are in thumb mode, we don't have RRX.
3595 if (ST->isThumb1Only()) return SDValue();
3597 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3598 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3599 DAG.getConstant(0, MVT::i32));
3600 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3601 DAG.getConstant(1, MVT::i32));
3603 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3604 // captures the result into a carry flag.
3605 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3606 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3608 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3609 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3611 // Merge the pieces into a single i64 value.
3612 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3615 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3616 SDValue TmpOp0, TmpOp1;
3617 bool Invert = false;
3621 SDValue Op0 = Op.getOperand(0);
3622 SDValue Op1 = Op.getOperand(1);
3623 SDValue CC = Op.getOperand(2);
3624 EVT VT = Op.getValueType();
3625 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3626 DebugLoc dl = Op.getDebugLoc();
3628 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3629 switch (SetCCOpcode) {
3630 default: llvm_unreachable("Illegal FP comparison");
3632 case ISD::SETNE: Invert = true; // Fallthrough
3634 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3636 case ISD::SETLT: Swap = true; // Fallthrough
3638 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3640 case ISD::SETLE: Swap = true; // Fallthrough
3642 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3643 case ISD::SETUGE: Swap = true; // Fallthrough
3644 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3645 case ISD::SETUGT: Swap = true; // Fallthrough
3646 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3647 case ISD::SETUEQ: Invert = true; // Fallthrough
3649 // Expand this to (OLT | OGT).
3653 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3654 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3656 case ISD::SETUO: Invert = true; // Fallthrough
3658 // Expand this to (OLT | OGE).
3662 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3663 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3667 // Integer comparisons.
3668 switch (SetCCOpcode) {
3669 default: llvm_unreachable("Illegal integer comparison");
3670 case ISD::SETNE: Invert = true;
3671 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3672 case ISD::SETLT: Swap = true;
3673 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3674 case ISD::SETLE: Swap = true;
3675 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3676 case ISD::SETULT: Swap = true;
3677 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3678 case ISD::SETULE: Swap = true;
3679 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3682 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3683 if (Opc == ARMISD::VCEQ) {
3686 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3688 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3691 // Ignore bitconvert.
3692 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3693 AndOp = AndOp.getOperand(0);
3695 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3697 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3698 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3705 std::swap(Op0, Op1);
3707 // If one of the operands is a constant vector zero, attempt to fold the
3708 // comparison to a specialized compare-against-zero form.
3710 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3712 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3713 if (Opc == ARMISD::VCGE)
3714 Opc = ARMISD::VCLEZ;
3715 else if (Opc == ARMISD::VCGT)
3716 Opc = ARMISD::VCLTZ;
3721 if (SingleOp.getNode()) {
3724 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3726 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3728 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3730 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3732 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3734 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3737 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3741 Result = DAG.getNOT(dl, Result, VT);
3746 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3747 /// valid vector constant for a NEON instruction with a "modified immediate"
3748 /// operand (e.g., VMOV). If so, return the encoded value.
3749 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3750 unsigned SplatBitSize, SelectionDAG &DAG,
3751 EVT &VT, bool is128Bits, NEONModImmType type) {
3752 unsigned OpCmode, Imm;
3754 // SplatBitSize is set to the smallest size that splats the vector, so a
3755 // zero vector will always have SplatBitSize == 8. However, NEON modified
3756 // immediate instructions others than VMOV do not support the 8-bit encoding
3757 // of a zero vector, and the default encoding of zero is supposed to be the
3762 switch (SplatBitSize) {
3764 if (type != VMOVModImm)
3766 // Any 1-byte value is OK. Op=0, Cmode=1110.
3767 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3770 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3774 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3775 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3776 if ((SplatBits & ~0xff) == 0) {
3777 // Value = 0x00nn: Op=x, Cmode=100x.
3782 if ((SplatBits & ~0xff00) == 0) {
3783 // Value = 0xnn00: Op=x, Cmode=101x.
3785 Imm = SplatBits >> 8;
3791 // NEON's 32-bit VMOV supports splat values where:
3792 // * only one byte is nonzero, or
3793 // * the least significant byte is 0xff and the second byte is nonzero, or
3794 // * the least significant 2 bytes are 0xff and the third is nonzero.
3795 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3796 if ((SplatBits & ~0xff) == 0) {
3797 // Value = 0x000000nn: Op=x, Cmode=000x.
3802 if ((SplatBits & ~0xff00) == 0) {
3803 // Value = 0x0000nn00: Op=x, Cmode=001x.
3805 Imm = SplatBits >> 8;
3808 if ((SplatBits & ~0xff0000) == 0) {
3809 // Value = 0x00nn0000: Op=x, Cmode=010x.
3811 Imm = SplatBits >> 16;
3814 if ((SplatBits & ~0xff000000) == 0) {
3815 // Value = 0xnn000000: Op=x, Cmode=011x.
3817 Imm = SplatBits >> 24;
3821 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3822 if (type == OtherModImm) return SDValue();
3824 if ((SplatBits & ~0xffff) == 0 &&
3825 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3826 // Value = 0x0000nnff: Op=x, Cmode=1100.
3828 Imm = SplatBits >> 8;
3833 if ((SplatBits & ~0xffffff) == 0 &&
3834 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3835 // Value = 0x00nnffff: Op=x, Cmode=1101.
3837 Imm = SplatBits >> 16;
3838 SplatBits |= 0xffff;
3842 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3843 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3844 // VMOV.I32. A (very) minor optimization would be to replicate the value
3845 // and fall through here to test for a valid 64-bit splat. But, then the
3846 // caller would also need to check and handle the change in size.
3850 if (type != VMOVModImm)
3852 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3853 uint64_t BitMask = 0xff;
3855 unsigned ImmMask = 1;
3857 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3858 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3861 } else if ((SplatBits & BitMask) != 0) {
3867 // Op=1, Cmode=1110.
3870 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3875 llvm_unreachable("unexpected size for isNEONModifiedImm");
3878 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3879 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3882 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3883 const ARMSubtarget *ST) const {
3884 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3887 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3888 assert(Op.getValueType() == MVT::f32 &&
3889 "ConstantFP custom lowering should only occur for f32.");
3891 // Try splatting with a VMOV.f32...
3892 APFloat FPVal = CFP->getValueAPF();
3893 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3895 DebugLoc DL = Op.getDebugLoc();
3896 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3897 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3899 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3900 DAG.getConstant(0, MVT::i32));
3903 // If that fails, try a VMOV.i32
3905 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3906 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3908 if (NewVal != SDValue()) {
3909 DebugLoc DL = Op.getDebugLoc();
3910 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3912 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3914 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3915 DAG.getConstant(0, MVT::i32));
3918 // Finally, try a VMVN.i32
3919 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3921 if (NewVal != SDValue()) {
3922 DebugLoc DL = Op.getDebugLoc();
3923 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3924 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3926 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3927 DAG.getConstant(0, MVT::i32));
3933 // check if an VEXT instruction can handle the shuffle mask when the
3934 // vector sources of the shuffle are the same.
3935 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3936 unsigned NumElts = VT.getVectorNumElements();
3938 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3944 // If this is a VEXT shuffle, the immediate value is the index of the first
3945 // element. The other shuffle indices must be the successive elements after
3947 unsigned ExpectedElt = Imm;
3948 for (unsigned i = 1; i < NumElts; ++i) {
3949 // Increment the expected index. If it wraps around, just follow it
3950 // back to index zero and keep going.
3952 if (ExpectedElt == NumElts)
3955 if (M[i] < 0) continue; // ignore UNDEF indices
3956 if (ExpectedElt != static_cast<unsigned>(M[i]))
3964 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3965 bool &ReverseVEXT, unsigned &Imm) {
3966 unsigned NumElts = VT.getVectorNumElements();
3967 ReverseVEXT = false;
3969 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3975 // If this is a VEXT shuffle, the immediate value is the index of the first
3976 // element. The other shuffle indices must be the successive elements after
3978 unsigned ExpectedElt = Imm;
3979 for (unsigned i = 1; i < NumElts; ++i) {
3980 // Increment the expected index. If it wraps around, it may still be
3981 // a VEXT but the source vectors must be swapped.
3983 if (ExpectedElt == NumElts * 2) {
3988 if (M[i] < 0) continue; // ignore UNDEF indices
3989 if (ExpectedElt != static_cast<unsigned>(M[i]))
3993 // Adjust the index value if the source operands will be swapped.
4000 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4001 /// instruction with the specified blocksize. (The order of the elements
4002 /// within each block of the vector is reversed.)
4003 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4004 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4005 "Only possible block sizes for VREV are: 16, 32, 64");
4007 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4011 unsigned NumElts = VT.getVectorNumElements();
4012 unsigned BlockElts = M[0] + 1;
4013 // If the first shuffle index is UNDEF, be optimistic.
4015 BlockElts = BlockSize / EltSz;
4017 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4020 for (unsigned i = 0; i < NumElts; ++i) {
4021 if (M[i] < 0) continue; // ignore UNDEF indices
4022 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4029 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4030 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4031 // range, then 0 is placed into the resulting vector. So pretty much any mask
4032 // of 8 elements can work here.
4033 return VT == MVT::v8i8 && M.size() == 8;
4036 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4037 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4041 unsigned NumElts = VT.getVectorNumElements();
4042 WhichResult = (M[0] == 0 ? 0 : 1);
4043 for (unsigned i = 0; i < NumElts; i += 2) {
4044 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4045 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4051 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4052 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4053 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4054 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4055 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4059 unsigned NumElts = VT.getVectorNumElements();
4060 WhichResult = (M[0] == 0 ? 0 : 1);
4061 for (unsigned i = 0; i < NumElts; i += 2) {
4062 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4063 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4069 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4070 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4074 unsigned NumElts = VT.getVectorNumElements();
4075 WhichResult = (M[0] == 0 ? 0 : 1);
4076 for (unsigned i = 0; i != NumElts; ++i) {
4077 if (M[i] < 0) continue; // ignore UNDEF indices
4078 if ((unsigned) M[i] != 2 * i + WhichResult)
4082 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4083 if (VT.is64BitVector() && EltSz == 32)
4089 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4090 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4091 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4092 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4093 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4097 unsigned Half = VT.getVectorNumElements() / 2;
4098 WhichResult = (M[0] == 0 ? 0 : 1);
4099 for (unsigned j = 0; j != 2; ++j) {
4100 unsigned Idx = WhichResult;
4101 for (unsigned i = 0; i != Half; ++i) {
4102 int MIdx = M[i + j * Half];
4103 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4109 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4110 if (VT.is64BitVector() && EltSz == 32)
4116 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4117 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4121 unsigned NumElts = VT.getVectorNumElements();
4122 WhichResult = (M[0] == 0 ? 0 : 1);
4123 unsigned Idx = WhichResult * NumElts / 2;
4124 for (unsigned i = 0; i != NumElts; i += 2) {
4125 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4126 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4131 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4132 if (VT.is64BitVector() && EltSz == 32)
4138 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4139 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4140 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4141 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4142 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4146 unsigned NumElts = VT.getVectorNumElements();
4147 WhichResult = (M[0] == 0 ? 0 : 1);
4148 unsigned Idx = WhichResult * NumElts / 2;
4149 for (unsigned i = 0; i != NumElts; i += 2) {
4150 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4151 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4156 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4157 if (VT.is64BitVector() && EltSz == 32)
4163 // If N is an integer constant that can be moved into a register in one
4164 // instruction, return an SDValue of such a constant (will become a MOV
4165 // instruction). Otherwise return null.
4166 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4167 const ARMSubtarget *ST, DebugLoc dl) {
4169 if (!isa<ConstantSDNode>(N))
4171 Val = cast<ConstantSDNode>(N)->getZExtValue();
4173 if (ST->isThumb1Only()) {
4174 if (Val <= 255 || ~Val <= 255)
4175 return DAG.getConstant(Val, MVT::i32);
4177 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4178 return DAG.getConstant(Val, MVT::i32);
4183 // If this is a case we can't handle, return null and let the default
4184 // expansion code take care of it.
4185 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4186 const ARMSubtarget *ST) const {
4187 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4188 DebugLoc dl = Op.getDebugLoc();
4189 EVT VT = Op.getValueType();
4191 APInt SplatBits, SplatUndef;
4192 unsigned SplatBitSize;
4194 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4195 if (SplatBitSize <= 64) {
4196 // Check if an immediate VMOV works.
4198 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4199 SplatUndef.getZExtValue(), SplatBitSize,
4200 DAG, VmovVT, VT.is128BitVector(),
4202 if (Val.getNode()) {
4203 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4204 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4207 // Try an immediate VMVN.
4208 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4209 Val = isNEONModifiedImm(NegatedImm,
4210 SplatUndef.getZExtValue(), SplatBitSize,
4211 DAG, VmovVT, VT.is128BitVector(),
4213 if (Val.getNode()) {
4214 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4215 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4218 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4219 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4220 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4222 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4223 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4229 // Scan through the operands to see if only one value is used.
4231 // As an optimisation, even if more than one value is used it may be more
4232 // profitable to splat with one value then change some lanes.
4234 // Heuristically we decide to do this if the vector has a "dominant" value,
4235 // defined as splatted to more than half of the lanes.
4236 unsigned NumElts = VT.getVectorNumElements();
4237 bool isOnlyLowElement = true;
4238 bool usesOnlyOneValue = true;
4239 bool hasDominantValue = false;
4240 bool isConstant = true;
4242 // Map of the number of times a particular SDValue appears in the
4244 DenseMap<SDValue, unsigned> ValueCounts;
4246 for (unsigned i = 0; i < NumElts; ++i) {
4247 SDValue V = Op.getOperand(i);
4248 if (V.getOpcode() == ISD::UNDEF)
4251 isOnlyLowElement = false;
4252 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4255 ValueCounts.insert(std::make_pair(V, 0));
4256 unsigned &Count = ValueCounts[V];
4258 // Is this value dominant? (takes up more than half of the lanes)
4259 if (++Count > (NumElts / 2)) {
4260 hasDominantValue = true;
4264 if (ValueCounts.size() != 1)
4265 usesOnlyOneValue = false;
4266 if (!Value.getNode() && ValueCounts.size() > 0)
4267 Value = ValueCounts.begin()->first;
4269 if (ValueCounts.size() == 0)
4270 return DAG.getUNDEF(VT);
4272 if (isOnlyLowElement)
4273 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4275 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4277 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4278 // i32 and try again.
4279 if (hasDominantValue && EltSize <= 32) {
4283 // If we are VDUPing a value that comes directly from a vector, that will
4284 // cause an unnecessary move to and from a GPR, where instead we could
4285 // just use VDUPLANE.
4286 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4287 // We need to create a new undef vector to use for the VDUPLANE if the
4288 // size of the vector from which we get the value is different than the
4289 // size of the vector that we need to create. We will insert the element
4290 // such that the register coalescer will remove unnecessary copies.
4291 if (VT != Value->getOperand(0).getValueType()) {
4292 ConstantSDNode *constIndex;
4293 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4294 assert(constIndex && "The index is not a constant!");
4295 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4296 VT.getVectorNumElements();
4297 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4298 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4299 Value, DAG.getConstant(index, MVT::i32)),
4300 DAG.getConstant(index, MVT::i32));
4302 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4303 Value->getOperand(0), Value->getOperand(1));
4307 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4309 if (!usesOnlyOneValue) {
4310 // The dominant value was splatted as 'N', but we now have to insert
4311 // all differing elements.
4312 for (unsigned I = 0; I < NumElts; ++I) {
4313 if (Op.getOperand(I) == Value)
4315 SmallVector<SDValue, 3> Ops;
4317 Ops.push_back(Op.getOperand(I));
4318 Ops.push_back(DAG.getConstant(I, MVT::i32));
4319 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4324 if (VT.getVectorElementType().isFloatingPoint()) {
4325 SmallVector<SDValue, 8> Ops;
4326 for (unsigned i = 0; i < NumElts; ++i)
4327 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4329 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4330 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4331 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4333 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4335 if (usesOnlyOneValue) {
4336 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4337 if (isConstant && Val.getNode())
4338 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4342 // If all elements are constants and the case above didn't get hit, fall back
4343 // to the default expansion, which will generate a load from the constant
4348 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4350 SDValue shuffle = ReconstructShuffle(Op, DAG);
4351 if (shuffle != SDValue())
4355 // Vectors with 32- or 64-bit elements can be built by directly assigning
4356 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4357 // will be legalized.
4358 if (EltSize >= 32) {
4359 // Do the expansion with floating-point types, since that is what the VFP
4360 // registers are defined to use, and since i64 is not legal.
4361 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4362 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4363 SmallVector<SDValue, 8> Ops;
4364 for (unsigned i = 0; i < NumElts; ++i)
4365 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4366 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4367 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4373 // Gather data to see if the operation can be modelled as a
4374 // shuffle in combination with VEXTs.
4375 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4376 SelectionDAG &DAG) const {
4377 DebugLoc dl = Op.getDebugLoc();
4378 EVT VT = Op.getValueType();
4379 unsigned NumElts = VT.getVectorNumElements();
4381 SmallVector<SDValue, 2> SourceVecs;
4382 SmallVector<unsigned, 2> MinElts;
4383 SmallVector<unsigned, 2> MaxElts;
4385 for (unsigned i = 0; i < NumElts; ++i) {
4386 SDValue V = Op.getOperand(i);
4387 if (V.getOpcode() == ISD::UNDEF)
4389 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4390 // A shuffle can only come from building a vector from various
4391 // elements of other vectors.
4393 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4394 VT.getVectorElementType()) {
4395 // This code doesn't know how to handle shuffles where the vector
4396 // element types do not match (this happens because type legalization
4397 // promotes the return type of EXTRACT_VECTOR_ELT).
4398 // FIXME: It might be appropriate to extend this code to handle
4399 // mismatched types.
4403 // Record this extraction against the appropriate vector if possible...
4404 SDValue SourceVec = V.getOperand(0);
4405 // If the element number isn't a constant, we can't effectively
4406 // analyze what's going on.
4407 if (!isa<ConstantSDNode>(V.getOperand(1)))
4409 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4410 bool FoundSource = false;
4411 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4412 if (SourceVecs[j] == SourceVec) {
4413 if (MinElts[j] > EltNo)
4415 if (MaxElts[j] < EltNo)
4422 // Or record a new source if not...
4424 SourceVecs.push_back(SourceVec);
4425 MinElts.push_back(EltNo);
4426 MaxElts.push_back(EltNo);
4430 // Currently only do something sane when at most two source vectors
4432 if (SourceVecs.size() > 2)
4435 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4436 int VEXTOffsets[2] = {0, 0};
4438 // This loop extracts the usage patterns of the source vectors
4439 // and prepares appropriate SDValues for a shuffle if possible.
4440 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4441 if (SourceVecs[i].getValueType() == VT) {
4442 // No VEXT necessary
4443 ShuffleSrcs[i] = SourceVecs[i];
4446 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4447 // It probably isn't worth padding out a smaller vector just to
4448 // break it down again in a shuffle.
4452 // Since only 64-bit and 128-bit vectors are legal on ARM and
4453 // we've eliminated the other cases...
4454 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4455 "unexpected vector sizes in ReconstructShuffle");
4457 if (MaxElts[i] - MinElts[i] >= NumElts) {
4458 // Span too large for a VEXT to cope
4462 if (MinElts[i] >= NumElts) {
4463 // The extraction can just take the second half
4464 VEXTOffsets[i] = NumElts;
4465 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4467 DAG.getIntPtrConstant(NumElts));
4468 } else if (MaxElts[i] < NumElts) {
4469 // The extraction can just take the first half
4471 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4473 DAG.getIntPtrConstant(0));
4475 // An actual VEXT is needed
4476 VEXTOffsets[i] = MinElts[i];
4477 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4479 DAG.getIntPtrConstant(0));
4480 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4482 DAG.getIntPtrConstant(NumElts));
4483 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4484 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4488 SmallVector<int, 8> Mask;
4490 for (unsigned i = 0; i < NumElts; ++i) {
4491 SDValue Entry = Op.getOperand(i);
4492 if (Entry.getOpcode() == ISD::UNDEF) {
4497 SDValue ExtractVec = Entry.getOperand(0);
4498 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4499 .getOperand(1))->getSExtValue();
4500 if (ExtractVec == SourceVecs[0]) {
4501 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4503 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4507 // Final check before we try to produce nonsense...
4508 if (isShuffleMaskLegal(Mask, VT))
4509 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4515 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4516 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4517 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4518 /// are assumed to be legal.
4520 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4522 if (VT.getVectorNumElements() == 4 &&
4523 (VT.is128BitVector() || VT.is64BitVector())) {
4524 unsigned PFIndexes[4];
4525 for (unsigned i = 0; i != 4; ++i) {
4529 PFIndexes[i] = M[i];
4532 // Compute the index in the perfect shuffle table.
4533 unsigned PFTableIndex =
4534 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4535 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4536 unsigned Cost = (PFEntry >> 30);
4543 unsigned Imm, WhichResult;
4545 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4546 return (EltSize >= 32 ||
4547 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4548 isVREVMask(M, VT, 64) ||
4549 isVREVMask(M, VT, 32) ||
4550 isVREVMask(M, VT, 16) ||
4551 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4552 isVTBLMask(M, VT) ||
4553 isVTRNMask(M, VT, WhichResult) ||
4554 isVUZPMask(M, VT, WhichResult) ||
4555 isVZIPMask(M, VT, WhichResult) ||
4556 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4557 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4558 isVZIP_v_undef_Mask(M, VT, WhichResult));
4561 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4562 /// the specified operations to build the shuffle.
4563 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4564 SDValue RHS, SelectionDAG &DAG,
4566 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4567 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4568 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4571 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4580 OP_VUZPL, // VUZP, left result
4581 OP_VUZPR, // VUZP, right result
4582 OP_VZIPL, // VZIP, left result
4583 OP_VZIPR, // VZIP, right result
4584 OP_VTRNL, // VTRN, left result
4585 OP_VTRNR // VTRN, right result
4588 if (OpNum == OP_COPY) {
4589 if (LHSID == (1*9+2)*9+3) return LHS;
4590 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4594 SDValue OpLHS, OpRHS;
4595 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4596 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4597 EVT VT = OpLHS.getValueType();
4600 default: llvm_unreachable("Unknown shuffle opcode!");
4602 // VREV divides the vector in half and swaps within the half.
4603 if (VT.getVectorElementType() == MVT::i32 ||
4604 VT.getVectorElementType() == MVT::f32)
4605 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4606 // vrev <4 x i16> -> VREV32
4607 if (VT.getVectorElementType() == MVT::i16)
4608 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4609 // vrev <4 x i8> -> VREV16
4610 assert(VT.getVectorElementType() == MVT::i8);
4611 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4616 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4617 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4621 return DAG.getNode(ARMISD::VEXT, dl, VT,
4623 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4626 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4627 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4630 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4631 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4634 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4635 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4639 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4640 ArrayRef<int> ShuffleMask,
4641 SelectionDAG &DAG) {
4642 // Check to see if we can use the VTBL instruction.
4643 SDValue V1 = Op.getOperand(0);
4644 SDValue V2 = Op.getOperand(1);
4645 DebugLoc DL = Op.getDebugLoc();
4647 SmallVector<SDValue, 8> VTBLMask;
4648 for (ArrayRef<int>::iterator
4649 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4650 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4652 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4653 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4654 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4657 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4658 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4662 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4663 SDValue V1 = Op.getOperand(0);
4664 SDValue V2 = Op.getOperand(1);
4665 DebugLoc dl = Op.getDebugLoc();
4666 EVT VT = Op.getValueType();
4667 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4669 // Convert shuffles that are directly supported on NEON to target-specific
4670 // DAG nodes, instead of keeping them as shuffles and matching them again
4671 // during code selection. This is more efficient and avoids the possibility
4672 // of inconsistencies between legalization and selection.
4673 // FIXME: floating-point vectors should be canonicalized to integer vectors
4674 // of the same time so that they get CSEd properly.
4675 ArrayRef<int> ShuffleMask = SVN->getMask();
4677 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4678 if (EltSize <= 32) {
4679 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4680 int Lane = SVN->getSplatIndex();
4681 // If this is undef splat, generate it via "just" vdup, if possible.
4682 if (Lane == -1) Lane = 0;
4684 // Test if V1 is a SCALAR_TO_VECTOR.
4685 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4686 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4688 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4689 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4691 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4692 !isa<ConstantSDNode>(V1.getOperand(0))) {
4693 bool IsScalarToVector = true;
4694 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4695 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4696 IsScalarToVector = false;
4699 if (IsScalarToVector)
4700 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4702 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4703 DAG.getConstant(Lane, MVT::i32));
4708 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4711 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4712 DAG.getConstant(Imm, MVT::i32));
4715 if (isVREVMask(ShuffleMask, VT, 64))
4716 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4717 if (isVREVMask(ShuffleMask, VT, 32))
4718 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4719 if (isVREVMask(ShuffleMask, VT, 16))
4720 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4722 if (V2->getOpcode() == ISD::UNDEF &&
4723 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4724 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4725 DAG.getConstant(Imm, MVT::i32));
4728 // Check for Neon shuffles that modify both input vectors in place.
4729 // If both results are used, i.e., if there are two shuffles with the same
4730 // source operands and with masks corresponding to both results of one of
4731 // these operations, DAG memoization will ensure that a single node is
4732 // used for both shuffles.
4733 unsigned WhichResult;
4734 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4735 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4736 V1, V2).getValue(WhichResult);
4737 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4738 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4739 V1, V2).getValue(WhichResult);
4740 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4741 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4742 V1, V2).getValue(WhichResult);
4744 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4745 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4746 V1, V1).getValue(WhichResult);
4747 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4748 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4749 V1, V1).getValue(WhichResult);
4750 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4751 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4752 V1, V1).getValue(WhichResult);
4755 // If the shuffle is not directly supported and it has 4 elements, use
4756 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4757 unsigned NumElts = VT.getVectorNumElements();
4759 unsigned PFIndexes[4];
4760 for (unsigned i = 0; i != 4; ++i) {
4761 if (ShuffleMask[i] < 0)
4764 PFIndexes[i] = ShuffleMask[i];
4767 // Compute the index in the perfect shuffle table.
4768 unsigned PFTableIndex =
4769 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4770 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4771 unsigned Cost = (PFEntry >> 30);
4774 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4777 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4778 if (EltSize >= 32) {
4779 // Do the expansion with floating-point types, since that is what the VFP
4780 // registers are defined to use, and since i64 is not legal.
4781 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4782 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4783 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4784 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4785 SmallVector<SDValue, 8> Ops;
4786 for (unsigned i = 0; i < NumElts; ++i) {
4787 if (ShuffleMask[i] < 0)
4788 Ops.push_back(DAG.getUNDEF(EltVT));
4790 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4791 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4792 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4795 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4796 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4799 if (VT == MVT::v8i8) {
4800 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4801 if (NewOp.getNode())
4808 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4809 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4810 SDValue Lane = Op.getOperand(2);
4811 if (!isa<ConstantSDNode>(Lane))
4817 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4818 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4819 SDValue Lane = Op.getOperand(1);
4820 if (!isa<ConstantSDNode>(Lane))
4823 SDValue Vec = Op.getOperand(0);
4824 if (Op.getValueType() == MVT::i32 &&
4825 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4826 DebugLoc dl = Op.getDebugLoc();
4827 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4833 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4834 // The only time a CONCAT_VECTORS operation can have legal types is when
4835 // two 64-bit vectors are concatenated to a 128-bit vector.
4836 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4837 "unexpected CONCAT_VECTORS");
4838 DebugLoc dl = Op.getDebugLoc();
4839 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4840 SDValue Op0 = Op.getOperand(0);
4841 SDValue Op1 = Op.getOperand(1);
4842 if (Op0.getOpcode() != ISD::UNDEF)
4843 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4844 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4845 DAG.getIntPtrConstant(0));
4846 if (Op1.getOpcode() != ISD::UNDEF)
4847 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4848 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4849 DAG.getIntPtrConstant(1));
4850 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4853 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4854 /// element has been zero/sign-extended, depending on the isSigned parameter,
4855 /// from an integer type half its size.
4856 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4858 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4859 EVT VT = N->getValueType(0);
4860 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4861 SDNode *BVN = N->getOperand(0).getNode();
4862 if (BVN->getValueType(0) != MVT::v4i32 ||
4863 BVN->getOpcode() != ISD::BUILD_VECTOR)
4865 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4866 unsigned HiElt = 1 - LoElt;
4867 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4868 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4869 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4870 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4871 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4874 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4875 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4878 if (Hi0->isNullValue() && Hi1->isNullValue())
4884 if (N->getOpcode() != ISD::BUILD_VECTOR)
4887 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4888 SDNode *Elt = N->getOperand(i).getNode();
4889 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4890 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4891 unsigned HalfSize = EltSize / 2;
4893 if (!isIntN(HalfSize, C->getSExtValue()))
4896 if (!isUIntN(HalfSize, C->getZExtValue()))
4907 /// isSignExtended - Check if a node is a vector value that is sign-extended
4908 /// or a constant BUILD_VECTOR with sign-extended elements.
4909 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4910 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4912 if (isExtendedBUILD_VECTOR(N, DAG, true))
4917 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4918 /// or a constant BUILD_VECTOR with zero-extended elements.
4919 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4920 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4922 if (isExtendedBUILD_VECTOR(N, DAG, false))
4927 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4928 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4929 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4930 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4931 return N->getOperand(0);
4932 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4933 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4934 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4935 LD->isNonTemporal(), LD->isInvariant(),
4936 LD->getAlignment());
4937 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4938 // have been legalized as a BITCAST from v4i32.
4939 if (N->getOpcode() == ISD::BITCAST) {
4940 SDNode *BVN = N->getOperand(0).getNode();
4941 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4942 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4943 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4944 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4945 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4947 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4948 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4949 EVT VT = N->getValueType(0);
4950 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4951 unsigned NumElts = VT.getVectorNumElements();
4952 MVT TruncVT = MVT::getIntegerVT(EltSize);
4953 SmallVector<SDValue, 8> Ops;
4954 for (unsigned i = 0; i != NumElts; ++i) {
4955 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4956 const APInt &CInt = C->getAPIntValue();
4957 // Element types smaller than 32 bits are not legal, so use i32 elements.
4958 // The values are implicitly truncated so sext vs. zext doesn't matter.
4959 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4961 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4962 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4965 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4966 unsigned Opcode = N->getOpcode();
4967 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4968 SDNode *N0 = N->getOperand(0).getNode();
4969 SDNode *N1 = N->getOperand(1).getNode();
4970 return N0->hasOneUse() && N1->hasOneUse() &&
4971 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4976 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4977 unsigned Opcode = N->getOpcode();
4978 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4979 SDNode *N0 = N->getOperand(0).getNode();
4980 SDNode *N1 = N->getOperand(1).getNode();
4981 return N0->hasOneUse() && N1->hasOneUse() &&
4982 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4987 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4988 // Multiplications are only custom-lowered for 128-bit vectors so that
4989 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4990 EVT VT = Op.getValueType();
4991 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4992 SDNode *N0 = Op.getOperand(0).getNode();
4993 SDNode *N1 = Op.getOperand(1).getNode();
4994 unsigned NewOpc = 0;
4996 bool isN0SExt = isSignExtended(N0, DAG);
4997 bool isN1SExt = isSignExtended(N1, DAG);
4998 if (isN0SExt && isN1SExt)
4999 NewOpc = ARMISD::VMULLs;
5001 bool isN0ZExt = isZeroExtended(N0, DAG);
5002 bool isN1ZExt = isZeroExtended(N1, DAG);
5003 if (isN0ZExt && isN1ZExt)
5004 NewOpc = ARMISD::VMULLu;
5005 else if (isN1SExt || isN1ZExt) {
5006 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5007 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5008 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5009 NewOpc = ARMISD::VMULLs;
5011 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5012 NewOpc = ARMISD::VMULLu;
5014 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5016 NewOpc = ARMISD::VMULLu;
5022 if (VT == MVT::v2i64)
5023 // Fall through to expand this. It is not legal.
5026 // Other vector multiplications are legal.
5031 // Legalize to a VMULL instruction.
5032 DebugLoc DL = Op.getDebugLoc();
5034 SDValue Op1 = SkipExtension(N1, DAG);
5036 Op0 = SkipExtension(N0, DAG);
5037 assert(Op0.getValueType().is64BitVector() &&
5038 Op1.getValueType().is64BitVector() &&
5039 "unexpected types for extended operands to VMULL");
5040 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5043 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5044 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5051 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5052 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5053 EVT Op1VT = Op1.getValueType();
5054 return DAG.getNode(N0->getOpcode(), DL, VT,
5055 DAG.getNode(NewOpc, DL, VT,
5056 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5057 DAG.getNode(NewOpc, DL, VT,
5058 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5062 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5064 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5065 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5066 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5067 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5068 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5069 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5070 // Get reciprocal estimate.
5071 // float4 recip = vrecpeq_f32(yf);
5072 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5073 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5074 // Because char has a smaller range than uchar, we can actually get away
5075 // without any newton steps. This requires that we use a weird bias
5076 // of 0xb000, however (again, this has been exhaustively tested).
5077 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5078 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5079 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5080 Y = DAG.getConstant(0xb000, MVT::i32);
5081 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5082 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5083 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5084 // Convert back to short.
5085 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5086 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5091 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5093 // Convert to float.
5094 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5095 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5096 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5097 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5098 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5099 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5101 // Use reciprocal estimate and one refinement step.
5102 // float4 recip = vrecpeq_f32(yf);
5103 // recip *= vrecpsq_f32(yf, recip);
5104 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5105 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5106 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5107 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5109 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5110 // Because short has a smaller range than ushort, we can actually get away
5111 // with only a single newton step. This requires that we use a weird bias
5112 // of 89, however (again, this has been exhaustively tested).
5113 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5114 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5115 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5116 N1 = DAG.getConstant(0x89, MVT::i32);
5117 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5118 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5119 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5120 // Convert back to integer and return.
5121 // return vmovn_s32(vcvt_s32_f32(result));
5122 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5123 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5127 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5128 EVT VT = Op.getValueType();
5129 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5130 "unexpected type for custom-lowering ISD::SDIV");
5132 DebugLoc dl = Op.getDebugLoc();
5133 SDValue N0 = Op.getOperand(0);
5134 SDValue N1 = Op.getOperand(1);
5137 if (VT == MVT::v8i8) {
5138 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5139 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5141 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5142 DAG.getIntPtrConstant(4));
5143 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5144 DAG.getIntPtrConstant(4));
5145 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5146 DAG.getIntPtrConstant(0));
5147 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5148 DAG.getIntPtrConstant(0));
5150 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5151 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5153 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5154 N0 = LowerCONCAT_VECTORS(N0, DAG);
5156 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5159 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5162 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5163 EVT VT = Op.getValueType();
5164 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5165 "unexpected type for custom-lowering ISD::UDIV");
5167 DebugLoc dl = Op.getDebugLoc();
5168 SDValue N0 = Op.getOperand(0);
5169 SDValue N1 = Op.getOperand(1);
5172 if (VT == MVT::v8i8) {
5173 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5174 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5176 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5177 DAG.getIntPtrConstant(4));
5178 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5179 DAG.getIntPtrConstant(4));
5180 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5181 DAG.getIntPtrConstant(0));
5182 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5183 DAG.getIntPtrConstant(0));
5185 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5186 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5188 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5189 N0 = LowerCONCAT_VECTORS(N0, DAG);
5191 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5192 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5197 // v4i16 sdiv ... Convert to float.
5198 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5199 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5200 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5201 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5202 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5203 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5205 // Use reciprocal estimate and two refinement steps.
5206 // float4 recip = vrecpeq_f32(yf);
5207 // recip *= vrecpsq_f32(yf, recip);
5208 // recip *= vrecpsq_f32(yf, recip);
5209 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5210 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5211 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5212 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5214 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5215 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5216 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5218 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5219 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5220 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5221 // and that it will never cause us to return an answer too large).
5222 // float4 result = as_float4(as_int4(xf*recip) + 2);
5223 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5224 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5225 N1 = DAG.getConstant(2, MVT::i32);
5226 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5227 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5228 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5229 // Convert back to integer and return.
5230 // return vmovn_u32(vcvt_s32_f32(result));
5231 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5232 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5236 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5237 EVT VT = Op.getNode()->getValueType(0);
5238 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5241 bool ExtraOp = false;
5242 switch (Op.getOpcode()) {
5243 default: llvm_unreachable("Invalid code");
5244 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5245 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5246 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5247 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5251 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5253 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5254 Op.getOperand(1), Op.getOperand(2));
5257 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5258 // Monotonic load/store is legal for all targets
5259 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5262 // Aquire/Release load/store is not legal for targets without a
5263 // dmb or equivalent available.
5269 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5270 SelectionDAG &DAG, unsigned NewOp) {
5271 DebugLoc dl = Node->getDebugLoc();
5272 assert (Node->getValueType(0) == MVT::i64 &&
5273 "Only know how to expand i64 atomics");
5275 SmallVector<SDValue, 6> Ops;
5276 Ops.push_back(Node->getOperand(0)); // Chain
5277 Ops.push_back(Node->getOperand(1)); // Ptr
5279 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5280 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5281 // High part of Val1
5282 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5283 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5284 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5285 // High part of Val1
5286 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5287 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5288 // High part of Val2
5289 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5290 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5292 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5294 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5295 cast<MemSDNode>(Node)->getMemOperand());
5296 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5297 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5298 Results.push_back(Result.getValue(2));
5301 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5302 switch (Op.getOpcode()) {
5303 default: llvm_unreachable("Don't know how to custom lower this!");
5304 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5305 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5306 case ISD::GlobalAddress:
5307 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5308 LowerGlobalAddressELF(Op, DAG);
5309 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5310 case ISD::SELECT: return LowerSELECT(Op, DAG);
5311 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5312 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5313 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5314 case ISD::VASTART: return LowerVASTART(Op, DAG);
5315 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5316 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5317 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5318 case ISD::SINT_TO_FP:
5319 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5320 case ISD::FP_TO_SINT:
5321 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5322 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5323 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5324 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5325 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5326 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5327 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5328 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5330 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5333 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5334 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5335 case ISD::SRL_PARTS:
5336 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5337 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5338 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5339 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5340 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5341 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5342 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5343 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5344 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5345 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5346 case ISD::MUL: return LowerMUL(Op, DAG);
5347 case ISD::SDIV: return LowerSDIV(Op, DAG);
5348 case ISD::UDIV: return LowerUDIV(Op, DAG);
5352 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5353 case ISD::ATOMIC_LOAD:
5354 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5358 /// ReplaceNodeResults - Replace the results of node with an illegal result
5359 /// type with new values built out of custom code.
5360 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5361 SmallVectorImpl<SDValue>&Results,
5362 SelectionDAG &DAG) const {
5364 switch (N->getOpcode()) {
5366 llvm_unreachable("Don't know how to custom expand this!");
5368 Res = ExpandBITCAST(N, DAG);
5372 Res = Expand64BitShift(N, DAG, Subtarget);
5374 case ISD::ATOMIC_LOAD_ADD:
5375 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5377 case ISD::ATOMIC_LOAD_AND:
5378 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5380 case ISD::ATOMIC_LOAD_NAND:
5381 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5383 case ISD::ATOMIC_LOAD_OR:
5384 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5386 case ISD::ATOMIC_LOAD_SUB:
5387 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5389 case ISD::ATOMIC_LOAD_XOR:
5390 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5392 case ISD::ATOMIC_SWAP:
5393 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5395 case ISD::ATOMIC_CMP_SWAP:
5396 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5400 Results.push_back(Res);
5403 //===----------------------------------------------------------------------===//
5404 // ARM Scheduler Hooks
5405 //===----------------------------------------------------------------------===//
5408 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5409 MachineBasicBlock *BB,
5410 unsigned Size) const {
5411 unsigned dest = MI->getOperand(0).getReg();
5412 unsigned ptr = MI->getOperand(1).getReg();
5413 unsigned oldval = MI->getOperand(2).getReg();
5414 unsigned newval = MI->getOperand(3).getReg();
5415 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5416 DebugLoc dl = MI->getDebugLoc();
5417 bool isThumb2 = Subtarget->isThumb2();
5419 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5420 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5421 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5422 (const TargetRegisterClass*)&ARM::GPRRegClass);
5425 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5426 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5427 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5430 unsigned ldrOpc, strOpc;
5432 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5434 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5435 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5438 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5439 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5442 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5443 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5447 MachineFunction *MF = BB->getParent();
5448 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5449 MachineFunction::iterator It = BB;
5450 ++It; // insert the new blocks after the current block
5452 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5453 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5454 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5455 MF->insert(It, loop1MBB);
5456 MF->insert(It, loop2MBB);
5457 MF->insert(It, exitMBB);
5459 // Transfer the remainder of BB and its successor edges to exitMBB.
5460 exitMBB->splice(exitMBB->begin(), BB,
5461 llvm::next(MachineBasicBlock::iterator(MI)),
5463 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5467 // fallthrough --> loop1MBB
5468 BB->addSuccessor(loop1MBB);
5471 // ldrex dest, [ptr]
5475 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5476 if (ldrOpc == ARM::t2LDREX)
5478 AddDefaultPred(MIB);
5479 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5480 .addReg(dest).addReg(oldval));
5481 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5482 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5483 BB->addSuccessor(loop2MBB);
5484 BB->addSuccessor(exitMBB);
5487 // strex scratch, newval, [ptr]
5491 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5492 if (strOpc == ARM::t2STREX)
5494 AddDefaultPred(MIB);
5495 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5496 .addReg(scratch).addImm(0));
5497 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5498 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5499 BB->addSuccessor(loop1MBB);
5500 BB->addSuccessor(exitMBB);
5506 MI->eraseFromParent(); // The instruction is gone now.
5512 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5513 unsigned Size, unsigned BinOpcode) const {
5514 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5515 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5517 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5518 MachineFunction *MF = BB->getParent();
5519 MachineFunction::iterator It = BB;
5522 unsigned dest = MI->getOperand(0).getReg();
5523 unsigned ptr = MI->getOperand(1).getReg();
5524 unsigned incr = MI->getOperand(2).getReg();
5525 DebugLoc dl = MI->getDebugLoc();
5526 bool isThumb2 = Subtarget->isThumb2();
5528 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5530 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5531 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5534 unsigned ldrOpc, strOpc;
5536 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5538 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5539 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5542 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5543 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5546 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5547 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5551 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5552 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5553 MF->insert(It, loopMBB);
5554 MF->insert(It, exitMBB);
5556 // Transfer the remainder of BB and its successor edges to exitMBB.
5557 exitMBB->splice(exitMBB->begin(), BB,
5558 llvm::next(MachineBasicBlock::iterator(MI)),
5560 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5562 const TargetRegisterClass *TRC = isThumb2 ?
5563 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5564 (const TargetRegisterClass*)&ARM::GPRRegClass;
5565 unsigned scratch = MRI.createVirtualRegister(TRC);
5566 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5570 // fallthrough --> loopMBB
5571 BB->addSuccessor(loopMBB);
5575 // <binop> scratch2, dest, incr
5576 // strex scratch, scratch2, ptr
5579 // fallthrough --> exitMBB
5581 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5582 if (ldrOpc == ARM::t2LDREX)
5584 AddDefaultPred(MIB);
5586 // operand order needs to go the other way for NAND
5587 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5588 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5589 addReg(incr).addReg(dest)).addReg(0);
5591 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5592 addReg(dest).addReg(incr)).addReg(0);
5595 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5596 if (strOpc == ARM::t2STREX)
5598 AddDefaultPred(MIB);
5599 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5600 .addReg(scratch).addImm(0));
5601 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5602 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5604 BB->addSuccessor(loopMBB);
5605 BB->addSuccessor(exitMBB);
5611 MI->eraseFromParent(); // The instruction is gone now.
5617 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5618 MachineBasicBlock *BB,
5621 ARMCC::CondCodes Cond) const {
5622 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5624 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5625 MachineFunction *MF = BB->getParent();
5626 MachineFunction::iterator It = BB;
5629 unsigned dest = MI->getOperand(0).getReg();
5630 unsigned ptr = MI->getOperand(1).getReg();
5631 unsigned incr = MI->getOperand(2).getReg();
5632 unsigned oldval = dest;
5633 DebugLoc dl = MI->getDebugLoc();
5634 bool isThumb2 = Subtarget->isThumb2();
5636 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5638 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5639 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5642 unsigned ldrOpc, strOpc, extendOpc;
5644 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5646 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5647 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5648 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5651 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5652 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5653 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5656 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5657 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5662 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5663 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5664 MF->insert(It, loopMBB);
5665 MF->insert(It, exitMBB);
5667 // Transfer the remainder of BB and its successor edges to exitMBB.
5668 exitMBB->splice(exitMBB->begin(), BB,
5669 llvm::next(MachineBasicBlock::iterator(MI)),
5671 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5673 const TargetRegisterClass *TRC = isThumb2 ?
5674 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5675 (const TargetRegisterClass*)&ARM::GPRRegClass;
5676 unsigned scratch = MRI.createVirtualRegister(TRC);
5677 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5681 // fallthrough --> loopMBB
5682 BB->addSuccessor(loopMBB);
5686 // (sign extend dest, if required)
5688 // cmov.cond scratch2, incr, dest
5689 // strex scratch, scratch2, ptr
5692 // fallthrough --> exitMBB
5694 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5695 if (ldrOpc == ARM::t2LDREX)
5697 AddDefaultPred(MIB);
5699 // Sign extend the value, if necessary.
5700 if (signExtend && extendOpc) {
5701 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5702 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5707 // Build compare and cmov instructions.
5708 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5709 .addReg(oldval).addReg(incr));
5710 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5711 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5713 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5714 if (strOpc == ARM::t2STREX)
5716 AddDefaultPred(MIB);
5717 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5718 .addReg(scratch).addImm(0));
5719 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5720 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5722 BB->addSuccessor(loopMBB);
5723 BB->addSuccessor(exitMBB);
5729 MI->eraseFromParent(); // The instruction is gone now.
5735 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5736 unsigned Op1, unsigned Op2,
5737 bool NeedsCarry, bool IsCmpxchg) const {
5738 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5739 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5741 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5742 MachineFunction *MF = BB->getParent();
5743 MachineFunction::iterator It = BB;
5746 unsigned destlo = MI->getOperand(0).getReg();
5747 unsigned desthi = MI->getOperand(1).getReg();
5748 unsigned ptr = MI->getOperand(2).getReg();
5749 unsigned vallo = MI->getOperand(3).getReg();
5750 unsigned valhi = MI->getOperand(4).getReg();
5751 DebugLoc dl = MI->getDebugLoc();
5752 bool isThumb2 = Subtarget->isThumb2();
5754 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5756 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5757 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5758 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5761 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5762 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5764 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5765 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5767 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5768 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5770 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5771 MF->insert(It, loopMBB);
5773 MF->insert(It, contBB);
5774 MF->insert(It, cont2BB);
5776 MF->insert(It, exitMBB);
5778 // Transfer the remainder of BB and its successor edges to exitMBB.
5779 exitMBB->splice(exitMBB->begin(), BB,
5780 llvm::next(MachineBasicBlock::iterator(MI)),
5782 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5784 const TargetRegisterClass *TRC = isThumb2 ?
5785 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5786 (const TargetRegisterClass*)&ARM::GPRRegClass;
5787 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5791 // fallthrough --> loopMBB
5792 BB->addSuccessor(loopMBB);
5795 // ldrexd r2, r3, ptr
5796 // <binopa> r0, r2, incr
5797 // <binopb> r1, r3, incr
5798 // strexd storesuccess, r0, r1, ptr
5799 // cmp storesuccess, #0
5801 // fallthrough --> exitMBB
5803 // Note that the registers are explicitly specified because there is not any
5804 // way to force the register allocator to allocate a register pair.
5806 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5807 // need to properly enforce the restriction that the two output registers
5808 // for ldrexd must be different.
5811 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5812 .addReg(ARM::R2, RegState::Define)
5813 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5814 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5815 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5816 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5820 for (unsigned i = 0; i < 2; i++) {
5821 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5823 .addReg(i == 0 ? destlo : desthi)
5824 .addReg(i == 0 ? vallo : valhi));
5825 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5826 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5827 BB->addSuccessor(exitMBB);
5828 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5829 BB = (i == 0 ? contBB : cont2BB);
5832 // Copy to physregs for strexd
5833 unsigned setlo = MI->getOperand(5).getReg();
5834 unsigned sethi = MI->getOperand(6).getReg();
5835 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5836 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5838 // Perform binary operation
5839 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5840 .addReg(destlo).addReg(vallo))
5841 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5842 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5843 .addReg(desthi).addReg(valhi)).addReg(0);
5845 // Copy to physregs for strexd
5846 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5847 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5851 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5852 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5854 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5855 .addReg(storesuccess).addImm(0));
5856 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5857 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5859 BB->addSuccessor(loopMBB);
5860 BB->addSuccessor(exitMBB);
5866 MI->eraseFromParent(); // The instruction is gone now.
5871 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5872 /// registers the function context.
5873 void ARMTargetLowering::
5874 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5875 MachineBasicBlock *DispatchBB, int FI) const {
5876 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5877 DebugLoc dl = MI->getDebugLoc();
5878 MachineFunction *MF = MBB->getParent();
5879 MachineRegisterInfo *MRI = &MF->getRegInfo();
5880 MachineConstantPool *MCP = MF->getConstantPool();
5881 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5882 const Function *F = MF->getFunction();
5884 bool isThumb = Subtarget->isThumb();
5885 bool isThumb2 = Subtarget->isThumb2();
5887 unsigned PCLabelId = AFI->createPICLabelUId();
5888 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5889 ARMConstantPoolValue *CPV =
5890 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5891 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5893 const TargetRegisterClass *TRC = isThumb ?
5894 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5895 (const TargetRegisterClass*)&ARM::GPRRegClass;
5897 // Grab constant pool and fixed stack memory operands.
5898 MachineMemOperand *CPMMO =
5899 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5900 MachineMemOperand::MOLoad, 4, 4);
5902 MachineMemOperand *FIMMOSt =
5903 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5904 MachineMemOperand::MOStore, 4, 4);
5906 // Load the address of the dispatch MBB into the jump buffer.
5908 // Incoming value: jbuf
5909 // ldr.n r5, LCPI1_1
5912 // str r5, [$jbuf, #+4] ; &jbuf[1]
5913 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5914 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5915 .addConstantPoolIndex(CPI)
5916 .addMemOperand(CPMMO));
5917 // Set the low bit because of thumb mode.
5918 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5920 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5921 .addReg(NewVReg1, RegState::Kill)
5923 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5924 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5925 .addReg(NewVReg2, RegState::Kill)
5927 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5928 .addReg(NewVReg3, RegState::Kill)
5930 .addImm(36) // &jbuf[1] :: pc
5931 .addMemOperand(FIMMOSt));
5932 } else if (isThumb) {
5933 // Incoming value: jbuf
5934 // ldr.n r1, LCPI1_4
5938 // add r2, $jbuf, #+4 ; &jbuf[1]
5940 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5941 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5942 .addConstantPoolIndex(CPI)
5943 .addMemOperand(CPMMO));
5944 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5945 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5946 .addReg(NewVReg1, RegState::Kill)
5948 // Set the low bit because of thumb mode.
5949 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5950 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5951 .addReg(ARM::CPSR, RegState::Define)
5953 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5954 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5955 .addReg(ARM::CPSR, RegState::Define)
5956 .addReg(NewVReg2, RegState::Kill)
5957 .addReg(NewVReg3, RegState::Kill));
5958 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5959 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5961 .addImm(36)); // &jbuf[1] :: pc
5962 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5963 .addReg(NewVReg4, RegState::Kill)
5964 .addReg(NewVReg5, RegState::Kill)
5966 .addMemOperand(FIMMOSt));
5968 // Incoming value: jbuf
5971 // str r1, [$jbuf, #+4] ; &jbuf[1]
5972 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5973 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5974 .addConstantPoolIndex(CPI)
5976 .addMemOperand(CPMMO));
5977 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5978 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5979 .addReg(NewVReg1, RegState::Kill)
5980 .addImm(PCLabelId));
5981 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5982 .addReg(NewVReg2, RegState::Kill)
5984 .addImm(36) // &jbuf[1] :: pc
5985 .addMemOperand(FIMMOSt));
5989 MachineBasicBlock *ARMTargetLowering::
5990 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5991 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5992 DebugLoc dl = MI->getDebugLoc();
5993 MachineFunction *MF = MBB->getParent();
5994 MachineRegisterInfo *MRI = &MF->getRegInfo();
5995 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5996 MachineFrameInfo *MFI = MF->getFrameInfo();
5997 int FI = MFI->getFunctionContextIndex();
5999 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6000 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6001 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6003 // Get a mapping of the call site numbers to all of the landing pads they're
6005 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6006 unsigned MaxCSNum = 0;
6007 MachineModuleInfo &MMI = MF->getMMI();
6008 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6010 if (!BB->isLandingPad()) continue;
6012 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6014 for (MachineBasicBlock::iterator
6015 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6016 if (!II->isEHLabel()) continue;
6018 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6019 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6021 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6022 for (SmallVectorImpl<unsigned>::iterator
6023 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6024 CSI != CSE; ++CSI) {
6025 CallSiteNumToLPad[*CSI].push_back(BB);
6026 MaxCSNum = std::max(MaxCSNum, *CSI);
6032 // Get an ordered list of the machine basic blocks for the jump table.
6033 std::vector<MachineBasicBlock*> LPadList;
6034 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6035 LPadList.reserve(CallSiteNumToLPad.size());
6036 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6037 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6038 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6039 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6040 LPadList.push_back(*II);
6041 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6045 assert(!LPadList.empty() &&
6046 "No landing pad destinations for the dispatch jump table!");
6048 // Create the jump table and associated information.
6049 MachineJumpTableInfo *JTI =
6050 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6051 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6052 unsigned UId = AFI->createJumpTableUId();
6054 // Create the MBBs for the dispatch code.
6056 // Shove the dispatch's address into the return slot in the function context.
6057 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6058 DispatchBB->setIsLandingPad();
6060 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6061 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6062 DispatchBB->addSuccessor(TrapBB);
6064 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6065 DispatchBB->addSuccessor(DispContBB);
6068 MF->insert(MF->end(), DispatchBB);
6069 MF->insert(MF->end(), DispContBB);
6070 MF->insert(MF->end(), TrapBB);
6072 // Insert code into the entry block that creates and registers the function
6074 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6076 MachineMemOperand *FIMMOLd =
6077 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6078 MachineMemOperand::MOLoad |
6079 MachineMemOperand::MOVolatile, 4, 4);
6081 MachineInstrBuilder MIB;
6082 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6084 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6085 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6087 // Add a register mask with no preserved registers. This results in all
6088 // registers being marked as clobbered.
6089 MIB.addRegMask(RI.getNoPreservedMask());
6091 unsigned NumLPads = LPadList.size();
6092 if (Subtarget->isThumb2()) {
6093 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6094 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6097 .addMemOperand(FIMMOLd));
6099 if (NumLPads < 256) {
6100 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6102 .addImm(LPadList.size()));
6104 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6105 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6106 .addImm(NumLPads & 0xFFFF));
6108 unsigned VReg2 = VReg1;
6109 if ((NumLPads & 0xFFFF0000) != 0) {
6110 VReg2 = MRI->createVirtualRegister(TRC);
6111 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6113 .addImm(NumLPads >> 16));
6116 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6121 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6126 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6127 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6128 .addJumpTableIndex(MJTI)
6131 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6134 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6135 .addReg(NewVReg3, RegState::Kill)
6137 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6139 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6140 .addReg(NewVReg4, RegState::Kill)
6142 .addJumpTableIndex(MJTI)
6144 } else if (Subtarget->isThumb()) {
6145 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6146 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6149 .addMemOperand(FIMMOLd));
6151 if (NumLPads < 256) {
6152 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6156 MachineConstantPool *ConstantPool = MF->getConstantPool();
6157 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6158 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6160 // MachineConstantPool wants an explicit alignment.
6161 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6163 Align = getDataLayout()->getTypeAllocSize(C->getType());
6164 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6166 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6167 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6168 .addReg(VReg1, RegState::Define)
6169 .addConstantPoolIndex(Idx));
6170 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6175 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6180 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6181 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6182 .addReg(ARM::CPSR, RegState::Define)
6186 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6187 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6188 .addJumpTableIndex(MJTI)
6191 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6192 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6193 .addReg(ARM::CPSR, RegState::Define)
6194 .addReg(NewVReg2, RegState::Kill)
6197 MachineMemOperand *JTMMOLd =
6198 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6199 MachineMemOperand::MOLoad, 4, 4);
6201 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6202 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6203 .addReg(NewVReg4, RegState::Kill)
6205 .addMemOperand(JTMMOLd));
6207 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6208 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6209 .addReg(ARM::CPSR, RegState::Define)
6210 .addReg(NewVReg5, RegState::Kill)
6213 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6214 .addReg(NewVReg6, RegState::Kill)
6215 .addJumpTableIndex(MJTI)
6218 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6219 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6222 .addMemOperand(FIMMOLd));
6224 if (NumLPads < 256) {
6225 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6228 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6229 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6230 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6231 .addImm(NumLPads & 0xFFFF));
6233 unsigned VReg2 = VReg1;
6234 if ((NumLPads & 0xFFFF0000) != 0) {
6235 VReg2 = MRI->createVirtualRegister(TRC);
6236 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6238 .addImm(NumLPads >> 16));
6241 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6245 MachineConstantPool *ConstantPool = MF->getConstantPool();
6246 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6247 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6249 // MachineConstantPool wants an explicit alignment.
6250 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6252 Align = getDataLayout()->getTypeAllocSize(C->getType());
6253 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6255 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6256 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6257 .addReg(VReg1, RegState::Define)
6258 .addConstantPoolIndex(Idx)
6260 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6262 .addReg(VReg1, RegState::Kill));
6265 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6270 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6272 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6274 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6275 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6276 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6277 .addJumpTableIndex(MJTI)
6280 MachineMemOperand *JTMMOLd =
6281 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6282 MachineMemOperand::MOLoad, 4, 4);
6283 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6285 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6286 .addReg(NewVReg3, RegState::Kill)
6289 .addMemOperand(JTMMOLd));
6291 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6292 .addReg(NewVReg5, RegState::Kill)
6294 .addJumpTableIndex(MJTI)
6298 // Add the jump table entries as successors to the MBB.
6299 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6300 for (std::vector<MachineBasicBlock*>::iterator
6301 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6302 MachineBasicBlock *CurMBB = *I;
6303 if (SeenMBBs.insert(CurMBB))
6304 DispContBB->addSuccessor(CurMBB);
6307 // N.B. the order the invoke BBs are processed in doesn't matter here.
6308 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6309 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6310 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6311 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6312 MachineBasicBlock *BB = *I;
6314 // Remove the landing pad successor from the invoke block and replace it
6315 // with the new dispatch block.
6316 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6318 while (!Successors.empty()) {
6319 MachineBasicBlock *SMBB = Successors.pop_back_val();
6320 if (SMBB->isLandingPad()) {
6321 BB->removeSuccessor(SMBB);
6322 MBBLPads.push_back(SMBB);
6326 BB->addSuccessor(DispatchBB);
6328 // Find the invoke call and mark all of the callee-saved registers as
6329 // 'implicit defined' so that they're spilled. This prevents code from
6330 // moving instructions to before the EH block, where they will never be
6332 for (MachineBasicBlock::reverse_iterator
6333 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6334 if (!II->isCall()) continue;
6336 DenseMap<unsigned, bool> DefRegs;
6337 for (MachineInstr::mop_iterator
6338 OI = II->operands_begin(), OE = II->operands_end();
6340 if (!OI->isReg()) continue;
6341 DefRegs[OI->getReg()] = true;
6344 MachineInstrBuilder MIB(&*II);
6346 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6347 unsigned Reg = SavedRegs[i];
6348 if (Subtarget->isThumb2() &&
6349 !ARM::tGPRRegClass.contains(Reg) &&
6350 !ARM::hGPRRegClass.contains(Reg))
6352 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6354 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6357 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6364 // Mark all former landing pads as non-landing pads. The dispatch is the only
6366 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6367 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6368 (*I)->setIsLandingPad(false);
6370 // The instruction is gone now.
6371 MI->eraseFromParent();
6377 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6378 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6379 E = MBB->succ_end(); I != E; ++I)
6382 llvm_unreachable("Expecting a BB with two successors!");
6385 MachineBasicBlock *ARMTargetLowering::
6386 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6387 // This pseudo instruction has 3 operands: dst, src, size
6388 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6389 // Otherwise, we will generate unrolled scalar copies.
6390 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6391 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6392 MachineFunction::iterator It = BB;
6395 unsigned dest = MI->getOperand(0).getReg();
6396 unsigned src = MI->getOperand(1).getReg();
6397 unsigned SizeVal = MI->getOperand(2).getImm();
6398 unsigned Align = MI->getOperand(3).getImm();
6399 DebugLoc dl = MI->getDebugLoc();
6401 bool isThumb2 = Subtarget->isThumb2();
6402 MachineFunction *MF = BB->getParent();
6403 MachineRegisterInfo &MRI = MF->getRegInfo();
6404 unsigned ldrOpc, strOpc, UnitSize = 0;
6406 const TargetRegisterClass *TRC = isThumb2 ?
6407 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6408 (const TargetRegisterClass*)&ARM::GPRRegClass;
6409 const TargetRegisterClass *TRC_Vec = 0;
6412 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6413 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6415 } else if (Align & 2) {
6416 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6417 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6420 // Check whether we can use NEON instructions.
6421 if (!MF->getFunction()->getFnAttributes().
6422 hasAttribute(Attributes::NoImplicitFloat) &&
6423 Subtarget->hasNEON()) {
6424 if ((Align % 16 == 0) && SizeVal >= 16) {
6425 ldrOpc = ARM::VLD1q32wb_fixed;
6426 strOpc = ARM::VST1q32wb_fixed;
6428 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6430 else if ((Align % 8 == 0) && SizeVal >= 8) {
6431 ldrOpc = ARM::VLD1d32wb_fixed;
6432 strOpc = ARM::VST1d32wb_fixed;
6434 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6437 // Can't use NEON instructions.
6438 if (UnitSize == 0) {
6439 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6440 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6445 unsigned BytesLeft = SizeVal % UnitSize;
6446 unsigned LoopSize = SizeVal - BytesLeft;
6448 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6449 // Use LDR and STR to copy.
6450 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6451 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6452 unsigned srcIn = src;
6453 unsigned destIn = dest;
6454 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6455 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6456 unsigned srcOut = MRI.createVirtualRegister(TRC);
6457 unsigned destOut = MRI.createVirtualRegister(TRC);
6458 if (UnitSize >= 8) {
6459 AddDefaultPred(BuildMI(*BB, MI, dl,
6460 TII->get(ldrOpc), scratch)
6461 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6463 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6464 .addReg(destIn).addImm(0).addReg(scratch));
6465 } else if (isThumb2) {
6466 AddDefaultPred(BuildMI(*BB, MI, dl,
6467 TII->get(ldrOpc), scratch)
6468 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6470 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6471 .addReg(scratch).addReg(destIn)
6474 AddDefaultPred(BuildMI(*BB, MI, dl,
6475 TII->get(ldrOpc), scratch)
6476 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6479 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6480 .addReg(scratch).addReg(destIn)
6481 .addReg(0).addImm(UnitSize));
6487 // Handle the leftover bytes with LDRB and STRB.
6488 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6489 // [destOut] = STRB_POST(scratch, destIn, 1)
6490 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6491 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6492 for (unsigned i = 0; i < BytesLeft; i++) {
6493 unsigned scratch = MRI.createVirtualRegister(TRC);
6494 unsigned srcOut = MRI.createVirtualRegister(TRC);
6495 unsigned destOut = MRI.createVirtualRegister(TRC);
6497 AddDefaultPred(BuildMI(*BB, MI, dl,
6498 TII->get(ldrOpc),scratch)
6499 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6501 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6502 .addReg(scratch).addReg(destIn)
6503 .addReg(0).addImm(1));
6505 AddDefaultPred(BuildMI(*BB, MI, dl,
6506 TII->get(ldrOpc),scratch)
6507 .addReg(srcOut, RegState::Define).addReg(srcIn)
6508 .addReg(0).addImm(1));
6510 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6511 .addReg(scratch).addReg(destIn)
6512 .addReg(0).addImm(1));
6517 MI->eraseFromParent(); // The instruction is gone now.
6521 // Expand the pseudo op to a loop.
6524 // movw varEnd, # --> with thumb2
6526 // ldrcp varEnd, idx --> without thumb2
6527 // fallthrough --> loopMBB
6529 // PHI varPhi, varEnd, varLoop
6530 // PHI srcPhi, src, srcLoop
6531 // PHI destPhi, dst, destLoop
6532 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6533 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6534 // subs varLoop, varPhi, #UnitSize
6536 // fallthrough --> exitMBB
6538 // epilogue to handle left-over bytes
6539 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6540 // [destOut] = STRB_POST(scratch, destLoop, 1)
6541 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6542 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6543 MF->insert(It, loopMBB);
6544 MF->insert(It, exitMBB);
6546 // Transfer the remainder of BB and its successor edges to exitMBB.
6547 exitMBB->splice(exitMBB->begin(), BB,
6548 llvm::next(MachineBasicBlock::iterator(MI)),
6550 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6552 // Load an immediate to varEnd.
6553 unsigned varEnd = MRI.createVirtualRegister(TRC);
6555 unsigned VReg1 = varEnd;
6556 if ((LoopSize & 0xFFFF0000) != 0)
6557 VReg1 = MRI.createVirtualRegister(TRC);
6558 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6559 .addImm(LoopSize & 0xFFFF));
6561 if ((LoopSize & 0xFFFF0000) != 0)
6562 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6564 .addImm(LoopSize >> 16));
6566 MachineConstantPool *ConstantPool = MF->getConstantPool();
6567 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6568 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6570 // MachineConstantPool wants an explicit alignment.
6571 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6573 Align = getDataLayout()->getTypeAllocSize(C->getType());
6574 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6576 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6577 .addReg(varEnd, RegState::Define)
6578 .addConstantPoolIndex(Idx)
6581 BB->addSuccessor(loopMBB);
6583 // Generate the loop body:
6584 // varPhi = PHI(varLoop, varEnd)
6585 // srcPhi = PHI(srcLoop, src)
6586 // destPhi = PHI(destLoop, dst)
6587 MachineBasicBlock *entryBB = BB;
6589 unsigned varLoop = MRI.createVirtualRegister(TRC);
6590 unsigned varPhi = MRI.createVirtualRegister(TRC);
6591 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6592 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6593 unsigned destLoop = MRI.createVirtualRegister(TRC);
6594 unsigned destPhi = MRI.createVirtualRegister(TRC);
6596 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6597 .addReg(varLoop).addMBB(loopMBB)
6598 .addReg(varEnd).addMBB(entryBB);
6599 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6600 .addReg(srcLoop).addMBB(loopMBB)
6601 .addReg(src).addMBB(entryBB);
6602 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6603 .addReg(destLoop).addMBB(loopMBB)
6604 .addReg(dest).addMBB(entryBB);
6606 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6607 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6608 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6609 if (UnitSize >= 8) {
6610 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6611 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6613 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6614 .addReg(destPhi).addImm(0).addReg(scratch));
6615 } else if (isThumb2) {
6616 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6617 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6619 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6620 .addReg(scratch).addReg(destPhi)
6623 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6624 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6627 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6628 .addReg(scratch).addReg(destPhi)
6629 .addReg(0).addImm(UnitSize));
6632 // Decrement loop variable by UnitSize.
6633 MachineInstrBuilder MIB = BuildMI(BB, dl,
6634 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6635 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6636 MIB->getOperand(5).setReg(ARM::CPSR);
6637 MIB->getOperand(5).setIsDef(true);
6639 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6640 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6642 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6643 BB->addSuccessor(loopMBB);
6644 BB->addSuccessor(exitMBB);
6646 // Add epilogue to handle BytesLeft.
6648 MachineInstr *StartOfExit = exitMBB->begin();
6649 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6650 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6652 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6653 // [destOut] = STRB_POST(scratch, destLoop, 1)
6654 unsigned srcIn = srcLoop;
6655 unsigned destIn = destLoop;
6656 for (unsigned i = 0; i < BytesLeft; i++) {
6657 unsigned scratch = MRI.createVirtualRegister(TRC);
6658 unsigned srcOut = MRI.createVirtualRegister(TRC);
6659 unsigned destOut = MRI.createVirtualRegister(TRC);
6661 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6662 TII->get(ldrOpc),scratch)
6663 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6665 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6666 .addReg(scratch).addReg(destIn)
6669 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6670 TII->get(ldrOpc),scratch)
6671 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6673 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6674 .addReg(scratch).addReg(destIn)
6675 .addReg(0).addImm(1));
6681 MI->eraseFromParent(); // The instruction is gone now.
6686 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6687 MachineBasicBlock *BB) const {
6688 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6689 DebugLoc dl = MI->getDebugLoc();
6690 bool isThumb2 = Subtarget->isThumb2();
6691 switch (MI->getOpcode()) {
6694 llvm_unreachable("Unexpected instr type to insert");
6696 // The Thumb2 pre-indexed stores have the same MI operands, they just
6697 // define them differently in the .td files from the isel patterns, so
6698 // they need pseudos.
6699 case ARM::t2STR_preidx:
6700 MI->setDesc(TII->get(ARM::t2STR_PRE));
6702 case ARM::t2STRB_preidx:
6703 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6705 case ARM::t2STRH_preidx:
6706 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6709 case ARM::STRi_preidx:
6710 case ARM::STRBi_preidx: {
6711 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6712 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6713 // Decode the offset.
6714 unsigned Offset = MI->getOperand(4).getImm();
6715 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6716 Offset = ARM_AM::getAM2Offset(Offset);
6720 MachineMemOperand *MMO = *MI->memoperands_begin();
6721 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6722 .addOperand(MI->getOperand(0)) // Rn_wb
6723 .addOperand(MI->getOperand(1)) // Rt
6724 .addOperand(MI->getOperand(2)) // Rn
6725 .addImm(Offset) // offset (skip GPR==zero_reg)
6726 .addOperand(MI->getOperand(5)) // pred
6727 .addOperand(MI->getOperand(6))
6728 .addMemOperand(MMO);
6729 MI->eraseFromParent();
6732 case ARM::STRr_preidx:
6733 case ARM::STRBr_preidx:
6734 case ARM::STRH_preidx: {
6736 switch (MI->getOpcode()) {
6737 default: llvm_unreachable("unexpected opcode!");
6738 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6739 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6740 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6742 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6743 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6744 MIB.addOperand(MI->getOperand(i));
6745 MI->eraseFromParent();
6748 case ARM::ATOMIC_LOAD_ADD_I8:
6749 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6750 case ARM::ATOMIC_LOAD_ADD_I16:
6751 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6752 case ARM::ATOMIC_LOAD_ADD_I32:
6753 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6755 case ARM::ATOMIC_LOAD_AND_I8:
6756 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6757 case ARM::ATOMIC_LOAD_AND_I16:
6758 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6759 case ARM::ATOMIC_LOAD_AND_I32:
6760 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6762 case ARM::ATOMIC_LOAD_OR_I8:
6763 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6764 case ARM::ATOMIC_LOAD_OR_I16:
6765 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6766 case ARM::ATOMIC_LOAD_OR_I32:
6767 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6769 case ARM::ATOMIC_LOAD_XOR_I8:
6770 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6771 case ARM::ATOMIC_LOAD_XOR_I16:
6772 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6773 case ARM::ATOMIC_LOAD_XOR_I32:
6774 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6776 case ARM::ATOMIC_LOAD_NAND_I8:
6777 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6778 case ARM::ATOMIC_LOAD_NAND_I16:
6779 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6780 case ARM::ATOMIC_LOAD_NAND_I32:
6781 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6783 case ARM::ATOMIC_LOAD_SUB_I8:
6784 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6785 case ARM::ATOMIC_LOAD_SUB_I16:
6786 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6787 case ARM::ATOMIC_LOAD_SUB_I32:
6788 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6790 case ARM::ATOMIC_LOAD_MIN_I8:
6791 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6792 case ARM::ATOMIC_LOAD_MIN_I16:
6793 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6794 case ARM::ATOMIC_LOAD_MIN_I32:
6795 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6797 case ARM::ATOMIC_LOAD_MAX_I8:
6798 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6799 case ARM::ATOMIC_LOAD_MAX_I16:
6800 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6801 case ARM::ATOMIC_LOAD_MAX_I32:
6802 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6804 case ARM::ATOMIC_LOAD_UMIN_I8:
6805 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6806 case ARM::ATOMIC_LOAD_UMIN_I16:
6807 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6808 case ARM::ATOMIC_LOAD_UMIN_I32:
6809 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6811 case ARM::ATOMIC_LOAD_UMAX_I8:
6812 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6813 case ARM::ATOMIC_LOAD_UMAX_I16:
6814 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6815 case ARM::ATOMIC_LOAD_UMAX_I32:
6816 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6818 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6819 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6820 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6822 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6823 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6824 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6827 case ARM::ATOMADD6432:
6828 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6829 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6830 /*NeedsCarry*/ true);
6831 case ARM::ATOMSUB6432:
6832 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6833 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6834 /*NeedsCarry*/ true);
6835 case ARM::ATOMOR6432:
6836 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6837 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6838 case ARM::ATOMXOR6432:
6839 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6840 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6841 case ARM::ATOMAND6432:
6842 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6843 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6844 case ARM::ATOMSWAP6432:
6845 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6846 case ARM::ATOMCMPXCHG6432:
6847 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6848 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6849 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6851 case ARM::tMOVCCr_pseudo: {
6852 // To "insert" a SELECT_CC instruction, we actually have to insert the
6853 // diamond control-flow pattern. The incoming instruction knows the
6854 // destination vreg to set, the condition code register to branch on, the
6855 // true/false values to select between, and a branch opcode to use.
6856 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6857 MachineFunction::iterator It = BB;
6863 // cmpTY ccX, r1, r2
6865 // fallthrough --> copy0MBB
6866 MachineBasicBlock *thisMBB = BB;
6867 MachineFunction *F = BB->getParent();
6868 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6869 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6870 F->insert(It, copy0MBB);
6871 F->insert(It, sinkMBB);
6873 // Transfer the remainder of BB and its successor edges to sinkMBB.
6874 sinkMBB->splice(sinkMBB->begin(), BB,
6875 llvm::next(MachineBasicBlock::iterator(MI)),
6877 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6879 BB->addSuccessor(copy0MBB);
6880 BB->addSuccessor(sinkMBB);
6882 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6883 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6886 // %FalseValue = ...
6887 // # fallthrough to sinkMBB
6890 // Update machine-CFG edges
6891 BB->addSuccessor(sinkMBB);
6894 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6897 BuildMI(*BB, BB->begin(), dl,
6898 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6899 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6900 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6902 MI->eraseFromParent(); // The pseudo instruction is gone now.
6907 case ARM::BCCZi64: {
6908 // If there is an unconditional branch to the other successor, remove it.
6909 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6911 // Compare both parts that make up the double comparison separately for
6913 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6915 unsigned LHS1 = MI->getOperand(1).getReg();
6916 unsigned LHS2 = MI->getOperand(2).getReg();
6918 AddDefaultPred(BuildMI(BB, dl,
6919 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6920 .addReg(LHS1).addImm(0));
6921 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6922 .addReg(LHS2).addImm(0)
6923 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6925 unsigned RHS1 = MI->getOperand(3).getReg();
6926 unsigned RHS2 = MI->getOperand(4).getReg();
6927 AddDefaultPred(BuildMI(BB, dl,
6928 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6929 .addReg(LHS1).addReg(RHS1));
6930 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6931 .addReg(LHS2).addReg(RHS2)
6932 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6935 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6936 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6937 if (MI->getOperand(0).getImm() == ARMCC::NE)
6938 std::swap(destMBB, exitMBB);
6940 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6941 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6943 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6945 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6947 MI->eraseFromParent(); // The pseudo instruction is gone now.
6951 case ARM::Int_eh_sjlj_setjmp:
6952 case ARM::Int_eh_sjlj_setjmp_nofp:
6953 case ARM::tInt_eh_sjlj_setjmp:
6954 case ARM::t2Int_eh_sjlj_setjmp:
6955 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6956 EmitSjLjDispatchBlock(MI, BB);
6961 // To insert an ABS instruction, we have to insert the
6962 // diamond control-flow pattern. The incoming instruction knows the
6963 // source vreg to test against 0, the destination vreg to set,
6964 // the condition code register to branch on, the
6965 // true/false values to select between, and a branch opcode to use.
6970 // BCC (branch to SinkBB if V0 >= 0)
6971 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6972 // SinkBB: V1 = PHI(V2, V3)
6973 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6974 MachineFunction::iterator BBI = BB;
6976 MachineFunction *Fn = BB->getParent();
6977 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6978 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6979 Fn->insert(BBI, RSBBB);
6980 Fn->insert(BBI, SinkBB);
6982 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6983 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6984 bool isThumb2 = Subtarget->isThumb2();
6985 MachineRegisterInfo &MRI = Fn->getRegInfo();
6986 // In Thumb mode S must not be specified if source register is the SP or
6987 // PC and if destination register is the SP, so restrict register class
6988 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6989 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6990 (const TargetRegisterClass*)&ARM::GPRRegClass);
6992 // Transfer the remainder of BB and its successor edges to sinkMBB.
6993 SinkBB->splice(SinkBB->begin(), BB,
6994 llvm::next(MachineBasicBlock::iterator(MI)),
6996 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6998 BB->addSuccessor(RSBBB);
6999 BB->addSuccessor(SinkBB);
7001 // fall through to SinkMBB
7002 RSBBB->addSuccessor(SinkBB);
7004 // insert a cmp at the end of BB
7005 AddDefaultPred(BuildMI(BB, dl,
7006 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7007 .addReg(ABSSrcReg).addImm(0));
7009 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7011 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7012 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7014 // insert rsbri in RSBBB
7015 // Note: BCC and rsbri will be converted into predicated rsbmi
7016 // by if-conversion pass
7017 BuildMI(*RSBBB, RSBBB->begin(), dl,
7018 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7019 .addReg(ABSSrcReg, RegState::Kill)
7020 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7022 // insert PHI in SinkBB,
7023 // reuse ABSDstReg to not change uses of ABS instruction
7024 BuildMI(*SinkBB, SinkBB->begin(), dl,
7025 TII->get(ARM::PHI), ABSDstReg)
7026 .addReg(NewRsbDstReg).addMBB(RSBBB)
7027 .addReg(ABSSrcReg).addMBB(BB);
7029 // remove ABS instruction
7030 MI->eraseFromParent();
7032 // return last added BB
7035 case ARM::COPY_STRUCT_BYVAL_I32:
7037 return EmitStructByval(MI, BB);
7041 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7042 SDNode *Node) const {
7043 if (!MI->hasPostISelHook()) {
7044 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7045 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7049 const MCInstrDesc *MCID = &MI->getDesc();
7050 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7051 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7052 // operand is still set to noreg. If needed, set the optional operand's
7053 // register to CPSR, and remove the redundant implicit def.
7055 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7057 // Rename pseudo opcodes.
7058 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7060 const ARMBaseInstrInfo *TII =
7061 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7062 MCID = &TII->get(NewOpc);
7064 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7065 "converted opcode should be the same except for cc_out");
7069 // Add the optional cc_out operand
7070 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7072 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7074 // Any ARM instruction that sets the 's' bit should specify an optional
7075 // "cc_out" operand in the last operand position.
7076 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7077 assert(!NewOpc && "Optional cc_out operand required");
7080 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7081 // since we already have an optional CPSR def.
7082 bool definesCPSR = false;
7083 bool deadCPSR = false;
7084 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7086 const MachineOperand &MO = MI->getOperand(i);
7087 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7091 MI->RemoveOperand(i);
7096 assert(!NewOpc && "Optional cc_out operand required");
7099 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7101 assert(!MI->getOperand(ccOutIdx).getReg() &&
7102 "expect uninitialized optional cc_out operand");
7106 // If this instruction was defined with an optional CPSR def and its dag node
7107 // had a live implicit CPSR def, then activate the optional CPSR def.
7108 MachineOperand &MO = MI->getOperand(ccOutIdx);
7109 MO.setReg(ARM::CPSR);
7113 //===----------------------------------------------------------------------===//
7114 // ARM Optimization Hooks
7115 //===----------------------------------------------------------------------===//
7117 // Helper function that checks if N is a null or all ones constant.
7118 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7119 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7122 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7125 // Return true if N is conditionally 0 or all ones.
7126 // Detects these expressions where cc is an i1 value:
7128 // (select cc 0, y) [AllOnes=0]
7129 // (select cc y, 0) [AllOnes=0]
7130 // (zext cc) [AllOnes=0]
7131 // (sext cc) [AllOnes=0/1]
7132 // (select cc -1, y) [AllOnes=1]
7133 // (select cc y, -1) [AllOnes=1]
7135 // Invert is set when N is the null/all ones constant when CC is false.
7136 // OtherOp is set to the alternative value of N.
7137 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7138 SDValue &CC, bool &Invert,
7140 SelectionDAG &DAG) {
7141 switch (N->getOpcode()) {
7142 default: return false;
7144 CC = N->getOperand(0);
7145 SDValue N1 = N->getOperand(1);
7146 SDValue N2 = N->getOperand(2);
7147 if (isZeroOrAllOnes(N1, AllOnes)) {
7152 if (isZeroOrAllOnes(N2, AllOnes)) {
7159 case ISD::ZERO_EXTEND:
7160 // (zext cc) can never be the all ones value.
7164 case ISD::SIGN_EXTEND: {
7165 EVT VT = N->getValueType(0);
7166 CC = N->getOperand(0);
7167 if (CC.getValueType() != MVT::i1)
7171 // When looking for an AllOnes constant, N is an sext, and the 'other'
7173 OtherOp = DAG.getConstant(0, VT);
7174 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7175 // When looking for a 0 constant, N can be zext or sext.
7176 OtherOp = DAG.getConstant(1, VT);
7178 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7184 // Combine a constant select operand into its use:
7186 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7187 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7188 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7189 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7190 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7192 // The transform is rejected if the select doesn't have a constant operand that
7193 // is null, or all ones when AllOnes is set.
7195 // Also recognize sext/zext from i1:
7197 // (add (zext cc), x) -> (select cc (add x, 1), x)
7198 // (add (sext cc), x) -> (select cc (add x, -1), x)
7200 // These transformations eventually create predicated instructions.
7202 // @param N The node to transform.
7203 // @param Slct The N operand that is a select.
7204 // @param OtherOp The other N operand (x above).
7205 // @param DCI Context.
7206 // @param AllOnes Require the select constant to be all ones instead of null.
7207 // @returns The new node, or SDValue() on failure.
7209 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7210 TargetLowering::DAGCombinerInfo &DCI,
7211 bool AllOnes = false) {
7212 SelectionDAG &DAG = DCI.DAG;
7213 EVT VT = N->getValueType(0);
7214 SDValue NonConstantVal;
7217 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7218 NonConstantVal, DAG))
7221 // Slct is now know to be the desired identity constant when CC is true.
7222 SDValue TrueVal = OtherOp;
7223 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7224 OtherOp, NonConstantVal);
7225 // Unless SwapSelectOps says CC should be false.
7227 std::swap(TrueVal, FalseVal);
7229 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7230 CCOp, TrueVal, FalseVal);
7233 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7235 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7236 TargetLowering::DAGCombinerInfo &DCI) {
7237 SDValue N0 = N->getOperand(0);
7238 SDValue N1 = N->getOperand(1);
7239 if (N0.getNode()->hasOneUse()) {
7240 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7241 if (Result.getNode())
7244 if (N1.getNode()->hasOneUse()) {
7245 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7246 if (Result.getNode())
7252 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7253 // (only after legalization).
7254 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7255 TargetLowering::DAGCombinerInfo &DCI,
7256 const ARMSubtarget *Subtarget) {
7258 // Only perform optimization if after legalize, and if NEON is available. We
7259 // also expected both operands to be BUILD_VECTORs.
7260 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7261 || N0.getOpcode() != ISD::BUILD_VECTOR
7262 || N1.getOpcode() != ISD::BUILD_VECTOR)
7265 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7266 EVT VT = N->getValueType(0);
7267 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7270 // Check that the vector operands are of the right form.
7271 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7272 // operands, where N is the size of the formed vector.
7273 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7274 // index such that we have a pair wise add pattern.
7276 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7277 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7279 SDValue Vec = N0->getOperand(0)->getOperand(0);
7280 SDNode *V = Vec.getNode();
7281 unsigned nextIndex = 0;
7283 // For each operands to the ADD which are BUILD_VECTORs,
7284 // check to see if each of their operands are an EXTRACT_VECTOR with
7285 // the same vector and appropriate index.
7286 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7287 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7288 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7290 SDValue ExtVec0 = N0->getOperand(i);
7291 SDValue ExtVec1 = N1->getOperand(i);
7293 // First operand is the vector, verify its the same.
7294 if (V != ExtVec0->getOperand(0).getNode() ||
7295 V != ExtVec1->getOperand(0).getNode())
7298 // Second is the constant, verify its correct.
7299 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7300 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7302 // For the constant, we want to see all the even or all the odd.
7303 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7304 || C1->getZExtValue() != nextIndex+1)
7313 // Create VPADDL node.
7314 SelectionDAG &DAG = DCI.DAG;
7315 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7317 // Build operand list.
7318 SmallVector<SDValue, 8> Ops;
7319 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7320 TLI.getPointerTy()));
7322 // Input is the vector.
7325 // Get widened type and narrowed type.
7327 unsigned numElem = VT.getVectorNumElements();
7328 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7329 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7330 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7331 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7333 llvm_unreachable("Invalid vector element type for padd optimization.");
7336 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7337 widenType, &Ops[0], Ops.size());
7338 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7341 static SDValue findMUL_LOHI(SDValue V) {
7342 if (V->getOpcode() == ISD::UMUL_LOHI ||
7343 V->getOpcode() == ISD::SMUL_LOHI)
7348 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7349 TargetLowering::DAGCombinerInfo &DCI,
7350 const ARMSubtarget *Subtarget) {
7352 if (Subtarget->isThumb1Only()) return SDValue();
7354 // Only perform the checks after legalize when the pattern is available.
7355 if (DCI.isBeforeLegalize()) return SDValue();
7357 // Look for multiply add opportunities.
7358 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7359 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7360 // a glue link from the first add to the second add.
7361 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7362 // a S/UMLAL instruction.
7365 // \ / \ [no multiline comment]
7371 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7372 SDValue AddcOp0 = AddcNode->getOperand(0);
7373 SDValue AddcOp1 = AddcNode->getOperand(1);
7375 // Check if the two operands are from the same mul_lohi node.
7376 if (AddcOp0.getNode() == AddcOp1.getNode())
7379 assert(AddcNode->getNumValues() == 2 &&
7380 AddcNode->getValueType(0) == MVT::i32 &&
7381 AddcNode->getValueType(1) == MVT::Glue &&
7382 "Expect ADDC with two result values: i32, glue");
7384 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7385 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7386 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7387 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7388 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7391 // Look for the glued ADDE.
7392 SDNode* AddeNode = AddcNode->getGluedUser();
7393 if (AddeNode == NULL)
7396 // Make sure it is really an ADDE.
7397 if (AddeNode->getOpcode() != ISD::ADDE)
7400 assert(AddeNode->getNumOperands() == 3 &&
7401 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7402 "ADDE node has the wrong inputs");
7404 // Check for the triangle shape.
7405 SDValue AddeOp0 = AddeNode->getOperand(0);
7406 SDValue AddeOp1 = AddeNode->getOperand(1);
7408 // Make sure that the ADDE operands are not coming from the same node.
7409 if (AddeOp0.getNode() == AddeOp1.getNode())
7412 // Find the MUL_LOHI node walking up ADDE's operands.
7413 bool IsLeftOperandMUL = false;
7414 SDValue MULOp = findMUL_LOHI(AddeOp0);
7415 if (MULOp == SDValue())
7416 MULOp = findMUL_LOHI(AddeOp1);
7418 IsLeftOperandMUL = true;
7419 if (MULOp == SDValue())
7422 // Figure out the right opcode.
7423 unsigned Opc = MULOp->getOpcode();
7424 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7426 // Figure out the high and low input values to the MLAL node.
7427 SDValue* HiMul = &MULOp;
7428 SDValue* HiAdd = NULL;
7429 SDValue* LoMul = NULL;
7430 SDValue* LowAdd = NULL;
7432 if (IsLeftOperandMUL)
7438 if (AddcOp0->getOpcode() == Opc) {
7442 if (AddcOp1->getOpcode() == Opc) {
7450 if (LoMul->getNode() != HiMul->getNode())
7453 // Create the merged node.
7454 SelectionDAG &DAG = DCI.DAG;
7456 // Build operand list.
7457 SmallVector<SDValue, 8> Ops;
7458 Ops.push_back(LoMul->getOperand(0));
7459 Ops.push_back(LoMul->getOperand(1));
7460 Ops.push_back(*LowAdd);
7461 Ops.push_back(*HiAdd);
7463 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7464 DAG.getVTList(MVT::i32, MVT::i32),
7465 &Ops[0], Ops.size());
7467 // Replace the ADDs' nodes uses by the MLA node's values.
7468 SDValue HiMLALResult(MLALNode.getNode(), 1);
7469 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7471 SDValue LoMLALResult(MLALNode.getNode(), 0);
7472 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7474 // Return original node to notify the driver to stop replacing.
7475 SDValue resNode(AddcNode, 0);
7479 /// PerformADDCCombine - Target-specific dag combine transform from
7480 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7481 static SDValue PerformADDCCombine(SDNode *N,
7482 TargetLowering::DAGCombinerInfo &DCI,
7483 const ARMSubtarget *Subtarget) {
7485 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7489 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7490 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7491 /// called with the default operands, and if that fails, with commuted
7493 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7494 TargetLowering::DAGCombinerInfo &DCI,
7495 const ARMSubtarget *Subtarget){
7497 // Attempt to create vpaddl for this add.
7498 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7499 if (Result.getNode())
7502 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7503 if (N0.getNode()->hasOneUse()) {
7504 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7505 if (Result.getNode()) return Result;
7510 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7512 static SDValue PerformADDCombine(SDNode *N,
7513 TargetLowering::DAGCombinerInfo &DCI,
7514 const ARMSubtarget *Subtarget) {
7515 SDValue N0 = N->getOperand(0);
7516 SDValue N1 = N->getOperand(1);
7518 // First try with the default operand order.
7519 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7520 if (Result.getNode())
7523 // If that didn't work, try again with the operands commuted.
7524 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7527 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7529 static SDValue PerformSUBCombine(SDNode *N,
7530 TargetLowering::DAGCombinerInfo &DCI) {
7531 SDValue N0 = N->getOperand(0);
7532 SDValue N1 = N->getOperand(1);
7534 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7535 if (N1.getNode()->hasOneUse()) {
7536 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7537 if (Result.getNode()) return Result;
7543 /// PerformVMULCombine
7544 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7545 /// special multiplier accumulator forwarding.
7551 static SDValue PerformVMULCombine(SDNode *N,
7552 TargetLowering::DAGCombinerInfo &DCI,
7553 const ARMSubtarget *Subtarget) {
7554 if (!Subtarget->hasVMLxForwarding())
7557 SelectionDAG &DAG = DCI.DAG;
7558 SDValue N0 = N->getOperand(0);
7559 SDValue N1 = N->getOperand(1);
7560 unsigned Opcode = N0.getOpcode();
7561 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7562 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7563 Opcode = N1.getOpcode();
7564 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7565 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7570 EVT VT = N->getValueType(0);
7571 DebugLoc DL = N->getDebugLoc();
7572 SDValue N00 = N0->getOperand(0);
7573 SDValue N01 = N0->getOperand(1);
7574 return DAG.getNode(Opcode, DL, VT,
7575 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7576 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7579 static SDValue PerformMULCombine(SDNode *N,
7580 TargetLowering::DAGCombinerInfo &DCI,
7581 const ARMSubtarget *Subtarget) {
7582 SelectionDAG &DAG = DCI.DAG;
7584 if (Subtarget->isThumb1Only())
7587 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7590 EVT VT = N->getValueType(0);
7591 if (VT.is64BitVector() || VT.is128BitVector())
7592 return PerformVMULCombine(N, DCI, Subtarget);
7596 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7600 int64_t MulAmt = C->getSExtValue();
7601 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7603 ShiftAmt = ShiftAmt & (32 - 1);
7604 SDValue V = N->getOperand(0);
7605 DebugLoc DL = N->getDebugLoc();
7608 MulAmt >>= ShiftAmt;
7611 if (isPowerOf2_32(MulAmt - 1)) {
7612 // (mul x, 2^N + 1) => (add (shl x, N), x)
7613 Res = DAG.getNode(ISD::ADD, DL, VT,
7615 DAG.getNode(ISD::SHL, DL, VT,
7617 DAG.getConstant(Log2_32(MulAmt - 1),
7619 } else if (isPowerOf2_32(MulAmt + 1)) {
7620 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7621 Res = DAG.getNode(ISD::SUB, DL, VT,
7622 DAG.getNode(ISD::SHL, DL, VT,
7624 DAG.getConstant(Log2_32(MulAmt + 1),
7630 uint64_t MulAmtAbs = -MulAmt;
7631 if (isPowerOf2_32(MulAmtAbs + 1)) {
7632 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7633 Res = DAG.getNode(ISD::SUB, DL, VT,
7635 DAG.getNode(ISD::SHL, DL, VT,
7637 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7639 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7640 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7641 Res = DAG.getNode(ISD::ADD, DL, VT,
7643 DAG.getNode(ISD::SHL, DL, VT,
7645 DAG.getConstant(Log2_32(MulAmtAbs-1),
7647 Res = DAG.getNode(ISD::SUB, DL, VT,
7648 DAG.getConstant(0, MVT::i32),Res);
7655 Res = DAG.getNode(ISD::SHL, DL, VT,
7656 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7658 // Do not add new nodes to DAG combiner worklist.
7659 DCI.CombineTo(N, Res, false);
7663 static SDValue PerformANDCombine(SDNode *N,
7664 TargetLowering::DAGCombinerInfo &DCI,
7665 const ARMSubtarget *Subtarget) {
7667 // Attempt to use immediate-form VBIC
7668 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7669 DebugLoc dl = N->getDebugLoc();
7670 EVT VT = N->getValueType(0);
7671 SelectionDAG &DAG = DCI.DAG;
7673 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7676 APInt SplatBits, SplatUndef;
7677 unsigned SplatBitSize;
7680 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7681 if (SplatBitSize <= 64) {
7683 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7684 SplatUndef.getZExtValue(), SplatBitSize,
7685 DAG, VbicVT, VT.is128BitVector(),
7687 if (Val.getNode()) {
7689 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7690 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7691 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7696 if (!Subtarget->isThumb1Only()) {
7697 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7698 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7699 if (Result.getNode())
7706 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7707 static SDValue PerformORCombine(SDNode *N,
7708 TargetLowering::DAGCombinerInfo &DCI,
7709 const ARMSubtarget *Subtarget) {
7710 // Attempt to use immediate-form VORR
7711 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7712 DebugLoc dl = N->getDebugLoc();
7713 EVT VT = N->getValueType(0);
7714 SelectionDAG &DAG = DCI.DAG;
7716 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7719 APInt SplatBits, SplatUndef;
7720 unsigned SplatBitSize;
7722 if (BVN && Subtarget->hasNEON() &&
7723 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7724 if (SplatBitSize <= 64) {
7726 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7727 SplatUndef.getZExtValue(), SplatBitSize,
7728 DAG, VorrVT, VT.is128BitVector(),
7730 if (Val.getNode()) {
7732 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7733 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7734 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7739 if (!Subtarget->isThumb1Only()) {
7740 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7741 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7742 if (Result.getNode())
7746 // The code below optimizes (or (and X, Y), Z).
7747 // The AND operand needs to have a single user to make these optimizations
7749 SDValue N0 = N->getOperand(0);
7750 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7752 SDValue N1 = N->getOperand(1);
7754 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7755 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7756 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7758 unsigned SplatBitSize;
7761 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7763 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7764 HasAnyUndefs) && !HasAnyUndefs) {
7765 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7767 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7768 HasAnyUndefs) && !HasAnyUndefs &&
7769 SplatBits0 == ~SplatBits1) {
7770 // Canonicalize the vector type to make instruction selection simpler.
7771 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7772 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7773 N0->getOperand(1), N0->getOperand(0),
7775 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7780 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7783 // BFI is only available on V6T2+
7784 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7787 DebugLoc DL = N->getDebugLoc();
7788 // 1) or (and A, mask), val => ARMbfi A, val, mask
7789 // iff (val & mask) == val
7791 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7792 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7793 // && mask == ~mask2
7794 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7795 // && ~mask == mask2
7796 // (i.e., copy a bitfield value into another bitfield of the same width)
7801 SDValue N00 = N0.getOperand(0);
7803 // The value and the mask need to be constants so we can verify this is
7804 // actually a bitfield set. If the mask is 0xffff, we can do better
7805 // via a movt instruction, so don't use BFI in that case.
7806 SDValue MaskOp = N0.getOperand(1);
7807 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7810 unsigned Mask = MaskC->getZExtValue();
7814 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7815 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7817 unsigned Val = N1C->getZExtValue();
7818 if ((Val & ~Mask) != Val)
7821 if (ARM::isBitFieldInvertedMask(Mask)) {
7822 Val >>= CountTrailingZeros_32(~Mask);
7824 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7825 DAG.getConstant(Val, MVT::i32),
7826 DAG.getConstant(Mask, MVT::i32));
7828 // Do not add new nodes to DAG combiner worklist.
7829 DCI.CombineTo(N, Res, false);
7832 } else if (N1.getOpcode() == ISD::AND) {
7833 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7834 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7837 unsigned Mask2 = N11C->getZExtValue();
7839 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7841 if (ARM::isBitFieldInvertedMask(Mask) &&
7843 // The pack halfword instruction works better for masks that fit it,
7844 // so use that when it's available.
7845 if (Subtarget->hasT2ExtractPack() &&
7846 (Mask == 0xffff || Mask == 0xffff0000))
7849 unsigned amt = CountTrailingZeros_32(Mask2);
7850 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7851 DAG.getConstant(amt, MVT::i32));
7852 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7853 DAG.getConstant(Mask, MVT::i32));
7854 // Do not add new nodes to DAG combiner worklist.
7855 DCI.CombineTo(N, Res, false);
7857 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7859 // The pack halfword instruction works better for masks that fit it,
7860 // so use that when it's available.
7861 if (Subtarget->hasT2ExtractPack() &&
7862 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7865 unsigned lsb = CountTrailingZeros_32(Mask);
7866 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7867 DAG.getConstant(lsb, MVT::i32));
7868 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7869 DAG.getConstant(Mask2, MVT::i32));
7870 // Do not add new nodes to DAG combiner worklist.
7871 DCI.CombineTo(N, Res, false);
7876 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7877 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7878 ARM::isBitFieldInvertedMask(~Mask)) {
7879 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7880 // where lsb(mask) == #shamt and masked bits of B are known zero.
7881 SDValue ShAmt = N00.getOperand(1);
7882 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7883 unsigned LSB = CountTrailingZeros_32(Mask);
7887 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7888 DAG.getConstant(~Mask, MVT::i32));
7890 // Do not add new nodes to DAG combiner worklist.
7891 DCI.CombineTo(N, Res, false);
7897 static SDValue PerformXORCombine(SDNode *N,
7898 TargetLowering::DAGCombinerInfo &DCI,
7899 const ARMSubtarget *Subtarget) {
7900 EVT VT = N->getValueType(0);
7901 SelectionDAG &DAG = DCI.DAG;
7903 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7906 if (!Subtarget->isThumb1Only()) {
7907 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7908 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7909 if (Result.getNode())
7916 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7917 /// the bits being cleared by the AND are not demanded by the BFI.
7918 static SDValue PerformBFICombine(SDNode *N,
7919 TargetLowering::DAGCombinerInfo &DCI) {
7920 SDValue N1 = N->getOperand(1);
7921 if (N1.getOpcode() == ISD::AND) {
7922 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7925 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7926 unsigned LSB = CountTrailingZeros_32(~InvMask);
7927 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7928 unsigned Mask = (1 << Width)-1;
7929 unsigned Mask2 = N11C->getZExtValue();
7930 if ((Mask & (~Mask2)) == 0)
7931 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7932 N->getOperand(0), N1.getOperand(0),
7938 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7939 /// ARMISD::VMOVRRD.
7940 static SDValue PerformVMOVRRDCombine(SDNode *N,
7941 TargetLowering::DAGCombinerInfo &DCI) {
7942 // vmovrrd(vmovdrr x, y) -> x,y
7943 SDValue InDouble = N->getOperand(0);
7944 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7945 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7947 // vmovrrd(load f64) -> (load i32), (load i32)
7948 SDNode *InNode = InDouble.getNode();
7949 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7950 InNode->getValueType(0) == MVT::f64 &&
7951 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7952 !cast<LoadSDNode>(InNode)->isVolatile()) {
7953 // TODO: Should this be done for non-FrameIndex operands?
7954 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7956 SelectionDAG &DAG = DCI.DAG;
7957 DebugLoc DL = LD->getDebugLoc();
7958 SDValue BasePtr = LD->getBasePtr();
7959 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7960 LD->getPointerInfo(), LD->isVolatile(),
7961 LD->isNonTemporal(), LD->isInvariant(),
7962 LD->getAlignment());
7964 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7965 DAG.getConstant(4, MVT::i32));
7966 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7967 LD->getPointerInfo(), LD->isVolatile(),
7968 LD->isNonTemporal(), LD->isInvariant(),
7969 std::min(4U, LD->getAlignment() / 2));
7971 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7972 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7973 DCI.RemoveFromWorklist(LD);
7981 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7982 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7983 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7984 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7985 SDValue Op0 = N->getOperand(0);
7986 SDValue Op1 = N->getOperand(1);
7987 if (Op0.getOpcode() == ISD::BITCAST)
7988 Op0 = Op0.getOperand(0);
7989 if (Op1.getOpcode() == ISD::BITCAST)
7990 Op1 = Op1.getOperand(0);
7991 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7992 Op0.getNode() == Op1.getNode() &&
7993 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7994 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7995 N->getValueType(0), Op0.getOperand(0));
7999 /// PerformSTORECombine - Target-specific dag combine xforms for
8001 static SDValue PerformSTORECombine(SDNode *N,
8002 TargetLowering::DAGCombinerInfo &DCI) {
8003 StoreSDNode *St = cast<StoreSDNode>(N);
8004 if (St->isVolatile())
8007 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8008 // pack all of the elements in one place. Next, store to memory in fewer
8010 SDValue StVal = St->getValue();
8011 EVT VT = StVal.getValueType();
8012 if (St->isTruncatingStore() && VT.isVector()) {
8013 SelectionDAG &DAG = DCI.DAG;
8014 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8015 EVT StVT = St->getMemoryVT();
8016 unsigned NumElems = VT.getVectorNumElements();
8017 assert(StVT != VT && "Cannot truncate to the same type");
8018 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8019 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8021 // From, To sizes and ElemCount must be pow of two
8022 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8024 // We are going to use the original vector elt for storing.
8025 // Accumulated smaller vector elements must be a multiple of the store size.
8026 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8028 unsigned SizeRatio = FromEltSz / ToEltSz;
8029 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8031 // Create a type on which we perform the shuffle.
8032 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8033 NumElems*SizeRatio);
8034 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8036 DebugLoc DL = St->getDebugLoc();
8037 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8038 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8039 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8041 // Can't shuffle using an illegal type.
8042 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8044 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8045 DAG.getUNDEF(WideVec.getValueType()),
8047 // At this point all of the data is stored at the bottom of the
8048 // register. We now need to save it to mem.
8050 // Find the largest store unit
8051 MVT StoreType = MVT::i8;
8052 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8053 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8054 MVT Tp = (MVT::SimpleValueType)tp;
8055 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8058 // Didn't find a legal store type.
8059 if (!TLI.isTypeLegal(StoreType))
8062 // Bitcast the original vector into a vector of store-size units
8063 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8064 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8065 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8066 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8067 SmallVector<SDValue, 8> Chains;
8068 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8069 TLI.getPointerTy());
8070 SDValue BasePtr = St->getBasePtr();
8072 // Perform one or more big stores into memory.
8073 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8074 for (unsigned I = 0; I < E; I++) {
8075 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8076 StoreType, ShuffWide,
8077 DAG.getIntPtrConstant(I));
8078 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8079 St->getPointerInfo(), St->isVolatile(),
8080 St->isNonTemporal(), St->getAlignment());
8081 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8083 Chains.push_back(Ch);
8085 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8089 if (!ISD::isNormalStore(St))
8092 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8093 // ARM stores of arguments in the same cache line.
8094 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8095 StVal.getNode()->hasOneUse()) {
8096 SelectionDAG &DAG = DCI.DAG;
8097 DebugLoc DL = St->getDebugLoc();
8098 SDValue BasePtr = St->getBasePtr();
8099 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8100 StVal.getNode()->getOperand(0), BasePtr,
8101 St->getPointerInfo(), St->isVolatile(),
8102 St->isNonTemporal(), St->getAlignment());
8104 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8105 DAG.getConstant(4, MVT::i32));
8106 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8107 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8108 St->isNonTemporal(),
8109 std::min(4U, St->getAlignment() / 2));
8112 if (StVal.getValueType() != MVT::i64 ||
8113 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8116 // Bitcast an i64 store extracted from a vector to f64.
8117 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8118 SelectionDAG &DAG = DCI.DAG;
8119 DebugLoc dl = StVal.getDebugLoc();
8120 SDValue IntVec = StVal.getOperand(0);
8121 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8122 IntVec.getValueType().getVectorNumElements());
8123 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8124 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8125 Vec, StVal.getOperand(1));
8126 dl = N->getDebugLoc();
8127 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8128 // Make the DAGCombiner fold the bitcasts.
8129 DCI.AddToWorklist(Vec.getNode());
8130 DCI.AddToWorklist(ExtElt.getNode());
8131 DCI.AddToWorklist(V.getNode());
8132 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8133 St->getPointerInfo(), St->isVolatile(),
8134 St->isNonTemporal(), St->getAlignment(),
8138 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8139 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8140 /// i64 vector to have f64 elements, since the value can then be loaded
8141 /// directly into a VFP register.
8142 static bool hasNormalLoadOperand(SDNode *N) {
8143 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8144 for (unsigned i = 0; i < NumElts; ++i) {
8145 SDNode *Elt = N->getOperand(i).getNode();
8146 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8152 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8153 /// ISD::BUILD_VECTOR.
8154 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8155 TargetLowering::DAGCombinerInfo &DCI){
8156 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8157 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8158 // into a pair of GPRs, which is fine when the value is used as a scalar,
8159 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8160 SelectionDAG &DAG = DCI.DAG;
8161 if (N->getNumOperands() == 2) {
8162 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8167 // Load i64 elements as f64 values so that type legalization does not split
8168 // them up into i32 values.
8169 EVT VT = N->getValueType(0);
8170 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8172 DebugLoc dl = N->getDebugLoc();
8173 SmallVector<SDValue, 8> Ops;
8174 unsigned NumElts = VT.getVectorNumElements();
8175 for (unsigned i = 0; i < NumElts; ++i) {
8176 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8178 // Make the DAGCombiner fold the bitcast.
8179 DCI.AddToWorklist(V.getNode());
8181 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8182 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8183 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8186 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8187 /// ISD::INSERT_VECTOR_ELT.
8188 static SDValue PerformInsertEltCombine(SDNode *N,
8189 TargetLowering::DAGCombinerInfo &DCI) {
8190 // Bitcast an i64 load inserted into a vector to f64.
8191 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8192 EVT VT = N->getValueType(0);
8193 SDNode *Elt = N->getOperand(1).getNode();
8194 if (VT.getVectorElementType() != MVT::i64 ||
8195 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8198 SelectionDAG &DAG = DCI.DAG;
8199 DebugLoc dl = N->getDebugLoc();
8200 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8201 VT.getVectorNumElements());
8202 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8203 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8204 // Make the DAGCombiner fold the bitcasts.
8205 DCI.AddToWorklist(Vec.getNode());
8206 DCI.AddToWorklist(V.getNode());
8207 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8208 Vec, V, N->getOperand(2));
8209 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8212 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8213 /// ISD::VECTOR_SHUFFLE.
8214 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8215 // The LLVM shufflevector instruction does not require the shuffle mask
8216 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8217 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8218 // operands do not match the mask length, they are extended by concatenating
8219 // them with undef vectors. That is probably the right thing for other
8220 // targets, but for NEON it is better to concatenate two double-register
8221 // size vector operands into a single quad-register size vector. Do that
8222 // transformation here:
8223 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8224 // shuffle(concat(v1, v2), undef)
8225 SDValue Op0 = N->getOperand(0);
8226 SDValue Op1 = N->getOperand(1);
8227 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8228 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8229 Op0.getNumOperands() != 2 ||
8230 Op1.getNumOperands() != 2)
8232 SDValue Concat0Op1 = Op0.getOperand(1);
8233 SDValue Concat1Op1 = Op1.getOperand(1);
8234 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8235 Concat1Op1.getOpcode() != ISD::UNDEF)
8237 // Skip the transformation if any of the types are illegal.
8238 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8239 EVT VT = N->getValueType(0);
8240 if (!TLI.isTypeLegal(VT) ||
8241 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8242 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8245 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8246 Op0.getOperand(0), Op1.getOperand(0));
8247 // Translate the shuffle mask.
8248 SmallVector<int, 16> NewMask;
8249 unsigned NumElts = VT.getVectorNumElements();
8250 unsigned HalfElts = NumElts/2;
8251 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8252 for (unsigned n = 0; n < NumElts; ++n) {
8253 int MaskElt = SVN->getMaskElt(n);
8255 if (MaskElt < (int)HalfElts)
8257 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8258 NewElt = HalfElts + MaskElt - NumElts;
8259 NewMask.push_back(NewElt);
8261 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8262 DAG.getUNDEF(VT), NewMask.data());
8265 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8266 /// NEON load/store intrinsics to merge base address updates.
8267 static SDValue CombineBaseUpdate(SDNode *N,
8268 TargetLowering::DAGCombinerInfo &DCI) {
8269 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8272 SelectionDAG &DAG = DCI.DAG;
8273 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8274 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8275 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8276 SDValue Addr = N->getOperand(AddrOpIdx);
8278 // Search for a use of the address operand that is an increment.
8279 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8280 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8282 if (User->getOpcode() != ISD::ADD ||
8283 UI.getUse().getResNo() != Addr.getResNo())
8286 // Check that the add is independent of the load/store. Otherwise, folding
8287 // it would create a cycle.
8288 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8291 // Find the new opcode for the updating load/store.
8293 bool isLaneOp = false;
8294 unsigned NewOpc = 0;
8295 unsigned NumVecs = 0;
8297 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8299 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8300 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8302 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8304 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8306 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8308 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8309 NumVecs = 2; isLaneOp = true; break;
8310 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8311 NumVecs = 3; isLaneOp = true; break;
8312 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8313 NumVecs = 4; isLaneOp = true; break;
8314 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8315 NumVecs = 1; isLoad = false; break;
8316 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8317 NumVecs = 2; isLoad = false; break;
8318 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8319 NumVecs = 3; isLoad = false; break;
8320 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8321 NumVecs = 4; isLoad = false; break;
8322 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8323 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8324 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8325 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8326 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8327 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8331 switch (N->getOpcode()) {
8332 default: llvm_unreachable("unexpected opcode for Neon base update");
8333 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8334 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8335 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8339 // Find the size of memory referenced by the load/store.
8342 VecTy = N->getValueType(0);
8344 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8345 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8347 NumBytes /= VecTy.getVectorNumElements();
8349 // If the increment is a constant, it must match the memory ref size.
8350 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8351 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8352 uint64_t IncVal = CInc->getZExtValue();
8353 if (IncVal != NumBytes)
8355 } else if (NumBytes >= 3 * 16) {
8356 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8357 // separate instructions that make it harder to use a non-constant update.
8361 // Create the new updating load/store node.
8363 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8365 for (n = 0; n < NumResultVecs; ++n)
8367 Tys[n++] = MVT::i32;
8368 Tys[n] = MVT::Other;
8369 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8370 SmallVector<SDValue, 8> Ops;
8371 Ops.push_back(N->getOperand(0)); // incoming chain
8372 Ops.push_back(N->getOperand(AddrOpIdx));
8374 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8375 Ops.push_back(N->getOperand(i));
8377 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8378 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8379 Ops.data(), Ops.size(),
8380 MemInt->getMemoryVT(),
8381 MemInt->getMemOperand());
8384 std::vector<SDValue> NewResults;
8385 for (unsigned i = 0; i < NumResultVecs; ++i) {
8386 NewResults.push_back(SDValue(UpdN.getNode(), i));
8388 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8389 DCI.CombineTo(N, NewResults);
8390 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8397 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8398 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8399 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8401 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8402 SelectionDAG &DAG = DCI.DAG;
8403 EVT VT = N->getValueType(0);
8404 // vldN-dup instructions only support 64-bit vectors for N > 1.
8405 if (!VT.is64BitVector())
8408 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8409 SDNode *VLD = N->getOperand(0).getNode();
8410 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8412 unsigned NumVecs = 0;
8413 unsigned NewOpc = 0;
8414 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8415 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8417 NewOpc = ARMISD::VLD2DUP;
8418 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8420 NewOpc = ARMISD::VLD3DUP;
8421 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8423 NewOpc = ARMISD::VLD4DUP;
8428 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8429 // numbers match the load.
8430 unsigned VLDLaneNo =
8431 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8432 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8434 // Ignore uses of the chain result.
8435 if (UI.getUse().getResNo() == NumVecs)
8438 if (User->getOpcode() != ARMISD::VDUPLANE ||
8439 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8443 // Create the vldN-dup node.
8446 for (n = 0; n < NumVecs; ++n)
8448 Tys[n] = MVT::Other;
8449 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8450 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8451 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8452 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8453 Ops, 2, VLDMemInt->getMemoryVT(),
8454 VLDMemInt->getMemOperand());
8457 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8459 unsigned ResNo = UI.getUse().getResNo();
8460 // Ignore uses of the chain result.
8461 if (ResNo == NumVecs)
8464 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8467 // Now the vldN-lane intrinsic is dead except for its chain result.
8468 // Update uses of the chain.
8469 std::vector<SDValue> VLDDupResults;
8470 for (unsigned n = 0; n < NumVecs; ++n)
8471 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8472 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8473 DCI.CombineTo(VLD, VLDDupResults);
8478 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8479 /// ARMISD::VDUPLANE.
8480 static SDValue PerformVDUPLANECombine(SDNode *N,
8481 TargetLowering::DAGCombinerInfo &DCI) {
8482 SDValue Op = N->getOperand(0);
8484 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8485 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8486 if (CombineVLDDUP(N, DCI))
8487 return SDValue(N, 0);
8489 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8490 // redundant. Ignore bit_converts for now; element sizes are checked below.
8491 while (Op.getOpcode() == ISD::BITCAST)
8492 Op = Op.getOperand(0);
8493 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8496 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8497 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8498 // The canonical VMOV for a zero vector uses a 32-bit element size.
8499 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8501 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8503 EVT VT = N->getValueType(0);
8504 if (EltSize > VT.getVectorElementType().getSizeInBits())
8507 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8510 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8511 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8512 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8516 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8518 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8523 APFloat APF = C->getValueAPF();
8524 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8525 != APFloat::opOK || !isExact)
8528 c0 = (I == 0) ? cN : c0;
8529 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8536 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8537 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8538 /// when the VMUL has a constant operand that is a power of 2.
8540 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8541 /// vmul.f32 d16, d17, d16
8542 /// vcvt.s32.f32 d16, d16
8544 /// vcvt.s32.f32 d16, d16, #3
8545 static SDValue PerformVCVTCombine(SDNode *N,
8546 TargetLowering::DAGCombinerInfo &DCI,
8547 const ARMSubtarget *Subtarget) {
8548 SelectionDAG &DAG = DCI.DAG;
8549 SDValue Op = N->getOperand(0);
8551 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8552 Op.getOpcode() != ISD::FMUL)
8556 SDValue N0 = Op->getOperand(0);
8557 SDValue ConstVec = Op->getOperand(1);
8558 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8560 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8561 !isConstVecPow2(ConstVec, isSigned, C))
8564 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8565 Intrinsic::arm_neon_vcvtfp2fxu;
8566 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8568 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8569 DAG.getConstant(Log2_64(C), MVT::i32));
8572 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8573 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8574 /// when the VDIV has a constant operand that is a power of 2.
8576 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8577 /// vcvt.f32.s32 d16, d16
8578 /// vdiv.f32 d16, d17, d16
8580 /// vcvt.f32.s32 d16, d16, #3
8581 static SDValue PerformVDIVCombine(SDNode *N,
8582 TargetLowering::DAGCombinerInfo &DCI,
8583 const ARMSubtarget *Subtarget) {
8584 SelectionDAG &DAG = DCI.DAG;
8585 SDValue Op = N->getOperand(0);
8586 unsigned OpOpcode = Op.getNode()->getOpcode();
8588 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8589 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8593 SDValue ConstVec = N->getOperand(1);
8594 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8596 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8597 !isConstVecPow2(ConstVec, isSigned, C))
8600 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8601 Intrinsic::arm_neon_vcvtfxu2fp;
8602 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8604 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8605 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8608 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8609 /// operand of a vector shift operation, where all the elements of the
8610 /// build_vector must have the same constant integer value.
8611 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8612 // Ignore bit_converts.
8613 while (Op.getOpcode() == ISD::BITCAST)
8614 Op = Op.getOperand(0);
8615 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8616 APInt SplatBits, SplatUndef;
8617 unsigned SplatBitSize;
8619 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8620 HasAnyUndefs, ElementBits) ||
8621 SplatBitSize > ElementBits)
8623 Cnt = SplatBits.getSExtValue();
8627 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8628 /// operand of a vector shift left operation. That value must be in the range:
8629 /// 0 <= Value < ElementBits for a left shift; or
8630 /// 0 <= Value <= ElementBits for a long left shift.
8631 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8632 assert(VT.isVector() && "vector shift count is not a vector type");
8633 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8634 if (! getVShiftImm(Op, ElementBits, Cnt))
8636 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8639 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8640 /// operand of a vector shift right operation. For a shift opcode, the value
8641 /// is positive, but for an intrinsic the value count must be negative. The
8642 /// absolute value must be in the range:
8643 /// 1 <= |Value| <= ElementBits for a right shift; or
8644 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8645 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8647 assert(VT.isVector() && "vector shift count is not a vector type");
8648 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8649 if (! getVShiftImm(Op, ElementBits, Cnt))
8653 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8656 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8657 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8658 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8661 // Don't do anything for most intrinsics.
8664 // Vector shifts: check for immediate versions and lower them.
8665 // Note: This is done during DAG combining instead of DAG legalizing because
8666 // the build_vectors for 64-bit vector element shift counts are generally
8667 // not legal, and it is hard to see their values after they get legalized to
8668 // loads from a constant pool.
8669 case Intrinsic::arm_neon_vshifts:
8670 case Intrinsic::arm_neon_vshiftu:
8671 case Intrinsic::arm_neon_vshiftls:
8672 case Intrinsic::arm_neon_vshiftlu:
8673 case Intrinsic::arm_neon_vshiftn:
8674 case Intrinsic::arm_neon_vrshifts:
8675 case Intrinsic::arm_neon_vrshiftu:
8676 case Intrinsic::arm_neon_vrshiftn:
8677 case Intrinsic::arm_neon_vqshifts:
8678 case Intrinsic::arm_neon_vqshiftu:
8679 case Intrinsic::arm_neon_vqshiftsu:
8680 case Intrinsic::arm_neon_vqshiftns:
8681 case Intrinsic::arm_neon_vqshiftnu:
8682 case Intrinsic::arm_neon_vqshiftnsu:
8683 case Intrinsic::arm_neon_vqrshiftns:
8684 case Intrinsic::arm_neon_vqrshiftnu:
8685 case Intrinsic::arm_neon_vqrshiftnsu: {
8686 EVT VT = N->getOperand(1).getValueType();
8688 unsigned VShiftOpc = 0;
8691 case Intrinsic::arm_neon_vshifts:
8692 case Intrinsic::arm_neon_vshiftu:
8693 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8694 VShiftOpc = ARMISD::VSHL;
8697 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8698 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8699 ARMISD::VSHRs : ARMISD::VSHRu);
8704 case Intrinsic::arm_neon_vshiftls:
8705 case Intrinsic::arm_neon_vshiftlu:
8706 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8708 llvm_unreachable("invalid shift count for vshll intrinsic");
8710 case Intrinsic::arm_neon_vrshifts:
8711 case Intrinsic::arm_neon_vrshiftu:
8712 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8716 case Intrinsic::arm_neon_vqshifts:
8717 case Intrinsic::arm_neon_vqshiftu:
8718 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8722 case Intrinsic::arm_neon_vqshiftsu:
8723 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8725 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8727 case Intrinsic::arm_neon_vshiftn:
8728 case Intrinsic::arm_neon_vrshiftn:
8729 case Intrinsic::arm_neon_vqshiftns:
8730 case Intrinsic::arm_neon_vqshiftnu:
8731 case Intrinsic::arm_neon_vqshiftnsu:
8732 case Intrinsic::arm_neon_vqrshiftns:
8733 case Intrinsic::arm_neon_vqrshiftnu:
8734 case Intrinsic::arm_neon_vqrshiftnsu:
8735 // Narrowing shifts require an immediate right shift.
8736 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8738 llvm_unreachable("invalid shift count for narrowing vector shift "
8742 llvm_unreachable("unhandled vector shift");
8746 case Intrinsic::arm_neon_vshifts:
8747 case Intrinsic::arm_neon_vshiftu:
8748 // Opcode already set above.
8750 case Intrinsic::arm_neon_vshiftls:
8751 case Intrinsic::arm_neon_vshiftlu:
8752 if (Cnt == VT.getVectorElementType().getSizeInBits())
8753 VShiftOpc = ARMISD::VSHLLi;
8755 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8756 ARMISD::VSHLLs : ARMISD::VSHLLu);
8758 case Intrinsic::arm_neon_vshiftn:
8759 VShiftOpc = ARMISD::VSHRN; break;
8760 case Intrinsic::arm_neon_vrshifts:
8761 VShiftOpc = ARMISD::VRSHRs; break;
8762 case Intrinsic::arm_neon_vrshiftu:
8763 VShiftOpc = ARMISD::VRSHRu; break;
8764 case Intrinsic::arm_neon_vrshiftn:
8765 VShiftOpc = ARMISD::VRSHRN; break;
8766 case Intrinsic::arm_neon_vqshifts:
8767 VShiftOpc = ARMISD::VQSHLs; break;
8768 case Intrinsic::arm_neon_vqshiftu:
8769 VShiftOpc = ARMISD::VQSHLu; break;
8770 case Intrinsic::arm_neon_vqshiftsu:
8771 VShiftOpc = ARMISD::VQSHLsu; break;
8772 case Intrinsic::arm_neon_vqshiftns:
8773 VShiftOpc = ARMISD::VQSHRNs; break;
8774 case Intrinsic::arm_neon_vqshiftnu:
8775 VShiftOpc = ARMISD::VQSHRNu; break;
8776 case Intrinsic::arm_neon_vqshiftnsu:
8777 VShiftOpc = ARMISD::VQSHRNsu; break;
8778 case Intrinsic::arm_neon_vqrshiftns:
8779 VShiftOpc = ARMISD::VQRSHRNs; break;
8780 case Intrinsic::arm_neon_vqrshiftnu:
8781 VShiftOpc = ARMISD::VQRSHRNu; break;
8782 case Intrinsic::arm_neon_vqrshiftnsu:
8783 VShiftOpc = ARMISD::VQRSHRNsu; break;
8786 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8787 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8790 case Intrinsic::arm_neon_vshiftins: {
8791 EVT VT = N->getOperand(1).getValueType();
8793 unsigned VShiftOpc = 0;
8795 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8796 VShiftOpc = ARMISD::VSLI;
8797 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8798 VShiftOpc = ARMISD::VSRI;
8800 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8803 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8804 N->getOperand(1), N->getOperand(2),
8805 DAG.getConstant(Cnt, MVT::i32));
8808 case Intrinsic::arm_neon_vqrshifts:
8809 case Intrinsic::arm_neon_vqrshiftu:
8810 // No immediate versions of these to check for.
8817 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8818 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8819 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8820 /// vector element shift counts are generally not legal, and it is hard to see
8821 /// their values after they get legalized to loads from a constant pool.
8822 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8823 const ARMSubtarget *ST) {
8824 EVT VT = N->getValueType(0);
8825 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8826 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8827 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8828 SDValue N1 = N->getOperand(1);
8829 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8830 SDValue N0 = N->getOperand(0);
8831 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8832 DAG.MaskedValueIsZero(N0.getOperand(0),
8833 APInt::getHighBitsSet(32, 16)))
8834 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8838 // Nothing to be done for scalar shifts.
8839 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8840 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8843 assert(ST->hasNEON() && "unexpected vector shift");
8846 switch (N->getOpcode()) {
8847 default: llvm_unreachable("unexpected shift opcode");
8850 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8851 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8852 DAG.getConstant(Cnt, MVT::i32));
8857 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8858 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8859 ARMISD::VSHRs : ARMISD::VSHRu);
8860 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8861 DAG.getConstant(Cnt, MVT::i32));
8867 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8868 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8869 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8870 const ARMSubtarget *ST) {
8871 SDValue N0 = N->getOperand(0);
8873 // Check for sign- and zero-extensions of vector extract operations of 8-
8874 // and 16-bit vector elements. NEON supports these directly. They are
8875 // handled during DAG combining because type legalization will promote them
8876 // to 32-bit types and it is messy to recognize the operations after that.
8877 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8878 SDValue Vec = N0.getOperand(0);
8879 SDValue Lane = N0.getOperand(1);
8880 EVT VT = N->getValueType(0);
8881 EVT EltVT = N0.getValueType();
8882 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8884 if (VT == MVT::i32 &&
8885 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8886 TLI.isTypeLegal(Vec.getValueType()) &&
8887 isa<ConstantSDNode>(Lane)) {
8890 switch (N->getOpcode()) {
8891 default: llvm_unreachable("unexpected opcode");
8892 case ISD::SIGN_EXTEND:
8893 Opc = ARMISD::VGETLANEs;
8895 case ISD::ZERO_EXTEND:
8896 case ISD::ANY_EXTEND:
8897 Opc = ARMISD::VGETLANEu;
8900 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8907 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8908 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8909 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8910 const ARMSubtarget *ST) {
8911 // If the target supports NEON, try to use vmax/vmin instructions for f32
8912 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8913 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8914 // a NaN; only do the transformation when it matches that behavior.
8916 // For now only do this when using NEON for FP operations; if using VFP, it
8917 // is not obvious that the benefit outweighs the cost of switching to the
8919 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8920 N->getValueType(0) != MVT::f32)
8923 SDValue CondLHS = N->getOperand(0);
8924 SDValue CondRHS = N->getOperand(1);
8925 SDValue LHS = N->getOperand(2);
8926 SDValue RHS = N->getOperand(3);
8927 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8929 unsigned Opcode = 0;
8931 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8932 IsReversed = false; // x CC y ? x : y
8933 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8934 IsReversed = true ; // x CC y ? y : x
8948 // If LHS is NaN, an ordered comparison will be false and the result will
8949 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8950 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8951 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8952 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8954 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8955 // will return -0, so vmin can only be used for unsafe math or if one of
8956 // the operands is known to be nonzero.
8957 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8958 !DAG.getTarget().Options.UnsafeFPMath &&
8959 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8961 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8970 // If LHS is NaN, an ordered comparison will be false and the result will
8971 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8972 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8973 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8974 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8976 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8977 // will return +0, so vmax can only be used for unsafe math or if one of
8978 // the operands is known to be nonzero.
8979 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8980 !DAG.getTarget().Options.UnsafeFPMath &&
8981 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8983 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8989 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8992 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8994 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8995 SDValue Cmp = N->getOperand(4);
8996 if (Cmp.getOpcode() != ARMISD::CMPZ)
8997 // Only looking at EQ and NE cases.
9000 EVT VT = N->getValueType(0);
9001 DebugLoc dl = N->getDebugLoc();
9002 SDValue LHS = Cmp.getOperand(0);
9003 SDValue RHS = Cmp.getOperand(1);
9004 SDValue FalseVal = N->getOperand(0);
9005 SDValue TrueVal = N->getOperand(1);
9006 SDValue ARMcc = N->getOperand(2);
9007 ARMCC::CondCodes CC =
9008 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9026 /// FIXME: Turn this into a target neutral optimization?
9028 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9029 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9030 N->getOperand(3), Cmp);
9031 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9033 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9034 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9035 N->getOperand(3), NewCmp);
9038 if (Res.getNode()) {
9039 APInt KnownZero, KnownOne;
9040 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9041 // Capture demanded bits information that would be otherwise lost.
9042 if (KnownZero == 0xfffffffe)
9043 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9044 DAG.getValueType(MVT::i1));
9045 else if (KnownZero == 0xffffff00)
9046 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9047 DAG.getValueType(MVT::i8));
9048 else if (KnownZero == 0xffff0000)
9049 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9050 DAG.getValueType(MVT::i16));
9056 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9057 DAGCombinerInfo &DCI) const {
9058 switch (N->getOpcode()) {
9060 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9061 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9062 case ISD::SUB: return PerformSUBCombine(N, DCI);
9063 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9064 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9065 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9066 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9067 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9068 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9069 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9070 case ISD::STORE: return PerformSTORECombine(N, DCI);
9071 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9072 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9073 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9074 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9075 case ISD::FP_TO_SINT:
9076 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9077 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9078 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9081 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9082 case ISD::SIGN_EXTEND:
9083 case ISD::ZERO_EXTEND:
9084 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9085 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9086 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9087 case ARMISD::VLD2DUP:
9088 case ARMISD::VLD3DUP:
9089 case ARMISD::VLD4DUP:
9090 return CombineBaseUpdate(N, DCI);
9091 case ISD::INTRINSIC_VOID:
9092 case ISD::INTRINSIC_W_CHAIN:
9093 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9094 case Intrinsic::arm_neon_vld1:
9095 case Intrinsic::arm_neon_vld2:
9096 case Intrinsic::arm_neon_vld3:
9097 case Intrinsic::arm_neon_vld4:
9098 case Intrinsic::arm_neon_vld2lane:
9099 case Intrinsic::arm_neon_vld3lane:
9100 case Intrinsic::arm_neon_vld4lane:
9101 case Intrinsic::arm_neon_vst1:
9102 case Intrinsic::arm_neon_vst2:
9103 case Intrinsic::arm_neon_vst3:
9104 case Intrinsic::arm_neon_vst4:
9105 case Intrinsic::arm_neon_vst2lane:
9106 case Intrinsic::arm_neon_vst3lane:
9107 case Intrinsic::arm_neon_vst4lane:
9108 return CombineBaseUpdate(N, DCI);
9116 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9118 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9121 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9122 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9123 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9125 switch (VT.getSimpleVT().SimpleTy) {
9131 // Unaligned access can use (for example) LRDB, LRDH, LDR
9132 return AllowsUnaligned;
9135 // For any little-endian targets with neon, we can support unaligned ld/st
9136 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9137 // A big-endian target may also explictly support unaligned accesses
9138 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9142 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9143 unsigned AlignCheck) {
9144 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9145 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9148 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9149 unsigned DstAlign, unsigned SrcAlign,
9152 MachineFunction &MF) const {
9153 const Function *F = MF.getFunction();
9155 // See if we can use NEON instructions for this...
9157 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9158 Subtarget->hasNEON()) {
9159 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9161 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9166 // Lowering to i32/i16 if the size permits.
9169 } else if (Size >= 2) {
9173 // Let the target-independent logic figure it out.
9177 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9182 switch (VT.getSimpleVT().SimpleTy) {
9183 default: return false;
9198 if ((V & (Scale - 1)) != 0)
9201 return V == (V & ((1LL << 5) - 1));
9204 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9205 const ARMSubtarget *Subtarget) {
9212 switch (VT.getSimpleVT().SimpleTy) {
9213 default: return false;
9218 // + imm12 or - imm8
9220 return V == (V & ((1LL << 8) - 1));
9221 return V == (V & ((1LL << 12) - 1));
9224 // Same as ARM mode. FIXME: NEON?
9225 if (!Subtarget->hasVFP2())
9230 return V == (V & ((1LL << 8) - 1));
9234 /// isLegalAddressImmediate - Return true if the integer value can be used
9235 /// as the offset of the target addressing mode for load / store of the
9237 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9238 const ARMSubtarget *Subtarget) {
9245 if (Subtarget->isThumb1Only())
9246 return isLegalT1AddressImmediate(V, VT);
9247 else if (Subtarget->isThumb2())
9248 return isLegalT2AddressImmediate(V, VT, Subtarget);
9253 switch (VT.getSimpleVT().SimpleTy) {
9254 default: return false;
9259 return V == (V & ((1LL << 12) - 1));
9262 return V == (V & ((1LL << 8) - 1));
9265 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9270 return V == (V & ((1LL << 8) - 1));
9274 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9276 int Scale = AM.Scale;
9280 switch (VT.getSimpleVT().SimpleTy) {
9281 default: return false;
9290 return Scale == 2 || Scale == 4 || Scale == 8;
9293 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9297 // Note, we allow "void" uses (basically, uses that aren't loads or
9298 // stores), because arm allows folding a scale into many arithmetic
9299 // operations. This should be made more precise and revisited later.
9301 // Allow r << imm, but the imm has to be a multiple of two.
9302 if (Scale & 1) return false;
9303 return isPowerOf2_32(Scale);
9307 /// isLegalAddressingMode - Return true if the addressing mode represented
9308 /// by AM is legal for this target, for a load/store of the specified type.
9309 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9311 EVT VT = getValueType(Ty, true);
9312 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9315 // Can never fold addr of global into load/store.
9320 case 0: // no scale reg, must be "r+i" or "r", or "i".
9323 if (Subtarget->isThumb1Only())
9327 // ARM doesn't support any R+R*scale+imm addr modes.
9334 if (Subtarget->isThumb2())
9335 return isLegalT2ScaledAddressingMode(AM, VT);
9337 int Scale = AM.Scale;
9338 switch (VT.getSimpleVT().SimpleTy) {
9339 default: return false;
9343 if (Scale < 0) Scale = -Scale;
9347 return isPowerOf2_32(Scale & ~1);
9351 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9356 // Note, we allow "void" uses (basically, uses that aren't loads or
9357 // stores), because arm allows folding a scale into many arithmetic
9358 // operations. This should be made more precise and revisited later.
9360 // Allow r << imm, but the imm has to be a multiple of two.
9361 if (Scale & 1) return false;
9362 return isPowerOf2_32(Scale);
9368 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9369 /// icmp immediate, that is the target has icmp instructions which can compare
9370 /// a register against the immediate without having to materialize the
9371 /// immediate into a register.
9372 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9373 // Thumb2 and ARM modes can use cmn for negative immediates.
9374 if (!Subtarget->isThumb())
9375 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9376 if (Subtarget->isThumb2())
9377 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9378 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9379 return Imm >= 0 && Imm <= 255;
9382 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9383 /// *or sub* immediate, that is the target has add or sub instructions which can
9384 /// add a register with the immediate without having to materialize the
9385 /// immediate into a register.
9386 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9387 // Same encoding for add/sub, just flip the sign.
9388 int64_t AbsImm = llvm::abs64(Imm);
9389 if (!Subtarget->isThumb())
9390 return ARM_AM::getSOImmVal(AbsImm) != -1;
9391 if (Subtarget->isThumb2())
9392 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9393 // Thumb1 only has 8-bit unsigned immediate.
9394 return AbsImm >= 0 && AbsImm <= 255;
9397 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9398 bool isSEXTLoad, SDValue &Base,
9399 SDValue &Offset, bool &isInc,
9400 SelectionDAG &DAG) {
9401 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9404 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9406 Base = Ptr->getOperand(0);
9407 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9408 int RHSC = (int)RHS->getZExtValue();
9409 if (RHSC < 0 && RHSC > -256) {
9410 assert(Ptr->getOpcode() == ISD::ADD);
9412 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9416 isInc = (Ptr->getOpcode() == ISD::ADD);
9417 Offset = Ptr->getOperand(1);
9419 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9421 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9422 int RHSC = (int)RHS->getZExtValue();
9423 if (RHSC < 0 && RHSC > -0x1000) {
9424 assert(Ptr->getOpcode() == ISD::ADD);
9426 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9427 Base = Ptr->getOperand(0);
9432 if (Ptr->getOpcode() == ISD::ADD) {
9434 ARM_AM::ShiftOpc ShOpcVal=
9435 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9436 if (ShOpcVal != ARM_AM::no_shift) {
9437 Base = Ptr->getOperand(1);
9438 Offset = Ptr->getOperand(0);
9440 Base = Ptr->getOperand(0);
9441 Offset = Ptr->getOperand(1);
9446 isInc = (Ptr->getOpcode() == ISD::ADD);
9447 Base = Ptr->getOperand(0);
9448 Offset = Ptr->getOperand(1);
9452 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9456 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9457 bool isSEXTLoad, SDValue &Base,
9458 SDValue &Offset, bool &isInc,
9459 SelectionDAG &DAG) {
9460 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9463 Base = Ptr->getOperand(0);
9464 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9465 int RHSC = (int)RHS->getZExtValue();
9466 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9467 assert(Ptr->getOpcode() == ISD::ADD);
9469 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9471 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9472 isInc = Ptr->getOpcode() == ISD::ADD;
9473 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9481 /// getPreIndexedAddressParts - returns true by value, base pointer and
9482 /// offset pointer and addressing mode by reference if the node's address
9483 /// can be legally represented as pre-indexed load / store address.
9485 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9487 ISD::MemIndexedMode &AM,
9488 SelectionDAG &DAG) const {
9489 if (Subtarget->isThumb1Only())
9494 bool isSEXTLoad = false;
9495 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9496 Ptr = LD->getBasePtr();
9497 VT = LD->getMemoryVT();
9498 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9499 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9500 Ptr = ST->getBasePtr();
9501 VT = ST->getMemoryVT();
9506 bool isLegal = false;
9507 if (Subtarget->isThumb2())
9508 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9509 Offset, isInc, DAG);
9511 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9512 Offset, isInc, DAG);
9516 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9520 /// getPostIndexedAddressParts - returns true by value, base pointer and
9521 /// offset pointer and addressing mode by reference if this node can be
9522 /// combined with a load / store to form a post-indexed load / store.
9523 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9526 ISD::MemIndexedMode &AM,
9527 SelectionDAG &DAG) const {
9528 if (Subtarget->isThumb1Only())
9533 bool isSEXTLoad = false;
9534 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9535 VT = LD->getMemoryVT();
9536 Ptr = LD->getBasePtr();
9537 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9538 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9539 VT = ST->getMemoryVT();
9540 Ptr = ST->getBasePtr();
9545 bool isLegal = false;
9546 if (Subtarget->isThumb2())
9547 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9550 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9556 // Swap base ptr and offset to catch more post-index load / store when
9557 // it's legal. In Thumb2 mode, offset must be an immediate.
9558 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9559 !Subtarget->isThumb2())
9560 std::swap(Base, Offset);
9562 // Post-indexed load / store update the base pointer.
9567 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9571 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9574 const SelectionDAG &DAG,
9575 unsigned Depth) const {
9576 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9577 switch (Op.getOpcode()) {
9579 case ARMISD::CMOV: {
9580 // Bits are known zero/one if known on the LHS and RHS.
9581 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9582 if (KnownZero == 0 && KnownOne == 0) return;
9584 APInt KnownZeroRHS, KnownOneRHS;
9585 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9586 KnownZero &= KnownZeroRHS;
9587 KnownOne &= KnownOneRHS;
9593 //===----------------------------------------------------------------------===//
9594 // ARM Inline Assembly Support
9595 //===----------------------------------------------------------------------===//
9597 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9598 // Looking for "rev" which is V6+.
9599 if (!Subtarget->hasV6Ops())
9602 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9603 std::string AsmStr = IA->getAsmString();
9604 SmallVector<StringRef, 4> AsmPieces;
9605 SplitString(AsmStr, AsmPieces, ";\n");
9607 switch (AsmPieces.size()) {
9608 default: return false;
9610 AsmStr = AsmPieces[0];
9612 SplitString(AsmStr, AsmPieces, " \t,");
9615 if (AsmPieces.size() == 3 &&
9616 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9617 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9618 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9619 if (Ty && Ty->getBitWidth() == 32)
9620 return IntrinsicLowering::LowerToByteSwap(CI);
9628 /// getConstraintType - Given a constraint letter, return the type of
9629 /// constraint it is for this target.
9630 ARMTargetLowering::ConstraintType
9631 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9632 if (Constraint.size() == 1) {
9633 switch (Constraint[0]) {
9635 case 'l': return C_RegisterClass;
9636 case 'w': return C_RegisterClass;
9637 case 'h': return C_RegisterClass;
9638 case 'x': return C_RegisterClass;
9639 case 't': return C_RegisterClass;
9640 case 'j': return C_Other; // Constant for movw.
9641 // An address with a single base register. Due to the way we
9642 // currently handle addresses it is the same as an 'r' memory constraint.
9643 case 'Q': return C_Memory;
9645 } else if (Constraint.size() == 2) {
9646 switch (Constraint[0]) {
9648 // All 'U+' constraints are addresses.
9649 case 'U': return C_Memory;
9652 return TargetLowering::getConstraintType(Constraint);
9655 /// Examine constraint type and operand type and determine a weight value.
9656 /// This object must already have been set up with the operand type
9657 /// and the current alternative constraint selected.
9658 TargetLowering::ConstraintWeight
9659 ARMTargetLowering::getSingleConstraintMatchWeight(
9660 AsmOperandInfo &info, const char *constraint) const {
9661 ConstraintWeight weight = CW_Invalid;
9662 Value *CallOperandVal = info.CallOperandVal;
9663 // If we don't have a value, we can't do a match,
9664 // but allow it at the lowest weight.
9665 if (CallOperandVal == NULL)
9667 Type *type = CallOperandVal->getType();
9668 // Look at the constraint type.
9669 switch (*constraint) {
9671 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9674 if (type->isIntegerTy()) {
9675 if (Subtarget->isThumb())
9676 weight = CW_SpecificReg;
9678 weight = CW_Register;
9682 if (type->isFloatingPointTy())
9683 weight = CW_Register;
9689 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9691 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9693 if (Constraint.size() == 1) {
9694 // GCC ARM Constraint Letters
9695 switch (Constraint[0]) {
9696 case 'l': // Low regs or general regs.
9697 if (Subtarget->isThumb())
9698 return RCPair(0U, &ARM::tGPRRegClass);
9699 return RCPair(0U, &ARM::GPRRegClass);
9700 case 'h': // High regs or no regs.
9701 if (Subtarget->isThumb())
9702 return RCPair(0U, &ARM::hGPRRegClass);
9705 return RCPair(0U, &ARM::GPRRegClass);
9708 return RCPair(0U, &ARM::SPRRegClass);
9709 if (VT.getSizeInBits() == 64)
9710 return RCPair(0U, &ARM::DPRRegClass);
9711 if (VT.getSizeInBits() == 128)
9712 return RCPair(0U, &ARM::QPRRegClass);
9716 return RCPair(0U, &ARM::SPR_8RegClass);
9717 if (VT.getSizeInBits() == 64)
9718 return RCPair(0U, &ARM::DPR_8RegClass);
9719 if (VT.getSizeInBits() == 128)
9720 return RCPair(0U, &ARM::QPR_8RegClass);
9724 return RCPair(0U, &ARM::SPRRegClass);
9728 if (StringRef("{cc}").equals_lower(Constraint))
9729 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9731 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9734 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9735 /// vector. If it is invalid, don't add anything to Ops.
9736 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9737 std::string &Constraint,
9738 std::vector<SDValue>&Ops,
9739 SelectionDAG &DAG) const {
9740 SDValue Result(0, 0);
9742 // Currently only support length 1 constraints.
9743 if (Constraint.length() != 1) return;
9745 char ConstraintLetter = Constraint[0];
9746 switch (ConstraintLetter) {
9749 case 'I': case 'J': case 'K': case 'L':
9750 case 'M': case 'N': case 'O':
9751 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9755 int64_t CVal64 = C->getSExtValue();
9756 int CVal = (int) CVal64;
9757 // None of these constraints allow values larger than 32 bits. Check
9758 // that the value fits in an int.
9762 switch (ConstraintLetter) {
9764 // Constant suitable for movw, must be between 0 and
9766 if (Subtarget->hasV6T2Ops())
9767 if (CVal >= 0 && CVal <= 65535)
9771 if (Subtarget->isThumb1Only()) {
9772 // This must be a constant between 0 and 255, for ADD
9774 if (CVal >= 0 && CVal <= 255)
9776 } else if (Subtarget->isThumb2()) {
9777 // A constant that can be used as an immediate value in a
9778 // data-processing instruction.
9779 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9782 // A constant that can be used as an immediate value in a
9783 // data-processing instruction.
9784 if (ARM_AM::getSOImmVal(CVal) != -1)
9790 if (Subtarget->isThumb()) { // FIXME thumb2
9791 // This must be a constant between -255 and -1, for negated ADD
9792 // immediates. This can be used in GCC with an "n" modifier that
9793 // prints the negated value, for use with SUB instructions. It is
9794 // not useful otherwise but is implemented for compatibility.
9795 if (CVal >= -255 && CVal <= -1)
9798 // This must be a constant between -4095 and 4095. It is not clear
9799 // what this constraint is intended for. Implemented for
9800 // compatibility with GCC.
9801 if (CVal >= -4095 && CVal <= 4095)
9807 if (Subtarget->isThumb1Only()) {
9808 // A 32-bit value where only one byte has a nonzero value. Exclude
9809 // zero to match GCC. This constraint is used by GCC internally for
9810 // constants that can be loaded with a move/shift combination.
9811 // It is not useful otherwise but is implemented for compatibility.
9812 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9814 } else if (Subtarget->isThumb2()) {
9815 // A constant whose bitwise inverse can be used as an immediate
9816 // value in a data-processing instruction. This can be used in GCC
9817 // with a "B" modifier that prints the inverted value, for use with
9818 // BIC and MVN instructions. It is not useful otherwise but is
9819 // implemented for compatibility.
9820 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9823 // A constant whose bitwise inverse can be used as an immediate
9824 // value in a data-processing instruction. This can be used in GCC
9825 // with a "B" modifier that prints the inverted value, for use with
9826 // BIC and MVN instructions. It is not useful otherwise but is
9827 // implemented for compatibility.
9828 if (ARM_AM::getSOImmVal(~CVal) != -1)
9834 if (Subtarget->isThumb1Only()) {
9835 // This must be a constant between -7 and 7,
9836 // for 3-operand ADD/SUB immediate instructions.
9837 if (CVal >= -7 && CVal < 7)
9839 } else if (Subtarget->isThumb2()) {
9840 // A constant whose negation can be used as an immediate value in a
9841 // data-processing instruction. This can be used in GCC with an "n"
9842 // modifier that prints the negated value, for use with SUB
9843 // instructions. It is not useful otherwise but is implemented for
9845 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9848 // A constant whose negation can be used as an immediate value in a
9849 // data-processing instruction. This can be used in GCC with an "n"
9850 // modifier that prints the negated value, for use with SUB
9851 // instructions. It is not useful otherwise but is implemented for
9853 if (ARM_AM::getSOImmVal(-CVal) != -1)
9859 if (Subtarget->isThumb()) { // FIXME thumb2
9860 // This must be a multiple of 4 between 0 and 1020, for
9861 // ADD sp + immediate.
9862 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9865 // A power of two or a constant between 0 and 32. This is used in
9866 // GCC for the shift amount on shifted register operands, but it is
9867 // useful in general for any shift amounts.
9868 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9874 if (Subtarget->isThumb()) { // FIXME thumb2
9875 // This must be a constant between 0 and 31, for shift amounts.
9876 if (CVal >= 0 && CVal <= 31)
9882 if (Subtarget->isThumb()) { // FIXME thumb2
9883 // This must be a multiple of 4 between -508 and 508, for
9884 // ADD/SUB sp = sp + immediate.
9885 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9890 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9894 if (Result.getNode()) {
9895 Ops.push_back(Result);
9898 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9902 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9903 // The ARM target isn't yet aware of offsets.
9907 bool ARM::isBitFieldInvertedMask(unsigned v) {
9908 if (v == 0xffffffff)
9910 // there can be 1's on either or both "outsides", all the "inside"
9912 unsigned int lsb = 0, msb = 31;
9913 while (v & (1 << msb)) --msb;
9914 while (v & (1 << lsb)) ++lsb;
9915 for (unsigned int i = lsb; i <= msb; ++i) {
9922 /// isFPImmLegal - Returns true if the target can instruction select the
9923 /// specified FP immediate natively. If false, the legalizer will
9924 /// materialize the FP immediate as a load from a constant pool.
9925 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9926 if (!Subtarget->hasVFP3())
9929 return ARM_AM::getFP32Imm(Imm) != -1;
9931 return ARM_AM::getFP64Imm(Imm) != -1;
9935 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9936 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9937 /// specified in the intrinsic calls.
9938 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9940 unsigned Intrinsic) const {
9941 switch (Intrinsic) {
9942 case Intrinsic::arm_neon_vld1:
9943 case Intrinsic::arm_neon_vld2:
9944 case Intrinsic::arm_neon_vld3:
9945 case Intrinsic::arm_neon_vld4:
9946 case Intrinsic::arm_neon_vld2lane:
9947 case Intrinsic::arm_neon_vld3lane:
9948 case Intrinsic::arm_neon_vld4lane: {
9949 Info.opc = ISD::INTRINSIC_W_CHAIN;
9950 // Conservatively set memVT to the entire set of vectors loaded.
9951 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9952 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9953 Info.ptrVal = I.getArgOperand(0);
9955 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9956 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9957 Info.vol = false; // volatile loads with NEON intrinsics not supported
9958 Info.readMem = true;
9959 Info.writeMem = false;
9962 case Intrinsic::arm_neon_vst1:
9963 case Intrinsic::arm_neon_vst2:
9964 case Intrinsic::arm_neon_vst3:
9965 case Intrinsic::arm_neon_vst4:
9966 case Intrinsic::arm_neon_vst2lane:
9967 case Intrinsic::arm_neon_vst3lane:
9968 case Intrinsic::arm_neon_vst4lane: {
9969 Info.opc = ISD::INTRINSIC_VOID;
9970 // Conservatively set memVT to the entire set of vectors stored.
9971 unsigned NumElts = 0;
9972 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9973 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9974 if (!ArgTy->isVectorTy())
9976 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9978 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9979 Info.ptrVal = I.getArgOperand(0);
9981 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9982 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9983 Info.vol = false; // volatile stores with NEON intrinsics not supported
9984 Info.readMem = false;
9985 Info.writeMem = true;
9988 case Intrinsic::arm_strexd: {
9989 Info.opc = ISD::INTRINSIC_W_CHAIN;
9990 Info.memVT = MVT::i64;
9991 Info.ptrVal = I.getArgOperand(2);
9995 Info.readMem = false;
9996 Info.writeMem = true;
9999 case Intrinsic::arm_ldrexd: {
10000 Info.opc = ISD::INTRINSIC_W_CHAIN;
10001 Info.memVT = MVT::i64;
10002 Info.ptrVal = I.getArgOperand(0);
10006 Info.readMem = true;
10007 Info.writeMem = false;