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::FSQRT, MVT::v4f32, Expand);
509 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
510 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
518 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
520 // Neon does not support some operations on v1i64 and v2i64 types.
521 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
522 // Custom handling for some quad-vector types to detect VMULL.
523 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
524 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
525 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
526 // Custom handling for some vector types to avoid expensive expansions
527 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
528 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
529 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
532 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
533 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
534 // a destination type that is wider than the source, and nor does
535 // it have a FP_TO_[SU]INT instruction with a narrower destination than
537 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
538 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
539 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
540 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
542 setTargetDAGCombine(ISD::INTRINSIC_VOID);
543 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
544 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
545 setTargetDAGCombine(ISD::SHL);
546 setTargetDAGCombine(ISD::SRL);
547 setTargetDAGCombine(ISD::SRA);
548 setTargetDAGCombine(ISD::SIGN_EXTEND);
549 setTargetDAGCombine(ISD::ZERO_EXTEND);
550 setTargetDAGCombine(ISD::ANY_EXTEND);
551 setTargetDAGCombine(ISD::SELECT_CC);
552 setTargetDAGCombine(ISD::BUILD_VECTOR);
553 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
554 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
555 setTargetDAGCombine(ISD::STORE);
556 setTargetDAGCombine(ISD::FP_TO_SINT);
557 setTargetDAGCombine(ISD::FP_TO_UINT);
558 setTargetDAGCombine(ISD::FDIV);
560 // It is legal to extload from v4i8 to v4i16 or v4i32.
561 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
562 MVT::v4i16, MVT::v2i16,
564 for (unsigned i = 0; i < 6; ++i) {
565 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
566 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
567 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
571 // ARM and Thumb2 support UMLAL/SMLAL.
572 if (!Subtarget->isThumb1Only())
573 setTargetDAGCombine(ISD::ADDC);
576 computeRegisterProperties();
578 // ARM does not have f32 extending load.
579 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
581 // ARM does not have i1 sign extending load.
582 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
584 // ARM supports all 4 flavors of integer indexed load / store.
585 if (!Subtarget->isThumb1Only()) {
586 for (unsigned im = (unsigned)ISD::PRE_INC;
587 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
588 setIndexedLoadAction(im, MVT::i1, Legal);
589 setIndexedLoadAction(im, MVT::i8, Legal);
590 setIndexedLoadAction(im, MVT::i16, Legal);
591 setIndexedLoadAction(im, MVT::i32, Legal);
592 setIndexedStoreAction(im, MVT::i1, Legal);
593 setIndexedStoreAction(im, MVT::i8, Legal);
594 setIndexedStoreAction(im, MVT::i16, Legal);
595 setIndexedStoreAction(im, MVT::i32, Legal);
599 // i64 operation support.
600 setOperationAction(ISD::MUL, MVT::i64, Expand);
601 setOperationAction(ISD::MULHU, MVT::i32, Expand);
602 if (Subtarget->isThumb1Only()) {
603 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
604 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
606 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
607 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
608 setOperationAction(ISD::MULHS, MVT::i32, Expand);
610 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
612 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
613 setOperationAction(ISD::SRL, MVT::i64, Custom);
614 setOperationAction(ISD::SRA, MVT::i64, Custom);
616 if (!Subtarget->isThumb1Only()) {
617 // FIXME: We should do this for Thumb1 as well.
618 setOperationAction(ISD::ADDC, MVT::i32, Custom);
619 setOperationAction(ISD::ADDE, MVT::i32, Custom);
620 setOperationAction(ISD::SUBC, MVT::i32, Custom);
621 setOperationAction(ISD::SUBE, MVT::i32, Custom);
624 // ARM does not have ROTL.
625 setOperationAction(ISD::ROTL, MVT::i32, Expand);
626 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
627 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
628 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
629 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
631 // These just redirect to CTTZ and CTLZ on ARM.
632 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
633 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
635 // Only ARMv6 has BSWAP.
636 if (!Subtarget->hasV6Ops())
637 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
639 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
640 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
641 // These are expanded into libcalls if the cpu doesn't have HW divider.
642 setOperationAction(ISD::SDIV, MVT::i32, Expand);
643 setOperationAction(ISD::UDIV, MVT::i32, Expand);
645 setOperationAction(ISD::SREM, MVT::i32, Expand);
646 setOperationAction(ISD::UREM, MVT::i32, Expand);
647 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
648 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
650 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
651 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
652 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
653 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
654 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
656 setOperationAction(ISD::TRAP, MVT::Other, Legal);
658 // Use the default implementation.
659 setOperationAction(ISD::VASTART, MVT::Other, Custom);
660 setOperationAction(ISD::VAARG, MVT::Other, Expand);
661 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
662 setOperationAction(ISD::VAEND, MVT::Other, Expand);
663 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
664 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
666 if (!Subtarget->isTargetDarwin()) {
667 // Non-Darwin platforms may return values in these registers via the
668 // personality function.
669 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
670 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
671 setExceptionPointerRegister(ARM::R0);
672 setExceptionSelectorRegister(ARM::R1);
675 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
676 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
677 // the default expansion.
678 // FIXME: This should be checking for v6k, not just v6.
679 if (Subtarget->hasDataBarrier() ||
680 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
681 // membarrier needs custom lowering; the rest are legal and handled
683 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
684 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
685 // Custom lowering for 64-bit ops
686 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
691 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
692 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
693 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
694 setInsertFencesForAtomic(true);
696 // Set them all for expansion, which will force libcalls.
697 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
698 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
699 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
711 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
712 // Unordered/Monotonic case.
713 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
714 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
715 // Since the libcalls include locking, fold in the fences
716 setShouldFoldAtomicFences(true);
719 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
721 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
722 if (!Subtarget->hasV6Ops()) {
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
726 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
728 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
729 !Subtarget->isThumb1Only()) {
730 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
731 // iff target supports vfp2.
732 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
733 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
736 // We want to custom lower some of our intrinsics.
737 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
738 if (Subtarget->isTargetDarwin()) {
739 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
740 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
741 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
744 setOperationAction(ISD::SETCC, MVT::i32, Expand);
745 setOperationAction(ISD::SETCC, MVT::f32, Expand);
746 setOperationAction(ISD::SETCC, MVT::f64, Expand);
747 setOperationAction(ISD::SELECT, MVT::i32, Custom);
748 setOperationAction(ISD::SELECT, MVT::f32, Custom);
749 setOperationAction(ISD::SELECT, MVT::f64, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
751 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
752 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
754 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
755 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
756 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
757 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
758 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
760 // We don't support sin/cos/fmod/copysign/pow
761 setOperationAction(ISD::FSIN, MVT::f64, Expand);
762 setOperationAction(ISD::FSIN, MVT::f32, Expand);
763 setOperationAction(ISD::FCOS, MVT::f32, Expand);
764 setOperationAction(ISD::FCOS, MVT::f64, Expand);
765 setOperationAction(ISD::FREM, MVT::f64, Expand);
766 setOperationAction(ISD::FREM, MVT::f32, Expand);
767 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
768 !Subtarget->isThumb1Only()) {
769 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
770 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
772 setOperationAction(ISD::FPOW, MVT::f64, Expand);
773 setOperationAction(ISD::FPOW, MVT::f32, Expand);
775 if (!Subtarget->hasVFP4()) {
776 setOperationAction(ISD::FMA, MVT::f64, Expand);
777 setOperationAction(ISD::FMA, MVT::f32, Expand);
780 // Various VFP goodness
781 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
782 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
783 if (Subtarget->hasVFP2()) {
784 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
785 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
786 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
787 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
789 // Special handling for half-precision FP.
790 if (!Subtarget->hasFP16()) {
791 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
792 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
796 // We have target-specific dag combine patterns for the following nodes:
797 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
798 setTargetDAGCombine(ISD::ADD);
799 setTargetDAGCombine(ISD::SUB);
800 setTargetDAGCombine(ISD::MUL);
801 setTargetDAGCombine(ISD::AND);
802 setTargetDAGCombine(ISD::OR);
803 setTargetDAGCombine(ISD::XOR);
805 if (Subtarget->hasV6Ops())
806 setTargetDAGCombine(ISD::SRL);
808 setStackPointerRegisterToSaveRestore(ARM::SP);
810 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
811 !Subtarget->hasVFP2())
812 setSchedulingPreference(Sched::RegPressure);
814 setSchedulingPreference(Sched::Hybrid);
816 //// temporary - rewrite interface to use type
817 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
818 maxStoresPerMemset = 16;
819 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
821 // On ARM arguments smaller than 4 bytes are extended, so all arguments
822 // are at least 4 bytes aligned.
823 setMinStackArgumentAlignment(4);
825 benefitFromCodePlacementOpt = true;
827 // Prefer likely predicted branches to selects on out-of-order cores.
828 predictableSelectIsExpensive = Subtarget->isLikeA9();
830 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
833 // FIXME: It might make sense to define the representative register class as the
834 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
835 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
836 // SPR's representative would be DPR_VFP2. This should work well if register
837 // pressure tracking were modified such that a register use would increment the
838 // pressure of the register class's representative and all of it's super
839 // classes' representatives transitively. We have not implemented this because
840 // of the difficulty prior to coalescing of modeling operand register classes
841 // due to the common occurrence of cross class copies and subregister insertions
843 std::pair<const TargetRegisterClass*, uint8_t>
844 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
845 const TargetRegisterClass *RRC = 0;
847 switch (VT.getSimpleVT().SimpleTy) {
849 return TargetLowering::findRepresentativeClass(VT);
850 // Use DPR as representative register class for all floating point
851 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
852 // the cost is 1 for both f32 and f64.
853 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
854 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
855 RRC = &ARM::DPRRegClass;
856 // When NEON is used for SP, only half of the register file is available
857 // because operations that define both SP and DP results will be constrained
858 // to the VFP2 class (D0-D15). We currently model this constraint prior to
859 // coalescing by double-counting the SP regs. See the FIXME above.
860 if (Subtarget->useNEONForSinglePrecisionFP())
863 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
864 case MVT::v4f32: case MVT::v2f64:
865 RRC = &ARM::DPRRegClass;
869 RRC = &ARM::DPRRegClass;
873 RRC = &ARM::DPRRegClass;
877 return std::make_pair(RRC, Cost);
880 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
883 case ARMISD::Wrapper: return "ARMISD::Wrapper";
884 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
885 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
886 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
887 case ARMISD::CALL: return "ARMISD::CALL";
888 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
889 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
890 case ARMISD::tCALL: return "ARMISD::tCALL";
891 case ARMISD::BRCOND: return "ARMISD::BRCOND";
892 case ARMISD::BR_JT: return "ARMISD::BR_JT";
893 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
894 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
895 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
896 case ARMISD::CMP: return "ARMISD::CMP";
897 case ARMISD::CMN: return "ARMISD::CMN";
898 case ARMISD::CMPZ: return "ARMISD::CMPZ";
899 case ARMISD::CMPFP: return "ARMISD::CMPFP";
900 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
901 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
902 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
904 case ARMISD::CMOV: return "ARMISD::CMOV";
906 case ARMISD::RBIT: return "ARMISD::RBIT";
908 case ARMISD::FTOSI: return "ARMISD::FTOSI";
909 case ARMISD::FTOUI: return "ARMISD::FTOUI";
910 case ARMISD::SITOF: return "ARMISD::SITOF";
911 case ARMISD::UITOF: return "ARMISD::UITOF";
913 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
914 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
915 case ARMISD::RRX: return "ARMISD::RRX";
917 case ARMISD::ADDC: return "ARMISD::ADDC";
918 case ARMISD::ADDE: return "ARMISD::ADDE";
919 case ARMISD::SUBC: return "ARMISD::SUBC";
920 case ARMISD::SUBE: return "ARMISD::SUBE";
922 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
923 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
925 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
926 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
928 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
930 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
932 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
934 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
935 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
937 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
939 case ARMISD::VCEQ: return "ARMISD::VCEQ";
940 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
941 case ARMISD::VCGE: return "ARMISD::VCGE";
942 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
943 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
944 case ARMISD::VCGEU: return "ARMISD::VCGEU";
945 case ARMISD::VCGT: return "ARMISD::VCGT";
946 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
947 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
948 case ARMISD::VCGTU: return "ARMISD::VCGTU";
949 case ARMISD::VTST: return "ARMISD::VTST";
951 case ARMISD::VSHL: return "ARMISD::VSHL";
952 case ARMISD::VSHRs: return "ARMISD::VSHRs";
953 case ARMISD::VSHRu: return "ARMISD::VSHRu";
954 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
955 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
956 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
957 case ARMISD::VSHRN: return "ARMISD::VSHRN";
958 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
959 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
960 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
961 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
962 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
963 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
964 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
965 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
966 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
967 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
968 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
969 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
970 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
971 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
972 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
973 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
974 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
975 case ARMISD::VDUP: return "ARMISD::VDUP";
976 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
977 case ARMISD::VEXT: return "ARMISD::VEXT";
978 case ARMISD::VREV64: return "ARMISD::VREV64";
979 case ARMISD::VREV32: return "ARMISD::VREV32";
980 case ARMISD::VREV16: return "ARMISD::VREV16";
981 case ARMISD::VZIP: return "ARMISD::VZIP";
982 case ARMISD::VUZP: return "ARMISD::VUZP";
983 case ARMISD::VTRN: return "ARMISD::VTRN";
984 case ARMISD::VTBL1: return "ARMISD::VTBL1";
985 case ARMISD::VTBL2: return "ARMISD::VTBL2";
986 case ARMISD::VMULLs: return "ARMISD::VMULLs";
987 case ARMISD::VMULLu: return "ARMISD::VMULLu";
988 case ARMISD::UMLAL: return "ARMISD::UMLAL";
989 case ARMISD::SMLAL: return "ARMISD::SMLAL";
990 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
991 case ARMISD::FMAX: return "ARMISD::FMAX";
992 case ARMISD::FMIN: return "ARMISD::FMIN";
993 case ARMISD::BFI: return "ARMISD::BFI";
994 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
995 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
996 case ARMISD::VBSL: return "ARMISD::VBSL";
997 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
998 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
999 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1000 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1001 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1002 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1003 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1004 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1005 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1006 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1007 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1008 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1009 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1010 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1011 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1012 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1013 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1014 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1015 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1016 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1020 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1021 if (!VT.isVector()) return getPointerTy();
1022 return VT.changeVectorElementTypeToInteger();
1025 /// getRegClassFor - Return the register class that should be used for the
1026 /// specified value type.
1027 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1028 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1029 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1030 // load / store 4 to 8 consecutive D registers.
1031 if (Subtarget->hasNEON()) {
1032 if (VT == MVT::v4i64)
1033 return &ARM::QQPRRegClass;
1034 if (VT == MVT::v8i64)
1035 return &ARM::QQQQPRRegClass;
1037 return TargetLowering::getRegClassFor(VT);
1040 // Create a fast isel object.
1042 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1043 const TargetLibraryInfo *libInfo) const {
1044 return ARM::createFastISel(funcInfo, libInfo);
1047 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1048 /// be used for loads / stores from the global.
1049 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1050 return (Subtarget->isThumb1Only() ? 127 : 4095);
1053 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1054 unsigned NumVals = N->getNumValues();
1056 return Sched::RegPressure;
1058 for (unsigned i = 0; i != NumVals; ++i) {
1059 EVT VT = N->getValueType(i);
1060 if (VT == MVT::Glue || VT == MVT::Other)
1062 if (VT.isFloatingPoint() || VT.isVector())
1066 if (!N->isMachineOpcode())
1067 return Sched::RegPressure;
1069 // Load are scheduled for latency even if there instruction itinerary
1070 // is not available.
1071 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1072 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1074 if (MCID.getNumDefs() == 0)
1075 return Sched::RegPressure;
1076 if (!Itins->isEmpty() &&
1077 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1080 return Sched::RegPressure;
1083 //===----------------------------------------------------------------------===//
1085 //===----------------------------------------------------------------------===//
1087 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1088 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1090 default: llvm_unreachable("Unknown condition code!");
1091 case ISD::SETNE: return ARMCC::NE;
1092 case ISD::SETEQ: return ARMCC::EQ;
1093 case ISD::SETGT: return ARMCC::GT;
1094 case ISD::SETGE: return ARMCC::GE;
1095 case ISD::SETLT: return ARMCC::LT;
1096 case ISD::SETLE: return ARMCC::LE;
1097 case ISD::SETUGT: return ARMCC::HI;
1098 case ISD::SETUGE: return ARMCC::HS;
1099 case ISD::SETULT: return ARMCC::LO;
1100 case ISD::SETULE: return ARMCC::LS;
1104 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1105 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1106 ARMCC::CondCodes &CondCode2) {
1107 CondCode2 = ARMCC::AL;
1109 default: llvm_unreachable("Unknown FP condition!");
1111 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1113 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1115 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1116 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1117 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1118 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1119 case ISD::SETO: CondCode = ARMCC::VC; break;
1120 case ISD::SETUO: CondCode = ARMCC::VS; break;
1121 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1122 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1123 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1125 case ISD::SETULT: CondCode = ARMCC::LT; break;
1127 case ISD::SETULE: CondCode = ARMCC::LE; break;
1129 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1133 //===----------------------------------------------------------------------===//
1134 // Calling Convention Implementation
1135 //===----------------------------------------------------------------------===//
1137 #include "ARMGenCallingConv.inc"
1139 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1140 /// given CallingConvention value.
1141 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1143 bool isVarArg) const {
1146 llvm_unreachable("Unsupported calling convention");
1147 case CallingConv::Fast:
1148 if (Subtarget->hasVFP2() && !isVarArg) {
1149 if (!Subtarget->isAAPCS_ABI())
1150 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1151 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1152 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1155 case CallingConv::C: {
1156 // Use target triple & subtarget features to do actual dispatch.
1157 if (!Subtarget->isAAPCS_ABI())
1158 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1159 else if (Subtarget->hasVFP2() &&
1160 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1162 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_AAPCS_VFP:
1167 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1169 case CallingConv::ARM_AAPCS:
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1171 case CallingConv::ARM_APCS:
1172 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1173 case CallingConv::GHC:
1174 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1178 /// LowerCallResult - Lower the result values of a call into the
1179 /// appropriate copies out of appropriate physical registers.
1181 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1182 CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<ISD::InputArg> &Ins,
1184 DebugLoc dl, SelectionDAG &DAG,
1185 SmallVectorImpl<SDValue> &InVals) const {
1187 // Assign locations to each value returned by this call.
1188 SmallVector<CCValAssign, 16> RVLocs;
1189 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1190 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1191 CCInfo.AnalyzeCallResult(Ins,
1192 CCAssignFnForNode(CallConv, /* Return*/ true,
1195 // Copy all of the result registers out of their specified physreg.
1196 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1197 CCValAssign VA = RVLocs[i];
1200 if (VA.needsCustom()) {
1201 // Handle f64 or half of a v2f64.
1202 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1204 Chain = Lo.getValue(1);
1205 InFlag = Lo.getValue(2);
1206 VA = RVLocs[++i]; // skip ahead to next loc
1207 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1209 Chain = Hi.getValue(1);
1210 InFlag = Hi.getValue(2);
1211 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1213 if (VA.getLocVT() == MVT::v2f64) {
1214 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1215 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1216 DAG.getConstant(0, MVT::i32));
1218 VA = RVLocs[++i]; // skip ahead to next loc
1219 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1220 Chain = Lo.getValue(1);
1221 InFlag = Lo.getValue(2);
1222 VA = RVLocs[++i]; // skip ahead to next loc
1223 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1224 Chain = Hi.getValue(1);
1225 InFlag = Hi.getValue(2);
1226 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1227 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1228 DAG.getConstant(1, MVT::i32));
1231 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1233 Chain = Val.getValue(1);
1234 InFlag = Val.getValue(2);
1237 switch (VA.getLocInfo()) {
1238 default: llvm_unreachable("Unknown loc info!");
1239 case CCValAssign::Full: break;
1240 case CCValAssign::BCvt:
1241 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1245 InVals.push_back(Val);
1251 /// LowerMemOpCallTo - Store the argument to the stack.
1253 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1254 SDValue StackPtr, SDValue Arg,
1255 DebugLoc dl, SelectionDAG &DAG,
1256 const CCValAssign &VA,
1257 ISD::ArgFlagsTy Flags) const {
1258 unsigned LocMemOffset = VA.getLocMemOffset();
1259 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1260 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1261 return DAG.getStore(Chain, dl, Arg, PtrOff,
1262 MachinePointerInfo::getStack(LocMemOffset),
1266 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1267 SDValue Chain, SDValue &Arg,
1268 RegsToPassVector &RegsToPass,
1269 CCValAssign &VA, CCValAssign &NextVA,
1271 SmallVector<SDValue, 8> &MemOpChains,
1272 ISD::ArgFlagsTy Flags) const {
1274 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1275 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1276 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1278 if (NextVA.isRegLoc())
1279 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1281 assert(NextVA.isMemLoc());
1282 if (StackPtr.getNode() == 0)
1283 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1285 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1291 /// LowerCall - Lowering a call into a callseq_start <-
1292 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1295 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1296 SmallVectorImpl<SDValue> &InVals) const {
1297 SelectionDAG &DAG = CLI.DAG;
1298 DebugLoc &dl = CLI.DL;
1299 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1300 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1301 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1302 SDValue Chain = CLI.Chain;
1303 SDValue Callee = CLI.Callee;
1304 bool &isTailCall = CLI.IsTailCall;
1305 CallingConv::ID CallConv = CLI.CallConv;
1306 bool doesNotRet = CLI.DoesNotReturn;
1307 bool isVarArg = CLI.IsVarArg;
1309 MachineFunction &MF = DAG.getMachineFunction();
1310 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1311 bool IsSibCall = false;
1312 // Disable tail calls if they're not supported.
1313 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1316 // Check if it's really possible to do a tail call.
1317 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1318 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1319 Outs, OutVals, Ins, DAG);
1320 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1321 // detected sibcalls.
1328 // Analyze operands of the call, assigning locations to each operand.
1329 SmallVector<CCValAssign, 16> ArgLocs;
1330 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1331 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1332 CCInfo.AnalyzeCallOperands(Outs,
1333 CCAssignFnForNode(CallConv, /* Return*/ false,
1336 // Get a count of how many bytes are to be pushed on the stack.
1337 unsigned NumBytes = CCInfo.getNextStackOffset();
1339 // For tail calls, memory operands are available in our caller's stack.
1343 // Adjust the stack pointer for the new arguments...
1344 // These operations are automatically eliminated by the prolog/epilog pass
1346 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1348 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1350 RegsToPassVector RegsToPass;
1351 SmallVector<SDValue, 8> MemOpChains;
1353 // Walk the register/memloc assignments, inserting copies/loads. In the case
1354 // of tail call optimization, arguments are handled later.
1355 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1357 ++i, ++realArgIdx) {
1358 CCValAssign &VA = ArgLocs[i];
1359 SDValue Arg = OutVals[realArgIdx];
1360 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1361 bool isByVal = Flags.isByVal();
1363 // Promote the value if needed.
1364 switch (VA.getLocInfo()) {
1365 default: llvm_unreachable("Unknown loc info!");
1366 case CCValAssign::Full: break;
1367 case CCValAssign::SExt:
1368 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1370 case CCValAssign::ZExt:
1371 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1373 case CCValAssign::AExt:
1374 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1376 case CCValAssign::BCvt:
1377 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1381 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1382 if (VA.needsCustom()) {
1383 if (VA.getLocVT() == MVT::v2f64) {
1384 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1385 DAG.getConstant(0, MVT::i32));
1386 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1387 DAG.getConstant(1, MVT::i32));
1389 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1390 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1392 VA = ArgLocs[++i]; // skip ahead to next loc
1393 if (VA.isRegLoc()) {
1394 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1395 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1397 assert(VA.isMemLoc());
1399 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1400 dl, DAG, VA, Flags));
1403 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1404 StackPtr, MemOpChains, Flags);
1406 } else if (VA.isRegLoc()) {
1407 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1408 } else if (isByVal) {
1409 assert(VA.isMemLoc());
1410 unsigned offset = 0;
1412 // True if this byval aggregate will be split between registers
1414 if (CCInfo.isFirstByValRegValid()) {
1415 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1417 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1418 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1419 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1420 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1421 MachinePointerInfo(),
1422 false, false, false, 0);
1423 MemOpChains.push_back(Load.getValue(1));
1424 RegsToPass.push_back(std::make_pair(j, Load));
1426 offset = ARM::R4 - CCInfo.getFirstByValReg();
1427 CCInfo.clearFirstByValReg();
1430 if (Flags.getByValSize() - 4*offset > 0) {
1431 unsigned LocMemOffset = VA.getLocMemOffset();
1432 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1433 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1435 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1436 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1437 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1439 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1441 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1442 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1443 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1444 Ops, array_lengthof(Ops)));
1446 } else if (!IsSibCall) {
1447 assert(VA.isMemLoc());
1449 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1450 dl, DAG, VA, Flags));
1454 if (!MemOpChains.empty())
1455 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1456 &MemOpChains[0], MemOpChains.size());
1458 // Build a sequence of copy-to-reg nodes chained together with token chain
1459 // and flag operands which copy the outgoing args into the appropriate regs.
1461 // Tail call byval lowering might overwrite argument registers so in case of
1462 // tail call optimization the copies to registers are lowered later.
1464 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1465 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1466 RegsToPass[i].second, InFlag);
1467 InFlag = Chain.getValue(1);
1470 // For tail calls lower the arguments to the 'real' stack slot.
1472 // Force all the incoming stack arguments to be loaded from the stack
1473 // before any new outgoing arguments are stored to the stack, because the
1474 // outgoing stack slots may alias the incoming argument stack slots, and
1475 // the alias isn't otherwise explicit. This is slightly more conservative
1476 // than necessary, because it means that each store effectively depends
1477 // on every argument instead of just those arguments it would clobber.
1479 // Do not flag preceding copytoreg stuff together with the following stuff.
1481 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1482 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1483 RegsToPass[i].second, InFlag);
1484 InFlag = Chain.getValue(1);
1489 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1490 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1491 // node so that legalize doesn't hack it.
1492 bool isDirect = false;
1493 bool isARMFunc = false;
1494 bool isLocalARMFunc = false;
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1497 if (EnableARMLongCalls) {
1498 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1499 && "long-calls with non-static relocation model!");
1500 // Handle a global address or an external symbol. If it's not one of
1501 // those, the target's already in a register, so we don't need to do
1503 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1504 const GlobalValue *GV = G->getGlobal();
1505 // Create a constant pool entry for the callee address
1506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507 ARMConstantPoolValue *CPV =
1508 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1510 // Get the address of the callee into a register
1511 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513 Callee = DAG.getLoad(getPointerTy(), dl,
1514 DAG.getEntryNode(), CPAddr,
1515 MachinePointerInfo::getConstantPool(),
1516 false, false, false, 0);
1517 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1518 const char *Sym = S->getSymbol();
1520 // Create a constant pool entry for the callee address
1521 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1522 ARMConstantPoolValue *CPV =
1523 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1524 ARMPCLabelIndex, 0);
1525 // Get the address of the callee into a register
1526 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1527 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1528 Callee = DAG.getLoad(getPointerTy(), dl,
1529 DAG.getEntryNode(), CPAddr,
1530 MachinePointerInfo::getConstantPool(),
1531 false, false, false, 0);
1533 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1534 const GlobalValue *GV = G->getGlobal();
1536 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1537 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1538 getTargetMachine().getRelocationModel() != Reloc::Static;
1539 isARMFunc = !Subtarget->isThumb() || isStub;
1540 // ARM call to a local ARM function is predicable.
1541 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1542 // tBX takes a register source operand.
1543 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1544 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1545 ARMConstantPoolValue *CPV =
1546 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1547 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1548 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1549 Callee = DAG.getLoad(getPointerTy(), dl,
1550 DAG.getEntryNode(), CPAddr,
1551 MachinePointerInfo::getConstantPool(),
1552 false, false, false, 0);
1553 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1554 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1555 getPointerTy(), Callee, PICLabel);
1557 // On ELF targets for PIC code, direct calls should go through the PLT
1558 unsigned OpFlags = 0;
1559 if (Subtarget->isTargetELF() &&
1560 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1561 OpFlags = ARMII::MO_PLT;
1562 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1564 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1566 bool isStub = Subtarget->isTargetDarwin() &&
1567 getTargetMachine().getRelocationModel() != Reloc::Static;
1568 isARMFunc = !Subtarget->isThumb() || isStub;
1569 // tBX takes a register source operand.
1570 const char *Sym = S->getSymbol();
1571 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1572 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1573 ARMConstantPoolValue *CPV =
1574 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1575 ARMPCLabelIndex, 4);
1576 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1577 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1578 Callee = DAG.getLoad(getPointerTy(), dl,
1579 DAG.getEntryNode(), CPAddr,
1580 MachinePointerInfo::getConstantPool(),
1581 false, false, false, 0);
1582 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1583 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1584 getPointerTy(), Callee, PICLabel);
1586 unsigned OpFlags = 0;
1587 // On ELF targets for PIC code, direct calls should go through the PLT
1588 if (Subtarget->isTargetELF() &&
1589 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1590 OpFlags = ARMII::MO_PLT;
1591 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1595 // FIXME: handle tail calls differently.
1597 bool HasMinSizeAttr = MF.getFunction()->getFnAttributes().
1598 hasAttribute(Attributes::MinSize);
1599 if (Subtarget->isThumb()) {
1600 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1601 CallOpc = ARMISD::CALL_NOLINK;
1602 else if (doesNotRet && isDirect && !isARMFunc &&
1603 Subtarget->hasRAS() && !Subtarget->isThumb1Only() &&
1604 // Emit regular call when code size is the priority
1606 // "mov lr, pc; b _foo" to avoid confusing the RSP
1607 CallOpc = ARMISD::CALL_NOLINK;
1609 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1611 if (!isDirect && !Subtarget->hasV5TOps()) {
1612 CallOpc = ARMISD::CALL_NOLINK;
1613 } else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1614 // Emit regular call when code size is the priority
1616 // "mov lr, pc; b _foo" to avoid confusing the RSP
1617 CallOpc = ARMISD::CALL_NOLINK;
1619 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1622 std::vector<SDValue> Ops;
1623 Ops.push_back(Chain);
1624 Ops.push_back(Callee);
1626 // Add argument registers to the end of the list so that they are known live
1628 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1629 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1630 RegsToPass[i].second.getValueType()));
1632 // Add a register mask operand representing the call-preserved registers.
1633 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1634 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1635 assert(Mask && "Missing call preserved mask for calling convention");
1636 Ops.push_back(DAG.getRegisterMask(Mask));
1638 if (InFlag.getNode())
1639 Ops.push_back(InFlag);
1641 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1643 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1645 // Returns a chain and a flag for retval copy to use.
1646 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1647 InFlag = Chain.getValue(1);
1649 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1650 DAG.getIntPtrConstant(0, true), InFlag);
1652 InFlag = Chain.getValue(1);
1654 // Handle result values, copying them out of physregs into vregs that we
1656 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1660 /// HandleByVal - Every parameter *after* a byval parameter is passed
1661 /// on the stack. Remember the next parameter register to allocate,
1662 /// and then confiscate the rest of the parameter registers to insure
1665 ARMTargetLowering::HandleByVal(
1666 CCState *State, unsigned &size, unsigned Align) const {
1667 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1668 assert((State->getCallOrPrologue() == Prologue ||
1669 State->getCallOrPrologue() == Call) &&
1670 "unhandled ParmContext");
1671 if ((!State->isFirstByValRegValid()) &&
1672 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1673 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1674 unsigned AlignInRegs = Align / 4;
1675 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1676 for (unsigned i = 0; i < Waste; ++i)
1677 reg = State->AllocateReg(GPRArgRegs, 4);
1680 State->setFirstByValReg(reg);
1681 // At a call site, a byval parameter that is split between
1682 // registers and memory needs its size truncated here. In a
1683 // function prologue, such byval parameters are reassembled in
1684 // memory, and are not truncated.
1685 if (State->getCallOrPrologue() == Call) {
1686 unsigned excess = 4 * (ARM::R4 - reg);
1687 assert(size >= excess && "expected larger existing stack allocation");
1692 // Confiscate any remaining parameter registers to preclude their
1693 // assignment to subsequent parameters.
1694 while (State->AllocateReg(GPRArgRegs, 4))
1698 /// MatchingStackOffset - Return true if the given stack call argument is
1699 /// already available in the same position (relatively) of the caller's
1700 /// incoming argument stack.
1702 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1703 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1704 const TargetInstrInfo *TII) {
1705 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1707 if (Arg.getOpcode() == ISD::CopyFromReg) {
1708 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1709 if (!TargetRegisterInfo::isVirtualRegister(VR))
1711 MachineInstr *Def = MRI->getVRegDef(VR);
1714 if (!Flags.isByVal()) {
1715 if (!TII->isLoadFromStackSlot(Def, FI))
1720 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1721 if (Flags.isByVal())
1722 // ByVal argument is passed in as a pointer but it's now being
1723 // dereferenced. e.g.
1724 // define @foo(%struct.X* %A) {
1725 // tail call @bar(%struct.X* byval %A)
1728 SDValue Ptr = Ld->getBasePtr();
1729 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1732 FI = FINode->getIndex();
1736 assert(FI != INT_MAX);
1737 if (!MFI->isFixedObjectIndex(FI))
1739 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1742 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1743 /// for tail call optimization. Targets which want to do tail call
1744 /// optimization should implement this function.
1746 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1747 CallingConv::ID CalleeCC,
1749 bool isCalleeStructRet,
1750 bool isCallerStructRet,
1751 const SmallVectorImpl<ISD::OutputArg> &Outs,
1752 const SmallVectorImpl<SDValue> &OutVals,
1753 const SmallVectorImpl<ISD::InputArg> &Ins,
1754 SelectionDAG& DAG) const {
1755 const Function *CallerF = DAG.getMachineFunction().getFunction();
1756 CallingConv::ID CallerCC = CallerF->getCallingConv();
1757 bool CCMatch = CallerCC == CalleeCC;
1759 // Look for obvious safe cases to perform tail call optimization that do not
1760 // require ABI changes. This is what gcc calls sibcall.
1762 // Do not sibcall optimize vararg calls unless the call site is not passing
1764 if (isVarArg && !Outs.empty())
1767 // Also avoid sibcall optimization if either caller or callee uses struct
1768 // return semantics.
1769 if (isCalleeStructRet || isCallerStructRet)
1772 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1773 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1774 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1775 // support in the assembler and linker to be used. This would need to be
1776 // fixed to fully support tail calls in Thumb1.
1778 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1779 // LR. This means if we need to reload LR, it takes an extra instructions,
1780 // which outweighs the value of the tail call; but here we don't know yet
1781 // whether LR is going to be used. Probably the right approach is to
1782 // generate the tail call here and turn it back into CALL/RET in
1783 // emitEpilogue if LR is used.
1785 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1786 // but we need to make sure there are enough registers; the only valid
1787 // registers are the 4 used for parameters. We don't currently do this
1789 if (Subtarget->isThumb1Only())
1792 // If the calling conventions do not match, then we'd better make sure the
1793 // results are returned in the same way as what the caller expects.
1795 SmallVector<CCValAssign, 16> RVLocs1;
1796 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1797 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1798 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1800 SmallVector<CCValAssign, 16> RVLocs2;
1801 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1802 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1803 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1805 if (RVLocs1.size() != RVLocs2.size())
1807 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1808 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1810 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1812 if (RVLocs1[i].isRegLoc()) {
1813 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1816 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1822 // If Caller's vararg or byval argument has been split between registers and
1823 // stack, do not perform tail call, since part of the argument is in caller's
1825 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1826 getInfo<ARMFunctionInfo>();
1827 if (AFI_Caller->getVarArgsRegSaveSize())
1830 // If the callee takes no arguments then go on to check the results of the
1832 if (!Outs.empty()) {
1833 // Check if stack adjustment is needed. For now, do not do this if any
1834 // argument is passed on the stack.
1835 SmallVector<CCValAssign, 16> ArgLocs;
1836 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1837 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1838 CCInfo.AnalyzeCallOperands(Outs,
1839 CCAssignFnForNode(CalleeCC, false, isVarArg));
1840 if (CCInfo.getNextStackOffset()) {
1841 MachineFunction &MF = DAG.getMachineFunction();
1843 // Check if the arguments are already laid out in the right way as
1844 // the caller's fixed stack objects.
1845 MachineFrameInfo *MFI = MF.getFrameInfo();
1846 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1847 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1848 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1850 ++i, ++realArgIdx) {
1851 CCValAssign &VA = ArgLocs[i];
1852 EVT RegVT = VA.getLocVT();
1853 SDValue Arg = OutVals[realArgIdx];
1854 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1855 if (VA.getLocInfo() == CCValAssign::Indirect)
1857 if (VA.needsCustom()) {
1858 // f64 and vector types are split into multiple registers or
1859 // register/stack-slot combinations. The types will not match
1860 // the registers; give up on memory f64 refs until we figure
1861 // out what to do about this.
1864 if (!ArgLocs[++i].isRegLoc())
1866 if (RegVT == MVT::v2f64) {
1867 if (!ArgLocs[++i].isRegLoc())
1869 if (!ArgLocs[++i].isRegLoc())
1872 } else if (!VA.isRegLoc()) {
1873 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1885 ARMTargetLowering::LowerReturn(SDValue Chain,
1886 CallingConv::ID CallConv, bool isVarArg,
1887 const SmallVectorImpl<ISD::OutputArg> &Outs,
1888 const SmallVectorImpl<SDValue> &OutVals,
1889 DebugLoc dl, SelectionDAG &DAG) const {
1891 // CCValAssign - represent the assignment of the return value to a location.
1892 SmallVector<CCValAssign, 16> RVLocs;
1894 // CCState - Info about the registers and stack slots.
1895 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1896 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1898 // Analyze outgoing return values.
1899 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1902 // If this is the first return lowered for this function, add
1903 // the regs to the liveout set for the function.
1904 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1905 for (unsigned i = 0; i != RVLocs.size(); ++i)
1906 if (RVLocs[i].isRegLoc())
1907 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1912 // Copy the result values into the output registers.
1913 for (unsigned i = 0, realRVLocIdx = 0;
1915 ++i, ++realRVLocIdx) {
1916 CCValAssign &VA = RVLocs[i];
1917 assert(VA.isRegLoc() && "Can only return in registers!");
1919 SDValue Arg = OutVals[realRVLocIdx];
1921 switch (VA.getLocInfo()) {
1922 default: llvm_unreachable("Unknown loc info!");
1923 case CCValAssign::Full: break;
1924 case CCValAssign::BCvt:
1925 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1929 if (VA.needsCustom()) {
1930 if (VA.getLocVT() == MVT::v2f64) {
1931 // Extract the first half and return it in two registers.
1932 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1933 DAG.getConstant(0, MVT::i32));
1934 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1935 DAG.getVTList(MVT::i32, MVT::i32), Half);
1937 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1938 Flag = Chain.getValue(1);
1939 VA = RVLocs[++i]; // skip ahead to next loc
1940 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1941 HalfGPRs.getValue(1), Flag);
1942 Flag = Chain.getValue(1);
1943 VA = RVLocs[++i]; // skip ahead to next loc
1945 // Extract the 2nd half and fall through to handle it as an f64 value.
1946 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1947 DAG.getConstant(1, MVT::i32));
1949 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1951 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1952 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1953 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1954 Flag = Chain.getValue(1);
1955 VA = RVLocs[++i]; // skip ahead to next loc
1956 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1959 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1961 // Guarantee that all emitted copies are
1962 // stuck together, avoiding something bad.
1963 Flag = Chain.getValue(1);
1968 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1970 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1975 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1976 if (N->getNumValues() != 1)
1978 if (!N->hasNUsesOfValue(1, 0))
1981 SDValue TCChain = Chain;
1982 SDNode *Copy = *N->use_begin();
1983 if (Copy->getOpcode() == ISD::CopyToReg) {
1984 // If the copy has a glue operand, we conservatively assume it isn't safe to
1985 // perform a tail call.
1986 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1988 TCChain = Copy->getOperand(0);
1989 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1990 SDNode *VMov = Copy;
1991 // f64 returned in a pair of GPRs.
1992 SmallPtrSet<SDNode*, 2> Copies;
1993 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1995 if (UI->getOpcode() != ISD::CopyToReg)
1999 if (Copies.size() > 2)
2002 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2004 SDValue UseChain = UI->getOperand(0);
2005 if (Copies.count(UseChain.getNode()))
2012 } else if (Copy->getOpcode() == ISD::BITCAST) {
2013 // f32 returned in a single GPR.
2014 if (!Copy->hasOneUse())
2016 Copy = *Copy->use_begin();
2017 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2019 Chain = Copy->getOperand(0);
2024 bool HasRet = false;
2025 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2027 if (UI->getOpcode() != ARMISD::RET_FLAG)
2039 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2040 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2043 if (!CI->isTailCall())
2046 return !Subtarget->isThumb1Only();
2049 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2050 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2051 // one of the above mentioned nodes. It has to be wrapped because otherwise
2052 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2053 // be used to form addressing mode. These wrapped nodes will be selected
2055 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2056 EVT PtrVT = Op.getValueType();
2057 // FIXME there is no actual debug info here
2058 DebugLoc dl = Op.getDebugLoc();
2059 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2061 if (CP->isMachineConstantPoolEntry())
2062 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2063 CP->getAlignment());
2065 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2066 CP->getAlignment());
2067 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2070 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2071 return MachineJumpTableInfo::EK_Inline;
2074 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2075 SelectionDAG &DAG) const {
2076 MachineFunction &MF = DAG.getMachineFunction();
2077 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2078 unsigned ARMPCLabelIndex = 0;
2079 DebugLoc DL = Op.getDebugLoc();
2080 EVT PtrVT = getPointerTy();
2081 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2082 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2084 if (RelocM == Reloc::Static) {
2085 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2087 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2088 ARMPCLabelIndex = AFI->createPICLabelUId();
2089 ARMConstantPoolValue *CPV =
2090 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2091 ARMCP::CPBlockAddress, PCAdj);
2092 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2094 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2095 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2096 MachinePointerInfo::getConstantPool(),
2097 false, false, false, 0);
2098 if (RelocM == Reloc::Static)
2100 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2101 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2104 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2106 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2107 SelectionDAG &DAG) const {
2108 DebugLoc dl = GA->getDebugLoc();
2109 EVT PtrVT = getPointerTy();
2110 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2111 MachineFunction &MF = DAG.getMachineFunction();
2112 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2113 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2114 ARMConstantPoolValue *CPV =
2115 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2116 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2117 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2118 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2119 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2120 MachinePointerInfo::getConstantPool(),
2121 false, false, false, 0);
2122 SDValue Chain = Argument.getValue(1);
2124 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2125 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2127 // call __tls_get_addr.
2130 Entry.Node = Argument;
2131 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2132 Args.push_back(Entry);
2133 // FIXME: is there useful debug info available here?
2134 TargetLowering::CallLoweringInfo CLI(Chain,
2135 (Type *) Type::getInt32Ty(*DAG.getContext()),
2136 false, false, false, false,
2137 0, CallingConv::C, /*isTailCall=*/false,
2138 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2139 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2140 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2141 return CallResult.first;
2144 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2145 // "local exec" model.
2147 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2149 TLSModel::Model model) const {
2150 const GlobalValue *GV = GA->getGlobal();
2151 DebugLoc dl = GA->getDebugLoc();
2153 SDValue Chain = DAG.getEntryNode();
2154 EVT PtrVT = getPointerTy();
2155 // Get the Thread Pointer
2156 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2158 if (model == TLSModel::InitialExec) {
2159 MachineFunction &MF = DAG.getMachineFunction();
2160 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2161 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2162 // Initial exec model.
2163 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2164 ARMConstantPoolValue *CPV =
2165 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2166 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2168 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2169 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2170 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2171 MachinePointerInfo::getConstantPool(),
2172 false, false, false, 0);
2173 Chain = Offset.getValue(1);
2175 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2176 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2178 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2179 MachinePointerInfo::getConstantPool(),
2180 false, false, false, 0);
2183 assert(model == TLSModel::LocalExec);
2184 ARMConstantPoolValue *CPV =
2185 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2186 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2187 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2188 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2189 MachinePointerInfo::getConstantPool(),
2190 false, false, false, 0);
2193 // The address of the thread local variable is the add of the thread
2194 // pointer with the offset of the variable.
2195 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2199 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2200 // TODO: implement the "local dynamic" model
2201 assert(Subtarget->isTargetELF() &&
2202 "TLS not implemented for non-ELF targets");
2203 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2205 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2208 case TLSModel::GeneralDynamic:
2209 case TLSModel::LocalDynamic:
2210 return LowerToTLSGeneralDynamicModel(GA, DAG);
2211 case TLSModel::InitialExec:
2212 case TLSModel::LocalExec:
2213 return LowerToTLSExecModels(GA, DAG, model);
2215 llvm_unreachable("bogus TLS model");
2218 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2219 SelectionDAG &DAG) const {
2220 EVT PtrVT = getPointerTy();
2221 DebugLoc dl = Op.getDebugLoc();
2222 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2223 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2224 if (RelocM == Reloc::PIC_) {
2225 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2226 ARMConstantPoolValue *CPV =
2227 ARMConstantPoolConstant::Create(GV,
2228 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2229 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2230 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2231 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2233 MachinePointerInfo::getConstantPool(),
2234 false, false, false, 0);
2235 SDValue Chain = Result.getValue(1);
2236 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2237 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2239 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2240 MachinePointerInfo::getGOT(),
2241 false, false, false, 0);
2245 // If we have T2 ops, we can materialize the address directly via movt/movw
2246 // pair. This is always cheaper.
2247 if (Subtarget->useMovt()) {
2249 // FIXME: Once remat is capable of dealing with instructions with register
2250 // operands, expand this into two nodes.
2251 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2252 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2254 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2255 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2256 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2257 MachinePointerInfo::getConstantPool(),
2258 false, false, false, 0);
2262 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2263 SelectionDAG &DAG) const {
2264 EVT PtrVT = getPointerTy();
2265 DebugLoc dl = Op.getDebugLoc();
2266 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2267 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2268 MachineFunction &MF = DAG.getMachineFunction();
2269 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2271 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2272 // update ARMFastISel::ARMMaterializeGV.
2273 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2275 // FIXME: Once remat is capable of dealing with instructions with register
2276 // operands, expand this into two nodes.
2277 if (RelocM == Reloc::Static)
2278 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2279 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2281 unsigned Wrapper = (RelocM == Reloc::PIC_)
2282 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2283 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2284 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2285 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2286 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2287 MachinePointerInfo::getGOT(),
2288 false, false, false, 0);
2292 unsigned ARMPCLabelIndex = 0;
2294 if (RelocM == Reloc::Static) {
2295 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2297 ARMPCLabelIndex = AFI->createPICLabelUId();
2298 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2299 ARMConstantPoolValue *CPV =
2300 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2302 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2304 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2306 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2307 MachinePointerInfo::getConstantPool(),
2308 false, false, false, 0);
2309 SDValue Chain = Result.getValue(1);
2311 if (RelocM == Reloc::PIC_) {
2312 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2313 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2316 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2317 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2318 false, false, false, 0);
2323 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2324 SelectionDAG &DAG) const {
2325 assert(Subtarget->isTargetELF() &&
2326 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2327 MachineFunction &MF = DAG.getMachineFunction();
2328 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2329 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2330 EVT PtrVT = getPointerTy();
2331 DebugLoc dl = Op.getDebugLoc();
2332 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2333 ARMConstantPoolValue *CPV =
2334 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2335 ARMPCLabelIndex, PCAdj);
2336 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2337 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2338 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2339 MachinePointerInfo::getConstantPool(),
2340 false, false, false, 0);
2341 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2342 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2346 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2347 DebugLoc dl = Op.getDebugLoc();
2348 SDValue Val = DAG.getConstant(0, MVT::i32);
2349 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2350 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2351 Op.getOperand(1), Val);
2355 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2356 DebugLoc dl = Op.getDebugLoc();
2357 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2358 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2362 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2363 const ARMSubtarget *Subtarget) const {
2364 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2365 DebugLoc dl = Op.getDebugLoc();
2367 default: return SDValue(); // Don't custom lower most intrinsics.
2368 case Intrinsic::arm_thread_pointer: {
2369 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2370 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2372 case Intrinsic::eh_sjlj_lsda: {
2373 MachineFunction &MF = DAG.getMachineFunction();
2374 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2375 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2376 EVT PtrVT = getPointerTy();
2377 DebugLoc dl = Op.getDebugLoc();
2378 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2380 unsigned PCAdj = (RelocM != Reloc::PIC_)
2381 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2382 ARMConstantPoolValue *CPV =
2383 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2384 ARMCP::CPLSDA, PCAdj);
2385 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2386 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2388 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2389 MachinePointerInfo::getConstantPool(),
2390 false, false, false, 0);
2392 if (RelocM == Reloc::PIC_) {
2393 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2394 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2398 case Intrinsic::arm_neon_vmulls:
2399 case Intrinsic::arm_neon_vmullu: {
2400 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2401 ? ARMISD::VMULLs : ARMISD::VMULLu;
2402 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2403 Op.getOperand(1), Op.getOperand(2));
2408 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2409 const ARMSubtarget *Subtarget) {
2410 DebugLoc dl = Op.getDebugLoc();
2411 if (!Subtarget->hasDataBarrier()) {
2412 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2413 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2415 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2416 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2417 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2418 DAG.getConstant(0, MVT::i32));
2421 SDValue Op5 = Op.getOperand(5);
2422 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2423 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2424 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2425 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2427 ARM_MB::MemBOpt DMBOpt;
2428 if (isDeviceBarrier)
2429 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2431 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2432 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2433 DAG.getConstant(DMBOpt, MVT::i32));
2437 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2438 const ARMSubtarget *Subtarget) {
2439 // FIXME: handle "fence singlethread" more efficiently.
2440 DebugLoc dl = Op.getDebugLoc();
2441 if (!Subtarget->hasDataBarrier()) {
2442 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2443 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2445 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2446 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2447 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2448 DAG.getConstant(0, MVT::i32));
2451 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2452 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2455 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2456 const ARMSubtarget *Subtarget) {
2457 // ARM pre v5TE and Thumb1 does not have preload instructions.
2458 if (!(Subtarget->isThumb2() ||
2459 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2460 // Just preserve the chain.
2461 return Op.getOperand(0);
2463 DebugLoc dl = Op.getDebugLoc();
2464 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2466 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2467 // ARMv7 with MP extension has PLDW.
2468 return Op.getOperand(0);
2470 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2471 if (Subtarget->isThumb()) {
2473 isRead = ~isRead & 1;
2474 isData = ~isData & 1;
2477 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2478 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2479 DAG.getConstant(isData, MVT::i32));
2482 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2483 MachineFunction &MF = DAG.getMachineFunction();
2484 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2486 // vastart just stores the address of the VarArgsFrameIndex slot into the
2487 // memory location argument.
2488 DebugLoc dl = Op.getDebugLoc();
2489 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2490 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2491 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2492 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2493 MachinePointerInfo(SV), false, false, 0);
2497 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2498 SDValue &Root, SelectionDAG &DAG,
2499 DebugLoc dl) const {
2500 MachineFunction &MF = DAG.getMachineFunction();
2501 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2503 const TargetRegisterClass *RC;
2504 if (AFI->isThumb1OnlyFunction())
2505 RC = &ARM::tGPRRegClass;
2507 RC = &ARM::GPRRegClass;
2509 // Transform the arguments stored in physical registers into virtual ones.
2510 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2511 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2514 if (NextVA.isMemLoc()) {
2515 MachineFrameInfo *MFI = MF.getFrameInfo();
2516 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2518 // Create load node to retrieve arguments from the stack.
2519 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2520 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2521 MachinePointerInfo::getFixedStack(FI),
2522 false, false, false, 0);
2524 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2525 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2528 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2532 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2533 unsigned &VARegSize, unsigned &VARegSaveSize)
2536 if (CCInfo.isFirstByValRegValid())
2537 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2539 unsigned int firstUnalloced;
2540 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2541 sizeof(GPRArgRegs) /
2542 sizeof(GPRArgRegs[0]));
2543 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2546 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2547 VARegSize = NumGPRs * 4;
2548 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2551 // The remaining GPRs hold either the beginning of variable-argument
2552 // data, or the beginning of an aggregate passed by value (usuall
2553 // byval). Either way, we allocate stack slots adjacent to the data
2554 // provided by our caller, and store the unallocated registers there.
2555 // If this is a variadic function, the va_list pointer will begin with
2556 // these values; otherwise, this reassembles a (byval) structure that
2557 // was split between registers and memory.
2559 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2560 DebugLoc dl, SDValue &Chain,
2561 const Value *OrigArg,
2562 unsigned OffsetFromOrigArg,
2564 bool ForceMutable) const {
2565 MachineFunction &MF = DAG.getMachineFunction();
2566 MachineFrameInfo *MFI = MF.getFrameInfo();
2567 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2568 unsigned firstRegToSaveIndex;
2569 if (CCInfo.isFirstByValRegValid())
2570 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2572 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2573 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2576 unsigned VARegSize, VARegSaveSize;
2577 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2578 if (VARegSaveSize) {
2579 // If this function is vararg, store any remaining integer argument regs
2580 // to their spots on the stack so that they may be loaded by deferencing
2581 // the result of va_next.
2582 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2583 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2584 ArgOffset + VARegSaveSize
2587 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2590 SmallVector<SDValue, 4> MemOps;
2591 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2592 const TargetRegisterClass *RC;
2593 if (AFI->isThumb1OnlyFunction())
2594 RC = &ARM::tGPRRegClass;
2596 RC = &ARM::GPRRegClass;
2598 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2599 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2601 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2602 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2604 MemOps.push_back(Store);
2605 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2606 DAG.getConstant(4, getPointerTy()));
2608 if (!MemOps.empty())
2609 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2610 &MemOps[0], MemOps.size());
2612 // This will point to the next argument passed via stack.
2613 AFI->setVarArgsFrameIndex(
2614 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2618 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2619 CallingConv::ID CallConv, bool isVarArg,
2620 const SmallVectorImpl<ISD::InputArg>
2622 DebugLoc dl, SelectionDAG &DAG,
2623 SmallVectorImpl<SDValue> &InVals)
2625 MachineFunction &MF = DAG.getMachineFunction();
2626 MachineFrameInfo *MFI = MF.getFrameInfo();
2628 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2630 // Assign locations to all of the incoming arguments.
2631 SmallVector<CCValAssign, 16> ArgLocs;
2632 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2633 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2634 CCInfo.AnalyzeFormalArguments(Ins,
2635 CCAssignFnForNode(CallConv, /* Return*/ false,
2638 SmallVector<SDValue, 16> ArgValues;
2639 int lastInsIndex = -1;
2641 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2642 unsigned CurArgIdx = 0;
2643 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2644 CCValAssign &VA = ArgLocs[i];
2645 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2646 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2647 // Arguments stored in registers.
2648 if (VA.isRegLoc()) {
2649 EVT RegVT = VA.getLocVT();
2651 if (VA.needsCustom()) {
2652 // f64 and vector types are split up into multiple registers or
2653 // combinations of registers and stack slots.
2654 if (VA.getLocVT() == MVT::v2f64) {
2655 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2657 VA = ArgLocs[++i]; // skip ahead to next loc
2659 if (VA.isMemLoc()) {
2660 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2661 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2662 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2663 MachinePointerInfo::getFixedStack(FI),
2664 false, false, false, 0);
2666 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2669 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2670 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2671 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2672 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2673 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2675 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2678 const TargetRegisterClass *RC;
2680 if (RegVT == MVT::f32)
2681 RC = &ARM::SPRRegClass;
2682 else if (RegVT == MVT::f64)
2683 RC = &ARM::DPRRegClass;
2684 else if (RegVT == MVT::v2f64)
2685 RC = &ARM::QPRRegClass;
2686 else if (RegVT == MVT::i32)
2687 RC = AFI->isThumb1OnlyFunction() ?
2688 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2689 (const TargetRegisterClass*)&ARM::GPRRegClass;
2691 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2693 // Transform the arguments in physical registers into virtual ones.
2694 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2695 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2698 // If this is an 8 or 16-bit value, it is really passed promoted
2699 // to 32 bits. Insert an assert[sz]ext to capture this, then
2700 // truncate to the right size.
2701 switch (VA.getLocInfo()) {
2702 default: llvm_unreachable("Unknown loc info!");
2703 case CCValAssign::Full: break;
2704 case CCValAssign::BCvt:
2705 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2707 case CCValAssign::SExt:
2708 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2709 DAG.getValueType(VA.getValVT()));
2710 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2712 case CCValAssign::ZExt:
2713 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2714 DAG.getValueType(VA.getValVT()));
2715 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2719 InVals.push_back(ArgValue);
2721 } else { // VA.isRegLoc()
2724 assert(VA.isMemLoc());
2725 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2727 int index = ArgLocs[i].getValNo();
2729 // Some Ins[] entries become multiple ArgLoc[] entries.
2730 // Process them only once.
2731 if (index != lastInsIndex)
2733 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2734 // FIXME: For now, all byval parameter objects are marked mutable.
2735 // This can be changed with more analysis.
2736 // In case of tail call optimization mark all arguments mutable.
2737 // Since they could be overwritten by lowering of arguments in case of
2739 if (Flags.isByVal()) {
2740 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2741 if (!AFI->getVarArgsFrameIndex()) {
2742 VarArgStyleRegisters(CCInfo, DAG,
2743 dl, Chain, CurOrigArg,
2744 Ins[VA.getValNo()].PartOffset,
2745 VA.getLocMemOffset(),
2746 true /*force mutable frames*/);
2747 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2748 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2750 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2751 VA.getLocMemOffset(), false);
2752 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2755 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2756 VA.getLocMemOffset(), true);
2758 // Create load nodes to retrieve arguments from the stack.
2759 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2760 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2761 MachinePointerInfo::getFixedStack(FI),
2762 false, false, false, 0));
2764 lastInsIndex = index;
2771 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2772 CCInfo.getNextStackOffset());
2777 /// isFloatingPointZero - Return true if this is +0.0.
2778 static bool isFloatingPointZero(SDValue Op) {
2779 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2780 return CFP->getValueAPF().isPosZero();
2781 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2782 // Maybe this has already been legalized into the constant pool?
2783 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2784 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2785 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2786 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2787 return CFP->getValueAPF().isPosZero();
2793 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2794 /// the given operands.
2796 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2797 SDValue &ARMcc, SelectionDAG &DAG,
2798 DebugLoc dl) const {
2799 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2800 unsigned C = RHSC->getZExtValue();
2801 if (!isLegalICmpImmediate(C)) {
2802 // Constant does not fit, try adjusting it by one?
2807 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2808 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2809 RHS = DAG.getConstant(C-1, MVT::i32);
2814 if (C != 0 && isLegalICmpImmediate(C-1)) {
2815 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2816 RHS = DAG.getConstant(C-1, MVT::i32);
2821 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2822 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2823 RHS = DAG.getConstant(C+1, MVT::i32);
2828 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2829 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2830 RHS = DAG.getConstant(C+1, MVT::i32);
2837 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2838 ARMISD::NodeType CompareType;
2841 CompareType = ARMISD::CMP;
2846 CompareType = ARMISD::CMPZ;
2849 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2850 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2853 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2855 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2856 DebugLoc dl) const {
2858 if (!isFloatingPointZero(RHS))
2859 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2861 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2862 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2865 /// duplicateCmp - Glue values can have only one use, so this function
2866 /// duplicates a comparison node.
2868 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2869 unsigned Opc = Cmp.getOpcode();
2870 DebugLoc DL = Cmp.getDebugLoc();
2871 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2872 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2874 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2875 Cmp = Cmp.getOperand(0);
2876 Opc = Cmp.getOpcode();
2877 if (Opc == ARMISD::CMPFP)
2878 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2880 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2881 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2883 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2886 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2887 SDValue Cond = Op.getOperand(0);
2888 SDValue SelectTrue = Op.getOperand(1);
2889 SDValue SelectFalse = Op.getOperand(2);
2890 DebugLoc dl = Op.getDebugLoc();
2894 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2895 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2897 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2898 const ConstantSDNode *CMOVTrue =
2899 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2900 const ConstantSDNode *CMOVFalse =
2901 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2903 if (CMOVTrue && CMOVFalse) {
2904 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2905 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2909 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2911 False = SelectFalse;
2912 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2917 if (True.getNode() && False.getNode()) {
2918 EVT VT = Op.getValueType();
2919 SDValue ARMcc = Cond.getOperand(2);
2920 SDValue CCR = Cond.getOperand(3);
2921 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2922 assert(True.getValueType() == VT);
2923 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2928 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2929 // undefined bits before doing a full-word comparison with zero.
2930 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2931 DAG.getConstant(1, Cond.getValueType()));
2933 return DAG.getSelectCC(dl, Cond,
2934 DAG.getConstant(0, Cond.getValueType()),
2935 SelectTrue, SelectFalse, ISD::SETNE);
2938 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2939 EVT VT = Op.getValueType();
2940 SDValue LHS = Op.getOperand(0);
2941 SDValue RHS = Op.getOperand(1);
2942 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2943 SDValue TrueVal = Op.getOperand(2);
2944 SDValue FalseVal = Op.getOperand(3);
2945 DebugLoc dl = Op.getDebugLoc();
2947 if (LHS.getValueType() == MVT::i32) {
2949 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2950 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2951 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2954 ARMCC::CondCodes CondCode, CondCode2;
2955 FPCCToARMCC(CC, CondCode, CondCode2);
2957 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2958 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2959 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2960 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2962 if (CondCode2 != ARMCC::AL) {
2963 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2964 // FIXME: Needs another CMP because flag can have but one use.
2965 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2966 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2967 Result, TrueVal, ARMcc2, CCR, Cmp2);
2972 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2973 /// to morph to an integer compare sequence.
2974 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2975 const ARMSubtarget *Subtarget) {
2976 SDNode *N = Op.getNode();
2977 if (!N->hasOneUse())
2978 // Otherwise it requires moving the value from fp to integer registers.
2980 if (!N->getNumValues())
2982 EVT VT = Op.getValueType();
2983 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2984 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2985 // vmrs are very slow, e.g. cortex-a8.
2988 if (isFloatingPointZero(Op)) {
2992 return ISD::isNormalLoad(N);
2995 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2996 if (isFloatingPointZero(Op))
2997 return DAG.getConstant(0, MVT::i32);
2999 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3000 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3001 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3002 Ld->isVolatile(), Ld->isNonTemporal(),
3003 Ld->isInvariant(), Ld->getAlignment());
3005 llvm_unreachable("Unknown VFP cmp argument!");
3008 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3009 SDValue &RetVal1, SDValue &RetVal2) {
3010 if (isFloatingPointZero(Op)) {
3011 RetVal1 = DAG.getConstant(0, MVT::i32);
3012 RetVal2 = DAG.getConstant(0, MVT::i32);
3016 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3017 SDValue Ptr = Ld->getBasePtr();
3018 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3019 Ld->getChain(), Ptr,
3020 Ld->getPointerInfo(),
3021 Ld->isVolatile(), Ld->isNonTemporal(),
3022 Ld->isInvariant(), Ld->getAlignment());
3024 EVT PtrType = Ptr.getValueType();
3025 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3026 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3027 PtrType, Ptr, DAG.getConstant(4, PtrType));
3028 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3029 Ld->getChain(), NewPtr,
3030 Ld->getPointerInfo().getWithOffset(4),
3031 Ld->isVolatile(), Ld->isNonTemporal(),
3032 Ld->isInvariant(), NewAlign);
3036 llvm_unreachable("Unknown VFP cmp argument!");
3039 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3040 /// f32 and even f64 comparisons to integer ones.
3042 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3043 SDValue Chain = Op.getOperand(0);
3044 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3045 SDValue LHS = Op.getOperand(2);
3046 SDValue RHS = Op.getOperand(3);
3047 SDValue Dest = Op.getOperand(4);
3048 DebugLoc dl = Op.getDebugLoc();
3050 bool LHSSeenZero = false;
3051 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3052 bool RHSSeenZero = false;
3053 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3054 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3055 // If unsafe fp math optimization is enabled and there are no other uses of
3056 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3057 // to an integer comparison.
3058 if (CC == ISD::SETOEQ)
3060 else if (CC == ISD::SETUNE)
3063 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3065 if (LHS.getValueType() == MVT::f32) {
3066 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3067 bitcastf32Toi32(LHS, DAG), Mask);
3068 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3069 bitcastf32Toi32(RHS, DAG), Mask);
3070 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3071 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3072 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3073 Chain, Dest, ARMcc, CCR, Cmp);
3078 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3079 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3080 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3081 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3082 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3083 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3084 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3085 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3086 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3092 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3093 SDValue Chain = Op.getOperand(0);
3094 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3095 SDValue LHS = Op.getOperand(2);
3096 SDValue RHS = Op.getOperand(3);
3097 SDValue Dest = Op.getOperand(4);
3098 DebugLoc dl = Op.getDebugLoc();
3100 if (LHS.getValueType() == MVT::i32) {
3102 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3103 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3104 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3105 Chain, Dest, ARMcc, CCR, Cmp);
3108 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3110 if (getTargetMachine().Options.UnsafeFPMath &&
3111 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3112 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3113 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3114 if (Result.getNode())
3118 ARMCC::CondCodes CondCode, CondCode2;
3119 FPCCToARMCC(CC, CondCode, CondCode2);
3121 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3122 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3123 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3124 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3125 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3126 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3127 if (CondCode2 != ARMCC::AL) {
3128 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3129 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3130 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3135 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3136 SDValue Chain = Op.getOperand(0);
3137 SDValue Table = Op.getOperand(1);
3138 SDValue Index = Op.getOperand(2);
3139 DebugLoc dl = Op.getDebugLoc();
3141 EVT PTy = getPointerTy();
3142 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3143 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3144 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3145 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3146 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3147 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3148 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3149 if (Subtarget->isThumb2()) {
3150 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3151 // which does another jump to the destination. This also makes it easier
3152 // to translate it to TBB / TBH later.
3153 // FIXME: This might not work if the function is extremely large.
3154 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3155 Addr, Op.getOperand(2), JTI, UId);
3157 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3158 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3159 MachinePointerInfo::getJumpTable(),
3160 false, false, false, 0);
3161 Chain = Addr.getValue(1);
3162 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3163 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3165 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3166 MachinePointerInfo::getJumpTable(),
3167 false, false, false, 0);
3168 Chain = Addr.getValue(1);
3169 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3173 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3174 EVT VT = Op.getValueType();
3175 DebugLoc dl = Op.getDebugLoc();
3177 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3178 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3180 return DAG.UnrollVectorOp(Op.getNode());
3183 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3184 "Invalid type for custom lowering!");
3185 if (VT != MVT::v4i16)
3186 return DAG.UnrollVectorOp(Op.getNode());
3188 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3189 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3192 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3193 EVT VT = Op.getValueType();
3195 return LowerVectorFP_TO_INT(Op, DAG);
3197 DebugLoc dl = Op.getDebugLoc();
3200 switch (Op.getOpcode()) {
3201 default: llvm_unreachable("Invalid opcode!");
3202 case ISD::FP_TO_SINT:
3203 Opc = ARMISD::FTOSI;
3205 case ISD::FP_TO_UINT:
3206 Opc = ARMISD::FTOUI;
3209 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3210 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3213 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3214 EVT VT = Op.getValueType();
3215 DebugLoc dl = Op.getDebugLoc();
3217 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3218 if (VT.getVectorElementType() == MVT::f32)
3220 return DAG.UnrollVectorOp(Op.getNode());
3223 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3224 "Invalid type for custom lowering!");
3225 if (VT != MVT::v4f32)
3226 return DAG.UnrollVectorOp(Op.getNode());
3230 switch (Op.getOpcode()) {
3231 default: llvm_unreachable("Invalid opcode!");
3232 case ISD::SINT_TO_FP:
3233 CastOpc = ISD::SIGN_EXTEND;
3234 Opc = ISD::SINT_TO_FP;
3236 case ISD::UINT_TO_FP:
3237 CastOpc = ISD::ZERO_EXTEND;
3238 Opc = ISD::UINT_TO_FP;
3242 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3243 return DAG.getNode(Opc, dl, VT, Op);
3246 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3247 EVT VT = Op.getValueType();
3249 return LowerVectorINT_TO_FP(Op, DAG);
3251 DebugLoc dl = Op.getDebugLoc();
3254 switch (Op.getOpcode()) {
3255 default: llvm_unreachable("Invalid opcode!");
3256 case ISD::SINT_TO_FP:
3257 Opc = ARMISD::SITOF;
3259 case ISD::UINT_TO_FP:
3260 Opc = ARMISD::UITOF;
3264 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3265 return DAG.getNode(Opc, dl, VT, Op);
3268 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3269 // Implement fcopysign with a fabs and a conditional fneg.
3270 SDValue Tmp0 = Op.getOperand(0);
3271 SDValue Tmp1 = Op.getOperand(1);
3272 DebugLoc dl = Op.getDebugLoc();
3273 EVT VT = Op.getValueType();
3274 EVT SrcVT = Tmp1.getValueType();
3275 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3276 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3277 bool UseNEON = !InGPR && Subtarget->hasNEON();
3280 // Use VBSL to copy the sign bit.
3281 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3282 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3283 DAG.getTargetConstant(EncodedVal, MVT::i32));
3284 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3286 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3287 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3288 DAG.getConstant(32, MVT::i32));
3289 else /*if (VT == MVT::f32)*/
3290 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3291 if (SrcVT == MVT::f32) {
3292 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3294 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3295 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3296 DAG.getConstant(32, MVT::i32));
3297 } else if (VT == MVT::f32)
3298 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3299 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3300 DAG.getConstant(32, MVT::i32));
3301 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3302 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3304 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3306 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3307 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3308 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3310 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3311 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3312 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3313 if (VT == MVT::f32) {
3314 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3315 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3316 DAG.getConstant(0, MVT::i32));
3318 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3324 // Bitcast operand 1 to i32.
3325 if (SrcVT == MVT::f64)
3326 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3327 &Tmp1, 1).getValue(1);
3328 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3330 // Or in the signbit with integer operations.
3331 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3332 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3333 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3334 if (VT == MVT::f32) {
3335 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3336 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3337 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3338 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3341 // f64: Or the high part with signbit and then combine two parts.
3342 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3344 SDValue Lo = Tmp0.getValue(0);
3345 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3346 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3347 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3350 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3351 MachineFunction &MF = DAG.getMachineFunction();
3352 MachineFrameInfo *MFI = MF.getFrameInfo();
3353 MFI->setReturnAddressIsTaken(true);
3355 EVT VT = Op.getValueType();
3356 DebugLoc dl = Op.getDebugLoc();
3357 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3359 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3360 SDValue Offset = DAG.getConstant(4, MVT::i32);
3361 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3362 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3363 MachinePointerInfo(), false, false, false, 0);
3366 // Return LR, which contains the return address. Mark it an implicit live-in.
3367 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3368 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3371 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3372 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3373 MFI->setFrameAddressIsTaken(true);
3375 EVT VT = Op.getValueType();
3376 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3377 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3378 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3379 ? ARM::R7 : ARM::R11;
3380 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3382 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3383 MachinePointerInfo(),
3384 false, false, false, 0);
3388 /// ExpandBITCAST - If the target supports VFP, this function is called to
3389 /// expand a bit convert where either the source or destination type is i64 to
3390 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3391 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3392 /// vectors), since the legalizer won't know what to do with that.
3393 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3394 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3395 DebugLoc dl = N->getDebugLoc();
3396 SDValue Op = N->getOperand(0);
3398 // This function is only supposed to be called for i64 types, either as the
3399 // source or destination of the bit convert.
3400 EVT SrcVT = Op.getValueType();
3401 EVT DstVT = N->getValueType(0);
3402 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3403 "ExpandBITCAST called for non-i64 type");
3405 // Turn i64->f64 into VMOVDRR.
3406 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3407 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3408 DAG.getConstant(0, MVT::i32));
3409 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3410 DAG.getConstant(1, MVT::i32));
3411 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3412 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3415 // Turn f64->i64 into VMOVRRD.
3416 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3417 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3418 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3419 // Merge the pieces into a single i64 value.
3420 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3426 /// getZeroVector - Returns a vector of specified type with all zero elements.
3427 /// Zero vectors are used to represent vector negation and in those cases
3428 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3429 /// not support i64 elements, so sometimes the zero vectors will need to be
3430 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3432 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3433 assert(VT.isVector() && "Expected a vector type");
3434 // The canonical modified immediate encoding of a zero vector is....0!
3435 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3436 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3437 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3438 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3441 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3442 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3443 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3444 SelectionDAG &DAG) const {
3445 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3446 EVT VT = Op.getValueType();
3447 unsigned VTBits = VT.getSizeInBits();
3448 DebugLoc dl = Op.getDebugLoc();
3449 SDValue ShOpLo = Op.getOperand(0);
3450 SDValue ShOpHi = Op.getOperand(1);
3451 SDValue ShAmt = Op.getOperand(2);
3453 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3455 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3457 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3458 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3459 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3460 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3461 DAG.getConstant(VTBits, MVT::i32));
3462 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3463 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3464 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3466 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3467 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3469 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3470 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3473 SDValue Ops[2] = { Lo, Hi };
3474 return DAG.getMergeValues(Ops, 2, dl);
3477 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3478 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3479 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3480 SelectionDAG &DAG) const {
3481 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3482 EVT VT = Op.getValueType();
3483 unsigned VTBits = VT.getSizeInBits();
3484 DebugLoc dl = Op.getDebugLoc();
3485 SDValue ShOpLo = Op.getOperand(0);
3486 SDValue ShOpHi = Op.getOperand(1);
3487 SDValue ShAmt = Op.getOperand(2);
3490 assert(Op.getOpcode() == ISD::SHL_PARTS);
3491 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3492 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3493 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3494 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3495 DAG.getConstant(VTBits, MVT::i32));
3496 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3497 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3499 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3500 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3501 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3503 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3504 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3507 SDValue Ops[2] = { Lo, Hi };
3508 return DAG.getMergeValues(Ops, 2, dl);
3511 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3512 SelectionDAG &DAG) const {
3513 // The rounding mode is in bits 23:22 of the FPSCR.
3514 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3515 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3516 // so that the shift + and get folded into a bitfield extract.
3517 DebugLoc dl = Op.getDebugLoc();
3518 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3519 DAG.getConstant(Intrinsic::arm_get_fpscr,
3521 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3522 DAG.getConstant(1U << 22, MVT::i32));
3523 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3524 DAG.getConstant(22, MVT::i32));
3525 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3526 DAG.getConstant(3, MVT::i32));
3529 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3530 const ARMSubtarget *ST) {
3531 EVT VT = N->getValueType(0);
3532 DebugLoc dl = N->getDebugLoc();
3534 if (!ST->hasV6T2Ops())
3537 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3538 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3541 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3542 const ARMSubtarget *ST) {
3543 EVT VT = N->getValueType(0);
3544 DebugLoc dl = N->getDebugLoc();
3549 // Lower vector shifts on NEON to use VSHL.
3550 assert(ST->hasNEON() && "unexpected vector shift");
3552 // Left shifts translate directly to the vshiftu intrinsic.
3553 if (N->getOpcode() == ISD::SHL)
3554 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3555 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3556 N->getOperand(0), N->getOperand(1));
3558 assert((N->getOpcode() == ISD::SRA ||
3559 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3561 // NEON uses the same intrinsics for both left and right shifts. For
3562 // right shifts, the shift amounts are negative, so negate the vector of
3564 EVT ShiftVT = N->getOperand(1).getValueType();
3565 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3566 getZeroVector(ShiftVT, DAG, dl),
3568 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3569 Intrinsic::arm_neon_vshifts :
3570 Intrinsic::arm_neon_vshiftu);
3571 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3572 DAG.getConstant(vshiftInt, MVT::i32),
3573 N->getOperand(0), NegatedCount);
3576 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3577 const ARMSubtarget *ST) {
3578 EVT VT = N->getValueType(0);
3579 DebugLoc dl = N->getDebugLoc();
3581 // We can get here for a node like i32 = ISD::SHL i32, i64
3585 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3586 "Unknown shift to lower!");
3588 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3589 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3590 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3593 // If we are in thumb mode, we don't have RRX.
3594 if (ST->isThumb1Only()) return SDValue();
3596 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3597 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3598 DAG.getConstant(0, MVT::i32));
3599 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3600 DAG.getConstant(1, MVT::i32));
3602 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3603 // captures the result into a carry flag.
3604 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3605 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3607 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3608 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3610 // Merge the pieces into a single i64 value.
3611 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3614 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3615 SDValue TmpOp0, TmpOp1;
3616 bool Invert = false;
3620 SDValue Op0 = Op.getOperand(0);
3621 SDValue Op1 = Op.getOperand(1);
3622 SDValue CC = Op.getOperand(2);
3623 EVT VT = Op.getValueType();
3624 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3625 DebugLoc dl = Op.getDebugLoc();
3627 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3628 switch (SetCCOpcode) {
3629 default: llvm_unreachable("Illegal FP comparison");
3631 case ISD::SETNE: Invert = true; // Fallthrough
3633 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3635 case ISD::SETLT: Swap = true; // Fallthrough
3637 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3639 case ISD::SETLE: Swap = true; // Fallthrough
3641 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3642 case ISD::SETUGE: Swap = true; // Fallthrough
3643 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3644 case ISD::SETUGT: Swap = true; // Fallthrough
3645 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3646 case ISD::SETUEQ: Invert = true; // Fallthrough
3648 // Expand this to (OLT | OGT).
3652 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3653 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3655 case ISD::SETUO: Invert = true; // Fallthrough
3657 // Expand this to (OLT | OGE).
3661 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3662 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3666 // Integer comparisons.
3667 switch (SetCCOpcode) {
3668 default: llvm_unreachable("Illegal integer comparison");
3669 case ISD::SETNE: Invert = true;
3670 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3671 case ISD::SETLT: Swap = true;
3672 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3673 case ISD::SETLE: Swap = true;
3674 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3675 case ISD::SETULT: Swap = true;
3676 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3677 case ISD::SETULE: Swap = true;
3678 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3681 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3682 if (Opc == ARMISD::VCEQ) {
3685 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3687 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3690 // Ignore bitconvert.
3691 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3692 AndOp = AndOp.getOperand(0);
3694 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3696 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3697 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3704 std::swap(Op0, Op1);
3706 // If one of the operands is a constant vector zero, attempt to fold the
3707 // comparison to a specialized compare-against-zero form.
3709 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3711 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3712 if (Opc == ARMISD::VCGE)
3713 Opc = ARMISD::VCLEZ;
3714 else if (Opc == ARMISD::VCGT)
3715 Opc = ARMISD::VCLTZ;
3720 if (SingleOp.getNode()) {
3723 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3725 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3727 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3729 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3731 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3733 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3736 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3740 Result = DAG.getNOT(dl, Result, VT);
3745 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3746 /// valid vector constant for a NEON instruction with a "modified immediate"
3747 /// operand (e.g., VMOV). If so, return the encoded value.
3748 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3749 unsigned SplatBitSize, SelectionDAG &DAG,
3750 EVT &VT, bool is128Bits, NEONModImmType type) {
3751 unsigned OpCmode, Imm;
3753 // SplatBitSize is set to the smallest size that splats the vector, so a
3754 // zero vector will always have SplatBitSize == 8. However, NEON modified
3755 // immediate instructions others than VMOV do not support the 8-bit encoding
3756 // of a zero vector, and the default encoding of zero is supposed to be the
3761 switch (SplatBitSize) {
3763 if (type != VMOVModImm)
3765 // Any 1-byte value is OK. Op=0, Cmode=1110.
3766 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3769 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3773 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3774 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3775 if ((SplatBits & ~0xff) == 0) {
3776 // Value = 0x00nn: Op=x, Cmode=100x.
3781 if ((SplatBits & ~0xff00) == 0) {
3782 // Value = 0xnn00: Op=x, Cmode=101x.
3784 Imm = SplatBits >> 8;
3790 // NEON's 32-bit VMOV supports splat values where:
3791 // * only one byte is nonzero, or
3792 // * the least significant byte is 0xff and the second byte is nonzero, or
3793 // * the least significant 2 bytes are 0xff and the third is nonzero.
3794 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3795 if ((SplatBits & ~0xff) == 0) {
3796 // Value = 0x000000nn: Op=x, Cmode=000x.
3801 if ((SplatBits & ~0xff00) == 0) {
3802 // Value = 0x0000nn00: Op=x, Cmode=001x.
3804 Imm = SplatBits >> 8;
3807 if ((SplatBits & ~0xff0000) == 0) {
3808 // Value = 0x00nn0000: Op=x, Cmode=010x.
3810 Imm = SplatBits >> 16;
3813 if ((SplatBits & ~0xff000000) == 0) {
3814 // Value = 0xnn000000: Op=x, Cmode=011x.
3816 Imm = SplatBits >> 24;
3820 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3821 if (type == OtherModImm) return SDValue();
3823 if ((SplatBits & ~0xffff) == 0 &&
3824 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3825 // Value = 0x0000nnff: Op=x, Cmode=1100.
3827 Imm = SplatBits >> 8;
3832 if ((SplatBits & ~0xffffff) == 0 &&
3833 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3834 // Value = 0x00nnffff: Op=x, Cmode=1101.
3836 Imm = SplatBits >> 16;
3837 SplatBits |= 0xffff;
3841 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3842 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3843 // VMOV.I32. A (very) minor optimization would be to replicate the value
3844 // and fall through here to test for a valid 64-bit splat. But, then the
3845 // caller would also need to check and handle the change in size.
3849 if (type != VMOVModImm)
3851 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3852 uint64_t BitMask = 0xff;
3854 unsigned ImmMask = 1;
3856 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3857 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3860 } else if ((SplatBits & BitMask) != 0) {
3866 // Op=1, Cmode=1110.
3869 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3874 llvm_unreachable("unexpected size for isNEONModifiedImm");
3877 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3878 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3881 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3882 const ARMSubtarget *ST) const {
3883 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3886 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3887 assert(Op.getValueType() == MVT::f32 &&
3888 "ConstantFP custom lowering should only occur for f32.");
3890 // Try splatting with a VMOV.f32...
3891 APFloat FPVal = CFP->getValueAPF();
3892 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3894 DebugLoc DL = Op.getDebugLoc();
3895 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3896 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3898 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3899 DAG.getConstant(0, MVT::i32));
3902 // If that fails, try a VMOV.i32
3904 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3905 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3907 if (NewVal != SDValue()) {
3908 DebugLoc DL = Op.getDebugLoc();
3909 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3911 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3913 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3914 DAG.getConstant(0, MVT::i32));
3917 // Finally, try a VMVN.i32
3918 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3920 if (NewVal != SDValue()) {
3921 DebugLoc DL = Op.getDebugLoc();
3922 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3923 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3925 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3926 DAG.getConstant(0, MVT::i32));
3933 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3934 bool &ReverseVEXT, unsigned &Imm) {
3935 unsigned NumElts = VT.getVectorNumElements();
3936 ReverseVEXT = false;
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, it may still be
3950 // a VEXT but the source vectors must be swapped.
3952 if (ExpectedElt == NumElts * 2) {
3957 if (M[i] < 0) continue; // ignore UNDEF indices
3958 if (ExpectedElt != static_cast<unsigned>(M[i]))
3962 // Adjust the index value if the source operands will be swapped.
3969 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3970 /// instruction with the specified blocksize. (The order of the elements
3971 /// within each block of the vector is reversed.)
3972 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3973 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3974 "Only possible block sizes for VREV are: 16, 32, 64");
3976 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3980 unsigned NumElts = VT.getVectorNumElements();
3981 unsigned BlockElts = M[0] + 1;
3982 // If the first shuffle index is UNDEF, be optimistic.
3984 BlockElts = BlockSize / EltSz;
3986 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3989 for (unsigned i = 0; i < NumElts; ++i) {
3990 if (M[i] < 0) continue; // ignore UNDEF indices
3991 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3998 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3999 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4000 // range, then 0 is placed into the resulting vector. So pretty much any mask
4001 // of 8 elements can work here.
4002 return VT == MVT::v8i8 && M.size() == 8;
4005 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4006 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4010 unsigned NumElts = VT.getVectorNumElements();
4011 WhichResult = (M[0] == 0 ? 0 : 1);
4012 for (unsigned i = 0; i < NumElts; i += 2) {
4013 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4014 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4020 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4021 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4022 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4023 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4024 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4028 unsigned NumElts = VT.getVectorNumElements();
4029 WhichResult = (M[0] == 0 ? 0 : 1);
4030 for (unsigned i = 0; i < NumElts; i += 2) {
4031 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4032 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4038 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4039 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4043 unsigned NumElts = VT.getVectorNumElements();
4044 WhichResult = (M[0] == 0 ? 0 : 1);
4045 for (unsigned i = 0; i != NumElts; ++i) {
4046 if (M[i] < 0) continue; // ignore UNDEF indices
4047 if ((unsigned) M[i] != 2 * i + WhichResult)
4051 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4052 if (VT.is64BitVector() && EltSz == 32)
4058 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4059 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4060 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4061 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4062 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4066 unsigned Half = VT.getVectorNumElements() / 2;
4067 WhichResult = (M[0] == 0 ? 0 : 1);
4068 for (unsigned j = 0; j != 2; ++j) {
4069 unsigned Idx = WhichResult;
4070 for (unsigned i = 0; i != Half; ++i) {
4071 int MIdx = M[i + j * Half];
4072 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4078 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4079 if (VT.is64BitVector() && EltSz == 32)
4085 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4086 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4090 unsigned NumElts = VT.getVectorNumElements();
4091 WhichResult = (M[0] == 0 ? 0 : 1);
4092 unsigned Idx = WhichResult * NumElts / 2;
4093 for (unsigned i = 0; i != NumElts; i += 2) {
4094 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4095 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4100 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4101 if (VT.is64BitVector() && EltSz == 32)
4107 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4108 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4109 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4110 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4111 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4115 unsigned NumElts = VT.getVectorNumElements();
4116 WhichResult = (M[0] == 0 ? 0 : 1);
4117 unsigned Idx = WhichResult * NumElts / 2;
4118 for (unsigned i = 0; i != NumElts; i += 2) {
4119 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4120 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4125 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4126 if (VT.is64BitVector() && EltSz == 32)
4132 // If N is an integer constant that can be moved into a register in one
4133 // instruction, return an SDValue of such a constant (will become a MOV
4134 // instruction). Otherwise return null.
4135 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4136 const ARMSubtarget *ST, DebugLoc dl) {
4138 if (!isa<ConstantSDNode>(N))
4140 Val = cast<ConstantSDNode>(N)->getZExtValue();
4142 if (ST->isThumb1Only()) {
4143 if (Val <= 255 || ~Val <= 255)
4144 return DAG.getConstant(Val, MVT::i32);
4146 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4147 return DAG.getConstant(Val, MVT::i32);
4152 // If this is a case we can't handle, return null and let the default
4153 // expansion code take care of it.
4154 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4155 const ARMSubtarget *ST) const {
4156 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4157 DebugLoc dl = Op.getDebugLoc();
4158 EVT VT = Op.getValueType();
4160 APInt SplatBits, SplatUndef;
4161 unsigned SplatBitSize;
4163 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4164 if (SplatBitSize <= 64) {
4165 // Check if an immediate VMOV works.
4167 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4168 SplatUndef.getZExtValue(), SplatBitSize,
4169 DAG, VmovVT, VT.is128BitVector(),
4171 if (Val.getNode()) {
4172 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4173 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4176 // Try an immediate VMVN.
4177 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4178 Val = isNEONModifiedImm(NegatedImm,
4179 SplatUndef.getZExtValue(), SplatBitSize,
4180 DAG, VmovVT, VT.is128BitVector(),
4182 if (Val.getNode()) {
4183 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4184 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4187 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4188 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4189 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4191 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4192 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4198 // Scan through the operands to see if only one value is used.
4200 // As an optimisation, even if more than one value is used it may be more
4201 // profitable to splat with one value then change some lanes.
4203 // Heuristically we decide to do this if the vector has a "dominant" value,
4204 // defined as splatted to more than half of the lanes.
4205 unsigned NumElts = VT.getVectorNumElements();
4206 bool isOnlyLowElement = true;
4207 bool usesOnlyOneValue = true;
4208 bool hasDominantValue = false;
4209 bool isConstant = true;
4211 // Map of the number of times a particular SDValue appears in the
4213 DenseMap<SDValue, unsigned> ValueCounts;
4215 for (unsigned i = 0; i < NumElts; ++i) {
4216 SDValue V = Op.getOperand(i);
4217 if (V.getOpcode() == ISD::UNDEF)
4220 isOnlyLowElement = false;
4221 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4224 ValueCounts.insert(std::make_pair(V, 0));
4225 unsigned &Count = ValueCounts[V];
4227 // Is this value dominant? (takes up more than half of the lanes)
4228 if (++Count > (NumElts / 2)) {
4229 hasDominantValue = true;
4233 if (ValueCounts.size() != 1)
4234 usesOnlyOneValue = false;
4235 if (!Value.getNode() && ValueCounts.size() > 0)
4236 Value = ValueCounts.begin()->first;
4238 if (ValueCounts.size() == 0)
4239 return DAG.getUNDEF(VT);
4241 if (isOnlyLowElement)
4242 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4244 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4246 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4247 // i32 and try again.
4248 if (hasDominantValue && EltSize <= 32) {
4252 // If we are VDUPing a value that comes directly from a vector, that will
4253 // cause an unnecessary move to and from a GPR, where instead we could
4254 // just use VDUPLANE.
4255 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4256 // We need to create a new undef vector to use for the VDUPLANE if the
4257 // size of the vector from which we get the value is different than the
4258 // size of the vector that we need to create. We will insert the element
4259 // such that the register coalescer will remove unnecessary copies.
4260 if (VT != Value->getOperand(0).getValueType()) {
4261 ConstantSDNode *constIndex;
4262 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4263 assert(constIndex && "The index is not a constant!");
4264 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4265 VT.getVectorNumElements();
4266 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4267 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4268 Value, DAG.getConstant(index, MVT::i32)),
4269 DAG.getConstant(index, MVT::i32));
4271 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4272 Value->getOperand(0), Value->getOperand(1));
4276 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4278 if (!usesOnlyOneValue) {
4279 // The dominant value was splatted as 'N', but we now have to insert
4280 // all differing elements.
4281 for (unsigned I = 0; I < NumElts; ++I) {
4282 if (Op.getOperand(I) == Value)
4284 SmallVector<SDValue, 3> Ops;
4286 Ops.push_back(Op.getOperand(I));
4287 Ops.push_back(DAG.getConstant(I, MVT::i32));
4288 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4293 if (VT.getVectorElementType().isFloatingPoint()) {
4294 SmallVector<SDValue, 8> Ops;
4295 for (unsigned i = 0; i < NumElts; ++i)
4296 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4298 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4299 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4300 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4302 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4304 if (usesOnlyOneValue) {
4305 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4306 if (isConstant && Val.getNode())
4307 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4311 // If all elements are constants and the case above didn't get hit, fall back
4312 // to the default expansion, which will generate a load from the constant
4317 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4319 SDValue shuffle = ReconstructShuffle(Op, DAG);
4320 if (shuffle != SDValue())
4324 // Vectors with 32- or 64-bit elements can be built by directly assigning
4325 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4326 // will be legalized.
4327 if (EltSize >= 32) {
4328 // Do the expansion with floating-point types, since that is what the VFP
4329 // registers are defined to use, and since i64 is not legal.
4330 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4331 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4332 SmallVector<SDValue, 8> Ops;
4333 for (unsigned i = 0; i < NumElts; ++i)
4334 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4335 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4336 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4342 // Gather data to see if the operation can be modelled as a
4343 // shuffle in combination with VEXTs.
4344 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4345 SelectionDAG &DAG) const {
4346 DebugLoc dl = Op.getDebugLoc();
4347 EVT VT = Op.getValueType();
4348 unsigned NumElts = VT.getVectorNumElements();
4350 SmallVector<SDValue, 2> SourceVecs;
4351 SmallVector<unsigned, 2> MinElts;
4352 SmallVector<unsigned, 2> MaxElts;
4354 for (unsigned i = 0; i < NumElts; ++i) {
4355 SDValue V = Op.getOperand(i);
4356 if (V.getOpcode() == ISD::UNDEF)
4358 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4359 // A shuffle can only come from building a vector from various
4360 // elements of other vectors.
4362 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4363 VT.getVectorElementType()) {
4364 // This code doesn't know how to handle shuffles where the vector
4365 // element types do not match (this happens because type legalization
4366 // promotes the return type of EXTRACT_VECTOR_ELT).
4367 // FIXME: It might be appropriate to extend this code to handle
4368 // mismatched types.
4372 // Record this extraction against the appropriate vector if possible...
4373 SDValue SourceVec = V.getOperand(0);
4374 // If the element number isn't a constant, we can't effectively
4375 // analyze what's going on.
4376 if (!isa<ConstantSDNode>(V.getOperand(1)))
4378 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4379 bool FoundSource = false;
4380 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4381 if (SourceVecs[j] == SourceVec) {
4382 if (MinElts[j] > EltNo)
4384 if (MaxElts[j] < EltNo)
4391 // Or record a new source if not...
4393 SourceVecs.push_back(SourceVec);
4394 MinElts.push_back(EltNo);
4395 MaxElts.push_back(EltNo);
4399 // Currently only do something sane when at most two source vectors
4401 if (SourceVecs.size() > 2)
4404 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4405 int VEXTOffsets[2] = {0, 0};
4407 // This loop extracts the usage patterns of the source vectors
4408 // and prepares appropriate SDValues for a shuffle if possible.
4409 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4410 if (SourceVecs[i].getValueType() == VT) {
4411 // No VEXT necessary
4412 ShuffleSrcs[i] = SourceVecs[i];
4415 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4416 // It probably isn't worth padding out a smaller vector just to
4417 // break it down again in a shuffle.
4421 // Since only 64-bit and 128-bit vectors are legal on ARM and
4422 // we've eliminated the other cases...
4423 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4424 "unexpected vector sizes in ReconstructShuffle");
4426 if (MaxElts[i] - MinElts[i] >= NumElts) {
4427 // Span too large for a VEXT to cope
4431 if (MinElts[i] >= NumElts) {
4432 // The extraction can just take the second half
4433 VEXTOffsets[i] = NumElts;
4434 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4436 DAG.getIntPtrConstant(NumElts));
4437 } else if (MaxElts[i] < NumElts) {
4438 // The extraction can just take the first half
4440 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4442 DAG.getIntPtrConstant(0));
4444 // An actual VEXT is needed
4445 VEXTOffsets[i] = MinElts[i];
4446 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4448 DAG.getIntPtrConstant(0));
4449 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4451 DAG.getIntPtrConstant(NumElts));
4452 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4453 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4457 SmallVector<int, 8> Mask;
4459 for (unsigned i = 0; i < NumElts; ++i) {
4460 SDValue Entry = Op.getOperand(i);
4461 if (Entry.getOpcode() == ISD::UNDEF) {
4466 SDValue ExtractVec = Entry.getOperand(0);
4467 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4468 .getOperand(1))->getSExtValue();
4469 if (ExtractVec == SourceVecs[0]) {
4470 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4472 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4476 // Final check before we try to produce nonsense...
4477 if (isShuffleMaskLegal(Mask, VT))
4478 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4484 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4485 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4486 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4487 /// are assumed to be legal.
4489 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4491 if (VT.getVectorNumElements() == 4 &&
4492 (VT.is128BitVector() || VT.is64BitVector())) {
4493 unsigned PFIndexes[4];
4494 for (unsigned i = 0; i != 4; ++i) {
4498 PFIndexes[i] = M[i];
4501 // Compute the index in the perfect shuffle table.
4502 unsigned PFTableIndex =
4503 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4504 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4505 unsigned Cost = (PFEntry >> 30);
4512 unsigned Imm, WhichResult;
4514 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4515 return (EltSize >= 32 ||
4516 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4517 isVREVMask(M, VT, 64) ||
4518 isVREVMask(M, VT, 32) ||
4519 isVREVMask(M, VT, 16) ||
4520 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4521 isVTBLMask(M, VT) ||
4522 isVTRNMask(M, VT, WhichResult) ||
4523 isVUZPMask(M, VT, WhichResult) ||
4524 isVZIPMask(M, VT, WhichResult) ||
4525 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4526 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4527 isVZIP_v_undef_Mask(M, VT, WhichResult));
4530 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4531 /// the specified operations to build the shuffle.
4532 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4533 SDValue RHS, SelectionDAG &DAG,
4535 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4536 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4537 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4540 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4549 OP_VUZPL, // VUZP, left result
4550 OP_VUZPR, // VUZP, right result
4551 OP_VZIPL, // VZIP, left result
4552 OP_VZIPR, // VZIP, right result
4553 OP_VTRNL, // VTRN, left result
4554 OP_VTRNR // VTRN, right result
4557 if (OpNum == OP_COPY) {
4558 if (LHSID == (1*9+2)*9+3) return LHS;
4559 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4563 SDValue OpLHS, OpRHS;
4564 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4565 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4566 EVT VT = OpLHS.getValueType();
4569 default: llvm_unreachable("Unknown shuffle opcode!");
4571 // VREV divides the vector in half and swaps within the half.
4572 if (VT.getVectorElementType() == MVT::i32 ||
4573 VT.getVectorElementType() == MVT::f32)
4574 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4575 // vrev <4 x i16> -> VREV32
4576 if (VT.getVectorElementType() == MVT::i16)
4577 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4578 // vrev <4 x i8> -> VREV16
4579 assert(VT.getVectorElementType() == MVT::i8);
4580 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4585 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4586 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4590 return DAG.getNode(ARMISD::VEXT, dl, VT,
4592 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4595 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4596 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4599 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4600 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4603 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4604 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4608 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4609 ArrayRef<int> ShuffleMask,
4610 SelectionDAG &DAG) {
4611 // Check to see if we can use the VTBL instruction.
4612 SDValue V1 = Op.getOperand(0);
4613 SDValue V2 = Op.getOperand(1);
4614 DebugLoc DL = Op.getDebugLoc();
4616 SmallVector<SDValue, 8> VTBLMask;
4617 for (ArrayRef<int>::iterator
4618 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4619 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4621 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4622 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4623 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4626 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4627 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4631 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4632 SDValue V1 = Op.getOperand(0);
4633 SDValue V2 = Op.getOperand(1);
4634 DebugLoc dl = Op.getDebugLoc();
4635 EVT VT = Op.getValueType();
4636 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4638 // Convert shuffles that are directly supported on NEON to target-specific
4639 // DAG nodes, instead of keeping them as shuffles and matching them again
4640 // during code selection. This is more efficient and avoids the possibility
4641 // of inconsistencies between legalization and selection.
4642 // FIXME: floating-point vectors should be canonicalized to integer vectors
4643 // of the same time so that they get CSEd properly.
4644 ArrayRef<int> ShuffleMask = SVN->getMask();
4646 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4647 if (EltSize <= 32) {
4648 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4649 int Lane = SVN->getSplatIndex();
4650 // If this is undef splat, generate it via "just" vdup, if possible.
4651 if (Lane == -1) Lane = 0;
4653 // Test if V1 is a SCALAR_TO_VECTOR.
4654 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4655 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4657 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4658 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4660 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4661 !isa<ConstantSDNode>(V1.getOperand(0))) {
4662 bool IsScalarToVector = true;
4663 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4664 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4665 IsScalarToVector = false;
4668 if (IsScalarToVector)
4669 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4671 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4672 DAG.getConstant(Lane, MVT::i32));
4677 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4680 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4681 DAG.getConstant(Imm, MVT::i32));
4684 if (isVREVMask(ShuffleMask, VT, 64))
4685 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4686 if (isVREVMask(ShuffleMask, VT, 32))
4687 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4688 if (isVREVMask(ShuffleMask, VT, 16))
4689 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4691 // Check for Neon shuffles that modify both input vectors in place.
4692 // If both results are used, i.e., if there are two shuffles with the same
4693 // source operands and with masks corresponding to both results of one of
4694 // these operations, DAG memoization will ensure that a single node is
4695 // used for both shuffles.
4696 unsigned WhichResult;
4697 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4698 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4699 V1, V2).getValue(WhichResult);
4700 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4701 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4702 V1, V2).getValue(WhichResult);
4703 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4704 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4705 V1, V2).getValue(WhichResult);
4707 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4708 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4709 V1, V1).getValue(WhichResult);
4710 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4711 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4712 V1, V1).getValue(WhichResult);
4713 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4714 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4715 V1, V1).getValue(WhichResult);
4718 // If the shuffle is not directly supported and it has 4 elements, use
4719 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4720 unsigned NumElts = VT.getVectorNumElements();
4722 unsigned PFIndexes[4];
4723 for (unsigned i = 0; i != 4; ++i) {
4724 if (ShuffleMask[i] < 0)
4727 PFIndexes[i] = ShuffleMask[i];
4730 // Compute the index in the perfect shuffle table.
4731 unsigned PFTableIndex =
4732 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4733 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4734 unsigned Cost = (PFEntry >> 30);
4737 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4740 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4741 if (EltSize >= 32) {
4742 // Do the expansion with floating-point types, since that is what the VFP
4743 // registers are defined to use, and since i64 is not legal.
4744 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4745 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4746 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4747 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4748 SmallVector<SDValue, 8> Ops;
4749 for (unsigned i = 0; i < NumElts; ++i) {
4750 if (ShuffleMask[i] < 0)
4751 Ops.push_back(DAG.getUNDEF(EltVT));
4753 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4754 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4755 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4758 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4759 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4762 if (VT == MVT::v8i8) {
4763 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4764 if (NewOp.getNode())
4771 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4772 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4773 SDValue Lane = Op.getOperand(2);
4774 if (!isa<ConstantSDNode>(Lane))
4780 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4781 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4782 SDValue Lane = Op.getOperand(1);
4783 if (!isa<ConstantSDNode>(Lane))
4786 SDValue Vec = Op.getOperand(0);
4787 if (Op.getValueType() == MVT::i32 &&
4788 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4789 DebugLoc dl = Op.getDebugLoc();
4790 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4796 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4797 // The only time a CONCAT_VECTORS operation can have legal types is when
4798 // two 64-bit vectors are concatenated to a 128-bit vector.
4799 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4800 "unexpected CONCAT_VECTORS");
4801 DebugLoc dl = Op.getDebugLoc();
4802 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4803 SDValue Op0 = Op.getOperand(0);
4804 SDValue Op1 = Op.getOperand(1);
4805 if (Op0.getOpcode() != ISD::UNDEF)
4806 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4807 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4808 DAG.getIntPtrConstant(0));
4809 if (Op1.getOpcode() != ISD::UNDEF)
4810 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4811 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4812 DAG.getIntPtrConstant(1));
4813 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4816 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4817 /// element has been zero/sign-extended, depending on the isSigned parameter,
4818 /// from an integer type half its size.
4819 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4821 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4822 EVT VT = N->getValueType(0);
4823 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4824 SDNode *BVN = N->getOperand(0).getNode();
4825 if (BVN->getValueType(0) != MVT::v4i32 ||
4826 BVN->getOpcode() != ISD::BUILD_VECTOR)
4828 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4829 unsigned HiElt = 1 - LoElt;
4830 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4831 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4832 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4833 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4834 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4837 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4838 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4841 if (Hi0->isNullValue() && Hi1->isNullValue())
4847 if (N->getOpcode() != ISD::BUILD_VECTOR)
4850 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4851 SDNode *Elt = N->getOperand(i).getNode();
4852 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4853 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4854 unsigned HalfSize = EltSize / 2;
4856 if (!isIntN(HalfSize, C->getSExtValue()))
4859 if (!isUIntN(HalfSize, C->getZExtValue()))
4870 /// isSignExtended - Check if a node is a vector value that is sign-extended
4871 /// or a constant BUILD_VECTOR with sign-extended elements.
4872 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4873 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4875 if (isExtendedBUILD_VECTOR(N, DAG, true))
4880 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4881 /// or a constant BUILD_VECTOR with zero-extended elements.
4882 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4883 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4885 if (isExtendedBUILD_VECTOR(N, DAG, false))
4890 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4891 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4892 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4893 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4894 return N->getOperand(0);
4895 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4896 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4897 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4898 LD->isNonTemporal(), LD->isInvariant(),
4899 LD->getAlignment());
4900 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4901 // have been legalized as a BITCAST from v4i32.
4902 if (N->getOpcode() == ISD::BITCAST) {
4903 SDNode *BVN = N->getOperand(0).getNode();
4904 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4905 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4906 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4907 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4908 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4910 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4911 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4912 EVT VT = N->getValueType(0);
4913 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4914 unsigned NumElts = VT.getVectorNumElements();
4915 MVT TruncVT = MVT::getIntegerVT(EltSize);
4916 SmallVector<SDValue, 8> Ops;
4917 for (unsigned i = 0; i != NumElts; ++i) {
4918 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4919 const APInt &CInt = C->getAPIntValue();
4920 // Element types smaller than 32 bits are not legal, so use i32 elements.
4921 // The values are implicitly truncated so sext vs. zext doesn't matter.
4922 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4924 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4925 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4928 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4929 unsigned Opcode = N->getOpcode();
4930 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4931 SDNode *N0 = N->getOperand(0).getNode();
4932 SDNode *N1 = N->getOperand(1).getNode();
4933 return N0->hasOneUse() && N1->hasOneUse() &&
4934 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4939 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4940 unsigned Opcode = N->getOpcode();
4941 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4942 SDNode *N0 = N->getOperand(0).getNode();
4943 SDNode *N1 = N->getOperand(1).getNode();
4944 return N0->hasOneUse() && N1->hasOneUse() &&
4945 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4950 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4951 // Multiplications are only custom-lowered for 128-bit vectors so that
4952 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4953 EVT VT = Op.getValueType();
4954 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4955 SDNode *N0 = Op.getOperand(0).getNode();
4956 SDNode *N1 = Op.getOperand(1).getNode();
4957 unsigned NewOpc = 0;
4959 bool isN0SExt = isSignExtended(N0, DAG);
4960 bool isN1SExt = isSignExtended(N1, DAG);
4961 if (isN0SExt && isN1SExt)
4962 NewOpc = ARMISD::VMULLs;
4964 bool isN0ZExt = isZeroExtended(N0, DAG);
4965 bool isN1ZExt = isZeroExtended(N1, DAG);
4966 if (isN0ZExt && isN1ZExt)
4967 NewOpc = ARMISD::VMULLu;
4968 else if (isN1SExt || isN1ZExt) {
4969 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4970 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4971 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4972 NewOpc = ARMISD::VMULLs;
4974 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4975 NewOpc = ARMISD::VMULLu;
4977 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4979 NewOpc = ARMISD::VMULLu;
4985 if (VT == MVT::v2i64)
4986 // Fall through to expand this. It is not legal.
4989 // Other vector multiplications are legal.
4994 // Legalize to a VMULL instruction.
4995 DebugLoc DL = Op.getDebugLoc();
4997 SDValue Op1 = SkipExtension(N1, DAG);
4999 Op0 = SkipExtension(N0, DAG);
5000 assert(Op0.getValueType().is64BitVector() &&
5001 Op1.getValueType().is64BitVector() &&
5002 "unexpected types for extended operands to VMULL");
5003 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5006 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5007 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5014 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5015 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5016 EVT Op1VT = Op1.getValueType();
5017 return DAG.getNode(N0->getOpcode(), DL, VT,
5018 DAG.getNode(NewOpc, DL, VT,
5019 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5020 DAG.getNode(NewOpc, DL, VT,
5021 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5025 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5027 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5028 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5029 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5030 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5031 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5032 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5033 // Get reciprocal estimate.
5034 // float4 recip = vrecpeq_f32(yf);
5035 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5036 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5037 // Because char has a smaller range than uchar, we can actually get away
5038 // without any newton steps. This requires that we use a weird bias
5039 // of 0xb000, however (again, this has been exhaustively tested).
5040 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5041 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5042 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5043 Y = DAG.getConstant(0xb000, MVT::i32);
5044 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5045 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5046 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5047 // Convert back to short.
5048 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5049 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5054 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5056 // Convert to float.
5057 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5058 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5059 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5060 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5061 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5062 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5064 // Use reciprocal estimate and one refinement step.
5065 // float4 recip = vrecpeq_f32(yf);
5066 // recip *= vrecpsq_f32(yf, recip);
5067 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5068 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5069 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5070 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5072 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5073 // Because short has a smaller range than ushort, we can actually get away
5074 // with only a single newton step. This requires that we use a weird bias
5075 // of 89, however (again, this has been exhaustively tested).
5076 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5077 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5078 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5079 N1 = DAG.getConstant(0x89, MVT::i32);
5080 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5081 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5082 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5083 // Convert back to integer and return.
5084 // return vmovn_s32(vcvt_s32_f32(result));
5085 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5086 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5090 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5091 EVT VT = Op.getValueType();
5092 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5093 "unexpected type for custom-lowering ISD::SDIV");
5095 DebugLoc dl = Op.getDebugLoc();
5096 SDValue N0 = Op.getOperand(0);
5097 SDValue N1 = Op.getOperand(1);
5100 if (VT == MVT::v8i8) {
5101 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5102 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5104 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5105 DAG.getIntPtrConstant(4));
5106 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5107 DAG.getIntPtrConstant(4));
5108 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5109 DAG.getIntPtrConstant(0));
5110 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5111 DAG.getIntPtrConstant(0));
5113 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5114 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5116 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5117 N0 = LowerCONCAT_VECTORS(N0, DAG);
5119 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5122 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5125 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5126 EVT VT = Op.getValueType();
5127 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5128 "unexpected type for custom-lowering ISD::UDIV");
5130 DebugLoc dl = Op.getDebugLoc();
5131 SDValue N0 = Op.getOperand(0);
5132 SDValue N1 = Op.getOperand(1);
5135 if (VT == MVT::v8i8) {
5136 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5137 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5139 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5140 DAG.getIntPtrConstant(4));
5141 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5142 DAG.getIntPtrConstant(4));
5143 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5144 DAG.getIntPtrConstant(0));
5145 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5146 DAG.getIntPtrConstant(0));
5148 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5149 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5151 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5152 N0 = LowerCONCAT_VECTORS(N0, DAG);
5154 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5155 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5160 // v4i16 sdiv ... Convert to float.
5161 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5162 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5163 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5164 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5165 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5166 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5168 // Use reciprocal estimate and two refinement steps.
5169 // float4 recip = vrecpeq_f32(yf);
5170 // recip *= vrecpsq_f32(yf, recip);
5171 // recip *= vrecpsq_f32(yf, recip);
5172 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5173 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5174 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5175 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5177 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5178 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5179 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5181 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5182 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5183 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5184 // and that it will never cause us to return an answer too large).
5185 // float4 result = as_float4(as_int4(xf*recip) + 2);
5186 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5187 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5188 N1 = DAG.getConstant(2, MVT::i32);
5189 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5190 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5191 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5192 // Convert back to integer and return.
5193 // return vmovn_u32(vcvt_s32_f32(result));
5194 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5195 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5199 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5200 EVT VT = Op.getNode()->getValueType(0);
5201 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5204 bool ExtraOp = false;
5205 switch (Op.getOpcode()) {
5206 default: llvm_unreachable("Invalid code");
5207 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5208 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5209 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5210 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5214 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5216 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5217 Op.getOperand(1), Op.getOperand(2));
5220 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5221 // Monotonic load/store is legal for all targets
5222 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5225 // Aquire/Release load/store is not legal for targets without a
5226 // dmb or equivalent available.
5232 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5233 SelectionDAG &DAG, unsigned NewOp) {
5234 DebugLoc dl = Node->getDebugLoc();
5235 assert (Node->getValueType(0) == MVT::i64 &&
5236 "Only know how to expand i64 atomics");
5238 SmallVector<SDValue, 6> Ops;
5239 Ops.push_back(Node->getOperand(0)); // Chain
5240 Ops.push_back(Node->getOperand(1)); // Ptr
5242 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5243 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5244 // High part of Val1
5245 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5246 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5247 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5248 // High part of Val1
5249 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5250 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5251 // High part of Val2
5252 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5253 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5255 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5257 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5258 cast<MemSDNode>(Node)->getMemOperand());
5259 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5260 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5261 Results.push_back(Result.getValue(2));
5264 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5265 switch (Op.getOpcode()) {
5266 default: llvm_unreachable("Don't know how to custom lower this!");
5267 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5268 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5269 case ISD::GlobalAddress:
5270 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5271 LowerGlobalAddressELF(Op, DAG);
5272 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5273 case ISD::SELECT: return LowerSELECT(Op, DAG);
5274 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5275 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5276 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5277 case ISD::VASTART: return LowerVASTART(Op, DAG);
5278 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5279 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5280 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5281 case ISD::SINT_TO_FP:
5282 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5283 case ISD::FP_TO_SINT:
5284 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5285 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5286 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5287 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5288 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5289 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5290 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5291 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5293 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5296 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5297 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5298 case ISD::SRL_PARTS:
5299 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5300 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5301 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5302 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5303 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5304 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5305 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5306 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5307 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5308 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5309 case ISD::MUL: return LowerMUL(Op, DAG);
5310 case ISD::SDIV: return LowerSDIV(Op, DAG);
5311 case ISD::UDIV: return LowerUDIV(Op, DAG);
5315 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5316 case ISD::ATOMIC_LOAD:
5317 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5321 /// ReplaceNodeResults - Replace the results of node with an illegal result
5322 /// type with new values built out of custom code.
5323 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5324 SmallVectorImpl<SDValue>&Results,
5325 SelectionDAG &DAG) const {
5327 switch (N->getOpcode()) {
5329 llvm_unreachable("Don't know how to custom expand this!");
5331 Res = ExpandBITCAST(N, DAG);
5335 Res = Expand64BitShift(N, DAG, Subtarget);
5337 case ISD::ATOMIC_LOAD_ADD:
5338 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5340 case ISD::ATOMIC_LOAD_AND:
5341 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5343 case ISD::ATOMIC_LOAD_NAND:
5344 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5346 case ISD::ATOMIC_LOAD_OR:
5347 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5349 case ISD::ATOMIC_LOAD_SUB:
5350 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5352 case ISD::ATOMIC_LOAD_XOR:
5353 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5355 case ISD::ATOMIC_SWAP:
5356 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5358 case ISD::ATOMIC_CMP_SWAP:
5359 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5363 Results.push_back(Res);
5366 //===----------------------------------------------------------------------===//
5367 // ARM Scheduler Hooks
5368 //===----------------------------------------------------------------------===//
5371 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5372 MachineBasicBlock *BB,
5373 unsigned Size) const {
5374 unsigned dest = MI->getOperand(0).getReg();
5375 unsigned ptr = MI->getOperand(1).getReg();
5376 unsigned oldval = MI->getOperand(2).getReg();
5377 unsigned newval = MI->getOperand(3).getReg();
5378 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5379 DebugLoc dl = MI->getDebugLoc();
5380 bool isThumb2 = Subtarget->isThumb2();
5382 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5383 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5384 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5385 (const TargetRegisterClass*)&ARM::GPRRegClass);
5388 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5389 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5390 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5393 unsigned ldrOpc, strOpc;
5395 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5397 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5398 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5401 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5402 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5405 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5406 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5410 MachineFunction *MF = BB->getParent();
5411 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5412 MachineFunction::iterator It = BB;
5413 ++It; // insert the new blocks after the current block
5415 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5416 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5417 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5418 MF->insert(It, loop1MBB);
5419 MF->insert(It, loop2MBB);
5420 MF->insert(It, exitMBB);
5422 // Transfer the remainder of BB and its successor edges to exitMBB.
5423 exitMBB->splice(exitMBB->begin(), BB,
5424 llvm::next(MachineBasicBlock::iterator(MI)),
5426 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5430 // fallthrough --> loop1MBB
5431 BB->addSuccessor(loop1MBB);
5434 // ldrex dest, [ptr]
5438 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5439 if (ldrOpc == ARM::t2LDREX)
5441 AddDefaultPred(MIB);
5442 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5443 .addReg(dest).addReg(oldval));
5444 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5445 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5446 BB->addSuccessor(loop2MBB);
5447 BB->addSuccessor(exitMBB);
5450 // strex scratch, newval, [ptr]
5454 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5455 if (strOpc == ARM::t2STREX)
5457 AddDefaultPred(MIB);
5458 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5459 .addReg(scratch).addImm(0));
5460 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5461 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5462 BB->addSuccessor(loop1MBB);
5463 BB->addSuccessor(exitMBB);
5469 MI->eraseFromParent(); // The instruction is gone now.
5475 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5476 unsigned Size, unsigned BinOpcode) const {
5477 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5478 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5480 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5481 MachineFunction *MF = BB->getParent();
5482 MachineFunction::iterator It = BB;
5485 unsigned dest = MI->getOperand(0).getReg();
5486 unsigned ptr = MI->getOperand(1).getReg();
5487 unsigned incr = MI->getOperand(2).getReg();
5488 DebugLoc dl = MI->getDebugLoc();
5489 bool isThumb2 = Subtarget->isThumb2();
5491 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5493 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5494 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5497 unsigned ldrOpc, strOpc;
5499 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5501 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5502 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5505 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5506 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5509 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5510 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5514 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5515 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5516 MF->insert(It, loopMBB);
5517 MF->insert(It, exitMBB);
5519 // Transfer the remainder of BB and its successor edges to exitMBB.
5520 exitMBB->splice(exitMBB->begin(), BB,
5521 llvm::next(MachineBasicBlock::iterator(MI)),
5523 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5525 const TargetRegisterClass *TRC = isThumb2 ?
5526 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5527 (const TargetRegisterClass*)&ARM::GPRRegClass;
5528 unsigned scratch = MRI.createVirtualRegister(TRC);
5529 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5533 // fallthrough --> loopMBB
5534 BB->addSuccessor(loopMBB);
5538 // <binop> scratch2, dest, incr
5539 // strex scratch, scratch2, ptr
5542 // fallthrough --> exitMBB
5544 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5545 if (ldrOpc == ARM::t2LDREX)
5547 AddDefaultPred(MIB);
5549 // operand order needs to go the other way for NAND
5550 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5551 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5552 addReg(incr).addReg(dest)).addReg(0);
5554 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5555 addReg(dest).addReg(incr)).addReg(0);
5558 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5559 if (strOpc == ARM::t2STREX)
5561 AddDefaultPred(MIB);
5562 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5563 .addReg(scratch).addImm(0));
5564 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5565 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5567 BB->addSuccessor(loopMBB);
5568 BB->addSuccessor(exitMBB);
5574 MI->eraseFromParent(); // The instruction is gone now.
5580 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5581 MachineBasicBlock *BB,
5584 ARMCC::CondCodes Cond) const {
5585 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5587 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5588 MachineFunction *MF = BB->getParent();
5589 MachineFunction::iterator It = BB;
5592 unsigned dest = MI->getOperand(0).getReg();
5593 unsigned ptr = MI->getOperand(1).getReg();
5594 unsigned incr = MI->getOperand(2).getReg();
5595 unsigned oldval = dest;
5596 DebugLoc dl = MI->getDebugLoc();
5597 bool isThumb2 = Subtarget->isThumb2();
5599 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5601 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5602 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5605 unsigned ldrOpc, strOpc, extendOpc;
5607 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5609 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5610 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5611 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5614 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5615 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5616 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5619 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5620 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5625 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5626 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5627 MF->insert(It, loopMBB);
5628 MF->insert(It, exitMBB);
5630 // Transfer the remainder of BB and its successor edges to exitMBB.
5631 exitMBB->splice(exitMBB->begin(), BB,
5632 llvm::next(MachineBasicBlock::iterator(MI)),
5634 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5636 const TargetRegisterClass *TRC = isThumb2 ?
5637 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5638 (const TargetRegisterClass*)&ARM::GPRRegClass;
5639 unsigned scratch = MRI.createVirtualRegister(TRC);
5640 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5644 // fallthrough --> loopMBB
5645 BB->addSuccessor(loopMBB);
5649 // (sign extend dest, if required)
5651 // cmov.cond scratch2, incr, dest
5652 // strex scratch, scratch2, ptr
5655 // fallthrough --> exitMBB
5657 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5658 if (ldrOpc == ARM::t2LDREX)
5660 AddDefaultPred(MIB);
5662 // Sign extend the value, if necessary.
5663 if (signExtend && extendOpc) {
5664 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5665 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5670 // Build compare and cmov instructions.
5671 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5672 .addReg(oldval).addReg(incr));
5673 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5674 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5676 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5677 if (strOpc == ARM::t2STREX)
5679 AddDefaultPred(MIB);
5680 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5681 .addReg(scratch).addImm(0));
5682 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5683 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5685 BB->addSuccessor(loopMBB);
5686 BB->addSuccessor(exitMBB);
5692 MI->eraseFromParent(); // The instruction is gone now.
5698 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5699 unsigned Op1, unsigned Op2,
5700 bool NeedsCarry, bool IsCmpxchg) const {
5701 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5702 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5704 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5705 MachineFunction *MF = BB->getParent();
5706 MachineFunction::iterator It = BB;
5709 unsigned destlo = MI->getOperand(0).getReg();
5710 unsigned desthi = MI->getOperand(1).getReg();
5711 unsigned ptr = MI->getOperand(2).getReg();
5712 unsigned vallo = MI->getOperand(3).getReg();
5713 unsigned valhi = MI->getOperand(4).getReg();
5714 DebugLoc dl = MI->getDebugLoc();
5715 bool isThumb2 = Subtarget->isThumb2();
5717 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5719 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5720 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5721 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5724 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5725 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5727 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5728 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5730 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5731 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5733 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5734 MF->insert(It, loopMBB);
5736 MF->insert(It, contBB);
5737 MF->insert(It, cont2BB);
5739 MF->insert(It, exitMBB);
5741 // Transfer the remainder of BB and its successor edges to exitMBB.
5742 exitMBB->splice(exitMBB->begin(), BB,
5743 llvm::next(MachineBasicBlock::iterator(MI)),
5745 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5747 const TargetRegisterClass *TRC = isThumb2 ?
5748 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5749 (const TargetRegisterClass*)&ARM::GPRRegClass;
5750 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5754 // fallthrough --> loopMBB
5755 BB->addSuccessor(loopMBB);
5758 // ldrexd r2, r3, ptr
5759 // <binopa> r0, r2, incr
5760 // <binopb> r1, r3, incr
5761 // strexd storesuccess, r0, r1, ptr
5762 // cmp storesuccess, #0
5764 // fallthrough --> exitMBB
5766 // Note that the registers are explicitly specified because there is not any
5767 // way to force the register allocator to allocate a register pair.
5769 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5770 // need to properly enforce the restriction that the two output registers
5771 // for ldrexd must be different.
5774 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5775 .addReg(ARM::R2, RegState::Define)
5776 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5777 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5778 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5779 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5783 for (unsigned i = 0; i < 2; i++) {
5784 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5786 .addReg(i == 0 ? destlo : desthi)
5787 .addReg(i == 0 ? vallo : valhi));
5788 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5789 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5790 BB->addSuccessor(exitMBB);
5791 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5792 BB = (i == 0 ? contBB : cont2BB);
5795 // Copy to physregs for strexd
5796 unsigned setlo = MI->getOperand(5).getReg();
5797 unsigned sethi = MI->getOperand(6).getReg();
5798 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5799 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5801 // Perform binary operation
5802 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5803 .addReg(destlo).addReg(vallo))
5804 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5805 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5806 .addReg(desthi).addReg(valhi)).addReg(0);
5808 // Copy to physregs for strexd
5809 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5810 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5814 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5815 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5817 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5818 .addReg(storesuccess).addImm(0));
5819 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5820 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5822 BB->addSuccessor(loopMBB);
5823 BB->addSuccessor(exitMBB);
5829 MI->eraseFromParent(); // The instruction is gone now.
5834 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5835 /// registers the function context.
5836 void ARMTargetLowering::
5837 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5838 MachineBasicBlock *DispatchBB, int FI) const {
5839 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5840 DebugLoc dl = MI->getDebugLoc();
5841 MachineFunction *MF = MBB->getParent();
5842 MachineRegisterInfo *MRI = &MF->getRegInfo();
5843 MachineConstantPool *MCP = MF->getConstantPool();
5844 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5845 const Function *F = MF->getFunction();
5847 bool isThumb = Subtarget->isThumb();
5848 bool isThumb2 = Subtarget->isThumb2();
5850 unsigned PCLabelId = AFI->createPICLabelUId();
5851 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5852 ARMConstantPoolValue *CPV =
5853 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5854 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5856 const TargetRegisterClass *TRC = isThumb ?
5857 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5858 (const TargetRegisterClass*)&ARM::GPRRegClass;
5860 // Grab constant pool and fixed stack memory operands.
5861 MachineMemOperand *CPMMO =
5862 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5863 MachineMemOperand::MOLoad, 4, 4);
5865 MachineMemOperand *FIMMOSt =
5866 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5867 MachineMemOperand::MOStore, 4, 4);
5869 // Load the address of the dispatch MBB into the jump buffer.
5871 // Incoming value: jbuf
5872 // ldr.n r5, LCPI1_1
5875 // str r5, [$jbuf, #+4] ; &jbuf[1]
5876 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5877 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5878 .addConstantPoolIndex(CPI)
5879 .addMemOperand(CPMMO));
5880 // Set the low bit because of thumb mode.
5881 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5883 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5884 .addReg(NewVReg1, RegState::Kill)
5886 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5887 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5888 .addReg(NewVReg2, RegState::Kill)
5890 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5891 .addReg(NewVReg3, RegState::Kill)
5893 .addImm(36) // &jbuf[1] :: pc
5894 .addMemOperand(FIMMOSt));
5895 } else if (isThumb) {
5896 // Incoming value: jbuf
5897 // ldr.n r1, LCPI1_4
5901 // add r2, $jbuf, #+4 ; &jbuf[1]
5903 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5904 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5905 .addConstantPoolIndex(CPI)
5906 .addMemOperand(CPMMO));
5907 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5908 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5909 .addReg(NewVReg1, RegState::Kill)
5911 // Set the low bit because of thumb mode.
5912 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5913 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5914 .addReg(ARM::CPSR, RegState::Define)
5916 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5917 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5918 .addReg(ARM::CPSR, RegState::Define)
5919 .addReg(NewVReg2, RegState::Kill)
5920 .addReg(NewVReg3, RegState::Kill));
5921 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5922 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5924 .addImm(36)); // &jbuf[1] :: pc
5925 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5926 .addReg(NewVReg4, RegState::Kill)
5927 .addReg(NewVReg5, RegState::Kill)
5929 .addMemOperand(FIMMOSt));
5931 // Incoming value: jbuf
5934 // str r1, [$jbuf, #+4] ; &jbuf[1]
5935 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5936 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5937 .addConstantPoolIndex(CPI)
5939 .addMemOperand(CPMMO));
5940 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5941 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5942 .addReg(NewVReg1, RegState::Kill)
5943 .addImm(PCLabelId));
5944 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5945 .addReg(NewVReg2, RegState::Kill)
5947 .addImm(36) // &jbuf[1] :: pc
5948 .addMemOperand(FIMMOSt));
5952 MachineBasicBlock *ARMTargetLowering::
5953 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5954 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5955 DebugLoc dl = MI->getDebugLoc();
5956 MachineFunction *MF = MBB->getParent();
5957 MachineRegisterInfo *MRI = &MF->getRegInfo();
5958 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5959 MachineFrameInfo *MFI = MF->getFrameInfo();
5960 int FI = MFI->getFunctionContextIndex();
5962 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5963 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5964 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
5966 // Get a mapping of the call site numbers to all of the landing pads they're
5968 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5969 unsigned MaxCSNum = 0;
5970 MachineModuleInfo &MMI = MF->getMMI();
5971 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5973 if (!BB->isLandingPad()) continue;
5975 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5977 for (MachineBasicBlock::iterator
5978 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5979 if (!II->isEHLabel()) continue;
5981 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5982 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5984 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5985 for (SmallVectorImpl<unsigned>::iterator
5986 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5987 CSI != CSE; ++CSI) {
5988 CallSiteNumToLPad[*CSI].push_back(BB);
5989 MaxCSNum = std::max(MaxCSNum, *CSI);
5995 // Get an ordered list of the machine basic blocks for the jump table.
5996 std::vector<MachineBasicBlock*> LPadList;
5997 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5998 LPadList.reserve(CallSiteNumToLPad.size());
5999 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6000 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6001 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6002 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6003 LPadList.push_back(*II);
6004 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6008 assert(!LPadList.empty() &&
6009 "No landing pad destinations for the dispatch jump table!");
6011 // Create the jump table and associated information.
6012 MachineJumpTableInfo *JTI =
6013 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6014 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6015 unsigned UId = AFI->createJumpTableUId();
6017 // Create the MBBs for the dispatch code.
6019 // Shove the dispatch's address into the return slot in the function context.
6020 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6021 DispatchBB->setIsLandingPad();
6023 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6024 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6025 DispatchBB->addSuccessor(TrapBB);
6027 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6028 DispatchBB->addSuccessor(DispContBB);
6031 MF->insert(MF->end(), DispatchBB);
6032 MF->insert(MF->end(), DispContBB);
6033 MF->insert(MF->end(), TrapBB);
6035 // Insert code into the entry block that creates and registers the function
6037 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6039 MachineMemOperand *FIMMOLd =
6040 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6041 MachineMemOperand::MOLoad |
6042 MachineMemOperand::MOVolatile, 4, 4);
6044 if (AFI->isThumb1OnlyFunction())
6045 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
6046 else if (!Subtarget->hasVFP2())
6047 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
6049 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6051 unsigned NumLPads = LPadList.size();
6052 if (Subtarget->isThumb2()) {
6053 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6054 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6057 .addMemOperand(FIMMOLd));
6059 if (NumLPads < 256) {
6060 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6062 .addImm(LPadList.size()));
6064 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6065 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6066 .addImm(NumLPads & 0xFFFF));
6068 unsigned VReg2 = VReg1;
6069 if ((NumLPads & 0xFFFF0000) != 0) {
6070 VReg2 = MRI->createVirtualRegister(TRC);
6071 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6073 .addImm(NumLPads >> 16));
6076 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6081 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6086 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6087 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6088 .addJumpTableIndex(MJTI)
6091 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6094 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6095 .addReg(NewVReg3, RegState::Kill)
6097 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6099 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6100 .addReg(NewVReg4, RegState::Kill)
6102 .addJumpTableIndex(MJTI)
6104 } else if (Subtarget->isThumb()) {
6105 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6106 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6109 .addMemOperand(FIMMOLd));
6111 if (NumLPads < 256) {
6112 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6116 MachineConstantPool *ConstantPool = MF->getConstantPool();
6117 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6118 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6120 // MachineConstantPool wants an explicit alignment.
6121 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6123 Align = getDataLayout()->getTypeAllocSize(C->getType());
6124 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6126 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6127 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6128 .addReg(VReg1, RegState::Define)
6129 .addConstantPoolIndex(Idx));
6130 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6135 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6140 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6141 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6142 .addReg(ARM::CPSR, RegState::Define)
6146 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6147 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6148 .addJumpTableIndex(MJTI)
6151 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6152 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6153 .addReg(ARM::CPSR, RegState::Define)
6154 .addReg(NewVReg2, RegState::Kill)
6157 MachineMemOperand *JTMMOLd =
6158 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6159 MachineMemOperand::MOLoad, 4, 4);
6161 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6162 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6163 .addReg(NewVReg4, RegState::Kill)
6165 .addMemOperand(JTMMOLd));
6167 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6168 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6169 .addReg(ARM::CPSR, RegState::Define)
6170 .addReg(NewVReg5, RegState::Kill)
6173 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6174 .addReg(NewVReg6, RegState::Kill)
6175 .addJumpTableIndex(MJTI)
6178 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6179 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6182 .addMemOperand(FIMMOLd));
6184 if (NumLPads < 256) {
6185 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6188 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6189 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6190 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6191 .addImm(NumLPads & 0xFFFF));
6193 unsigned VReg2 = VReg1;
6194 if ((NumLPads & 0xFFFF0000) != 0) {
6195 VReg2 = MRI->createVirtualRegister(TRC);
6196 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6198 .addImm(NumLPads >> 16));
6201 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6205 MachineConstantPool *ConstantPool = MF->getConstantPool();
6206 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6207 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6209 // MachineConstantPool wants an explicit alignment.
6210 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6212 Align = getDataLayout()->getTypeAllocSize(C->getType());
6213 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6215 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6216 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6217 .addReg(VReg1, RegState::Define)
6218 .addConstantPoolIndex(Idx)
6220 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6222 .addReg(VReg1, RegState::Kill));
6225 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6230 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6232 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6234 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6235 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6236 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6237 .addJumpTableIndex(MJTI)
6240 MachineMemOperand *JTMMOLd =
6241 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6242 MachineMemOperand::MOLoad, 4, 4);
6243 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6245 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6246 .addReg(NewVReg3, RegState::Kill)
6249 .addMemOperand(JTMMOLd));
6251 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6252 .addReg(NewVReg5, RegState::Kill)
6254 .addJumpTableIndex(MJTI)
6258 // Add the jump table entries as successors to the MBB.
6259 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6260 for (std::vector<MachineBasicBlock*>::iterator
6261 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6262 MachineBasicBlock *CurMBB = *I;
6263 if (SeenMBBs.insert(CurMBB))
6264 DispContBB->addSuccessor(CurMBB);
6267 // N.B. the order the invoke BBs are processed in doesn't matter here.
6268 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6269 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6270 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6271 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6272 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6273 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6274 MachineBasicBlock *BB = *I;
6276 // Remove the landing pad successor from the invoke block and replace it
6277 // with the new dispatch block.
6278 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6280 while (!Successors.empty()) {
6281 MachineBasicBlock *SMBB = Successors.pop_back_val();
6282 if (SMBB->isLandingPad()) {
6283 BB->removeSuccessor(SMBB);
6284 MBBLPads.push_back(SMBB);
6288 BB->addSuccessor(DispatchBB);
6290 // Find the invoke call and mark all of the callee-saved registers as
6291 // 'implicit defined' so that they're spilled. This prevents code from
6292 // moving instructions to before the EH block, where they will never be
6294 for (MachineBasicBlock::reverse_iterator
6295 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6296 if (!II->isCall()) continue;
6298 DenseMap<unsigned, bool> DefRegs;
6299 for (MachineInstr::mop_iterator
6300 OI = II->operands_begin(), OE = II->operands_end();
6302 if (!OI->isReg()) continue;
6303 DefRegs[OI->getReg()] = true;
6306 MachineInstrBuilder MIB(&*II);
6308 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6309 unsigned Reg = SavedRegs[i];
6310 if (Subtarget->isThumb2() &&
6311 !ARM::tGPRRegClass.contains(Reg) &&
6312 !ARM::hGPRRegClass.contains(Reg))
6314 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6316 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6319 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6326 // Mark all former landing pads as non-landing pads. The dispatch is the only
6328 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6329 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6330 (*I)->setIsLandingPad(false);
6332 // The instruction is gone now.
6333 MI->eraseFromParent();
6339 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6340 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6341 E = MBB->succ_end(); I != E; ++I)
6344 llvm_unreachable("Expecting a BB with two successors!");
6347 MachineBasicBlock *ARMTargetLowering::
6348 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6349 // This pseudo instruction has 3 operands: dst, src, size
6350 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6351 // Otherwise, we will generate unrolled scalar copies.
6352 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6353 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6354 MachineFunction::iterator It = BB;
6357 unsigned dest = MI->getOperand(0).getReg();
6358 unsigned src = MI->getOperand(1).getReg();
6359 unsigned SizeVal = MI->getOperand(2).getImm();
6360 unsigned Align = MI->getOperand(3).getImm();
6361 DebugLoc dl = MI->getDebugLoc();
6363 bool isThumb2 = Subtarget->isThumb2();
6364 MachineFunction *MF = BB->getParent();
6365 MachineRegisterInfo &MRI = MF->getRegInfo();
6366 unsigned ldrOpc, strOpc, UnitSize = 0;
6368 const TargetRegisterClass *TRC = isThumb2 ?
6369 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6370 (const TargetRegisterClass*)&ARM::GPRRegClass;
6371 const TargetRegisterClass *TRC_Vec = 0;
6374 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6375 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6377 } else if (Align & 2) {
6378 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6379 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6382 // Check whether we can use NEON instructions.
6383 if (!MF->getFunction()->getFnAttributes().
6384 hasAttribute(Attributes::NoImplicitFloat) &&
6385 Subtarget->hasNEON()) {
6386 if ((Align % 16 == 0) && SizeVal >= 16) {
6387 ldrOpc = ARM::VLD1q32wb_fixed;
6388 strOpc = ARM::VST1q32wb_fixed;
6390 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6392 else if ((Align % 8 == 0) && SizeVal >= 8) {
6393 ldrOpc = ARM::VLD1d32wb_fixed;
6394 strOpc = ARM::VST1d32wb_fixed;
6396 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6399 // Can't use NEON instructions.
6400 if (UnitSize == 0) {
6401 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6402 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6407 unsigned BytesLeft = SizeVal % UnitSize;
6408 unsigned LoopSize = SizeVal - BytesLeft;
6410 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6411 // Use LDR and STR to copy.
6412 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6413 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6414 unsigned srcIn = src;
6415 unsigned destIn = dest;
6416 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6417 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6418 unsigned srcOut = MRI.createVirtualRegister(TRC);
6419 unsigned destOut = MRI.createVirtualRegister(TRC);
6420 if (UnitSize >= 8) {
6421 AddDefaultPred(BuildMI(*BB, MI, dl,
6422 TII->get(ldrOpc), scratch)
6423 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6425 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6426 .addReg(destIn).addImm(0).addReg(scratch));
6427 } else if (isThumb2) {
6428 AddDefaultPred(BuildMI(*BB, MI, dl,
6429 TII->get(ldrOpc), scratch)
6430 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6432 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6433 .addReg(scratch).addReg(destIn)
6436 AddDefaultPred(BuildMI(*BB, MI, dl,
6437 TII->get(ldrOpc), scratch)
6438 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6441 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6442 .addReg(scratch).addReg(destIn)
6443 .addReg(0).addImm(UnitSize));
6449 // Handle the leftover bytes with LDRB and STRB.
6450 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6451 // [destOut] = STRB_POST(scratch, destIn, 1)
6452 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6453 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6454 for (unsigned i = 0; i < BytesLeft; i++) {
6455 unsigned scratch = MRI.createVirtualRegister(TRC);
6456 unsigned srcOut = MRI.createVirtualRegister(TRC);
6457 unsigned destOut = MRI.createVirtualRegister(TRC);
6459 AddDefaultPred(BuildMI(*BB, MI, dl,
6460 TII->get(ldrOpc),scratch)
6461 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6463 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6464 .addReg(scratch).addReg(destIn)
6465 .addReg(0).addImm(1));
6467 AddDefaultPred(BuildMI(*BB, MI, dl,
6468 TII->get(ldrOpc),scratch)
6469 .addReg(srcOut, RegState::Define).addReg(srcIn)
6470 .addReg(0).addImm(1));
6472 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6473 .addReg(scratch).addReg(destIn)
6474 .addReg(0).addImm(1));
6479 MI->eraseFromParent(); // The instruction is gone now.
6483 // Expand the pseudo op to a loop.
6486 // movw varEnd, # --> with thumb2
6488 // ldrcp varEnd, idx --> without thumb2
6489 // fallthrough --> loopMBB
6491 // PHI varPhi, varEnd, varLoop
6492 // PHI srcPhi, src, srcLoop
6493 // PHI destPhi, dst, destLoop
6494 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6495 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6496 // subs varLoop, varPhi, #UnitSize
6498 // fallthrough --> exitMBB
6500 // epilogue to handle left-over bytes
6501 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6502 // [destOut] = STRB_POST(scratch, destLoop, 1)
6503 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6504 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6505 MF->insert(It, loopMBB);
6506 MF->insert(It, exitMBB);
6508 // Transfer the remainder of BB and its successor edges to exitMBB.
6509 exitMBB->splice(exitMBB->begin(), BB,
6510 llvm::next(MachineBasicBlock::iterator(MI)),
6512 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6514 // Load an immediate to varEnd.
6515 unsigned varEnd = MRI.createVirtualRegister(TRC);
6517 unsigned VReg1 = varEnd;
6518 if ((LoopSize & 0xFFFF0000) != 0)
6519 VReg1 = MRI.createVirtualRegister(TRC);
6520 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6521 .addImm(LoopSize & 0xFFFF));
6523 if ((LoopSize & 0xFFFF0000) != 0)
6524 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6526 .addImm(LoopSize >> 16));
6528 MachineConstantPool *ConstantPool = MF->getConstantPool();
6529 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6530 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6532 // MachineConstantPool wants an explicit alignment.
6533 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6535 Align = getDataLayout()->getTypeAllocSize(C->getType());
6536 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6538 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6539 .addReg(varEnd, RegState::Define)
6540 .addConstantPoolIndex(Idx)
6543 BB->addSuccessor(loopMBB);
6545 // Generate the loop body:
6546 // varPhi = PHI(varLoop, varEnd)
6547 // srcPhi = PHI(srcLoop, src)
6548 // destPhi = PHI(destLoop, dst)
6549 MachineBasicBlock *entryBB = BB;
6551 unsigned varLoop = MRI.createVirtualRegister(TRC);
6552 unsigned varPhi = MRI.createVirtualRegister(TRC);
6553 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6554 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6555 unsigned destLoop = MRI.createVirtualRegister(TRC);
6556 unsigned destPhi = MRI.createVirtualRegister(TRC);
6558 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6559 .addReg(varLoop).addMBB(loopMBB)
6560 .addReg(varEnd).addMBB(entryBB);
6561 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6562 .addReg(srcLoop).addMBB(loopMBB)
6563 .addReg(src).addMBB(entryBB);
6564 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6565 .addReg(destLoop).addMBB(loopMBB)
6566 .addReg(dest).addMBB(entryBB);
6568 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6569 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6570 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6571 if (UnitSize >= 8) {
6572 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6573 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6575 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6576 .addReg(destPhi).addImm(0).addReg(scratch));
6577 } else if (isThumb2) {
6578 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6579 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6581 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6582 .addReg(scratch).addReg(destPhi)
6585 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6586 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6589 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6590 .addReg(scratch).addReg(destPhi)
6591 .addReg(0).addImm(UnitSize));
6594 // Decrement loop variable by UnitSize.
6595 MachineInstrBuilder MIB = BuildMI(BB, dl,
6596 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6597 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6598 MIB->getOperand(5).setReg(ARM::CPSR);
6599 MIB->getOperand(5).setIsDef(true);
6601 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6602 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6604 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6605 BB->addSuccessor(loopMBB);
6606 BB->addSuccessor(exitMBB);
6608 // Add epilogue to handle BytesLeft.
6610 MachineInstr *StartOfExit = exitMBB->begin();
6611 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6612 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6614 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6615 // [destOut] = STRB_POST(scratch, destLoop, 1)
6616 unsigned srcIn = srcLoop;
6617 unsigned destIn = destLoop;
6618 for (unsigned i = 0; i < BytesLeft; i++) {
6619 unsigned scratch = MRI.createVirtualRegister(TRC);
6620 unsigned srcOut = MRI.createVirtualRegister(TRC);
6621 unsigned destOut = MRI.createVirtualRegister(TRC);
6623 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6624 TII->get(ldrOpc),scratch)
6625 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6627 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6628 .addReg(scratch).addReg(destIn)
6631 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6632 TII->get(ldrOpc),scratch)
6633 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6635 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6636 .addReg(scratch).addReg(destIn)
6637 .addReg(0).addImm(1));
6643 MI->eraseFromParent(); // The instruction is gone now.
6648 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6649 MachineBasicBlock *BB) const {
6650 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6651 DebugLoc dl = MI->getDebugLoc();
6652 bool isThumb2 = Subtarget->isThumb2();
6653 switch (MI->getOpcode()) {
6656 llvm_unreachable("Unexpected instr type to insert");
6658 // The Thumb2 pre-indexed stores have the same MI operands, they just
6659 // define them differently in the .td files from the isel patterns, so
6660 // they need pseudos.
6661 case ARM::t2STR_preidx:
6662 MI->setDesc(TII->get(ARM::t2STR_PRE));
6664 case ARM::t2STRB_preidx:
6665 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6667 case ARM::t2STRH_preidx:
6668 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6671 case ARM::STRi_preidx:
6672 case ARM::STRBi_preidx: {
6673 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6674 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6675 // Decode the offset.
6676 unsigned Offset = MI->getOperand(4).getImm();
6677 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6678 Offset = ARM_AM::getAM2Offset(Offset);
6682 MachineMemOperand *MMO = *MI->memoperands_begin();
6683 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6684 .addOperand(MI->getOperand(0)) // Rn_wb
6685 .addOperand(MI->getOperand(1)) // Rt
6686 .addOperand(MI->getOperand(2)) // Rn
6687 .addImm(Offset) // offset (skip GPR==zero_reg)
6688 .addOperand(MI->getOperand(5)) // pred
6689 .addOperand(MI->getOperand(6))
6690 .addMemOperand(MMO);
6691 MI->eraseFromParent();
6694 case ARM::STRr_preidx:
6695 case ARM::STRBr_preidx:
6696 case ARM::STRH_preidx: {
6698 switch (MI->getOpcode()) {
6699 default: llvm_unreachable("unexpected opcode!");
6700 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6701 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6702 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6704 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6705 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6706 MIB.addOperand(MI->getOperand(i));
6707 MI->eraseFromParent();
6710 case ARM::ATOMIC_LOAD_ADD_I8:
6711 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6712 case ARM::ATOMIC_LOAD_ADD_I16:
6713 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6714 case ARM::ATOMIC_LOAD_ADD_I32:
6715 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6717 case ARM::ATOMIC_LOAD_AND_I8:
6718 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6719 case ARM::ATOMIC_LOAD_AND_I16:
6720 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6721 case ARM::ATOMIC_LOAD_AND_I32:
6722 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6724 case ARM::ATOMIC_LOAD_OR_I8:
6725 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6726 case ARM::ATOMIC_LOAD_OR_I16:
6727 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6728 case ARM::ATOMIC_LOAD_OR_I32:
6729 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6731 case ARM::ATOMIC_LOAD_XOR_I8:
6732 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6733 case ARM::ATOMIC_LOAD_XOR_I16:
6734 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6735 case ARM::ATOMIC_LOAD_XOR_I32:
6736 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6738 case ARM::ATOMIC_LOAD_NAND_I8:
6739 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6740 case ARM::ATOMIC_LOAD_NAND_I16:
6741 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6742 case ARM::ATOMIC_LOAD_NAND_I32:
6743 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6745 case ARM::ATOMIC_LOAD_SUB_I8:
6746 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6747 case ARM::ATOMIC_LOAD_SUB_I16:
6748 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6749 case ARM::ATOMIC_LOAD_SUB_I32:
6750 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6752 case ARM::ATOMIC_LOAD_MIN_I8:
6753 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6754 case ARM::ATOMIC_LOAD_MIN_I16:
6755 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6756 case ARM::ATOMIC_LOAD_MIN_I32:
6757 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6759 case ARM::ATOMIC_LOAD_MAX_I8:
6760 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6761 case ARM::ATOMIC_LOAD_MAX_I16:
6762 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6763 case ARM::ATOMIC_LOAD_MAX_I32:
6764 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6766 case ARM::ATOMIC_LOAD_UMIN_I8:
6767 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6768 case ARM::ATOMIC_LOAD_UMIN_I16:
6769 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6770 case ARM::ATOMIC_LOAD_UMIN_I32:
6771 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6773 case ARM::ATOMIC_LOAD_UMAX_I8:
6774 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6775 case ARM::ATOMIC_LOAD_UMAX_I16:
6776 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6777 case ARM::ATOMIC_LOAD_UMAX_I32:
6778 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6780 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6781 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6782 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6784 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6785 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6786 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6789 case ARM::ATOMADD6432:
6790 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6791 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6792 /*NeedsCarry*/ true);
6793 case ARM::ATOMSUB6432:
6794 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6795 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6796 /*NeedsCarry*/ true);
6797 case ARM::ATOMOR6432:
6798 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6799 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6800 case ARM::ATOMXOR6432:
6801 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6802 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6803 case ARM::ATOMAND6432:
6804 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6805 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6806 case ARM::ATOMSWAP6432:
6807 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6808 case ARM::ATOMCMPXCHG6432:
6809 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6810 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6811 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6813 case ARM::tMOVCCr_pseudo: {
6814 // To "insert" a SELECT_CC instruction, we actually have to insert the
6815 // diamond control-flow pattern. The incoming instruction knows the
6816 // destination vreg to set, the condition code register to branch on, the
6817 // true/false values to select between, and a branch opcode to use.
6818 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6819 MachineFunction::iterator It = BB;
6825 // cmpTY ccX, r1, r2
6827 // fallthrough --> copy0MBB
6828 MachineBasicBlock *thisMBB = BB;
6829 MachineFunction *F = BB->getParent();
6830 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6831 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6832 F->insert(It, copy0MBB);
6833 F->insert(It, sinkMBB);
6835 // Transfer the remainder of BB and its successor edges to sinkMBB.
6836 sinkMBB->splice(sinkMBB->begin(), BB,
6837 llvm::next(MachineBasicBlock::iterator(MI)),
6839 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6841 BB->addSuccessor(copy0MBB);
6842 BB->addSuccessor(sinkMBB);
6844 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6845 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6848 // %FalseValue = ...
6849 // # fallthrough to sinkMBB
6852 // Update machine-CFG edges
6853 BB->addSuccessor(sinkMBB);
6856 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6859 BuildMI(*BB, BB->begin(), dl,
6860 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6861 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6862 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6864 MI->eraseFromParent(); // The pseudo instruction is gone now.
6869 case ARM::BCCZi64: {
6870 // If there is an unconditional branch to the other successor, remove it.
6871 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6873 // Compare both parts that make up the double comparison separately for
6875 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6877 unsigned LHS1 = MI->getOperand(1).getReg();
6878 unsigned LHS2 = MI->getOperand(2).getReg();
6880 AddDefaultPred(BuildMI(BB, dl,
6881 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6882 .addReg(LHS1).addImm(0));
6883 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6884 .addReg(LHS2).addImm(0)
6885 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6887 unsigned RHS1 = MI->getOperand(3).getReg();
6888 unsigned RHS2 = MI->getOperand(4).getReg();
6889 AddDefaultPred(BuildMI(BB, dl,
6890 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6891 .addReg(LHS1).addReg(RHS1));
6892 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6893 .addReg(LHS2).addReg(RHS2)
6894 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6897 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6898 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6899 if (MI->getOperand(0).getImm() == ARMCC::NE)
6900 std::swap(destMBB, exitMBB);
6902 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6903 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6905 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6907 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6909 MI->eraseFromParent(); // The pseudo instruction is gone now.
6913 case ARM::Int_eh_sjlj_setjmp:
6914 case ARM::Int_eh_sjlj_setjmp_nofp:
6915 case ARM::tInt_eh_sjlj_setjmp:
6916 case ARM::t2Int_eh_sjlj_setjmp:
6917 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6918 EmitSjLjDispatchBlock(MI, BB);
6923 // To insert an ABS instruction, we have to insert the
6924 // diamond control-flow pattern. The incoming instruction knows the
6925 // source vreg to test against 0, the destination vreg to set,
6926 // the condition code register to branch on, the
6927 // true/false values to select between, and a branch opcode to use.
6932 // BCC (branch to SinkBB if V0 >= 0)
6933 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6934 // SinkBB: V1 = PHI(V2, V3)
6935 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6936 MachineFunction::iterator BBI = BB;
6938 MachineFunction *Fn = BB->getParent();
6939 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6940 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6941 Fn->insert(BBI, RSBBB);
6942 Fn->insert(BBI, SinkBB);
6944 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6945 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6946 bool isThumb2 = Subtarget->isThumb2();
6947 MachineRegisterInfo &MRI = Fn->getRegInfo();
6948 // In Thumb mode S must not be specified if source register is the SP or
6949 // PC and if destination register is the SP, so restrict register class
6950 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6951 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6952 (const TargetRegisterClass*)&ARM::GPRRegClass);
6954 // Transfer the remainder of BB and its successor edges to sinkMBB.
6955 SinkBB->splice(SinkBB->begin(), BB,
6956 llvm::next(MachineBasicBlock::iterator(MI)),
6958 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6960 BB->addSuccessor(RSBBB);
6961 BB->addSuccessor(SinkBB);
6963 // fall through to SinkMBB
6964 RSBBB->addSuccessor(SinkBB);
6966 // insert a cmp at the end of BB
6967 AddDefaultPred(BuildMI(BB, dl,
6968 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6969 .addReg(ABSSrcReg).addImm(0));
6971 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6973 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6974 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6976 // insert rsbri in RSBBB
6977 // Note: BCC and rsbri will be converted into predicated rsbmi
6978 // by if-conversion pass
6979 BuildMI(*RSBBB, RSBBB->begin(), dl,
6980 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6981 .addReg(ABSSrcReg, RegState::Kill)
6982 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6984 // insert PHI in SinkBB,
6985 // reuse ABSDstReg to not change uses of ABS instruction
6986 BuildMI(*SinkBB, SinkBB->begin(), dl,
6987 TII->get(ARM::PHI), ABSDstReg)
6988 .addReg(NewRsbDstReg).addMBB(RSBBB)
6989 .addReg(ABSSrcReg).addMBB(BB);
6991 // remove ABS instruction
6992 MI->eraseFromParent();
6994 // return last added BB
6997 case ARM::COPY_STRUCT_BYVAL_I32:
6999 return EmitStructByval(MI, BB);
7003 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7004 SDNode *Node) const {
7005 if (!MI->hasPostISelHook()) {
7006 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7007 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7011 const MCInstrDesc *MCID = &MI->getDesc();
7012 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7013 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7014 // operand is still set to noreg. If needed, set the optional operand's
7015 // register to CPSR, and remove the redundant implicit def.
7017 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7019 // Rename pseudo opcodes.
7020 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7022 const ARMBaseInstrInfo *TII =
7023 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7024 MCID = &TII->get(NewOpc);
7026 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7027 "converted opcode should be the same except for cc_out");
7031 // Add the optional cc_out operand
7032 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7034 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7036 // Any ARM instruction that sets the 's' bit should specify an optional
7037 // "cc_out" operand in the last operand position.
7038 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7039 assert(!NewOpc && "Optional cc_out operand required");
7042 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7043 // since we already have an optional CPSR def.
7044 bool definesCPSR = false;
7045 bool deadCPSR = false;
7046 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7048 const MachineOperand &MO = MI->getOperand(i);
7049 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7053 MI->RemoveOperand(i);
7058 assert(!NewOpc && "Optional cc_out operand required");
7061 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7063 assert(!MI->getOperand(ccOutIdx).getReg() &&
7064 "expect uninitialized optional cc_out operand");
7068 // If this instruction was defined with an optional CPSR def and its dag node
7069 // had a live implicit CPSR def, then activate the optional CPSR def.
7070 MachineOperand &MO = MI->getOperand(ccOutIdx);
7071 MO.setReg(ARM::CPSR);
7075 //===----------------------------------------------------------------------===//
7076 // ARM Optimization Hooks
7077 //===----------------------------------------------------------------------===//
7079 // Helper function that checks if N is a null or all ones constant.
7080 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7081 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7084 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7087 // Return true if N is conditionally 0 or all ones.
7088 // Detects these expressions where cc is an i1 value:
7090 // (select cc 0, y) [AllOnes=0]
7091 // (select cc y, 0) [AllOnes=0]
7092 // (zext cc) [AllOnes=0]
7093 // (sext cc) [AllOnes=0/1]
7094 // (select cc -1, y) [AllOnes=1]
7095 // (select cc y, -1) [AllOnes=1]
7097 // Invert is set when N is the null/all ones constant when CC is false.
7098 // OtherOp is set to the alternative value of N.
7099 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7100 SDValue &CC, bool &Invert,
7102 SelectionDAG &DAG) {
7103 switch (N->getOpcode()) {
7104 default: return false;
7106 CC = N->getOperand(0);
7107 SDValue N1 = N->getOperand(1);
7108 SDValue N2 = N->getOperand(2);
7109 if (isZeroOrAllOnes(N1, AllOnes)) {
7114 if (isZeroOrAllOnes(N2, AllOnes)) {
7121 case ISD::ZERO_EXTEND:
7122 // (zext cc) can never be the all ones value.
7126 case ISD::SIGN_EXTEND: {
7127 EVT VT = N->getValueType(0);
7128 CC = N->getOperand(0);
7129 if (CC.getValueType() != MVT::i1)
7133 // When looking for an AllOnes constant, N is an sext, and the 'other'
7135 OtherOp = DAG.getConstant(0, VT);
7136 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7137 // When looking for a 0 constant, N can be zext or sext.
7138 OtherOp = DAG.getConstant(1, VT);
7140 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7146 // Combine a constant select operand into its use:
7148 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7149 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7150 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7151 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7152 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7154 // The transform is rejected if the select doesn't have a constant operand that
7155 // is null, or all ones when AllOnes is set.
7157 // Also recognize sext/zext from i1:
7159 // (add (zext cc), x) -> (select cc (add x, 1), x)
7160 // (add (sext cc), x) -> (select cc (add x, -1), x)
7162 // These transformations eventually create predicated instructions.
7164 // @param N The node to transform.
7165 // @param Slct The N operand that is a select.
7166 // @param OtherOp The other N operand (x above).
7167 // @param DCI Context.
7168 // @param AllOnes Require the select constant to be all ones instead of null.
7169 // @returns The new node, or SDValue() on failure.
7171 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7172 TargetLowering::DAGCombinerInfo &DCI,
7173 bool AllOnes = false) {
7174 SelectionDAG &DAG = DCI.DAG;
7175 EVT VT = N->getValueType(0);
7176 SDValue NonConstantVal;
7179 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7180 NonConstantVal, DAG))
7183 // Slct is now know to be the desired identity constant when CC is true.
7184 SDValue TrueVal = OtherOp;
7185 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7186 OtherOp, NonConstantVal);
7187 // Unless SwapSelectOps says CC should be false.
7189 std::swap(TrueVal, FalseVal);
7191 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7192 CCOp, TrueVal, FalseVal);
7195 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7197 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7198 TargetLowering::DAGCombinerInfo &DCI) {
7199 SDValue N0 = N->getOperand(0);
7200 SDValue N1 = N->getOperand(1);
7201 if (N0.getNode()->hasOneUse()) {
7202 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7203 if (Result.getNode())
7206 if (N1.getNode()->hasOneUse()) {
7207 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7208 if (Result.getNode())
7214 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7215 // (only after legalization).
7216 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7217 TargetLowering::DAGCombinerInfo &DCI,
7218 const ARMSubtarget *Subtarget) {
7220 // Only perform optimization if after legalize, and if NEON is available. We
7221 // also expected both operands to be BUILD_VECTORs.
7222 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7223 || N0.getOpcode() != ISD::BUILD_VECTOR
7224 || N1.getOpcode() != ISD::BUILD_VECTOR)
7227 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7228 EVT VT = N->getValueType(0);
7229 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7232 // Check that the vector operands are of the right form.
7233 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7234 // operands, where N is the size of the formed vector.
7235 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7236 // index such that we have a pair wise add pattern.
7238 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7239 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7241 SDValue Vec = N0->getOperand(0)->getOperand(0);
7242 SDNode *V = Vec.getNode();
7243 unsigned nextIndex = 0;
7245 // For each operands to the ADD which are BUILD_VECTORs,
7246 // check to see if each of their operands are an EXTRACT_VECTOR with
7247 // the same vector and appropriate index.
7248 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7249 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7250 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7252 SDValue ExtVec0 = N0->getOperand(i);
7253 SDValue ExtVec1 = N1->getOperand(i);
7255 // First operand is the vector, verify its the same.
7256 if (V != ExtVec0->getOperand(0).getNode() ||
7257 V != ExtVec1->getOperand(0).getNode())
7260 // Second is the constant, verify its correct.
7261 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7262 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7264 // For the constant, we want to see all the even or all the odd.
7265 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7266 || C1->getZExtValue() != nextIndex+1)
7275 // Create VPADDL node.
7276 SelectionDAG &DAG = DCI.DAG;
7277 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7279 // Build operand list.
7280 SmallVector<SDValue, 8> Ops;
7281 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7282 TLI.getPointerTy()));
7284 // Input is the vector.
7287 // Get widened type and narrowed type.
7289 unsigned numElem = VT.getVectorNumElements();
7290 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7291 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7292 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7293 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7295 llvm_unreachable("Invalid vector element type for padd optimization.");
7298 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7299 widenType, &Ops[0], Ops.size());
7300 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7303 static SDValue findMUL_LOHI(SDValue V) {
7304 if (V->getOpcode() == ISD::UMUL_LOHI ||
7305 V->getOpcode() == ISD::SMUL_LOHI)
7310 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7311 TargetLowering::DAGCombinerInfo &DCI,
7312 const ARMSubtarget *Subtarget) {
7314 if (Subtarget->isThumb1Only()) return SDValue();
7316 // Only perform the checks after legalize when the pattern is available.
7317 if (DCI.isBeforeLegalize()) return SDValue();
7319 // Look for multiply add opportunities.
7320 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7321 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7322 // a glue link from the first add to the second add.
7323 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7324 // a S/UMLAL instruction.
7327 // \ / \ [no multiline comment]
7333 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7334 SDValue AddcOp0 = AddcNode->getOperand(0);
7335 SDValue AddcOp1 = AddcNode->getOperand(1);
7337 // Check if the two operands are from the same mul_lohi node.
7338 if (AddcOp0.getNode() == AddcOp1.getNode())
7341 assert(AddcNode->getNumValues() == 2 &&
7342 AddcNode->getValueType(0) == MVT::i32 &&
7343 AddcNode->getValueType(1) == MVT::Glue &&
7344 "Expect ADDC with two result values: i32, glue");
7346 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7347 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7348 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7349 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7350 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7353 // Look for the glued ADDE.
7354 SDNode* AddeNode = AddcNode->getGluedUser();
7355 if (AddeNode == NULL)
7358 // Make sure it is really an ADDE.
7359 if (AddeNode->getOpcode() != ISD::ADDE)
7362 assert(AddeNode->getNumOperands() == 3 &&
7363 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7364 "ADDE node has the wrong inputs");
7366 // Check for the triangle shape.
7367 SDValue AddeOp0 = AddeNode->getOperand(0);
7368 SDValue AddeOp1 = AddeNode->getOperand(1);
7370 // Make sure that the ADDE operands are not coming from the same node.
7371 if (AddeOp0.getNode() == AddeOp1.getNode())
7374 // Find the MUL_LOHI node walking up ADDE's operands.
7375 bool IsLeftOperandMUL = false;
7376 SDValue MULOp = findMUL_LOHI(AddeOp0);
7377 if (MULOp == SDValue())
7378 MULOp = findMUL_LOHI(AddeOp1);
7380 IsLeftOperandMUL = true;
7381 if (MULOp == SDValue())
7384 // Figure out the right opcode.
7385 unsigned Opc = MULOp->getOpcode();
7386 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7388 // Figure out the high and low input values to the MLAL node.
7389 SDValue* HiMul = &MULOp;
7390 SDValue* HiAdd = NULL;
7391 SDValue* LoMul = NULL;
7392 SDValue* LowAdd = NULL;
7394 if (IsLeftOperandMUL)
7400 if (AddcOp0->getOpcode() == Opc) {
7404 if (AddcOp1->getOpcode() == Opc) {
7412 if (LoMul->getNode() != HiMul->getNode())
7415 // Create the merged node.
7416 SelectionDAG &DAG = DCI.DAG;
7418 // Build operand list.
7419 SmallVector<SDValue, 8> Ops;
7420 Ops.push_back(LoMul->getOperand(0));
7421 Ops.push_back(LoMul->getOperand(1));
7422 Ops.push_back(*LowAdd);
7423 Ops.push_back(*HiAdd);
7425 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7426 DAG.getVTList(MVT::i32, MVT::i32),
7427 &Ops[0], Ops.size());
7429 // Replace the ADDs' nodes uses by the MLA node's values.
7430 SDValue HiMLALResult(MLALNode.getNode(), 1);
7431 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7433 SDValue LoMLALResult(MLALNode.getNode(), 0);
7434 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7436 // Return original node to notify the driver to stop replacing.
7437 SDValue resNode(AddcNode, 0);
7441 /// PerformADDCCombine - Target-specific dag combine transform from
7442 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7443 static SDValue PerformADDCCombine(SDNode *N,
7444 TargetLowering::DAGCombinerInfo &DCI,
7445 const ARMSubtarget *Subtarget) {
7447 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7451 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7452 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7453 /// called with the default operands, and if that fails, with commuted
7455 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7456 TargetLowering::DAGCombinerInfo &DCI,
7457 const ARMSubtarget *Subtarget){
7459 // Attempt to create vpaddl for this add.
7460 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7461 if (Result.getNode())
7464 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7465 if (N0.getNode()->hasOneUse()) {
7466 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7467 if (Result.getNode()) return Result;
7472 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7474 static SDValue PerformADDCombine(SDNode *N,
7475 TargetLowering::DAGCombinerInfo &DCI,
7476 const ARMSubtarget *Subtarget) {
7477 SDValue N0 = N->getOperand(0);
7478 SDValue N1 = N->getOperand(1);
7480 // First try with the default operand order.
7481 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7482 if (Result.getNode())
7485 // If that didn't work, try again with the operands commuted.
7486 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7489 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7491 static SDValue PerformSUBCombine(SDNode *N,
7492 TargetLowering::DAGCombinerInfo &DCI) {
7493 SDValue N0 = N->getOperand(0);
7494 SDValue N1 = N->getOperand(1);
7496 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7497 if (N1.getNode()->hasOneUse()) {
7498 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7499 if (Result.getNode()) return Result;
7505 /// PerformVMULCombine
7506 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7507 /// special multiplier accumulator forwarding.
7513 static SDValue PerformVMULCombine(SDNode *N,
7514 TargetLowering::DAGCombinerInfo &DCI,
7515 const ARMSubtarget *Subtarget) {
7516 if (!Subtarget->hasVMLxForwarding())
7519 SelectionDAG &DAG = DCI.DAG;
7520 SDValue N0 = N->getOperand(0);
7521 SDValue N1 = N->getOperand(1);
7522 unsigned Opcode = N0.getOpcode();
7523 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7524 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7525 Opcode = N1.getOpcode();
7526 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7527 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7532 EVT VT = N->getValueType(0);
7533 DebugLoc DL = N->getDebugLoc();
7534 SDValue N00 = N0->getOperand(0);
7535 SDValue N01 = N0->getOperand(1);
7536 return DAG.getNode(Opcode, DL, VT,
7537 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7538 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7541 static SDValue PerformMULCombine(SDNode *N,
7542 TargetLowering::DAGCombinerInfo &DCI,
7543 const ARMSubtarget *Subtarget) {
7544 SelectionDAG &DAG = DCI.DAG;
7546 if (Subtarget->isThumb1Only())
7549 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7552 EVT VT = N->getValueType(0);
7553 if (VT.is64BitVector() || VT.is128BitVector())
7554 return PerformVMULCombine(N, DCI, Subtarget);
7558 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7562 int64_t MulAmt = C->getSExtValue();
7563 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7565 ShiftAmt = ShiftAmt & (32 - 1);
7566 SDValue V = N->getOperand(0);
7567 DebugLoc DL = N->getDebugLoc();
7570 MulAmt >>= ShiftAmt;
7573 if (isPowerOf2_32(MulAmt - 1)) {
7574 // (mul x, 2^N + 1) => (add (shl x, N), x)
7575 Res = DAG.getNode(ISD::ADD, DL, VT,
7577 DAG.getNode(ISD::SHL, DL, VT,
7579 DAG.getConstant(Log2_32(MulAmt - 1),
7581 } else if (isPowerOf2_32(MulAmt + 1)) {
7582 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7583 Res = DAG.getNode(ISD::SUB, DL, VT,
7584 DAG.getNode(ISD::SHL, DL, VT,
7586 DAG.getConstant(Log2_32(MulAmt + 1),
7592 uint64_t MulAmtAbs = -MulAmt;
7593 if (isPowerOf2_32(MulAmtAbs + 1)) {
7594 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7595 Res = DAG.getNode(ISD::SUB, DL, VT,
7597 DAG.getNode(ISD::SHL, DL, VT,
7599 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7601 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7602 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7603 Res = DAG.getNode(ISD::ADD, DL, VT,
7605 DAG.getNode(ISD::SHL, DL, VT,
7607 DAG.getConstant(Log2_32(MulAmtAbs-1),
7609 Res = DAG.getNode(ISD::SUB, DL, VT,
7610 DAG.getConstant(0, MVT::i32),Res);
7617 Res = DAG.getNode(ISD::SHL, DL, VT,
7618 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7620 // Do not add new nodes to DAG combiner worklist.
7621 DCI.CombineTo(N, Res, false);
7625 static SDValue PerformANDCombine(SDNode *N,
7626 TargetLowering::DAGCombinerInfo &DCI,
7627 const ARMSubtarget *Subtarget) {
7629 // Attempt to use immediate-form VBIC
7630 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7631 DebugLoc dl = N->getDebugLoc();
7632 EVT VT = N->getValueType(0);
7633 SelectionDAG &DAG = DCI.DAG;
7635 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7638 APInt SplatBits, SplatUndef;
7639 unsigned SplatBitSize;
7642 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7643 if (SplatBitSize <= 64) {
7645 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7646 SplatUndef.getZExtValue(), SplatBitSize,
7647 DAG, VbicVT, VT.is128BitVector(),
7649 if (Val.getNode()) {
7651 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7652 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7653 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7658 if (!Subtarget->isThumb1Only()) {
7659 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7660 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7661 if (Result.getNode())
7668 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7669 static SDValue PerformORCombine(SDNode *N,
7670 TargetLowering::DAGCombinerInfo &DCI,
7671 const ARMSubtarget *Subtarget) {
7672 // Attempt to use immediate-form VORR
7673 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7674 DebugLoc dl = N->getDebugLoc();
7675 EVT VT = N->getValueType(0);
7676 SelectionDAG &DAG = DCI.DAG;
7678 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7681 APInt SplatBits, SplatUndef;
7682 unsigned SplatBitSize;
7684 if (BVN && Subtarget->hasNEON() &&
7685 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7686 if (SplatBitSize <= 64) {
7688 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7689 SplatUndef.getZExtValue(), SplatBitSize,
7690 DAG, VorrVT, VT.is128BitVector(),
7692 if (Val.getNode()) {
7694 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7695 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7696 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7701 if (!Subtarget->isThumb1Only()) {
7702 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7703 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7704 if (Result.getNode())
7708 // The code below optimizes (or (and X, Y), Z).
7709 // The AND operand needs to have a single user to make these optimizations
7711 SDValue N0 = N->getOperand(0);
7712 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7714 SDValue N1 = N->getOperand(1);
7716 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7717 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7718 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7720 unsigned SplatBitSize;
7723 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7725 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7726 HasAnyUndefs) && !HasAnyUndefs) {
7727 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7729 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7730 HasAnyUndefs) && !HasAnyUndefs &&
7731 SplatBits0 == ~SplatBits1) {
7732 // Canonicalize the vector type to make instruction selection simpler.
7733 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7734 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7735 N0->getOperand(1), N0->getOperand(0),
7737 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7742 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7745 // BFI is only available on V6T2+
7746 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7749 DebugLoc DL = N->getDebugLoc();
7750 // 1) or (and A, mask), val => ARMbfi A, val, mask
7751 // iff (val & mask) == val
7753 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7754 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7755 // && mask == ~mask2
7756 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7757 // && ~mask == mask2
7758 // (i.e., copy a bitfield value into another bitfield of the same width)
7763 SDValue N00 = N0.getOperand(0);
7765 // The value and the mask need to be constants so we can verify this is
7766 // actually a bitfield set. If the mask is 0xffff, we can do better
7767 // via a movt instruction, so don't use BFI in that case.
7768 SDValue MaskOp = N0.getOperand(1);
7769 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7772 unsigned Mask = MaskC->getZExtValue();
7776 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7777 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7779 unsigned Val = N1C->getZExtValue();
7780 if ((Val & ~Mask) != Val)
7783 if (ARM::isBitFieldInvertedMask(Mask)) {
7784 Val >>= CountTrailingZeros_32(~Mask);
7786 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7787 DAG.getConstant(Val, MVT::i32),
7788 DAG.getConstant(Mask, MVT::i32));
7790 // Do not add new nodes to DAG combiner worklist.
7791 DCI.CombineTo(N, Res, false);
7794 } else if (N1.getOpcode() == ISD::AND) {
7795 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7796 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7799 unsigned Mask2 = N11C->getZExtValue();
7801 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7803 if (ARM::isBitFieldInvertedMask(Mask) &&
7805 // The pack halfword instruction works better for masks that fit it,
7806 // so use that when it's available.
7807 if (Subtarget->hasT2ExtractPack() &&
7808 (Mask == 0xffff || Mask == 0xffff0000))
7811 unsigned amt = CountTrailingZeros_32(Mask2);
7812 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7813 DAG.getConstant(amt, MVT::i32));
7814 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7815 DAG.getConstant(Mask, MVT::i32));
7816 // Do not add new nodes to DAG combiner worklist.
7817 DCI.CombineTo(N, Res, false);
7819 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7821 // The pack halfword instruction works better for masks that fit it,
7822 // so use that when it's available.
7823 if (Subtarget->hasT2ExtractPack() &&
7824 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7827 unsigned lsb = CountTrailingZeros_32(Mask);
7828 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7829 DAG.getConstant(lsb, MVT::i32));
7830 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7831 DAG.getConstant(Mask2, MVT::i32));
7832 // Do not add new nodes to DAG combiner worklist.
7833 DCI.CombineTo(N, Res, false);
7838 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7839 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7840 ARM::isBitFieldInvertedMask(~Mask)) {
7841 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7842 // where lsb(mask) == #shamt and masked bits of B are known zero.
7843 SDValue ShAmt = N00.getOperand(1);
7844 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7845 unsigned LSB = CountTrailingZeros_32(Mask);
7849 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7850 DAG.getConstant(~Mask, MVT::i32));
7852 // Do not add new nodes to DAG combiner worklist.
7853 DCI.CombineTo(N, Res, false);
7859 static SDValue PerformXORCombine(SDNode *N,
7860 TargetLowering::DAGCombinerInfo &DCI,
7861 const ARMSubtarget *Subtarget) {
7862 EVT VT = N->getValueType(0);
7863 SelectionDAG &DAG = DCI.DAG;
7865 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7868 if (!Subtarget->isThumb1Only()) {
7869 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7870 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7871 if (Result.getNode())
7878 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7879 /// the bits being cleared by the AND are not demanded by the BFI.
7880 static SDValue PerformBFICombine(SDNode *N,
7881 TargetLowering::DAGCombinerInfo &DCI) {
7882 SDValue N1 = N->getOperand(1);
7883 if (N1.getOpcode() == ISD::AND) {
7884 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7887 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7888 unsigned LSB = CountTrailingZeros_32(~InvMask);
7889 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7890 unsigned Mask = (1 << Width)-1;
7891 unsigned Mask2 = N11C->getZExtValue();
7892 if ((Mask & (~Mask2)) == 0)
7893 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7894 N->getOperand(0), N1.getOperand(0),
7900 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7901 /// ARMISD::VMOVRRD.
7902 static SDValue PerformVMOVRRDCombine(SDNode *N,
7903 TargetLowering::DAGCombinerInfo &DCI) {
7904 // vmovrrd(vmovdrr x, y) -> x,y
7905 SDValue InDouble = N->getOperand(0);
7906 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7907 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7909 // vmovrrd(load f64) -> (load i32), (load i32)
7910 SDNode *InNode = InDouble.getNode();
7911 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7912 InNode->getValueType(0) == MVT::f64 &&
7913 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7914 !cast<LoadSDNode>(InNode)->isVolatile()) {
7915 // TODO: Should this be done for non-FrameIndex operands?
7916 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7918 SelectionDAG &DAG = DCI.DAG;
7919 DebugLoc DL = LD->getDebugLoc();
7920 SDValue BasePtr = LD->getBasePtr();
7921 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7922 LD->getPointerInfo(), LD->isVolatile(),
7923 LD->isNonTemporal(), LD->isInvariant(),
7924 LD->getAlignment());
7926 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7927 DAG.getConstant(4, MVT::i32));
7928 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7929 LD->getPointerInfo(), LD->isVolatile(),
7930 LD->isNonTemporal(), LD->isInvariant(),
7931 std::min(4U, LD->getAlignment() / 2));
7933 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7934 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7935 DCI.RemoveFromWorklist(LD);
7943 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7944 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7945 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7946 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7947 SDValue Op0 = N->getOperand(0);
7948 SDValue Op1 = N->getOperand(1);
7949 if (Op0.getOpcode() == ISD::BITCAST)
7950 Op0 = Op0.getOperand(0);
7951 if (Op1.getOpcode() == ISD::BITCAST)
7952 Op1 = Op1.getOperand(0);
7953 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7954 Op0.getNode() == Op1.getNode() &&
7955 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7956 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7957 N->getValueType(0), Op0.getOperand(0));
7961 /// PerformSTORECombine - Target-specific dag combine xforms for
7963 static SDValue PerformSTORECombine(SDNode *N,
7964 TargetLowering::DAGCombinerInfo &DCI) {
7965 StoreSDNode *St = cast<StoreSDNode>(N);
7966 if (St->isVolatile())
7969 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7970 // pack all of the elements in one place. Next, store to memory in fewer
7972 SDValue StVal = St->getValue();
7973 EVT VT = StVal.getValueType();
7974 if (St->isTruncatingStore() && VT.isVector()) {
7975 SelectionDAG &DAG = DCI.DAG;
7976 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7977 EVT StVT = St->getMemoryVT();
7978 unsigned NumElems = VT.getVectorNumElements();
7979 assert(StVT != VT && "Cannot truncate to the same type");
7980 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7981 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7983 // From, To sizes and ElemCount must be pow of two
7984 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7986 // We are going to use the original vector elt for storing.
7987 // Accumulated smaller vector elements must be a multiple of the store size.
7988 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7990 unsigned SizeRatio = FromEltSz / ToEltSz;
7991 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7993 // Create a type on which we perform the shuffle.
7994 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7995 NumElems*SizeRatio);
7996 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7998 DebugLoc DL = St->getDebugLoc();
7999 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8000 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8001 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8003 // Can't shuffle using an illegal type.
8004 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8006 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8007 DAG.getUNDEF(WideVec.getValueType()),
8009 // At this point all of the data is stored at the bottom of the
8010 // register. We now need to save it to mem.
8012 // Find the largest store unit
8013 MVT StoreType = MVT::i8;
8014 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8015 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8016 MVT Tp = (MVT::SimpleValueType)tp;
8017 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8020 // Didn't find a legal store type.
8021 if (!TLI.isTypeLegal(StoreType))
8024 // Bitcast the original vector into a vector of store-size units
8025 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8026 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8027 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8028 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8029 SmallVector<SDValue, 8> Chains;
8030 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8031 TLI.getPointerTy());
8032 SDValue BasePtr = St->getBasePtr();
8034 // Perform one or more big stores into memory.
8035 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8036 for (unsigned I = 0; I < E; I++) {
8037 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8038 StoreType, ShuffWide,
8039 DAG.getIntPtrConstant(I));
8040 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8041 St->getPointerInfo(), St->isVolatile(),
8042 St->isNonTemporal(), St->getAlignment());
8043 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8045 Chains.push_back(Ch);
8047 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8051 if (!ISD::isNormalStore(St))
8054 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8055 // ARM stores of arguments in the same cache line.
8056 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8057 StVal.getNode()->hasOneUse()) {
8058 SelectionDAG &DAG = DCI.DAG;
8059 DebugLoc DL = St->getDebugLoc();
8060 SDValue BasePtr = St->getBasePtr();
8061 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8062 StVal.getNode()->getOperand(0), BasePtr,
8063 St->getPointerInfo(), St->isVolatile(),
8064 St->isNonTemporal(), St->getAlignment());
8066 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8067 DAG.getConstant(4, MVT::i32));
8068 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8069 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8070 St->isNonTemporal(),
8071 std::min(4U, St->getAlignment() / 2));
8074 if (StVal.getValueType() != MVT::i64 ||
8075 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8078 // Bitcast an i64 store extracted from a vector to f64.
8079 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8080 SelectionDAG &DAG = DCI.DAG;
8081 DebugLoc dl = StVal.getDebugLoc();
8082 SDValue IntVec = StVal.getOperand(0);
8083 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8084 IntVec.getValueType().getVectorNumElements());
8085 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8086 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8087 Vec, StVal.getOperand(1));
8088 dl = N->getDebugLoc();
8089 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8090 // Make the DAGCombiner fold the bitcasts.
8091 DCI.AddToWorklist(Vec.getNode());
8092 DCI.AddToWorklist(ExtElt.getNode());
8093 DCI.AddToWorklist(V.getNode());
8094 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8095 St->getPointerInfo(), St->isVolatile(),
8096 St->isNonTemporal(), St->getAlignment(),
8100 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8101 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8102 /// i64 vector to have f64 elements, since the value can then be loaded
8103 /// directly into a VFP register.
8104 static bool hasNormalLoadOperand(SDNode *N) {
8105 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8106 for (unsigned i = 0; i < NumElts; ++i) {
8107 SDNode *Elt = N->getOperand(i).getNode();
8108 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8114 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8115 /// ISD::BUILD_VECTOR.
8116 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8117 TargetLowering::DAGCombinerInfo &DCI){
8118 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8119 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8120 // into a pair of GPRs, which is fine when the value is used as a scalar,
8121 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8122 SelectionDAG &DAG = DCI.DAG;
8123 if (N->getNumOperands() == 2) {
8124 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8129 // Load i64 elements as f64 values so that type legalization does not split
8130 // them up into i32 values.
8131 EVT VT = N->getValueType(0);
8132 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8134 DebugLoc dl = N->getDebugLoc();
8135 SmallVector<SDValue, 8> Ops;
8136 unsigned NumElts = VT.getVectorNumElements();
8137 for (unsigned i = 0; i < NumElts; ++i) {
8138 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8140 // Make the DAGCombiner fold the bitcast.
8141 DCI.AddToWorklist(V.getNode());
8143 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8144 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8145 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8148 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8149 /// ISD::INSERT_VECTOR_ELT.
8150 static SDValue PerformInsertEltCombine(SDNode *N,
8151 TargetLowering::DAGCombinerInfo &DCI) {
8152 // Bitcast an i64 load inserted into a vector to f64.
8153 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8154 EVT VT = N->getValueType(0);
8155 SDNode *Elt = N->getOperand(1).getNode();
8156 if (VT.getVectorElementType() != MVT::i64 ||
8157 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8160 SelectionDAG &DAG = DCI.DAG;
8161 DebugLoc dl = N->getDebugLoc();
8162 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8163 VT.getVectorNumElements());
8164 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8165 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8166 // Make the DAGCombiner fold the bitcasts.
8167 DCI.AddToWorklist(Vec.getNode());
8168 DCI.AddToWorklist(V.getNode());
8169 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8170 Vec, V, N->getOperand(2));
8171 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8174 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8175 /// ISD::VECTOR_SHUFFLE.
8176 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8177 // The LLVM shufflevector instruction does not require the shuffle mask
8178 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8179 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8180 // operands do not match the mask length, they are extended by concatenating
8181 // them with undef vectors. That is probably the right thing for other
8182 // targets, but for NEON it is better to concatenate two double-register
8183 // size vector operands into a single quad-register size vector. Do that
8184 // transformation here:
8185 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8186 // shuffle(concat(v1, v2), undef)
8187 SDValue Op0 = N->getOperand(0);
8188 SDValue Op1 = N->getOperand(1);
8189 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8190 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8191 Op0.getNumOperands() != 2 ||
8192 Op1.getNumOperands() != 2)
8194 SDValue Concat0Op1 = Op0.getOperand(1);
8195 SDValue Concat1Op1 = Op1.getOperand(1);
8196 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8197 Concat1Op1.getOpcode() != ISD::UNDEF)
8199 // Skip the transformation if any of the types are illegal.
8200 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8201 EVT VT = N->getValueType(0);
8202 if (!TLI.isTypeLegal(VT) ||
8203 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8204 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8207 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8208 Op0.getOperand(0), Op1.getOperand(0));
8209 // Translate the shuffle mask.
8210 SmallVector<int, 16> NewMask;
8211 unsigned NumElts = VT.getVectorNumElements();
8212 unsigned HalfElts = NumElts/2;
8213 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8214 for (unsigned n = 0; n < NumElts; ++n) {
8215 int MaskElt = SVN->getMaskElt(n);
8217 if (MaskElt < (int)HalfElts)
8219 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8220 NewElt = HalfElts + MaskElt - NumElts;
8221 NewMask.push_back(NewElt);
8223 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8224 DAG.getUNDEF(VT), NewMask.data());
8227 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8228 /// NEON load/store intrinsics to merge base address updates.
8229 static SDValue CombineBaseUpdate(SDNode *N,
8230 TargetLowering::DAGCombinerInfo &DCI) {
8231 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8234 SelectionDAG &DAG = DCI.DAG;
8235 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8236 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8237 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8238 SDValue Addr = N->getOperand(AddrOpIdx);
8240 // Search for a use of the address operand that is an increment.
8241 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8242 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8244 if (User->getOpcode() != ISD::ADD ||
8245 UI.getUse().getResNo() != Addr.getResNo())
8248 // Check that the add is independent of the load/store. Otherwise, folding
8249 // it would create a cycle.
8250 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8253 // Find the new opcode for the updating load/store.
8255 bool isLaneOp = false;
8256 unsigned NewOpc = 0;
8257 unsigned NumVecs = 0;
8259 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8261 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8262 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8264 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8266 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8268 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8270 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8271 NumVecs = 2; isLaneOp = true; break;
8272 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8273 NumVecs = 3; isLaneOp = true; break;
8274 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8275 NumVecs = 4; isLaneOp = true; break;
8276 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8277 NumVecs = 1; isLoad = false; break;
8278 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8279 NumVecs = 2; isLoad = false; break;
8280 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8281 NumVecs = 3; isLoad = false; break;
8282 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8283 NumVecs = 4; isLoad = false; break;
8284 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8285 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8286 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8287 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8288 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8289 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8293 switch (N->getOpcode()) {
8294 default: llvm_unreachable("unexpected opcode for Neon base update");
8295 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8296 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8297 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8301 // Find the size of memory referenced by the load/store.
8304 VecTy = N->getValueType(0);
8306 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8307 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8309 NumBytes /= VecTy.getVectorNumElements();
8311 // If the increment is a constant, it must match the memory ref size.
8312 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8313 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8314 uint64_t IncVal = CInc->getZExtValue();
8315 if (IncVal != NumBytes)
8317 } else if (NumBytes >= 3 * 16) {
8318 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8319 // separate instructions that make it harder to use a non-constant update.
8323 // Create the new updating load/store node.
8325 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8327 for (n = 0; n < NumResultVecs; ++n)
8329 Tys[n++] = MVT::i32;
8330 Tys[n] = MVT::Other;
8331 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8332 SmallVector<SDValue, 8> Ops;
8333 Ops.push_back(N->getOperand(0)); // incoming chain
8334 Ops.push_back(N->getOperand(AddrOpIdx));
8336 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8337 Ops.push_back(N->getOperand(i));
8339 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8340 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8341 Ops.data(), Ops.size(),
8342 MemInt->getMemoryVT(),
8343 MemInt->getMemOperand());
8346 std::vector<SDValue> NewResults;
8347 for (unsigned i = 0; i < NumResultVecs; ++i) {
8348 NewResults.push_back(SDValue(UpdN.getNode(), i));
8350 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8351 DCI.CombineTo(N, NewResults);
8352 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8359 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8360 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8361 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8363 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8364 SelectionDAG &DAG = DCI.DAG;
8365 EVT VT = N->getValueType(0);
8366 // vldN-dup instructions only support 64-bit vectors for N > 1.
8367 if (!VT.is64BitVector())
8370 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8371 SDNode *VLD = N->getOperand(0).getNode();
8372 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8374 unsigned NumVecs = 0;
8375 unsigned NewOpc = 0;
8376 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8377 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8379 NewOpc = ARMISD::VLD2DUP;
8380 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8382 NewOpc = ARMISD::VLD3DUP;
8383 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8385 NewOpc = ARMISD::VLD4DUP;
8390 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8391 // numbers match the load.
8392 unsigned VLDLaneNo =
8393 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8394 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8396 // Ignore uses of the chain result.
8397 if (UI.getUse().getResNo() == NumVecs)
8400 if (User->getOpcode() != ARMISD::VDUPLANE ||
8401 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8405 // Create the vldN-dup node.
8408 for (n = 0; n < NumVecs; ++n)
8410 Tys[n] = MVT::Other;
8411 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8412 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8413 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8414 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8415 Ops, 2, VLDMemInt->getMemoryVT(),
8416 VLDMemInt->getMemOperand());
8419 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8421 unsigned ResNo = UI.getUse().getResNo();
8422 // Ignore uses of the chain result.
8423 if (ResNo == NumVecs)
8426 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8429 // Now the vldN-lane intrinsic is dead except for its chain result.
8430 // Update uses of the chain.
8431 std::vector<SDValue> VLDDupResults;
8432 for (unsigned n = 0; n < NumVecs; ++n)
8433 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8434 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8435 DCI.CombineTo(VLD, VLDDupResults);
8440 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8441 /// ARMISD::VDUPLANE.
8442 static SDValue PerformVDUPLANECombine(SDNode *N,
8443 TargetLowering::DAGCombinerInfo &DCI) {
8444 SDValue Op = N->getOperand(0);
8446 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8447 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8448 if (CombineVLDDUP(N, DCI))
8449 return SDValue(N, 0);
8451 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8452 // redundant. Ignore bit_converts for now; element sizes are checked below.
8453 while (Op.getOpcode() == ISD::BITCAST)
8454 Op = Op.getOperand(0);
8455 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8458 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8459 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8460 // The canonical VMOV for a zero vector uses a 32-bit element size.
8461 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8463 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8465 EVT VT = N->getValueType(0);
8466 if (EltSize > VT.getVectorElementType().getSizeInBits())
8469 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8472 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8473 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8474 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8478 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8480 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8485 APFloat APF = C->getValueAPF();
8486 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8487 != APFloat::opOK || !isExact)
8490 c0 = (I == 0) ? cN : c0;
8491 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8498 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8499 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8500 /// when the VMUL has a constant operand that is a power of 2.
8502 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8503 /// vmul.f32 d16, d17, d16
8504 /// vcvt.s32.f32 d16, d16
8506 /// vcvt.s32.f32 d16, d16, #3
8507 static SDValue PerformVCVTCombine(SDNode *N,
8508 TargetLowering::DAGCombinerInfo &DCI,
8509 const ARMSubtarget *Subtarget) {
8510 SelectionDAG &DAG = DCI.DAG;
8511 SDValue Op = N->getOperand(0);
8513 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8514 Op.getOpcode() != ISD::FMUL)
8518 SDValue N0 = Op->getOperand(0);
8519 SDValue ConstVec = Op->getOperand(1);
8520 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8522 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8523 !isConstVecPow2(ConstVec, isSigned, C))
8526 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8527 Intrinsic::arm_neon_vcvtfp2fxu;
8528 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8530 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8531 DAG.getConstant(Log2_64(C), MVT::i32));
8534 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8535 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8536 /// when the VDIV has a constant operand that is a power of 2.
8538 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8539 /// vcvt.f32.s32 d16, d16
8540 /// vdiv.f32 d16, d17, d16
8542 /// vcvt.f32.s32 d16, d16, #3
8543 static SDValue PerformVDIVCombine(SDNode *N,
8544 TargetLowering::DAGCombinerInfo &DCI,
8545 const ARMSubtarget *Subtarget) {
8546 SelectionDAG &DAG = DCI.DAG;
8547 SDValue Op = N->getOperand(0);
8548 unsigned OpOpcode = Op.getNode()->getOpcode();
8550 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8551 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8555 SDValue ConstVec = N->getOperand(1);
8556 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8558 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8559 !isConstVecPow2(ConstVec, isSigned, C))
8562 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8563 Intrinsic::arm_neon_vcvtfxu2fp;
8564 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8566 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8567 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8570 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8571 /// operand of a vector shift operation, where all the elements of the
8572 /// build_vector must have the same constant integer value.
8573 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8574 // Ignore bit_converts.
8575 while (Op.getOpcode() == ISD::BITCAST)
8576 Op = Op.getOperand(0);
8577 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8578 APInt SplatBits, SplatUndef;
8579 unsigned SplatBitSize;
8581 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8582 HasAnyUndefs, ElementBits) ||
8583 SplatBitSize > ElementBits)
8585 Cnt = SplatBits.getSExtValue();
8589 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8590 /// operand of a vector shift left operation. That value must be in the range:
8591 /// 0 <= Value < ElementBits for a left shift; or
8592 /// 0 <= Value <= ElementBits for a long left shift.
8593 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8594 assert(VT.isVector() && "vector shift count is not a vector type");
8595 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8596 if (! getVShiftImm(Op, ElementBits, Cnt))
8598 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8601 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8602 /// operand of a vector shift right operation. For a shift opcode, the value
8603 /// is positive, but for an intrinsic the value count must be negative. The
8604 /// absolute value must be in the range:
8605 /// 1 <= |Value| <= ElementBits for a right shift; or
8606 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8607 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8609 assert(VT.isVector() && "vector shift count is not a vector type");
8610 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8611 if (! getVShiftImm(Op, ElementBits, Cnt))
8615 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8618 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8619 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8620 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8623 // Don't do anything for most intrinsics.
8626 // Vector shifts: check for immediate versions and lower them.
8627 // Note: This is done during DAG combining instead of DAG legalizing because
8628 // the build_vectors for 64-bit vector element shift counts are generally
8629 // not legal, and it is hard to see their values after they get legalized to
8630 // loads from a constant pool.
8631 case Intrinsic::arm_neon_vshifts:
8632 case Intrinsic::arm_neon_vshiftu:
8633 case Intrinsic::arm_neon_vshiftls:
8634 case Intrinsic::arm_neon_vshiftlu:
8635 case Intrinsic::arm_neon_vshiftn:
8636 case Intrinsic::arm_neon_vrshifts:
8637 case Intrinsic::arm_neon_vrshiftu:
8638 case Intrinsic::arm_neon_vrshiftn:
8639 case Intrinsic::arm_neon_vqshifts:
8640 case Intrinsic::arm_neon_vqshiftu:
8641 case Intrinsic::arm_neon_vqshiftsu:
8642 case Intrinsic::arm_neon_vqshiftns:
8643 case Intrinsic::arm_neon_vqshiftnu:
8644 case Intrinsic::arm_neon_vqshiftnsu:
8645 case Intrinsic::arm_neon_vqrshiftns:
8646 case Intrinsic::arm_neon_vqrshiftnu:
8647 case Intrinsic::arm_neon_vqrshiftnsu: {
8648 EVT VT = N->getOperand(1).getValueType();
8650 unsigned VShiftOpc = 0;
8653 case Intrinsic::arm_neon_vshifts:
8654 case Intrinsic::arm_neon_vshiftu:
8655 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8656 VShiftOpc = ARMISD::VSHL;
8659 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8660 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8661 ARMISD::VSHRs : ARMISD::VSHRu);
8666 case Intrinsic::arm_neon_vshiftls:
8667 case Intrinsic::arm_neon_vshiftlu:
8668 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8670 llvm_unreachable("invalid shift count for vshll intrinsic");
8672 case Intrinsic::arm_neon_vrshifts:
8673 case Intrinsic::arm_neon_vrshiftu:
8674 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8678 case Intrinsic::arm_neon_vqshifts:
8679 case Intrinsic::arm_neon_vqshiftu:
8680 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8684 case Intrinsic::arm_neon_vqshiftsu:
8685 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8687 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8689 case Intrinsic::arm_neon_vshiftn:
8690 case Intrinsic::arm_neon_vrshiftn:
8691 case Intrinsic::arm_neon_vqshiftns:
8692 case Intrinsic::arm_neon_vqshiftnu:
8693 case Intrinsic::arm_neon_vqshiftnsu:
8694 case Intrinsic::arm_neon_vqrshiftns:
8695 case Intrinsic::arm_neon_vqrshiftnu:
8696 case Intrinsic::arm_neon_vqrshiftnsu:
8697 // Narrowing shifts require an immediate right shift.
8698 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8700 llvm_unreachable("invalid shift count for narrowing vector shift "
8704 llvm_unreachable("unhandled vector shift");
8708 case Intrinsic::arm_neon_vshifts:
8709 case Intrinsic::arm_neon_vshiftu:
8710 // Opcode already set above.
8712 case Intrinsic::arm_neon_vshiftls:
8713 case Intrinsic::arm_neon_vshiftlu:
8714 if (Cnt == VT.getVectorElementType().getSizeInBits())
8715 VShiftOpc = ARMISD::VSHLLi;
8717 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8718 ARMISD::VSHLLs : ARMISD::VSHLLu);
8720 case Intrinsic::arm_neon_vshiftn:
8721 VShiftOpc = ARMISD::VSHRN; break;
8722 case Intrinsic::arm_neon_vrshifts:
8723 VShiftOpc = ARMISD::VRSHRs; break;
8724 case Intrinsic::arm_neon_vrshiftu:
8725 VShiftOpc = ARMISD::VRSHRu; break;
8726 case Intrinsic::arm_neon_vrshiftn:
8727 VShiftOpc = ARMISD::VRSHRN; break;
8728 case Intrinsic::arm_neon_vqshifts:
8729 VShiftOpc = ARMISD::VQSHLs; break;
8730 case Intrinsic::arm_neon_vqshiftu:
8731 VShiftOpc = ARMISD::VQSHLu; break;
8732 case Intrinsic::arm_neon_vqshiftsu:
8733 VShiftOpc = ARMISD::VQSHLsu; break;
8734 case Intrinsic::arm_neon_vqshiftns:
8735 VShiftOpc = ARMISD::VQSHRNs; break;
8736 case Intrinsic::arm_neon_vqshiftnu:
8737 VShiftOpc = ARMISD::VQSHRNu; break;
8738 case Intrinsic::arm_neon_vqshiftnsu:
8739 VShiftOpc = ARMISD::VQSHRNsu; break;
8740 case Intrinsic::arm_neon_vqrshiftns:
8741 VShiftOpc = ARMISD::VQRSHRNs; break;
8742 case Intrinsic::arm_neon_vqrshiftnu:
8743 VShiftOpc = ARMISD::VQRSHRNu; break;
8744 case Intrinsic::arm_neon_vqrshiftnsu:
8745 VShiftOpc = ARMISD::VQRSHRNsu; break;
8748 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8749 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8752 case Intrinsic::arm_neon_vshiftins: {
8753 EVT VT = N->getOperand(1).getValueType();
8755 unsigned VShiftOpc = 0;
8757 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8758 VShiftOpc = ARMISD::VSLI;
8759 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8760 VShiftOpc = ARMISD::VSRI;
8762 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8765 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8766 N->getOperand(1), N->getOperand(2),
8767 DAG.getConstant(Cnt, MVT::i32));
8770 case Intrinsic::arm_neon_vqrshifts:
8771 case Intrinsic::arm_neon_vqrshiftu:
8772 // No immediate versions of these to check for.
8779 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8780 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8781 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8782 /// vector element shift counts are generally not legal, and it is hard to see
8783 /// their values after they get legalized to loads from a constant pool.
8784 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8785 const ARMSubtarget *ST) {
8786 EVT VT = N->getValueType(0);
8787 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8788 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8789 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8790 SDValue N1 = N->getOperand(1);
8791 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8792 SDValue N0 = N->getOperand(0);
8793 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8794 DAG.MaskedValueIsZero(N0.getOperand(0),
8795 APInt::getHighBitsSet(32, 16)))
8796 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8800 // Nothing to be done for scalar shifts.
8801 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8802 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8805 assert(ST->hasNEON() && "unexpected vector shift");
8808 switch (N->getOpcode()) {
8809 default: llvm_unreachable("unexpected shift opcode");
8812 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8813 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8814 DAG.getConstant(Cnt, MVT::i32));
8819 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8820 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8821 ARMISD::VSHRs : ARMISD::VSHRu);
8822 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8823 DAG.getConstant(Cnt, MVT::i32));
8829 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8830 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8831 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8832 const ARMSubtarget *ST) {
8833 SDValue N0 = N->getOperand(0);
8835 // Check for sign- and zero-extensions of vector extract operations of 8-
8836 // and 16-bit vector elements. NEON supports these directly. They are
8837 // handled during DAG combining because type legalization will promote them
8838 // to 32-bit types and it is messy to recognize the operations after that.
8839 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8840 SDValue Vec = N0.getOperand(0);
8841 SDValue Lane = N0.getOperand(1);
8842 EVT VT = N->getValueType(0);
8843 EVT EltVT = N0.getValueType();
8844 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8846 if (VT == MVT::i32 &&
8847 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8848 TLI.isTypeLegal(Vec.getValueType()) &&
8849 isa<ConstantSDNode>(Lane)) {
8852 switch (N->getOpcode()) {
8853 default: llvm_unreachable("unexpected opcode");
8854 case ISD::SIGN_EXTEND:
8855 Opc = ARMISD::VGETLANEs;
8857 case ISD::ZERO_EXTEND:
8858 case ISD::ANY_EXTEND:
8859 Opc = ARMISD::VGETLANEu;
8862 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8869 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8870 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8871 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8872 const ARMSubtarget *ST) {
8873 // If the target supports NEON, try to use vmax/vmin instructions for f32
8874 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8875 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8876 // a NaN; only do the transformation when it matches that behavior.
8878 // For now only do this when using NEON for FP operations; if using VFP, it
8879 // is not obvious that the benefit outweighs the cost of switching to the
8881 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8882 N->getValueType(0) != MVT::f32)
8885 SDValue CondLHS = N->getOperand(0);
8886 SDValue CondRHS = N->getOperand(1);
8887 SDValue LHS = N->getOperand(2);
8888 SDValue RHS = N->getOperand(3);
8889 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8891 unsigned Opcode = 0;
8893 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8894 IsReversed = false; // x CC y ? x : y
8895 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8896 IsReversed = true ; // x CC y ? y : x
8910 // If LHS is NaN, an ordered comparison will be false and the result will
8911 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8912 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8913 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8914 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8916 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8917 // will return -0, so vmin can only be used for unsafe math or if one of
8918 // the operands is known to be nonzero.
8919 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8920 !DAG.getTarget().Options.UnsafeFPMath &&
8921 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8923 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8932 // If LHS is NaN, an ordered comparison will be false and the result will
8933 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8934 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8935 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8936 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8938 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8939 // will return +0, so vmax can only be used for unsafe math or if one of
8940 // the operands is known to be nonzero.
8941 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8942 !DAG.getTarget().Options.UnsafeFPMath &&
8943 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8945 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8951 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8954 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8956 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8957 SDValue Cmp = N->getOperand(4);
8958 if (Cmp.getOpcode() != ARMISD::CMPZ)
8959 // Only looking at EQ and NE cases.
8962 EVT VT = N->getValueType(0);
8963 DebugLoc dl = N->getDebugLoc();
8964 SDValue LHS = Cmp.getOperand(0);
8965 SDValue RHS = Cmp.getOperand(1);
8966 SDValue FalseVal = N->getOperand(0);
8967 SDValue TrueVal = N->getOperand(1);
8968 SDValue ARMcc = N->getOperand(2);
8969 ARMCC::CondCodes CC =
8970 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8988 /// FIXME: Turn this into a target neutral optimization?
8990 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8991 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8992 N->getOperand(3), Cmp);
8993 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8995 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8996 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8997 N->getOperand(3), NewCmp);
9000 if (Res.getNode()) {
9001 APInt KnownZero, KnownOne;
9002 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9003 // Capture demanded bits information that would be otherwise lost.
9004 if (KnownZero == 0xfffffffe)
9005 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9006 DAG.getValueType(MVT::i1));
9007 else if (KnownZero == 0xffffff00)
9008 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9009 DAG.getValueType(MVT::i8));
9010 else if (KnownZero == 0xffff0000)
9011 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9012 DAG.getValueType(MVT::i16));
9018 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9019 DAGCombinerInfo &DCI) const {
9020 switch (N->getOpcode()) {
9022 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9023 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9024 case ISD::SUB: return PerformSUBCombine(N, DCI);
9025 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9026 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9027 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9028 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9029 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9030 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9031 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9032 case ISD::STORE: return PerformSTORECombine(N, DCI);
9033 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9034 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9035 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9036 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9037 case ISD::FP_TO_SINT:
9038 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9039 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9040 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9043 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9044 case ISD::SIGN_EXTEND:
9045 case ISD::ZERO_EXTEND:
9046 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9047 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9048 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9049 case ARMISD::VLD2DUP:
9050 case ARMISD::VLD3DUP:
9051 case ARMISD::VLD4DUP:
9052 return CombineBaseUpdate(N, DCI);
9053 case ISD::INTRINSIC_VOID:
9054 case ISD::INTRINSIC_W_CHAIN:
9055 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9056 case Intrinsic::arm_neon_vld1:
9057 case Intrinsic::arm_neon_vld2:
9058 case Intrinsic::arm_neon_vld3:
9059 case Intrinsic::arm_neon_vld4:
9060 case Intrinsic::arm_neon_vld2lane:
9061 case Intrinsic::arm_neon_vld3lane:
9062 case Intrinsic::arm_neon_vld4lane:
9063 case Intrinsic::arm_neon_vst1:
9064 case Intrinsic::arm_neon_vst2:
9065 case Intrinsic::arm_neon_vst3:
9066 case Intrinsic::arm_neon_vst4:
9067 case Intrinsic::arm_neon_vst2lane:
9068 case Intrinsic::arm_neon_vst3lane:
9069 case Intrinsic::arm_neon_vst4lane:
9070 return CombineBaseUpdate(N, DCI);
9078 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9080 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9083 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9084 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9085 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9087 switch (VT.getSimpleVT().SimpleTy) {
9093 // Unaligned access can use (for example) LRDB, LRDH, LDR
9094 return AllowsUnaligned;
9097 // For any little-endian targets with neon, we can support unaligned ld/st
9098 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9099 // A big-endian target may also explictly support unaligned accesses
9100 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9104 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9105 unsigned AlignCheck) {
9106 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9107 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9110 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9111 unsigned DstAlign, unsigned SrcAlign,
9114 MachineFunction &MF) const {
9115 const Function *F = MF.getFunction();
9117 // See if we can use NEON instructions for this...
9119 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9120 Subtarget->hasNEON()) {
9121 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9123 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9128 // Lowering to i32/i16 if the size permits.
9131 } else if (Size >= 2) {
9135 // Let the target-independent logic figure it out.
9139 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9144 switch (VT.getSimpleVT().SimpleTy) {
9145 default: return false;
9160 if ((V & (Scale - 1)) != 0)
9163 return V == (V & ((1LL << 5) - 1));
9166 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9167 const ARMSubtarget *Subtarget) {
9174 switch (VT.getSimpleVT().SimpleTy) {
9175 default: return false;
9180 // + imm12 or - imm8
9182 return V == (V & ((1LL << 8) - 1));
9183 return V == (V & ((1LL << 12) - 1));
9186 // Same as ARM mode. FIXME: NEON?
9187 if (!Subtarget->hasVFP2())
9192 return V == (V & ((1LL << 8) - 1));
9196 /// isLegalAddressImmediate - Return true if the integer value can be used
9197 /// as the offset of the target addressing mode for load / store of the
9199 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9200 const ARMSubtarget *Subtarget) {
9207 if (Subtarget->isThumb1Only())
9208 return isLegalT1AddressImmediate(V, VT);
9209 else if (Subtarget->isThumb2())
9210 return isLegalT2AddressImmediate(V, VT, Subtarget);
9215 switch (VT.getSimpleVT().SimpleTy) {
9216 default: return false;
9221 return V == (V & ((1LL << 12) - 1));
9224 return V == (V & ((1LL << 8) - 1));
9227 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9232 return V == (V & ((1LL << 8) - 1));
9236 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9238 int Scale = AM.Scale;
9242 switch (VT.getSimpleVT().SimpleTy) {
9243 default: return false;
9252 return Scale == 2 || Scale == 4 || Scale == 8;
9255 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9259 // Note, we allow "void" uses (basically, uses that aren't loads or
9260 // stores), because arm allows folding a scale into many arithmetic
9261 // operations. This should be made more precise and revisited later.
9263 // Allow r << imm, but the imm has to be a multiple of two.
9264 if (Scale & 1) return false;
9265 return isPowerOf2_32(Scale);
9269 /// isLegalAddressingMode - Return true if the addressing mode represented
9270 /// by AM is legal for this target, for a load/store of the specified type.
9271 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9273 EVT VT = getValueType(Ty, true);
9274 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9277 // Can never fold addr of global into load/store.
9282 case 0: // no scale reg, must be "r+i" or "r", or "i".
9285 if (Subtarget->isThumb1Only())
9289 // ARM doesn't support any R+R*scale+imm addr modes.
9296 if (Subtarget->isThumb2())
9297 return isLegalT2ScaledAddressingMode(AM, VT);
9299 int Scale = AM.Scale;
9300 switch (VT.getSimpleVT().SimpleTy) {
9301 default: return false;
9305 if (Scale < 0) Scale = -Scale;
9309 return isPowerOf2_32(Scale & ~1);
9313 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9318 // Note, we allow "void" uses (basically, uses that aren't loads or
9319 // stores), because arm allows folding a scale into many arithmetic
9320 // operations. This should be made more precise and revisited later.
9322 // Allow r << imm, but the imm has to be a multiple of two.
9323 if (Scale & 1) return false;
9324 return isPowerOf2_32(Scale);
9330 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9331 /// icmp immediate, that is the target has icmp instructions which can compare
9332 /// a register against the immediate without having to materialize the
9333 /// immediate into a register.
9334 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9335 // Thumb2 and ARM modes can use cmn for negative immediates.
9336 if (!Subtarget->isThumb())
9337 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9338 if (Subtarget->isThumb2())
9339 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9340 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9341 return Imm >= 0 && Imm <= 255;
9344 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9345 /// *or sub* immediate, that is the target has add or sub instructions which can
9346 /// add a register with the immediate without having to materialize the
9347 /// immediate into a register.
9348 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9349 // Same encoding for add/sub, just flip the sign.
9350 int64_t AbsImm = llvm::abs64(Imm);
9351 if (!Subtarget->isThumb())
9352 return ARM_AM::getSOImmVal(AbsImm) != -1;
9353 if (Subtarget->isThumb2())
9354 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9355 // Thumb1 only has 8-bit unsigned immediate.
9356 return AbsImm >= 0 && AbsImm <= 255;
9359 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9360 bool isSEXTLoad, SDValue &Base,
9361 SDValue &Offset, bool &isInc,
9362 SelectionDAG &DAG) {
9363 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9366 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9368 Base = Ptr->getOperand(0);
9369 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9370 int RHSC = (int)RHS->getZExtValue();
9371 if (RHSC < 0 && RHSC > -256) {
9372 assert(Ptr->getOpcode() == ISD::ADD);
9374 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9378 isInc = (Ptr->getOpcode() == ISD::ADD);
9379 Offset = Ptr->getOperand(1);
9381 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9383 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9384 int RHSC = (int)RHS->getZExtValue();
9385 if (RHSC < 0 && RHSC > -0x1000) {
9386 assert(Ptr->getOpcode() == ISD::ADD);
9388 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9389 Base = Ptr->getOperand(0);
9394 if (Ptr->getOpcode() == ISD::ADD) {
9396 ARM_AM::ShiftOpc ShOpcVal=
9397 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9398 if (ShOpcVal != ARM_AM::no_shift) {
9399 Base = Ptr->getOperand(1);
9400 Offset = Ptr->getOperand(0);
9402 Base = Ptr->getOperand(0);
9403 Offset = Ptr->getOperand(1);
9408 isInc = (Ptr->getOpcode() == ISD::ADD);
9409 Base = Ptr->getOperand(0);
9410 Offset = Ptr->getOperand(1);
9414 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9418 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9419 bool isSEXTLoad, SDValue &Base,
9420 SDValue &Offset, bool &isInc,
9421 SelectionDAG &DAG) {
9422 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9425 Base = Ptr->getOperand(0);
9426 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9427 int RHSC = (int)RHS->getZExtValue();
9428 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9429 assert(Ptr->getOpcode() == ISD::ADD);
9431 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9433 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9434 isInc = Ptr->getOpcode() == ISD::ADD;
9435 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9443 /// getPreIndexedAddressParts - returns true by value, base pointer and
9444 /// offset pointer and addressing mode by reference if the node's address
9445 /// can be legally represented as pre-indexed load / store address.
9447 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9449 ISD::MemIndexedMode &AM,
9450 SelectionDAG &DAG) const {
9451 if (Subtarget->isThumb1Only())
9456 bool isSEXTLoad = false;
9457 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9458 Ptr = LD->getBasePtr();
9459 VT = LD->getMemoryVT();
9460 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9461 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9462 Ptr = ST->getBasePtr();
9463 VT = ST->getMemoryVT();
9468 bool isLegal = false;
9469 if (Subtarget->isThumb2())
9470 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9471 Offset, isInc, DAG);
9473 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9474 Offset, isInc, DAG);
9478 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9482 /// getPostIndexedAddressParts - returns true by value, base pointer and
9483 /// offset pointer and addressing mode by reference if this node can be
9484 /// combined with a load / store to form a post-indexed load / store.
9485 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9488 ISD::MemIndexedMode &AM,
9489 SelectionDAG &DAG) const {
9490 if (Subtarget->isThumb1Only())
9495 bool isSEXTLoad = false;
9496 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9497 VT = LD->getMemoryVT();
9498 Ptr = LD->getBasePtr();
9499 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9500 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9501 VT = ST->getMemoryVT();
9502 Ptr = ST->getBasePtr();
9507 bool isLegal = false;
9508 if (Subtarget->isThumb2())
9509 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9512 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9518 // Swap base ptr and offset to catch more post-index load / store when
9519 // it's legal. In Thumb2 mode, offset must be an immediate.
9520 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9521 !Subtarget->isThumb2())
9522 std::swap(Base, Offset);
9524 // Post-indexed load / store update the base pointer.
9529 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9533 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9536 const SelectionDAG &DAG,
9537 unsigned Depth) const {
9538 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9539 switch (Op.getOpcode()) {
9541 case ARMISD::CMOV: {
9542 // Bits are known zero/one if known on the LHS and RHS.
9543 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9544 if (KnownZero == 0 && KnownOne == 0) return;
9546 APInt KnownZeroRHS, KnownOneRHS;
9547 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9548 KnownZero &= KnownZeroRHS;
9549 KnownOne &= KnownOneRHS;
9555 //===----------------------------------------------------------------------===//
9556 // ARM Inline Assembly Support
9557 //===----------------------------------------------------------------------===//
9559 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9560 // Looking for "rev" which is V6+.
9561 if (!Subtarget->hasV6Ops())
9564 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9565 std::string AsmStr = IA->getAsmString();
9566 SmallVector<StringRef, 4> AsmPieces;
9567 SplitString(AsmStr, AsmPieces, ";\n");
9569 switch (AsmPieces.size()) {
9570 default: return false;
9572 AsmStr = AsmPieces[0];
9574 SplitString(AsmStr, AsmPieces, " \t,");
9577 if (AsmPieces.size() == 3 &&
9578 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9579 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9580 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9581 if (Ty && Ty->getBitWidth() == 32)
9582 return IntrinsicLowering::LowerToByteSwap(CI);
9590 /// getConstraintType - Given a constraint letter, return the type of
9591 /// constraint it is for this target.
9592 ARMTargetLowering::ConstraintType
9593 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9594 if (Constraint.size() == 1) {
9595 switch (Constraint[0]) {
9597 case 'l': return C_RegisterClass;
9598 case 'w': return C_RegisterClass;
9599 case 'h': return C_RegisterClass;
9600 case 'x': return C_RegisterClass;
9601 case 't': return C_RegisterClass;
9602 case 'j': return C_Other; // Constant for movw.
9603 // An address with a single base register. Due to the way we
9604 // currently handle addresses it is the same as an 'r' memory constraint.
9605 case 'Q': return C_Memory;
9607 } else if (Constraint.size() == 2) {
9608 switch (Constraint[0]) {
9610 // All 'U+' constraints are addresses.
9611 case 'U': return C_Memory;
9614 return TargetLowering::getConstraintType(Constraint);
9617 /// Examine constraint type and operand type and determine a weight value.
9618 /// This object must already have been set up with the operand type
9619 /// and the current alternative constraint selected.
9620 TargetLowering::ConstraintWeight
9621 ARMTargetLowering::getSingleConstraintMatchWeight(
9622 AsmOperandInfo &info, const char *constraint) const {
9623 ConstraintWeight weight = CW_Invalid;
9624 Value *CallOperandVal = info.CallOperandVal;
9625 // If we don't have a value, we can't do a match,
9626 // but allow it at the lowest weight.
9627 if (CallOperandVal == NULL)
9629 Type *type = CallOperandVal->getType();
9630 // Look at the constraint type.
9631 switch (*constraint) {
9633 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9636 if (type->isIntegerTy()) {
9637 if (Subtarget->isThumb())
9638 weight = CW_SpecificReg;
9640 weight = CW_Register;
9644 if (type->isFloatingPointTy())
9645 weight = CW_Register;
9651 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9653 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9655 if (Constraint.size() == 1) {
9656 // GCC ARM Constraint Letters
9657 switch (Constraint[0]) {
9658 case 'l': // Low regs or general regs.
9659 if (Subtarget->isThumb())
9660 return RCPair(0U, &ARM::tGPRRegClass);
9661 return RCPair(0U, &ARM::GPRRegClass);
9662 case 'h': // High regs or no regs.
9663 if (Subtarget->isThumb())
9664 return RCPair(0U, &ARM::hGPRRegClass);
9667 return RCPair(0U, &ARM::GPRRegClass);
9670 return RCPair(0U, &ARM::SPRRegClass);
9671 if (VT.getSizeInBits() == 64)
9672 return RCPair(0U, &ARM::DPRRegClass);
9673 if (VT.getSizeInBits() == 128)
9674 return RCPair(0U, &ARM::QPRRegClass);
9678 return RCPair(0U, &ARM::SPR_8RegClass);
9679 if (VT.getSizeInBits() == 64)
9680 return RCPair(0U, &ARM::DPR_8RegClass);
9681 if (VT.getSizeInBits() == 128)
9682 return RCPair(0U, &ARM::QPR_8RegClass);
9686 return RCPair(0U, &ARM::SPRRegClass);
9690 if (StringRef("{cc}").equals_lower(Constraint))
9691 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9693 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9696 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9697 /// vector. If it is invalid, don't add anything to Ops.
9698 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9699 std::string &Constraint,
9700 std::vector<SDValue>&Ops,
9701 SelectionDAG &DAG) const {
9702 SDValue Result(0, 0);
9704 // Currently only support length 1 constraints.
9705 if (Constraint.length() != 1) return;
9707 char ConstraintLetter = Constraint[0];
9708 switch (ConstraintLetter) {
9711 case 'I': case 'J': case 'K': case 'L':
9712 case 'M': case 'N': case 'O':
9713 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9717 int64_t CVal64 = C->getSExtValue();
9718 int CVal = (int) CVal64;
9719 // None of these constraints allow values larger than 32 bits. Check
9720 // that the value fits in an int.
9724 switch (ConstraintLetter) {
9726 // Constant suitable for movw, must be between 0 and
9728 if (Subtarget->hasV6T2Ops())
9729 if (CVal >= 0 && CVal <= 65535)
9733 if (Subtarget->isThumb1Only()) {
9734 // This must be a constant between 0 and 255, for ADD
9736 if (CVal >= 0 && CVal <= 255)
9738 } else if (Subtarget->isThumb2()) {
9739 // A constant that can be used as an immediate value in a
9740 // data-processing instruction.
9741 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9744 // A constant that can be used as an immediate value in a
9745 // data-processing instruction.
9746 if (ARM_AM::getSOImmVal(CVal) != -1)
9752 if (Subtarget->isThumb()) { // FIXME thumb2
9753 // This must be a constant between -255 and -1, for negated ADD
9754 // immediates. This can be used in GCC with an "n" modifier that
9755 // prints the negated value, for use with SUB instructions. It is
9756 // not useful otherwise but is implemented for compatibility.
9757 if (CVal >= -255 && CVal <= -1)
9760 // This must be a constant between -4095 and 4095. It is not clear
9761 // what this constraint is intended for. Implemented for
9762 // compatibility with GCC.
9763 if (CVal >= -4095 && CVal <= 4095)
9769 if (Subtarget->isThumb1Only()) {
9770 // A 32-bit value where only one byte has a nonzero value. Exclude
9771 // zero to match GCC. This constraint is used by GCC internally for
9772 // constants that can be loaded with a move/shift combination.
9773 // It is not useful otherwise but is implemented for compatibility.
9774 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9776 } else if (Subtarget->isThumb2()) {
9777 // A constant whose bitwise inverse can be used as an immediate
9778 // value in a data-processing instruction. This can be used in GCC
9779 // with a "B" modifier that prints the inverted value, for use with
9780 // BIC and MVN instructions. It is not useful otherwise but is
9781 // implemented for compatibility.
9782 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9785 // A constant whose bitwise inverse can be used as an immediate
9786 // value in a data-processing instruction. This can be used in GCC
9787 // with a "B" modifier that prints the inverted value, for use with
9788 // BIC and MVN instructions. It is not useful otherwise but is
9789 // implemented for compatibility.
9790 if (ARM_AM::getSOImmVal(~CVal) != -1)
9796 if (Subtarget->isThumb1Only()) {
9797 // This must be a constant between -7 and 7,
9798 // for 3-operand ADD/SUB immediate instructions.
9799 if (CVal >= -7 && CVal < 7)
9801 } else if (Subtarget->isThumb2()) {
9802 // A constant whose negation can be used as an immediate value in a
9803 // data-processing instruction. This can be used in GCC with an "n"
9804 // modifier that prints the negated value, for use with SUB
9805 // instructions. It is not useful otherwise but is implemented for
9807 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9810 // A constant whose negation can be used as an immediate value in a
9811 // data-processing instruction. This can be used in GCC with an "n"
9812 // modifier that prints the negated value, for use with SUB
9813 // instructions. It is not useful otherwise but is implemented for
9815 if (ARM_AM::getSOImmVal(-CVal) != -1)
9821 if (Subtarget->isThumb()) { // FIXME thumb2
9822 // This must be a multiple of 4 between 0 and 1020, for
9823 // ADD sp + immediate.
9824 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9827 // A power of two or a constant between 0 and 32. This is used in
9828 // GCC for the shift amount on shifted register operands, but it is
9829 // useful in general for any shift amounts.
9830 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9836 if (Subtarget->isThumb()) { // FIXME thumb2
9837 // This must be a constant between 0 and 31, for shift amounts.
9838 if (CVal >= 0 && CVal <= 31)
9844 if (Subtarget->isThumb()) { // FIXME thumb2
9845 // This must be a multiple of 4 between -508 and 508, for
9846 // ADD/SUB sp = sp + immediate.
9847 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9852 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9856 if (Result.getNode()) {
9857 Ops.push_back(Result);
9860 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9864 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9865 // The ARM target isn't yet aware of offsets.
9869 bool ARM::isBitFieldInvertedMask(unsigned v) {
9870 if (v == 0xffffffff)
9872 // there can be 1's on either or both "outsides", all the "inside"
9874 unsigned int lsb = 0, msb = 31;
9875 while (v & (1 << msb)) --msb;
9876 while (v & (1 << lsb)) ++lsb;
9877 for (unsigned int i = lsb; i <= msb; ++i) {
9884 /// isFPImmLegal - Returns true if the target can instruction select the
9885 /// specified FP immediate natively. If false, the legalizer will
9886 /// materialize the FP immediate as a load from a constant pool.
9887 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9888 if (!Subtarget->hasVFP3())
9891 return ARM_AM::getFP32Imm(Imm) != -1;
9893 return ARM_AM::getFP64Imm(Imm) != -1;
9897 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9898 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9899 /// specified in the intrinsic calls.
9900 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9902 unsigned Intrinsic) const {
9903 switch (Intrinsic) {
9904 case Intrinsic::arm_neon_vld1:
9905 case Intrinsic::arm_neon_vld2:
9906 case Intrinsic::arm_neon_vld3:
9907 case Intrinsic::arm_neon_vld4:
9908 case Intrinsic::arm_neon_vld2lane:
9909 case Intrinsic::arm_neon_vld3lane:
9910 case Intrinsic::arm_neon_vld4lane: {
9911 Info.opc = ISD::INTRINSIC_W_CHAIN;
9912 // Conservatively set memVT to the entire set of vectors loaded.
9913 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9914 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9915 Info.ptrVal = I.getArgOperand(0);
9917 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9918 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9919 Info.vol = false; // volatile loads with NEON intrinsics not supported
9920 Info.readMem = true;
9921 Info.writeMem = false;
9924 case Intrinsic::arm_neon_vst1:
9925 case Intrinsic::arm_neon_vst2:
9926 case Intrinsic::arm_neon_vst3:
9927 case Intrinsic::arm_neon_vst4:
9928 case Intrinsic::arm_neon_vst2lane:
9929 case Intrinsic::arm_neon_vst3lane:
9930 case Intrinsic::arm_neon_vst4lane: {
9931 Info.opc = ISD::INTRINSIC_VOID;
9932 // Conservatively set memVT to the entire set of vectors stored.
9933 unsigned NumElts = 0;
9934 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9935 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9936 if (!ArgTy->isVectorTy())
9938 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9940 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9941 Info.ptrVal = I.getArgOperand(0);
9943 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9944 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9945 Info.vol = false; // volatile stores with NEON intrinsics not supported
9946 Info.readMem = false;
9947 Info.writeMem = true;
9950 case Intrinsic::arm_strexd: {
9951 Info.opc = ISD::INTRINSIC_W_CHAIN;
9952 Info.memVT = MVT::i64;
9953 Info.ptrVal = I.getArgOperand(2);
9957 Info.readMem = false;
9958 Info.writeMem = true;
9961 case Intrinsic::arm_ldrexd: {
9962 Info.opc = ISD::INTRINSIC_W_CHAIN;
9963 Info.memVT = MVT::i64;
9964 Info.ptrVal = I.getArgOperand(0);
9968 Info.readMem = true;
9969 Info.writeMem = false;