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::FCEIL, MVT::v4f32, Expand);
519 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
520 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
521 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
524 // Neon does not support some operations on v1i64 and v2i64 types.
525 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
526 // Custom handling for some quad-vector types to detect VMULL.
527 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
528 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
529 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
530 // Custom handling for some vector types to avoid expensive expansions
531 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
532 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
533 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
534 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
535 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
536 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
537 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
538 // a destination type that is wider than the source, and nor does
539 // it have a FP_TO_[SU]INT instruction with a narrower destination than
541 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
542 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
543 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
544 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
546 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
547 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
549 setTargetDAGCombine(ISD::INTRINSIC_VOID);
550 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
551 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
552 setTargetDAGCombine(ISD::SHL);
553 setTargetDAGCombine(ISD::SRL);
554 setTargetDAGCombine(ISD::SRA);
555 setTargetDAGCombine(ISD::SIGN_EXTEND);
556 setTargetDAGCombine(ISD::ZERO_EXTEND);
557 setTargetDAGCombine(ISD::ANY_EXTEND);
558 setTargetDAGCombine(ISD::SELECT_CC);
559 setTargetDAGCombine(ISD::BUILD_VECTOR);
560 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
561 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
562 setTargetDAGCombine(ISD::STORE);
563 setTargetDAGCombine(ISD::FP_TO_SINT);
564 setTargetDAGCombine(ISD::FP_TO_UINT);
565 setTargetDAGCombine(ISD::FDIV);
567 // It is legal to extload from v4i8 to v4i16 or v4i32.
568 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
569 MVT::v4i16, MVT::v2i16,
571 for (unsigned i = 0; i < 6; ++i) {
572 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
573 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
574 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
578 // ARM and Thumb2 support UMLAL/SMLAL.
579 if (!Subtarget->isThumb1Only())
580 setTargetDAGCombine(ISD::ADDC);
583 computeRegisterProperties();
585 // ARM does not have f32 extending load.
586 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
588 // ARM does not have i1 sign extending load.
589 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
591 // ARM supports all 4 flavors of integer indexed load / store.
592 if (!Subtarget->isThumb1Only()) {
593 for (unsigned im = (unsigned)ISD::PRE_INC;
594 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
595 setIndexedLoadAction(im, MVT::i1, Legal);
596 setIndexedLoadAction(im, MVT::i8, Legal);
597 setIndexedLoadAction(im, MVT::i16, Legal);
598 setIndexedLoadAction(im, MVT::i32, Legal);
599 setIndexedStoreAction(im, MVT::i1, Legal);
600 setIndexedStoreAction(im, MVT::i8, Legal);
601 setIndexedStoreAction(im, MVT::i16, Legal);
602 setIndexedStoreAction(im, MVT::i32, Legal);
606 // i64 operation support.
607 setOperationAction(ISD::MUL, MVT::i64, Expand);
608 setOperationAction(ISD::MULHU, MVT::i32, Expand);
609 if (Subtarget->isThumb1Only()) {
610 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
611 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
613 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
614 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
615 setOperationAction(ISD::MULHS, MVT::i32, Expand);
617 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
618 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
619 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
620 setOperationAction(ISD::SRL, MVT::i64, Custom);
621 setOperationAction(ISD::SRA, MVT::i64, Custom);
623 if (!Subtarget->isThumb1Only()) {
624 // FIXME: We should do this for Thumb1 as well.
625 setOperationAction(ISD::ADDC, MVT::i32, Custom);
626 setOperationAction(ISD::ADDE, MVT::i32, Custom);
627 setOperationAction(ISD::SUBC, MVT::i32, Custom);
628 setOperationAction(ISD::SUBE, MVT::i32, Custom);
631 // ARM does not have ROTL.
632 setOperationAction(ISD::ROTL, MVT::i32, Expand);
633 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
634 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
635 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
636 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
638 // These just redirect to CTTZ and CTLZ on ARM.
639 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
640 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
642 // Only ARMv6 has BSWAP.
643 if (!Subtarget->hasV6Ops())
644 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
646 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
647 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
648 // These are expanded into libcalls if the cpu doesn't have HW divider.
649 setOperationAction(ISD::SDIV, MVT::i32, Expand);
650 setOperationAction(ISD::UDIV, MVT::i32, Expand);
652 setOperationAction(ISD::SREM, MVT::i32, Expand);
653 setOperationAction(ISD::UREM, MVT::i32, Expand);
654 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
655 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
657 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
658 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
659 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
660 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
661 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
663 setOperationAction(ISD::TRAP, MVT::Other, Legal);
665 // Use the default implementation.
666 setOperationAction(ISD::VASTART, MVT::Other, Custom);
667 setOperationAction(ISD::VAARG, MVT::Other, Expand);
668 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
669 setOperationAction(ISD::VAEND, MVT::Other, Expand);
670 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
671 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
673 if (!Subtarget->isTargetDarwin()) {
674 // Non-Darwin platforms may return values in these registers via the
675 // personality function.
676 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
677 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
678 setExceptionPointerRegister(ARM::R0);
679 setExceptionSelectorRegister(ARM::R1);
682 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
683 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
684 // the default expansion.
685 // FIXME: This should be checking for v6k, not just v6.
686 if (Subtarget->hasDataBarrier() ||
687 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
688 // membarrier needs custom lowering; the rest are legal and handled
690 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
691 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
692 // Custom lowering for 64-bit ops
693 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
694 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
695 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
696 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
697 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
698 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
699 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
700 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
701 setInsertFencesForAtomic(true);
703 // Set them all for expansion, which will force libcalls.
704 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
705 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
706 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
711 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
712 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
713 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
714 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
715 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
716 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
717 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
718 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
719 // Unordered/Monotonic case.
720 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
721 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
722 // Since the libcalls include locking, fold in the fences
723 setShouldFoldAtomicFences(true);
726 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
728 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
729 if (!Subtarget->hasV6Ops()) {
730 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
731 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
733 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
735 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
736 !Subtarget->isThumb1Only()) {
737 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
738 // iff target supports vfp2.
739 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
740 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
743 // We want to custom lower some of our intrinsics.
744 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
745 if (Subtarget->isTargetDarwin()) {
746 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
747 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
748 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
751 setOperationAction(ISD::SETCC, MVT::i32, Expand);
752 setOperationAction(ISD::SETCC, MVT::f32, Expand);
753 setOperationAction(ISD::SETCC, MVT::f64, Expand);
754 setOperationAction(ISD::SELECT, MVT::i32, Custom);
755 setOperationAction(ISD::SELECT, MVT::f32, Custom);
756 setOperationAction(ISD::SELECT, MVT::f64, Custom);
757 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
758 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
759 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
761 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
762 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
763 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
764 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
765 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
767 // We don't support sin/cos/fmod/copysign/pow
768 setOperationAction(ISD::FSIN, MVT::f64, Expand);
769 setOperationAction(ISD::FSIN, MVT::f32, Expand);
770 setOperationAction(ISD::FCOS, MVT::f32, Expand);
771 setOperationAction(ISD::FCOS, MVT::f64, Expand);
772 setOperationAction(ISD::FREM, MVT::f64, Expand);
773 setOperationAction(ISD::FREM, MVT::f32, Expand);
774 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
775 !Subtarget->isThumb1Only()) {
776 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
777 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
779 setOperationAction(ISD::FPOW, MVT::f64, Expand);
780 setOperationAction(ISD::FPOW, MVT::f32, Expand);
782 if (!Subtarget->hasVFP4()) {
783 setOperationAction(ISD::FMA, MVT::f64, Expand);
784 setOperationAction(ISD::FMA, MVT::f32, Expand);
787 // Various VFP goodness
788 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
789 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
790 if (Subtarget->hasVFP2()) {
791 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
792 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
793 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
794 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
796 // Special handling for half-precision FP.
797 if (!Subtarget->hasFP16()) {
798 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
799 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
803 // We have target-specific dag combine patterns for the following nodes:
804 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
805 setTargetDAGCombine(ISD::ADD);
806 setTargetDAGCombine(ISD::SUB);
807 setTargetDAGCombine(ISD::MUL);
808 setTargetDAGCombine(ISD::AND);
809 setTargetDAGCombine(ISD::OR);
810 setTargetDAGCombine(ISD::XOR);
812 if (Subtarget->hasV6Ops())
813 setTargetDAGCombine(ISD::SRL);
815 setStackPointerRegisterToSaveRestore(ARM::SP);
817 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
818 !Subtarget->hasVFP2())
819 setSchedulingPreference(Sched::RegPressure);
821 setSchedulingPreference(Sched::Hybrid);
823 //// temporary - rewrite interface to use type
824 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
825 maxStoresPerMemset = 16;
826 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
828 // On ARM arguments smaller than 4 bytes are extended, so all arguments
829 // are at least 4 bytes aligned.
830 setMinStackArgumentAlignment(4);
832 benefitFromCodePlacementOpt = true;
834 // Prefer likely predicted branches to selects on out-of-order cores.
835 predictableSelectIsExpensive = Subtarget->isLikeA9();
837 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
840 // FIXME: It might make sense to define the representative register class as the
841 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
842 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
843 // SPR's representative would be DPR_VFP2. This should work well if register
844 // pressure tracking were modified such that a register use would increment the
845 // pressure of the register class's representative and all of it's super
846 // classes' representatives transitively. We have not implemented this because
847 // of the difficulty prior to coalescing of modeling operand register classes
848 // due to the common occurrence of cross class copies and subregister insertions
850 std::pair<const TargetRegisterClass*, uint8_t>
851 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
852 const TargetRegisterClass *RRC = 0;
854 switch (VT.getSimpleVT().SimpleTy) {
856 return TargetLowering::findRepresentativeClass(VT);
857 // Use DPR as representative register class for all floating point
858 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
859 // the cost is 1 for both f32 and f64.
860 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
861 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
862 RRC = &ARM::DPRRegClass;
863 // When NEON is used for SP, only half of the register file is available
864 // because operations that define both SP and DP results will be constrained
865 // to the VFP2 class (D0-D15). We currently model this constraint prior to
866 // coalescing by double-counting the SP regs. See the FIXME above.
867 if (Subtarget->useNEONForSinglePrecisionFP())
870 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
871 case MVT::v4f32: case MVT::v2f64:
872 RRC = &ARM::DPRRegClass;
876 RRC = &ARM::DPRRegClass;
880 RRC = &ARM::DPRRegClass;
884 return std::make_pair(RRC, Cost);
887 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
890 case ARMISD::Wrapper: return "ARMISD::Wrapper";
891 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
892 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
893 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
894 case ARMISD::CALL: return "ARMISD::CALL";
895 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
896 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
897 case ARMISD::tCALL: return "ARMISD::tCALL";
898 case ARMISD::BRCOND: return "ARMISD::BRCOND";
899 case ARMISD::BR_JT: return "ARMISD::BR_JT";
900 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
901 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
902 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
903 case ARMISD::CMP: return "ARMISD::CMP";
904 case ARMISD::CMN: return "ARMISD::CMN";
905 case ARMISD::CMPZ: return "ARMISD::CMPZ";
906 case ARMISD::CMPFP: return "ARMISD::CMPFP";
907 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
908 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
909 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
911 case ARMISD::CMOV: return "ARMISD::CMOV";
913 case ARMISD::RBIT: return "ARMISD::RBIT";
915 case ARMISD::FTOSI: return "ARMISD::FTOSI";
916 case ARMISD::FTOUI: return "ARMISD::FTOUI";
917 case ARMISD::SITOF: return "ARMISD::SITOF";
918 case ARMISD::UITOF: return "ARMISD::UITOF";
920 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
921 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
922 case ARMISD::RRX: return "ARMISD::RRX";
924 case ARMISD::ADDC: return "ARMISD::ADDC";
925 case ARMISD::ADDE: return "ARMISD::ADDE";
926 case ARMISD::SUBC: return "ARMISD::SUBC";
927 case ARMISD::SUBE: return "ARMISD::SUBE";
929 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
930 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
932 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
933 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
935 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
937 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
939 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
941 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
942 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
944 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
946 case ARMISD::VCEQ: return "ARMISD::VCEQ";
947 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
948 case ARMISD::VCGE: return "ARMISD::VCGE";
949 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
950 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
951 case ARMISD::VCGEU: return "ARMISD::VCGEU";
952 case ARMISD::VCGT: return "ARMISD::VCGT";
953 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
954 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
955 case ARMISD::VCGTU: return "ARMISD::VCGTU";
956 case ARMISD::VTST: return "ARMISD::VTST";
958 case ARMISD::VSHL: return "ARMISD::VSHL";
959 case ARMISD::VSHRs: return "ARMISD::VSHRs";
960 case ARMISD::VSHRu: return "ARMISD::VSHRu";
961 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
962 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
963 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
964 case ARMISD::VSHRN: return "ARMISD::VSHRN";
965 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
966 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
967 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
968 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
969 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
970 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
971 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
972 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
973 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
974 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
975 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
976 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
977 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
978 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
979 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
980 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
981 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
982 case ARMISD::VDUP: return "ARMISD::VDUP";
983 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
984 case ARMISD::VEXT: return "ARMISD::VEXT";
985 case ARMISD::VREV64: return "ARMISD::VREV64";
986 case ARMISD::VREV32: return "ARMISD::VREV32";
987 case ARMISD::VREV16: return "ARMISD::VREV16";
988 case ARMISD::VZIP: return "ARMISD::VZIP";
989 case ARMISD::VUZP: return "ARMISD::VUZP";
990 case ARMISD::VTRN: return "ARMISD::VTRN";
991 case ARMISD::VTBL1: return "ARMISD::VTBL1";
992 case ARMISD::VTBL2: return "ARMISD::VTBL2";
993 case ARMISD::VMULLs: return "ARMISD::VMULLs";
994 case ARMISD::VMULLu: return "ARMISD::VMULLu";
995 case ARMISD::UMLAL: return "ARMISD::UMLAL";
996 case ARMISD::SMLAL: return "ARMISD::SMLAL";
997 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
998 case ARMISD::FMAX: return "ARMISD::FMAX";
999 case ARMISD::FMIN: return "ARMISD::FMIN";
1000 case ARMISD::BFI: return "ARMISD::BFI";
1001 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1002 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1003 case ARMISD::VBSL: return "ARMISD::VBSL";
1004 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1005 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1006 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1007 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1008 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1009 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1010 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1011 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1012 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1013 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1014 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1015 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1016 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1017 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1018 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1019 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1020 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1021 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1022 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1023 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1027 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1028 if (!VT.isVector()) return getPointerTy();
1029 return VT.changeVectorElementTypeToInteger();
1032 /// getRegClassFor - Return the register class that should be used for the
1033 /// specified value type.
1034 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1035 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1036 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1037 // load / store 4 to 8 consecutive D registers.
1038 if (Subtarget->hasNEON()) {
1039 if (VT == MVT::v4i64)
1040 return &ARM::QQPRRegClass;
1041 if (VT == MVT::v8i64)
1042 return &ARM::QQQQPRRegClass;
1044 return TargetLowering::getRegClassFor(VT);
1047 // Create a fast isel object.
1049 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1050 const TargetLibraryInfo *libInfo) const {
1051 return ARM::createFastISel(funcInfo, libInfo);
1054 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1055 /// be used for loads / stores from the global.
1056 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1057 return (Subtarget->isThumb1Only() ? 127 : 4095);
1060 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1061 unsigned NumVals = N->getNumValues();
1063 return Sched::RegPressure;
1065 for (unsigned i = 0; i != NumVals; ++i) {
1066 EVT VT = N->getValueType(i);
1067 if (VT == MVT::Glue || VT == MVT::Other)
1069 if (VT.isFloatingPoint() || VT.isVector())
1073 if (!N->isMachineOpcode())
1074 return Sched::RegPressure;
1076 // Load are scheduled for latency even if there instruction itinerary
1077 // is not available.
1078 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1079 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1081 if (MCID.getNumDefs() == 0)
1082 return Sched::RegPressure;
1083 if (!Itins->isEmpty() &&
1084 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1087 return Sched::RegPressure;
1090 //===----------------------------------------------------------------------===//
1092 //===----------------------------------------------------------------------===//
1094 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1095 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1097 default: llvm_unreachable("Unknown condition code!");
1098 case ISD::SETNE: return ARMCC::NE;
1099 case ISD::SETEQ: return ARMCC::EQ;
1100 case ISD::SETGT: return ARMCC::GT;
1101 case ISD::SETGE: return ARMCC::GE;
1102 case ISD::SETLT: return ARMCC::LT;
1103 case ISD::SETLE: return ARMCC::LE;
1104 case ISD::SETUGT: return ARMCC::HI;
1105 case ISD::SETUGE: return ARMCC::HS;
1106 case ISD::SETULT: return ARMCC::LO;
1107 case ISD::SETULE: return ARMCC::LS;
1111 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1112 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1113 ARMCC::CondCodes &CondCode2) {
1114 CondCode2 = ARMCC::AL;
1116 default: llvm_unreachable("Unknown FP condition!");
1118 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1120 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1122 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1123 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1124 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1125 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1126 case ISD::SETO: CondCode = ARMCC::VC; break;
1127 case ISD::SETUO: CondCode = ARMCC::VS; break;
1128 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1129 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1130 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1132 case ISD::SETULT: CondCode = ARMCC::LT; break;
1134 case ISD::SETULE: CondCode = ARMCC::LE; break;
1136 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1140 //===----------------------------------------------------------------------===//
1141 // Calling Convention Implementation
1142 //===----------------------------------------------------------------------===//
1144 #include "ARMGenCallingConv.inc"
1146 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1147 /// given CallingConvention value.
1148 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1150 bool isVarArg) const {
1153 llvm_unreachable("Unsupported calling convention");
1154 case CallingConv::Fast:
1155 if (Subtarget->hasVFP2() && !isVarArg) {
1156 if (!Subtarget->isAAPCS_ABI())
1157 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1158 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1159 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1162 case CallingConv::C: {
1163 // Use target triple & subtarget features to do actual dispatch.
1164 if (!Subtarget->isAAPCS_ABI())
1165 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1166 else if (Subtarget->hasVFP2() &&
1167 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1169 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1172 case CallingConv::ARM_AAPCS_VFP:
1174 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1176 case CallingConv::ARM_AAPCS:
1177 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1178 case CallingConv::ARM_APCS:
1179 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1180 case CallingConv::GHC:
1181 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1185 /// LowerCallResult - Lower the result values of a call into the
1186 /// appropriate copies out of appropriate physical registers.
1188 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1189 CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<ISD::InputArg> &Ins,
1191 DebugLoc dl, SelectionDAG &DAG,
1192 SmallVectorImpl<SDValue> &InVals) const {
1194 // Assign locations to each value returned by this call.
1195 SmallVector<CCValAssign, 16> RVLocs;
1196 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1197 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1198 CCInfo.AnalyzeCallResult(Ins,
1199 CCAssignFnForNode(CallConv, /* Return*/ true,
1202 // Copy all of the result registers out of their specified physreg.
1203 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1204 CCValAssign VA = RVLocs[i];
1207 if (VA.needsCustom()) {
1208 // Handle f64 or half of a v2f64.
1209 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1211 Chain = Lo.getValue(1);
1212 InFlag = Lo.getValue(2);
1213 VA = RVLocs[++i]; // skip ahead to next loc
1214 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1216 Chain = Hi.getValue(1);
1217 InFlag = Hi.getValue(2);
1218 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1220 if (VA.getLocVT() == MVT::v2f64) {
1221 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1222 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1223 DAG.getConstant(0, MVT::i32));
1225 VA = RVLocs[++i]; // skip ahead to next loc
1226 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1227 Chain = Lo.getValue(1);
1228 InFlag = Lo.getValue(2);
1229 VA = RVLocs[++i]; // skip ahead to next loc
1230 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1231 Chain = Hi.getValue(1);
1232 InFlag = Hi.getValue(2);
1233 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1234 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1235 DAG.getConstant(1, MVT::i32));
1238 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1240 Chain = Val.getValue(1);
1241 InFlag = Val.getValue(2);
1244 switch (VA.getLocInfo()) {
1245 default: llvm_unreachable("Unknown loc info!");
1246 case CCValAssign::Full: break;
1247 case CCValAssign::BCvt:
1248 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1252 InVals.push_back(Val);
1258 /// LowerMemOpCallTo - Store the argument to the stack.
1260 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1261 SDValue StackPtr, SDValue Arg,
1262 DebugLoc dl, SelectionDAG &DAG,
1263 const CCValAssign &VA,
1264 ISD::ArgFlagsTy Flags) const {
1265 unsigned LocMemOffset = VA.getLocMemOffset();
1266 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1267 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1268 return DAG.getStore(Chain, dl, Arg, PtrOff,
1269 MachinePointerInfo::getStack(LocMemOffset),
1273 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1274 SDValue Chain, SDValue &Arg,
1275 RegsToPassVector &RegsToPass,
1276 CCValAssign &VA, CCValAssign &NextVA,
1278 SmallVector<SDValue, 8> &MemOpChains,
1279 ISD::ArgFlagsTy Flags) const {
1281 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1282 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1283 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1285 if (NextVA.isRegLoc())
1286 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1288 assert(NextVA.isMemLoc());
1289 if (StackPtr.getNode() == 0)
1290 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1292 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1298 /// LowerCall - Lowering a call into a callseq_start <-
1299 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1302 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1303 SmallVectorImpl<SDValue> &InVals) const {
1304 SelectionDAG &DAG = CLI.DAG;
1305 DebugLoc &dl = CLI.DL;
1306 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1307 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1308 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1309 SDValue Chain = CLI.Chain;
1310 SDValue Callee = CLI.Callee;
1311 bool &isTailCall = CLI.IsTailCall;
1312 CallingConv::ID CallConv = CLI.CallConv;
1313 bool doesNotRet = CLI.DoesNotReturn;
1314 bool isVarArg = CLI.IsVarArg;
1316 MachineFunction &MF = DAG.getMachineFunction();
1317 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1318 bool IsSibCall = false;
1319 // Disable tail calls if they're not supported.
1320 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1323 // Check if it's really possible to do a tail call.
1324 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1325 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1326 Outs, OutVals, Ins, DAG);
1327 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1328 // detected sibcalls.
1335 // Analyze operands of the call, assigning locations to each operand.
1336 SmallVector<CCValAssign, 16> ArgLocs;
1337 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1338 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1339 CCInfo.AnalyzeCallOperands(Outs,
1340 CCAssignFnForNode(CallConv, /* Return*/ false,
1343 // Get a count of how many bytes are to be pushed on the stack.
1344 unsigned NumBytes = CCInfo.getNextStackOffset();
1346 // For tail calls, memory operands are available in our caller's stack.
1350 // Adjust the stack pointer for the new arguments...
1351 // These operations are automatically eliminated by the prolog/epilog pass
1353 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1355 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1357 RegsToPassVector RegsToPass;
1358 SmallVector<SDValue, 8> MemOpChains;
1360 // Walk the register/memloc assignments, inserting copies/loads. In the case
1361 // of tail call optimization, arguments are handled later.
1362 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1364 ++i, ++realArgIdx) {
1365 CCValAssign &VA = ArgLocs[i];
1366 SDValue Arg = OutVals[realArgIdx];
1367 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1368 bool isByVal = Flags.isByVal();
1370 // Promote the value if needed.
1371 switch (VA.getLocInfo()) {
1372 default: llvm_unreachable("Unknown loc info!");
1373 case CCValAssign::Full: break;
1374 case CCValAssign::SExt:
1375 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1377 case CCValAssign::ZExt:
1378 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1380 case CCValAssign::AExt:
1381 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1383 case CCValAssign::BCvt:
1384 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1388 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1389 if (VA.needsCustom()) {
1390 if (VA.getLocVT() == MVT::v2f64) {
1391 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1392 DAG.getConstant(0, MVT::i32));
1393 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1394 DAG.getConstant(1, MVT::i32));
1396 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1397 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1399 VA = ArgLocs[++i]; // skip ahead to next loc
1400 if (VA.isRegLoc()) {
1401 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1402 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1404 assert(VA.isMemLoc());
1406 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1407 dl, DAG, VA, Flags));
1410 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1411 StackPtr, MemOpChains, Flags);
1413 } else if (VA.isRegLoc()) {
1414 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1415 } else if (isByVal) {
1416 assert(VA.isMemLoc());
1417 unsigned offset = 0;
1419 // True if this byval aggregate will be split between registers
1421 if (CCInfo.isFirstByValRegValid()) {
1422 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1424 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1425 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1426 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1427 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1428 MachinePointerInfo(),
1429 false, false, false, 0);
1430 MemOpChains.push_back(Load.getValue(1));
1431 RegsToPass.push_back(std::make_pair(j, Load));
1433 offset = ARM::R4 - CCInfo.getFirstByValReg();
1434 CCInfo.clearFirstByValReg();
1437 if (Flags.getByValSize() - 4*offset > 0) {
1438 unsigned LocMemOffset = VA.getLocMemOffset();
1439 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1440 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1442 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1443 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1444 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1446 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1448 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1449 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1450 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1451 Ops, array_lengthof(Ops)));
1453 } else if (!IsSibCall) {
1454 assert(VA.isMemLoc());
1456 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1457 dl, DAG, VA, Flags));
1461 if (!MemOpChains.empty())
1462 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1463 &MemOpChains[0], MemOpChains.size());
1465 // Build a sequence of copy-to-reg nodes chained together with token chain
1466 // and flag operands which copy the outgoing args into the appropriate regs.
1468 // Tail call byval lowering might overwrite argument registers so in case of
1469 // tail call optimization the copies to registers are lowered later.
1471 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1472 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1473 RegsToPass[i].second, InFlag);
1474 InFlag = Chain.getValue(1);
1477 // For tail calls lower the arguments to the 'real' stack slot.
1479 // Force all the incoming stack arguments to be loaded from the stack
1480 // before any new outgoing arguments are stored to the stack, because the
1481 // outgoing stack slots may alias the incoming argument stack slots, and
1482 // the alias isn't otherwise explicit. This is slightly more conservative
1483 // than necessary, because it means that each store effectively depends
1484 // on every argument instead of just those arguments it would clobber.
1486 // Do not flag preceding copytoreg stuff together with the following stuff.
1488 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1489 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1490 RegsToPass[i].second, InFlag);
1491 InFlag = Chain.getValue(1);
1496 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1497 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1498 // node so that legalize doesn't hack it.
1499 bool isDirect = false;
1500 bool isARMFunc = false;
1501 bool isLocalARMFunc = false;
1502 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1504 if (EnableARMLongCalls) {
1505 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1506 && "long-calls with non-static relocation model!");
1507 // Handle a global address or an external symbol. If it's not one of
1508 // those, the target's already in a register, so we don't need to do
1510 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1511 const GlobalValue *GV = G->getGlobal();
1512 // Create a constant pool entry for the callee address
1513 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1514 ARMConstantPoolValue *CPV =
1515 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1517 // Get the address of the callee into a register
1518 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1519 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1520 Callee = DAG.getLoad(getPointerTy(), dl,
1521 DAG.getEntryNode(), CPAddr,
1522 MachinePointerInfo::getConstantPool(),
1523 false, false, false, 0);
1524 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1525 const char *Sym = S->getSymbol();
1527 // Create a constant pool entry for the callee address
1528 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1529 ARMConstantPoolValue *CPV =
1530 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1531 ARMPCLabelIndex, 0);
1532 // Get the address of the callee into a register
1533 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1534 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1535 Callee = DAG.getLoad(getPointerTy(), dl,
1536 DAG.getEntryNode(), CPAddr,
1537 MachinePointerInfo::getConstantPool(),
1538 false, false, false, 0);
1540 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1541 const GlobalValue *GV = G->getGlobal();
1543 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1544 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1545 getTargetMachine().getRelocationModel() != Reloc::Static;
1546 isARMFunc = !Subtarget->isThumb() || isStub;
1547 // ARM call to a local ARM function is predicable.
1548 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1549 // tBX takes a register source operand.
1550 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1551 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1552 ARMConstantPoolValue *CPV =
1553 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1554 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1555 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1556 Callee = DAG.getLoad(getPointerTy(), dl,
1557 DAG.getEntryNode(), CPAddr,
1558 MachinePointerInfo::getConstantPool(),
1559 false, false, false, 0);
1560 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1561 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1562 getPointerTy(), Callee, PICLabel);
1564 // On ELF targets for PIC code, direct calls should go through the PLT
1565 unsigned OpFlags = 0;
1566 if (Subtarget->isTargetELF() &&
1567 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1568 OpFlags = ARMII::MO_PLT;
1569 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1571 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1573 bool isStub = Subtarget->isTargetDarwin() &&
1574 getTargetMachine().getRelocationModel() != Reloc::Static;
1575 isARMFunc = !Subtarget->isThumb() || isStub;
1576 // tBX takes a register source operand.
1577 const char *Sym = S->getSymbol();
1578 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1579 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1580 ARMConstantPoolValue *CPV =
1581 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1582 ARMPCLabelIndex, 4);
1583 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1584 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1585 Callee = DAG.getLoad(getPointerTy(), dl,
1586 DAG.getEntryNode(), CPAddr,
1587 MachinePointerInfo::getConstantPool(),
1588 false, false, false, 0);
1589 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1590 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1591 getPointerTy(), Callee, PICLabel);
1593 unsigned OpFlags = 0;
1594 // On ELF targets for PIC code, direct calls should go through the PLT
1595 if (Subtarget->isTargetELF() &&
1596 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1597 OpFlags = ARMII::MO_PLT;
1598 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1602 // FIXME: handle tail calls differently.
1604 bool HasMinSizeAttr = MF.getFunction()->getFnAttributes().
1605 hasAttribute(Attributes::MinSize);
1606 if (Subtarget->isThumb()) {
1607 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1608 CallOpc = ARMISD::CALL_NOLINK;
1610 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1612 if (!isDirect && !Subtarget->hasV5TOps())
1613 CallOpc = ARMISD::CALL_NOLINK;
1614 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1615 // Emit regular call when code size is the priority
1617 // "mov lr, pc; b _foo" to avoid confusing the RSP
1618 CallOpc = ARMISD::CALL_NOLINK;
1620 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1623 std::vector<SDValue> Ops;
1624 Ops.push_back(Chain);
1625 Ops.push_back(Callee);
1627 // Add argument registers to the end of the list so that they are known live
1629 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1630 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1631 RegsToPass[i].second.getValueType()));
1633 // Add a register mask operand representing the call-preserved registers.
1634 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1635 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1636 assert(Mask && "Missing call preserved mask for calling convention");
1637 Ops.push_back(DAG.getRegisterMask(Mask));
1639 if (InFlag.getNode())
1640 Ops.push_back(InFlag);
1642 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1644 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1646 // Returns a chain and a flag for retval copy to use.
1647 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1648 InFlag = Chain.getValue(1);
1650 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1651 DAG.getIntPtrConstant(0, true), InFlag);
1653 InFlag = Chain.getValue(1);
1655 // Handle result values, copying them out of physregs into vregs that we
1657 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1661 /// HandleByVal - Every parameter *after* a byval parameter is passed
1662 /// on the stack. Remember the next parameter register to allocate,
1663 /// and then confiscate the rest of the parameter registers to insure
1666 ARMTargetLowering::HandleByVal(
1667 CCState *State, unsigned &size, unsigned Align) const {
1668 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1669 assert((State->getCallOrPrologue() == Prologue ||
1670 State->getCallOrPrologue() == Call) &&
1671 "unhandled ParmContext");
1672 if ((!State->isFirstByValRegValid()) &&
1673 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1674 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1675 unsigned AlignInRegs = Align / 4;
1676 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1677 for (unsigned i = 0; i < Waste; ++i)
1678 reg = State->AllocateReg(GPRArgRegs, 4);
1681 State->setFirstByValReg(reg);
1682 // At a call site, a byval parameter that is split between
1683 // registers and memory needs its size truncated here. In a
1684 // function prologue, such byval parameters are reassembled in
1685 // memory, and are not truncated.
1686 if (State->getCallOrPrologue() == Call) {
1687 unsigned excess = 4 * (ARM::R4 - reg);
1688 assert(size >= excess && "expected larger existing stack allocation");
1693 // Confiscate any remaining parameter registers to preclude their
1694 // assignment to subsequent parameters.
1695 while (State->AllocateReg(GPRArgRegs, 4))
1699 /// MatchingStackOffset - Return true if the given stack call argument is
1700 /// already available in the same position (relatively) of the caller's
1701 /// incoming argument stack.
1703 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1704 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1705 const TargetInstrInfo *TII) {
1706 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1708 if (Arg.getOpcode() == ISD::CopyFromReg) {
1709 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1710 if (!TargetRegisterInfo::isVirtualRegister(VR))
1712 MachineInstr *Def = MRI->getVRegDef(VR);
1715 if (!Flags.isByVal()) {
1716 if (!TII->isLoadFromStackSlot(Def, FI))
1721 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1722 if (Flags.isByVal())
1723 // ByVal argument is passed in as a pointer but it's now being
1724 // dereferenced. e.g.
1725 // define @foo(%struct.X* %A) {
1726 // tail call @bar(%struct.X* byval %A)
1729 SDValue Ptr = Ld->getBasePtr();
1730 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1733 FI = FINode->getIndex();
1737 assert(FI != INT_MAX);
1738 if (!MFI->isFixedObjectIndex(FI))
1740 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1743 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1744 /// for tail call optimization. Targets which want to do tail call
1745 /// optimization should implement this function.
1747 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1748 CallingConv::ID CalleeCC,
1750 bool isCalleeStructRet,
1751 bool isCallerStructRet,
1752 const SmallVectorImpl<ISD::OutputArg> &Outs,
1753 const SmallVectorImpl<SDValue> &OutVals,
1754 const SmallVectorImpl<ISD::InputArg> &Ins,
1755 SelectionDAG& DAG) const {
1756 const Function *CallerF = DAG.getMachineFunction().getFunction();
1757 CallingConv::ID CallerCC = CallerF->getCallingConv();
1758 bool CCMatch = CallerCC == CalleeCC;
1760 // Look for obvious safe cases to perform tail call optimization that do not
1761 // require ABI changes. This is what gcc calls sibcall.
1763 // Do not sibcall optimize vararg calls unless the call site is not passing
1765 if (isVarArg && !Outs.empty())
1768 // Also avoid sibcall optimization if either caller or callee uses struct
1769 // return semantics.
1770 if (isCalleeStructRet || isCallerStructRet)
1773 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1774 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1775 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1776 // support in the assembler and linker to be used. This would need to be
1777 // fixed to fully support tail calls in Thumb1.
1779 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1780 // LR. This means if we need to reload LR, it takes an extra instructions,
1781 // which outweighs the value of the tail call; but here we don't know yet
1782 // whether LR is going to be used. Probably the right approach is to
1783 // generate the tail call here and turn it back into CALL/RET in
1784 // emitEpilogue if LR is used.
1786 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1787 // but we need to make sure there are enough registers; the only valid
1788 // registers are the 4 used for parameters. We don't currently do this
1790 if (Subtarget->isThumb1Only())
1793 // If the calling conventions do not match, then we'd better make sure the
1794 // results are returned in the same way as what the caller expects.
1796 SmallVector<CCValAssign, 16> RVLocs1;
1797 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1798 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1799 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1801 SmallVector<CCValAssign, 16> RVLocs2;
1802 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1803 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1804 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1806 if (RVLocs1.size() != RVLocs2.size())
1808 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1809 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1811 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1813 if (RVLocs1[i].isRegLoc()) {
1814 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1817 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1823 // If Caller's vararg or byval argument has been split between registers and
1824 // stack, do not perform tail call, since part of the argument is in caller's
1826 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1827 getInfo<ARMFunctionInfo>();
1828 if (AFI_Caller->getVarArgsRegSaveSize())
1831 // If the callee takes no arguments then go on to check the results of the
1833 if (!Outs.empty()) {
1834 // Check if stack adjustment is needed. For now, do not do this if any
1835 // argument is passed on the stack.
1836 SmallVector<CCValAssign, 16> ArgLocs;
1837 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1838 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1839 CCInfo.AnalyzeCallOperands(Outs,
1840 CCAssignFnForNode(CalleeCC, false, isVarArg));
1841 if (CCInfo.getNextStackOffset()) {
1842 MachineFunction &MF = DAG.getMachineFunction();
1844 // Check if the arguments are already laid out in the right way as
1845 // the caller's fixed stack objects.
1846 MachineFrameInfo *MFI = MF.getFrameInfo();
1847 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1848 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1849 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1851 ++i, ++realArgIdx) {
1852 CCValAssign &VA = ArgLocs[i];
1853 EVT RegVT = VA.getLocVT();
1854 SDValue Arg = OutVals[realArgIdx];
1855 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1856 if (VA.getLocInfo() == CCValAssign::Indirect)
1858 if (VA.needsCustom()) {
1859 // f64 and vector types are split into multiple registers or
1860 // register/stack-slot combinations. The types will not match
1861 // the registers; give up on memory f64 refs until we figure
1862 // out what to do about this.
1865 if (!ArgLocs[++i].isRegLoc())
1867 if (RegVT == MVT::v2f64) {
1868 if (!ArgLocs[++i].isRegLoc())
1870 if (!ArgLocs[++i].isRegLoc())
1873 } else if (!VA.isRegLoc()) {
1874 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1886 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1887 MachineFunction &MF, bool isVarArg,
1888 const SmallVectorImpl<ISD::OutputArg> &Outs,
1889 LLVMContext &Context) const {
1890 SmallVector<CCValAssign, 16> RVLocs;
1891 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
1892 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
1897 ARMTargetLowering::LowerReturn(SDValue Chain,
1898 CallingConv::ID CallConv, bool isVarArg,
1899 const SmallVectorImpl<ISD::OutputArg> &Outs,
1900 const SmallVectorImpl<SDValue> &OutVals,
1901 DebugLoc dl, SelectionDAG &DAG) const {
1903 // CCValAssign - represent the assignment of the return value to a location.
1904 SmallVector<CCValAssign, 16> RVLocs;
1906 // CCState - Info about the registers and stack slots.
1907 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1908 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1910 // Analyze outgoing return values.
1911 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1914 // If this is the first return lowered for this function, add
1915 // the regs to the liveout set for the function.
1916 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1917 for (unsigned i = 0; i != RVLocs.size(); ++i)
1918 if (RVLocs[i].isRegLoc())
1919 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1924 // Copy the result values into the output registers.
1925 for (unsigned i = 0, realRVLocIdx = 0;
1927 ++i, ++realRVLocIdx) {
1928 CCValAssign &VA = RVLocs[i];
1929 assert(VA.isRegLoc() && "Can only return in registers!");
1931 SDValue Arg = OutVals[realRVLocIdx];
1933 switch (VA.getLocInfo()) {
1934 default: llvm_unreachable("Unknown loc info!");
1935 case CCValAssign::Full: break;
1936 case CCValAssign::BCvt:
1937 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1941 if (VA.needsCustom()) {
1942 if (VA.getLocVT() == MVT::v2f64) {
1943 // Extract the first half and return it in two registers.
1944 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1945 DAG.getConstant(0, MVT::i32));
1946 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1947 DAG.getVTList(MVT::i32, MVT::i32), Half);
1949 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1950 Flag = Chain.getValue(1);
1951 VA = RVLocs[++i]; // skip ahead to next loc
1952 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1953 HalfGPRs.getValue(1), Flag);
1954 Flag = Chain.getValue(1);
1955 VA = RVLocs[++i]; // skip ahead to next loc
1957 // Extract the 2nd half and fall through to handle it as an f64 value.
1958 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1959 DAG.getConstant(1, MVT::i32));
1961 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1963 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1964 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1965 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1966 Flag = Chain.getValue(1);
1967 VA = RVLocs[++i]; // skip ahead to next loc
1968 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1971 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1973 // Guarantee that all emitted copies are
1974 // stuck together, avoiding something bad.
1975 Flag = Chain.getValue(1);
1980 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1982 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1987 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1988 if (N->getNumValues() != 1)
1990 if (!N->hasNUsesOfValue(1, 0))
1993 SDValue TCChain = Chain;
1994 SDNode *Copy = *N->use_begin();
1995 if (Copy->getOpcode() == ISD::CopyToReg) {
1996 // If the copy has a glue operand, we conservatively assume it isn't safe to
1997 // perform a tail call.
1998 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2000 TCChain = Copy->getOperand(0);
2001 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2002 SDNode *VMov = Copy;
2003 // f64 returned in a pair of GPRs.
2004 SmallPtrSet<SDNode*, 2> Copies;
2005 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2007 if (UI->getOpcode() != ISD::CopyToReg)
2011 if (Copies.size() > 2)
2014 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2016 SDValue UseChain = UI->getOperand(0);
2017 if (Copies.count(UseChain.getNode()))
2024 } else if (Copy->getOpcode() == ISD::BITCAST) {
2025 // f32 returned in a single GPR.
2026 if (!Copy->hasOneUse())
2028 Copy = *Copy->use_begin();
2029 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2031 Chain = Copy->getOperand(0);
2036 bool HasRet = false;
2037 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2039 if (UI->getOpcode() != ARMISD::RET_FLAG)
2051 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2052 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2055 if (!CI->isTailCall())
2058 return !Subtarget->isThumb1Only();
2061 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2062 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2063 // one of the above mentioned nodes. It has to be wrapped because otherwise
2064 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2065 // be used to form addressing mode. These wrapped nodes will be selected
2067 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2068 EVT PtrVT = Op.getValueType();
2069 // FIXME there is no actual debug info here
2070 DebugLoc dl = Op.getDebugLoc();
2071 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2073 if (CP->isMachineConstantPoolEntry())
2074 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2075 CP->getAlignment());
2077 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2078 CP->getAlignment());
2079 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2082 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2083 return MachineJumpTableInfo::EK_Inline;
2086 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2087 SelectionDAG &DAG) const {
2088 MachineFunction &MF = DAG.getMachineFunction();
2089 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2090 unsigned ARMPCLabelIndex = 0;
2091 DebugLoc DL = Op.getDebugLoc();
2092 EVT PtrVT = getPointerTy();
2093 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2094 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2096 if (RelocM == Reloc::Static) {
2097 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2099 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2100 ARMPCLabelIndex = AFI->createPICLabelUId();
2101 ARMConstantPoolValue *CPV =
2102 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2103 ARMCP::CPBlockAddress, PCAdj);
2104 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2106 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2107 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2108 MachinePointerInfo::getConstantPool(),
2109 false, false, false, 0);
2110 if (RelocM == Reloc::Static)
2112 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2113 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2116 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2118 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2119 SelectionDAG &DAG) const {
2120 DebugLoc dl = GA->getDebugLoc();
2121 EVT PtrVT = getPointerTy();
2122 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2123 MachineFunction &MF = DAG.getMachineFunction();
2124 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2125 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2126 ARMConstantPoolValue *CPV =
2127 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2128 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2129 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2130 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2131 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2132 MachinePointerInfo::getConstantPool(),
2133 false, false, false, 0);
2134 SDValue Chain = Argument.getValue(1);
2136 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2137 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2139 // call __tls_get_addr.
2142 Entry.Node = Argument;
2143 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2144 Args.push_back(Entry);
2145 // FIXME: is there useful debug info available here?
2146 TargetLowering::CallLoweringInfo CLI(Chain,
2147 (Type *) Type::getInt32Ty(*DAG.getContext()),
2148 false, false, false, false,
2149 0, CallingConv::C, /*isTailCall=*/false,
2150 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2151 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2152 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2153 return CallResult.first;
2156 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2157 // "local exec" model.
2159 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2161 TLSModel::Model model) const {
2162 const GlobalValue *GV = GA->getGlobal();
2163 DebugLoc dl = GA->getDebugLoc();
2165 SDValue Chain = DAG.getEntryNode();
2166 EVT PtrVT = getPointerTy();
2167 // Get the Thread Pointer
2168 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2170 if (model == TLSModel::InitialExec) {
2171 MachineFunction &MF = DAG.getMachineFunction();
2172 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2173 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2174 // Initial exec model.
2175 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2176 ARMConstantPoolValue *CPV =
2177 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2178 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2180 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2181 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2182 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2183 MachinePointerInfo::getConstantPool(),
2184 false, false, false, 0);
2185 Chain = Offset.getValue(1);
2187 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2188 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2190 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2191 MachinePointerInfo::getConstantPool(),
2192 false, false, false, 0);
2195 assert(model == TLSModel::LocalExec);
2196 ARMConstantPoolValue *CPV =
2197 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2198 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2199 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2200 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2201 MachinePointerInfo::getConstantPool(),
2202 false, false, false, 0);
2205 // The address of the thread local variable is the add of the thread
2206 // pointer with the offset of the variable.
2207 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2211 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2212 // TODO: implement the "local dynamic" model
2213 assert(Subtarget->isTargetELF() &&
2214 "TLS not implemented for non-ELF targets");
2215 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2217 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2220 case TLSModel::GeneralDynamic:
2221 case TLSModel::LocalDynamic:
2222 return LowerToTLSGeneralDynamicModel(GA, DAG);
2223 case TLSModel::InitialExec:
2224 case TLSModel::LocalExec:
2225 return LowerToTLSExecModels(GA, DAG, model);
2227 llvm_unreachable("bogus TLS model");
2230 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2231 SelectionDAG &DAG) const {
2232 EVT PtrVT = getPointerTy();
2233 DebugLoc dl = Op.getDebugLoc();
2234 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2235 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2236 if (RelocM == Reloc::PIC_) {
2237 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2238 ARMConstantPoolValue *CPV =
2239 ARMConstantPoolConstant::Create(GV,
2240 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2241 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2242 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2243 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2245 MachinePointerInfo::getConstantPool(),
2246 false, false, false, 0);
2247 SDValue Chain = Result.getValue(1);
2248 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2249 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2251 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2252 MachinePointerInfo::getGOT(),
2253 false, false, false, 0);
2257 // If we have T2 ops, we can materialize the address directly via movt/movw
2258 // pair. This is always cheaper.
2259 if (Subtarget->useMovt()) {
2261 // FIXME: Once remat is capable of dealing with instructions with register
2262 // operands, expand this into two nodes.
2263 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2264 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2266 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2267 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2268 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2269 MachinePointerInfo::getConstantPool(),
2270 false, false, false, 0);
2274 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2275 SelectionDAG &DAG) const {
2276 EVT PtrVT = getPointerTy();
2277 DebugLoc dl = Op.getDebugLoc();
2278 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2279 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2280 MachineFunction &MF = DAG.getMachineFunction();
2281 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2283 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2284 // update ARMFastISel::ARMMaterializeGV.
2285 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2287 // FIXME: Once remat is capable of dealing with instructions with register
2288 // operands, expand this into two nodes.
2289 if (RelocM == Reloc::Static)
2290 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2291 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2293 unsigned Wrapper = (RelocM == Reloc::PIC_)
2294 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2295 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2296 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2297 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2298 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2299 MachinePointerInfo::getGOT(),
2300 false, false, false, 0);
2304 unsigned ARMPCLabelIndex = 0;
2306 if (RelocM == Reloc::Static) {
2307 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2309 ARMPCLabelIndex = AFI->createPICLabelUId();
2310 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2311 ARMConstantPoolValue *CPV =
2312 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2314 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2316 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2318 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2319 MachinePointerInfo::getConstantPool(),
2320 false, false, false, 0);
2321 SDValue Chain = Result.getValue(1);
2323 if (RelocM == Reloc::PIC_) {
2324 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2325 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2328 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2329 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2330 false, false, false, 0);
2335 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2336 SelectionDAG &DAG) const {
2337 assert(Subtarget->isTargetELF() &&
2338 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2339 MachineFunction &MF = DAG.getMachineFunction();
2340 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2341 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2342 EVT PtrVT = getPointerTy();
2343 DebugLoc dl = Op.getDebugLoc();
2344 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2345 ARMConstantPoolValue *CPV =
2346 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2347 ARMPCLabelIndex, PCAdj);
2348 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2349 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2350 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2351 MachinePointerInfo::getConstantPool(),
2352 false, false, false, 0);
2353 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2354 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2358 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2359 DebugLoc dl = Op.getDebugLoc();
2360 SDValue Val = DAG.getConstant(0, MVT::i32);
2361 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2362 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2363 Op.getOperand(1), Val);
2367 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2368 DebugLoc dl = Op.getDebugLoc();
2369 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2370 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2374 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2375 const ARMSubtarget *Subtarget) const {
2376 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2377 DebugLoc dl = Op.getDebugLoc();
2379 default: return SDValue(); // Don't custom lower most intrinsics.
2380 case Intrinsic::arm_thread_pointer: {
2381 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2382 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2384 case Intrinsic::eh_sjlj_lsda: {
2385 MachineFunction &MF = DAG.getMachineFunction();
2386 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2387 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2388 EVT PtrVT = getPointerTy();
2389 DebugLoc dl = Op.getDebugLoc();
2390 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2392 unsigned PCAdj = (RelocM != Reloc::PIC_)
2393 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2394 ARMConstantPoolValue *CPV =
2395 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2396 ARMCP::CPLSDA, PCAdj);
2397 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2398 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2400 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2401 MachinePointerInfo::getConstantPool(),
2402 false, false, false, 0);
2404 if (RelocM == Reloc::PIC_) {
2405 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2406 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2410 case Intrinsic::arm_neon_vmulls:
2411 case Intrinsic::arm_neon_vmullu: {
2412 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2413 ? ARMISD::VMULLs : ARMISD::VMULLu;
2414 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2415 Op.getOperand(1), Op.getOperand(2));
2420 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2421 const ARMSubtarget *Subtarget) {
2422 DebugLoc dl = Op.getDebugLoc();
2423 if (!Subtarget->hasDataBarrier()) {
2424 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2425 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2427 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2428 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2429 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2430 DAG.getConstant(0, MVT::i32));
2433 SDValue Op5 = Op.getOperand(5);
2434 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2435 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2436 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2437 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2439 ARM_MB::MemBOpt DMBOpt;
2440 if (isDeviceBarrier)
2441 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2443 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2444 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2445 DAG.getConstant(DMBOpt, MVT::i32));
2449 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2450 const ARMSubtarget *Subtarget) {
2451 // FIXME: handle "fence singlethread" more efficiently.
2452 DebugLoc dl = Op.getDebugLoc();
2453 if (!Subtarget->hasDataBarrier()) {
2454 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2455 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2457 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2458 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2459 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2460 DAG.getConstant(0, MVT::i32));
2463 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2464 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2467 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2468 const ARMSubtarget *Subtarget) {
2469 // ARM pre v5TE and Thumb1 does not have preload instructions.
2470 if (!(Subtarget->isThumb2() ||
2471 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2472 // Just preserve the chain.
2473 return Op.getOperand(0);
2475 DebugLoc dl = Op.getDebugLoc();
2476 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2478 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2479 // ARMv7 with MP extension has PLDW.
2480 return Op.getOperand(0);
2482 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2483 if (Subtarget->isThumb()) {
2485 isRead = ~isRead & 1;
2486 isData = ~isData & 1;
2489 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2490 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2491 DAG.getConstant(isData, MVT::i32));
2494 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2495 MachineFunction &MF = DAG.getMachineFunction();
2496 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2498 // vastart just stores the address of the VarArgsFrameIndex slot into the
2499 // memory location argument.
2500 DebugLoc dl = Op.getDebugLoc();
2501 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2502 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2503 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2504 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2505 MachinePointerInfo(SV), false, false, 0);
2509 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2510 SDValue &Root, SelectionDAG &DAG,
2511 DebugLoc dl) const {
2512 MachineFunction &MF = DAG.getMachineFunction();
2513 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2515 const TargetRegisterClass *RC;
2516 if (AFI->isThumb1OnlyFunction())
2517 RC = &ARM::tGPRRegClass;
2519 RC = &ARM::GPRRegClass;
2521 // Transform the arguments stored in physical registers into virtual ones.
2522 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2523 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2526 if (NextVA.isMemLoc()) {
2527 MachineFrameInfo *MFI = MF.getFrameInfo();
2528 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2530 // Create load node to retrieve arguments from the stack.
2531 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2532 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2533 MachinePointerInfo::getFixedStack(FI),
2534 false, false, false, 0);
2536 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2537 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2540 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2544 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2545 unsigned &VARegSize, unsigned &VARegSaveSize)
2548 if (CCInfo.isFirstByValRegValid())
2549 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2551 unsigned int firstUnalloced;
2552 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2553 sizeof(GPRArgRegs) /
2554 sizeof(GPRArgRegs[0]));
2555 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2558 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2559 VARegSize = NumGPRs * 4;
2560 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2563 // The remaining GPRs hold either the beginning of variable-argument
2564 // data, or the beginning of an aggregate passed by value (usuall
2565 // byval). Either way, we allocate stack slots adjacent to the data
2566 // provided by our caller, and store the unallocated registers there.
2567 // If this is a variadic function, the va_list pointer will begin with
2568 // these values; otherwise, this reassembles a (byval) structure that
2569 // was split between registers and memory.
2571 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2572 DebugLoc dl, SDValue &Chain,
2573 const Value *OrigArg,
2574 unsigned OffsetFromOrigArg,
2576 bool ForceMutable) const {
2577 MachineFunction &MF = DAG.getMachineFunction();
2578 MachineFrameInfo *MFI = MF.getFrameInfo();
2579 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2580 unsigned firstRegToSaveIndex;
2581 if (CCInfo.isFirstByValRegValid())
2582 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2584 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2585 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2588 unsigned VARegSize, VARegSaveSize;
2589 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2590 if (VARegSaveSize) {
2591 // If this function is vararg, store any remaining integer argument regs
2592 // to their spots on the stack so that they may be loaded by deferencing
2593 // the result of va_next.
2594 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2595 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2596 ArgOffset + VARegSaveSize
2599 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2602 SmallVector<SDValue, 4> MemOps;
2603 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2604 const TargetRegisterClass *RC;
2605 if (AFI->isThumb1OnlyFunction())
2606 RC = &ARM::tGPRRegClass;
2608 RC = &ARM::GPRRegClass;
2610 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2611 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2613 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2614 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2616 MemOps.push_back(Store);
2617 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2618 DAG.getConstant(4, getPointerTy()));
2620 if (!MemOps.empty())
2621 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2622 &MemOps[0], MemOps.size());
2624 // This will point to the next argument passed via stack.
2625 AFI->setVarArgsFrameIndex(
2626 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2630 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2631 CallingConv::ID CallConv, bool isVarArg,
2632 const SmallVectorImpl<ISD::InputArg>
2634 DebugLoc dl, SelectionDAG &DAG,
2635 SmallVectorImpl<SDValue> &InVals)
2637 MachineFunction &MF = DAG.getMachineFunction();
2638 MachineFrameInfo *MFI = MF.getFrameInfo();
2640 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2642 // Assign locations to all of the incoming arguments.
2643 SmallVector<CCValAssign, 16> ArgLocs;
2644 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2645 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2646 CCInfo.AnalyzeFormalArguments(Ins,
2647 CCAssignFnForNode(CallConv, /* Return*/ false,
2650 SmallVector<SDValue, 16> ArgValues;
2651 int lastInsIndex = -1;
2653 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2654 unsigned CurArgIdx = 0;
2655 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2656 CCValAssign &VA = ArgLocs[i];
2657 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2658 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2659 // Arguments stored in registers.
2660 if (VA.isRegLoc()) {
2661 EVT RegVT = VA.getLocVT();
2663 if (VA.needsCustom()) {
2664 // f64 and vector types are split up into multiple registers or
2665 // combinations of registers and stack slots.
2666 if (VA.getLocVT() == MVT::v2f64) {
2667 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2669 VA = ArgLocs[++i]; // skip ahead to next loc
2671 if (VA.isMemLoc()) {
2672 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2673 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2674 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2675 MachinePointerInfo::getFixedStack(FI),
2676 false, false, false, 0);
2678 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2681 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2682 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2683 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2684 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2685 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2687 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2690 const TargetRegisterClass *RC;
2692 if (RegVT == MVT::f32)
2693 RC = &ARM::SPRRegClass;
2694 else if (RegVT == MVT::f64)
2695 RC = &ARM::DPRRegClass;
2696 else if (RegVT == MVT::v2f64)
2697 RC = &ARM::QPRRegClass;
2698 else if (RegVT == MVT::i32)
2699 RC = AFI->isThumb1OnlyFunction() ?
2700 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2701 (const TargetRegisterClass*)&ARM::GPRRegClass;
2703 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2705 // Transform the arguments in physical registers into virtual ones.
2706 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2707 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2710 // If this is an 8 or 16-bit value, it is really passed promoted
2711 // to 32 bits. Insert an assert[sz]ext to capture this, then
2712 // truncate to the right size.
2713 switch (VA.getLocInfo()) {
2714 default: llvm_unreachable("Unknown loc info!");
2715 case CCValAssign::Full: break;
2716 case CCValAssign::BCvt:
2717 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2719 case CCValAssign::SExt:
2720 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2721 DAG.getValueType(VA.getValVT()));
2722 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2724 case CCValAssign::ZExt:
2725 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2726 DAG.getValueType(VA.getValVT()));
2727 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2731 InVals.push_back(ArgValue);
2733 } else { // VA.isRegLoc()
2736 assert(VA.isMemLoc());
2737 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2739 int index = ArgLocs[i].getValNo();
2741 // Some Ins[] entries become multiple ArgLoc[] entries.
2742 // Process them only once.
2743 if (index != lastInsIndex)
2745 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2746 // FIXME: For now, all byval parameter objects are marked mutable.
2747 // This can be changed with more analysis.
2748 // In case of tail call optimization mark all arguments mutable.
2749 // Since they could be overwritten by lowering of arguments in case of
2751 if (Flags.isByVal()) {
2752 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2753 if (!AFI->getVarArgsFrameIndex()) {
2754 VarArgStyleRegisters(CCInfo, DAG,
2755 dl, Chain, CurOrigArg,
2756 Ins[VA.getValNo()].PartOffset,
2757 VA.getLocMemOffset(),
2758 true /*force mutable frames*/);
2759 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2760 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2762 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2763 VA.getLocMemOffset(), false);
2764 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2767 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2768 VA.getLocMemOffset(), true);
2770 // Create load nodes to retrieve arguments from the stack.
2771 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2772 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2773 MachinePointerInfo::getFixedStack(FI),
2774 false, false, false, 0));
2776 lastInsIndex = index;
2783 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2784 CCInfo.getNextStackOffset());
2789 /// isFloatingPointZero - Return true if this is +0.0.
2790 static bool isFloatingPointZero(SDValue Op) {
2791 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2792 return CFP->getValueAPF().isPosZero();
2793 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2794 // Maybe this has already been legalized into the constant pool?
2795 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2796 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2797 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2798 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2799 return CFP->getValueAPF().isPosZero();
2805 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2806 /// the given operands.
2808 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2809 SDValue &ARMcc, SelectionDAG &DAG,
2810 DebugLoc dl) const {
2811 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2812 unsigned C = RHSC->getZExtValue();
2813 if (!isLegalICmpImmediate(C)) {
2814 // Constant does not fit, try adjusting it by one?
2819 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2820 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2821 RHS = DAG.getConstant(C-1, MVT::i32);
2826 if (C != 0 && isLegalICmpImmediate(C-1)) {
2827 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2828 RHS = DAG.getConstant(C-1, MVT::i32);
2833 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2834 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2835 RHS = DAG.getConstant(C+1, MVT::i32);
2840 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2841 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2842 RHS = DAG.getConstant(C+1, MVT::i32);
2849 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2850 ARMISD::NodeType CompareType;
2853 CompareType = ARMISD::CMP;
2858 CompareType = ARMISD::CMPZ;
2861 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2862 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2865 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2867 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2868 DebugLoc dl) const {
2870 if (!isFloatingPointZero(RHS))
2871 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2873 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2874 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2877 /// duplicateCmp - Glue values can have only one use, so this function
2878 /// duplicates a comparison node.
2880 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2881 unsigned Opc = Cmp.getOpcode();
2882 DebugLoc DL = Cmp.getDebugLoc();
2883 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2884 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2886 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2887 Cmp = Cmp.getOperand(0);
2888 Opc = Cmp.getOpcode();
2889 if (Opc == ARMISD::CMPFP)
2890 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2892 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2893 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2895 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2898 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2899 SDValue Cond = Op.getOperand(0);
2900 SDValue SelectTrue = Op.getOperand(1);
2901 SDValue SelectFalse = Op.getOperand(2);
2902 DebugLoc dl = Op.getDebugLoc();
2906 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2907 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2909 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2910 const ConstantSDNode *CMOVTrue =
2911 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2912 const ConstantSDNode *CMOVFalse =
2913 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2915 if (CMOVTrue && CMOVFalse) {
2916 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2917 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2921 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2923 False = SelectFalse;
2924 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2929 if (True.getNode() && False.getNode()) {
2930 EVT VT = Op.getValueType();
2931 SDValue ARMcc = Cond.getOperand(2);
2932 SDValue CCR = Cond.getOperand(3);
2933 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2934 assert(True.getValueType() == VT);
2935 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2940 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2941 // undefined bits before doing a full-word comparison with zero.
2942 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2943 DAG.getConstant(1, Cond.getValueType()));
2945 return DAG.getSelectCC(dl, Cond,
2946 DAG.getConstant(0, Cond.getValueType()),
2947 SelectTrue, SelectFalse, ISD::SETNE);
2950 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2951 EVT VT = Op.getValueType();
2952 SDValue LHS = Op.getOperand(0);
2953 SDValue RHS = Op.getOperand(1);
2954 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2955 SDValue TrueVal = Op.getOperand(2);
2956 SDValue FalseVal = Op.getOperand(3);
2957 DebugLoc dl = Op.getDebugLoc();
2959 if (LHS.getValueType() == MVT::i32) {
2961 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2962 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2963 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2966 ARMCC::CondCodes CondCode, CondCode2;
2967 FPCCToARMCC(CC, CondCode, CondCode2);
2969 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2970 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2971 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2972 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2974 if (CondCode2 != ARMCC::AL) {
2975 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2976 // FIXME: Needs another CMP because flag can have but one use.
2977 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2978 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2979 Result, TrueVal, ARMcc2, CCR, Cmp2);
2984 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2985 /// to morph to an integer compare sequence.
2986 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2987 const ARMSubtarget *Subtarget) {
2988 SDNode *N = Op.getNode();
2989 if (!N->hasOneUse())
2990 // Otherwise it requires moving the value from fp to integer registers.
2992 if (!N->getNumValues())
2994 EVT VT = Op.getValueType();
2995 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2996 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2997 // vmrs are very slow, e.g. cortex-a8.
3000 if (isFloatingPointZero(Op)) {
3004 return ISD::isNormalLoad(N);
3007 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3008 if (isFloatingPointZero(Op))
3009 return DAG.getConstant(0, MVT::i32);
3011 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3012 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3013 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3014 Ld->isVolatile(), Ld->isNonTemporal(),
3015 Ld->isInvariant(), Ld->getAlignment());
3017 llvm_unreachable("Unknown VFP cmp argument!");
3020 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3021 SDValue &RetVal1, SDValue &RetVal2) {
3022 if (isFloatingPointZero(Op)) {
3023 RetVal1 = DAG.getConstant(0, MVT::i32);
3024 RetVal2 = DAG.getConstant(0, MVT::i32);
3028 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3029 SDValue Ptr = Ld->getBasePtr();
3030 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3031 Ld->getChain(), Ptr,
3032 Ld->getPointerInfo(),
3033 Ld->isVolatile(), Ld->isNonTemporal(),
3034 Ld->isInvariant(), Ld->getAlignment());
3036 EVT PtrType = Ptr.getValueType();
3037 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3038 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3039 PtrType, Ptr, DAG.getConstant(4, PtrType));
3040 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3041 Ld->getChain(), NewPtr,
3042 Ld->getPointerInfo().getWithOffset(4),
3043 Ld->isVolatile(), Ld->isNonTemporal(),
3044 Ld->isInvariant(), NewAlign);
3048 llvm_unreachable("Unknown VFP cmp argument!");
3051 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3052 /// f32 and even f64 comparisons to integer ones.
3054 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3055 SDValue Chain = Op.getOperand(0);
3056 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3057 SDValue LHS = Op.getOperand(2);
3058 SDValue RHS = Op.getOperand(3);
3059 SDValue Dest = Op.getOperand(4);
3060 DebugLoc dl = Op.getDebugLoc();
3062 bool LHSSeenZero = false;
3063 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3064 bool RHSSeenZero = false;
3065 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3066 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3067 // If unsafe fp math optimization is enabled and there are no other uses of
3068 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3069 // to an integer comparison.
3070 if (CC == ISD::SETOEQ)
3072 else if (CC == ISD::SETUNE)
3075 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3077 if (LHS.getValueType() == MVT::f32) {
3078 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3079 bitcastf32Toi32(LHS, DAG), Mask);
3080 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3081 bitcastf32Toi32(RHS, DAG), Mask);
3082 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3083 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3084 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3085 Chain, Dest, ARMcc, CCR, Cmp);
3090 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3091 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3092 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3093 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3094 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3095 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3096 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3097 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3098 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3104 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3105 SDValue Chain = Op.getOperand(0);
3106 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3107 SDValue LHS = Op.getOperand(2);
3108 SDValue RHS = Op.getOperand(3);
3109 SDValue Dest = Op.getOperand(4);
3110 DebugLoc dl = Op.getDebugLoc();
3112 if (LHS.getValueType() == MVT::i32) {
3114 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3115 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3116 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3117 Chain, Dest, ARMcc, CCR, Cmp);
3120 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3122 if (getTargetMachine().Options.UnsafeFPMath &&
3123 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3124 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3125 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3126 if (Result.getNode())
3130 ARMCC::CondCodes CondCode, CondCode2;
3131 FPCCToARMCC(CC, CondCode, CondCode2);
3133 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3134 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3135 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3136 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3137 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3138 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3139 if (CondCode2 != ARMCC::AL) {
3140 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3141 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3142 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3147 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3148 SDValue Chain = Op.getOperand(0);
3149 SDValue Table = Op.getOperand(1);
3150 SDValue Index = Op.getOperand(2);
3151 DebugLoc dl = Op.getDebugLoc();
3153 EVT PTy = getPointerTy();
3154 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3155 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3156 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3157 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3158 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3159 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3160 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3161 if (Subtarget->isThumb2()) {
3162 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3163 // which does another jump to the destination. This also makes it easier
3164 // to translate it to TBB / TBH later.
3165 // FIXME: This might not work if the function is extremely large.
3166 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3167 Addr, Op.getOperand(2), JTI, UId);
3169 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3170 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3171 MachinePointerInfo::getJumpTable(),
3172 false, false, false, 0);
3173 Chain = Addr.getValue(1);
3174 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3175 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3177 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3178 MachinePointerInfo::getJumpTable(),
3179 false, false, false, 0);
3180 Chain = Addr.getValue(1);
3181 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3185 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3186 EVT VT = Op.getValueType();
3187 DebugLoc dl = Op.getDebugLoc();
3189 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3190 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3192 return DAG.UnrollVectorOp(Op.getNode());
3195 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3196 "Invalid type for custom lowering!");
3197 if (VT != MVT::v4i16)
3198 return DAG.UnrollVectorOp(Op.getNode());
3200 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3201 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3204 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3205 EVT VT = Op.getValueType();
3207 return LowerVectorFP_TO_INT(Op, DAG);
3209 DebugLoc dl = Op.getDebugLoc();
3212 switch (Op.getOpcode()) {
3213 default: llvm_unreachable("Invalid opcode!");
3214 case ISD::FP_TO_SINT:
3215 Opc = ARMISD::FTOSI;
3217 case ISD::FP_TO_UINT:
3218 Opc = ARMISD::FTOUI;
3221 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3222 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3225 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3226 EVT VT = Op.getValueType();
3227 DebugLoc dl = Op.getDebugLoc();
3229 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3230 if (VT.getVectorElementType() == MVT::f32)
3232 return DAG.UnrollVectorOp(Op.getNode());
3235 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3236 "Invalid type for custom lowering!");
3237 if (VT != MVT::v4f32)
3238 return DAG.UnrollVectorOp(Op.getNode());
3242 switch (Op.getOpcode()) {
3243 default: llvm_unreachable("Invalid opcode!");
3244 case ISD::SINT_TO_FP:
3245 CastOpc = ISD::SIGN_EXTEND;
3246 Opc = ISD::SINT_TO_FP;
3248 case ISD::UINT_TO_FP:
3249 CastOpc = ISD::ZERO_EXTEND;
3250 Opc = ISD::UINT_TO_FP;
3254 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3255 return DAG.getNode(Opc, dl, VT, Op);
3258 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3259 EVT VT = Op.getValueType();
3261 return LowerVectorINT_TO_FP(Op, DAG);
3263 DebugLoc dl = Op.getDebugLoc();
3266 switch (Op.getOpcode()) {
3267 default: llvm_unreachable("Invalid opcode!");
3268 case ISD::SINT_TO_FP:
3269 Opc = ARMISD::SITOF;
3271 case ISD::UINT_TO_FP:
3272 Opc = ARMISD::UITOF;
3276 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3277 return DAG.getNode(Opc, dl, VT, Op);
3280 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3281 // Implement fcopysign with a fabs and a conditional fneg.
3282 SDValue Tmp0 = Op.getOperand(0);
3283 SDValue Tmp1 = Op.getOperand(1);
3284 DebugLoc dl = Op.getDebugLoc();
3285 EVT VT = Op.getValueType();
3286 EVT SrcVT = Tmp1.getValueType();
3287 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3288 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3289 bool UseNEON = !InGPR && Subtarget->hasNEON();
3292 // Use VBSL to copy the sign bit.
3293 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3294 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3295 DAG.getTargetConstant(EncodedVal, MVT::i32));
3296 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3298 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3299 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3300 DAG.getConstant(32, MVT::i32));
3301 else /*if (VT == MVT::f32)*/
3302 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3303 if (SrcVT == MVT::f32) {
3304 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3306 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3307 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3308 DAG.getConstant(32, MVT::i32));
3309 } else if (VT == MVT::f32)
3310 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3311 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3312 DAG.getConstant(32, MVT::i32));
3313 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3314 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3316 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3318 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3319 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3320 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3322 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3323 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3324 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3325 if (VT == MVT::f32) {
3326 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3327 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3328 DAG.getConstant(0, MVT::i32));
3330 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3336 // Bitcast operand 1 to i32.
3337 if (SrcVT == MVT::f64)
3338 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3339 &Tmp1, 1).getValue(1);
3340 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3342 // Or in the signbit with integer operations.
3343 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3344 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3345 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3346 if (VT == MVT::f32) {
3347 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3348 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3349 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3350 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3353 // f64: Or the high part with signbit and then combine two parts.
3354 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3356 SDValue Lo = Tmp0.getValue(0);
3357 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3358 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3359 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3362 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3363 MachineFunction &MF = DAG.getMachineFunction();
3364 MachineFrameInfo *MFI = MF.getFrameInfo();
3365 MFI->setReturnAddressIsTaken(true);
3367 EVT VT = Op.getValueType();
3368 DebugLoc dl = Op.getDebugLoc();
3369 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3371 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3372 SDValue Offset = DAG.getConstant(4, MVT::i32);
3373 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3374 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3375 MachinePointerInfo(), false, false, false, 0);
3378 // Return LR, which contains the return address. Mark it an implicit live-in.
3379 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3380 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3383 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3384 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3385 MFI->setFrameAddressIsTaken(true);
3387 EVT VT = Op.getValueType();
3388 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3389 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3390 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3391 ? ARM::R7 : ARM::R11;
3392 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3394 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3395 MachinePointerInfo(),
3396 false, false, false, 0);
3400 /// ExpandBITCAST - If the target supports VFP, this function is called to
3401 /// expand a bit convert where either the source or destination type is i64 to
3402 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3403 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3404 /// vectors), since the legalizer won't know what to do with that.
3405 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3406 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3407 DebugLoc dl = N->getDebugLoc();
3408 SDValue Op = N->getOperand(0);
3410 // This function is only supposed to be called for i64 types, either as the
3411 // source or destination of the bit convert.
3412 EVT SrcVT = Op.getValueType();
3413 EVT DstVT = N->getValueType(0);
3414 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3415 "ExpandBITCAST called for non-i64 type");
3417 // Turn i64->f64 into VMOVDRR.
3418 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3419 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3420 DAG.getConstant(0, MVT::i32));
3421 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3422 DAG.getConstant(1, MVT::i32));
3423 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3424 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3427 // Turn f64->i64 into VMOVRRD.
3428 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3429 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3430 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3431 // Merge the pieces into a single i64 value.
3432 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3438 /// getZeroVector - Returns a vector of specified type with all zero elements.
3439 /// Zero vectors are used to represent vector negation and in those cases
3440 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3441 /// not support i64 elements, so sometimes the zero vectors will need to be
3442 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3444 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3445 assert(VT.isVector() && "Expected a vector type");
3446 // The canonical modified immediate encoding of a zero vector is....0!
3447 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3448 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3449 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3450 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3453 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3454 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3455 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3456 SelectionDAG &DAG) const {
3457 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3458 EVT VT = Op.getValueType();
3459 unsigned VTBits = VT.getSizeInBits();
3460 DebugLoc dl = Op.getDebugLoc();
3461 SDValue ShOpLo = Op.getOperand(0);
3462 SDValue ShOpHi = Op.getOperand(1);
3463 SDValue ShAmt = Op.getOperand(2);
3465 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3467 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3469 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3470 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3471 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3472 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3473 DAG.getConstant(VTBits, MVT::i32));
3474 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3475 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3476 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3478 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3479 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3481 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3482 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3485 SDValue Ops[2] = { Lo, Hi };
3486 return DAG.getMergeValues(Ops, 2, dl);
3489 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3490 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3491 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3492 SelectionDAG &DAG) const {
3493 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3494 EVT VT = Op.getValueType();
3495 unsigned VTBits = VT.getSizeInBits();
3496 DebugLoc dl = Op.getDebugLoc();
3497 SDValue ShOpLo = Op.getOperand(0);
3498 SDValue ShOpHi = Op.getOperand(1);
3499 SDValue ShAmt = Op.getOperand(2);
3502 assert(Op.getOpcode() == ISD::SHL_PARTS);
3503 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3504 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3505 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3506 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3507 DAG.getConstant(VTBits, MVT::i32));
3508 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3509 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3511 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3512 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3513 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3515 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3516 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3519 SDValue Ops[2] = { Lo, Hi };
3520 return DAG.getMergeValues(Ops, 2, dl);
3523 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3524 SelectionDAG &DAG) const {
3525 // The rounding mode is in bits 23:22 of the FPSCR.
3526 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3527 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3528 // so that the shift + and get folded into a bitfield extract.
3529 DebugLoc dl = Op.getDebugLoc();
3530 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3531 DAG.getConstant(Intrinsic::arm_get_fpscr,
3533 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3534 DAG.getConstant(1U << 22, MVT::i32));
3535 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3536 DAG.getConstant(22, MVT::i32));
3537 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3538 DAG.getConstant(3, MVT::i32));
3541 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3542 const ARMSubtarget *ST) {
3543 EVT VT = N->getValueType(0);
3544 DebugLoc dl = N->getDebugLoc();
3546 if (!ST->hasV6T2Ops())
3549 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3550 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3553 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3554 const ARMSubtarget *ST) {
3555 EVT VT = N->getValueType(0);
3556 DebugLoc dl = N->getDebugLoc();
3561 // Lower vector shifts on NEON to use VSHL.
3562 assert(ST->hasNEON() && "unexpected vector shift");
3564 // Left shifts translate directly to the vshiftu intrinsic.
3565 if (N->getOpcode() == ISD::SHL)
3566 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3567 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3568 N->getOperand(0), N->getOperand(1));
3570 assert((N->getOpcode() == ISD::SRA ||
3571 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3573 // NEON uses the same intrinsics for both left and right shifts. For
3574 // right shifts, the shift amounts are negative, so negate the vector of
3576 EVT ShiftVT = N->getOperand(1).getValueType();
3577 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3578 getZeroVector(ShiftVT, DAG, dl),
3580 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3581 Intrinsic::arm_neon_vshifts :
3582 Intrinsic::arm_neon_vshiftu);
3583 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3584 DAG.getConstant(vshiftInt, MVT::i32),
3585 N->getOperand(0), NegatedCount);
3588 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3589 const ARMSubtarget *ST) {
3590 EVT VT = N->getValueType(0);
3591 DebugLoc dl = N->getDebugLoc();
3593 // We can get here for a node like i32 = ISD::SHL i32, i64
3597 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3598 "Unknown shift to lower!");
3600 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3601 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3602 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3605 // If we are in thumb mode, we don't have RRX.
3606 if (ST->isThumb1Only()) return SDValue();
3608 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3609 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3610 DAG.getConstant(0, MVT::i32));
3611 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3612 DAG.getConstant(1, MVT::i32));
3614 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3615 // captures the result into a carry flag.
3616 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3617 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3619 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3620 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3622 // Merge the pieces into a single i64 value.
3623 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3626 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3627 SDValue TmpOp0, TmpOp1;
3628 bool Invert = false;
3632 SDValue Op0 = Op.getOperand(0);
3633 SDValue Op1 = Op.getOperand(1);
3634 SDValue CC = Op.getOperand(2);
3635 EVT VT = Op.getValueType();
3636 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3637 DebugLoc dl = Op.getDebugLoc();
3639 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3640 switch (SetCCOpcode) {
3641 default: llvm_unreachable("Illegal FP comparison");
3643 case ISD::SETNE: Invert = true; // Fallthrough
3645 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3647 case ISD::SETLT: Swap = true; // Fallthrough
3649 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3651 case ISD::SETLE: Swap = true; // Fallthrough
3653 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3654 case ISD::SETUGE: Swap = true; // Fallthrough
3655 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3656 case ISD::SETUGT: Swap = true; // Fallthrough
3657 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3658 case ISD::SETUEQ: Invert = true; // Fallthrough
3660 // Expand this to (OLT | OGT).
3664 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3665 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3667 case ISD::SETUO: Invert = true; // Fallthrough
3669 // Expand this to (OLT | OGE).
3673 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3674 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3678 // Integer comparisons.
3679 switch (SetCCOpcode) {
3680 default: llvm_unreachable("Illegal integer comparison");
3681 case ISD::SETNE: Invert = true;
3682 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3683 case ISD::SETLT: Swap = true;
3684 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3685 case ISD::SETLE: Swap = true;
3686 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3687 case ISD::SETULT: Swap = true;
3688 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3689 case ISD::SETULE: Swap = true;
3690 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3693 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3694 if (Opc == ARMISD::VCEQ) {
3697 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3699 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3702 // Ignore bitconvert.
3703 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3704 AndOp = AndOp.getOperand(0);
3706 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3708 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3709 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3716 std::swap(Op0, Op1);
3718 // If one of the operands is a constant vector zero, attempt to fold the
3719 // comparison to a specialized compare-against-zero form.
3721 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3723 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3724 if (Opc == ARMISD::VCGE)
3725 Opc = ARMISD::VCLEZ;
3726 else if (Opc == ARMISD::VCGT)
3727 Opc = ARMISD::VCLTZ;
3732 if (SingleOp.getNode()) {
3735 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3737 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3739 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3741 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3743 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3745 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3748 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3752 Result = DAG.getNOT(dl, Result, VT);
3757 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3758 /// valid vector constant for a NEON instruction with a "modified immediate"
3759 /// operand (e.g., VMOV). If so, return the encoded value.
3760 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3761 unsigned SplatBitSize, SelectionDAG &DAG,
3762 EVT &VT, bool is128Bits, NEONModImmType type) {
3763 unsigned OpCmode, Imm;
3765 // SplatBitSize is set to the smallest size that splats the vector, so a
3766 // zero vector will always have SplatBitSize == 8. However, NEON modified
3767 // immediate instructions others than VMOV do not support the 8-bit encoding
3768 // of a zero vector, and the default encoding of zero is supposed to be the
3773 switch (SplatBitSize) {
3775 if (type != VMOVModImm)
3777 // Any 1-byte value is OK. Op=0, Cmode=1110.
3778 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3781 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3785 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3786 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3787 if ((SplatBits & ~0xff) == 0) {
3788 // Value = 0x00nn: Op=x, Cmode=100x.
3793 if ((SplatBits & ~0xff00) == 0) {
3794 // Value = 0xnn00: Op=x, Cmode=101x.
3796 Imm = SplatBits >> 8;
3802 // NEON's 32-bit VMOV supports splat values where:
3803 // * only one byte is nonzero, or
3804 // * the least significant byte is 0xff and the second byte is nonzero, or
3805 // * the least significant 2 bytes are 0xff and the third is nonzero.
3806 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3807 if ((SplatBits & ~0xff) == 0) {
3808 // Value = 0x000000nn: Op=x, Cmode=000x.
3813 if ((SplatBits & ~0xff00) == 0) {
3814 // Value = 0x0000nn00: Op=x, Cmode=001x.
3816 Imm = SplatBits >> 8;
3819 if ((SplatBits & ~0xff0000) == 0) {
3820 // Value = 0x00nn0000: Op=x, Cmode=010x.
3822 Imm = SplatBits >> 16;
3825 if ((SplatBits & ~0xff000000) == 0) {
3826 // Value = 0xnn000000: Op=x, Cmode=011x.
3828 Imm = SplatBits >> 24;
3832 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3833 if (type == OtherModImm) return SDValue();
3835 if ((SplatBits & ~0xffff) == 0 &&
3836 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3837 // Value = 0x0000nnff: Op=x, Cmode=1100.
3839 Imm = SplatBits >> 8;
3844 if ((SplatBits & ~0xffffff) == 0 &&
3845 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3846 // Value = 0x00nnffff: Op=x, Cmode=1101.
3848 Imm = SplatBits >> 16;
3849 SplatBits |= 0xffff;
3853 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3854 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3855 // VMOV.I32. A (very) minor optimization would be to replicate the value
3856 // and fall through here to test for a valid 64-bit splat. But, then the
3857 // caller would also need to check and handle the change in size.
3861 if (type != VMOVModImm)
3863 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3864 uint64_t BitMask = 0xff;
3866 unsigned ImmMask = 1;
3868 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3869 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3872 } else if ((SplatBits & BitMask) != 0) {
3878 // Op=1, Cmode=1110.
3881 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3886 llvm_unreachable("unexpected size for isNEONModifiedImm");
3889 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3890 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3893 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3894 const ARMSubtarget *ST) const {
3895 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3898 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3899 assert(Op.getValueType() == MVT::f32 &&
3900 "ConstantFP custom lowering should only occur for f32.");
3902 // Try splatting with a VMOV.f32...
3903 APFloat FPVal = CFP->getValueAPF();
3904 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3906 DebugLoc DL = Op.getDebugLoc();
3907 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3908 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3910 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3911 DAG.getConstant(0, MVT::i32));
3914 // If that fails, try a VMOV.i32
3916 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3917 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3919 if (NewVal != SDValue()) {
3920 DebugLoc DL = Op.getDebugLoc();
3921 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
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));
3929 // Finally, try a VMVN.i32
3930 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3932 if (NewVal != SDValue()) {
3933 DebugLoc DL = Op.getDebugLoc();
3934 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3935 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3937 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3938 DAG.getConstant(0, MVT::i32));
3944 // check if an VEXT instruction can handle the shuffle mask when the
3945 // vector sources of the shuffle are the same.
3946 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3947 unsigned NumElts = VT.getVectorNumElements();
3949 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3955 // If this is a VEXT shuffle, the immediate value is the index of the first
3956 // element. The other shuffle indices must be the successive elements after
3958 unsigned ExpectedElt = Imm;
3959 for (unsigned i = 1; i < NumElts; ++i) {
3960 // Increment the expected index. If it wraps around, just follow it
3961 // back to index zero and keep going.
3963 if (ExpectedElt == NumElts)
3966 if (M[i] < 0) continue; // ignore UNDEF indices
3967 if (ExpectedElt != static_cast<unsigned>(M[i]))
3975 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3976 bool &ReverseVEXT, unsigned &Imm) {
3977 unsigned NumElts = VT.getVectorNumElements();
3978 ReverseVEXT = false;
3980 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3986 // If this is a VEXT shuffle, the immediate value is the index of the first
3987 // element. The other shuffle indices must be the successive elements after
3989 unsigned ExpectedElt = Imm;
3990 for (unsigned i = 1; i < NumElts; ++i) {
3991 // Increment the expected index. If it wraps around, it may still be
3992 // a VEXT but the source vectors must be swapped.
3994 if (ExpectedElt == NumElts * 2) {
3999 if (M[i] < 0) continue; // ignore UNDEF indices
4000 if (ExpectedElt != static_cast<unsigned>(M[i]))
4004 // Adjust the index value if the source operands will be swapped.
4011 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4012 /// instruction with the specified blocksize. (The order of the elements
4013 /// within each block of the vector is reversed.)
4014 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4015 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4016 "Only possible block sizes for VREV are: 16, 32, 64");
4018 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4022 unsigned NumElts = VT.getVectorNumElements();
4023 unsigned BlockElts = M[0] + 1;
4024 // If the first shuffle index is UNDEF, be optimistic.
4026 BlockElts = BlockSize / EltSz;
4028 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4031 for (unsigned i = 0; i < NumElts; ++i) {
4032 if (M[i] < 0) continue; // ignore UNDEF indices
4033 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4040 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4041 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4042 // range, then 0 is placed into the resulting vector. So pretty much any mask
4043 // of 8 elements can work here.
4044 return VT == MVT::v8i8 && M.size() == 8;
4047 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4048 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4052 unsigned NumElts = VT.getVectorNumElements();
4053 WhichResult = (M[0] == 0 ? 0 : 1);
4054 for (unsigned i = 0; i < NumElts; i += 2) {
4055 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4056 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4062 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4063 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4064 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4065 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4066 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4070 unsigned NumElts = VT.getVectorNumElements();
4071 WhichResult = (M[0] == 0 ? 0 : 1);
4072 for (unsigned i = 0; i < NumElts; i += 2) {
4073 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4074 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4080 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4081 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4085 unsigned NumElts = VT.getVectorNumElements();
4086 WhichResult = (M[0] == 0 ? 0 : 1);
4087 for (unsigned i = 0; i != NumElts; ++i) {
4088 if (M[i] < 0) continue; // ignore UNDEF indices
4089 if ((unsigned) M[i] != 2 * i + WhichResult)
4093 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4094 if (VT.is64BitVector() && EltSz == 32)
4100 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4101 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4102 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4103 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4104 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4108 unsigned Half = VT.getVectorNumElements() / 2;
4109 WhichResult = (M[0] == 0 ? 0 : 1);
4110 for (unsigned j = 0; j != 2; ++j) {
4111 unsigned Idx = WhichResult;
4112 for (unsigned i = 0; i != Half; ++i) {
4113 int MIdx = M[i + j * Half];
4114 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4120 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4121 if (VT.is64BitVector() && EltSz == 32)
4127 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4128 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4132 unsigned NumElts = VT.getVectorNumElements();
4133 WhichResult = (M[0] == 0 ? 0 : 1);
4134 unsigned Idx = WhichResult * NumElts / 2;
4135 for (unsigned i = 0; i != NumElts; i += 2) {
4136 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4137 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4142 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4143 if (VT.is64BitVector() && EltSz == 32)
4149 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4150 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4151 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4152 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4153 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4157 unsigned NumElts = VT.getVectorNumElements();
4158 WhichResult = (M[0] == 0 ? 0 : 1);
4159 unsigned Idx = WhichResult * NumElts / 2;
4160 for (unsigned i = 0; i != NumElts; i += 2) {
4161 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4162 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4167 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4168 if (VT.is64BitVector() && EltSz == 32)
4174 // If N is an integer constant that can be moved into a register in one
4175 // instruction, return an SDValue of such a constant (will become a MOV
4176 // instruction). Otherwise return null.
4177 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4178 const ARMSubtarget *ST, DebugLoc dl) {
4180 if (!isa<ConstantSDNode>(N))
4182 Val = cast<ConstantSDNode>(N)->getZExtValue();
4184 if (ST->isThumb1Only()) {
4185 if (Val <= 255 || ~Val <= 255)
4186 return DAG.getConstant(Val, MVT::i32);
4188 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4189 return DAG.getConstant(Val, MVT::i32);
4194 // If this is a case we can't handle, return null and let the default
4195 // expansion code take care of it.
4196 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4197 const ARMSubtarget *ST) const {
4198 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4199 DebugLoc dl = Op.getDebugLoc();
4200 EVT VT = Op.getValueType();
4202 APInt SplatBits, SplatUndef;
4203 unsigned SplatBitSize;
4205 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4206 if (SplatBitSize <= 64) {
4207 // Check if an immediate VMOV works.
4209 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4210 SplatUndef.getZExtValue(), SplatBitSize,
4211 DAG, VmovVT, VT.is128BitVector(),
4213 if (Val.getNode()) {
4214 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4215 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4218 // Try an immediate VMVN.
4219 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4220 Val = isNEONModifiedImm(NegatedImm,
4221 SplatUndef.getZExtValue(), SplatBitSize,
4222 DAG, VmovVT, VT.is128BitVector(),
4224 if (Val.getNode()) {
4225 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4226 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4229 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4230 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4231 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4233 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4234 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4240 // Scan through the operands to see if only one value is used.
4242 // As an optimisation, even if more than one value is used it may be more
4243 // profitable to splat with one value then change some lanes.
4245 // Heuristically we decide to do this if the vector has a "dominant" value,
4246 // defined as splatted to more than half of the lanes.
4247 unsigned NumElts = VT.getVectorNumElements();
4248 bool isOnlyLowElement = true;
4249 bool usesOnlyOneValue = true;
4250 bool hasDominantValue = false;
4251 bool isConstant = true;
4253 // Map of the number of times a particular SDValue appears in the
4255 DenseMap<SDValue, unsigned> ValueCounts;
4257 for (unsigned i = 0; i < NumElts; ++i) {
4258 SDValue V = Op.getOperand(i);
4259 if (V.getOpcode() == ISD::UNDEF)
4262 isOnlyLowElement = false;
4263 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4266 ValueCounts.insert(std::make_pair(V, 0));
4267 unsigned &Count = ValueCounts[V];
4269 // Is this value dominant? (takes up more than half of the lanes)
4270 if (++Count > (NumElts / 2)) {
4271 hasDominantValue = true;
4275 if (ValueCounts.size() != 1)
4276 usesOnlyOneValue = false;
4277 if (!Value.getNode() && ValueCounts.size() > 0)
4278 Value = ValueCounts.begin()->first;
4280 if (ValueCounts.size() == 0)
4281 return DAG.getUNDEF(VT);
4283 if (isOnlyLowElement)
4284 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4286 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4288 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4289 // i32 and try again.
4290 if (hasDominantValue && EltSize <= 32) {
4294 // If we are VDUPing a value that comes directly from a vector, that will
4295 // cause an unnecessary move to and from a GPR, where instead we could
4296 // just use VDUPLANE.
4297 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4298 // We need to create a new undef vector to use for the VDUPLANE if the
4299 // size of the vector from which we get the value is different than the
4300 // size of the vector that we need to create. We will insert the element
4301 // such that the register coalescer will remove unnecessary copies.
4302 if (VT != Value->getOperand(0).getValueType()) {
4303 ConstantSDNode *constIndex;
4304 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4305 assert(constIndex && "The index is not a constant!");
4306 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4307 VT.getVectorNumElements();
4308 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4309 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4310 Value, DAG.getConstant(index, MVT::i32)),
4311 DAG.getConstant(index, MVT::i32));
4313 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4314 Value->getOperand(0), Value->getOperand(1));
4318 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4320 if (!usesOnlyOneValue) {
4321 // The dominant value was splatted as 'N', but we now have to insert
4322 // all differing elements.
4323 for (unsigned I = 0; I < NumElts; ++I) {
4324 if (Op.getOperand(I) == Value)
4326 SmallVector<SDValue, 3> Ops;
4328 Ops.push_back(Op.getOperand(I));
4329 Ops.push_back(DAG.getConstant(I, MVT::i32));
4330 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4335 if (VT.getVectorElementType().isFloatingPoint()) {
4336 SmallVector<SDValue, 8> Ops;
4337 for (unsigned i = 0; i < NumElts; ++i)
4338 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4340 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4341 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4342 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4344 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4346 if (usesOnlyOneValue) {
4347 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4348 if (isConstant && Val.getNode())
4349 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4353 // If all elements are constants and the case above didn't get hit, fall back
4354 // to the default expansion, which will generate a load from the constant
4359 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4361 SDValue shuffle = ReconstructShuffle(Op, DAG);
4362 if (shuffle != SDValue())
4366 // Vectors with 32- or 64-bit elements can be built by directly assigning
4367 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4368 // will be legalized.
4369 if (EltSize >= 32) {
4370 // Do the expansion with floating-point types, since that is what the VFP
4371 // registers are defined to use, and since i64 is not legal.
4372 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4373 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4374 SmallVector<SDValue, 8> Ops;
4375 for (unsigned i = 0; i < NumElts; ++i)
4376 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4377 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4378 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4384 // Gather data to see if the operation can be modelled as a
4385 // shuffle in combination with VEXTs.
4386 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4387 SelectionDAG &DAG) const {
4388 DebugLoc dl = Op.getDebugLoc();
4389 EVT VT = Op.getValueType();
4390 unsigned NumElts = VT.getVectorNumElements();
4392 SmallVector<SDValue, 2> SourceVecs;
4393 SmallVector<unsigned, 2> MinElts;
4394 SmallVector<unsigned, 2> MaxElts;
4396 for (unsigned i = 0; i < NumElts; ++i) {
4397 SDValue V = Op.getOperand(i);
4398 if (V.getOpcode() == ISD::UNDEF)
4400 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4401 // A shuffle can only come from building a vector from various
4402 // elements of other vectors.
4404 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4405 VT.getVectorElementType()) {
4406 // This code doesn't know how to handle shuffles where the vector
4407 // element types do not match (this happens because type legalization
4408 // promotes the return type of EXTRACT_VECTOR_ELT).
4409 // FIXME: It might be appropriate to extend this code to handle
4410 // mismatched types.
4414 // Record this extraction against the appropriate vector if possible...
4415 SDValue SourceVec = V.getOperand(0);
4416 // If the element number isn't a constant, we can't effectively
4417 // analyze what's going on.
4418 if (!isa<ConstantSDNode>(V.getOperand(1)))
4420 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4421 bool FoundSource = false;
4422 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4423 if (SourceVecs[j] == SourceVec) {
4424 if (MinElts[j] > EltNo)
4426 if (MaxElts[j] < EltNo)
4433 // Or record a new source if not...
4435 SourceVecs.push_back(SourceVec);
4436 MinElts.push_back(EltNo);
4437 MaxElts.push_back(EltNo);
4441 // Currently only do something sane when at most two source vectors
4443 if (SourceVecs.size() > 2)
4446 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4447 int VEXTOffsets[2] = {0, 0};
4449 // This loop extracts the usage patterns of the source vectors
4450 // and prepares appropriate SDValues for a shuffle if possible.
4451 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4452 if (SourceVecs[i].getValueType() == VT) {
4453 // No VEXT necessary
4454 ShuffleSrcs[i] = SourceVecs[i];
4457 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4458 // It probably isn't worth padding out a smaller vector just to
4459 // break it down again in a shuffle.
4463 // Since only 64-bit and 128-bit vectors are legal on ARM and
4464 // we've eliminated the other cases...
4465 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4466 "unexpected vector sizes in ReconstructShuffle");
4468 if (MaxElts[i] - MinElts[i] >= NumElts) {
4469 // Span too large for a VEXT to cope
4473 if (MinElts[i] >= NumElts) {
4474 // The extraction can just take the second half
4475 VEXTOffsets[i] = NumElts;
4476 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4478 DAG.getIntPtrConstant(NumElts));
4479 } else if (MaxElts[i] < NumElts) {
4480 // The extraction can just take the first half
4482 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4484 DAG.getIntPtrConstant(0));
4486 // An actual VEXT is needed
4487 VEXTOffsets[i] = MinElts[i];
4488 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4490 DAG.getIntPtrConstant(0));
4491 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4493 DAG.getIntPtrConstant(NumElts));
4494 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4495 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4499 SmallVector<int, 8> Mask;
4501 for (unsigned i = 0; i < NumElts; ++i) {
4502 SDValue Entry = Op.getOperand(i);
4503 if (Entry.getOpcode() == ISD::UNDEF) {
4508 SDValue ExtractVec = Entry.getOperand(0);
4509 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4510 .getOperand(1))->getSExtValue();
4511 if (ExtractVec == SourceVecs[0]) {
4512 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4514 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4518 // Final check before we try to produce nonsense...
4519 if (isShuffleMaskLegal(Mask, VT))
4520 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4526 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4527 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4528 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4529 /// are assumed to be legal.
4531 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4533 if (VT.getVectorNumElements() == 4 &&
4534 (VT.is128BitVector() || VT.is64BitVector())) {
4535 unsigned PFIndexes[4];
4536 for (unsigned i = 0; i != 4; ++i) {
4540 PFIndexes[i] = M[i];
4543 // Compute the index in the perfect shuffle table.
4544 unsigned PFTableIndex =
4545 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4546 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4547 unsigned Cost = (PFEntry >> 30);
4554 unsigned Imm, WhichResult;
4556 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4557 return (EltSize >= 32 ||
4558 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4559 isVREVMask(M, VT, 64) ||
4560 isVREVMask(M, VT, 32) ||
4561 isVREVMask(M, VT, 16) ||
4562 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4563 isVTBLMask(M, VT) ||
4564 isVTRNMask(M, VT, WhichResult) ||
4565 isVUZPMask(M, VT, WhichResult) ||
4566 isVZIPMask(M, VT, WhichResult) ||
4567 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4568 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4569 isVZIP_v_undef_Mask(M, VT, WhichResult));
4572 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4573 /// the specified operations to build the shuffle.
4574 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4575 SDValue RHS, SelectionDAG &DAG,
4577 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4578 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4579 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4582 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4591 OP_VUZPL, // VUZP, left result
4592 OP_VUZPR, // VUZP, right result
4593 OP_VZIPL, // VZIP, left result
4594 OP_VZIPR, // VZIP, right result
4595 OP_VTRNL, // VTRN, left result
4596 OP_VTRNR // VTRN, right result
4599 if (OpNum == OP_COPY) {
4600 if (LHSID == (1*9+2)*9+3) return LHS;
4601 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4605 SDValue OpLHS, OpRHS;
4606 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4607 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4608 EVT VT = OpLHS.getValueType();
4611 default: llvm_unreachable("Unknown shuffle opcode!");
4613 // VREV divides the vector in half and swaps within the half.
4614 if (VT.getVectorElementType() == MVT::i32 ||
4615 VT.getVectorElementType() == MVT::f32)
4616 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4617 // vrev <4 x i16> -> VREV32
4618 if (VT.getVectorElementType() == MVT::i16)
4619 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4620 // vrev <4 x i8> -> VREV16
4621 assert(VT.getVectorElementType() == MVT::i8);
4622 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4627 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4628 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4632 return DAG.getNode(ARMISD::VEXT, dl, VT,
4634 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4637 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4638 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4641 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4642 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4645 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4646 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4650 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4651 ArrayRef<int> ShuffleMask,
4652 SelectionDAG &DAG) {
4653 // Check to see if we can use the VTBL instruction.
4654 SDValue V1 = Op.getOperand(0);
4655 SDValue V2 = Op.getOperand(1);
4656 DebugLoc DL = Op.getDebugLoc();
4658 SmallVector<SDValue, 8> VTBLMask;
4659 for (ArrayRef<int>::iterator
4660 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4661 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4663 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4664 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4665 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4668 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4669 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4673 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4674 SDValue V1 = Op.getOperand(0);
4675 SDValue V2 = Op.getOperand(1);
4676 DebugLoc dl = Op.getDebugLoc();
4677 EVT VT = Op.getValueType();
4678 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4680 // Convert shuffles that are directly supported on NEON to target-specific
4681 // DAG nodes, instead of keeping them as shuffles and matching them again
4682 // during code selection. This is more efficient and avoids the possibility
4683 // of inconsistencies between legalization and selection.
4684 // FIXME: floating-point vectors should be canonicalized to integer vectors
4685 // of the same time so that they get CSEd properly.
4686 ArrayRef<int> ShuffleMask = SVN->getMask();
4688 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4689 if (EltSize <= 32) {
4690 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4691 int Lane = SVN->getSplatIndex();
4692 // If this is undef splat, generate it via "just" vdup, if possible.
4693 if (Lane == -1) Lane = 0;
4695 // Test if V1 is a SCALAR_TO_VECTOR.
4696 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4697 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4699 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4700 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4702 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4703 !isa<ConstantSDNode>(V1.getOperand(0))) {
4704 bool IsScalarToVector = true;
4705 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4706 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4707 IsScalarToVector = false;
4710 if (IsScalarToVector)
4711 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4713 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4714 DAG.getConstant(Lane, MVT::i32));
4719 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4722 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4723 DAG.getConstant(Imm, MVT::i32));
4726 if (isVREVMask(ShuffleMask, VT, 64))
4727 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4728 if (isVREVMask(ShuffleMask, VT, 32))
4729 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4730 if (isVREVMask(ShuffleMask, VT, 16))
4731 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4733 if (V2->getOpcode() == ISD::UNDEF &&
4734 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4735 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4736 DAG.getConstant(Imm, MVT::i32));
4739 // Check for Neon shuffles that modify both input vectors in place.
4740 // If both results are used, i.e., if there are two shuffles with the same
4741 // source operands and with masks corresponding to both results of one of
4742 // these operations, DAG memoization will ensure that a single node is
4743 // used for both shuffles.
4744 unsigned WhichResult;
4745 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4746 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4747 V1, V2).getValue(WhichResult);
4748 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4749 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4750 V1, V2).getValue(WhichResult);
4751 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4752 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4753 V1, V2).getValue(WhichResult);
4755 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4756 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4757 V1, V1).getValue(WhichResult);
4758 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4759 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4760 V1, V1).getValue(WhichResult);
4761 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4762 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4763 V1, V1).getValue(WhichResult);
4766 // If the shuffle is not directly supported and it has 4 elements, use
4767 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4768 unsigned NumElts = VT.getVectorNumElements();
4770 unsigned PFIndexes[4];
4771 for (unsigned i = 0; i != 4; ++i) {
4772 if (ShuffleMask[i] < 0)
4775 PFIndexes[i] = ShuffleMask[i];
4778 // Compute the index in the perfect shuffle table.
4779 unsigned PFTableIndex =
4780 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4781 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4782 unsigned Cost = (PFEntry >> 30);
4785 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4788 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4789 if (EltSize >= 32) {
4790 // Do the expansion with floating-point types, since that is what the VFP
4791 // registers are defined to use, and since i64 is not legal.
4792 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4793 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4794 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4795 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4796 SmallVector<SDValue, 8> Ops;
4797 for (unsigned i = 0; i < NumElts; ++i) {
4798 if (ShuffleMask[i] < 0)
4799 Ops.push_back(DAG.getUNDEF(EltVT));
4801 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4802 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4803 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4806 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4807 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4810 if (VT == MVT::v8i8) {
4811 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4812 if (NewOp.getNode())
4819 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4820 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4821 SDValue Lane = Op.getOperand(2);
4822 if (!isa<ConstantSDNode>(Lane))
4828 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4829 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4830 SDValue Lane = Op.getOperand(1);
4831 if (!isa<ConstantSDNode>(Lane))
4834 SDValue Vec = Op.getOperand(0);
4835 if (Op.getValueType() == MVT::i32 &&
4836 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4837 DebugLoc dl = Op.getDebugLoc();
4838 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4844 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4845 // The only time a CONCAT_VECTORS operation can have legal types is when
4846 // two 64-bit vectors are concatenated to a 128-bit vector.
4847 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4848 "unexpected CONCAT_VECTORS");
4849 DebugLoc dl = Op.getDebugLoc();
4850 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4851 SDValue Op0 = Op.getOperand(0);
4852 SDValue Op1 = Op.getOperand(1);
4853 if (Op0.getOpcode() != ISD::UNDEF)
4854 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4855 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4856 DAG.getIntPtrConstant(0));
4857 if (Op1.getOpcode() != ISD::UNDEF)
4858 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4859 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4860 DAG.getIntPtrConstant(1));
4861 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4864 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4865 /// element has been zero/sign-extended, depending on the isSigned parameter,
4866 /// from an integer type half its size.
4867 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4869 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4870 EVT VT = N->getValueType(0);
4871 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4872 SDNode *BVN = N->getOperand(0).getNode();
4873 if (BVN->getValueType(0) != MVT::v4i32 ||
4874 BVN->getOpcode() != ISD::BUILD_VECTOR)
4876 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4877 unsigned HiElt = 1 - LoElt;
4878 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4879 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4880 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4881 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4882 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4885 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4886 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4889 if (Hi0->isNullValue() && Hi1->isNullValue())
4895 if (N->getOpcode() != ISD::BUILD_VECTOR)
4898 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4899 SDNode *Elt = N->getOperand(i).getNode();
4900 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4901 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4902 unsigned HalfSize = EltSize / 2;
4904 if (!isIntN(HalfSize, C->getSExtValue()))
4907 if (!isUIntN(HalfSize, C->getZExtValue()))
4918 /// isSignExtended - Check if a node is a vector value that is sign-extended
4919 /// or a constant BUILD_VECTOR with sign-extended elements.
4920 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4921 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4923 if (isExtendedBUILD_VECTOR(N, DAG, true))
4928 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4929 /// or a constant BUILD_VECTOR with zero-extended elements.
4930 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4931 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4933 if (isExtendedBUILD_VECTOR(N, DAG, false))
4938 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4939 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4940 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4941 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4942 return N->getOperand(0);
4943 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4944 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4945 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4946 LD->isNonTemporal(), LD->isInvariant(),
4947 LD->getAlignment());
4948 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4949 // have been legalized as a BITCAST from v4i32.
4950 if (N->getOpcode() == ISD::BITCAST) {
4951 SDNode *BVN = N->getOperand(0).getNode();
4952 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4953 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4954 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4955 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4956 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4958 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4959 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4960 EVT VT = N->getValueType(0);
4961 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4962 unsigned NumElts = VT.getVectorNumElements();
4963 MVT TruncVT = MVT::getIntegerVT(EltSize);
4964 SmallVector<SDValue, 8> Ops;
4965 for (unsigned i = 0; i != NumElts; ++i) {
4966 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4967 const APInt &CInt = C->getAPIntValue();
4968 // Element types smaller than 32 bits are not legal, so use i32 elements.
4969 // The values are implicitly truncated so sext vs. zext doesn't matter.
4970 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4972 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4973 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4976 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4977 unsigned Opcode = N->getOpcode();
4978 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4979 SDNode *N0 = N->getOperand(0).getNode();
4980 SDNode *N1 = N->getOperand(1).getNode();
4981 return N0->hasOneUse() && N1->hasOneUse() &&
4982 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4987 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4988 unsigned Opcode = N->getOpcode();
4989 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4990 SDNode *N0 = N->getOperand(0).getNode();
4991 SDNode *N1 = N->getOperand(1).getNode();
4992 return N0->hasOneUse() && N1->hasOneUse() &&
4993 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4998 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4999 // Multiplications are only custom-lowered for 128-bit vectors so that
5000 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5001 EVT VT = Op.getValueType();
5002 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
5003 SDNode *N0 = Op.getOperand(0).getNode();
5004 SDNode *N1 = Op.getOperand(1).getNode();
5005 unsigned NewOpc = 0;
5007 bool isN0SExt = isSignExtended(N0, DAG);
5008 bool isN1SExt = isSignExtended(N1, DAG);
5009 if (isN0SExt && isN1SExt)
5010 NewOpc = ARMISD::VMULLs;
5012 bool isN0ZExt = isZeroExtended(N0, DAG);
5013 bool isN1ZExt = isZeroExtended(N1, DAG);
5014 if (isN0ZExt && isN1ZExt)
5015 NewOpc = ARMISD::VMULLu;
5016 else if (isN1SExt || isN1ZExt) {
5017 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5018 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5019 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5020 NewOpc = ARMISD::VMULLs;
5022 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5023 NewOpc = ARMISD::VMULLu;
5025 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5027 NewOpc = ARMISD::VMULLu;
5033 if (VT == MVT::v2i64)
5034 // Fall through to expand this. It is not legal.
5037 // Other vector multiplications are legal.
5042 // Legalize to a VMULL instruction.
5043 DebugLoc DL = Op.getDebugLoc();
5045 SDValue Op1 = SkipExtension(N1, DAG);
5047 Op0 = SkipExtension(N0, DAG);
5048 assert(Op0.getValueType().is64BitVector() &&
5049 Op1.getValueType().is64BitVector() &&
5050 "unexpected types for extended operands to VMULL");
5051 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5054 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5055 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5062 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5063 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5064 EVT Op1VT = Op1.getValueType();
5065 return DAG.getNode(N0->getOpcode(), DL, VT,
5066 DAG.getNode(NewOpc, DL, VT,
5067 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5068 DAG.getNode(NewOpc, DL, VT,
5069 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5073 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5075 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5076 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5077 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5078 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5079 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5080 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5081 // Get reciprocal estimate.
5082 // float4 recip = vrecpeq_f32(yf);
5083 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5084 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5085 // Because char has a smaller range than uchar, we can actually get away
5086 // without any newton steps. This requires that we use a weird bias
5087 // of 0xb000, however (again, this has been exhaustively tested).
5088 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5089 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5090 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5091 Y = DAG.getConstant(0xb000, MVT::i32);
5092 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5093 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5094 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5095 // Convert back to short.
5096 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5097 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5102 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5104 // Convert to float.
5105 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5106 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5107 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5108 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5109 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5110 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5112 // Use reciprocal estimate and one refinement step.
5113 // float4 recip = vrecpeq_f32(yf);
5114 // recip *= vrecpsq_f32(yf, recip);
5115 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5116 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5117 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5118 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5120 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5121 // Because short has a smaller range than ushort, we can actually get away
5122 // with only a single newton step. This requires that we use a weird bias
5123 // of 89, however (again, this has been exhaustively tested).
5124 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5125 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5126 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5127 N1 = DAG.getConstant(0x89, MVT::i32);
5128 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5129 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5130 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5131 // Convert back to integer and return.
5132 // return vmovn_s32(vcvt_s32_f32(result));
5133 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5134 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5138 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5139 EVT VT = Op.getValueType();
5140 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5141 "unexpected type for custom-lowering ISD::SDIV");
5143 DebugLoc dl = Op.getDebugLoc();
5144 SDValue N0 = Op.getOperand(0);
5145 SDValue N1 = Op.getOperand(1);
5148 if (VT == MVT::v8i8) {
5149 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5150 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5152 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5153 DAG.getIntPtrConstant(4));
5154 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5155 DAG.getIntPtrConstant(4));
5156 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5157 DAG.getIntPtrConstant(0));
5158 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5159 DAG.getIntPtrConstant(0));
5161 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5162 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5164 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5165 N0 = LowerCONCAT_VECTORS(N0, DAG);
5167 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5170 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5173 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5174 EVT VT = Op.getValueType();
5175 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5176 "unexpected type for custom-lowering ISD::UDIV");
5178 DebugLoc dl = Op.getDebugLoc();
5179 SDValue N0 = Op.getOperand(0);
5180 SDValue N1 = Op.getOperand(1);
5183 if (VT == MVT::v8i8) {
5184 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5185 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5187 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5188 DAG.getIntPtrConstant(4));
5189 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5190 DAG.getIntPtrConstant(4));
5191 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5192 DAG.getIntPtrConstant(0));
5193 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5194 DAG.getIntPtrConstant(0));
5196 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5197 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5199 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5200 N0 = LowerCONCAT_VECTORS(N0, DAG);
5202 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5203 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5208 // v4i16 sdiv ... Convert to float.
5209 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5210 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5211 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5212 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5213 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5214 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5216 // Use reciprocal estimate and two refinement steps.
5217 // float4 recip = vrecpeq_f32(yf);
5218 // recip *= vrecpsq_f32(yf, recip);
5219 // recip *= vrecpsq_f32(yf, recip);
5220 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5221 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5222 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5223 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5225 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5226 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5227 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5229 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5230 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5231 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5232 // and that it will never cause us to return an answer too large).
5233 // float4 result = as_float4(as_int4(xf*recip) + 2);
5234 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5235 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5236 N1 = DAG.getConstant(2, MVT::i32);
5237 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5238 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5239 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5240 // Convert back to integer and return.
5241 // return vmovn_u32(vcvt_s32_f32(result));
5242 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5243 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5247 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5248 EVT VT = Op.getNode()->getValueType(0);
5249 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5252 bool ExtraOp = false;
5253 switch (Op.getOpcode()) {
5254 default: llvm_unreachable("Invalid code");
5255 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5256 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5257 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5258 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5262 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5264 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5265 Op.getOperand(1), Op.getOperand(2));
5268 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5269 // Monotonic load/store is legal for all targets
5270 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5273 // Aquire/Release load/store is not legal for targets without a
5274 // dmb or equivalent available.
5280 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5281 SelectionDAG &DAG, unsigned NewOp) {
5282 DebugLoc dl = Node->getDebugLoc();
5283 assert (Node->getValueType(0) == MVT::i64 &&
5284 "Only know how to expand i64 atomics");
5286 SmallVector<SDValue, 6> Ops;
5287 Ops.push_back(Node->getOperand(0)); // Chain
5288 Ops.push_back(Node->getOperand(1)); // Ptr
5290 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5291 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5292 // High part of Val1
5293 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5294 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5295 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5296 // High part of Val1
5297 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5298 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5299 // High part of Val2
5300 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5301 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5303 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5305 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5306 cast<MemSDNode>(Node)->getMemOperand());
5307 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5308 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5309 Results.push_back(Result.getValue(2));
5312 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5313 switch (Op.getOpcode()) {
5314 default: llvm_unreachable("Don't know how to custom lower this!");
5315 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5316 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5317 case ISD::GlobalAddress:
5318 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5319 LowerGlobalAddressELF(Op, DAG);
5320 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5321 case ISD::SELECT: return LowerSELECT(Op, DAG);
5322 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5323 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5324 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5325 case ISD::VASTART: return LowerVASTART(Op, DAG);
5326 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5327 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5328 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5329 case ISD::SINT_TO_FP:
5330 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5331 case ISD::FP_TO_SINT:
5332 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5333 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5334 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5335 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5336 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5337 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5338 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5339 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5341 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5344 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5345 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5346 case ISD::SRL_PARTS:
5347 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5348 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5349 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5350 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5351 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5352 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5353 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5354 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5355 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5356 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5357 case ISD::MUL: return LowerMUL(Op, DAG);
5358 case ISD::SDIV: return LowerSDIV(Op, DAG);
5359 case ISD::UDIV: return LowerUDIV(Op, DAG);
5363 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5364 case ISD::ATOMIC_LOAD:
5365 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5369 /// ReplaceNodeResults - Replace the results of node with an illegal result
5370 /// type with new values built out of custom code.
5371 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5372 SmallVectorImpl<SDValue>&Results,
5373 SelectionDAG &DAG) const {
5375 switch (N->getOpcode()) {
5377 llvm_unreachable("Don't know how to custom expand this!");
5379 Res = ExpandBITCAST(N, DAG);
5383 Res = Expand64BitShift(N, DAG, Subtarget);
5385 case ISD::ATOMIC_LOAD_ADD:
5386 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5388 case ISD::ATOMIC_LOAD_AND:
5389 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5391 case ISD::ATOMIC_LOAD_NAND:
5392 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5394 case ISD::ATOMIC_LOAD_OR:
5395 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5397 case ISD::ATOMIC_LOAD_SUB:
5398 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5400 case ISD::ATOMIC_LOAD_XOR:
5401 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5403 case ISD::ATOMIC_SWAP:
5404 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5406 case ISD::ATOMIC_CMP_SWAP:
5407 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5411 Results.push_back(Res);
5414 //===----------------------------------------------------------------------===//
5415 // ARM Scheduler Hooks
5416 //===----------------------------------------------------------------------===//
5419 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5420 MachineBasicBlock *BB,
5421 unsigned Size) const {
5422 unsigned dest = MI->getOperand(0).getReg();
5423 unsigned ptr = MI->getOperand(1).getReg();
5424 unsigned oldval = MI->getOperand(2).getReg();
5425 unsigned newval = MI->getOperand(3).getReg();
5426 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5427 DebugLoc dl = MI->getDebugLoc();
5428 bool isThumb2 = Subtarget->isThumb2();
5430 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5431 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5432 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5433 (const TargetRegisterClass*)&ARM::GPRRegClass);
5436 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5437 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5438 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5441 unsigned ldrOpc, strOpc;
5443 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5445 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5446 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5449 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5450 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5453 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5454 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5458 MachineFunction *MF = BB->getParent();
5459 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5460 MachineFunction::iterator It = BB;
5461 ++It; // insert the new blocks after the current block
5463 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5464 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5465 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5466 MF->insert(It, loop1MBB);
5467 MF->insert(It, loop2MBB);
5468 MF->insert(It, exitMBB);
5470 // Transfer the remainder of BB and its successor edges to exitMBB.
5471 exitMBB->splice(exitMBB->begin(), BB,
5472 llvm::next(MachineBasicBlock::iterator(MI)),
5474 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5478 // fallthrough --> loop1MBB
5479 BB->addSuccessor(loop1MBB);
5482 // ldrex dest, [ptr]
5486 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5487 if (ldrOpc == ARM::t2LDREX)
5489 AddDefaultPred(MIB);
5490 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5491 .addReg(dest).addReg(oldval));
5492 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5493 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5494 BB->addSuccessor(loop2MBB);
5495 BB->addSuccessor(exitMBB);
5498 // strex scratch, newval, [ptr]
5502 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5503 if (strOpc == ARM::t2STREX)
5505 AddDefaultPred(MIB);
5506 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5507 .addReg(scratch).addImm(0));
5508 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5509 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5510 BB->addSuccessor(loop1MBB);
5511 BB->addSuccessor(exitMBB);
5517 MI->eraseFromParent(); // The instruction is gone now.
5523 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5524 unsigned Size, unsigned BinOpcode) const {
5525 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5526 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5528 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5529 MachineFunction *MF = BB->getParent();
5530 MachineFunction::iterator It = BB;
5533 unsigned dest = MI->getOperand(0).getReg();
5534 unsigned ptr = MI->getOperand(1).getReg();
5535 unsigned incr = MI->getOperand(2).getReg();
5536 DebugLoc dl = MI->getDebugLoc();
5537 bool isThumb2 = Subtarget->isThumb2();
5539 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5541 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5542 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5545 unsigned ldrOpc, strOpc;
5547 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5549 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5550 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5553 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5554 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5557 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5558 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5562 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5563 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5564 MF->insert(It, loopMBB);
5565 MF->insert(It, exitMBB);
5567 // Transfer the remainder of BB and its successor edges to exitMBB.
5568 exitMBB->splice(exitMBB->begin(), BB,
5569 llvm::next(MachineBasicBlock::iterator(MI)),
5571 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5573 const TargetRegisterClass *TRC = isThumb2 ?
5574 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5575 (const TargetRegisterClass*)&ARM::GPRRegClass;
5576 unsigned scratch = MRI.createVirtualRegister(TRC);
5577 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5581 // fallthrough --> loopMBB
5582 BB->addSuccessor(loopMBB);
5586 // <binop> scratch2, dest, incr
5587 // strex scratch, scratch2, ptr
5590 // fallthrough --> exitMBB
5592 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5593 if (ldrOpc == ARM::t2LDREX)
5595 AddDefaultPred(MIB);
5597 // operand order needs to go the other way for NAND
5598 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5599 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5600 addReg(incr).addReg(dest)).addReg(0);
5602 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5603 addReg(dest).addReg(incr)).addReg(0);
5606 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5607 if (strOpc == ARM::t2STREX)
5609 AddDefaultPred(MIB);
5610 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5611 .addReg(scratch).addImm(0));
5612 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5613 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5615 BB->addSuccessor(loopMBB);
5616 BB->addSuccessor(exitMBB);
5622 MI->eraseFromParent(); // The instruction is gone now.
5628 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5629 MachineBasicBlock *BB,
5632 ARMCC::CondCodes Cond) const {
5633 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5635 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5636 MachineFunction *MF = BB->getParent();
5637 MachineFunction::iterator It = BB;
5640 unsigned dest = MI->getOperand(0).getReg();
5641 unsigned ptr = MI->getOperand(1).getReg();
5642 unsigned incr = MI->getOperand(2).getReg();
5643 unsigned oldval = dest;
5644 DebugLoc dl = MI->getDebugLoc();
5645 bool isThumb2 = Subtarget->isThumb2();
5647 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5649 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5650 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5653 unsigned ldrOpc, strOpc, extendOpc;
5655 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5657 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5658 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5659 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5662 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5663 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5664 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5667 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5668 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5673 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5674 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5675 MF->insert(It, loopMBB);
5676 MF->insert(It, exitMBB);
5678 // Transfer the remainder of BB and its successor edges to exitMBB.
5679 exitMBB->splice(exitMBB->begin(), BB,
5680 llvm::next(MachineBasicBlock::iterator(MI)),
5682 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5684 const TargetRegisterClass *TRC = isThumb2 ?
5685 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5686 (const TargetRegisterClass*)&ARM::GPRRegClass;
5687 unsigned scratch = MRI.createVirtualRegister(TRC);
5688 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5692 // fallthrough --> loopMBB
5693 BB->addSuccessor(loopMBB);
5697 // (sign extend dest, if required)
5699 // cmov.cond scratch2, incr, dest
5700 // strex scratch, scratch2, ptr
5703 // fallthrough --> exitMBB
5705 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5706 if (ldrOpc == ARM::t2LDREX)
5708 AddDefaultPred(MIB);
5710 // Sign extend the value, if necessary.
5711 if (signExtend && extendOpc) {
5712 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5713 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5718 // Build compare and cmov instructions.
5719 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5720 .addReg(oldval).addReg(incr));
5721 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5722 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5724 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5725 if (strOpc == ARM::t2STREX)
5727 AddDefaultPred(MIB);
5728 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5729 .addReg(scratch).addImm(0));
5730 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5731 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5733 BB->addSuccessor(loopMBB);
5734 BB->addSuccessor(exitMBB);
5740 MI->eraseFromParent(); // The instruction is gone now.
5746 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5747 unsigned Op1, unsigned Op2,
5748 bool NeedsCarry, bool IsCmpxchg) const {
5749 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5750 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5752 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5753 MachineFunction *MF = BB->getParent();
5754 MachineFunction::iterator It = BB;
5757 unsigned destlo = MI->getOperand(0).getReg();
5758 unsigned desthi = MI->getOperand(1).getReg();
5759 unsigned ptr = MI->getOperand(2).getReg();
5760 unsigned vallo = MI->getOperand(3).getReg();
5761 unsigned valhi = MI->getOperand(4).getReg();
5762 DebugLoc dl = MI->getDebugLoc();
5763 bool isThumb2 = Subtarget->isThumb2();
5765 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5767 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5768 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5769 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5772 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5773 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5775 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5776 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5778 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5779 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5781 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5782 MF->insert(It, loopMBB);
5784 MF->insert(It, contBB);
5785 MF->insert(It, cont2BB);
5787 MF->insert(It, exitMBB);
5789 // Transfer the remainder of BB and its successor edges to exitMBB.
5790 exitMBB->splice(exitMBB->begin(), BB,
5791 llvm::next(MachineBasicBlock::iterator(MI)),
5793 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5795 const TargetRegisterClass *TRC = isThumb2 ?
5796 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5797 (const TargetRegisterClass*)&ARM::GPRRegClass;
5798 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5802 // fallthrough --> loopMBB
5803 BB->addSuccessor(loopMBB);
5806 // ldrexd r2, r3, ptr
5807 // <binopa> r0, r2, incr
5808 // <binopb> r1, r3, incr
5809 // strexd storesuccess, r0, r1, ptr
5810 // cmp storesuccess, #0
5812 // fallthrough --> exitMBB
5814 // Note that the registers are explicitly specified because there is not any
5815 // way to force the register allocator to allocate a register pair.
5817 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5818 // need to properly enforce the restriction that the two output registers
5819 // for ldrexd must be different.
5822 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5823 unsigned GPRPair1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5825 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5826 .addReg(GPRPair0, RegState::Define).addReg(ptr));
5827 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5828 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
5829 .addReg(GPRPair0, 0, ARM::gsub_0);
5830 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
5831 .addReg(GPRPair0, 0, ARM::gsub_1);
5835 for (unsigned i = 0; i < 2; i++) {
5836 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5838 .addReg(i == 0 ? destlo : desthi)
5839 .addReg(i == 0 ? vallo : valhi));
5840 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5841 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5842 BB->addSuccessor(exitMBB);
5843 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5844 BB = (i == 0 ? contBB : cont2BB);
5847 // Copy to physregs for strexd
5848 unsigned setlo = MI->getOperand(5).getReg();
5849 unsigned sethi = MI->getOperand(6).getReg();
5850 unsigned undef = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5851 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5852 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), undef);
5853 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5856 .addImm(ARM::gsub_0);
5857 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5860 .addImm(ARM::gsub_1);
5862 // Perform binary operation
5863 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
5864 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
5865 .addReg(destlo).addReg(vallo))
5866 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5867 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
5868 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
5869 .addReg(desthi).addReg(valhi)).addReg(0);
5871 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5872 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
5873 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5874 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5877 .addImm(ARM::gsub_0);
5878 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5881 .addImm(ARM::gsub_1);
5883 // Copy to physregs for strexd
5884 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5885 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5886 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
5887 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5890 .addImm(ARM::gsub_0);
5891 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5894 .addImm(ARM::gsub_1);
5898 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5899 .addReg(GPRPair1).addReg(ptr));
5901 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5902 .addReg(storesuccess).addImm(0));
5903 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5904 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5906 BB->addSuccessor(loopMBB);
5907 BB->addSuccessor(exitMBB);
5913 MI->eraseFromParent(); // The instruction is gone now.
5918 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5919 /// registers the function context.
5920 void ARMTargetLowering::
5921 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5922 MachineBasicBlock *DispatchBB, int FI) const {
5923 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5924 DebugLoc dl = MI->getDebugLoc();
5925 MachineFunction *MF = MBB->getParent();
5926 MachineRegisterInfo *MRI = &MF->getRegInfo();
5927 MachineConstantPool *MCP = MF->getConstantPool();
5928 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5929 const Function *F = MF->getFunction();
5931 bool isThumb = Subtarget->isThumb();
5932 bool isThumb2 = Subtarget->isThumb2();
5934 unsigned PCLabelId = AFI->createPICLabelUId();
5935 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5936 ARMConstantPoolValue *CPV =
5937 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5938 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5940 const TargetRegisterClass *TRC = isThumb ?
5941 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5942 (const TargetRegisterClass*)&ARM::GPRRegClass;
5944 // Grab constant pool and fixed stack memory operands.
5945 MachineMemOperand *CPMMO =
5946 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5947 MachineMemOperand::MOLoad, 4, 4);
5949 MachineMemOperand *FIMMOSt =
5950 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5951 MachineMemOperand::MOStore, 4, 4);
5953 // Load the address of the dispatch MBB into the jump buffer.
5955 // Incoming value: jbuf
5956 // ldr.n r5, LCPI1_1
5959 // str r5, [$jbuf, #+4] ; &jbuf[1]
5960 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5961 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5962 .addConstantPoolIndex(CPI)
5963 .addMemOperand(CPMMO));
5964 // Set the low bit because of thumb mode.
5965 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5967 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5968 .addReg(NewVReg1, RegState::Kill)
5970 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5971 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5972 .addReg(NewVReg2, RegState::Kill)
5974 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5975 .addReg(NewVReg3, RegState::Kill)
5977 .addImm(36) // &jbuf[1] :: pc
5978 .addMemOperand(FIMMOSt));
5979 } else if (isThumb) {
5980 // Incoming value: jbuf
5981 // ldr.n r1, LCPI1_4
5985 // add r2, $jbuf, #+4 ; &jbuf[1]
5987 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5988 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5989 .addConstantPoolIndex(CPI)
5990 .addMemOperand(CPMMO));
5991 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5992 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5993 .addReg(NewVReg1, RegState::Kill)
5995 // Set the low bit because of thumb mode.
5996 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5997 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5998 .addReg(ARM::CPSR, RegState::Define)
6000 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6001 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6002 .addReg(ARM::CPSR, RegState::Define)
6003 .addReg(NewVReg2, RegState::Kill)
6004 .addReg(NewVReg3, RegState::Kill));
6005 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6006 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6008 .addImm(36)); // &jbuf[1] :: pc
6009 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6010 .addReg(NewVReg4, RegState::Kill)
6011 .addReg(NewVReg5, RegState::Kill)
6013 .addMemOperand(FIMMOSt));
6015 // Incoming value: jbuf
6018 // str r1, [$jbuf, #+4] ; &jbuf[1]
6019 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6020 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6021 .addConstantPoolIndex(CPI)
6023 .addMemOperand(CPMMO));
6024 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6025 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6026 .addReg(NewVReg1, RegState::Kill)
6027 .addImm(PCLabelId));
6028 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6029 .addReg(NewVReg2, RegState::Kill)
6031 .addImm(36) // &jbuf[1] :: pc
6032 .addMemOperand(FIMMOSt));
6036 MachineBasicBlock *ARMTargetLowering::
6037 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6038 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6039 DebugLoc dl = MI->getDebugLoc();
6040 MachineFunction *MF = MBB->getParent();
6041 MachineRegisterInfo *MRI = &MF->getRegInfo();
6042 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6043 MachineFrameInfo *MFI = MF->getFrameInfo();
6044 int FI = MFI->getFunctionContextIndex();
6046 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6047 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6048 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6050 // Get a mapping of the call site numbers to all of the landing pads they're
6052 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6053 unsigned MaxCSNum = 0;
6054 MachineModuleInfo &MMI = MF->getMMI();
6055 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6057 if (!BB->isLandingPad()) continue;
6059 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6061 for (MachineBasicBlock::iterator
6062 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6063 if (!II->isEHLabel()) continue;
6065 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6066 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6068 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6069 for (SmallVectorImpl<unsigned>::iterator
6070 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6071 CSI != CSE; ++CSI) {
6072 CallSiteNumToLPad[*CSI].push_back(BB);
6073 MaxCSNum = std::max(MaxCSNum, *CSI);
6079 // Get an ordered list of the machine basic blocks for the jump table.
6080 std::vector<MachineBasicBlock*> LPadList;
6081 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6082 LPadList.reserve(CallSiteNumToLPad.size());
6083 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6084 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6085 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6086 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6087 LPadList.push_back(*II);
6088 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6092 assert(!LPadList.empty() &&
6093 "No landing pad destinations for the dispatch jump table!");
6095 // Create the jump table and associated information.
6096 MachineJumpTableInfo *JTI =
6097 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6098 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6099 unsigned UId = AFI->createJumpTableUId();
6101 // Create the MBBs for the dispatch code.
6103 // Shove the dispatch's address into the return slot in the function context.
6104 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6105 DispatchBB->setIsLandingPad();
6107 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6108 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6109 DispatchBB->addSuccessor(TrapBB);
6111 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6112 DispatchBB->addSuccessor(DispContBB);
6115 MF->insert(MF->end(), DispatchBB);
6116 MF->insert(MF->end(), DispContBB);
6117 MF->insert(MF->end(), TrapBB);
6119 // Insert code into the entry block that creates and registers the function
6121 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6123 MachineMemOperand *FIMMOLd =
6124 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6125 MachineMemOperand::MOLoad |
6126 MachineMemOperand::MOVolatile, 4, 4);
6128 MachineInstrBuilder MIB;
6129 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6131 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6132 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6134 // Add a register mask with no preserved registers. This results in all
6135 // registers being marked as clobbered.
6136 MIB.addRegMask(RI.getNoPreservedMask());
6138 unsigned NumLPads = LPadList.size();
6139 if (Subtarget->isThumb2()) {
6140 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6141 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6144 .addMemOperand(FIMMOLd));
6146 if (NumLPads < 256) {
6147 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6149 .addImm(LPadList.size()));
6151 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6152 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6153 .addImm(NumLPads & 0xFFFF));
6155 unsigned VReg2 = VReg1;
6156 if ((NumLPads & 0xFFFF0000) != 0) {
6157 VReg2 = MRI->createVirtualRegister(TRC);
6158 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6160 .addImm(NumLPads >> 16));
6163 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6168 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6173 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6174 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6175 .addJumpTableIndex(MJTI)
6178 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6181 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6182 .addReg(NewVReg3, RegState::Kill)
6184 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6186 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6187 .addReg(NewVReg4, RegState::Kill)
6189 .addJumpTableIndex(MJTI)
6191 } else if (Subtarget->isThumb()) {
6192 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6193 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6196 .addMemOperand(FIMMOLd));
6198 if (NumLPads < 256) {
6199 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6203 MachineConstantPool *ConstantPool = MF->getConstantPool();
6204 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6205 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6207 // MachineConstantPool wants an explicit alignment.
6208 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6210 Align = getDataLayout()->getTypeAllocSize(C->getType());
6211 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6213 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6214 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6215 .addReg(VReg1, RegState::Define)
6216 .addConstantPoolIndex(Idx));
6217 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6222 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6227 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6228 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6229 .addReg(ARM::CPSR, RegState::Define)
6233 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6234 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6235 .addJumpTableIndex(MJTI)
6238 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6239 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6240 .addReg(ARM::CPSR, RegState::Define)
6241 .addReg(NewVReg2, RegState::Kill)
6244 MachineMemOperand *JTMMOLd =
6245 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6246 MachineMemOperand::MOLoad, 4, 4);
6248 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6249 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6250 .addReg(NewVReg4, RegState::Kill)
6252 .addMemOperand(JTMMOLd));
6254 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6255 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6256 .addReg(ARM::CPSR, RegState::Define)
6257 .addReg(NewVReg5, RegState::Kill)
6260 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6261 .addReg(NewVReg6, RegState::Kill)
6262 .addJumpTableIndex(MJTI)
6265 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6266 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6269 .addMemOperand(FIMMOLd));
6271 if (NumLPads < 256) {
6272 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6275 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6276 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6277 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6278 .addImm(NumLPads & 0xFFFF));
6280 unsigned VReg2 = VReg1;
6281 if ((NumLPads & 0xFFFF0000) != 0) {
6282 VReg2 = MRI->createVirtualRegister(TRC);
6283 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6285 .addImm(NumLPads >> 16));
6288 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6292 MachineConstantPool *ConstantPool = MF->getConstantPool();
6293 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6294 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6296 // MachineConstantPool wants an explicit alignment.
6297 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6299 Align = getDataLayout()->getTypeAllocSize(C->getType());
6300 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6302 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6303 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6304 .addReg(VReg1, RegState::Define)
6305 .addConstantPoolIndex(Idx)
6307 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6309 .addReg(VReg1, RegState::Kill));
6312 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6317 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6319 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6321 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6322 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6323 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6324 .addJumpTableIndex(MJTI)
6327 MachineMemOperand *JTMMOLd =
6328 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6329 MachineMemOperand::MOLoad, 4, 4);
6330 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6332 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6333 .addReg(NewVReg3, RegState::Kill)
6336 .addMemOperand(JTMMOLd));
6338 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6339 .addReg(NewVReg5, RegState::Kill)
6341 .addJumpTableIndex(MJTI)
6345 // Add the jump table entries as successors to the MBB.
6346 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6347 for (std::vector<MachineBasicBlock*>::iterator
6348 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6349 MachineBasicBlock *CurMBB = *I;
6350 if (SeenMBBs.insert(CurMBB))
6351 DispContBB->addSuccessor(CurMBB);
6354 // N.B. the order the invoke BBs are processed in doesn't matter here.
6355 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6356 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6357 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6358 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6359 MachineBasicBlock *BB = *I;
6361 // Remove the landing pad successor from the invoke block and replace it
6362 // with the new dispatch block.
6363 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6365 while (!Successors.empty()) {
6366 MachineBasicBlock *SMBB = Successors.pop_back_val();
6367 if (SMBB->isLandingPad()) {
6368 BB->removeSuccessor(SMBB);
6369 MBBLPads.push_back(SMBB);
6373 BB->addSuccessor(DispatchBB);
6375 // Find the invoke call and mark all of the callee-saved registers as
6376 // 'implicit defined' so that they're spilled. This prevents code from
6377 // moving instructions to before the EH block, where they will never be
6379 for (MachineBasicBlock::reverse_iterator
6380 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6381 if (!II->isCall()) continue;
6383 DenseMap<unsigned, bool> DefRegs;
6384 for (MachineInstr::mop_iterator
6385 OI = II->operands_begin(), OE = II->operands_end();
6387 if (!OI->isReg()) continue;
6388 DefRegs[OI->getReg()] = true;
6391 MachineInstrBuilder MIB(&*II);
6393 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6394 unsigned Reg = SavedRegs[i];
6395 if (Subtarget->isThumb2() &&
6396 !ARM::tGPRRegClass.contains(Reg) &&
6397 !ARM::hGPRRegClass.contains(Reg))
6399 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6401 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6404 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6411 // Mark all former landing pads as non-landing pads. The dispatch is the only
6413 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6414 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6415 (*I)->setIsLandingPad(false);
6417 // The instruction is gone now.
6418 MI->eraseFromParent();
6424 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6425 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6426 E = MBB->succ_end(); I != E; ++I)
6429 llvm_unreachable("Expecting a BB with two successors!");
6432 MachineBasicBlock *ARMTargetLowering::
6433 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6434 // This pseudo instruction has 3 operands: dst, src, size
6435 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6436 // Otherwise, we will generate unrolled scalar copies.
6437 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6438 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6439 MachineFunction::iterator It = BB;
6442 unsigned dest = MI->getOperand(0).getReg();
6443 unsigned src = MI->getOperand(1).getReg();
6444 unsigned SizeVal = MI->getOperand(2).getImm();
6445 unsigned Align = MI->getOperand(3).getImm();
6446 DebugLoc dl = MI->getDebugLoc();
6448 bool isThumb2 = Subtarget->isThumb2();
6449 MachineFunction *MF = BB->getParent();
6450 MachineRegisterInfo &MRI = MF->getRegInfo();
6451 unsigned ldrOpc, strOpc, UnitSize = 0;
6453 const TargetRegisterClass *TRC = isThumb2 ?
6454 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6455 (const TargetRegisterClass*)&ARM::GPRRegClass;
6456 const TargetRegisterClass *TRC_Vec = 0;
6459 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6460 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6462 } else if (Align & 2) {
6463 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6464 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6467 // Check whether we can use NEON instructions.
6468 if (!MF->getFunction()->getFnAttributes().
6469 hasAttribute(Attributes::NoImplicitFloat) &&
6470 Subtarget->hasNEON()) {
6471 if ((Align % 16 == 0) && SizeVal >= 16) {
6472 ldrOpc = ARM::VLD1q32wb_fixed;
6473 strOpc = ARM::VST1q32wb_fixed;
6475 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6477 else if ((Align % 8 == 0) && SizeVal >= 8) {
6478 ldrOpc = ARM::VLD1d32wb_fixed;
6479 strOpc = ARM::VST1d32wb_fixed;
6481 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6484 // Can't use NEON instructions.
6485 if (UnitSize == 0) {
6486 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6487 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6492 unsigned BytesLeft = SizeVal % UnitSize;
6493 unsigned LoopSize = SizeVal - BytesLeft;
6495 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6496 // Use LDR and STR to copy.
6497 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6498 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6499 unsigned srcIn = src;
6500 unsigned destIn = dest;
6501 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6502 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6503 unsigned srcOut = MRI.createVirtualRegister(TRC);
6504 unsigned destOut = MRI.createVirtualRegister(TRC);
6505 if (UnitSize >= 8) {
6506 AddDefaultPred(BuildMI(*BB, MI, dl,
6507 TII->get(ldrOpc), scratch)
6508 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6510 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6511 .addReg(destIn).addImm(0).addReg(scratch));
6512 } else if (isThumb2) {
6513 AddDefaultPred(BuildMI(*BB, MI, dl,
6514 TII->get(ldrOpc), scratch)
6515 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6517 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6518 .addReg(scratch).addReg(destIn)
6521 AddDefaultPred(BuildMI(*BB, MI, dl,
6522 TII->get(ldrOpc), scratch)
6523 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6526 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6527 .addReg(scratch).addReg(destIn)
6528 .addReg(0).addImm(UnitSize));
6534 // Handle the leftover bytes with LDRB and STRB.
6535 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6536 // [destOut] = STRB_POST(scratch, destIn, 1)
6537 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6538 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6539 for (unsigned i = 0; i < BytesLeft; i++) {
6540 unsigned scratch = MRI.createVirtualRegister(TRC);
6541 unsigned srcOut = MRI.createVirtualRegister(TRC);
6542 unsigned destOut = MRI.createVirtualRegister(TRC);
6544 AddDefaultPred(BuildMI(*BB, MI, dl,
6545 TII->get(ldrOpc),scratch)
6546 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6548 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6549 .addReg(scratch).addReg(destIn)
6550 .addReg(0).addImm(1));
6552 AddDefaultPred(BuildMI(*BB, MI, dl,
6553 TII->get(ldrOpc),scratch)
6554 .addReg(srcOut, RegState::Define).addReg(srcIn)
6555 .addReg(0).addImm(1));
6557 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6558 .addReg(scratch).addReg(destIn)
6559 .addReg(0).addImm(1));
6564 MI->eraseFromParent(); // The instruction is gone now.
6568 // Expand the pseudo op to a loop.
6571 // movw varEnd, # --> with thumb2
6573 // ldrcp varEnd, idx --> without thumb2
6574 // fallthrough --> loopMBB
6576 // PHI varPhi, varEnd, varLoop
6577 // PHI srcPhi, src, srcLoop
6578 // PHI destPhi, dst, destLoop
6579 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6580 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6581 // subs varLoop, varPhi, #UnitSize
6583 // fallthrough --> exitMBB
6585 // epilogue to handle left-over bytes
6586 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6587 // [destOut] = STRB_POST(scratch, destLoop, 1)
6588 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6589 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6590 MF->insert(It, loopMBB);
6591 MF->insert(It, exitMBB);
6593 // Transfer the remainder of BB and its successor edges to exitMBB.
6594 exitMBB->splice(exitMBB->begin(), BB,
6595 llvm::next(MachineBasicBlock::iterator(MI)),
6597 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6599 // Load an immediate to varEnd.
6600 unsigned varEnd = MRI.createVirtualRegister(TRC);
6602 unsigned VReg1 = varEnd;
6603 if ((LoopSize & 0xFFFF0000) != 0)
6604 VReg1 = MRI.createVirtualRegister(TRC);
6605 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6606 .addImm(LoopSize & 0xFFFF));
6608 if ((LoopSize & 0xFFFF0000) != 0)
6609 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6611 .addImm(LoopSize >> 16));
6613 MachineConstantPool *ConstantPool = MF->getConstantPool();
6614 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6615 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6617 // MachineConstantPool wants an explicit alignment.
6618 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6620 Align = getDataLayout()->getTypeAllocSize(C->getType());
6621 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6623 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6624 .addReg(varEnd, RegState::Define)
6625 .addConstantPoolIndex(Idx)
6628 BB->addSuccessor(loopMBB);
6630 // Generate the loop body:
6631 // varPhi = PHI(varLoop, varEnd)
6632 // srcPhi = PHI(srcLoop, src)
6633 // destPhi = PHI(destLoop, dst)
6634 MachineBasicBlock *entryBB = BB;
6636 unsigned varLoop = MRI.createVirtualRegister(TRC);
6637 unsigned varPhi = MRI.createVirtualRegister(TRC);
6638 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6639 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6640 unsigned destLoop = MRI.createVirtualRegister(TRC);
6641 unsigned destPhi = MRI.createVirtualRegister(TRC);
6643 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6644 .addReg(varLoop).addMBB(loopMBB)
6645 .addReg(varEnd).addMBB(entryBB);
6646 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6647 .addReg(srcLoop).addMBB(loopMBB)
6648 .addReg(src).addMBB(entryBB);
6649 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6650 .addReg(destLoop).addMBB(loopMBB)
6651 .addReg(dest).addMBB(entryBB);
6653 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6654 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6655 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6656 if (UnitSize >= 8) {
6657 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6658 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6660 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6661 .addReg(destPhi).addImm(0).addReg(scratch));
6662 } else if (isThumb2) {
6663 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6664 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6666 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6667 .addReg(scratch).addReg(destPhi)
6670 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6671 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6674 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6675 .addReg(scratch).addReg(destPhi)
6676 .addReg(0).addImm(UnitSize));
6679 // Decrement loop variable by UnitSize.
6680 MachineInstrBuilder MIB = BuildMI(BB, dl,
6681 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6682 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6683 MIB->getOperand(5).setReg(ARM::CPSR);
6684 MIB->getOperand(5).setIsDef(true);
6686 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6687 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6689 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6690 BB->addSuccessor(loopMBB);
6691 BB->addSuccessor(exitMBB);
6693 // Add epilogue to handle BytesLeft.
6695 MachineInstr *StartOfExit = exitMBB->begin();
6696 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6697 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6699 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6700 // [destOut] = STRB_POST(scratch, destLoop, 1)
6701 unsigned srcIn = srcLoop;
6702 unsigned destIn = destLoop;
6703 for (unsigned i = 0; i < BytesLeft; i++) {
6704 unsigned scratch = MRI.createVirtualRegister(TRC);
6705 unsigned srcOut = MRI.createVirtualRegister(TRC);
6706 unsigned destOut = MRI.createVirtualRegister(TRC);
6708 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6709 TII->get(ldrOpc),scratch)
6710 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6712 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6713 .addReg(scratch).addReg(destIn)
6716 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6717 TII->get(ldrOpc),scratch)
6718 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6720 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6721 .addReg(scratch).addReg(destIn)
6722 .addReg(0).addImm(1));
6728 MI->eraseFromParent(); // The instruction is gone now.
6733 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6734 MachineBasicBlock *BB) const {
6735 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6736 DebugLoc dl = MI->getDebugLoc();
6737 bool isThumb2 = Subtarget->isThumb2();
6738 switch (MI->getOpcode()) {
6741 llvm_unreachable("Unexpected instr type to insert");
6743 // The Thumb2 pre-indexed stores have the same MI operands, they just
6744 // define them differently in the .td files from the isel patterns, so
6745 // they need pseudos.
6746 case ARM::t2STR_preidx:
6747 MI->setDesc(TII->get(ARM::t2STR_PRE));
6749 case ARM::t2STRB_preidx:
6750 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6752 case ARM::t2STRH_preidx:
6753 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6756 case ARM::STRi_preidx:
6757 case ARM::STRBi_preidx: {
6758 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6759 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6760 // Decode the offset.
6761 unsigned Offset = MI->getOperand(4).getImm();
6762 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6763 Offset = ARM_AM::getAM2Offset(Offset);
6767 MachineMemOperand *MMO = *MI->memoperands_begin();
6768 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6769 .addOperand(MI->getOperand(0)) // Rn_wb
6770 .addOperand(MI->getOperand(1)) // Rt
6771 .addOperand(MI->getOperand(2)) // Rn
6772 .addImm(Offset) // offset (skip GPR==zero_reg)
6773 .addOperand(MI->getOperand(5)) // pred
6774 .addOperand(MI->getOperand(6))
6775 .addMemOperand(MMO);
6776 MI->eraseFromParent();
6779 case ARM::STRr_preidx:
6780 case ARM::STRBr_preidx:
6781 case ARM::STRH_preidx: {
6783 switch (MI->getOpcode()) {
6784 default: llvm_unreachable("unexpected opcode!");
6785 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6786 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6787 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6789 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6790 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6791 MIB.addOperand(MI->getOperand(i));
6792 MI->eraseFromParent();
6795 case ARM::ATOMIC_LOAD_ADD_I8:
6796 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6797 case ARM::ATOMIC_LOAD_ADD_I16:
6798 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6799 case ARM::ATOMIC_LOAD_ADD_I32:
6800 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6802 case ARM::ATOMIC_LOAD_AND_I8:
6803 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6804 case ARM::ATOMIC_LOAD_AND_I16:
6805 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6806 case ARM::ATOMIC_LOAD_AND_I32:
6807 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6809 case ARM::ATOMIC_LOAD_OR_I8:
6810 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6811 case ARM::ATOMIC_LOAD_OR_I16:
6812 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6813 case ARM::ATOMIC_LOAD_OR_I32:
6814 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6816 case ARM::ATOMIC_LOAD_XOR_I8:
6817 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6818 case ARM::ATOMIC_LOAD_XOR_I16:
6819 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6820 case ARM::ATOMIC_LOAD_XOR_I32:
6821 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6823 case ARM::ATOMIC_LOAD_NAND_I8:
6824 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6825 case ARM::ATOMIC_LOAD_NAND_I16:
6826 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6827 case ARM::ATOMIC_LOAD_NAND_I32:
6828 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6830 case ARM::ATOMIC_LOAD_SUB_I8:
6831 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6832 case ARM::ATOMIC_LOAD_SUB_I16:
6833 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6834 case ARM::ATOMIC_LOAD_SUB_I32:
6835 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6837 case ARM::ATOMIC_LOAD_MIN_I8:
6838 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6839 case ARM::ATOMIC_LOAD_MIN_I16:
6840 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6841 case ARM::ATOMIC_LOAD_MIN_I32:
6842 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6844 case ARM::ATOMIC_LOAD_MAX_I8:
6845 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6846 case ARM::ATOMIC_LOAD_MAX_I16:
6847 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6848 case ARM::ATOMIC_LOAD_MAX_I32:
6849 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6851 case ARM::ATOMIC_LOAD_UMIN_I8:
6852 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6853 case ARM::ATOMIC_LOAD_UMIN_I16:
6854 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6855 case ARM::ATOMIC_LOAD_UMIN_I32:
6856 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6858 case ARM::ATOMIC_LOAD_UMAX_I8:
6859 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6860 case ARM::ATOMIC_LOAD_UMAX_I16:
6861 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6862 case ARM::ATOMIC_LOAD_UMAX_I32:
6863 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6865 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6866 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6867 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6869 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6870 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6871 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6874 case ARM::ATOMADD6432:
6875 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6876 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6877 /*NeedsCarry*/ true);
6878 case ARM::ATOMSUB6432:
6879 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6880 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6881 /*NeedsCarry*/ true);
6882 case ARM::ATOMOR6432:
6883 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6884 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6885 case ARM::ATOMXOR6432:
6886 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6887 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6888 case ARM::ATOMAND6432:
6889 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6890 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6891 case ARM::ATOMSWAP6432:
6892 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6893 case ARM::ATOMCMPXCHG6432:
6894 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6895 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6896 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6898 case ARM::tMOVCCr_pseudo: {
6899 // To "insert" a SELECT_CC instruction, we actually have to insert the
6900 // diamond control-flow pattern. The incoming instruction knows the
6901 // destination vreg to set, the condition code register to branch on, the
6902 // true/false values to select between, and a branch opcode to use.
6903 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6904 MachineFunction::iterator It = BB;
6910 // cmpTY ccX, r1, r2
6912 // fallthrough --> copy0MBB
6913 MachineBasicBlock *thisMBB = BB;
6914 MachineFunction *F = BB->getParent();
6915 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6916 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6917 F->insert(It, copy0MBB);
6918 F->insert(It, sinkMBB);
6920 // Transfer the remainder of BB and its successor edges to sinkMBB.
6921 sinkMBB->splice(sinkMBB->begin(), BB,
6922 llvm::next(MachineBasicBlock::iterator(MI)),
6924 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6926 BB->addSuccessor(copy0MBB);
6927 BB->addSuccessor(sinkMBB);
6929 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6930 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6933 // %FalseValue = ...
6934 // # fallthrough to sinkMBB
6937 // Update machine-CFG edges
6938 BB->addSuccessor(sinkMBB);
6941 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6944 BuildMI(*BB, BB->begin(), dl,
6945 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6946 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6947 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6949 MI->eraseFromParent(); // The pseudo instruction is gone now.
6954 case ARM::BCCZi64: {
6955 // If there is an unconditional branch to the other successor, remove it.
6956 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6958 // Compare both parts that make up the double comparison separately for
6960 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6962 unsigned LHS1 = MI->getOperand(1).getReg();
6963 unsigned LHS2 = MI->getOperand(2).getReg();
6965 AddDefaultPred(BuildMI(BB, dl,
6966 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6967 .addReg(LHS1).addImm(0));
6968 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6969 .addReg(LHS2).addImm(0)
6970 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6972 unsigned RHS1 = MI->getOperand(3).getReg();
6973 unsigned RHS2 = MI->getOperand(4).getReg();
6974 AddDefaultPred(BuildMI(BB, dl,
6975 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6976 .addReg(LHS1).addReg(RHS1));
6977 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6978 .addReg(LHS2).addReg(RHS2)
6979 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6982 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6983 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6984 if (MI->getOperand(0).getImm() == ARMCC::NE)
6985 std::swap(destMBB, exitMBB);
6987 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6988 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6990 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6992 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6994 MI->eraseFromParent(); // The pseudo instruction is gone now.
6998 case ARM::Int_eh_sjlj_setjmp:
6999 case ARM::Int_eh_sjlj_setjmp_nofp:
7000 case ARM::tInt_eh_sjlj_setjmp:
7001 case ARM::t2Int_eh_sjlj_setjmp:
7002 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7003 EmitSjLjDispatchBlock(MI, BB);
7008 // To insert an ABS instruction, we have to insert the
7009 // diamond control-flow pattern. The incoming instruction knows the
7010 // source vreg to test against 0, the destination vreg to set,
7011 // the condition code register to branch on, the
7012 // true/false values to select between, and a branch opcode to use.
7017 // BCC (branch to SinkBB if V0 >= 0)
7018 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7019 // SinkBB: V1 = PHI(V2, V3)
7020 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7021 MachineFunction::iterator BBI = BB;
7023 MachineFunction *Fn = BB->getParent();
7024 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7025 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7026 Fn->insert(BBI, RSBBB);
7027 Fn->insert(BBI, SinkBB);
7029 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7030 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7031 bool isThumb2 = Subtarget->isThumb2();
7032 MachineRegisterInfo &MRI = Fn->getRegInfo();
7033 // In Thumb mode S must not be specified if source register is the SP or
7034 // PC and if destination register is the SP, so restrict register class
7035 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7036 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7037 (const TargetRegisterClass*)&ARM::GPRRegClass);
7039 // Transfer the remainder of BB and its successor edges to sinkMBB.
7040 SinkBB->splice(SinkBB->begin(), BB,
7041 llvm::next(MachineBasicBlock::iterator(MI)),
7043 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7045 BB->addSuccessor(RSBBB);
7046 BB->addSuccessor(SinkBB);
7048 // fall through to SinkMBB
7049 RSBBB->addSuccessor(SinkBB);
7051 // insert a cmp at the end of BB
7052 AddDefaultPred(BuildMI(BB, dl,
7053 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7054 .addReg(ABSSrcReg).addImm(0));
7056 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7058 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7059 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7061 // insert rsbri in RSBBB
7062 // Note: BCC and rsbri will be converted into predicated rsbmi
7063 // by if-conversion pass
7064 BuildMI(*RSBBB, RSBBB->begin(), dl,
7065 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7066 .addReg(ABSSrcReg, RegState::Kill)
7067 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7069 // insert PHI in SinkBB,
7070 // reuse ABSDstReg to not change uses of ABS instruction
7071 BuildMI(*SinkBB, SinkBB->begin(), dl,
7072 TII->get(ARM::PHI), ABSDstReg)
7073 .addReg(NewRsbDstReg).addMBB(RSBBB)
7074 .addReg(ABSSrcReg).addMBB(BB);
7076 // remove ABS instruction
7077 MI->eraseFromParent();
7079 // return last added BB
7082 case ARM::COPY_STRUCT_BYVAL_I32:
7084 return EmitStructByval(MI, BB);
7088 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7089 SDNode *Node) const {
7090 if (!MI->hasPostISelHook()) {
7091 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7092 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7096 const MCInstrDesc *MCID = &MI->getDesc();
7097 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7098 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7099 // operand is still set to noreg. If needed, set the optional operand's
7100 // register to CPSR, and remove the redundant implicit def.
7102 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7104 // Rename pseudo opcodes.
7105 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7107 const ARMBaseInstrInfo *TII =
7108 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7109 MCID = &TII->get(NewOpc);
7111 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7112 "converted opcode should be the same except for cc_out");
7116 // Add the optional cc_out operand
7117 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7119 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7121 // Any ARM instruction that sets the 's' bit should specify an optional
7122 // "cc_out" operand in the last operand position.
7123 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7124 assert(!NewOpc && "Optional cc_out operand required");
7127 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7128 // since we already have an optional CPSR def.
7129 bool definesCPSR = false;
7130 bool deadCPSR = false;
7131 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7133 const MachineOperand &MO = MI->getOperand(i);
7134 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7138 MI->RemoveOperand(i);
7143 assert(!NewOpc && "Optional cc_out operand required");
7146 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7148 assert(!MI->getOperand(ccOutIdx).getReg() &&
7149 "expect uninitialized optional cc_out operand");
7153 // If this instruction was defined with an optional CPSR def and its dag node
7154 // had a live implicit CPSR def, then activate the optional CPSR def.
7155 MachineOperand &MO = MI->getOperand(ccOutIdx);
7156 MO.setReg(ARM::CPSR);
7160 //===----------------------------------------------------------------------===//
7161 // ARM Optimization Hooks
7162 //===----------------------------------------------------------------------===//
7164 // Helper function that checks if N is a null or all ones constant.
7165 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7166 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7169 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7172 // Return true if N is conditionally 0 or all ones.
7173 // Detects these expressions where cc is an i1 value:
7175 // (select cc 0, y) [AllOnes=0]
7176 // (select cc y, 0) [AllOnes=0]
7177 // (zext cc) [AllOnes=0]
7178 // (sext cc) [AllOnes=0/1]
7179 // (select cc -1, y) [AllOnes=1]
7180 // (select cc y, -1) [AllOnes=1]
7182 // Invert is set when N is the null/all ones constant when CC is false.
7183 // OtherOp is set to the alternative value of N.
7184 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7185 SDValue &CC, bool &Invert,
7187 SelectionDAG &DAG) {
7188 switch (N->getOpcode()) {
7189 default: return false;
7191 CC = N->getOperand(0);
7192 SDValue N1 = N->getOperand(1);
7193 SDValue N2 = N->getOperand(2);
7194 if (isZeroOrAllOnes(N1, AllOnes)) {
7199 if (isZeroOrAllOnes(N2, AllOnes)) {
7206 case ISD::ZERO_EXTEND:
7207 // (zext cc) can never be the all ones value.
7211 case ISD::SIGN_EXTEND: {
7212 EVT VT = N->getValueType(0);
7213 CC = N->getOperand(0);
7214 if (CC.getValueType() != MVT::i1)
7218 // When looking for an AllOnes constant, N is an sext, and the 'other'
7220 OtherOp = DAG.getConstant(0, VT);
7221 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7222 // When looking for a 0 constant, N can be zext or sext.
7223 OtherOp = DAG.getConstant(1, VT);
7225 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7231 // Combine a constant select operand into its use:
7233 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7234 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7235 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7236 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7237 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7239 // The transform is rejected if the select doesn't have a constant operand that
7240 // is null, or all ones when AllOnes is set.
7242 // Also recognize sext/zext from i1:
7244 // (add (zext cc), x) -> (select cc (add x, 1), x)
7245 // (add (sext cc), x) -> (select cc (add x, -1), x)
7247 // These transformations eventually create predicated instructions.
7249 // @param N The node to transform.
7250 // @param Slct The N operand that is a select.
7251 // @param OtherOp The other N operand (x above).
7252 // @param DCI Context.
7253 // @param AllOnes Require the select constant to be all ones instead of null.
7254 // @returns The new node, or SDValue() on failure.
7256 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7257 TargetLowering::DAGCombinerInfo &DCI,
7258 bool AllOnes = false) {
7259 SelectionDAG &DAG = DCI.DAG;
7260 EVT VT = N->getValueType(0);
7261 SDValue NonConstantVal;
7264 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7265 NonConstantVal, DAG))
7268 // Slct is now know to be the desired identity constant when CC is true.
7269 SDValue TrueVal = OtherOp;
7270 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7271 OtherOp, NonConstantVal);
7272 // Unless SwapSelectOps says CC should be false.
7274 std::swap(TrueVal, FalseVal);
7276 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7277 CCOp, TrueVal, FalseVal);
7280 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7282 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7283 TargetLowering::DAGCombinerInfo &DCI) {
7284 SDValue N0 = N->getOperand(0);
7285 SDValue N1 = N->getOperand(1);
7286 if (N0.getNode()->hasOneUse()) {
7287 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7288 if (Result.getNode())
7291 if (N1.getNode()->hasOneUse()) {
7292 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7293 if (Result.getNode())
7299 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7300 // (only after legalization).
7301 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7302 TargetLowering::DAGCombinerInfo &DCI,
7303 const ARMSubtarget *Subtarget) {
7305 // Only perform optimization if after legalize, and if NEON is available. We
7306 // also expected both operands to be BUILD_VECTORs.
7307 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7308 || N0.getOpcode() != ISD::BUILD_VECTOR
7309 || N1.getOpcode() != ISD::BUILD_VECTOR)
7312 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7313 EVT VT = N->getValueType(0);
7314 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7317 // Check that the vector operands are of the right form.
7318 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7319 // operands, where N is the size of the formed vector.
7320 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7321 // index such that we have a pair wise add pattern.
7323 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7324 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7326 SDValue Vec = N0->getOperand(0)->getOperand(0);
7327 SDNode *V = Vec.getNode();
7328 unsigned nextIndex = 0;
7330 // For each operands to the ADD which are BUILD_VECTORs,
7331 // check to see if each of their operands are an EXTRACT_VECTOR with
7332 // the same vector and appropriate index.
7333 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7334 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7335 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7337 SDValue ExtVec0 = N0->getOperand(i);
7338 SDValue ExtVec1 = N1->getOperand(i);
7340 // First operand is the vector, verify its the same.
7341 if (V != ExtVec0->getOperand(0).getNode() ||
7342 V != ExtVec1->getOperand(0).getNode())
7345 // Second is the constant, verify its correct.
7346 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7347 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7349 // For the constant, we want to see all the even or all the odd.
7350 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7351 || C1->getZExtValue() != nextIndex+1)
7360 // Create VPADDL node.
7361 SelectionDAG &DAG = DCI.DAG;
7362 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7364 // Build operand list.
7365 SmallVector<SDValue, 8> Ops;
7366 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7367 TLI.getPointerTy()));
7369 // Input is the vector.
7372 // Get widened type and narrowed type.
7374 unsigned numElem = VT.getVectorNumElements();
7375 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7376 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7377 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7378 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7380 llvm_unreachable("Invalid vector element type for padd optimization.");
7383 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7384 widenType, &Ops[0], Ops.size());
7385 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7388 static SDValue findMUL_LOHI(SDValue V) {
7389 if (V->getOpcode() == ISD::UMUL_LOHI ||
7390 V->getOpcode() == ISD::SMUL_LOHI)
7395 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7396 TargetLowering::DAGCombinerInfo &DCI,
7397 const ARMSubtarget *Subtarget) {
7399 if (Subtarget->isThumb1Only()) return SDValue();
7401 // Only perform the checks after legalize when the pattern is available.
7402 if (DCI.isBeforeLegalize()) return SDValue();
7404 // Look for multiply add opportunities.
7405 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7406 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7407 // a glue link from the first add to the second add.
7408 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7409 // a S/UMLAL instruction.
7412 // \ / \ [no multiline comment]
7418 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7419 SDValue AddcOp0 = AddcNode->getOperand(0);
7420 SDValue AddcOp1 = AddcNode->getOperand(1);
7422 // Check if the two operands are from the same mul_lohi node.
7423 if (AddcOp0.getNode() == AddcOp1.getNode())
7426 assert(AddcNode->getNumValues() == 2 &&
7427 AddcNode->getValueType(0) == MVT::i32 &&
7428 AddcNode->getValueType(1) == MVT::Glue &&
7429 "Expect ADDC with two result values: i32, glue");
7431 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7432 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7433 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7434 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7435 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7438 // Look for the glued ADDE.
7439 SDNode* AddeNode = AddcNode->getGluedUser();
7440 if (AddeNode == NULL)
7443 // Make sure it is really an ADDE.
7444 if (AddeNode->getOpcode() != ISD::ADDE)
7447 assert(AddeNode->getNumOperands() == 3 &&
7448 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7449 "ADDE node has the wrong inputs");
7451 // Check for the triangle shape.
7452 SDValue AddeOp0 = AddeNode->getOperand(0);
7453 SDValue AddeOp1 = AddeNode->getOperand(1);
7455 // Make sure that the ADDE operands are not coming from the same node.
7456 if (AddeOp0.getNode() == AddeOp1.getNode())
7459 // Find the MUL_LOHI node walking up ADDE's operands.
7460 bool IsLeftOperandMUL = false;
7461 SDValue MULOp = findMUL_LOHI(AddeOp0);
7462 if (MULOp == SDValue())
7463 MULOp = findMUL_LOHI(AddeOp1);
7465 IsLeftOperandMUL = true;
7466 if (MULOp == SDValue())
7469 // Figure out the right opcode.
7470 unsigned Opc = MULOp->getOpcode();
7471 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7473 // Figure out the high and low input values to the MLAL node.
7474 SDValue* HiMul = &MULOp;
7475 SDValue* HiAdd = NULL;
7476 SDValue* LoMul = NULL;
7477 SDValue* LowAdd = NULL;
7479 if (IsLeftOperandMUL)
7485 if (AddcOp0->getOpcode() == Opc) {
7489 if (AddcOp1->getOpcode() == Opc) {
7497 if (LoMul->getNode() != HiMul->getNode())
7500 // Create the merged node.
7501 SelectionDAG &DAG = DCI.DAG;
7503 // Build operand list.
7504 SmallVector<SDValue, 8> Ops;
7505 Ops.push_back(LoMul->getOperand(0));
7506 Ops.push_back(LoMul->getOperand(1));
7507 Ops.push_back(*LowAdd);
7508 Ops.push_back(*HiAdd);
7510 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7511 DAG.getVTList(MVT::i32, MVT::i32),
7512 &Ops[0], Ops.size());
7514 // Replace the ADDs' nodes uses by the MLA node's values.
7515 SDValue HiMLALResult(MLALNode.getNode(), 1);
7516 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7518 SDValue LoMLALResult(MLALNode.getNode(), 0);
7519 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7521 // Return original node to notify the driver to stop replacing.
7522 SDValue resNode(AddcNode, 0);
7526 /// PerformADDCCombine - Target-specific dag combine transform from
7527 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7528 static SDValue PerformADDCCombine(SDNode *N,
7529 TargetLowering::DAGCombinerInfo &DCI,
7530 const ARMSubtarget *Subtarget) {
7532 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7536 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7537 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7538 /// called with the default operands, and if that fails, with commuted
7540 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7541 TargetLowering::DAGCombinerInfo &DCI,
7542 const ARMSubtarget *Subtarget){
7544 // Attempt to create vpaddl for this add.
7545 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7546 if (Result.getNode())
7549 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7550 if (N0.getNode()->hasOneUse()) {
7551 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7552 if (Result.getNode()) return Result;
7557 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7559 static SDValue PerformADDCombine(SDNode *N,
7560 TargetLowering::DAGCombinerInfo &DCI,
7561 const ARMSubtarget *Subtarget) {
7562 SDValue N0 = N->getOperand(0);
7563 SDValue N1 = N->getOperand(1);
7565 // First try with the default operand order.
7566 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7567 if (Result.getNode())
7570 // If that didn't work, try again with the operands commuted.
7571 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7574 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7576 static SDValue PerformSUBCombine(SDNode *N,
7577 TargetLowering::DAGCombinerInfo &DCI) {
7578 SDValue N0 = N->getOperand(0);
7579 SDValue N1 = N->getOperand(1);
7581 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7582 if (N1.getNode()->hasOneUse()) {
7583 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7584 if (Result.getNode()) return Result;
7590 /// PerformVMULCombine
7591 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7592 /// special multiplier accumulator forwarding.
7598 static SDValue PerformVMULCombine(SDNode *N,
7599 TargetLowering::DAGCombinerInfo &DCI,
7600 const ARMSubtarget *Subtarget) {
7601 if (!Subtarget->hasVMLxForwarding())
7604 SelectionDAG &DAG = DCI.DAG;
7605 SDValue N0 = N->getOperand(0);
7606 SDValue N1 = N->getOperand(1);
7607 unsigned Opcode = N0.getOpcode();
7608 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7609 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7610 Opcode = N1.getOpcode();
7611 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7612 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7617 EVT VT = N->getValueType(0);
7618 DebugLoc DL = N->getDebugLoc();
7619 SDValue N00 = N0->getOperand(0);
7620 SDValue N01 = N0->getOperand(1);
7621 return DAG.getNode(Opcode, DL, VT,
7622 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7623 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7626 static SDValue PerformMULCombine(SDNode *N,
7627 TargetLowering::DAGCombinerInfo &DCI,
7628 const ARMSubtarget *Subtarget) {
7629 SelectionDAG &DAG = DCI.DAG;
7631 if (Subtarget->isThumb1Only())
7634 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7637 EVT VT = N->getValueType(0);
7638 if (VT.is64BitVector() || VT.is128BitVector())
7639 return PerformVMULCombine(N, DCI, Subtarget);
7643 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7647 int64_t MulAmt = C->getSExtValue();
7648 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7650 ShiftAmt = ShiftAmt & (32 - 1);
7651 SDValue V = N->getOperand(0);
7652 DebugLoc DL = N->getDebugLoc();
7655 MulAmt >>= ShiftAmt;
7658 if (isPowerOf2_32(MulAmt - 1)) {
7659 // (mul x, 2^N + 1) => (add (shl x, N), x)
7660 Res = DAG.getNode(ISD::ADD, DL, VT,
7662 DAG.getNode(ISD::SHL, DL, VT,
7664 DAG.getConstant(Log2_32(MulAmt - 1),
7666 } else if (isPowerOf2_32(MulAmt + 1)) {
7667 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7668 Res = DAG.getNode(ISD::SUB, DL, VT,
7669 DAG.getNode(ISD::SHL, DL, VT,
7671 DAG.getConstant(Log2_32(MulAmt + 1),
7677 uint64_t MulAmtAbs = -MulAmt;
7678 if (isPowerOf2_32(MulAmtAbs + 1)) {
7679 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7680 Res = DAG.getNode(ISD::SUB, DL, VT,
7682 DAG.getNode(ISD::SHL, DL, VT,
7684 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7686 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7687 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7688 Res = DAG.getNode(ISD::ADD, DL, VT,
7690 DAG.getNode(ISD::SHL, DL, VT,
7692 DAG.getConstant(Log2_32(MulAmtAbs-1),
7694 Res = DAG.getNode(ISD::SUB, DL, VT,
7695 DAG.getConstant(0, MVT::i32),Res);
7702 Res = DAG.getNode(ISD::SHL, DL, VT,
7703 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7705 // Do not add new nodes to DAG combiner worklist.
7706 DCI.CombineTo(N, Res, false);
7710 static SDValue PerformANDCombine(SDNode *N,
7711 TargetLowering::DAGCombinerInfo &DCI,
7712 const ARMSubtarget *Subtarget) {
7714 // Attempt to use immediate-form VBIC
7715 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7716 DebugLoc dl = N->getDebugLoc();
7717 EVT VT = N->getValueType(0);
7718 SelectionDAG &DAG = DCI.DAG;
7720 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7723 APInt SplatBits, SplatUndef;
7724 unsigned SplatBitSize;
7727 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7728 if (SplatBitSize <= 64) {
7730 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7731 SplatUndef.getZExtValue(), SplatBitSize,
7732 DAG, VbicVT, VT.is128BitVector(),
7734 if (Val.getNode()) {
7736 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7737 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7738 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7743 if (!Subtarget->isThumb1Only()) {
7744 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7745 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7746 if (Result.getNode())
7753 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7754 static SDValue PerformORCombine(SDNode *N,
7755 TargetLowering::DAGCombinerInfo &DCI,
7756 const ARMSubtarget *Subtarget) {
7757 // Attempt to use immediate-form VORR
7758 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7759 DebugLoc dl = N->getDebugLoc();
7760 EVT VT = N->getValueType(0);
7761 SelectionDAG &DAG = DCI.DAG;
7763 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7766 APInt SplatBits, SplatUndef;
7767 unsigned SplatBitSize;
7769 if (BVN && Subtarget->hasNEON() &&
7770 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7771 if (SplatBitSize <= 64) {
7773 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7774 SplatUndef.getZExtValue(), SplatBitSize,
7775 DAG, VorrVT, VT.is128BitVector(),
7777 if (Val.getNode()) {
7779 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7780 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7781 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7786 if (!Subtarget->isThumb1Only()) {
7787 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7788 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7789 if (Result.getNode())
7793 // The code below optimizes (or (and X, Y), Z).
7794 // The AND operand needs to have a single user to make these optimizations
7796 SDValue N0 = N->getOperand(0);
7797 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7799 SDValue N1 = N->getOperand(1);
7801 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7802 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7803 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7805 unsigned SplatBitSize;
7808 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7810 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7811 HasAnyUndefs) && !HasAnyUndefs) {
7812 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7814 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7815 HasAnyUndefs) && !HasAnyUndefs &&
7816 SplatBits0 == ~SplatBits1) {
7817 // Canonicalize the vector type to make instruction selection simpler.
7818 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7819 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7820 N0->getOperand(1), N0->getOperand(0),
7822 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7827 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7830 // BFI is only available on V6T2+
7831 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7834 DebugLoc DL = N->getDebugLoc();
7835 // 1) or (and A, mask), val => ARMbfi A, val, mask
7836 // iff (val & mask) == val
7838 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7839 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7840 // && mask == ~mask2
7841 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7842 // && ~mask == mask2
7843 // (i.e., copy a bitfield value into another bitfield of the same width)
7848 SDValue N00 = N0.getOperand(0);
7850 // The value and the mask need to be constants so we can verify this is
7851 // actually a bitfield set. If the mask is 0xffff, we can do better
7852 // via a movt instruction, so don't use BFI in that case.
7853 SDValue MaskOp = N0.getOperand(1);
7854 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7857 unsigned Mask = MaskC->getZExtValue();
7861 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7862 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7864 unsigned Val = N1C->getZExtValue();
7865 if ((Val & ~Mask) != Val)
7868 if (ARM::isBitFieldInvertedMask(Mask)) {
7869 Val >>= CountTrailingZeros_32(~Mask);
7871 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7872 DAG.getConstant(Val, MVT::i32),
7873 DAG.getConstant(Mask, MVT::i32));
7875 // Do not add new nodes to DAG combiner worklist.
7876 DCI.CombineTo(N, Res, false);
7879 } else if (N1.getOpcode() == ISD::AND) {
7880 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7881 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7884 unsigned Mask2 = N11C->getZExtValue();
7886 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7888 if (ARM::isBitFieldInvertedMask(Mask) &&
7890 // The pack halfword instruction works better for masks that fit it,
7891 // so use that when it's available.
7892 if (Subtarget->hasT2ExtractPack() &&
7893 (Mask == 0xffff || Mask == 0xffff0000))
7896 unsigned amt = CountTrailingZeros_32(Mask2);
7897 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7898 DAG.getConstant(amt, MVT::i32));
7899 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7900 DAG.getConstant(Mask, MVT::i32));
7901 // Do not add new nodes to DAG combiner worklist.
7902 DCI.CombineTo(N, Res, false);
7904 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7906 // The pack halfword instruction works better for masks that fit it,
7907 // so use that when it's available.
7908 if (Subtarget->hasT2ExtractPack() &&
7909 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7912 unsigned lsb = CountTrailingZeros_32(Mask);
7913 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7914 DAG.getConstant(lsb, MVT::i32));
7915 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7916 DAG.getConstant(Mask2, MVT::i32));
7917 // Do not add new nodes to DAG combiner worklist.
7918 DCI.CombineTo(N, Res, false);
7923 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7924 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7925 ARM::isBitFieldInvertedMask(~Mask)) {
7926 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7927 // where lsb(mask) == #shamt and masked bits of B are known zero.
7928 SDValue ShAmt = N00.getOperand(1);
7929 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7930 unsigned LSB = CountTrailingZeros_32(Mask);
7934 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7935 DAG.getConstant(~Mask, MVT::i32));
7937 // Do not add new nodes to DAG combiner worklist.
7938 DCI.CombineTo(N, Res, false);
7944 static SDValue PerformXORCombine(SDNode *N,
7945 TargetLowering::DAGCombinerInfo &DCI,
7946 const ARMSubtarget *Subtarget) {
7947 EVT VT = N->getValueType(0);
7948 SelectionDAG &DAG = DCI.DAG;
7950 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7953 if (!Subtarget->isThumb1Only()) {
7954 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7955 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7956 if (Result.getNode())
7963 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7964 /// the bits being cleared by the AND are not demanded by the BFI.
7965 static SDValue PerformBFICombine(SDNode *N,
7966 TargetLowering::DAGCombinerInfo &DCI) {
7967 SDValue N1 = N->getOperand(1);
7968 if (N1.getOpcode() == ISD::AND) {
7969 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7972 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7973 unsigned LSB = CountTrailingZeros_32(~InvMask);
7974 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7975 unsigned Mask = (1 << Width)-1;
7976 unsigned Mask2 = N11C->getZExtValue();
7977 if ((Mask & (~Mask2)) == 0)
7978 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7979 N->getOperand(0), N1.getOperand(0),
7985 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7986 /// ARMISD::VMOVRRD.
7987 static SDValue PerformVMOVRRDCombine(SDNode *N,
7988 TargetLowering::DAGCombinerInfo &DCI) {
7989 // vmovrrd(vmovdrr x, y) -> x,y
7990 SDValue InDouble = N->getOperand(0);
7991 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7992 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7994 // vmovrrd(load f64) -> (load i32), (load i32)
7995 SDNode *InNode = InDouble.getNode();
7996 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7997 InNode->getValueType(0) == MVT::f64 &&
7998 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7999 !cast<LoadSDNode>(InNode)->isVolatile()) {
8000 // TODO: Should this be done for non-FrameIndex operands?
8001 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8003 SelectionDAG &DAG = DCI.DAG;
8004 DebugLoc DL = LD->getDebugLoc();
8005 SDValue BasePtr = LD->getBasePtr();
8006 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8007 LD->getPointerInfo(), LD->isVolatile(),
8008 LD->isNonTemporal(), LD->isInvariant(),
8009 LD->getAlignment());
8011 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8012 DAG.getConstant(4, MVT::i32));
8013 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8014 LD->getPointerInfo(), LD->isVolatile(),
8015 LD->isNonTemporal(), LD->isInvariant(),
8016 std::min(4U, LD->getAlignment() / 2));
8018 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8019 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8020 DCI.RemoveFromWorklist(LD);
8028 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8029 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8030 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8031 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8032 SDValue Op0 = N->getOperand(0);
8033 SDValue Op1 = N->getOperand(1);
8034 if (Op0.getOpcode() == ISD::BITCAST)
8035 Op0 = Op0.getOperand(0);
8036 if (Op1.getOpcode() == ISD::BITCAST)
8037 Op1 = Op1.getOperand(0);
8038 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8039 Op0.getNode() == Op1.getNode() &&
8040 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8041 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
8042 N->getValueType(0), Op0.getOperand(0));
8046 /// PerformSTORECombine - Target-specific dag combine xforms for
8048 static SDValue PerformSTORECombine(SDNode *N,
8049 TargetLowering::DAGCombinerInfo &DCI) {
8050 StoreSDNode *St = cast<StoreSDNode>(N);
8051 if (St->isVolatile())
8054 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8055 // pack all of the elements in one place. Next, store to memory in fewer
8057 SDValue StVal = St->getValue();
8058 EVT VT = StVal.getValueType();
8059 if (St->isTruncatingStore() && VT.isVector()) {
8060 SelectionDAG &DAG = DCI.DAG;
8061 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8062 EVT StVT = St->getMemoryVT();
8063 unsigned NumElems = VT.getVectorNumElements();
8064 assert(StVT != VT && "Cannot truncate to the same type");
8065 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8066 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8068 // From, To sizes and ElemCount must be pow of two
8069 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8071 // We are going to use the original vector elt for storing.
8072 // Accumulated smaller vector elements must be a multiple of the store size.
8073 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8075 unsigned SizeRatio = FromEltSz / ToEltSz;
8076 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8078 // Create a type on which we perform the shuffle.
8079 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8080 NumElems*SizeRatio);
8081 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8083 DebugLoc DL = St->getDebugLoc();
8084 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8085 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8086 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8088 // Can't shuffle using an illegal type.
8089 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8091 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8092 DAG.getUNDEF(WideVec.getValueType()),
8094 // At this point all of the data is stored at the bottom of the
8095 // register. We now need to save it to mem.
8097 // Find the largest store unit
8098 MVT StoreType = MVT::i8;
8099 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8100 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8101 MVT Tp = (MVT::SimpleValueType)tp;
8102 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8105 // Didn't find a legal store type.
8106 if (!TLI.isTypeLegal(StoreType))
8109 // Bitcast the original vector into a vector of store-size units
8110 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8111 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8112 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8113 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8114 SmallVector<SDValue, 8> Chains;
8115 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8116 TLI.getPointerTy());
8117 SDValue BasePtr = St->getBasePtr();
8119 // Perform one or more big stores into memory.
8120 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8121 for (unsigned I = 0; I < E; I++) {
8122 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8123 StoreType, ShuffWide,
8124 DAG.getIntPtrConstant(I));
8125 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8126 St->getPointerInfo(), St->isVolatile(),
8127 St->isNonTemporal(), St->getAlignment());
8128 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8130 Chains.push_back(Ch);
8132 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8136 if (!ISD::isNormalStore(St))
8139 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8140 // ARM stores of arguments in the same cache line.
8141 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8142 StVal.getNode()->hasOneUse()) {
8143 SelectionDAG &DAG = DCI.DAG;
8144 DebugLoc DL = St->getDebugLoc();
8145 SDValue BasePtr = St->getBasePtr();
8146 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8147 StVal.getNode()->getOperand(0), BasePtr,
8148 St->getPointerInfo(), St->isVolatile(),
8149 St->isNonTemporal(), St->getAlignment());
8151 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8152 DAG.getConstant(4, MVT::i32));
8153 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8154 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8155 St->isNonTemporal(),
8156 std::min(4U, St->getAlignment() / 2));
8159 if (StVal.getValueType() != MVT::i64 ||
8160 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8163 // Bitcast an i64 store extracted from a vector to f64.
8164 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8165 SelectionDAG &DAG = DCI.DAG;
8166 DebugLoc dl = StVal.getDebugLoc();
8167 SDValue IntVec = StVal.getOperand(0);
8168 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8169 IntVec.getValueType().getVectorNumElements());
8170 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8171 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8172 Vec, StVal.getOperand(1));
8173 dl = N->getDebugLoc();
8174 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8175 // Make the DAGCombiner fold the bitcasts.
8176 DCI.AddToWorklist(Vec.getNode());
8177 DCI.AddToWorklist(ExtElt.getNode());
8178 DCI.AddToWorklist(V.getNode());
8179 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8180 St->getPointerInfo(), St->isVolatile(),
8181 St->isNonTemporal(), St->getAlignment(),
8185 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8186 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8187 /// i64 vector to have f64 elements, since the value can then be loaded
8188 /// directly into a VFP register.
8189 static bool hasNormalLoadOperand(SDNode *N) {
8190 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8191 for (unsigned i = 0; i < NumElts; ++i) {
8192 SDNode *Elt = N->getOperand(i).getNode();
8193 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8199 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8200 /// ISD::BUILD_VECTOR.
8201 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8202 TargetLowering::DAGCombinerInfo &DCI){
8203 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8204 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8205 // into a pair of GPRs, which is fine when the value is used as a scalar,
8206 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8207 SelectionDAG &DAG = DCI.DAG;
8208 if (N->getNumOperands() == 2) {
8209 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8214 // Load i64 elements as f64 values so that type legalization does not split
8215 // them up into i32 values.
8216 EVT VT = N->getValueType(0);
8217 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8219 DebugLoc dl = N->getDebugLoc();
8220 SmallVector<SDValue, 8> Ops;
8221 unsigned NumElts = VT.getVectorNumElements();
8222 for (unsigned i = 0; i < NumElts; ++i) {
8223 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8225 // Make the DAGCombiner fold the bitcast.
8226 DCI.AddToWorklist(V.getNode());
8228 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8229 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8230 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8233 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8234 /// ISD::INSERT_VECTOR_ELT.
8235 static SDValue PerformInsertEltCombine(SDNode *N,
8236 TargetLowering::DAGCombinerInfo &DCI) {
8237 // Bitcast an i64 load inserted into a vector to f64.
8238 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8239 EVT VT = N->getValueType(0);
8240 SDNode *Elt = N->getOperand(1).getNode();
8241 if (VT.getVectorElementType() != MVT::i64 ||
8242 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8245 SelectionDAG &DAG = DCI.DAG;
8246 DebugLoc dl = N->getDebugLoc();
8247 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8248 VT.getVectorNumElements());
8249 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8250 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8251 // Make the DAGCombiner fold the bitcasts.
8252 DCI.AddToWorklist(Vec.getNode());
8253 DCI.AddToWorklist(V.getNode());
8254 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8255 Vec, V, N->getOperand(2));
8256 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8259 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8260 /// ISD::VECTOR_SHUFFLE.
8261 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8262 // The LLVM shufflevector instruction does not require the shuffle mask
8263 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8264 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8265 // operands do not match the mask length, they are extended by concatenating
8266 // them with undef vectors. That is probably the right thing for other
8267 // targets, but for NEON it is better to concatenate two double-register
8268 // size vector operands into a single quad-register size vector. Do that
8269 // transformation here:
8270 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8271 // shuffle(concat(v1, v2), undef)
8272 SDValue Op0 = N->getOperand(0);
8273 SDValue Op1 = N->getOperand(1);
8274 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8275 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8276 Op0.getNumOperands() != 2 ||
8277 Op1.getNumOperands() != 2)
8279 SDValue Concat0Op1 = Op0.getOperand(1);
8280 SDValue Concat1Op1 = Op1.getOperand(1);
8281 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8282 Concat1Op1.getOpcode() != ISD::UNDEF)
8284 // Skip the transformation if any of the types are illegal.
8285 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8286 EVT VT = N->getValueType(0);
8287 if (!TLI.isTypeLegal(VT) ||
8288 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8289 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8292 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8293 Op0.getOperand(0), Op1.getOperand(0));
8294 // Translate the shuffle mask.
8295 SmallVector<int, 16> NewMask;
8296 unsigned NumElts = VT.getVectorNumElements();
8297 unsigned HalfElts = NumElts/2;
8298 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8299 for (unsigned n = 0; n < NumElts; ++n) {
8300 int MaskElt = SVN->getMaskElt(n);
8302 if (MaskElt < (int)HalfElts)
8304 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8305 NewElt = HalfElts + MaskElt - NumElts;
8306 NewMask.push_back(NewElt);
8308 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8309 DAG.getUNDEF(VT), NewMask.data());
8312 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8313 /// NEON load/store intrinsics to merge base address updates.
8314 static SDValue CombineBaseUpdate(SDNode *N,
8315 TargetLowering::DAGCombinerInfo &DCI) {
8316 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8319 SelectionDAG &DAG = DCI.DAG;
8320 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8321 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8322 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8323 SDValue Addr = N->getOperand(AddrOpIdx);
8325 // Search for a use of the address operand that is an increment.
8326 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8327 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8329 if (User->getOpcode() != ISD::ADD ||
8330 UI.getUse().getResNo() != Addr.getResNo())
8333 // Check that the add is independent of the load/store. Otherwise, folding
8334 // it would create a cycle.
8335 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8338 // Find the new opcode for the updating load/store.
8340 bool isLaneOp = false;
8341 unsigned NewOpc = 0;
8342 unsigned NumVecs = 0;
8344 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8346 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8347 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8349 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8351 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8353 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8355 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8356 NumVecs = 2; isLaneOp = true; break;
8357 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8358 NumVecs = 3; isLaneOp = true; break;
8359 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8360 NumVecs = 4; isLaneOp = true; break;
8361 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8362 NumVecs = 1; isLoad = false; break;
8363 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8364 NumVecs = 2; isLoad = false; break;
8365 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8366 NumVecs = 3; isLoad = false; break;
8367 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8368 NumVecs = 4; isLoad = false; break;
8369 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8370 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8371 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8372 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8373 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8374 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8378 switch (N->getOpcode()) {
8379 default: llvm_unreachable("unexpected opcode for Neon base update");
8380 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8381 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8382 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8386 // Find the size of memory referenced by the load/store.
8389 VecTy = N->getValueType(0);
8391 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8392 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8394 NumBytes /= VecTy.getVectorNumElements();
8396 // If the increment is a constant, it must match the memory ref size.
8397 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8398 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8399 uint64_t IncVal = CInc->getZExtValue();
8400 if (IncVal != NumBytes)
8402 } else if (NumBytes >= 3 * 16) {
8403 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8404 // separate instructions that make it harder to use a non-constant update.
8408 // Create the new updating load/store node.
8410 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8412 for (n = 0; n < NumResultVecs; ++n)
8414 Tys[n++] = MVT::i32;
8415 Tys[n] = MVT::Other;
8416 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8417 SmallVector<SDValue, 8> Ops;
8418 Ops.push_back(N->getOperand(0)); // incoming chain
8419 Ops.push_back(N->getOperand(AddrOpIdx));
8421 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8422 Ops.push_back(N->getOperand(i));
8424 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8425 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8426 Ops.data(), Ops.size(),
8427 MemInt->getMemoryVT(),
8428 MemInt->getMemOperand());
8431 std::vector<SDValue> NewResults;
8432 for (unsigned i = 0; i < NumResultVecs; ++i) {
8433 NewResults.push_back(SDValue(UpdN.getNode(), i));
8435 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8436 DCI.CombineTo(N, NewResults);
8437 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8444 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8445 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8446 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8448 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8449 SelectionDAG &DAG = DCI.DAG;
8450 EVT VT = N->getValueType(0);
8451 // vldN-dup instructions only support 64-bit vectors for N > 1.
8452 if (!VT.is64BitVector())
8455 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8456 SDNode *VLD = N->getOperand(0).getNode();
8457 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8459 unsigned NumVecs = 0;
8460 unsigned NewOpc = 0;
8461 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8462 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8464 NewOpc = ARMISD::VLD2DUP;
8465 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8467 NewOpc = ARMISD::VLD3DUP;
8468 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8470 NewOpc = ARMISD::VLD4DUP;
8475 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8476 // numbers match the load.
8477 unsigned VLDLaneNo =
8478 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8479 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8481 // Ignore uses of the chain result.
8482 if (UI.getUse().getResNo() == NumVecs)
8485 if (User->getOpcode() != ARMISD::VDUPLANE ||
8486 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8490 // Create the vldN-dup node.
8493 for (n = 0; n < NumVecs; ++n)
8495 Tys[n] = MVT::Other;
8496 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8497 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8498 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8499 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8500 Ops, 2, VLDMemInt->getMemoryVT(),
8501 VLDMemInt->getMemOperand());
8504 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8506 unsigned ResNo = UI.getUse().getResNo();
8507 // Ignore uses of the chain result.
8508 if (ResNo == NumVecs)
8511 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8514 // Now the vldN-lane intrinsic is dead except for its chain result.
8515 // Update uses of the chain.
8516 std::vector<SDValue> VLDDupResults;
8517 for (unsigned n = 0; n < NumVecs; ++n)
8518 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8519 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8520 DCI.CombineTo(VLD, VLDDupResults);
8525 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8526 /// ARMISD::VDUPLANE.
8527 static SDValue PerformVDUPLANECombine(SDNode *N,
8528 TargetLowering::DAGCombinerInfo &DCI) {
8529 SDValue Op = N->getOperand(0);
8531 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8532 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8533 if (CombineVLDDUP(N, DCI))
8534 return SDValue(N, 0);
8536 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8537 // redundant. Ignore bit_converts for now; element sizes are checked below.
8538 while (Op.getOpcode() == ISD::BITCAST)
8539 Op = Op.getOperand(0);
8540 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8543 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8544 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8545 // The canonical VMOV for a zero vector uses a 32-bit element size.
8546 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8548 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8550 EVT VT = N->getValueType(0);
8551 if (EltSize > VT.getVectorElementType().getSizeInBits())
8554 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8557 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8558 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8559 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8563 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8565 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8570 APFloat APF = C->getValueAPF();
8571 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8572 != APFloat::opOK || !isExact)
8575 c0 = (I == 0) ? cN : c0;
8576 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8583 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8584 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8585 /// when the VMUL has a constant operand that is a power of 2.
8587 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8588 /// vmul.f32 d16, d17, d16
8589 /// vcvt.s32.f32 d16, d16
8591 /// vcvt.s32.f32 d16, d16, #3
8592 static SDValue PerformVCVTCombine(SDNode *N,
8593 TargetLowering::DAGCombinerInfo &DCI,
8594 const ARMSubtarget *Subtarget) {
8595 SelectionDAG &DAG = DCI.DAG;
8596 SDValue Op = N->getOperand(0);
8598 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8599 Op.getOpcode() != ISD::FMUL)
8603 SDValue N0 = Op->getOperand(0);
8604 SDValue ConstVec = Op->getOperand(1);
8605 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8607 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8608 !isConstVecPow2(ConstVec, isSigned, C))
8611 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8612 Intrinsic::arm_neon_vcvtfp2fxu;
8613 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8615 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8616 DAG.getConstant(Log2_64(C), MVT::i32));
8619 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8620 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8621 /// when the VDIV has a constant operand that is a power of 2.
8623 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8624 /// vcvt.f32.s32 d16, d16
8625 /// vdiv.f32 d16, d17, d16
8627 /// vcvt.f32.s32 d16, d16, #3
8628 static SDValue PerformVDIVCombine(SDNode *N,
8629 TargetLowering::DAGCombinerInfo &DCI,
8630 const ARMSubtarget *Subtarget) {
8631 SelectionDAG &DAG = DCI.DAG;
8632 SDValue Op = N->getOperand(0);
8633 unsigned OpOpcode = Op.getNode()->getOpcode();
8635 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8636 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8640 SDValue ConstVec = N->getOperand(1);
8641 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8643 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8644 !isConstVecPow2(ConstVec, isSigned, C))
8647 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8648 Intrinsic::arm_neon_vcvtfxu2fp;
8649 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8651 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8652 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8655 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8656 /// operand of a vector shift operation, where all the elements of the
8657 /// build_vector must have the same constant integer value.
8658 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8659 // Ignore bit_converts.
8660 while (Op.getOpcode() == ISD::BITCAST)
8661 Op = Op.getOperand(0);
8662 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8663 APInt SplatBits, SplatUndef;
8664 unsigned SplatBitSize;
8666 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8667 HasAnyUndefs, ElementBits) ||
8668 SplatBitSize > ElementBits)
8670 Cnt = SplatBits.getSExtValue();
8674 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8675 /// operand of a vector shift left operation. That value must be in the range:
8676 /// 0 <= Value < ElementBits for a left shift; or
8677 /// 0 <= Value <= ElementBits for a long left shift.
8678 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8679 assert(VT.isVector() && "vector shift count is not a vector type");
8680 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8681 if (! getVShiftImm(Op, ElementBits, Cnt))
8683 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8686 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8687 /// operand of a vector shift right operation. For a shift opcode, the value
8688 /// is positive, but for an intrinsic the value count must be negative. The
8689 /// absolute value must be in the range:
8690 /// 1 <= |Value| <= ElementBits for a right shift; or
8691 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8692 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8694 assert(VT.isVector() && "vector shift count is not a vector type");
8695 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8696 if (! getVShiftImm(Op, ElementBits, Cnt))
8700 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8703 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8704 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8705 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8708 // Don't do anything for most intrinsics.
8711 // Vector shifts: check for immediate versions and lower them.
8712 // Note: This is done during DAG combining instead of DAG legalizing because
8713 // the build_vectors for 64-bit vector element shift counts are generally
8714 // not legal, and it is hard to see their values after they get legalized to
8715 // loads from a constant pool.
8716 case Intrinsic::arm_neon_vshifts:
8717 case Intrinsic::arm_neon_vshiftu:
8718 case Intrinsic::arm_neon_vshiftls:
8719 case Intrinsic::arm_neon_vshiftlu:
8720 case Intrinsic::arm_neon_vshiftn:
8721 case Intrinsic::arm_neon_vrshifts:
8722 case Intrinsic::arm_neon_vrshiftu:
8723 case Intrinsic::arm_neon_vrshiftn:
8724 case Intrinsic::arm_neon_vqshifts:
8725 case Intrinsic::arm_neon_vqshiftu:
8726 case Intrinsic::arm_neon_vqshiftsu:
8727 case Intrinsic::arm_neon_vqshiftns:
8728 case Intrinsic::arm_neon_vqshiftnu:
8729 case Intrinsic::arm_neon_vqshiftnsu:
8730 case Intrinsic::arm_neon_vqrshiftns:
8731 case Intrinsic::arm_neon_vqrshiftnu:
8732 case Intrinsic::arm_neon_vqrshiftnsu: {
8733 EVT VT = N->getOperand(1).getValueType();
8735 unsigned VShiftOpc = 0;
8738 case Intrinsic::arm_neon_vshifts:
8739 case Intrinsic::arm_neon_vshiftu:
8740 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8741 VShiftOpc = ARMISD::VSHL;
8744 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8745 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8746 ARMISD::VSHRs : ARMISD::VSHRu);
8751 case Intrinsic::arm_neon_vshiftls:
8752 case Intrinsic::arm_neon_vshiftlu:
8753 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8755 llvm_unreachable("invalid shift count for vshll intrinsic");
8757 case Intrinsic::arm_neon_vrshifts:
8758 case Intrinsic::arm_neon_vrshiftu:
8759 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8763 case Intrinsic::arm_neon_vqshifts:
8764 case Intrinsic::arm_neon_vqshiftu:
8765 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8769 case Intrinsic::arm_neon_vqshiftsu:
8770 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8772 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8774 case Intrinsic::arm_neon_vshiftn:
8775 case Intrinsic::arm_neon_vrshiftn:
8776 case Intrinsic::arm_neon_vqshiftns:
8777 case Intrinsic::arm_neon_vqshiftnu:
8778 case Intrinsic::arm_neon_vqshiftnsu:
8779 case Intrinsic::arm_neon_vqrshiftns:
8780 case Intrinsic::arm_neon_vqrshiftnu:
8781 case Intrinsic::arm_neon_vqrshiftnsu:
8782 // Narrowing shifts require an immediate right shift.
8783 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8785 llvm_unreachable("invalid shift count for narrowing vector shift "
8789 llvm_unreachable("unhandled vector shift");
8793 case Intrinsic::arm_neon_vshifts:
8794 case Intrinsic::arm_neon_vshiftu:
8795 // Opcode already set above.
8797 case Intrinsic::arm_neon_vshiftls:
8798 case Intrinsic::arm_neon_vshiftlu:
8799 if (Cnt == VT.getVectorElementType().getSizeInBits())
8800 VShiftOpc = ARMISD::VSHLLi;
8802 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8803 ARMISD::VSHLLs : ARMISD::VSHLLu);
8805 case Intrinsic::arm_neon_vshiftn:
8806 VShiftOpc = ARMISD::VSHRN; break;
8807 case Intrinsic::arm_neon_vrshifts:
8808 VShiftOpc = ARMISD::VRSHRs; break;
8809 case Intrinsic::arm_neon_vrshiftu:
8810 VShiftOpc = ARMISD::VRSHRu; break;
8811 case Intrinsic::arm_neon_vrshiftn:
8812 VShiftOpc = ARMISD::VRSHRN; break;
8813 case Intrinsic::arm_neon_vqshifts:
8814 VShiftOpc = ARMISD::VQSHLs; break;
8815 case Intrinsic::arm_neon_vqshiftu:
8816 VShiftOpc = ARMISD::VQSHLu; break;
8817 case Intrinsic::arm_neon_vqshiftsu:
8818 VShiftOpc = ARMISD::VQSHLsu; break;
8819 case Intrinsic::arm_neon_vqshiftns:
8820 VShiftOpc = ARMISD::VQSHRNs; break;
8821 case Intrinsic::arm_neon_vqshiftnu:
8822 VShiftOpc = ARMISD::VQSHRNu; break;
8823 case Intrinsic::arm_neon_vqshiftnsu:
8824 VShiftOpc = ARMISD::VQSHRNsu; break;
8825 case Intrinsic::arm_neon_vqrshiftns:
8826 VShiftOpc = ARMISD::VQRSHRNs; break;
8827 case Intrinsic::arm_neon_vqrshiftnu:
8828 VShiftOpc = ARMISD::VQRSHRNu; break;
8829 case Intrinsic::arm_neon_vqrshiftnsu:
8830 VShiftOpc = ARMISD::VQRSHRNsu; break;
8833 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8834 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8837 case Intrinsic::arm_neon_vshiftins: {
8838 EVT VT = N->getOperand(1).getValueType();
8840 unsigned VShiftOpc = 0;
8842 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8843 VShiftOpc = ARMISD::VSLI;
8844 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8845 VShiftOpc = ARMISD::VSRI;
8847 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8850 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8851 N->getOperand(1), N->getOperand(2),
8852 DAG.getConstant(Cnt, MVT::i32));
8855 case Intrinsic::arm_neon_vqrshifts:
8856 case Intrinsic::arm_neon_vqrshiftu:
8857 // No immediate versions of these to check for.
8864 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8865 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8866 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8867 /// vector element shift counts are generally not legal, and it is hard to see
8868 /// their values after they get legalized to loads from a constant pool.
8869 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8870 const ARMSubtarget *ST) {
8871 EVT VT = N->getValueType(0);
8872 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8873 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8874 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8875 SDValue N1 = N->getOperand(1);
8876 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8877 SDValue N0 = N->getOperand(0);
8878 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8879 DAG.MaskedValueIsZero(N0.getOperand(0),
8880 APInt::getHighBitsSet(32, 16)))
8881 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8885 // Nothing to be done for scalar shifts.
8886 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8887 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8890 assert(ST->hasNEON() && "unexpected vector shift");
8893 switch (N->getOpcode()) {
8894 default: llvm_unreachable("unexpected shift opcode");
8897 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8898 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8899 DAG.getConstant(Cnt, MVT::i32));
8904 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8905 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8906 ARMISD::VSHRs : ARMISD::VSHRu);
8907 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8908 DAG.getConstant(Cnt, MVT::i32));
8914 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8915 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8916 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8917 const ARMSubtarget *ST) {
8918 SDValue N0 = N->getOperand(0);
8920 // Check for sign- and zero-extensions of vector extract operations of 8-
8921 // and 16-bit vector elements. NEON supports these directly. They are
8922 // handled during DAG combining because type legalization will promote them
8923 // to 32-bit types and it is messy to recognize the operations after that.
8924 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8925 SDValue Vec = N0.getOperand(0);
8926 SDValue Lane = N0.getOperand(1);
8927 EVT VT = N->getValueType(0);
8928 EVT EltVT = N0.getValueType();
8929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8931 if (VT == MVT::i32 &&
8932 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8933 TLI.isTypeLegal(Vec.getValueType()) &&
8934 isa<ConstantSDNode>(Lane)) {
8937 switch (N->getOpcode()) {
8938 default: llvm_unreachable("unexpected opcode");
8939 case ISD::SIGN_EXTEND:
8940 Opc = ARMISD::VGETLANEs;
8942 case ISD::ZERO_EXTEND:
8943 case ISD::ANY_EXTEND:
8944 Opc = ARMISD::VGETLANEu;
8947 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8954 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8955 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8956 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8957 const ARMSubtarget *ST) {
8958 // If the target supports NEON, try to use vmax/vmin instructions for f32
8959 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8960 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8961 // a NaN; only do the transformation when it matches that behavior.
8963 // For now only do this when using NEON for FP operations; if using VFP, it
8964 // is not obvious that the benefit outweighs the cost of switching to the
8966 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8967 N->getValueType(0) != MVT::f32)
8970 SDValue CondLHS = N->getOperand(0);
8971 SDValue CondRHS = N->getOperand(1);
8972 SDValue LHS = N->getOperand(2);
8973 SDValue RHS = N->getOperand(3);
8974 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8976 unsigned Opcode = 0;
8978 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8979 IsReversed = false; // x CC y ? x : y
8980 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8981 IsReversed = true ; // x CC y ? y : x
8995 // If LHS is NaN, an ordered comparison will be false and the result will
8996 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8997 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8998 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8999 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9001 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9002 // will return -0, so vmin can only be used for unsafe math or if one of
9003 // the operands is known to be nonzero.
9004 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9005 !DAG.getTarget().Options.UnsafeFPMath &&
9006 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9008 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9017 // If LHS is NaN, an ordered comparison will be false and the result will
9018 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9019 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9020 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9021 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9023 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9024 // will return +0, so vmax can only be used for unsafe math or if one of
9025 // the operands is known to be nonzero.
9026 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9027 !DAG.getTarget().Options.UnsafeFPMath &&
9028 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9030 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9036 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
9039 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9041 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9042 SDValue Cmp = N->getOperand(4);
9043 if (Cmp.getOpcode() != ARMISD::CMPZ)
9044 // Only looking at EQ and NE cases.
9047 EVT VT = N->getValueType(0);
9048 DebugLoc dl = N->getDebugLoc();
9049 SDValue LHS = Cmp.getOperand(0);
9050 SDValue RHS = Cmp.getOperand(1);
9051 SDValue FalseVal = N->getOperand(0);
9052 SDValue TrueVal = N->getOperand(1);
9053 SDValue ARMcc = N->getOperand(2);
9054 ARMCC::CondCodes CC =
9055 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9073 /// FIXME: Turn this into a target neutral optimization?
9075 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9076 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9077 N->getOperand(3), Cmp);
9078 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9080 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9081 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9082 N->getOperand(3), NewCmp);
9085 if (Res.getNode()) {
9086 APInt KnownZero, KnownOne;
9087 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9088 // Capture demanded bits information that would be otherwise lost.
9089 if (KnownZero == 0xfffffffe)
9090 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9091 DAG.getValueType(MVT::i1));
9092 else if (KnownZero == 0xffffff00)
9093 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9094 DAG.getValueType(MVT::i8));
9095 else if (KnownZero == 0xffff0000)
9096 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9097 DAG.getValueType(MVT::i16));
9103 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9104 DAGCombinerInfo &DCI) const {
9105 switch (N->getOpcode()) {
9107 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9108 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9109 case ISD::SUB: return PerformSUBCombine(N, DCI);
9110 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9111 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9112 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9113 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9114 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9115 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9116 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9117 case ISD::STORE: return PerformSTORECombine(N, DCI);
9118 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9119 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9120 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9121 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9122 case ISD::FP_TO_SINT:
9123 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9124 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9125 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9128 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9129 case ISD::SIGN_EXTEND:
9130 case ISD::ZERO_EXTEND:
9131 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9132 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9133 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9134 case ARMISD::VLD2DUP:
9135 case ARMISD::VLD3DUP:
9136 case ARMISD::VLD4DUP:
9137 return CombineBaseUpdate(N, DCI);
9138 case ISD::INTRINSIC_VOID:
9139 case ISD::INTRINSIC_W_CHAIN:
9140 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9141 case Intrinsic::arm_neon_vld1:
9142 case Intrinsic::arm_neon_vld2:
9143 case Intrinsic::arm_neon_vld3:
9144 case Intrinsic::arm_neon_vld4:
9145 case Intrinsic::arm_neon_vld2lane:
9146 case Intrinsic::arm_neon_vld3lane:
9147 case Intrinsic::arm_neon_vld4lane:
9148 case Intrinsic::arm_neon_vst1:
9149 case Intrinsic::arm_neon_vst2:
9150 case Intrinsic::arm_neon_vst3:
9151 case Intrinsic::arm_neon_vst4:
9152 case Intrinsic::arm_neon_vst2lane:
9153 case Intrinsic::arm_neon_vst3lane:
9154 case Intrinsic::arm_neon_vst4lane:
9155 return CombineBaseUpdate(N, DCI);
9163 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9165 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9168 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9169 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9170 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9172 switch (VT.getSimpleVT().SimpleTy) {
9178 // Unaligned access can use (for example) LRDB, LRDH, LDR
9179 return AllowsUnaligned;
9182 // For any little-endian targets with neon, we can support unaligned ld/st
9183 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9184 // A big-endian target may also explictly support unaligned accesses
9185 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9189 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9190 unsigned AlignCheck) {
9191 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9192 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9195 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9196 unsigned DstAlign, unsigned SrcAlign,
9199 MachineFunction &MF) const {
9200 const Function *F = MF.getFunction();
9202 // See if we can use NEON instructions for this...
9204 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9205 Subtarget->hasNEON()) {
9206 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9208 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9213 // Lowering to i32/i16 if the size permits.
9216 } else if (Size >= 2) {
9220 // Let the target-independent logic figure it out.
9224 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9229 switch (VT.getSimpleVT().SimpleTy) {
9230 default: return false;
9245 if ((V & (Scale - 1)) != 0)
9248 return V == (V & ((1LL << 5) - 1));
9251 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9252 const ARMSubtarget *Subtarget) {
9259 switch (VT.getSimpleVT().SimpleTy) {
9260 default: return false;
9265 // + imm12 or - imm8
9267 return V == (V & ((1LL << 8) - 1));
9268 return V == (V & ((1LL << 12) - 1));
9271 // Same as ARM mode. FIXME: NEON?
9272 if (!Subtarget->hasVFP2())
9277 return V == (V & ((1LL << 8) - 1));
9281 /// isLegalAddressImmediate - Return true if the integer value can be used
9282 /// as the offset of the target addressing mode for load / store of the
9284 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9285 const ARMSubtarget *Subtarget) {
9292 if (Subtarget->isThumb1Only())
9293 return isLegalT1AddressImmediate(V, VT);
9294 else if (Subtarget->isThumb2())
9295 return isLegalT2AddressImmediate(V, VT, Subtarget);
9300 switch (VT.getSimpleVT().SimpleTy) {
9301 default: return false;
9306 return V == (V & ((1LL << 12) - 1));
9309 return V == (V & ((1LL << 8) - 1));
9312 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9317 return V == (V & ((1LL << 8) - 1));
9321 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9323 int Scale = AM.Scale;
9327 switch (VT.getSimpleVT().SimpleTy) {
9328 default: return false;
9337 return Scale == 2 || Scale == 4 || Scale == 8;
9340 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9344 // Note, we allow "void" uses (basically, uses that aren't loads or
9345 // stores), because arm allows folding a scale into many arithmetic
9346 // operations. This should be made more precise and revisited later.
9348 // Allow r << imm, but the imm has to be a multiple of two.
9349 if (Scale & 1) return false;
9350 return isPowerOf2_32(Scale);
9354 /// isLegalAddressingMode - Return true if the addressing mode represented
9355 /// by AM is legal for this target, for a load/store of the specified type.
9356 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9358 EVT VT = getValueType(Ty, true);
9359 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9362 // Can never fold addr of global into load/store.
9367 case 0: // no scale reg, must be "r+i" or "r", or "i".
9370 if (Subtarget->isThumb1Only())
9374 // ARM doesn't support any R+R*scale+imm addr modes.
9381 if (Subtarget->isThumb2())
9382 return isLegalT2ScaledAddressingMode(AM, VT);
9384 int Scale = AM.Scale;
9385 switch (VT.getSimpleVT().SimpleTy) {
9386 default: return false;
9390 if (Scale < 0) Scale = -Scale;
9394 return isPowerOf2_32(Scale & ~1);
9398 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9403 // Note, we allow "void" uses (basically, uses that aren't loads or
9404 // stores), because arm allows folding a scale into many arithmetic
9405 // operations. This should be made more precise and revisited later.
9407 // Allow r << imm, but the imm has to be a multiple of two.
9408 if (Scale & 1) return false;
9409 return isPowerOf2_32(Scale);
9415 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9416 /// icmp immediate, that is the target has icmp instructions which can compare
9417 /// a register against the immediate without having to materialize the
9418 /// immediate into a register.
9419 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9420 // Thumb2 and ARM modes can use cmn for negative immediates.
9421 if (!Subtarget->isThumb())
9422 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9423 if (Subtarget->isThumb2())
9424 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9425 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9426 return Imm >= 0 && Imm <= 255;
9429 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9430 /// *or sub* immediate, that is the target has add or sub instructions which can
9431 /// add a register with the immediate without having to materialize the
9432 /// immediate into a register.
9433 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9434 // Same encoding for add/sub, just flip the sign.
9435 int64_t AbsImm = llvm::abs64(Imm);
9436 if (!Subtarget->isThumb())
9437 return ARM_AM::getSOImmVal(AbsImm) != -1;
9438 if (Subtarget->isThumb2())
9439 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9440 // Thumb1 only has 8-bit unsigned immediate.
9441 return AbsImm >= 0 && AbsImm <= 255;
9444 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9445 bool isSEXTLoad, SDValue &Base,
9446 SDValue &Offset, bool &isInc,
9447 SelectionDAG &DAG) {
9448 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9451 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9453 Base = Ptr->getOperand(0);
9454 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9455 int RHSC = (int)RHS->getZExtValue();
9456 if (RHSC < 0 && RHSC > -256) {
9457 assert(Ptr->getOpcode() == ISD::ADD);
9459 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9463 isInc = (Ptr->getOpcode() == ISD::ADD);
9464 Offset = Ptr->getOperand(1);
9466 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9468 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9469 int RHSC = (int)RHS->getZExtValue();
9470 if (RHSC < 0 && RHSC > -0x1000) {
9471 assert(Ptr->getOpcode() == ISD::ADD);
9473 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9474 Base = Ptr->getOperand(0);
9479 if (Ptr->getOpcode() == ISD::ADD) {
9481 ARM_AM::ShiftOpc ShOpcVal=
9482 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9483 if (ShOpcVal != ARM_AM::no_shift) {
9484 Base = Ptr->getOperand(1);
9485 Offset = Ptr->getOperand(0);
9487 Base = Ptr->getOperand(0);
9488 Offset = Ptr->getOperand(1);
9493 isInc = (Ptr->getOpcode() == ISD::ADD);
9494 Base = Ptr->getOperand(0);
9495 Offset = Ptr->getOperand(1);
9499 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9503 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9504 bool isSEXTLoad, SDValue &Base,
9505 SDValue &Offset, bool &isInc,
9506 SelectionDAG &DAG) {
9507 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9510 Base = Ptr->getOperand(0);
9511 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9512 int RHSC = (int)RHS->getZExtValue();
9513 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9514 assert(Ptr->getOpcode() == ISD::ADD);
9516 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9518 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9519 isInc = Ptr->getOpcode() == ISD::ADD;
9520 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9528 /// getPreIndexedAddressParts - returns true by value, base pointer and
9529 /// offset pointer and addressing mode by reference if the node's address
9530 /// can be legally represented as pre-indexed load / store address.
9532 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9534 ISD::MemIndexedMode &AM,
9535 SelectionDAG &DAG) const {
9536 if (Subtarget->isThumb1Only())
9541 bool isSEXTLoad = false;
9542 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9543 Ptr = LD->getBasePtr();
9544 VT = LD->getMemoryVT();
9545 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9546 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9547 Ptr = ST->getBasePtr();
9548 VT = ST->getMemoryVT();
9553 bool isLegal = false;
9554 if (Subtarget->isThumb2())
9555 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9556 Offset, isInc, DAG);
9558 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9559 Offset, isInc, DAG);
9563 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9567 /// getPostIndexedAddressParts - returns true by value, base pointer and
9568 /// offset pointer and addressing mode by reference if this node can be
9569 /// combined with a load / store to form a post-indexed load / store.
9570 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9573 ISD::MemIndexedMode &AM,
9574 SelectionDAG &DAG) const {
9575 if (Subtarget->isThumb1Only())
9580 bool isSEXTLoad = false;
9581 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9582 VT = LD->getMemoryVT();
9583 Ptr = LD->getBasePtr();
9584 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9585 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9586 VT = ST->getMemoryVT();
9587 Ptr = ST->getBasePtr();
9592 bool isLegal = false;
9593 if (Subtarget->isThumb2())
9594 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9597 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9603 // Swap base ptr and offset to catch more post-index load / store when
9604 // it's legal. In Thumb2 mode, offset must be an immediate.
9605 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9606 !Subtarget->isThumb2())
9607 std::swap(Base, Offset);
9609 // Post-indexed load / store update the base pointer.
9614 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9618 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9621 const SelectionDAG &DAG,
9622 unsigned Depth) const {
9623 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9624 switch (Op.getOpcode()) {
9626 case ARMISD::CMOV: {
9627 // Bits are known zero/one if known on the LHS and RHS.
9628 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9629 if (KnownZero == 0 && KnownOne == 0) return;
9631 APInt KnownZeroRHS, KnownOneRHS;
9632 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9633 KnownZero &= KnownZeroRHS;
9634 KnownOne &= KnownOneRHS;
9640 //===----------------------------------------------------------------------===//
9641 // ARM Inline Assembly Support
9642 //===----------------------------------------------------------------------===//
9644 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9645 // Looking for "rev" which is V6+.
9646 if (!Subtarget->hasV6Ops())
9649 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9650 std::string AsmStr = IA->getAsmString();
9651 SmallVector<StringRef, 4> AsmPieces;
9652 SplitString(AsmStr, AsmPieces, ";\n");
9654 switch (AsmPieces.size()) {
9655 default: return false;
9657 AsmStr = AsmPieces[0];
9659 SplitString(AsmStr, AsmPieces, " \t,");
9662 if (AsmPieces.size() == 3 &&
9663 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9664 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9665 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9666 if (Ty && Ty->getBitWidth() == 32)
9667 return IntrinsicLowering::LowerToByteSwap(CI);
9675 /// getConstraintType - Given a constraint letter, return the type of
9676 /// constraint it is for this target.
9677 ARMTargetLowering::ConstraintType
9678 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9679 if (Constraint.size() == 1) {
9680 switch (Constraint[0]) {
9682 case 'l': return C_RegisterClass;
9683 case 'w': return C_RegisterClass;
9684 case 'h': return C_RegisterClass;
9685 case 'x': return C_RegisterClass;
9686 case 't': return C_RegisterClass;
9687 case 'j': return C_Other; // Constant for movw.
9688 // An address with a single base register. Due to the way we
9689 // currently handle addresses it is the same as an 'r' memory constraint.
9690 case 'Q': return C_Memory;
9692 } else if (Constraint.size() == 2) {
9693 switch (Constraint[0]) {
9695 // All 'U+' constraints are addresses.
9696 case 'U': return C_Memory;
9699 return TargetLowering::getConstraintType(Constraint);
9702 /// Examine constraint type and operand type and determine a weight value.
9703 /// This object must already have been set up with the operand type
9704 /// and the current alternative constraint selected.
9705 TargetLowering::ConstraintWeight
9706 ARMTargetLowering::getSingleConstraintMatchWeight(
9707 AsmOperandInfo &info, const char *constraint) const {
9708 ConstraintWeight weight = CW_Invalid;
9709 Value *CallOperandVal = info.CallOperandVal;
9710 // If we don't have a value, we can't do a match,
9711 // but allow it at the lowest weight.
9712 if (CallOperandVal == NULL)
9714 Type *type = CallOperandVal->getType();
9715 // Look at the constraint type.
9716 switch (*constraint) {
9718 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9721 if (type->isIntegerTy()) {
9722 if (Subtarget->isThumb())
9723 weight = CW_SpecificReg;
9725 weight = CW_Register;
9729 if (type->isFloatingPointTy())
9730 weight = CW_Register;
9736 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9738 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9740 if (Constraint.size() == 1) {
9741 // GCC ARM Constraint Letters
9742 switch (Constraint[0]) {
9743 case 'l': // Low regs or general regs.
9744 if (Subtarget->isThumb())
9745 return RCPair(0U, &ARM::tGPRRegClass);
9746 return RCPair(0U, &ARM::GPRRegClass);
9747 case 'h': // High regs or no regs.
9748 if (Subtarget->isThumb())
9749 return RCPair(0U, &ARM::hGPRRegClass);
9752 return RCPair(0U, &ARM::GPRRegClass);
9755 return RCPair(0U, &ARM::SPRRegClass);
9756 if (VT.getSizeInBits() == 64)
9757 return RCPair(0U, &ARM::DPRRegClass);
9758 if (VT.getSizeInBits() == 128)
9759 return RCPair(0U, &ARM::QPRRegClass);
9763 return RCPair(0U, &ARM::SPR_8RegClass);
9764 if (VT.getSizeInBits() == 64)
9765 return RCPair(0U, &ARM::DPR_8RegClass);
9766 if (VT.getSizeInBits() == 128)
9767 return RCPair(0U, &ARM::QPR_8RegClass);
9771 return RCPair(0U, &ARM::SPRRegClass);
9775 if (StringRef("{cc}").equals_lower(Constraint))
9776 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9778 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9781 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9782 /// vector. If it is invalid, don't add anything to Ops.
9783 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9784 std::string &Constraint,
9785 std::vector<SDValue>&Ops,
9786 SelectionDAG &DAG) const {
9787 SDValue Result(0, 0);
9789 // Currently only support length 1 constraints.
9790 if (Constraint.length() != 1) return;
9792 char ConstraintLetter = Constraint[0];
9793 switch (ConstraintLetter) {
9796 case 'I': case 'J': case 'K': case 'L':
9797 case 'M': case 'N': case 'O':
9798 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9802 int64_t CVal64 = C->getSExtValue();
9803 int CVal = (int) CVal64;
9804 // None of these constraints allow values larger than 32 bits. Check
9805 // that the value fits in an int.
9809 switch (ConstraintLetter) {
9811 // Constant suitable for movw, must be between 0 and
9813 if (Subtarget->hasV6T2Ops())
9814 if (CVal >= 0 && CVal <= 65535)
9818 if (Subtarget->isThumb1Only()) {
9819 // This must be a constant between 0 and 255, for ADD
9821 if (CVal >= 0 && CVal <= 255)
9823 } else if (Subtarget->isThumb2()) {
9824 // A constant that can be used as an immediate value in a
9825 // data-processing instruction.
9826 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9829 // A constant that can be used as an immediate value in a
9830 // data-processing instruction.
9831 if (ARM_AM::getSOImmVal(CVal) != -1)
9837 if (Subtarget->isThumb()) { // FIXME thumb2
9838 // This must be a constant between -255 and -1, for negated ADD
9839 // immediates. This can be used in GCC with an "n" modifier that
9840 // prints the negated value, for use with SUB instructions. It is
9841 // not useful otherwise but is implemented for compatibility.
9842 if (CVal >= -255 && CVal <= -1)
9845 // This must be a constant between -4095 and 4095. It is not clear
9846 // what this constraint is intended for. Implemented for
9847 // compatibility with GCC.
9848 if (CVal >= -4095 && CVal <= 4095)
9854 if (Subtarget->isThumb1Only()) {
9855 // A 32-bit value where only one byte has a nonzero value. Exclude
9856 // zero to match GCC. This constraint is used by GCC internally for
9857 // constants that can be loaded with a move/shift combination.
9858 // It is not useful otherwise but is implemented for compatibility.
9859 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9861 } else if (Subtarget->isThumb2()) {
9862 // A constant whose bitwise inverse can be used as an immediate
9863 // value in a data-processing instruction. This can be used in GCC
9864 // with a "B" modifier that prints the inverted value, for use with
9865 // BIC and MVN instructions. It is not useful otherwise but is
9866 // implemented for compatibility.
9867 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9870 // A constant whose bitwise inverse can be used as an immediate
9871 // value in a data-processing instruction. This can be used in GCC
9872 // with a "B" modifier that prints the inverted value, for use with
9873 // BIC and MVN instructions. It is not useful otherwise but is
9874 // implemented for compatibility.
9875 if (ARM_AM::getSOImmVal(~CVal) != -1)
9881 if (Subtarget->isThumb1Only()) {
9882 // This must be a constant between -7 and 7,
9883 // for 3-operand ADD/SUB immediate instructions.
9884 if (CVal >= -7 && CVal < 7)
9886 } else if (Subtarget->isThumb2()) {
9887 // A constant whose negation can be used as an immediate value in a
9888 // data-processing instruction. This can be used in GCC with an "n"
9889 // modifier that prints the negated value, for use with SUB
9890 // instructions. It is not useful otherwise but is implemented for
9892 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9895 // A constant whose negation can be used as an immediate value in a
9896 // data-processing instruction. This can be used in GCC with an "n"
9897 // modifier that prints the negated value, for use with SUB
9898 // instructions. It is not useful otherwise but is implemented for
9900 if (ARM_AM::getSOImmVal(-CVal) != -1)
9906 if (Subtarget->isThumb()) { // FIXME thumb2
9907 // This must be a multiple of 4 between 0 and 1020, for
9908 // ADD sp + immediate.
9909 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9912 // A power of two or a constant between 0 and 32. This is used in
9913 // GCC for the shift amount on shifted register operands, but it is
9914 // useful in general for any shift amounts.
9915 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9921 if (Subtarget->isThumb()) { // FIXME thumb2
9922 // This must be a constant between 0 and 31, for shift amounts.
9923 if (CVal >= 0 && CVal <= 31)
9929 if (Subtarget->isThumb()) { // FIXME thumb2
9930 // This must be a multiple of 4 between -508 and 508, for
9931 // ADD/SUB sp = sp + immediate.
9932 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9937 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9941 if (Result.getNode()) {
9942 Ops.push_back(Result);
9945 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9949 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9950 // The ARM target isn't yet aware of offsets.
9954 bool ARM::isBitFieldInvertedMask(unsigned v) {
9955 if (v == 0xffffffff)
9957 // there can be 1's on either or both "outsides", all the "inside"
9959 unsigned int lsb = 0, msb = 31;
9960 while (v & (1 << msb)) --msb;
9961 while (v & (1 << lsb)) ++lsb;
9962 for (unsigned int i = lsb; i <= msb; ++i) {
9969 /// isFPImmLegal - Returns true if the target can instruction select the
9970 /// specified FP immediate natively. If false, the legalizer will
9971 /// materialize the FP immediate as a load from a constant pool.
9972 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9973 if (!Subtarget->hasVFP3())
9976 return ARM_AM::getFP32Imm(Imm) != -1;
9978 return ARM_AM::getFP64Imm(Imm) != -1;
9982 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9983 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9984 /// specified in the intrinsic calls.
9985 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9987 unsigned Intrinsic) const {
9988 switch (Intrinsic) {
9989 case Intrinsic::arm_neon_vld1:
9990 case Intrinsic::arm_neon_vld2:
9991 case Intrinsic::arm_neon_vld3:
9992 case Intrinsic::arm_neon_vld4:
9993 case Intrinsic::arm_neon_vld2lane:
9994 case Intrinsic::arm_neon_vld3lane:
9995 case Intrinsic::arm_neon_vld4lane: {
9996 Info.opc = ISD::INTRINSIC_W_CHAIN;
9997 // Conservatively set memVT to the entire set of vectors loaded.
9998 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9999 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10000 Info.ptrVal = I.getArgOperand(0);
10002 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10003 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10004 Info.vol = false; // volatile loads with NEON intrinsics not supported
10005 Info.readMem = true;
10006 Info.writeMem = false;
10009 case Intrinsic::arm_neon_vst1:
10010 case Intrinsic::arm_neon_vst2:
10011 case Intrinsic::arm_neon_vst3:
10012 case Intrinsic::arm_neon_vst4:
10013 case Intrinsic::arm_neon_vst2lane:
10014 case Intrinsic::arm_neon_vst3lane:
10015 case Intrinsic::arm_neon_vst4lane: {
10016 Info.opc = ISD::INTRINSIC_VOID;
10017 // Conservatively set memVT to the entire set of vectors stored.
10018 unsigned NumElts = 0;
10019 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10020 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10021 if (!ArgTy->isVectorTy())
10023 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10025 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10026 Info.ptrVal = I.getArgOperand(0);
10028 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10029 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10030 Info.vol = false; // volatile stores with NEON intrinsics not supported
10031 Info.readMem = false;
10032 Info.writeMem = true;
10035 case Intrinsic::arm_strexd: {
10036 Info.opc = ISD::INTRINSIC_W_CHAIN;
10037 Info.memVT = MVT::i64;
10038 Info.ptrVal = I.getArgOperand(2);
10042 Info.readMem = false;
10043 Info.writeMem = true;
10046 case Intrinsic::arm_ldrexd: {
10047 Info.opc = ISD::INTRINSIC_W_CHAIN;
10048 Info.memVT = MVT::i64;
10049 Info.ptrVal = I.getArgOperand(0);
10053 Info.readMem = true;
10054 Info.writeMem = false;