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_LOAD_MIN, MVT::i64, Custom);
700 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
701 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
702 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
703 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
704 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
705 setInsertFencesForAtomic(true);
707 // Set them all for expansion, which will force libcalls.
708 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
709 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
710 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
711 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
712 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
713 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
714 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
715 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
716 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
717 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
718 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
719 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
720 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
721 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
722 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
723 // Unordered/Monotonic case.
724 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
725 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
726 // Since the libcalls include locking, fold in the fences
727 setShouldFoldAtomicFences(true);
730 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
732 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
733 if (!Subtarget->hasV6Ops()) {
734 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
735 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
737 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
739 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
740 !Subtarget->isThumb1Only()) {
741 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
742 // iff target supports vfp2.
743 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
744 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
747 // We want to custom lower some of our intrinsics.
748 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
749 if (Subtarget->isTargetDarwin()) {
750 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
751 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
752 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
755 setOperationAction(ISD::SETCC, MVT::i32, Expand);
756 setOperationAction(ISD::SETCC, MVT::f32, Expand);
757 setOperationAction(ISD::SETCC, MVT::f64, Expand);
758 setOperationAction(ISD::SELECT, MVT::i32, Custom);
759 setOperationAction(ISD::SELECT, MVT::f32, Custom);
760 setOperationAction(ISD::SELECT, MVT::f64, Custom);
761 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
762 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
763 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
765 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
766 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
767 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
768 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
769 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
771 // We don't support sin/cos/fmod/copysign/pow
772 setOperationAction(ISD::FSIN, MVT::f64, Expand);
773 setOperationAction(ISD::FSIN, MVT::f32, Expand);
774 setOperationAction(ISD::FCOS, MVT::f32, Expand);
775 setOperationAction(ISD::FCOS, MVT::f64, Expand);
776 setOperationAction(ISD::FREM, MVT::f64, Expand);
777 setOperationAction(ISD::FREM, MVT::f32, Expand);
778 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
779 !Subtarget->isThumb1Only()) {
780 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
781 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
783 setOperationAction(ISD::FPOW, MVT::f64, Expand);
784 setOperationAction(ISD::FPOW, MVT::f32, Expand);
786 if (!Subtarget->hasVFP4()) {
787 setOperationAction(ISD::FMA, MVT::f64, Expand);
788 setOperationAction(ISD::FMA, MVT::f32, Expand);
791 // Various VFP goodness
792 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
793 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
794 if (Subtarget->hasVFP2()) {
795 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
796 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
797 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
798 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
800 // Special handling for half-precision FP.
801 if (!Subtarget->hasFP16()) {
802 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
803 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
807 // We have target-specific dag combine patterns for the following nodes:
808 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
809 setTargetDAGCombine(ISD::ADD);
810 setTargetDAGCombine(ISD::SUB);
811 setTargetDAGCombine(ISD::MUL);
812 setTargetDAGCombine(ISD::AND);
813 setTargetDAGCombine(ISD::OR);
814 setTargetDAGCombine(ISD::XOR);
816 if (Subtarget->hasV6Ops())
817 setTargetDAGCombine(ISD::SRL);
819 setStackPointerRegisterToSaveRestore(ARM::SP);
821 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
822 !Subtarget->hasVFP2())
823 setSchedulingPreference(Sched::RegPressure);
825 setSchedulingPreference(Sched::Hybrid);
827 //// temporary - rewrite interface to use type
828 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
829 maxStoresPerMemset = 16;
830 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
832 // On ARM arguments smaller than 4 bytes are extended, so all arguments
833 // are at least 4 bytes aligned.
834 setMinStackArgumentAlignment(4);
836 benefitFromCodePlacementOpt = true;
838 // Prefer likely predicted branches to selects on out-of-order cores.
839 predictableSelectIsExpensive = Subtarget->isLikeA9();
841 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
844 // FIXME: It might make sense to define the representative register class as the
845 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
846 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
847 // SPR's representative would be DPR_VFP2. This should work well if register
848 // pressure tracking were modified such that a register use would increment the
849 // pressure of the register class's representative and all of it's super
850 // classes' representatives transitively. We have not implemented this because
851 // of the difficulty prior to coalescing of modeling operand register classes
852 // due to the common occurrence of cross class copies and subregister insertions
854 std::pair<const TargetRegisterClass*, uint8_t>
855 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
856 const TargetRegisterClass *RRC = 0;
858 switch (VT.getSimpleVT().SimpleTy) {
860 return TargetLowering::findRepresentativeClass(VT);
861 // Use DPR as representative register class for all floating point
862 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
863 // the cost is 1 for both f32 and f64.
864 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
865 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
866 RRC = &ARM::DPRRegClass;
867 // When NEON is used for SP, only half of the register file is available
868 // because operations that define both SP and DP results will be constrained
869 // to the VFP2 class (D0-D15). We currently model this constraint prior to
870 // coalescing by double-counting the SP regs. See the FIXME above.
871 if (Subtarget->useNEONForSinglePrecisionFP())
874 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
875 case MVT::v4f32: case MVT::v2f64:
876 RRC = &ARM::DPRRegClass;
880 RRC = &ARM::DPRRegClass;
884 RRC = &ARM::DPRRegClass;
888 return std::make_pair(RRC, Cost);
891 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
894 case ARMISD::Wrapper: return "ARMISD::Wrapper";
895 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
896 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
897 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
898 case ARMISD::CALL: return "ARMISD::CALL";
899 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
900 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
901 case ARMISD::tCALL: return "ARMISD::tCALL";
902 case ARMISD::BRCOND: return "ARMISD::BRCOND";
903 case ARMISD::BR_JT: return "ARMISD::BR_JT";
904 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
905 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
906 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
907 case ARMISD::CMP: return "ARMISD::CMP";
908 case ARMISD::CMN: return "ARMISD::CMN";
909 case ARMISD::CMPZ: return "ARMISD::CMPZ";
910 case ARMISD::CMPFP: return "ARMISD::CMPFP";
911 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
912 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
913 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
915 case ARMISD::CMOV: return "ARMISD::CMOV";
917 case ARMISD::RBIT: return "ARMISD::RBIT";
919 case ARMISD::FTOSI: return "ARMISD::FTOSI";
920 case ARMISD::FTOUI: return "ARMISD::FTOUI";
921 case ARMISD::SITOF: return "ARMISD::SITOF";
922 case ARMISD::UITOF: return "ARMISD::UITOF";
924 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
925 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
926 case ARMISD::RRX: return "ARMISD::RRX";
928 case ARMISD::ADDC: return "ARMISD::ADDC";
929 case ARMISD::ADDE: return "ARMISD::ADDE";
930 case ARMISD::SUBC: return "ARMISD::SUBC";
931 case ARMISD::SUBE: return "ARMISD::SUBE";
933 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
934 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
936 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
937 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
939 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
941 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
943 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
945 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
946 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
948 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
950 case ARMISD::VCEQ: return "ARMISD::VCEQ";
951 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
952 case ARMISD::VCGE: return "ARMISD::VCGE";
953 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
954 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
955 case ARMISD::VCGEU: return "ARMISD::VCGEU";
956 case ARMISD::VCGT: return "ARMISD::VCGT";
957 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
958 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
959 case ARMISD::VCGTU: return "ARMISD::VCGTU";
960 case ARMISD::VTST: return "ARMISD::VTST";
962 case ARMISD::VSHL: return "ARMISD::VSHL";
963 case ARMISD::VSHRs: return "ARMISD::VSHRs";
964 case ARMISD::VSHRu: return "ARMISD::VSHRu";
965 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
966 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
967 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
968 case ARMISD::VSHRN: return "ARMISD::VSHRN";
969 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
970 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
971 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
972 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
973 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
974 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
975 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
976 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
977 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
978 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
979 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
980 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
981 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
982 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
983 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
984 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
985 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
986 case ARMISD::VDUP: return "ARMISD::VDUP";
987 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
988 case ARMISD::VEXT: return "ARMISD::VEXT";
989 case ARMISD::VREV64: return "ARMISD::VREV64";
990 case ARMISD::VREV32: return "ARMISD::VREV32";
991 case ARMISD::VREV16: return "ARMISD::VREV16";
992 case ARMISD::VZIP: return "ARMISD::VZIP";
993 case ARMISD::VUZP: return "ARMISD::VUZP";
994 case ARMISD::VTRN: return "ARMISD::VTRN";
995 case ARMISD::VTBL1: return "ARMISD::VTBL1";
996 case ARMISD::VTBL2: return "ARMISD::VTBL2";
997 case ARMISD::VMULLs: return "ARMISD::VMULLs";
998 case ARMISD::VMULLu: return "ARMISD::VMULLu";
999 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1000 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1001 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1002 case ARMISD::FMAX: return "ARMISD::FMAX";
1003 case ARMISD::FMIN: return "ARMISD::FMIN";
1004 case ARMISD::BFI: return "ARMISD::BFI";
1005 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1006 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1007 case ARMISD::VBSL: return "ARMISD::VBSL";
1008 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1009 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1010 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1011 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1012 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1013 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1014 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1015 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1016 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1017 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1018 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1019 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1020 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1021 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1022 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1023 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1024 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1025 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1026 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1027 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1031 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1032 if (!VT.isVector()) return getPointerTy();
1033 return VT.changeVectorElementTypeToInteger();
1036 /// getRegClassFor - Return the register class that should be used for the
1037 /// specified value type.
1038 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1039 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1040 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1041 // load / store 4 to 8 consecutive D registers.
1042 if (Subtarget->hasNEON()) {
1043 if (VT == MVT::v4i64)
1044 return &ARM::QQPRRegClass;
1045 if (VT == MVT::v8i64)
1046 return &ARM::QQQQPRRegClass;
1048 return TargetLowering::getRegClassFor(VT);
1051 // Create a fast isel object.
1053 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1054 const TargetLibraryInfo *libInfo) const {
1055 return ARM::createFastISel(funcInfo, libInfo);
1058 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1059 /// be used for loads / stores from the global.
1060 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1061 return (Subtarget->isThumb1Only() ? 127 : 4095);
1064 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1065 unsigned NumVals = N->getNumValues();
1067 return Sched::RegPressure;
1069 for (unsigned i = 0; i != NumVals; ++i) {
1070 EVT VT = N->getValueType(i);
1071 if (VT == MVT::Glue || VT == MVT::Other)
1073 if (VT.isFloatingPoint() || VT.isVector())
1077 if (!N->isMachineOpcode())
1078 return Sched::RegPressure;
1080 // Load are scheduled for latency even if there instruction itinerary
1081 // is not available.
1082 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1083 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1085 if (MCID.getNumDefs() == 0)
1086 return Sched::RegPressure;
1087 if (!Itins->isEmpty() &&
1088 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1091 return Sched::RegPressure;
1094 //===----------------------------------------------------------------------===//
1096 //===----------------------------------------------------------------------===//
1098 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1099 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1101 default: llvm_unreachable("Unknown condition code!");
1102 case ISD::SETNE: return ARMCC::NE;
1103 case ISD::SETEQ: return ARMCC::EQ;
1104 case ISD::SETGT: return ARMCC::GT;
1105 case ISD::SETGE: return ARMCC::GE;
1106 case ISD::SETLT: return ARMCC::LT;
1107 case ISD::SETLE: return ARMCC::LE;
1108 case ISD::SETUGT: return ARMCC::HI;
1109 case ISD::SETUGE: return ARMCC::HS;
1110 case ISD::SETULT: return ARMCC::LO;
1111 case ISD::SETULE: return ARMCC::LS;
1115 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1116 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1117 ARMCC::CondCodes &CondCode2) {
1118 CondCode2 = ARMCC::AL;
1120 default: llvm_unreachable("Unknown FP condition!");
1122 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1124 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1126 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1127 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1128 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1129 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1130 case ISD::SETO: CondCode = ARMCC::VC; break;
1131 case ISD::SETUO: CondCode = ARMCC::VS; break;
1132 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1133 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1134 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1136 case ISD::SETULT: CondCode = ARMCC::LT; break;
1138 case ISD::SETULE: CondCode = ARMCC::LE; break;
1140 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1144 //===----------------------------------------------------------------------===//
1145 // Calling Convention Implementation
1146 //===----------------------------------------------------------------------===//
1148 #include "ARMGenCallingConv.inc"
1150 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1151 /// given CallingConvention value.
1152 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1154 bool isVarArg) const {
1157 llvm_unreachable("Unsupported calling convention");
1158 case CallingConv::Fast:
1159 if (Subtarget->hasVFP2() && !isVarArg) {
1160 if (!Subtarget->isAAPCS_ABI())
1161 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1162 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1163 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1166 case CallingConv::C: {
1167 // Use target triple & subtarget features to do actual dispatch.
1168 if (!Subtarget->isAAPCS_ABI())
1169 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1170 else if (Subtarget->hasVFP2() &&
1171 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1173 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1174 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1176 case CallingConv::ARM_AAPCS_VFP:
1178 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1180 case CallingConv::ARM_AAPCS:
1181 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1182 case CallingConv::ARM_APCS:
1183 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1184 case CallingConv::GHC:
1185 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1189 /// LowerCallResult - Lower the result values of a call into the
1190 /// appropriate copies out of appropriate physical registers.
1192 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1193 CallingConv::ID CallConv, bool isVarArg,
1194 const SmallVectorImpl<ISD::InputArg> &Ins,
1195 DebugLoc dl, SelectionDAG &DAG,
1196 SmallVectorImpl<SDValue> &InVals) const {
1198 // Assign locations to each value returned by this call.
1199 SmallVector<CCValAssign, 16> RVLocs;
1200 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1201 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1202 CCInfo.AnalyzeCallResult(Ins,
1203 CCAssignFnForNode(CallConv, /* Return*/ true,
1206 // Copy all of the result registers out of their specified physreg.
1207 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1208 CCValAssign VA = RVLocs[i];
1211 if (VA.needsCustom()) {
1212 // Handle f64 or half of a v2f64.
1213 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1215 Chain = Lo.getValue(1);
1216 InFlag = Lo.getValue(2);
1217 VA = RVLocs[++i]; // skip ahead to next loc
1218 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1220 Chain = Hi.getValue(1);
1221 InFlag = Hi.getValue(2);
1222 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1224 if (VA.getLocVT() == MVT::v2f64) {
1225 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1226 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1227 DAG.getConstant(0, MVT::i32));
1229 VA = RVLocs[++i]; // skip ahead to next loc
1230 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1231 Chain = Lo.getValue(1);
1232 InFlag = Lo.getValue(2);
1233 VA = RVLocs[++i]; // skip ahead to next loc
1234 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1235 Chain = Hi.getValue(1);
1236 InFlag = Hi.getValue(2);
1237 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1238 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1239 DAG.getConstant(1, MVT::i32));
1242 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1244 Chain = Val.getValue(1);
1245 InFlag = Val.getValue(2);
1248 switch (VA.getLocInfo()) {
1249 default: llvm_unreachable("Unknown loc info!");
1250 case CCValAssign::Full: break;
1251 case CCValAssign::BCvt:
1252 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1256 InVals.push_back(Val);
1262 /// LowerMemOpCallTo - Store the argument to the stack.
1264 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1265 SDValue StackPtr, SDValue Arg,
1266 DebugLoc dl, SelectionDAG &DAG,
1267 const CCValAssign &VA,
1268 ISD::ArgFlagsTy Flags) const {
1269 unsigned LocMemOffset = VA.getLocMemOffset();
1270 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1271 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1272 return DAG.getStore(Chain, dl, Arg, PtrOff,
1273 MachinePointerInfo::getStack(LocMemOffset),
1277 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1278 SDValue Chain, SDValue &Arg,
1279 RegsToPassVector &RegsToPass,
1280 CCValAssign &VA, CCValAssign &NextVA,
1282 SmallVector<SDValue, 8> &MemOpChains,
1283 ISD::ArgFlagsTy Flags) const {
1285 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1286 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1287 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1289 if (NextVA.isRegLoc())
1290 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1292 assert(NextVA.isMemLoc());
1293 if (StackPtr.getNode() == 0)
1294 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1296 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1302 /// LowerCall - Lowering a call into a callseq_start <-
1303 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1306 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1307 SmallVectorImpl<SDValue> &InVals) const {
1308 SelectionDAG &DAG = CLI.DAG;
1309 DebugLoc &dl = CLI.DL;
1310 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1311 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1312 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1313 SDValue Chain = CLI.Chain;
1314 SDValue Callee = CLI.Callee;
1315 bool &isTailCall = CLI.IsTailCall;
1316 CallingConv::ID CallConv = CLI.CallConv;
1317 bool doesNotRet = CLI.DoesNotReturn;
1318 bool isVarArg = CLI.IsVarArg;
1320 MachineFunction &MF = DAG.getMachineFunction();
1321 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1322 bool IsSibCall = false;
1323 // Disable tail calls if they're not supported.
1324 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1327 // Check if it's really possible to do a tail call.
1328 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1329 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1330 Outs, OutVals, Ins, DAG);
1331 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1332 // detected sibcalls.
1339 // Analyze operands of the call, assigning locations to each operand.
1340 SmallVector<CCValAssign, 16> ArgLocs;
1341 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1342 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1343 CCInfo.AnalyzeCallOperands(Outs,
1344 CCAssignFnForNode(CallConv, /* Return*/ false,
1347 // Get a count of how many bytes are to be pushed on the stack.
1348 unsigned NumBytes = CCInfo.getNextStackOffset();
1350 // For tail calls, memory operands are available in our caller's stack.
1354 // Adjust the stack pointer for the new arguments...
1355 // These operations are automatically eliminated by the prolog/epilog pass
1357 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1359 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1361 RegsToPassVector RegsToPass;
1362 SmallVector<SDValue, 8> MemOpChains;
1364 // Walk the register/memloc assignments, inserting copies/loads. In the case
1365 // of tail call optimization, arguments are handled later.
1366 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1368 ++i, ++realArgIdx) {
1369 CCValAssign &VA = ArgLocs[i];
1370 SDValue Arg = OutVals[realArgIdx];
1371 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1372 bool isByVal = Flags.isByVal();
1374 // Promote the value if needed.
1375 switch (VA.getLocInfo()) {
1376 default: llvm_unreachable("Unknown loc info!");
1377 case CCValAssign::Full: break;
1378 case CCValAssign::SExt:
1379 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1381 case CCValAssign::ZExt:
1382 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1384 case CCValAssign::AExt:
1385 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1387 case CCValAssign::BCvt:
1388 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1392 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1393 if (VA.needsCustom()) {
1394 if (VA.getLocVT() == MVT::v2f64) {
1395 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1396 DAG.getConstant(0, MVT::i32));
1397 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1398 DAG.getConstant(1, MVT::i32));
1400 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1401 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1403 VA = ArgLocs[++i]; // skip ahead to next loc
1404 if (VA.isRegLoc()) {
1405 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1406 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1408 assert(VA.isMemLoc());
1410 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1411 dl, DAG, VA, Flags));
1414 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1415 StackPtr, MemOpChains, Flags);
1417 } else if (VA.isRegLoc()) {
1418 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1419 } else if (isByVal) {
1420 assert(VA.isMemLoc());
1421 unsigned offset = 0;
1423 // True if this byval aggregate will be split between registers
1425 if (CCInfo.isFirstByValRegValid()) {
1426 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1428 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1429 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1430 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1431 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1432 MachinePointerInfo(),
1433 false, false, false, 0);
1434 MemOpChains.push_back(Load.getValue(1));
1435 RegsToPass.push_back(std::make_pair(j, Load));
1437 offset = ARM::R4 - CCInfo.getFirstByValReg();
1438 CCInfo.clearFirstByValReg();
1441 if (Flags.getByValSize() - 4*offset > 0) {
1442 unsigned LocMemOffset = VA.getLocMemOffset();
1443 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1444 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1446 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1447 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1448 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1450 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1452 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1453 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1454 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1455 Ops, array_lengthof(Ops)));
1457 } else if (!IsSibCall) {
1458 assert(VA.isMemLoc());
1460 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1461 dl, DAG, VA, Flags));
1465 if (!MemOpChains.empty())
1466 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1467 &MemOpChains[0], MemOpChains.size());
1469 // Build a sequence of copy-to-reg nodes chained together with token chain
1470 // and flag operands which copy the outgoing args into the appropriate regs.
1472 // Tail call byval lowering might overwrite argument registers so in case of
1473 // tail call optimization the copies to registers are lowered later.
1475 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1476 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1477 RegsToPass[i].second, InFlag);
1478 InFlag = Chain.getValue(1);
1481 // For tail calls lower the arguments to the 'real' stack slot.
1483 // Force all the incoming stack arguments to be loaded from the stack
1484 // before any new outgoing arguments are stored to the stack, because the
1485 // outgoing stack slots may alias the incoming argument stack slots, and
1486 // the alias isn't otherwise explicit. This is slightly more conservative
1487 // than necessary, because it means that each store effectively depends
1488 // on every argument instead of just those arguments it would clobber.
1490 // Do not flag preceding copytoreg stuff together with the following stuff.
1492 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1493 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1494 RegsToPass[i].second, InFlag);
1495 InFlag = Chain.getValue(1);
1500 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1501 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1502 // node so that legalize doesn't hack it.
1503 bool isDirect = false;
1504 bool isARMFunc = false;
1505 bool isLocalARMFunc = false;
1506 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1508 if (EnableARMLongCalls) {
1509 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1510 && "long-calls with non-static relocation model!");
1511 // Handle a global address or an external symbol. If it's not one of
1512 // those, the target's already in a register, so we don't need to do
1514 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1515 const GlobalValue *GV = G->getGlobal();
1516 // Create a constant pool entry for the callee address
1517 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1518 ARMConstantPoolValue *CPV =
1519 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1521 // Get the address of the callee into a register
1522 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1523 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1524 Callee = DAG.getLoad(getPointerTy(), dl,
1525 DAG.getEntryNode(), CPAddr,
1526 MachinePointerInfo::getConstantPool(),
1527 false, false, false, 0);
1528 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1529 const char *Sym = S->getSymbol();
1531 // Create a constant pool entry for the callee address
1532 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1533 ARMConstantPoolValue *CPV =
1534 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1535 ARMPCLabelIndex, 0);
1536 // Get the address of the callee into a register
1537 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1538 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1539 Callee = DAG.getLoad(getPointerTy(), dl,
1540 DAG.getEntryNode(), CPAddr,
1541 MachinePointerInfo::getConstantPool(),
1542 false, false, false, 0);
1544 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1545 const GlobalValue *GV = G->getGlobal();
1547 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1548 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1549 getTargetMachine().getRelocationModel() != Reloc::Static;
1550 isARMFunc = !Subtarget->isThumb() || isStub;
1551 // ARM call to a local ARM function is predicable.
1552 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1553 // tBX takes a register source operand.
1554 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1555 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1556 ARMConstantPoolValue *CPV =
1557 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1558 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1559 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1560 Callee = DAG.getLoad(getPointerTy(), dl,
1561 DAG.getEntryNode(), CPAddr,
1562 MachinePointerInfo::getConstantPool(),
1563 false, false, false, 0);
1564 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1565 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1566 getPointerTy(), Callee, PICLabel);
1568 // On ELF targets for PIC code, direct calls should go through the PLT
1569 unsigned OpFlags = 0;
1570 if (Subtarget->isTargetELF() &&
1571 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1572 OpFlags = ARMII::MO_PLT;
1573 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1575 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1577 bool isStub = Subtarget->isTargetDarwin() &&
1578 getTargetMachine().getRelocationModel() != Reloc::Static;
1579 isARMFunc = !Subtarget->isThumb() || isStub;
1580 // tBX takes a register source operand.
1581 const char *Sym = S->getSymbol();
1582 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1583 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1584 ARMConstantPoolValue *CPV =
1585 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1586 ARMPCLabelIndex, 4);
1587 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1588 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1589 Callee = DAG.getLoad(getPointerTy(), dl,
1590 DAG.getEntryNode(), CPAddr,
1591 MachinePointerInfo::getConstantPool(),
1592 false, false, false, 0);
1593 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1594 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1595 getPointerTy(), Callee, PICLabel);
1597 unsigned OpFlags = 0;
1598 // On ELF targets for PIC code, direct calls should go through the PLT
1599 if (Subtarget->isTargetELF() &&
1600 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1601 OpFlags = ARMII::MO_PLT;
1602 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1606 // FIXME: handle tail calls differently.
1608 bool HasMinSizeAttr = MF.getFunction()->getFnAttributes().
1609 hasAttribute(Attributes::MinSize);
1610 if (Subtarget->isThumb()) {
1611 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1612 CallOpc = ARMISD::CALL_NOLINK;
1614 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1616 if (!isDirect && !Subtarget->hasV5TOps())
1617 CallOpc = ARMISD::CALL_NOLINK;
1618 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1619 // Emit regular call when code size is the priority
1621 // "mov lr, pc; b _foo" to avoid confusing the RSP
1622 CallOpc = ARMISD::CALL_NOLINK;
1624 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1627 std::vector<SDValue> Ops;
1628 Ops.push_back(Chain);
1629 Ops.push_back(Callee);
1631 // Add argument registers to the end of the list so that they are known live
1633 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1634 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1635 RegsToPass[i].second.getValueType()));
1637 // Add a register mask operand representing the call-preserved registers.
1638 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1639 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1640 assert(Mask && "Missing call preserved mask for calling convention");
1641 Ops.push_back(DAG.getRegisterMask(Mask));
1643 if (InFlag.getNode())
1644 Ops.push_back(InFlag);
1646 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1648 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1650 // Returns a chain and a flag for retval copy to use.
1651 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1652 InFlag = Chain.getValue(1);
1654 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1655 DAG.getIntPtrConstant(0, true), InFlag);
1657 InFlag = Chain.getValue(1);
1659 // Handle result values, copying them out of physregs into vregs that we
1661 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1665 /// HandleByVal - Every parameter *after* a byval parameter is passed
1666 /// on the stack. Remember the next parameter register to allocate,
1667 /// and then confiscate the rest of the parameter registers to insure
1670 ARMTargetLowering::HandleByVal(
1671 CCState *State, unsigned &size, unsigned Align) const {
1672 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1673 assert((State->getCallOrPrologue() == Prologue ||
1674 State->getCallOrPrologue() == Call) &&
1675 "unhandled ParmContext");
1676 if ((!State->isFirstByValRegValid()) &&
1677 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1678 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1679 unsigned AlignInRegs = Align / 4;
1680 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1681 for (unsigned i = 0; i < Waste; ++i)
1682 reg = State->AllocateReg(GPRArgRegs, 4);
1685 State->setFirstByValReg(reg);
1686 // At a call site, a byval parameter that is split between
1687 // registers and memory needs its size truncated here. In a
1688 // function prologue, such byval parameters are reassembled in
1689 // memory, and are not truncated.
1690 if (State->getCallOrPrologue() == Call) {
1691 unsigned excess = 4 * (ARM::R4 - reg);
1692 assert(size >= excess && "expected larger existing stack allocation");
1697 // Confiscate any remaining parameter registers to preclude their
1698 // assignment to subsequent parameters.
1699 while (State->AllocateReg(GPRArgRegs, 4))
1703 /// MatchingStackOffset - Return true if the given stack call argument is
1704 /// already available in the same position (relatively) of the caller's
1705 /// incoming argument stack.
1707 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1708 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1709 const TargetInstrInfo *TII) {
1710 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1712 if (Arg.getOpcode() == ISD::CopyFromReg) {
1713 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1714 if (!TargetRegisterInfo::isVirtualRegister(VR))
1716 MachineInstr *Def = MRI->getVRegDef(VR);
1719 if (!Flags.isByVal()) {
1720 if (!TII->isLoadFromStackSlot(Def, FI))
1725 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1726 if (Flags.isByVal())
1727 // ByVal argument is passed in as a pointer but it's now being
1728 // dereferenced. e.g.
1729 // define @foo(%struct.X* %A) {
1730 // tail call @bar(%struct.X* byval %A)
1733 SDValue Ptr = Ld->getBasePtr();
1734 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1737 FI = FINode->getIndex();
1741 assert(FI != INT_MAX);
1742 if (!MFI->isFixedObjectIndex(FI))
1744 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1747 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1748 /// for tail call optimization. Targets which want to do tail call
1749 /// optimization should implement this function.
1751 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1752 CallingConv::ID CalleeCC,
1754 bool isCalleeStructRet,
1755 bool isCallerStructRet,
1756 const SmallVectorImpl<ISD::OutputArg> &Outs,
1757 const SmallVectorImpl<SDValue> &OutVals,
1758 const SmallVectorImpl<ISD::InputArg> &Ins,
1759 SelectionDAG& DAG) const {
1760 const Function *CallerF = DAG.getMachineFunction().getFunction();
1761 CallingConv::ID CallerCC = CallerF->getCallingConv();
1762 bool CCMatch = CallerCC == CalleeCC;
1764 // Look for obvious safe cases to perform tail call optimization that do not
1765 // require ABI changes. This is what gcc calls sibcall.
1767 // Do not sibcall optimize vararg calls unless the call site is not passing
1769 if (isVarArg && !Outs.empty())
1772 // Also avoid sibcall optimization if either caller or callee uses struct
1773 // return semantics.
1774 if (isCalleeStructRet || isCallerStructRet)
1777 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1778 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1779 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1780 // support in the assembler and linker to be used. This would need to be
1781 // fixed to fully support tail calls in Thumb1.
1783 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1784 // LR. This means if we need to reload LR, it takes an extra instructions,
1785 // which outweighs the value of the tail call; but here we don't know yet
1786 // whether LR is going to be used. Probably the right approach is to
1787 // generate the tail call here and turn it back into CALL/RET in
1788 // emitEpilogue if LR is used.
1790 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1791 // but we need to make sure there are enough registers; the only valid
1792 // registers are the 4 used for parameters. We don't currently do this
1794 if (Subtarget->isThumb1Only())
1797 // If the calling conventions do not match, then we'd better make sure the
1798 // results are returned in the same way as what the caller expects.
1800 SmallVector<CCValAssign, 16> RVLocs1;
1801 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1802 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1803 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1805 SmallVector<CCValAssign, 16> RVLocs2;
1806 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1807 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1808 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1810 if (RVLocs1.size() != RVLocs2.size())
1812 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1813 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1815 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1817 if (RVLocs1[i].isRegLoc()) {
1818 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1821 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1827 // If Caller's vararg or byval argument has been split between registers and
1828 // stack, do not perform tail call, since part of the argument is in caller's
1830 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1831 getInfo<ARMFunctionInfo>();
1832 if (AFI_Caller->getVarArgsRegSaveSize())
1835 // If the callee takes no arguments then go on to check the results of the
1837 if (!Outs.empty()) {
1838 // Check if stack adjustment is needed. For now, do not do this if any
1839 // argument is passed on the stack.
1840 SmallVector<CCValAssign, 16> ArgLocs;
1841 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1842 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1843 CCInfo.AnalyzeCallOperands(Outs,
1844 CCAssignFnForNode(CalleeCC, false, isVarArg));
1845 if (CCInfo.getNextStackOffset()) {
1846 MachineFunction &MF = DAG.getMachineFunction();
1848 // Check if the arguments are already laid out in the right way as
1849 // the caller's fixed stack objects.
1850 MachineFrameInfo *MFI = MF.getFrameInfo();
1851 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1852 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1853 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1855 ++i, ++realArgIdx) {
1856 CCValAssign &VA = ArgLocs[i];
1857 EVT RegVT = VA.getLocVT();
1858 SDValue Arg = OutVals[realArgIdx];
1859 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1860 if (VA.getLocInfo() == CCValAssign::Indirect)
1862 if (VA.needsCustom()) {
1863 // f64 and vector types are split into multiple registers or
1864 // register/stack-slot combinations. The types will not match
1865 // the registers; give up on memory f64 refs until we figure
1866 // out what to do about this.
1869 if (!ArgLocs[++i].isRegLoc())
1871 if (RegVT == MVT::v2f64) {
1872 if (!ArgLocs[++i].isRegLoc())
1874 if (!ArgLocs[++i].isRegLoc())
1877 } else if (!VA.isRegLoc()) {
1878 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1890 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1891 MachineFunction &MF, bool isVarArg,
1892 const SmallVectorImpl<ISD::OutputArg> &Outs,
1893 LLVMContext &Context) const {
1894 SmallVector<CCValAssign, 16> RVLocs;
1895 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
1896 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
1901 ARMTargetLowering::LowerReturn(SDValue Chain,
1902 CallingConv::ID CallConv, bool isVarArg,
1903 const SmallVectorImpl<ISD::OutputArg> &Outs,
1904 const SmallVectorImpl<SDValue> &OutVals,
1905 DebugLoc dl, SelectionDAG &DAG) const {
1907 // CCValAssign - represent the assignment of the return value to a location.
1908 SmallVector<CCValAssign, 16> RVLocs;
1910 // CCState - Info about the registers and stack slots.
1911 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1912 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1914 // Analyze outgoing return values.
1915 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1918 // If this is the first return lowered for this function, add
1919 // the regs to the liveout set for the function.
1920 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1921 for (unsigned i = 0; i != RVLocs.size(); ++i)
1922 if (RVLocs[i].isRegLoc())
1923 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1928 // Copy the result values into the output registers.
1929 for (unsigned i = 0, realRVLocIdx = 0;
1931 ++i, ++realRVLocIdx) {
1932 CCValAssign &VA = RVLocs[i];
1933 assert(VA.isRegLoc() && "Can only return in registers!");
1935 SDValue Arg = OutVals[realRVLocIdx];
1937 switch (VA.getLocInfo()) {
1938 default: llvm_unreachable("Unknown loc info!");
1939 case CCValAssign::Full: break;
1940 case CCValAssign::BCvt:
1941 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1945 if (VA.needsCustom()) {
1946 if (VA.getLocVT() == MVT::v2f64) {
1947 // Extract the first half and return it in two registers.
1948 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1949 DAG.getConstant(0, MVT::i32));
1950 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1951 DAG.getVTList(MVT::i32, MVT::i32), Half);
1953 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1954 Flag = Chain.getValue(1);
1955 VA = RVLocs[++i]; // skip ahead to next loc
1956 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1957 HalfGPRs.getValue(1), Flag);
1958 Flag = Chain.getValue(1);
1959 VA = RVLocs[++i]; // skip ahead to next loc
1961 // Extract the 2nd half and fall through to handle it as an f64 value.
1962 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1963 DAG.getConstant(1, MVT::i32));
1965 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1967 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1968 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1969 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1970 Flag = Chain.getValue(1);
1971 VA = RVLocs[++i]; // skip ahead to next loc
1972 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1975 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1977 // Guarantee that all emitted copies are
1978 // stuck together, avoiding something bad.
1979 Flag = Chain.getValue(1);
1984 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1986 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1991 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1992 if (N->getNumValues() != 1)
1994 if (!N->hasNUsesOfValue(1, 0))
1997 SDValue TCChain = Chain;
1998 SDNode *Copy = *N->use_begin();
1999 if (Copy->getOpcode() == ISD::CopyToReg) {
2000 // If the copy has a glue operand, we conservatively assume it isn't safe to
2001 // perform a tail call.
2002 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2004 TCChain = Copy->getOperand(0);
2005 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2006 SDNode *VMov = Copy;
2007 // f64 returned in a pair of GPRs.
2008 SmallPtrSet<SDNode*, 2> Copies;
2009 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2011 if (UI->getOpcode() != ISD::CopyToReg)
2015 if (Copies.size() > 2)
2018 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2020 SDValue UseChain = UI->getOperand(0);
2021 if (Copies.count(UseChain.getNode()))
2028 } else if (Copy->getOpcode() == ISD::BITCAST) {
2029 // f32 returned in a single GPR.
2030 if (!Copy->hasOneUse())
2032 Copy = *Copy->use_begin();
2033 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2035 Chain = Copy->getOperand(0);
2040 bool HasRet = false;
2041 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2043 if (UI->getOpcode() != ARMISD::RET_FLAG)
2055 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2056 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2059 if (!CI->isTailCall())
2062 return !Subtarget->isThumb1Only();
2065 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2066 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2067 // one of the above mentioned nodes. It has to be wrapped because otherwise
2068 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2069 // be used to form addressing mode. These wrapped nodes will be selected
2071 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2072 EVT PtrVT = Op.getValueType();
2073 // FIXME there is no actual debug info here
2074 DebugLoc dl = Op.getDebugLoc();
2075 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2077 if (CP->isMachineConstantPoolEntry())
2078 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2079 CP->getAlignment());
2081 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2082 CP->getAlignment());
2083 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2086 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2087 return MachineJumpTableInfo::EK_Inline;
2090 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2091 SelectionDAG &DAG) const {
2092 MachineFunction &MF = DAG.getMachineFunction();
2093 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2094 unsigned ARMPCLabelIndex = 0;
2095 DebugLoc DL = Op.getDebugLoc();
2096 EVT PtrVT = getPointerTy();
2097 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2098 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2100 if (RelocM == Reloc::Static) {
2101 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2103 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2104 ARMPCLabelIndex = AFI->createPICLabelUId();
2105 ARMConstantPoolValue *CPV =
2106 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2107 ARMCP::CPBlockAddress, PCAdj);
2108 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2110 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2111 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2112 MachinePointerInfo::getConstantPool(),
2113 false, false, false, 0);
2114 if (RelocM == Reloc::Static)
2116 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2117 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2120 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2122 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2123 SelectionDAG &DAG) const {
2124 DebugLoc dl = GA->getDebugLoc();
2125 EVT PtrVT = getPointerTy();
2126 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2127 MachineFunction &MF = DAG.getMachineFunction();
2128 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2129 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2130 ARMConstantPoolValue *CPV =
2131 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2132 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2133 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2134 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2135 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2136 MachinePointerInfo::getConstantPool(),
2137 false, false, false, 0);
2138 SDValue Chain = Argument.getValue(1);
2140 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2141 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2143 // call __tls_get_addr.
2146 Entry.Node = Argument;
2147 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2148 Args.push_back(Entry);
2149 // FIXME: is there useful debug info available here?
2150 TargetLowering::CallLoweringInfo CLI(Chain,
2151 (Type *) Type::getInt32Ty(*DAG.getContext()),
2152 false, false, false, false,
2153 0, CallingConv::C, /*isTailCall=*/false,
2154 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2155 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2156 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2157 return CallResult.first;
2160 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2161 // "local exec" model.
2163 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2165 TLSModel::Model model) const {
2166 const GlobalValue *GV = GA->getGlobal();
2167 DebugLoc dl = GA->getDebugLoc();
2169 SDValue Chain = DAG.getEntryNode();
2170 EVT PtrVT = getPointerTy();
2171 // Get the Thread Pointer
2172 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2174 if (model == TLSModel::InitialExec) {
2175 MachineFunction &MF = DAG.getMachineFunction();
2176 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2177 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2178 // Initial exec model.
2179 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2180 ARMConstantPoolValue *CPV =
2181 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2182 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2184 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2185 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2186 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2187 MachinePointerInfo::getConstantPool(),
2188 false, false, false, 0);
2189 Chain = Offset.getValue(1);
2191 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2192 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2194 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2195 MachinePointerInfo::getConstantPool(),
2196 false, false, false, 0);
2199 assert(model == TLSModel::LocalExec);
2200 ARMConstantPoolValue *CPV =
2201 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2202 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2203 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2204 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2205 MachinePointerInfo::getConstantPool(),
2206 false, false, false, 0);
2209 // The address of the thread local variable is the add of the thread
2210 // pointer with the offset of the variable.
2211 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2215 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2216 // TODO: implement the "local dynamic" model
2217 assert(Subtarget->isTargetELF() &&
2218 "TLS not implemented for non-ELF targets");
2219 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2221 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2224 case TLSModel::GeneralDynamic:
2225 case TLSModel::LocalDynamic:
2226 return LowerToTLSGeneralDynamicModel(GA, DAG);
2227 case TLSModel::InitialExec:
2228 case TLSModel::LocalExec:
2229 return LowerToTLSExecModels(GA, DAG, model);
2231 llvm_unreachable("bogus TLS model");
2234 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2235 SelectionDAG &DAG) const {
2236 EVT PtrVT = getPointerTy();
2237 DebugLoc dl = Op.getDebugLoc();
2238 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2239 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2240 if (RelocM == Reloc::PIC_) {
2241 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2242 ARMConstantPoolValue *CPV =
2243 ARMConstantPoolConstant::Create(GV,
2244 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2245 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2246 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2247 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2249 MachinePointerInfo::getConstantPool(),
2250 false, false, false, 0);
2251 SDValue Chain = Result.getValue(1);
2252 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2253 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2255 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2256 MachinePointerInfo::getGOT(),
2257 false, false, false, 0);
2261 // If we have T2 ops, we can materialize the address directly via movt/movw
2262 // pair. This is always cheaper.
2263 if (Subtarget->useMovt()) {
2265 // FIXME: Once remat is capable of dealing with instructions with register
2266 // operands, expand this into two nodes.
2267 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2268 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2270 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2271 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2272 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2273 MachinePointerInfo::getConstantPool(),
2274 false, false, false, 0);
2278 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2279 SelectionDAG &DAG) const {
2280 EVT PtrVT = getPointerTy();
2281 DebugLoc dl = Op.getDebugLoc();
2282 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2283 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2284 MachineFunction &MF = DAG.getMachineFunction();
2285 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2287 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2288 // update ARMFastISel::ARMMaterializeGV.
2289 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2291 // FIXME: Once remat is capable of dealing with instructions with register
2292 // operands, expand this into two nodes.
2293 if (RelocM == Reloc::Static)
2294 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2295 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2297 unsigned Wrapper = (RelocM == Reloc::PIC_)
2298 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2299 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2300 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2301 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2302 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2303 MachinePointerInfo::getGOT(),
2304 false, false, false, 0);
2308 unsigned ARMPCLabelIndex = 0;
2310 if (RelocM == Reloc::Static) {
2311 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2313 ARMPCLabelIndex = AFI->createPICLabelUId();
2314 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2315 ARMConstantPoolValue *CPV =
2316 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2318 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2320 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2322 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2323 MachinePointerInfo::getConstantPool(),
2324 false, false, false, 0);
2325 SDValue Chain = Result.getValue(1);
2327 if (RelocM == Reloc::PIC_) {
2328 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2329 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2332 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2333 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2334 false, false, false, 0);
2339 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2340 SelectionDAG &DAG) const {
2341 assert(Subtarget->isTargetELF() &&
2342 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2343 MachineFunction &MF = DAG.getMachineFunction();
2344 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2345 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2346 EVT PtrVT = getPointerTy();
2347 DebugLoc dl = Op.getDebugLoc();
2348 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2349 ARMConstantPoolValue *CPV =
2350 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2351 ARMPCLabelIndex, PCAdj);
2352 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2353 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2354 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2355 MachinePointerInfo::getConstantPool(),
2356 false, false, false, 0);
2357 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2358 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2362 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2363 DebugLoc dl = Op.getDebugLoc();
2364 SDValue Val = DAG.getConstant(0, MVT::i32);
2365 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2366 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2367 Op.getOperand(1), Val);
2371 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2372 DebugLoc dl = Op.getDebugLoc();
2373 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2374 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2378 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2379 const ARMSubtarget *Subtarget) const {
2380 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2381 DebugLoc dl = Op.getDebugLoc();
2383 default: return SDValue(); // Don't custom lower most intrinsics.
2384 case Intrinsic::arm_thread_pointer: {
2385 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2386 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2388 case Intrinsic::eh_sjlj_lsda: {
2389 MachineFunction &MF = DAG.getMachineFunction();
2390 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2391 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2392 EVT PtrVT = getPointerTy();
2393 DebugLoc dl = Op.getDebugLoc();
2394 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2396 unsigned PCAdj = (RelocM != Reloc::PIC_)
2397 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2398 ARMConstantPoolValue *CPV =
2399 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2400 ARMCP::CPLSDA, PCAdj);
2401 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2402 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2404 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2405 MachinePointerInfo::getConstantPool(),
2406 false, false, false, 0);
2408 if (RelocM == Reloc::PIC_) {
2409 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2410 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2414 case Intrinsic::arm_neon_vmulls:
2415 case Intrinsic::arm_neon_vmullu: {
2416 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2417 ? ARMISD::VMULLs : ARMISD::VMULLu;
2418 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2419 Op.getOperand(1), Op.getOperand(2));
2424 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2425 const ARMSubtarget *Subtarget) {
2426 DebugLoc dl = Op.getDebugLoc();
2427 if (!Subtarget->hasDataBarrier()) {
2428 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2429 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2431 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2432 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2433 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2434 DAG.getConstant(0, MVT::i32));
2437 SDValue Op5 = Op.getOperand(5);
2438 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2439 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2440 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2441 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2443 ARM_MB::MemBOpt DMBOpt;
2444 if (isDeviceBarrier)
2445 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2447 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2448 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2449 DAG.getConstant(DMBOpt, MVT::i32));
2453 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2454 const ARMSubtarget *Subtarget) {
2455 // FIXME: handle "fence singlethread" more efficiently.
2456 DebugLoc dl = Op.getDebugLoc();
2457 if (!Subtarget->hasDataBarrier()) {
2458 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2459 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2461 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2462 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2463 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2464 DAG.getConstant(0, MVT::i32));
2467 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2468 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2471 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2472 const ARMSubtarget *Subtarget) {
2473 // ARM pre v5TE and Thumb1 does not have preload instructions.
2474 if (!(Subtarget->isThumb2() ||
2475 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2476 // Just preserve the chain.
2477 return Op.getOperand(0);
2479 DebugLoc dl = Op.getDebugLoc();
2480 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2482 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2483 // ARMv7 with MP extension has PLDW.
2484 return Op.getOperand(0);
2486 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2487 if (Subtarget->isThumb()) {
2489 isRead = ~isRead & 1;
2490 isData = ~isData & 1;
2493 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2494 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2495 DAG.getConstant(isData, MVT::i32));
2498 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2499 MachineFunction &MF = DAG.getMachineFunction();
2500 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2502 // vastart just stores the address of the VarArgsFrameIndex slot into the
2503 // memory location argument.
2504 DebugLoc dl = Op.getDebugLoc();
2505 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2506 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2507 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2508 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2509 MachinePointerInfo(SV), false, false, 0);
2513 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2514 SDValue &Root, SelectionDAG &DAG,
2515 DebugLoc dl) const {
2516 MachineFunction &MF = DAG.getMachineFunction();
2517 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2519 const TargetRegisterClass *RC;
2520 if (AFI->isThumb1OnlyFunction())
2521 RC = &ARM::tGPRRegClass;
2523 RC = &ARM::GPRRegClass;
2525 // Transform the arguments stored in physical registers into virtual ones.
2526 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2527 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2530 if (NextVA.isMemLoc()) {
2531 MachineFrameInfo *MFI = MF.getFrameInfo();
2532 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2534 // Create load node to retrieve arguments from the stack.
2535 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2536 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2537 MachinePointerInfo::getFixedStack(FI),
2538 false, false, false, 0);
2540 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2541 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2544 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2548 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2549 unsigned &VARegSize, unsigned &VARegSaveSize)
2552 if (CCInfo.isFirstByValRegValid())
2553 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2555 unsigned int firstUnalloced;
2556 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2557 sizeof(GPRArgRegs) /
2558 sizeof(GPRArgRegs[0]));
2559 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2562 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2563 VARegSize = NumGPRs * 4;
2564 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2567 // The remaining GPRs hold either the beginning of variable-argument
2568 // data, or the beginning of an aggregate passed by value (usuall
2569 // byval). Either way, we allocate stack slots adjacent to the data
2570 // provided by our caller, and store the unallocated registers there.
2571 // If this is a variadic function, the va_list pointer will begin with
2572 // these values; otherwise, this reassembles a (byval) structure that
2573 // was split between registers and memory.
2575 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2576 DebugLoc dl, SDValue &Chain,
2577 const Value *OrigArg,
2578 unsigned OffsetFromOrigArg,
2580 bool ForceMutable) const {
2581 MachineFunction &MF = DAG.getMachineFunction();
2582 MachineFrameInfo *MFI = MF.getFrameInfo();
2583 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2584 unsigned firstRegToSaveIndex;
2585 if (CCInfo.isFirstByValRegValid())
2586 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2588 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2589 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2592 unsigned VARegSize, VARegSaveSize;
2593 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2594 if (VARegSaveSize) {
2595 // If this function is vararg, store any remaining integer argument regs
2596 // to their spots on the stack so that they may be loaded by deferencing
2597 // the result of va_next.
2598 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2599 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2600 ArgOffset + VARegSaveSize
2603 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2606 SmallVector<SDValue, 4> MemOps;
2607 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2608 const TargetRegisterClass *RC;
2609 if (AFI->isThumb1OnlyFunction())
2610 RC = &ARM::tGPRRegClass;
2612 RC = &ARM::GPRRegClass;
2614 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2615 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2617 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2618 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2620 MemOps.push_back(Store);
2621 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2622 DAG.getConstant(4, getPointerTy()));
2624 if (!MemOps.empty())
2625 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2626 &MemOps[0], MemOps.size());
2628 // This will point to the next argument passed via stack.
2629 AFI->setVarArgsFrameIndex(
2630 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2634 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2635 CallingConv::ID CallConv, bool isVarArg,
2636 const SmallVectorImpl<ISD::InputArg>
2638 DebugLoc dl, SelectionDAG &DAG,
2639 SmallVectorImpl<SDValue> &InVals)
2641 MachineFunction &MF = DAG.getMachineFunction();
2642 MachineFrameInfo *MFI = MF.getFrameInfo();
2644 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2646 // Assign locations to all of the incoming arguments.
2647 SmallVector<CCValAssign, 16> ArgLocs;
2648 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2649 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2650 CCInfo.AnalyzeFormalArguments(Ins,
2651 CCAssignFnForNode(CallConv, /* Return*/ false,
2654 SmallVector<SDValue, 16> ArgValues;
2655 int lastInsIndex = -1;
2657 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2658 unsigned CurArgIdx = 0;
2659 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2660 CCValAssign &VA = ArgLocs[i];
2661 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2662 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2663 // Arguments stored in registers.
2664 if (VA.isRegLoc()) {
2665 EVT RegVT = VA.getLocVT();
2667 if (VA.needsCustom()) {
2668 // f64 and vector types are split up into multiple registers or
2669 // combinations of registers and stack slots.
2670 if (VA.getLocVT() == MVT::v2f64) {
2671 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2673 VA = ArgLocs[++i]; // skip ahead to next loc
2675 if (VA.isMemLoc()) {
2676 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2677 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2678 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2679 MachinePointerInfo::getFixedStack(FI),
2680 false, false, false, 0);
2682 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2685 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2686 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2687 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2688 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2689 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2691 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2694 const TargetRegisterClass *RC;
2696 if (RegVT == MVT::f32)
2697 RC = &ARM::SPRRegClass;
2698 else if (RegVT == MVT::f64)
2699 RC = &ARM::DPRRegClass;
2700 else if (RegVT == MVT::v2f64)
2701 RC = &ARM::QPRRegClass;
2702 else if (RegVT == MVT::i32)
2703 RC = AFI->isThumb1OnlyFunction() ?
2704 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2705 (const TargetRegisterClass*)&ARM::GPRRegClass;
2707 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2709 // Transform the arguments in physical registers into virtual ones.
2710 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2711 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2714 // If this is an 8 or 16-bit value, it is really passed promoted
2715 // to 32 bits. Insert an assert[sz]ext to capture this, then
2716 // truncate to the right size.
2717 switch (VA.getLocInfo()) {
2718 default: llvm_unreachable("Unknown loc info!");
2719 case CCValAssign::Full: break;
2720 case CCValAssign::BCvt:
2721 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2723 case CCValAssign::SExt:
2724 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2725 DAG.getValueType(VA.getValVT()));
2726 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2728 case CCValAssign::ZExt:
2729 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2730 DAG.getValueType(VA.getValVT()));
2731 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2735 InVals.push_back(ArgValue);
2737 } else { // VA.isRegLoc()
2740 assert(VA.isMemLoc());
2741 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2743 int index = ArgLocs[i].getValNo();
2745 // Some Ins[] entries become multiple ArgLoc[] entries.
2746 // Process them only once.
2747 if (index != lastInsIndex)
2749 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2750 // FIXME: For now, all byval parameter objects are marked mutable.
2751 // This can be changed with more analysis.
2752 // In case of tail call optimization mark all arguments mutable.
2753 // Since they could be overwritten by lowering of arguments in case of
2755 if (Flags.isByVal()) {
2756 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2757 if (!AFI->getVarArgsFrameIndex()) {
2758 VarArgStyleRegisters(CCInfo, DAG,
2759 dl, Chain, CurOrigArg,
2760 Ins[VA.getValNo()].PartOffset,
2761 VA.getLocMemOffset(),
2762 true /*force mutable frames*/);
2763 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2764 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2766 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2767 VA.getLocMemOffset(), false);
2768 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2771 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2772 VA.getLocMemOffset(), true);
2774 // Create load nodes to retrieve arguments from the stack.
2775 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2776 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2777 MachinePointerInfo::getFixedStack(FI),
2778 false, false, false, 0));
2780 lastInsIndex = index;
2787 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2788 CCInfo.getNextStackOffset());
2793 /// isFloatingPointZero - Return true if this is +0.0.
2794 static bool isFloatingPointZero(SDValue Op) {
2795 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2796 return CFP->getValueAPF().isPosZero();
2797 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2798 // Maybe this has already been legalized into the constant pool?
2799 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2800 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2801 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2802 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2803 return CFP->getValueAPF().isPosZero();
2809 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2810 /// the given operands.
2812 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2813 SDValue &ARMcc, SelectionDAG &DAG,
2814 DebugLoc dl) const {
2815 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2816 unsigned C = RHSC->getZExtValue();
2817 if (!isLegalICmpImmediate(C)) {
2818 // Constant does not fit, try adjusting it by one?
2823 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2824 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2825 RHS = DAG.getConstant(C-1, MVT::i32);
2830 if (C != 0 && isLegalICmpImmediate(C-1)) {
2831 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2832 RHS = DAG.getConstant(C-1, MVT::i32);
2837 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2838 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2839 RHS = DAG.getConstant(C+1, MVT::i32);
2844 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2845 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2846 RHS = DAG.getConstant(C+1, MVT::i32);
2853 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2854 ARMISD::NodeType CompareType;
2857 CompareType = ARMISD::CMP;
2862 CompareType = ARMISD::CMPZ;
2865 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2866 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2869 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2871 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2872 DebugLoc dl) const {
2874 if (!isFloatingPointZero(RHS))
2875 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2877 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2878 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2881 /// duplicateCmp - Glue values can have only one use, so this function
2882 /// duplicates a comparison node.
2884 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2885 unsigned Opc = Cmp.getOpcode();
2886 DebugLoc DL = Cmp.getDebugLoc();
2887 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2888 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2890 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2891 Cmp = Cmp.getOperand(0);
2892 Opc = Cmp.getOpcode();
2893 if (Opc == ARMISD::CMPFP)
2894 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2896 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2897 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2899 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2902 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2903 SDValue Cond = Op.getOperand(0);
2904 SDValue SelectTrue = Op.getOperand(1);
2905 SDValue SelectFalse = Op.getOperand(2);
2906 DebugLoc dl = Op.getDebugLoc();
2910 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2911 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2913 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2914 const ConstantSDNode *CMOVTrue =
2915 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2916 const ConstantSDNode *CMOVFalse =
2917 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2919 if (CMOVTrue && CMOVFalse) {
2920 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2921 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2925 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2927 False = SelectFalse;
2928 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2933 if (True.getNode() && False.getNode()) {
2934 EVT VT = Op.getValueType();
2935 SDValue ARMcc = Cond.getOperand(2);
2936 SDValue CCR = Cond.getOperand(3);
2937 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2938 assert(True.getValueType() == VT);
2939 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2944 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2945 // undefined bits before doing a full-word comparison with zero.
2946 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2947 DAG.getConstant(1, Cond.getValueType()));
2949 return DAG.getSelectCC(dl, Cond,
2950 DAG.getConstant(0, Cond.getValueType()),
2951 SelectTrue, SelectFalse, ISD::SETNE);
2954 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2955 EVT VT = Op.getValueType();
2956 SDValue LHS = Op.getOperand(0);
2957 SDValue RHS = Op.getOperand(1);
2958 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2959 SDValue TrueVal = Op.getOperand(2);
2960 SDValue FalseVal = Op.getOperand(3);
2961 DebugLoc dl = Op.getDebugLoc();
2963 if (LHS.getValueType() == MVT::i32) {
2965 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2966 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2967 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2970 ARMCC::CondCodes CondCode, CondCode2;
2971 FPCCToARMCC(CC, CondCode, CondCode2);
2973 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2974 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2975 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2976 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2978 if (CondCode2 != ARMCC::AL) {
2979 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2980 // FIXME: Needs another CMP because flag can have but one use.
2981 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2982 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2983 Result, TrueVal, ARMcc2, CCR, Cmp2);
2988 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2989 /// to morph to an integer compare sequence.
2990 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2991 const ARMSubtarget *Subtarget) {
2992 SDNode *N = Op.getNode();
2993 if (!N->hasOneUse())
2994 // Otherwise it requires moving the value from fp to integer registers.
2996 if (!N->getNumValues())
2998 EVT VT = Op.getValueType();
2999 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3000 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3001 // vmrs are very slow, e.g. cortex-a8.
3004 if (isFloatingPointZero(Op)) {
3008 return ISD::isNormalLoad(N);
3011 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3012 if (isFloatingPointZero(Op))
3013 return DAG.getConstant(0, MVT::i32);
3015 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3016 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3017 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3018 Ld->isVolatile(), Ld->isNonTemporal(),
3019 Ld->isInvariant(), Ld->getAlignment());
3021 llvm_unreachable("Unknown VFP cmp argument!");
3024 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3025 SDValue &RetVal1, SDValue &RetVal2) {
3026 if (isFloatingPointZero(Op)) {
3027 RetVal1 = DAG.getConstant(0, MVT::i32);
3028 RetVal2 = DAG.getConstant(0, MVT::i32);
3032 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3033 SDValue Ptr = Ld->getBasePtr();
3034 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3035 Ld->getChain(), Ptr,
3036 Ld->getPointerInfo(),
3037 Ld->isVolatile(), Ld->isNonTemporal(),
3038 Ld->isInvariant(), Ld->getAlignment());
3040 EVT PtrType = Ptr.getValueType();
3041 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3042 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3043 PtrType, Ptr, DAG.getConstant(4, PtrType));
3044 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3045 Ld->getChain(), NewPtr,
3046 Ld->getPointerInfo().getWithOffset(4),
3047 Ld->isVolatile(), Ld->isNonTemporal(),
3048 Ld->isInvariant(), NewAlign);
3052 llvm_unreachable("Unknown VFP cmp argument!");
3055 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3056 /// f32 and even f64 comparisons to integer ones.
3058 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3059 SDValue Chain = Op.getOperand(0);
3060 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3061 SDValue LHS = Op.getOperand(2);
3062 SDValue RHS = Op.getOperand(3);
3063 SDValue Dest = Op.getOperand(4);
3064 DebugLoc dl = Op.getDebugLoc();
3066 bool LHSSeenZero = false;
3067 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3068 bool RHSSeenZero = false;
3069 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3070 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3071 // If unsafe fp math optimization is enabled and there are no other uses of
3072 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3073 // to an integer comparison.
3074 if (CC == ISD::SETOEQ)
3076 else if (CC == ISD::SETUNE)
3079 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3081 if (LHS.getValueType() == MVT::f32) {
3082 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3083 bitcastf32Toi32(LHS, DAG), Mask);
3084 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3085 bitcastf32Toi32(RHS, DAG), Mask);
3086 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3087 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3088 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3089 Chain, Dest, ARMcc, CCR, Cmp);
3094 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3095 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3096 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3097 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3098 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3099 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3100 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3101 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3102 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3108 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3109 SDValue Chain = Op.getOperand(0);
3110 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3111 SDValue LHS = Op.getOperand(2);
3112 SDValue RHS = Op.getOperand(3);
3113 SDValue Dest = Op.getOperand(4);
3114 DebugLoc dl = Op.getDebugLoc();
3116 if (LHS.getValueType() == MVT::i32) {
3118 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3119 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3120 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3121 Chain, Dest, ARMcc, CCR, Cmp);
3124 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3126 if (getTargetMachine().Options.UnsafeFPMath &&
3127 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3128 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3129 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3130 if (Result.getNode())
3134 ARMCC::CondCodes CondCode, CondCode2;
3135 FPCCToARMCC(CC, CondCode, CondCode2);
3137 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3138 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3139 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3140 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3141 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3142 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3143 if (CondCode2 != ARMCC::AL) {
3144 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3145 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3146 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3151 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3152 SDValue Chain = Op.getOperand(0);
3153 SDValue Table = Op.getOperand(1);
3154 SDValue Index = Op.getOperand(2);
3155 DebugLoc dl = Op.getDebugLoc();
3157 EVT PTy = getPointerTy();
3158 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3159 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3160 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3161 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3162 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3163 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3164 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3165 if (Subtarget->isThumb2()) {
3166 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3167 // which does another jump to the destination. This also makes it easier
3168 // to translate it to TBB / TBH later.
3169 // FIXME: This might not work if the function is extremely large.
3170 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3171 Addr, Op.getOperand(2), JTI, UId);
3173 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3174 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3175 MachinePointerInfo::getJumpTable(),
3176 false, false, false, 0);
3177 Chain = Addr.getValue(1);
3178 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3179 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3181 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3182 MachinePointerInfo::getJumpTable(),
3183 false, false, false, 0);
3184 Chain = Addr.getValue(1);
3185 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3189 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3190 EVT VT = Op.getValueType();
3191 DebugLoc dl = Op.getDebugLoc();
3193 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3194 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3196 return DAG.UnrollVectorOp(Op.getNode());
3199 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3200 "Invalid type for custom lowering!");
3201 if (VT != MVT::v4i16)
3202 return DAG.UnrollVectorOp(Op.getNode());
3204 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3205 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3208 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3209 EVT VT = Op.getValueType();
3211 return LowerVectorFP_TO_INT(Op, DAG);
3213 DebugLoc dl = Op.getDebugLoc();
3216 switch (Op.getOpcode()) {
3217 default: llvm_unreachable("Invalid opcode!");
3218 case ISD::FP_TO_SINT:
3219 Opc = ARMISD::FTOSI;
3221 case ISD::FP_TO_UINT:
3222 Opc = ARMISD::FTOUI;
3225 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3226 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3229 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3230 EVT VT = Op.getValueType();
3231 DebugLoc dl = Op.getDebugLoc();
3233 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3234 if (VT.getVectorElementType() == MVT::f32)
3236 return DAG.UnrollVectorOp(Op.getNode());
3239 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3240 "Invalid type for custom lowering!");
3241 if (VT != MVT::v4f32)
3242 return DAG.UnrollVectorOp(Op.getNode());
3246 switch (Op.getOpcode()) {
3247 default: llvm_unreachable("Invalid opcode!");
3248 case ISD::SINT_TO_FP:
3249 CastOpc = ISD::SIGN_EXTEND;
3250 Opc = ISD::SINT_TO_FP;
3252 case ISD::UINT_TO_FP:
3253 CastOpc = ISD::ZERO_EXTEND;
3254 Opc = ISD::UINT_TO_FP;
3258 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3259 return DAG.getNode(Opc, dl, VT, Op);
3262 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3263 EVT VT = Op.getValueType();
3265 return LowerVectorINT_TO_FP(Op, DAG);
3267 DebugLoc dl = Op.getDebugLoc();
3270 switch (Op.getOpcode()) {
3271 default: llvm_unreachable("Invalid opcode!");
3272 case ISD::SINT_TO_FP:
3273 Opc = ARMISD::SITOF;
3275 case ISD::UINT_TO_FP:
3276 Opc = ARMISD::UITOF;
3280 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3281 return DAG.getNode(Opc, dl, VT, Op);
3284 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3285 // Implement fcopysign with a fabs and a conditional fneg.
3286 SDValue Tmp0 = Op.getOperand(0);
3287 SDValue Tmp1 = Op.getOperand(1);
3288 DebugLoc dl = Op.getDebugLoc();
3289 EVT VT = Op.getValueType();
3290 EVT SrcVT = Tmp1.getValueType();
3291 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3292 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3293 bool UseNEON = !InGPR && Subtarget->hasNEON();
3296 // Use VBSL to copy the sign bit.
3297 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3298 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3299 DAG.getTargetConstant(EncodedVal, MVT::i32));
3300 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3302 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3303 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3304 DAG.getConstant(32, MVT::i32));
3305 else /*if (VT == MVT::f32)*/
3306 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3307 if (SrcVT == MVT::f32) {
3308 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3310 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3311 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3312 DAG.getConstant(32, MVT::i32));
3313 } else if (VT == MVT::f32)
3314 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3315 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3316 DAG.getConstant(32, MVT::i32));
3317 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3318 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3320 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3322 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3323 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3324 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3326 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3327 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3328 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3329 if (VT == MVT::f32) {
3330 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3331 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3332 DAG.getConstant(0, MVT::i32));
3334 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3340 // Bitcast operand 1 to i32.
3341 if (SrcVT == MVT::f64)
3342 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3343 &Tmp1, 1).getValue(1);
3344 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3346 // Or in the signbit with integer operations.
3347 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3348 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3349 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3350 if (VT == MVT::f32) {
3351 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3352 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3353 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3354 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3357 // f64: Or the high part with signbit and then combine two parts.
3358 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3360 SDValue Lo = Tmp0.getValue(0);
3361 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3362 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3363 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3366 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3367 MachineFunction &MF = DAG.getMachineFunction();
3368 MachineFrameInfo *MFI = MF.getFrameInfo();
3369 MFI->setReturnAddressIsTaken(true);
3371 EVT VT = Op.getValueType();
3372 DebugLoc dl = Op.getDebugLoc();
3373 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3375 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3376 SDValue Offset = DAG.getConstant(4, MVT::i32);
3377 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3378 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3379 MachinePointerInfo(), false, false, false, 0);
3382 // Return LR, which contains the return address. Mark it an implicit live-in.
3383 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3384 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3387 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3388 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3389 MFI->setFrameAddressIsTaken(true);
3391 EVT VT = Op.getValueType();
3392 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3393 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3394 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3395 ? ARM::R7 : ARM::R11;
3396 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3398 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3399 MachinePointerInfo(),
3400 false, false, false, 0);
3404 /// ExpandBITCAST - If the target supports VFP, this function is called to
3405 /// expand a bit convert where either the source or destination type is i64 to
3406 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3407 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3408 /// vectors), since the legalizer won't know what to do with that.
3409 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3410 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3411 DebugLoc dl = N->getDebugLoc();
3412 SDValue Op = N->getOperand(0);
3414 // This function is only supposed to be called for i64 types, either as the
3415 // source or destination of the bit convert.
3416 EVT SrcVT = Op.getValueType();
3417 EVT DstVT = N->getValueType(0);
3418 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3419 "ExpandBITCAST called for non-i64 type");
3421 // Turn i64->f64 into VMOVDRR.
3422 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3423 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3424 DAG.getConstant(0, MVT::i32));
3425 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3426 DAG.getConstant(1, MVT::i32));
3427 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3428 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3431 // Turn f64->i64 into VMOVRRD.
3432 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3433 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3434 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3435 // Merge the pieces into a single i64 value.
3436 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3442 /// getZeroVector - Returns a vector of specified type with all zero elements.
3443 /// Zero vectors are used to represent vector negation and in those cases
3444 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3445 /// not support i64 elements, so sometimes the zero vectors will need to be
3446 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3448 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3449 assert(VT.isVector() && "Expected a vector type");
3450 // The canonical modified immediate encoding of a zero vector is....0!
3451 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3452 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3453 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3454 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3457 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3458 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3459 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3460 SelectionDAG &DAG) const {
3461 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3462 EVT VT = Op.getValueType();
3463 unsigned VTBits = VT.getSizeInBits();
3464 DebugLoc dl = Op.getDebugLoc();
3465 SDValue ShOpLo = Op.getOperand(0);
3466 SDValue ShOpHi = Op.getOperand(1);
3467 SDValue ShAmt = Op.getOperand(2);
3469 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3471 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3473 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3474 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3475 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3476 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3477 DAG.getConstant(VTBits, MVT::i32));
3478 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3479 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3480 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3482 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3483 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3485 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3486 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3489 SDValue Ops[2] = { Lo, Hi };
3490 return DAG.getMergeValues(Ops, 2, dl);
3493 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3494 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3495 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3496 SelectionDAG &DAG) const {
3497 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3498 EVT VT = Op.getValueType();
3499 unsigned VTBits = VT.getSizeInBits();
3500 DebugLoc dl = Op.getDebugLoc();
3501 SDValue ShOpLo = Op.getOperand(0);
3502 SDValue ShOpHi = Op.getOperand(1);
3503 SDValue ShAmt = Op.getOperand(2);
3506 assert(Op.getOpcode() == ISD::SHL_PARTS);
3507 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3508 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3509 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3510 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3511 DAG.getConstant(VTBits, MVT::i32));
3512 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3513 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3515 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3516 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3517 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3519 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3520 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3523 SDValue Ops[2] = { Lo, Hi };
3524 return DAG.getMergeValues(Ops, 2, dl);
3527 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3528 SelectionDAG &DAG) const {
3529 // The rounding mode is in bits 23:22 of the FPSCR.
3530 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3531 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3532 // so that the shift + and get folded into a bitfield extract.
3533 DebugLoc dl = Op.getDebugLoc();
3534 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3535 DAG.getConstant(Intrinsic::arm_get_fpscr,
3537 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3538 DAG.getConstant(1U << 22, MVT::i32));
3539 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3540 DAG.getConstant(22, MVT::i32));
3541 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3542 DAG.getConstant(3, MVT::i32));
3545 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3546 const ARMSubtarget *ST) {
3547 EVT VT = N->getValueType(0);
3548 DebugLoc dl = N->getDebugLoc();
3550 if (!ST->hasV6T2Ops())
3553 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3554 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3557 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3558 const ARMSubtarget *ST) {
3559 EVT VT = N->getValueType(0);
3560 DebugLoc dl = N->getDebugLoc();
3565 // Lower vector shifts on NEON to use VSHL.
3566 assert(ST->hasNEON() && "unexpected vector shift");
3568 // Left shifts translate directly to the vshiftu intrinsic.
3569 if (N->getOpcode() == ISD::SHL)
3570 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3571 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3572 N->getOperand(0), N->getOperand(1));
3574 assert((N->getOpcode() == ISD::SRA ||
3575 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3577 // NEON uses the same intrinsics for both left and right shifts. For
3578 // right shifts, the shift amounts are negative, so negate the vector of
3580 EVT ShiftVT = N->getOperand(1).getValueType();
3581 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3582 getZeroVector(ShiftVT, DAG, dl),
3584 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3585 Intrinsic::arm_neon_vshifts :
3586 Intrinsic::arm_neon_vshiftu);
3587 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3588 DAG.getConstant(vshiftInt, MVT::i32),
3589 N->getOperand(0), NegatedCount);
3592 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3593 const ARMSubtarget *ST) {
3594 EVT VT = N->getValueType(0);
3595 DebugLoc dl = N->getDebugLoc();
3597 // We can get here for a node like i32 = ISD::SHL i32, i64
3601 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3602 "Unknown shift to lower!");
3604 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3605 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3606 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3609 // If we are in thumb mode, we don't have RRX.
3610 if (ST->isThumb1Only()) return SDValue();
3612 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3613 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3614 DAG.getConstant(0, MVT::i32));
3615 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3616 DAG.getConstant(1, MVT::i32));
3618 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3619 // captures the result into a carry flag.
3620 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3621 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3623 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3624 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3626 // Merge the pieces into a single i64 value.
3627 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3630 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3631 SDValue TmpOp0, TmpOp1;
3632 bool Invert = false;
3636 SDValue Op0 = Op.getOperand(0);
3637 SDValue Op1 = Op.getOperand(1);
3638 SDValue CC = Op.getOperand(2);
3639 EVT VT = Op.getValueType();
3640 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3641 DebugLoc dl = Op.getDebugLoc();
3643 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3644 switch (SetCCOpcode) {
3645 default: llvm_unreachable("Illegal FP comparison");
3647 case ISD::SETNE: Invert = true; // Fallthrough
3649 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3651 case ISD::SETLT: Swap = true; // Fallthrough
3653 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3655 case ISD::SETLE: Swap = true; // Fallthrough
3657 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3658 case ISD::SETUGE: Swap = true; // Fallthrough
3659 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3660 case ISD::SETUGT: Swap = true; // Fallthrough
3661 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3662 case ISD::SETUEQ: Invert = true; // Fallthrough
3664 // Expand this to (OLT | OGT).
3668 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3669 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3671 case ISD::SETUO: Invert = true; // Fallthrough
3673 // Expand this to (OLT | OGE).
3677 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3678 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3682 // Integer comparisons.
3683 switch (SetCCOpcode) {
3684 default: llvm_unreachable("Illegal integer comparison");
3685 case ISD::SETNE: Invert = true;
3686 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3687 case ISD::SETLT: Swap = true;
3688 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3689 case ISD::SETLE: Swap = true;
3690 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3691 case ISD::SETULT: Swap = true;
3692 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3693 case ISD::SETULE: Swap = true;
3694 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3697 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3698 if (Opc == ARMISD::VCEQ) {
3701 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3703 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3706 // Ignore bitconvert.
3707 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3708 AndOp = AndOp.getOperand(0);
3710 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3712 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3713 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3720 std::swap(Op0, Op1);
3722 // If one of the operands is a constant vector zero, attempt to fold the
3723 // comparison to a specialized compare-against-zero form.
3725 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3727 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3728 if (Opc == ARMISD::VCGE)
3729 Opc = ARMISD::VCLEZ;
3730 else if (Opc == ARMISD::VCGT)
3731 Opc = ARMISD::VCLTZ;
3736 if (SingleOp.getNode()) {
3739 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3741 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3743 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3745 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3747 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3749 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3752 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3756 Result = DAG.getNOT(dl, Result, VT);
3761 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3762 /// valid vector constant for a NEON instruction with a "modified immediate"
3763 /// operand (e.g., VMOV). If so, return the encoded value.
3764 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3765 unsigned SplatBitSize, SelectionDAG &DAG,
3766 EVT &VT, bool is128Bits, NEONModImmType type) {
3767 unsigned OpCmode, Imm;
3769 // SplatBitSize is set to the smallest size that splats the vector, so a
3770 // zero vector will always have SplatBitSize == 8. However, NEON modified
3771 // immediate instructions others than VMOV do not support the 8-bit encoding
3772 // of a zero vector, and the default encoding of zero is supposed to be the
3777 switch (SplatBitSize) {
3779 if (type != VMOVModImm)
3781 // Any 1-byte value is OK. Op=0, Cmode=1110.
3782 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3785 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3789 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3790 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3791 if ((SplatBits & ~0xff) == 0) {
3792 // Value = 0x00nn: Op=x, Cmode=100x.
3797 if ((SplatBits & ~0xff00) == 0) {
3798 // Value = 0xnn00: Op=x, Cmode=101x.
3800 Imm = SplatBits >> 8;
3806 // NEON's 32-bit VMOV supports splat values where:
3807 // * only one byte is nonzero, or
3808 // * the least significant byte is 0xff and the second byte is nonzero, or
3809 // * the least significant 2 bytes are 0xff and the third is nonzero.
3810 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3811 if ((SplatBits & ~0xff) == 0) {
3812 // Value = 0x000000nn: Op=x, Cmode=000x.
3817 if ((SplatBits & ~0xff00) == 0) {
3818 // Value = 0x0000nn00: Op=x, Cmode=001x.
3820 Imm = SplatBits >> 8;
3823 if ((SplatBits & ~0xff0000) == 0) {
3824 // Value = 0x00nn0000: Op=x, Cmode=010x.
3826 Imm = SplatBits >> 16;
3829 if ((SplatBits & ~0xff000000) == 0) {
3830 // Value = 0xnn000000: Op=x, Cmode=011x.
3832 Imm = SplatBits >> 24;
3836 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3837 if (type == OtherModImm) return SDValue();
3839 if ((SplatBits & ~0xffff) == 0 &&
3840 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3841 // Value = 0x0000nnff: Op=x, Cmode=1100.
3843 Imm = SplatBits >> 8;
3848 if ((SplatBits & ~0xffffff) == 0 &&
3849 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3850 // Value = 0x00nnffff: Op=x, Cmode=1101.
3852 Imm = SplatBits >> 16;
3853 SplatBits |= 0xffff;
3857 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3858 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3859 // VMOV.I32. A (very) minor optimization would be to replicate the value
3860 // and fall through here to test for a valid 64-bit splat. But, then the
3861 // caller would also need to check and handle the change in size.
3865 if (type != VMOVModImm)
3867 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3868 uint64_t BitMask = 0xff;
3870 unsigned ImmMask = 1;
3872 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3873 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3876 } else if ((SplatBits & BitMask) != 0) {
3882 // Op=1, Cmode=1110.
3885 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3890 llvm_unreachable("unexpected size for isNEONModifiedImm");
3893 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3894 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3897 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3898 const ARMSubtarget *ST) const {
3899 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3902 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3903 assert(Op.getValueType() == MVT::f32 &&
3904 "ConstantFP custom lowering should only occur for f32.");
3906 // Try splatting with a VMOV.f32...
3907 APFloat FPVal = CFP->getValueAPF();
3908 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3910 DebugLoc DL = Op.getDebugLoc();
3911 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3912 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3914 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3915 DAG.getConstant(0, MVT::i32));
3918 // If that fails, try a VMOV.i32
3920 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3921 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3923 if (NewVal != SDValue()) {
3924 DebugLoc DL = Op.getDebugLoc();
3925 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3927 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3929 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3930 DAG.getConstant(0, MVT::i32));
3933 // Finally, try a VMVN.i32
3934 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3936 if (NewVal != SDValue()) {
3937 DebugLoc DL = Op.getDebugLoc();
3938 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3939 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3941 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3942 DAG.getConstant(0, MVT::i32));
3948 // check if an VEXT instruction can handle the shuffle mask when the
3949 // vector sources of the shuffle are the same.
3950 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
3951 unsigned NumElts = VT.getVectorNumElements();
3953 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3959 // If this is a VEXT shuffle, the immediate value is the index of the first
3960 // element. The other shuffle indices must be the successive elements after
3962 unsigned ExpectedElt = Imm;
3963 for (unsigned i = 1; i < NumElts; ++i) {
3964 // Increment the expected index. If it wraps around, just follow it
3965 // back to index zero and keep going.
3967 if (ExpectedElt == NumElts)
3970 if (M[i] < 0) continue; // ignore UNDEF indices
3971 if (ExpectedElt != static_cast<unsigned>(M[i]))
3979 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3980 bool &ReverseVEXT, unsigned &Imm) {
3981 unsigned NumElts = VT.getVectorNumElements();
3982 ReverseVEXT = false;
3984 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3990 // If this is a VEXT shuffle, the immediate value is the index of the first
3991 // element. The other shuffle indices must be the successive elements after
3993 unsigned ExpectedElt = Imm;
3994 for (unsigned i = 1; i < NumElts; ++i) {
3995 // Increment the expected index. If it wraps around, it may still be
3996 // a VEXT but the source vectors must be swapped.
3998 if (ExpectedElt == NumElts * 2) {
4003 if (M[i] < 0) continue; // ignore UNDEF indices
4004 if (ExpectedElt != static_cast<unsigned>(M[i]))
4008 // Adjust the index value if the source operands will be swapped.
4015 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4016 /// instruction with the specified blocksize. (The order of the elements
4017 /// within each block of the vector is reversed.)
4018 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4019 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4020 "Only possible block sizes for VREV are: 16, 32, 64");
4022 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4026 unsigned NumElts = VT.getVectorNumElements();
4027 unsigned BlockElts = M[0] + 1;
4028 // If the first shuffle index is UNDEF, be optimistic.
4030 BlockElts = BlockSize / EltSz;
4032 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4035 for (unsigned i = 0; i < NumElts; ++i) {
4036 if (M[i] < 0) continue; // ignore UNDEF indices
4037 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4044 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4045 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4046 // range, then 0 is placed into the resulting vector. So pretty much any mask
4047 // of 8 elements can work here.
4048 return VT == MVT::v8i8 && M.size() == 8;
4051 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4052 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4056 unsigned NumElts = VT.getVectorNumElements();
4057 WhichResult = (M[0] == 0 ? 0 : 1);
4058 for (unsigned i = 0; i < NumElts; i += 2) {
4059 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4060 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4066 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4067 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4068 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4069 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4070 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4074 unsigned NumElts = VT.getVectorNumElements();
4075 WhichResult = (M[0] == 0 ? 0 : 1);
4076 for (unsigned i = 0; i < NumElts; i += 2) {
4077 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4078 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4084 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4085 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4089 unsigned NumElts = VT.getVectorNumElements();
4090 WhichResult = (M[0] == 0 ? 0 : 1);
4091 for (unsigned i = 0; i != NumElts; ++i) {
4092 if (M[i] < 0) continue; // ignore UNDEF indices
4093 if ((unsigned) M[i] != 2 * i + WhichResult)
4097 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4098 if (VT.is64BitVector() && EltSz == 32)
4104 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4105 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4106 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4107 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4108 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4112 unsigned Half = VT.getVectorNumElements() / 2;
4113 WhichResult = (M[0] == 0 ? 0 : 1);
4114 for (unsigned j = 0; j != 2; ++j) {
4115 unsigned Idx = WhichResult;
4116 for (unsigned i = 0; i != Half; ++i) {
4117 int MIdx = M[i + j * Half];
4118 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4124 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4125 if (VT.is64BitVector() && EltSz == 32)
4131 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4132 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4136 unsigned NumElts = VT.getVectorNumElements();
4137 WhichResult = (M[0] == 0 ? 0 : 1);
4138 unsigned Idx = WhichResult * NumElts / 2;
4139 for (unsigned i = 0; i != NumElts; i += 2) {
4140 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4141 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4146 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4147 if (VT.is64BitVector() && EltSz == 32)
4153 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4154 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4155 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4156 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4157 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4161 unsigned NumElts = VT.getVectorNumElements();
4162 WhichResult = (M[0] == 0 ? 0 : 1);
4163 unsigned Idx = WhichResult * NumElts / 2;
4164 for (unsigned i = 0; i != NumElts; i += 2) {
4165 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4166 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4171 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4172 if (VT.is64BitVector() && EltSz == 32)
4178 // If N is an integer constant that can be moved into a register in one
4179 // instruction, return an SDValue of such a constant (will become a MOV
4180 // instruction). Otherwise return null.
4181 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4182 const ARMSubtarget *ST, DebugLoc dl) {
4184 if (!isa<ConstantSDNode>(N))
4186 Val = cast<ConstantSDNode>(N)->getZExtValue();
4188 if (ST->isThumb1Only()) {
4189 if (Val <= 255 || ~Val <= 255)
4190 return DAG.getConstant(Val, MVT::i32);
4192 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4193 return DAG.getConstant(Val, MVT::i32);
4198 // If this is a case we can't handle, return null and let the default
4199 // expansion code take care of it.
4200 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4201 const ARMSubtarget *ST) const {
4202 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4203 DebugLoc dl = Op.getDebugLoc();
4204 EVT VT = Op.getValueType();
4206 APInt SplatBits, SplatUndef;
4207 unsigned SplatBitSize;
4209 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4210 if (SplatBitSize <= 64) {
4211 // Check if an immediate VMOV works.
4213 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4214 SplatUndef.getZExtValue(), SplatBitSize,
4215 DAG, VmovVT, VT.is128BitVector(),
4217 if (Val.getNode()) {
4218 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4219 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4222 // Try an immediate VMVN.
4223 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4224 Val = isNEONModifiedImm(NegatedImm,
4225 SplatUndef.getZExtValue(), SplatBitSize,
4226 DAG, VmovVT, VT.is128BitVector(),
4228 if (Val.getNode()) {
4229 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4230 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4233 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4234 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4235 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4237 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4238 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4244 // Scan through the operands to see if only one value is used.
4246 // As an optimisation, even if more than one value is used it may be more
4247 // profitable to splat with one value then change some lanes.
4249 // Heuristically we decide to do this if the vector has a "dominant" value,
4250 // defined as splatted to more than half of the lanes.
4251 unsigned NumElts = VT.getVectorNumElements();
4252 bool isOnlyLowElement = true;
4253 bool usesOnlyOneValue = true;
4254 bool hasDominantValue = false;
4255 bool isConstant = true;
4257 // Map of the number of times a particular SDValue appears in the
4259 DenseMap<SDValue, unsigned> ValueCounts;
4261 for (unsigned i = 0; i < NumElts; ++i) {
4262 SDValue V = Op.getOperand(i);
4263 if (V.getOpcode() == ISD::UNDEF)
4266 isOnlyLowElement = false;
4267 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4270 ValueCounts.insert(std::make_pair(V, 0));
4271 unsigned &Count = ValueCounts[V];
4273 // Is this value dominant? (takes up more than half of the lanes)
4274 if (++Count > (NumElts / 2)) {
4275 hasDominantValue = true;
4279 if (ValueCounts.size() != 1)
4280 usesOnlyOneValue = false;
4281 if (!Value.getNode() && ValueCounts.size() > 0)
4282 Value = ValueCounts.begin()->first;
4284 if (ValueCounts.size() == 0)
4285 return DAG.getUNDEF(VT);
4287 if (isOnlyLowElement)
4288 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4290 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4292 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4293 // i32 and try again.
4294 if (hasDominantValue && EltSize <= 32) {
4298 // If we are VDUPing a value that comes directly from a vector, that will
4299 // cause an unnecessary move to and from a GPR, where instead we could
4300 // just use VDUPLANE.
4301 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4302 // We need to create a new undef vector to use for the VDUPLANE if the
4303 // size of the vector from which we get the value is different than the
4304 // size of the vector that we need to create. We will insert the element
4305 // such that the register coalescer will remove unnecessary copies.
4306 if (VT != Value->getOperand(0).getValueType()) {
4307 ConstantSDNode *constIndex;
4308 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4309 assert(constIndex && "The index is not a constant!");
4310 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4311 VT.getVectorNumElements();
4312 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4313 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4314 Value, DAG.getConstant(index, MVT::i32)),
4315 DAG.getConstant(index, MVT::i32));
4317 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4318 Value->getOperand(0), Value->getOperand(1));
4322 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4324 if (!usesOnlyOneValue) {
4325 // The dominant value was splatted as 'N', but we now have to insert
4326 // all differing elements.
4327 for (unsigned I = 0; I < NumElts; ++I) {
4328 if (Op.getOperand(I) == Value)
4330 SmallVector<SDValue, 3> Ops;
4332 Ops.push_back(Op.getOperand(I));
4333 Ops.push_back(DAG.getConstant(I, MVT::i32));
4334 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4339 if (VT.getVectorElementType().isFloatingPoint()) {
4340 SmallVector<SDValue, 8> Ops;
4341 for (unsigned i = 0; i < NumElts; ++i)
4342 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4344 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4345 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4346 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4348 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4350 if (usesOnlyOneValue) {
4351 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4352 if (isConstant && Val.getNode())
4353 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4357 // If all elements are constants and the case above didn't get hit, fall back
4358 // to the default expansion, which will generate a load from the constant
4363 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4365 SDValue shuffle = ReconstructShuffle(Op, DAG);
4366 if (shuffle != SDValue())
4370 // Vectors with 32- or 64-bit elements can be built by directly assigning
4371 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4372 // will be legalized.
4373 if (EltSize >= 32) {
4374 // Do the expansion with floating-point types, since that is what the VFP
4375 // registers are defined to use, and since i64 is not legal.
4376 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4377 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4378 SmallVector<SDValue, 8> Ops;
4379 for (unsigned i = 0; i < NumElts; ++i)
4380 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4381 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4382 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4388 // Gather data to see if the operation can be modelled as a
4389 // shuffle in combination with VEXTs.
4390 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4391 SelectionDAG &DAG) const {
4392 DebugLoc dl = Op.getDebugLoc();
4393 EVT VT = Op.getValueType();
4394 unsigned NumElts = VT.getVectorNumElements();
4396 SmallVector<SDValue, 2> SourceVecs;
4397 SmallVector<unsigned, 2> MinElts;
4398 SmallVector<unsigned, 2> MaxElts;
4400 for (unsigned i = 0; i < NumElts; ++i) {
4401 SDValue V = Op.getOperand(i);
4402 if (V.getOpcode() == ISD::UNDEF)
4404 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4405 // A shuffle can only come from building a vector from various
4406 // elements of other vectors.
4408 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4409 VT.getVectorElementType()) {
4410 // This code doesn't know how to handle shuffles where the vector
4411 // element types do not match (this happens because type legalization
4412 // promotes the return type of EXTRACT_VECTOR_ELT).
4413 // FIXME: It might be appropriate to extend this code to handle
4414 // mismatched types.
4418 // Record this extraction against the appropriate vector if possible...
4419 SDValue SourceVec = V.getOperand(0);
4420 // If the element number isn't a constant, we can't effectively
4421 // analyze what's going on.
4422 if (!isa<ConstantSDNode>(V.getOperand(1)))
4424 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4425 bool FoundSource = false;
4426 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4427 if (SourceVecs[j] == SourceVec) {
4428 if (MinElts[j] > EltNo)
4430 if (MaxElts[j] < EltNo)
4437 // Or record a new source if not...
4439 SourceVecs.push_back(SourceVec);
4440 MinElts.push_back(EltNo);
4441 MaxElts.push_back(EltNo);
4445 // Currently only do something sane when at most two source vectors
4447 if (SourceVecs.size() > 2)
4450 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4451 int VEXTOffsets[2] = {0, 0};
4453 // This loop extracts the usage patterns of the source vectors
4454 // and prepares appropriate SDValues for a shuffle if possible.
4455 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4456 if (SourceVecs[i].getValueType() == VT) {
4457 // No VEXT necessary
4458 ShuffleSrcs[i] = SourceVecs[i];
4461 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4462 // It probably isn't worth padding out a smaller vector just to
4463 // break it down again in a shuffle.
4467 // Since only 64-bit and 128-bit vectors are legal on ARM and
4468 // we've eliminated the other cases...
4469 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4470 "unexpected vector sizes in ReconstructShuffle");
4472 if (MaxElts[i] - MinElts[i] >= NumElts) {
4473 // Span too large for a VEXT to cope
4477 if (MinElts[i] >= NumElts) {
4478 // The extraction can just take the second half
4479 VEXTOffsets[i] = NumElts;
4480 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4482 DAG.getIntPtrConstant(NumElts));
4483 } else if (MaxElts[i] < NumElts) {
4484 // The extraction can just take the first half
4486 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4488 DAG.getIntPtrConstant(0));
4490 // An actual VEXT is needed
4491 VEXTOffsets[i] = MinElts[i];
4492 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4494 DAG.getIntPtrConstant(0));
4495 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4497 DAG.getIntPtrConstant(NumElts));
4498 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4499 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4503 SmallVector<int, 8> Mask;
4505 for (unsigned i = 0; i < NumElts; ++i) {
4506 SDValue Entry = Op.getOperand(i);
4507 if (Entry.getOpcode() == ISD::UNDEF) {
4512 SDValue ExtractVec = Entry.getOperand(0);
4513 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4514 .getOperand(1))->getSExtValue();
4515 if (ExtractVec == SourceVecs[0]) {
4516 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4518 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4522 // Final check before we try to produce nonsense...
4523 if (isShuffleMaskLegal(Mask, VT))
4524 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4530 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4531 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4532 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4533 /// are assumed to be legal.
4535 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4537 if (VT.getVectorNumElements() == 4 &&
4538 (VT.is128BitVector() || VT.is64BitVector())) {
4539 unsigned PFIndexes[4];
4540 for (unsigned i = 0; i != 4; ++i) {
4544 PFIndexes[i] = M[i];
4547 // Compute the index in the perfect shuffle table.
4548 unsigned PFTableIndex =
4549 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4550 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4551 unsigned Cost = (PFEntry >> 30);
4558 unsigned Imm, WhichResult;
4560 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4561 return (EltSize >= 32 ||
4562 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4563 isVREVMask(M, VT, 64) ||
4564 isVREVMask(M, VT, 32) ||
4565 isVREVMask(M, VT, 16) ||
4566 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4567 isVTBLMask(M, VT) ||
4568 isVTRNMask(M, VT, WhichResult) ||
4569 isVUZPMask(M, VT, WhichResult) ||
4570 isVZIPMask(M, VT, WhichResult) ||
4571 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4572 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4573 isVZIP_v_undef_Mask(M, VT, WhichResult));
4576 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4577 /// the specified operations to build the shuffle.
4578 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4579 SDValue RHS, SelectionDAG &DAG,
4581 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4582 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4583 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4586 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4595 OP_VUZPL, // VUZP, left result
4596 OP_VUZPR, // VUZP, right result
4597 OP_VZIPL, // VZIP, left result
4598 OP_VZIPR, // VZIP, right result
4599 OP_VTRNL, // VTRN, left result
4600 OP_VTRNR // VTRN, right result
4603 if (OpNum == OP_COPY) {
4604 if (LHSID == (1*9+2)*9+3) return LHS;
4605 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4609 SDValue OpLHS, OpRHS;
4610 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4611 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4612 EVT VT = OpLHS.getValueType();
4615 default: llvm_unreachable("Unknown shuffle opcode!");
4617 // VREV divides the vector in half and swaps within the half.
4618 if (VT.getVectorElementType() == MVT::i32 ||
4619 VT.getVectorElementType() == MVT::f32)
4620 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4621 // vrev <4 x i16> -> VREV32
4622 if (VT.getVectorElementType() == MVT::i16)
4623 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4624 // vrev <4 x i8> -> VREV16
4625 assert(VT.getVectorElementType() == MVT::i8);
4626 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4631 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4632 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4636 return DAG.getNode(ARMISD::VEXT, dl, VT,
4638 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4641 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4642 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4645 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4646 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4649 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4650 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4654 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4655 ArrayRef<int> ShuffleMask,
4656 SelectionDAG &DAG) {
4657 // Check to see if we can use the VTBL instruction.
4658 SDValue V1 = Op.getOperand(0);
4659 SDValue V2 = Op.getOperand(1);
4660 DebugLoc DL = Op.getDebugLoc();
4662 SmallVector<SDValue, 8> VTBLMask;
4663 for (ArrayRef<int>::iterator
4664 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4665 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4667 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4668 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4669 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4672 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4673 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4677 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4678 SDValue V1 = Op.getOperand(0);
4679 SDValue V2 = Op.getOperand(1);
4680 DebugLoc dl = Op.getDebugLoc();
4681 EVT VT = Op.getValueType();
4682 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4684 // Convert shuffles that are directly supported on NEON to target-specific
4685 // DAG nodes, instead of keeping them as shuffles and matching them again
4686 // during code selection. This is more efficient and avoids the possibility
4687 // of inconsistencies between legalization and selection.
4688 // FIXME: floating-point vectors should be canonicalized to integer vectors
4689 // of the same time so that they get CSEd properly.
4690 ArrayRef<int> ShuffleMask = SVN->getMask();
4692 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4693 if (EltSize <= 32) {
4694 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4695 int Lane = SVN->getSplatIndex();
4696 // If this is undef splat, generate it via "just" vdup, if possible.
4697 if (Lane == -1) Lane = 0;
4699 // Test if V1 is a SCALAR_TO_VECTOR.
4700 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4701 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4703 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4704 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4706 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4707 !isa<ConstantSDNode>(V1.getOperand(0))) {
4708 bool IsScalarToVector = true;
4709 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4710 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4711 IsScalarToVector = false;
4714 if (IsScalarToVector)
4715 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4717 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4718 DAG.getConstant(Lane, MVT::i32));
4723 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4726 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4727 DAG.getConstant(Imm, MVT::i32));
4730 if (isVREVMask(ShuffleMask, VT, 64))
4731 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4732 if (isVREVMask(ShuffleMask, VT, 32))
4733 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4734 if (isVREVMask(ShuffleMask, VT, 16))
4735 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4737 if (V2->getOpcode() == ISD::UNDEF &&
4738 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4739 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4740 DAG.getConstant(Imm, MVT::i32));
4743 // Check for Neon shuffles that modify both input vectors in place.
4744 // If both results are used, i.e., if there are two shuffles with the same
4745 // source operands and with masks corresponding to both results of one of
4746 // these operations, DAG memoization will ensure that a single node is
4747 // used for both shuffles.
4748 unsigned WhichResult;
4749 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4750 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4751 V1, V2).getValue(WhichResult);
4752 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4753 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4754 V1, V2).getValue(WhichResult);
4755 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4756 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4757 V1, V2).getValue(WhichResult);
4759 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4760 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4761 V1, V1).getValue(WhichResult);
4762 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4763 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4764 V1, V1).getValue(WhichResult);
4765 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4766 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4767 V1, V1).getValue(WhichResult);
4770 // If the shuffle is not directly supported and it has 4 elements, use
4771 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4772 unsigned NumElts = VT.getVectorNumElements();
4774 unsigned PFIndexes[4];
4775 for (unsigned i = 0; i != 4; ++i) {
4776 if (ShuffleMask[i] < 0)
4779 PFIndexes[i] = ShuffleMask[i];
4782 // Compute the index in the perfect shuffle table.
4783 unsigned PFTableIndex =
4784 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4785 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4786 unsigned Cost = (PFEntry >> 30);
4789 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4792 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4793 if (EltSize >= 32) {
4794 // Do the expansion with floating-point types, since that is what the VFP
4795 // registers are defined to use, and since i64 is not legal.
4796 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4797 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4798 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4799 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4800 SmallVector<SDValue, 8> Ops;
4801 for (unsigned i = 0; i < NumElts; ++i) {
4802 if (ShuffleMask[i] < 0)
4803 Ops.push_back(DAG.getUNDEF(EltVT));
4805 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4806 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4807 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4810 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4811 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4814 if (VT == MVT::v8i8) {
4815 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4816 if (NewOp.getNode())
4823 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4824 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4825 SDValue Lane = Op.getOperand(2);
4826 if (!isa<ConstantSDNode>(Lane))
4832 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4833 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4834 SDValue Lane = Op.getOperand(1);
4835 if (!isa<ConstantSDNode>(Lane))
4838 SDValue Vec = Op.getOperand(0);
4839 if (Op.getValueType() == MVT::i32 &&
4840 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4841 DebugLoc dl = Op.getDebugLoc();
4842 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4848 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4849 // The only time a CONCAT_VECTORS operation can have legal types is when
4850 // two 64-bit vectors are concatenated to a 128-bit vector.
4851 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4852 "unexpected CONCAT_VECTORS");
4853 DebugLoc dl = Op.getDebugLoc();
4854 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4855 SDValue Op0 = Op.getOperand(0);
4856 SDValue Op1 = Op.getOperand(1);
4857 if (Op0.getOpcode() != ISD::UNDEF)
4858 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4859 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4860 DAG.getIntPtrConstant(0));
4861 if (Op1.getOpcode() != ISD::UNDEF)
4862 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4863 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4864 DAG.getIntPtrConstant(1));
4865 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4868 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4869 /// element has been zero/sign-extended, depending on the isSigned parameter,
4870 /// from an integer type half its size.
4871 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4873 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4874 EVT VT = N->getValueType(0);
4875 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4876 SDNode *BVN = N->getOperand(0).getNode();
4877 if (BVN->getValueType(0) != MVT::v4i32 ||
4878 BVN->getOpcode() != ISD::BUILD_VECTOR)
4880 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4881 unsigned HiElt = 1 - LoElt;
4882 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4883 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4884 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4885 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4886 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4889 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4890 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4893 if (Hi0->isNullValue() && Hi1->isNullValue())
4899 if (N->getOpcode() != ISD::BUILD_VECTOR)
4902 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4903 SDNode *Elt = N->getOperand(i).getNode();
4904 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4905 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4906 unsigned HalfSize = EltSize / 2;
4908 if (!isIntN(HalfSize, C->getSExtValue()))
4911 if (!isUIntN(HalfSize, C->getZExtValue()))
4922 /// isSignExtended - Check if a node is a vector value that is sign-extended
4923 /// or a constant BUILD_VECTOR with sign-extended elements.
4924 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4925 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4927 if (isExtendedBUILD_VECTOR(N, DAG, true))
4932 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4933 /// or a constant BUILD_VECTOR with zero-extended elements.
4934 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4935 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4937 if (isExtendedBUILD_VECTOR(N, DAG, false))
4942 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4943 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4944 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4945 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4946 return N->getOperand(0);
4947 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4948 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4949 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4950 LD->isNonTemporal(), LD->isInvariant(),
4951 LD->getAlignment());
4952 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4953 // have been legalized as a BITCAST from v4i32.
4954 if (N->getOpcode() == ISD::BITCAST) {
4955 SDNode *BVN = N->getOperand(0).getNode();
4956 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4957 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4958 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4959 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4960 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4962 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4963 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4964 EVT VT = N->getValueType(0);
4965 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4966 unsigned NumElts = VT.getVectorNumElements();
4967 MVT TruncVT = MVT::getIntegerVT(EltSize);
4968 SmallVector<SDValue, 8> Ops;
4969 for (unsigned i = 0; i != NumElts; ++i) {
4970 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4971 const APInt &CInt = C->getAPIntValue();
4972 // Element types smaller than 32 bits are not legal, so use i32 elements.
4973 // The values are implicitly truncated so sext vs. zext doesn't matter.
4974 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4976 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4977 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4980 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4981 unsigned Opcode = N->getOpcode();
4982 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4983 SDNode *N0 = N->getOperand(0).getNode();
4984 SDNode *N1 = N->getOperand(1).getNode();
4985 return N0->hasOneUse() && N1->hasOneUse() &&
4986 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4991 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4992 unsigned Opcode = N->getOpcode();
4993 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4994 SDNode *N0 = N->getOperand(0).getNode();
4995 SDNode *N1 = N->getOperand(1).getNode();
4996 return N0->hasOneUse() && N1->hasOneUse() &&
4997 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5002 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5003 // Multiplications are only custom-lowered for 128-bit vectors so that
5004 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5005 EVT VT = Op.getValueType();
5006 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
5007 SDNode *N0 = Op.getOperand(0).getNode();
5008 SDNode *N1 = Op.getOperand(1).getNode();
5009 unsigned NewOpc = 0;
5011 bool isN0SExt = isSignExtended(N0, DAG);
5012 bool isN1SExt = isSignExtended(N1, DAG);
5013 if (isN0SExt && isN1SExt)
5014 NewOpc = ARMISD::VMULLs;
5016 bool isN0ZExt = isZeroExtended(N0, DAG);
5017 bool isN1ZExt = isZeroExtended(N1, DAG);
5018 if (isN0ZExt && isN1ZExt)
5019 NewOpc = ARMISD::VMULLu;
5020 else if (isN1SExt || isN1ZExt) {
5021 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5022 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5023 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5024 NewOpc = ARMISD::VMULLs;
5026 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5027 NewOpc = ARMISD::VMULLu;
5029 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5031 NewOpc = ARMISD::VMULLu;
5037 if (VT == MVT::v2i64)
5038 // Fall through to expand this. It is not legal.
5041 // Other vector multiplications are legal.
5046 // Legalize to a VMULL instruction.
5047 DebugLoc DL = Op.getDebugLoc();
5049 SDValue Op1 = SkipExtension(N1, DAG);
5051 Op0 = SkipExtension(N0, DAG);
5052 assert(Op0.getValueType().is64BitVector() &&
5053 Op1.getValueType().is64BitVector() &&
5054 "unexpected types for extended operands to VMULL");
5055 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5058 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5059 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5066 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
5067 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
5068 EVT Op1VT = Op1.getValueType();
5069 return DAG.getNode(N0->getOpcode(), DL, VT,
5070 DAG.getNode(NewOpc, DL, VT,
5071 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5072 DAG.getNode(NewOpc, DL, VT,
5073 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5077 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5079 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5080 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5081 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5082 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5083 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5084 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5085 // Get reciprocal estimate.
5086 // float4 recip = vrecpeq_f32(yf);
5087 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5088 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5089 // Because char has a smaller range than uchar, we can actually get away
5090 // without any newton steps. This requires that we use a weird bias
5091 // of 0xb000, however (again, this has been exhaustively tested).
5092 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5093 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5094 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5095 Y = DAG.getConstant(0xb000, MVT::i32);
5096 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5097 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5098 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5099 // Convert back to short.
5100 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5101 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5106 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5108 // Convert to float.
5109 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5110 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5111 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5112 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5113 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5114 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5116 // Use reciprocal estimate and one refinement step.
5117 // float4 recip = vrecpeq_f32(yf);
5118 // recip *= vrecpsq_f32(yf, recip);
5119 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5120 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5121 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5122 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5124 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5125 // Because short has a smaller range than ushort, we can actually get away
5126 // with only a single newton step. This requires that we use a weird bias
5127 // of 89, however (again, this has been exhaustively tested).
5128 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5129 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5130 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5131 N1 = DAG.getConstant(0x89, MVT::i32);
5132 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5133 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5134 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5135 // Convert back to integer and return.
5136 // return vmovn_s32(vcvt_s32_f32(result));
5137 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5138 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5142 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5143 EVT VT = Op.getValueType();
5144 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5145 "unexpected type for custom-lowering ISD::SDIV");
5147 DebugLoc dl = Op.getDebugLoc();
5148 SDValue N0 = Op.getOperand(0);
5149 SDValue N1 = Op.getOperand(1);
5152 if (VT == MVT::v8i8) {
5153 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5154 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5156 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5157 DAG.getIntPtrConstant(4));
5158 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5159 DAG.getIntPtrConstant(4));
5160 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5161 DAG.getIntPtrConstant(0));
5162 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5163 DAG.getIntPtrConstant(0));
5165 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5166 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5168 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5169 N0 = LowerCONCAT_VECTORS(N0, DAG);
5171 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5174 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5177 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5178 EVT VT = Op.getValueType();
5179 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5180 "unexpected type for custom-lowering ISD::UDIV");
5182 DebugLoc dl = Op.getDebugLoc();
5183 SDValue N0 = Op.getOperand(0);
5184 SDValue N1 = Op.getOperand(1);
5187 if (VT == MVT::v8i8) {
5188 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5189 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5191 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5192 DAG.getIntPtrConstant(4));
5193 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5194 DAG.getIntPtrConstant(4));
5195 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5196 DAG.getIntPtrConstant(0));
5197 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5198 DAG.getIntPtrConstant(0));
5200 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5201 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5203 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5204 N0 = LowerCONCAT_VECTORS(N0, DAG);
5206 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5207 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5212 // v4i16 sdiv ... Convert to float.
5213 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5214 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5215 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5216 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5217 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5218 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5220 // Use reciprocal estimate and two refinement steps.
5221 // float4 recip = vrecpeq_f32(yf);
5222 // recip *= vrecpsq_f32(yf, recip);
5223 // recip *= vrecpsq_f32(yf, recip);
5224 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5225 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
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 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5231 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5233 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5234 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5235 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5236 // and that it will never cause us to return an answer too large).
5237 // float4 result = as_float4(as_int4(xf*recip) + 2);
5238 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5239 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5240 N1 = DAG.getConstant(2, MVT::i32);
5241 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5242 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5243 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5244 // Convert back to integer and return.
5245 // return vmovn_u32(vcvt_s32_f32(result));
5246 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5247 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5251 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5252 EVT VT = Op.getNode()->getValueType(0);
5253 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5256 bool ExtraOp = false;
5257 switch (Op.getOpcode()) {
5258 default: llvm_unreachable("Invalid code");
5259 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5260 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5261 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5262 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5266 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5268 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5269 Op.getOperand(1), Op.getOperand(2));
5272 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5273 // Monotonic load/store is legal for all targets
5274 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5277 // Aquire/Release load/store is not legal for targets without a
5278 // dmb or equivalent available.
5284 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5285 SelectionDAG &DAG, unsigned NewOp) {
5286 DebugLoc dl = Node->getDebugLoc();
5287 assert (Node->getValueType(0) == MVT::i64 &&
5288 "Only know how to expand i64 atomics");
5290 SmallVector<SDValue, 6> Ops;
5291 Ops.push_back(Node->getOperand(0)); // Chain
5292 Ops.push_back(Node->getOperand(1)); // Ptr
5294 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5295 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5296 // High part of Val1
5297 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5298 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5299 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5300 // High part of Val1
5301 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5302 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5303 // High part of Val2
5304 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5305 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5307 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5309 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5310 cast<MemSDNode>(Node)->getMemOperand());
5311 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5312 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5313 Results.push_back(Result.getValue(2));
5316 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5317 switch (Op.getOpcode()) {
5318 default: llvm_unreachable("Don't know how to custom lower this!");
5319 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5320 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5321 case ISD::GlobalAddress:
5322 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5323 LowerGlobalAddressELF(Op, DAG);
5324 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5325 case ISD::SELECT: return LowerSELECT(Op, DAG);
5326 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5327 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5328 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5329 case ISD::VASTART: return LowerVASTART(Op, DAG);
5330 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5331 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5332 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5333 case ISD::SINT_TO_FP:
5334 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5335 case ISD::FP_TO_SINT:
5336 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5337 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5338 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5339 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5340 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5341 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5342 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5343 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5345 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5348 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5349 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5350 case ISD::SRL_PARTS:
5351 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5352 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5353 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5354 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5355 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5356 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5357 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5358 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5359 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5360 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5361 case ISD::MUL: return LowerMUL(Op, DAG);
5362 case ISD::SDIV: return LowerSDIV(Op, DAG);
5363 case ISD::UDIV: return LowerUDIV(Op, DAG);
5367 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5368 case ISD::ATOMIC_LOAD:
5369 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5373 /// ReplaceNodeResults - Replace the results of node with an illegal result
5374 /// type with new values built out of custom code.
5375 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5376 SmallVectorImpl<SDValue>&Results,
5377 SelectionDAG &DAG) const {
5379 switch (N->getOpcode()) {
5381 llvm_unreachable("Don't know how to custom expand this!");
5383 Res = ExpandBITCAST(N, DAG);
5387 Res = Expand64BitShift(N, DAG, Subtarget);
5389 case ISD::ATOMIC_LOAD_ADD:
5390 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5392 case ISD::ATOMIC_LOAD_AND:
5393 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5395 case ISD::ATOMIC_LOAD_NAND:
5396 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5398 case ISD::ATOMIC_LOAD_OR:
5399 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5401 case ISD::ATOMIC_LOAD_SUB:
5402 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5404 case ISD::ATOMIC_LOAD_XOR:
5405 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5407 case ISD::ATOMIC_SWAP:
5408 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5410 case ISD::ATOMIC_CMP_SWAP:
5411 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5413 case ISD::ATOMIC_LOAD_MIN:
5414 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
5416 case ISD::ATOMIC_LOAD_UMIN:
5417 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
5419 case ISD::ATOMIC_LOAD_MAX:
5420 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
5422 case ISD::ATOMIC_LOAD_UMAX:
5423 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
5427 Results.push_back(Res);
5430 //===----------------------------------------------------------------------===//
5431 // ARM Scheduler Hooks
5432 //===----------------------------------------------------------------------===//
5435 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5436 MachineBasicBlock *BB,
5437 unsigned Size) const {
5438 unsigned dest = MI->getOperand(0).getReg();
5439 unsigned ptr = MI->getOperand(1).getReg();
5440 unsigned oldval = MI->getOperand(2).getReg();
5441 unsigned newval = MI->getOperand(3).getReg();
5442 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5443 DebugLoc dl = MI->getDebugLoc();
5444 bool isThumb2 = Subtarget->isThumb2();
5446 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5447 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5448 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5449 (const TargetRegisterClass*)&ARM::GPRRegClass);
5452 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5453 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5454 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5457 unsigned ldrOpc, strOpc;
5459 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5461 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5462 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5465 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5466 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5469 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5470 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5474 MachineFunction *MF = BB->getParent();
5475 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5476 MachineFunction::iterator It = BB;
5477 ++It; // insert the new blocks after the current block
5479 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5480 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5481 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5482 MF->insert(It, loop1MBB);
5483 MF->insert(It, loop2MBB);
5484 MF->insert(It, exitMBB);
5486 // Transfer the remainder of BB and its successor edges to exitMBB.
5487 exitMBB->splice(exitMBB->begin(), BB,
5488 llvm::next(MachineBasicBlock::iterator(MI)),
5490 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5494 // fallthrough --> loop1MBB
5495 BB->addSuccessor(loop1MBB);
5498 // ldrex dest, [ptr]
5502 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5503 if (ldrOpc == ARM::t2LDREX)
5505 AddDefaultPred(MIB);
5506 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5507 .addReg(dest).addReg(oldval));
5508 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5509 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5510 BB->addSuccessor(loop2MBB);
5511 BB->addSuccessor(exitMBB);
5514 // strex scratch, newval, [ptr]
5518 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5519 if (strOpc == ARM::t2STREX)
5521 AddDefaultPred(MIB);
5522 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5523 .addReg(scratch).addImm(0));
5524 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5525 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5526 BB->addSuccessor(loop1MBB);
5527 BB->addSuccessor(exitMBB);
5533 MI->eraseFromParent(); // The instruction is gone now.
5539 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5540 unsigned Size, unsigned BinOpcode) const {
5541 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5542 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5544 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5545 MachineFunction *MF = BB->getParent();
5546 MachineFunction::iterator It = BB;
5549 unsigned dest = MI->getOperand(0).getReg();
5550 unsigned ptr = MI->getOperand(1).getReg();
5551 unsigned incr = MI->getOperand(2).getReg();
5552 DebugLoc dl = MI->getDebugLoc();
5553 bool isThumb2 = Subtarget->isThumb2();
5555 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5557 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5558 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5561 unsigned ldrOpc, strOpc;
5563 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5565 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5566 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5569 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5570 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5573 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5574 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5578 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5579 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5580 MF->insert(It, loopMBB);
5581 MF->insert(It, exitMBB);
5583 // Transfer the remainder of BB and its successor edges to exitMBB.
5584 exitMBB->splice(exitMBB->begin(), BB,
5585 llvm::next(MachineBasicBlock::iterator(MI)),
5587 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5589 const TargetRegisterClass *TRC = isThumb2 ?
5590 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5591 (const TargetRegisterClass*)&ARM::GPRRegClass;
5592 unsigned scratch = MRI.createVirtualRegister(TRC);
5593 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5597 // fallthrough --> loopMBB
5598 BB->addSuccessor(loopMBB);
5602 // <binop> scratch2, dest, incr
5603 // strex scratch, scratch2, ptr
5606 // fallthrough --> exitMBB
5608 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5609 if (ldrOpc == ARM::t2LDREX)
5611 AddDefaultPred(MIB);
5613 // operand order needs to go the other way for NAND
5614 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5615 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5616 addReg(incr).addReg(dest)).addReg(0);
5618 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5619 addReg(dest).addReg(incr)).addReg(0);
5622 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5623 if (strOpc == ARM::t2STREX)
5625 AddDefaultPred(MIB);
5626 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5627 .addReg(scratch).addImm(0));
5628 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5629 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5631 BB->addSuccessor(loopMBB);
5632 BB->addSuccessor(exitMBB);
5638 MI->eraseFromParent(); // The instruction is gone now.
5644 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5645 MachineBasicBlock *BB,
5648 ARMCC::CondCodes Cond) const {
5649 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5651 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5652 MachineFunction *MF = BB->getParent();
5653 MachineFunction::iterator It = BB;
5656 unsigned dest = MI->getOperand(0).getReg();
5657 unsigned ptr = MI->getOperand(1).getReg();
5658 unsigned incr = MI->getOperand(2).getReg();
5659 unsigned oldval = dest;
5660 DebugLoc dl = MI->getDebugLoc();
5661 bool isThumb2 = Subtarget->isThumb2();
5663 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5665 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5666 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5669 unsigned ldrOpc, strOpc, extendOpc;
5671 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5673 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5674 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5675 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5678 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5679 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5680 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5683 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5684 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5689 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5690 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5691 MF->insert(It, loopMBB);
5692 MF->insert(It, exitMBB);
5694 // Transfer the remainder of BB and its successor edges to exitMBB.
5695 exitMBB->splice(exitMBB->begin(), BB,
5696 llvm::next(MachineBasicBlock::iterator(MI)),
5698 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5700 const TargetRegisterClass *TRC = isThumb2 ?
5701 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5702 (const TargetRegisterClass*)&ARM::GPRRegClass;
5703 unsigned scratch = MRI.createVirtualRegister(TRC);
5704 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5708 // fallthrough --> loopMBB
5709 BB->addSuccessor(loopMBB);
5713 // (sign extend dest, if required)
5715 // cmov.cond scratch2, incr, dest
5716 // strex scratch, scratch2, ptr
5719 // fallthrough --> exitMBB
5721 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5722 if (ldrOpc == ARM::t2LDREX)
5724 AddDefaultPred(MIB);
5726 // Sign extend the value, if necessary.
5727 if (signExtend && extendOpc) {
5728 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5729 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5734 // Build compare and cmov instructions.
5735 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5736 .addReg(oldval).addReg(incr));
5737 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5738 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5740 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5741 if (strOpc == ARM::t2STREX)
5743 AddDefaultPred(MIB);
5744 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5745 .addReg(scratch).addImm(0));
5746 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5747 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5749 BB->addSuccessor(loopMBB);
5750 BB->addSuccessor(exitMBB);
5756 MI->eraseFromParent(); // The instruction is gone now.
5762 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5763 unsigned Op1, unsigned Op2,
5764 bool NeedsCarry, bool IsCmpxchg,
5765 bool IsMinMax, ARMCC::CondCodes CC) const {
5766 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5767 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5769 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5770 MachineFunction *MF = BB->getParent();
5771 MachineFunction::iterator It = BB;
5774 unsigned destlo = MI->getOperand(0).getReg();
5775 unsigned desthi = MI->getOperand(1).getReg();
5776 unsigned ptr = MI->getOperand(2).getReg();
5777 unsigned vallo = MI->getOperand(3).getReg();
5778 unsigned valhi = MI->getOperand(4).getReg();
5779 DebugLoc dl = MI->getDebugLoc();
5780 bool isThumb2 = Subtarget->isThumb2();
5782 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5784 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5785 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5786 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5789 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5790 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5792 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5793 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5794 if (IsCmpxchg || IsMinMax)
5795 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5797 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5798 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5800 MF->insert(It, loopMBB);
5801 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
5802 if (IsCmpxchg) MF->insert(It, cont2BB);
5803 MF->insert(It, exitMBB);
5805 // Transfer the remainder of BB and its successor edges to exitMBB.
5806 exitMBB->splice(exitMBB->begin(), BB,
5807 llvm::next(MachineBasicBlock::iterator(MI)),
5809 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5811 const TargetRegisterClass *TRC = isThumb2 ?
5812 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5813 (const TargetRegisterClass*)&ARM::GPRRegClass;
5814 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5818 // fallthrough --> loopMBB
5819 BB->addSuccessor(loopMBB);
5822 // ldrexd r2, r3, ptr
5823 // <binopa> r0, r2, incr
5824 // <binopb> r1, r3, incr
5825 // strexd storesuccess, r0, r1, ptr
5826 // cmp storesuccess, #0
5828 // fallthrough --> exitMBB
5830 // Note that the registers are explicitly specified because there is not any
5831 // way to force the register allocator to allocate a register pair.
5833 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5834 // need to properly enforce the restriction that the two output registers
5835 // for ldrexd must be different.
5838 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5839 unsigned GPRPair1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5842 //We need an extra double register for doing min/max.
5843 unsigned undef = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5844 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5845 GPRPair2 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5846 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), undef);
5847 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5850 .addImm(ARM::gsub_0);
5851 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair2)
5854 .addImm(ARM::gsub_1);
5857 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5858 .addReg(GPRPair0, RegState::Define).addReg(ptr));
5859 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5860 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
5861 .addReg(GPRPair0, 0, ARM::gsub_0);
5862 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
5863 .addReg(GPRPair0, 0, ARM::gsub_1);
5867 for (unsigned i = 0; i < 2; i++) {
5868 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5870 .addReg(i == 0 ? destlo : desthi)
5871 .addReg(i == 0 ? vallo : valhi));
5872 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5873 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5874 BB->addSuccessor(exitMBB);
5875 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5876 BB = (i == 0 ? contBB : cont2BB);
5879 // Copy to physregs for strexd
5880 unsigned setlo = MI->getOperand(5).getReg();
5881 unsigned sethi = MI->getOperand(6).getReg();
5882 unsigned undef = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5883 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5884 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), undef);
5885 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5888 .addImm(ARM::gsub_0);
5889 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5892 .addImm(ARM::gsub_1);
5894 // Perform binary operation
5895 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
5896 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
5897 .addReg(destlo).addReg(vallo))
5898 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5899 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
5900 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
5901 .addReg(desthi).addReg(valhi))
5902 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
5904 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5905 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
5906 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5907 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5910 .addImm(ARM::gsub_0);
5911 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5914 .addImm(ARM::gsub_1);
5916 // Copy to physregs for strexd
5917 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5918 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5919 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
5920 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
5923 .addImm(ARM::gsub_0);
5924 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), GPRPair1)
5927 .addImm(ARM::gsub_1);
5929 unsigned GPRPairStore = GPRPair1;
5931 // Compare and branch to exit block.
5932 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5933 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
5934 BB->addSuccessor(exitMBB);
5935 BB->addSuccessor(contBB);
5937 GPRPairStore = GPRPair2;
5941 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5942 .addReg(GPRPairStore).addReg(ptr));
5944 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5945 .addReg(storesuccess).addImm(0));
5946 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5947 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5949 BB->addSuccessor(loopMBB);
5950 BB->addSuccessor(exitMBB);
5956 MI->eraseFromParent(); // The instruction is gone now.
5961 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5962 /// registers the function context.
5963 void ARMTargetLowering::
5964 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5965 MachineBasicBlock *DispatchBB, int FI) const {
5966 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5967 DebugLoc dl = MI->getDebugLoc();
5968 MachineFunction *MF = MBB->getParent();
5969 MachineRegisterInfo *MRI = &MF->getRegInfo();
5970 MachineConstantPool *MCP = MF->getConstantPool();
5971 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5972 const Function *F = MF->getFunction();
5974 bool isThumb = Subtarget->isThumb();
5975 bool isThumb2 = Subtarget->isThumb2();
5977 unsigned PCLabelId = AFI->createPICLabelUId();
5978 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5979 ARMConstantPoolValue *CPV =
5980 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5981 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5983 const TargetRegisterClass *TRC = isThumb ?
5984 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5985 (const TargetRegisterClass*)&ARM::GPRRegClass;
5987 // Grab constant pool and fixed stack memory operands.
5988 MachineMemOperand *CPMMO =
5989 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5990 MachineMemOperand::MOLoad, 4, 4);
5992 MachineMemOperand *FIMMOSt =
5993 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5994 MachineMemOperand::MOStore, 4, 4);
5996 // Load the address of the dispatch MBB into the jump buffer.
5998 // Incoming value: jbuf
5999 // ldr.n r5, LCPI1_1
6002 // str r5, [$jbuf, #+4] ; &jbuf[1]
6003 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6004 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6005 .addConstantPoolIndex(CPI)
6006 .addMemOperand(CPMMO));
6007 // Set the low bit because of thumb mode.
6008 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6010 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6011 .addReg(NewVReg1, RegState::Kill)
6013 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6014 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6015 .addReg(NewVReg2, RegState::Kill)
6017 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6018 .addReg(NewVReg3, RegState::Kill)
6020 .addImm(36) // &jbuf[1] :: pc
6021 .addMemOperand(FIMMOSt));
6022 } else if (isThumb) {
6023 // Incoming value: jbuf
6024 // ldr.n r1, LCPI1_4
6028 // add r2, $jbuf, #+4 ; &jbuf[1]
6030 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6031 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6032 .addConstantPoolIndex(CPI)
6033 .addMemOperand(CPMMO));
6034 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6035 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6036 .addReg(NewVReg1, RegState::Kill)
6038 // Set the low bit because of thumb mode.
6039 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6040 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6041 .addReg(ARM::CPSR, RegState::Define)
6043 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6044 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6045 .addReg(ARM::CPSR, RegState::Define)
6046 .addReg(NewVReg2, RegState::Kill)
6047 .addReg(NewVReg3, RegState::Kill));
6048 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6049 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6051 .addImm(36)); // &jbuf[1] :: pc
6052 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6053 .addReg(NewVReg4, RegState::Kill)
6054 .addReg(NewVReg5, RegState::Kill)
6056 .addMemOperand(FIMMOSt));
6058 // Incoming value: jbuf
6061 // str r1, [$jbuf, #+4] ; &jbuf[1]
6062 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6063 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6064 .addConstantPoolIndex(CPI)
6066 .addMemOperand(CPMMO));
6067 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6068 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6069 .addReg(NewVReg1, RegState::Kill)
6070 .addImm(PCLabelId));
6071 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6072 .addReg(NewVReg2, RegState::Kill)
6074 .addImm(36) // &jbuf[1] :: pc
6075 .addMemOperand(FIMMOSt));
6079 MachineBasicBlock *ARMTargetLowering::
6080 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6081 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6082 DebugLoc dl = MI->getDebugLoc();
6083 MachineFunction *MF = MBB->getParent();
6084 MachineRegisterInfo *MRI = &MF->getRegInfo();
6085 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6086 MachineFrameInfo *MFI = MF->getFrameInfo();
6087 int FI = MFI->getFunctionContextIndex();
6089 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6090 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6091 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6093 // Get a mapping of the call site numbers to all of the landing pads they're
6095 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6096 unsigned MaxCSNum = 0;
6097 MachineModuleInfo &MMI = MF->getMMI();
6098 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6100 if (!BB->isLandingPad()) continue;
6102 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6104 for (MachineBasicBlock::iterator
6105 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6106 if (!II->isEHLabel()) continue;
6108 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6109 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6111 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6112 for (SmallVectorImpl<unsigned>::iterator
6113 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6114 CSI != CSE; ++CSI) {
6115 CallSiteNumToLPad[*CSI].push_back(BB);
6116 MaxCSNum = std::max(MaxCSNum, *CSI);
6122 // Get an ordered list of the machine basic blocks for the jump table.
6123 std::vector<MachineBasicBlock*> LPadList;
6124 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6125 LPadList.reserve(CallSiteNumToLPad.size());
6126 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6127 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6128 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6129 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6130 LPadList.push_back(*II);
6131 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6135 assert(!LPadList.empty() &&
6136 "No landing pad destinations for the dispatch jump table!");
6138 // Create the jump table and associated information.
6139 MachineJumpTableInfo *JTI =
6140 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6141 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6142 unsigned UId = AFI->createJumpTableUId();
6144 // Create the MBBs for the dispatch code.
6146 // Shove the dispatch's address into the return slot in the function context.
6147 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6148 DispatchBB->setIsLandingPad();
6150 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6151 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6152 DispatchBB->addSuccessor(TrapBB);
6154 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6155 DispatchBB->addSuccessor(DispContBB);
6158 MF->insert(MF->end(), DispatchBB);
6159 MF->insert(MF->end(), DispContBB);
6160 MF->insert(MF->end(), TrapBB);
6162 // Insert code into the entry block that creates and registers the function
6164 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6166 MachineMemOperand *FIMMOLd =
6167 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6168 MachineMemOperand::MOLoad |
6169 MachineMemOperand::MOVolatile, 4, 4);
6171 MachineInstrBuilder MIB;
6172 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6174 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6175 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6177 // Add a register mask with no preserved registers. This results in all
6178 // registers being marked as clobbered.
6179 MIB.addRegMask(RI.getNoPreservedMask());
6181 unsigned NumLPads = LPadList.size();
6182 if (Subtarget->isThumb2()) {
6183 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6184 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6187 .addMemOperand(FIMMOLd));
6189 if (NumLPads < 256) {
6190 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6192 .addImm(LPadList.size()));
6194 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6195 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6196 .addImm(NumLPads & 0xFFFF));
6198 unsigned VReg2 = VReg1;
6199 if ((NumLPads & 0xFFFF0000) != 0) {
6200 VReg2 = MRI->createVirtualRegister(TRC);
6201 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6203 .addImm(NumLPads >> 16));
6206 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6211 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6216 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6217 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6218 .addJumpTableIndex(MJTI)
6221 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6224 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6225 .addReg(NewVReg3, RegState::Kill)
6227 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6229 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6230 .addReg(NewVReg4, RegState::Kill)
6232 .addJumpTableIndex(MJTI)
6234 } else if (Subtarget->isThumb()) {
6235 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6236 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6239 .addMemOperand(FIMMOLd));
6241 if (NumLPads < 256) {
6242 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6246 MachineConstantPool *ConstantPool = MF->getConstantPool();
6247 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6248 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6250 // MachineConstantPool wants an explicit alignment.
6251 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6253 Align = getDataLayout()->getTypeAllocSize(C->getType());
6254 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6256 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6257 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6258 .addReg(VReg1, RegState::Define)
6259 .addConstantPoolIndex(Idx));
6260 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6265 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6270 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6271 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6272 .addReg(ARM::CPSR, RegState::Define)
6276 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6277 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6278 .addJumpTableIndex(MJTI)
6281 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6282 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6283 .addReg(ARM::CPSR, RegState::Define)
6284 .addReg(NewVReg2, RegState::Kill)
6287 MachineMemOperand *JTMMOLd =
6288 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6289 MachineMemOperand::MOLoad, 4, 4);
6291 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6292 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6293 .addReg(NewVReg4, RegState::Kill)
6295 .addMemOperand(JTMMOLd));
6297 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6298 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6299 .addReg(ARM::CPSR, RegState::Define)
6300 .addReg(NewVReg5, RegState::Kill)
6303 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6304 .addReg(NewVReg6, RegState::Kill)
6305 .addJumpTableIndex(MJTI)
6308 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6309 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6312 .addMemOperand(FIMMOLd));
6314 if (NumLPads < 256) {
6315 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6318 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6319 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6320 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6321 .addImm(NumLPads & 0xFFFF));
6323 unsigned VReg2 = VReg1;
6324 if ((NumLPads & 0xFFFF0000) != 0) {
6325 VReg2 = MRI->createVirtualRegister(TRC);
6326 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6328 .addImm(NumLPads >> 16));
6331 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6335 MachineConstantPool *ConstantPool = MF->getConstantPool();
6336 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6337 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6339 // MachineConstantPool wants an explicit alignment.
6340 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6342 Align = getDataLayout()->getTypeAllocSize(C->getType());
6343 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6345 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6346 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6347 .addReg(VReg1, RegState::Define)
6348 .addConstantPoolIndex(Idx)
6350 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6352 .addReg(VReg1, RegState::Kill));
6355 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6360 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6362 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6364 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6365 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6366 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6367 .addJumpTableIndex(MJTI)
6370 MachineMemOperand *JTMMOLd =
6371 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6372 MachineMemOperand::MOLoad, 4, 4);
6373 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6375 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6376 .addReg(NewVReg3, RegState::Kill)
6379 .addMemOperand(JTMMOLd));
6381 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6382 .addReg(NewVReg5, RegState::Kill)
6384 .addJumpTableIndex(MJTI)
6388 // Add the jump table entries as successors to the MBB.
6389 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6390 for (std::vector<MachineBasicBlock*>::iterator
6391 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6392 MachineBasicBlock *CurMBB = *I;
6393 if (SeenMBBs.insert(CurMBB))
6394 DispContBB->addSuccessor(CurMBB);
6397 // N.B. the order the invoke BBs are processed in doesn't matter here.
6398 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6399 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6400 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6401 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6402 MachineBasicBlock *BB = *I;
6404 // Remove the landing pad successor from the invoke block and replace it
6405 // with the new dispatch block.
6406 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6408 while (!Successors.empty()) {
6409 MachineBasicBlock *SMBB = Successors.pop_back_val();
6410 if (SMBB->isLandingPad()) {
6411 BB->removeSuccessor(SMBB);
6412 MBBLPads.push_back(SMBB);
6416 BB->addSuccessor(DispatchBB);
6418 // Find the invoke call and mark all of the callee-saved registers as
6419 // 'implicit defined' so that they're spilled. This prevents code from
6420 // moving instructions to before the EH block, where they will never be
6422 for (MachineBasicBlock::reverse_iterator
6423 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6424 if (!II->isCall()) continue;
6426 DenseMap<unsigned, bool> DefRegs;
6427 for (MachineInstr::mop_iterator
6428 OI = II->operands_begin(), OE = II->operands_end();
6430 if (!OI->isReg()) continue;
6431 DefRegs[OI->getReg()] = true;
6434 MachineInstrBuilder MIB(&*II);
6436 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6437 unsigned Reg = SavedRegs[i];
6438 if (Subtarget->isThumb2() &&
6439 !ARM::tGPRRegClass.contains(Reg) &&
6440 !ARM::hGPRRegClass.contains(Reg))
6442 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6444 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6447 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6454 // Mark all former landing pads as non-landing pads. The dispatch is the only
6456 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6457 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6458 (*I)->setIsLandingPad(false);
6460 // The instruction is gone now.
6461 MI->eraseFromParent();
6467 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6468 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6469 E = MBB->succ_end(); I != E; ++I)
6472 llvm_unreachable("Expecting a BB with two successors!");
6475 MachineBasicBlock *ARMTargetLowering::
6476 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6477 // This pseudo instruction has 3 operands: dst, src, size
6478 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6479 // Otherwise, we will generate unrolled scalar copies.
6480 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6481 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6482 MachineFunction::iterator It = BB;
6485 unsigned dest = MI->getOperand(0).getReg();
6486 unsigned src = MI->getOperand(1).getReg();
6487 unsigned SizeVal = MI->getOperand(2).getImm();
6488 unsigned Align = MI->getOperand(3).getImm();
6489 DebugLoc dl = MI->getDebugLoc();
6491 bool isThumb2 = Subtarget->isThumb2();
6492 MachineFunction *MF = BB->getParent();
6493 MachineRegisterInfo &MRI = MF->getRegInfo();
6494 unsigned ldrOpc, strOpc, UnitSize = 0;
6496 const TargetRegisterClass *TRC = isThumb2 ?
6497 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6498 (const TargetRegisterClass*)&ARM::GPRRegClass;
6499 const TargetRegisterClass *TRC_Vec = 0;
6502 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6503 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6505 } else if (Align & 2) {
6506 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6507 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6510 // Check whether we can use NEON instructions.
6511 if (!MF->getFunction()->getFnAttributes().
6512 hasAttribute(Attributes::NoImplicitFloat) &&
6513 Subtarget->hasNEON()) {
6514 if ((Align % 16 == 0) && SizeVal >= 16) {
6515 ldrOpc = ARM::VLD1q32wb_fixed;
6516 strOpc = ARM::VST1q32wb_fixed;
6518 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6520 else if ((Align % 8 == 0) && SizeVal >= 8) {
6521 ldrOpc = ARM::VLD1d32wb_fixed;
6522 strOpc = ARM::VST1d32wb_fixed;
6524 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6527 // Can't use NEON instructions.
6528 if (UnitSize == 0) {
6529 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6530 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6535 unsigned BytesLeft = SizeVal % UnitSize;
6536 unsigned LoopSize = SizeVal - BytesLeft;
6538 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6539 // Use LDR and STR to copy.
6540 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6541 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6542 unsigned srcIn = src;
6543 unsigned destIn = dest;
6544 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6545 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6546 unsigned srcOut = MRI.createVirtualRegister(TRC);
6547 unsigned destOut = MRI.createVirtualRegister(TRC);
6548 if (UnitSize >= 8) {
6549 AddDefaultPred(BuildMI(*BB, MI, dl,
6550 TII->get(ldrOpc), scratch)
6551 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6553 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6554 .addReg(destIn).addImm(0).addReg(scratch));
6555 } else if (isThumb2) {
6556 AddDefaultPred(BuildMI(*BB, MI, dl,
6557 TII->get(ldrOpc), scratch)
6558 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6560 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6561 .addReg(scratch).addReg(destIn)
6564 AddDefaultPred(BuildMI(*BB, MI, dl,
6565 TII->get(ldrOpc), scratch)
6566 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6569 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6570 .addReg(scratch).addReg(destIn)
6571 .addReg(0).addImm(UnitSize));
6577 // Handle the leftover bytes with LDRB and STRB.
6578 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6579 // [destOut] = STRB_POST(scratch, destIn, 1)
6580 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6581 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6582 for (unsigned i = 0; i < BytesLeft; i++) {
6583 unsigned scratch = MRI.createVirtualRegister(TRC);
6584 unsigned srcOut = MRI.createVirtualRegister(TRC);
6585 unsigned destOut = MRI.createVirtualRegister(TRC);
6587 AddDefaultPred(BuildMI(*BB, MI, dl,
6588 TII->get(ldrOpc),scratch)
6589 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6591 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6592 .addReg(scratch).addReg(destIn)
6593 .addReg(0).addImm(1));
6595 AddDefaultPred(BuildMI(*BB, MI, dl,
6596 TII->get(ldrOpc),scratch)
6597 .addReg(srcOut, RegState::Define).addReg(srcIn)
6598 .addReg(0).addImm(1));
6600 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6601 .addReg(scratch).addReg(destIn)
6602 .addReg(0).addImm(1));
6607 MI->eraseFromParent(); // The instruction is gone now.
6611 // Expand the pseudo op to a loop.
6614 // movw varEnd, # --> with thumb2
6616 // ldrcp varEnd, idx --> without thumb2
6617 // fallthrough --> loopMBB
6619 // PHI varPhi, varEnd, varLoop
6620 // PHI srcPhi, src, srcLoop
6621 // PHI destPhi, dst, destLoop
6622 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6623 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6624 // subs varLoop, varPhi, #UnitSize
6626 // fallthrough --> exitMBB
6628 // epilogue to handle left-over bytes
6629 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6630 // [destOut] = STRB_POST(scratch, destLoop, 1)
6631 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6632 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6633 MF->insert(It, loopMBB);
6634 MF->insert(It, exitMBB);
6636 // Transfer the remainder of BB and its successor edges to exitMBB.
6637 exitMBB->splice(exitMBB->begin(), BB,
6638 llvm::next(MachineBasicBlock::iterator(MI)),
6640 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6642 // Load an immediate to varEnd.
6643 unsigned varEnd = MRI.createVirtualRegister(TRC);
6645 unsigned VReg1 = varEnd;
6646 if ((LoopSize & 0xFFFF0000) != 0)
6647 VReg1 = MRI.createVirtualRegister(TRC);
6648 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6649 .addImm(LoopSize & 0xFFFF));
6651 if ((LoopSize & 0xFFFF0000) != 0)
6652 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6654 .addImm(LoopSize >> 16));
6656 MachineConstantPool *ConstantPool = MF->getConstantPool();
6657 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6658 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6660 // MachineConstantPool wants an explicit alignment.
6661 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6663 Align = getDataLayout()->getTypeAllocSize(C->getType());
6664 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6666 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6667 .addReg(varEnd, RegState::Define)
6668 .addConstantPoolIndex(Idx)
6671 BB->addSuccessor(loopMBB);
6673 // Generate the loop body:
6674 // varPhi = PHI(varLoop, varEnd)
6675 // srcPhi = PHI(srcLoop, src)
6676 // destPhi = PHI(destLoop, dst)
6677 MachineBasicBlock *entryBB = BB;
6679 unsigned varLoop = MRI.createVirtualRegister(TRC);
6680 unsigned varPhi = MRI.createVirtualRegister(TRC);
6681 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6682 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6683 unsigned destLoop = MRI.createVirtualRegister(TRC);
6684 unsigned destPhi = MRI.createVirtualRegister(TRC);
6686 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6687 .addReg(varLoop).addMBB(loopMBB)
6688 .addReg(varEnd).addMBB(entryBB);
6689 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6690 .addReg(srcLoop).addMBB(loopMBB)
6691 .addReg(src).addMBB(entryBB);
6692 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6693 .addReg(destLoop).addMBB(loopMBB)
6694 .addReg(dest).addMBB(entryBB);
6696 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6697 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6698 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6699 if (UnitSize >= 8) {
6700 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6701 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6703 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6704 .addReg(destPhi).addImm(0).addReg(scratch));
6705 } else if (isThumb2) {
6706 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6707 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6709 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6710 .addReg(scratch).addReg(destPhi)
6713 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6714 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6717 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6718 .addReg(scratch).addReg(destPhi)
6719 .addReg(0).addImm(UnitSize));
6722 // Decrement loop variable by UnitSize.
6723 MachineInstrBuilder MIB = BuildMI(BB, dl,
6724 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6725 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6726 MIB->getOperand(5).setReg(ARM::CPSR);
6727 MIB->getOperand(5).setIsDef(true);
6729 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6730 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6732 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6733 BB->addSuccessor(loopMBB);
6734 BB->addSuccessor(exitMBB);
6736 // Add epilogue to handle BytesLeft.
6738 MachineInstr *StartOfExit = exitMBB->begin();
6739 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6740 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6742 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6743 // [destOut] = STRB_POST(scratch, destLoop, 1)
6744 unsigned srcIn = srcLoop;
6745 unsigned destIn = destLoop;
6746 for (unsigned i = 0; i < BytesLeft; i++) {
6747 unsigned scratch = MRI.createVirtualRegister(TRC);
6748 unsigned srcOut = MRI.createVirtualRegister(TRC);
6749 unsigned destOut = MRI.createVirtualRegister(TRC);
6751 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6752 TII->get(ldrOpc),scratch)
6753 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6755 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6756 .addReg(scratch).addReg(destIn)
6759 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6760 TII->get(ldrOpc),scratch)
6761 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6763 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6764 .addReg(scratch).addReg(destIn)
6765 .addReg(0).addImm(1));
6771 MI->eraseFromParent(); // The instruction is gone now.
6776 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6777 MachineBasicBlock *BB) const {
6778 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6779 DebugLoc dl = MI->getDebugLoc();
6780 bool isThumb2 = Subtarget->isThumb2();
6781 switch (MI->getOpcode()) {
6784 llvm_unreachable("Unexpected instr type to insert");
6786 // The Thumb2 pre-indexed stores have the same MI operands, they just
6787 // define them differently in the .td files from the isel patterns, so
6788 // they need pseudos.
6789 case ARM::t2STR_preidx:
6790 MI->setDesc(TII->get(ARM::t2STR_PRE));
6792 case ARM::t2STRB_preidx:
6793 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6795 case ARM::t2STRH_preidx:
6796 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6799 case ARM::STRi_preidx:
6800 case ARM::STRBi_preidx: {
6801 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6802 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6803 // Decode the offset.
6804 unsigned Offset = MI->getOperand(4).getImm();
6805 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6806 Offset = ARM_AM::getAM2Offset(Offset);
6810 MachineMemOperand *MMO = *MI->memoperands_begin();
6811 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6812 .addOperand(MI->getOperand(0)) // Rn_wb
6813 .addOperand(MI->getOperand(1)) // Rt
6814 .addOperand(MI->getOperand(2)) // Rn
6815 .addImm(Offset) // offset (skip GPR==zero_reg)
6816 .addOperand(MI->getOperand(5)) // pred
6817 .addOperand(MI->getOperand(6))
6818 .addMemOperand(MMO);
6819 MI->eraseFromParent();
6822 case ARM::STRr_preidx:
6823 case ARM::STRBr_preidx:
6824 case ARM::STRH_preidx: {
6826 switch (MI->getOpcode()) {
6827 default: llvm_unreachable("unexpected opcode!");
6828 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6829 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6830 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6832 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6833 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6834 MIB.addOperand(MI->getOperand(i));
6835 MI->eraseFromParent();
6838 case ARM::ATOMIC_LOAD_ADD_I8:
6839 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6840 case ARM::ATOMIC_LOAD_ADD_I16:
6841 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6842 case ARM::ATOMIC_LOAD_ADD_I32:
6843 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6845 case ARM::ATOMIC_LOAD_AND_I8:
6846 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6847 case ARM::ATOMIC_LOAD_AND_I16:
6848 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6849 case ARM::ATOMIC_LOAD_AND_I32:
6850 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6852 case ARM::ATOMIC_LOAD_OR_I8:
6853 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6854 case ARM::ATOMIC_LOAD_OR_I16:
6855 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6856 case ARM::ATOMIC_LOAD_OR_I32:
6857 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6859 case ARM::ATOMIC_LOAD_XOR_I8:
6860 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6861 case ARM::ATOMIC_LOAD_XOR_I16:
6862 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6863 case ARM::ATOMIC_LOAD_XOR_I32:
6864 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6866 case ARM::ATOMIC_LOAD_NAND_I8:
6867 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6868 case ARM::ATOMIC_LOAD_NAND_I16:
6869 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6870 case ARM::ATOMIC_LOAD_NAND_I32:
6871 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6873 case ARM::ATOMIC_LOAD_SUB_I8:
6874 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6875 case ARM::ATOMIC_LOAD_SUB_I16:
6876 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6877 case ARM::ATOMIC_LOAD_SUB_I32:
6878 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6880 case ARM::ATOMIC_LOAD_MIN_I8:
6881 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6882 case ARM::ATOMIC_LOAD_MIN_I16:
6883 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6884 case ARM::ATOMIC_LOAD_MIN_I32:
6885 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6887 case ARM::ATOMIC_LOAD_MAX_I8:
6888 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6889 case ARM::ATOMIC_LOAD_MAX_I16:
6890 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6891 case ARM::ATOMIC_LOAD_MAX_I32:
6892 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6894 case ARM::ATOMIC_LOAD_UMIN_I8:
6895 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6896 case ARM::ATOMIC_LOAD_UMIN_I16:
6897 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6898 case ARM::ATOMIC_LOAD_UMIN_I32:
6899 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6901 case ARM::ATOMIC_LOAD_UMAX_I8:
6902 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6903 case ARM::ATOMIC_LOAD_UMAX_I16:
6904 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6905 case ARM::ATOMIC_LOAD_UMAX_I32:
6906 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6908 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6909 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6910 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6912 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6913 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6914 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6917 case ARM::ATOMADD6432:
6918 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6919 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6920 /*NeedsCarry*/ true);
6921 case ARM::ATOMSUB6432:
6922 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6923 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6924 /*NeedsCarry*/ true);
6925 case ARM::ATOMOR6432:
6926 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6927 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6928 case ARM::ATOMXOR6432:
6929 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6930 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6931 case ARM::ATOMAND6432:
6932 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6933 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6934 case ARM::ATOMSWAP6432:
6935 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6936 case ARM::ATOMCMPXCHG6432:
6937 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6938 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6939 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6940 case ARM::ATOMMIN6432:
6941 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6942 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6943 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
6944 /*IsMinMax*/ true, ARMCC::LE);
6945 case ARM::ATOMMAX6432:
6946 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6947 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6948 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
6949 /*IsMinMax*/ true, ARMCC::GE);
6950 case ARM::ATOMUMIN6432:
6951 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6952 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6953 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
6954 /*IsMinMax*/ true, ARMCC::LS);
6955 case ARM::ATOMUMAX6432:
6956 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6957 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6958 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
6959 /*IsMinMax*/ true, ARMCC::HS);
6961 case ARM::tMOVCCr_pseudo: {
6962 // To "insert" a SELECT_CC instruction, we actually have to insert the
6963 // diamond control-flow pattern. The incoming instruction knows the
6964 // destination vreg to set, the condition code register to branch on, the
6965 // true/false values to select between, and a branch opcode to use.
6966 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6967 MachineFunction::iterator It = BB;
6973 // cmpTY ccX, r1, r2
6975 // fallthrough --> copy0MBB
6976 MachineBasicBlock *thisMBB = BB;
6977 MachineFunction *F = BB->getParent();
6978 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6979 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6980 F->insert(It, copy0MBB);
6981 F->insert(It, sinkMBB);
6983 // Transfer the remainder of BB and its successor edges to sinkMBB.
6984 sinkMBB->splice(sinkMBB->begin(), BB,
6985 llvm::next(MachineBasicBlock::iterator(MI)),
6987 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6989 BB->addSuccessor(copy0MBB);
6990 BB->addSuccessor(sinkMBB);
6992 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6993 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6996 // %FalseValue = ...
6997 // # fallthrough to sinkMBB
7000 // Update machine-CFG edges
7001 BB->addSuccessor(sinkMBB);
7004 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7007 BuildMI(*BB, BB->begin(), dl,
7008 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7009 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7010 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7012 MI->eraseFromParent(); // The pseudo instruction is gone now.
7017 case ARM::BCCZi64: {
7018 // If there is an unconditional branch to the other successor, remove it.
7019 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7021 // Compare both parts that make up the double comparison separately for
7023 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7025 unsigned LHS1 = MI->getOperand(1).getReg();
7026 unsigned LHS2 = MI->getOperand(2).getReg();
7028 AddDefaultPred(BuildMI(BB, dl,
7029 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7030 .addReg(LHS1).addImm(0));
7031 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7032 .addReg(LHS2).addImm(0)
7033 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7035 unsigned RHS1 = MI->getOperand(3).getReg();
7036 unsigned RHS2 = MI->getOperand(4).getReg();
7037 AddDefaultPred(BuildMI(BB, dl,
7038 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7039 .addReg(LHS1).addReg(RHS1));
7040 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7041 .addReg(LHS2).addReg(RHS2)
7042 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7045 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7046 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7047 if (MI->getOperand(0).getImm() == ARMCC::NE)
7048 std::swap(destMBB, exitMBB);
7050 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7051 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7053 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7055 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7057 MI->eraseFromParent(); // The pseudo instruction is gone now.
7061 case ARM::Int_eh_sjlj_setjmp:
7062 case ARM::Int_eh_sjlj_setjmp_nofp:
7063 case ARM::tInt_eh_sjlj_setjmp:
7064 case ARM::t2Int_eh_sjlj_setjmp:
7065 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7066 EmitSjLjDispatchBlock(MI, BB);
7071 // To insert an ABS instruction, we have to insert the
7072 // diamond control-flow pattern. The incoming instruction knows the
7073 // source vreg to test against 0, the destination vreg to set,
7074 // the condition code register to branch on, the
7075 // true/false values to select between, and a branch opcode to use.
7080 // BCC (branch to SinkBB if V0 >= 0)
7081 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7082 // SinkBB: V1 = PHI(V2, V3)
7083 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7084 MachineFunction::iterator BBI = BB;
7086 MachineFunction *Fn = BB->getParent();
7087 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7088 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7089 Fn->insert(BBI, RSBBB);
7090 Fn->insert(BBI, SinkBB);
7092 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7093 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7094 bool isThumb2 = Subtarget->isThumb2();
7095 MachineRegisterInfo &MRI = Fn->getRegInfo();
7096 // In Thumb mode S must not be specified if source register is the SP or
7097 // PC and if destination register is the SP, so restrict register class
7098 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7099 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7100 (const TargetRegisterClass*)&ARM::GPRRegClass);
7102 // Transfer the remainder of BB and its successor edges to sinkMBB.
7103 SinkBB->splice(SinkBB->begin(), BB,
7104 llvm::next(MachineBasicBlock::iterator(MI)),
7106 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7108 BB->addSuccessor(RSBBB);
7109 BB->addSuccessor(SinkBB);
7111 // fall through to SinkMBB
7112 RSBBB->addSuccessor(SinkBB);
7114 // insert a cmp at the end of BB
7115 AddDefaultPred(BuildMI(BB, dl,
7116 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7117 .addReg(ABSSrcReg).addImm(0));
7119 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7121 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7122 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7124 // insert rsbri in RSBBB
7125 // Note: BCC and rsbri will be converted into predicated rsbmi
7126 // by if-conversion pass
7127 BuildMI(*RSBBB, RSBBB->begin(), dl,
7128 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7129 .addReg(ABSSrcReg, RegState::Kill)
7130 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7132 // insert PHI in SinkBB,
7133 // reuse ABSDstReg to not change uses of ABS instruction
7134 BuildMI(*SinkBB, SinkBB->begin(), dl,
7135 TII->get(ARM::PHI), ABSDstReg)
7136 .addReg(NewRsbDstReg).addMBB(RSBBB)
7137 .addReg(ABSSrcReg).addMBB(BB);
7139 // remove ABS instruction
7140 MI->eraseFromParent();
7142 // return last added BB
7145 case ARM::COPY_STRUCT_BYVAL_I32:
7147 return EmitStructByval(MI, BB);
7151 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7152 SDNode *Node) const {
7153 if (!MI->hasPostISelHook()) {
7154 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7155 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7159 const MCInstrDesc *MCID = &MI->getDesc();
7160 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7161 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7162 // operand is still set to noreg. If needed, set the optional operand's
7163 // register to CPSR, and remove the redundant implicit def.
7165 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7167 // Rename pseudo opcodes.
7168 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7170 const ARMBaseInstrInfo *TII =
7171 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7172 MCID = &TII->get(NewOpc);
7174 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7175 "converted opcode should be the same except for cc_out");
7179 // Add the optional cc_out operand
7180 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7182 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7184 // Any ARM instruction that sets the 's' bit should specify an optional
7185 // "cc_out" operand in the last operand position.
7186 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7187 assert(!NewOpc && "Optional cc_out operand required");
7190 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7191 // since we already have an optional CPSR def.
7192 bool definesCPSR = false;
7193 bool deadCPSR = false;
7194 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7196 const MachineOperand &MO = MI->getOperand(i);
7197 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7201 MI->RemoveOperand(i);
7206 assert(!NewOpc && "Optional cc_out operand required");
7209 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7211 assert(!MI->getOperand(ccOutIdx).getReg() &&
7212 "expect uninitialized optional cc_out operand");
7216 // If this instruction was defined with an optional CPSR def and its dag node
7217 // had a live implicit CPSR def, then activate the optional CPSR def.
7218 MachineOperand &MO = MI->getOperand(ccOutIdx);
7219 MO.setReg(ARM::CPSR);
7223 //===----------------------------------------------------------------------===//
7224 // ARM Optimization Hooks
7225 //===----------------------------------------------------------------------===//
7227 // Helper function that checks if N is a null or all ones constant.
7228 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7229 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7232 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7235 // Return true if N is conditionally 0 or all ones.
7236 // Detects these expressions where cc is an i1 value:
7238 // (select cc 0, y) [AllOnes=0]
7239 // (select cc y, 0) [AllOnes=0]
7240 // (zext cc) [AllOnes=0]
7241 // (sext cc) [AllOnes=0/1]
7242 // (select cc -1, y) [AllOnes=1]
7243 // (select cc y, -1) [AllOnes=1]
7245 // Invert is set when N is the null/all ones constant when CC is false.
7246 // OtherOp is set to the alternative value of N.
7247 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7248 SDValue &CC, bool &Invert,
7250 SelectionDAG &DAG) {
7251 switch (N->getOpcode()) {
7252 default: return false;
7254 CC = N->getOperand(0);
7255 SDValue N1 = N->getOperand(1);
7256 SDValue N2 = N->getOperand(2);
7257 if (isZeroOrAllOnes(N1, AllOnes)) {
7262 if (isZeroOrAllOnes(N2, AllOnes)) {
7269 case ISD::ZERO_EXTEND:
7270 // (zext cc) can never be the all ones value.
7274 case ISD::SIGN_EXTEND: {
7275 EVT VT = N->getValueType(0);
7276 CC = N->getOperand(0);
7277 if (CC.getValueType() != MVT::i1)
7281 // When looking for an AllOnes constant, N is an sext, and the 'other'
7283 OtherOp = DAG.getConstant(0, VT);
7284 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7285 // When looking for a 0 constant, N can be zext or sext.
7286 OtherOp = DAG.getConstant(1, VT);
7288 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7294 // Combine a constant select operand into its use:
7296 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7297 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7298 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7299 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7300 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7302 // The transform is rejected if the select doesn't have a constant operand that
7303 // is null, or all ones when AllOnes is set.
7305 // Also recognize sext/zext from i1:
7307 // (add (zext cc), x) -> (select cc (add x, 1), x)
7308 // (add (sext cc), x) -> (select cc (add x, -1), x)
7310 // These transformations eventually create predicated instructions.
7312 // @param N The node to transform.
7313 // @param Slct The N operand that is a select.
7314 // @param OtherOp The other N operand (x above).
7315 // @param DCI Context.
7316 // @param AllOnes Require the select constant to be all ones instead of null.
7317 // @returns The new node, or SDValue() on failure.
7319 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7320 TargetLowering::DAGCombinerInfo &DCI,
7321 bool AllOnes = false) {
7322 SelectionDAG &DAG = DCI.DAG;
7323 EVT VT = N->getValueType(0);
7324 SDValue NonConstantVal;
7327 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7328 NonConstantVal, DAG))
7331 // Slct is now know to be the desired identity constant when CC is true.
7332 SDValue TrueVal = OtherOp;
7333 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7334 OtherOp, NonConstantVal);
7335 // Unless SwapSelectOps says CC should be false.
7337 std::swap(TrueVal, FalseVal);
7339 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7340 CCOp, TrueVal, FalseVal);
7343 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7345 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7346 TargetLowering::DAGCombinerInfo &DCI) {
7347 SDValue N0 = N->getOperand(0);
7348 SDValue N1 = N->getOperand(1);
7349 if (N0.getNode()->hasOneUse()) {
7350 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7351 if (Result.getNode())
7354 if (N1.getNode()->hasOneUse()) {
7355 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7356 if (Result.getNode())
7362 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7363 // (only after legalization).
7364 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7365 TargetLowering::DAGCombinerInfo &DCI,
7366 const ARMSubtarget *Subtarget) {
7368 // Only perform optimization if after legalize, and if NEON is available. We
7369 // also expected both operands to be BUILD_VECTORs.
7370 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7371 || N0.getOpcode() != ISD::BUILD_VECTOR
7372 || N1.getOpcode() != ISD::BUILD_VECTOR)
7375 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7376 EVT VT = N->getValueType(0);
7377 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7380 // Check that the vector operands are of the right form.
7381 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7382 // operands, where N is the size of the formed vector.
7383 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7384 // index such that we have a pair wise add pattern.
7386 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7387 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7389 SDValue Vec = N0->getOperand(0)->getOperand(0);
7390 SDNode *V = Vec.getNode();
7391 unsigned nextIndex = 0;
7393 // For each operands to the ADD which are BUILD_VECTORs,
7394 // check to see if each of their operands are an EXTRACT_VECTOR with
7395 // the same vector and appropriate index.
7396 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7397 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7398 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7400 SDValue ExtVec0 = N0->getOperand(i);
7401 SDValue ExtVec1 = N1->getOperand(i);
7403 // First operand is the vector, verify its the same.
7404 if (V != ExtVec0->getOperand(0).getNode() ||
7405 V != ExtVec1->getOperand(0).getNode())
7408 // Second is the constant, verify its correct.
7409 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7410 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7412 // For the constant, we want to see all the even or all the odd.
7413 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7414 || C1->getZExtValue() != nextIndex+1)
7423 // Create VPADDL node.
7424 SelectionDAG &DAG = DCI.DAG;
7425 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7427 // Build operand list.
7428 SmallVector<SDValue, 8> Ops;
7429 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7430 TLI.getPointerTy()));
7432 // Input is the vector.
7435 // Get widened type and narrowed type.
7437 unsigned numElem = VT.getVectorNumElements();
7438 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7439 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7440 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7441 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7443 llvm_unreachable("Invalid vector element type for padd optimization.");
7446 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7447 widenType, &Ops[0], Ops.size());
7448 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7451 static SDValue findMUL_LOHI(SDValue V) {
7452 if (V->getOpcode() == ISD::UMUL_LOHI ||
7453 V->getOpcode() == ISD::SMUL_LOHI)
7458 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7459 TargetLowering::DAGCombinerInfo &DCI,
7460 const ARMSubtarget *Subtarget) {
7462 if (Subtarget->isThumb1Only()) return SDValue();
7464 // Only perform the checks after legalize when the pattern is available.
7465 if (DCI.isBeforeLegalize()) return SDValue();
7467 // Look for multiply add opportunities.
7468 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7469 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7470 // a glue link from the first add to the second add.
7471 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7472 // a S/UMLAL instruction.
7475 // \ / \ [no multiline comment]
7481 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7482 SDValue AddcOp0 = AddcNode->getOperand(0);
7483 SDValue AddcOp1 = AddcNode->getOperand(1);
7485 // Check if the two operands are from the same mul_lohi node.
7486 if (AddcOp0.getNode() == AddcOp1.getNode())
7489 assert(AddcNode->getNumValues() == 2 &&
7490 AddcNode->getValueType(0) == MVT::i32 &&
7491 AddcNode->getValueType(1) == MVT::Glue &&
7492 "Expect ADDC with two result values: i32, glue");
7494 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7495 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7496 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7497 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7498 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7501 // Look for the glued ADDE.
7502 SDNode* AddeNode = AddcNode->getGluedUser();
7503 if (AddeNode == NULL)
7506 // Make sure it is really an ADDE.
7507 if (AddeNode->getOpcode() != ISD::ADDE)
7510 assert(AddeNode->getNumOperands() == 3 &&
7511 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7512 "ADDE node has the wrong inputs");
7514 // Check for the triangle shape.
7515 SDValue AddeOp0 = AddeNode->getOperand(0);
7516 SDValue AddeOp1 = AddeNode->getOperand(1);
7518 // Make sure that the ADDE operands are not coming from the same node.
7519 if (AddeOp0.getNode() == AddeOp1.getNode())
7522 // Find the MUL_LOHI node walking up ADDE's operands.
7523 bool IsLeftOperandMUL = false;
7524 SDValue MULOp = findMUL_LOHI(AddeOp0);
7525 if (MULOp == SDValue())
7526 MULOp = findMUL_LOHI(AddeOp1);
7528 IsLeftOperandMUL = true;
7529 if (MULOp == SDValue())
7532 // Figure out the right opcode.
7533 unsigned Opc = MULOp->getOpcode();
7534 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7536 // Figure out the high and low input values to the MLAL node.
7537 SDValue* HiMul = &MULOp;
7538 SDValue* HiAdd = NULL;
7539 SDValue* LoMul = NULL;
7540 SDValue* LowAdd = NULL;
7542 if (IsLeftOperandMUL)
7548 if (AddcOp0->getOpcode() == Opc) {
7552 if (AddcOp1->getOpcode() == Opc) {
7560 if (LoMul->getNode() != HiMul->getNode())
7563 // Create the merged node.
7564 SelectionDAG &DAG = DCI.DAG;
7566 // Build operand list.
7567 SmallVector<SDValue, 8> Ops;
7568 Ops.push_back(LoMul->getOperand(0));
7569 Ops.push_back(LoMul->getOperand(1));
7570 Ops.push_back(*LowAdd);
7571 Ops.push_back(*HiAdd);
7573 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7574 DAG.getVTList(MVT::i32, MVT::i32),
7575 &Ops[0], Ops.size());
7577 // Replace the ADDs' nodes uses by the MLA node's values.
7578 SDValue HiMLALResult(MLALNode.getNode(), 1);
7579 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7581 SDValue LoMLALResult(MLALNode.getNode(), 0);
7582 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7584 // Return original node to notify the driver to stop replacing.
7585 SDValue resNode(AddcNode, 0);
7589 /// PerformADDCCombine - Target-specific dag combine transform from
7590 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7591 static SDValue PerformADDCCombine(SDNode *N,
7592 TargetLowering::DAGCombinerInfo &DCI,
7593 const ARMSubtarget *Subtarget) {
7595 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7599 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7600 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7601 /// called with the default operands, and if that fails, with commuted
7603 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7604 TargetLowering::DAGCombinerInfo &DCI,
7605 const ARMSubtarget *Subtarget){
7607 // Attempt to create vpaddl for this add.
7608 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7609 if (Result.getNode())
7612 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7613 if (N0.getNode()->hasOneUse()) {
7614 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7615 if (Result.getNode()) return Result;
7620 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7622 static SDValue PerformADDCombine(SDNode *N,
7623 TargetLowering::DAGCombinerInfo &DCI,
7624 const ARMSubtarget *Subtarget) {
7625 SDValue N0 = N->getOperand(0);
7626 SDValue N1 = N->getOperand(1);
7628 // First try with the default operand order.
7629 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7630 if (Result.getNode())
7633 // If that didn't work, try again with the operands commuted.
7634 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7637 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7639 static SDValue PerformSUBCombine(SDNode *N,
7640 TargetLowering::DAGCombinerInfo &DCI) {
7641 SDValue N0 = N->getOperand(0);
7642 SDValue N1 = N->getOperand(1);
7644 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7645 if (N1.getNode()->hasOneUse()) {
7646 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7647 if (Result.getNode()) return Result;
7653 /// PerformVMULCombine
7654 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7655 /// special multiplier accumulator forwarding.
7661 static SDValue PerformVMULCombine(SDNode *N,
7662 TargetLowering::DAGCombinerInfo &DCI,
7663 const ARMSubtarget *Subtarget) {
7664 if (!Subtarget->hasVMLxForwarding())
7667 SelectionDAG &DAG = DCI.DAG;
7668 SDValue N0 = N->getOperand(0);
7669 SDValue N1 = N->getOperand(1);
7670 unsigned Opcode = N0.getOpcode();
7671 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7672 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7673 Opcode = N1.getOpcode();
7674 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7675 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7680 EVT VT = N->getValueType(0);
7681 DebugLoc DL = N->getDebugLoc();
7682 SDValue N00 = N0->getOperand(0);
7683 SDValue N01 = N0->getOperand(1);
7684 return DAG.getNode(Opcode, DL, VT,
7685 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7686 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7689 static SDValue PerformMULCombine(SDNode *N,
7690 TargetLowering::DAGCombinerInfo &DCI,
7691 const ARMSubtarget *Subtarget) {
7692 SelectionDAG &DAG = DCI.DAG;
7694 if (Subtarget->isThumb1Only())
7697 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7700 EVT VT = N->getValueType(0);
7701 if (VT.is64BitVector() || VT.is128BitVector())
7702 return PerformVMULCombine(N, DCI, Subtarget);
7706 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7710 int64_t MulAmt = C->getSExtValue();
7711 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7713 ShiftAmt = ShiftAmt & (32 - 1);
7714 SDValue V = N->getOperand(0);
7715 DebugLoc DL = N->getDebugLoc();
7718 MulAmt >>= ShiftAmt;
7721 if (isPowerOf2_32(MulAmt - 1)) {
7722 // (mul x, 2^N + 1) => (add (shl x, N), x)
7723 Res = DAG.getNode(ISD::ADD, DL, VT,
7725 DAG.getNode(ISD::SHL, DL, VT,
7727 DAG.getConstant(Log2_32(MulAmt - 1),
7729 } else if (isPowerOf2_32(MulAmt + 1)) {
7730 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7731 Res = DAG.getNode(ISD::SUB, DL, VT,
7732 DAG.getNode(ISD::SHL, DL, VT,
7734 DAG.getConstant(Log2_32(MulAmt + 1),
7740 uint64_t MulAmtAbs = -MulAmt;
7741 if (isPowerOf2_32(MulAmtAbs + 1)) {
7742 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7743 Res = DAG.getNode(ISD::SUB, DL, VT,
7745 DAG.getNode(ISD::SHL, DL, VT,
7747 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7749 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7750 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7751 Res = DAG.getNode(ISD::ADD, DL, VT,
7753 DAG.getNode(ISD::SHL, DL, VT,
7755 DAG.getConstant(Log2_32(MulAmtAbs-1),
7757 Res = DAG.getNode(ISD::SUB, DL, VT,
7758 DAG.getConstant(0, MVT::i32),Res);
7765 Res = DAG.getNode(ISD::SHL, DL, VT,
7766 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7768 // Do not add new nodes to DAG combiner worklist.
7769 DCI.CombineTo(N, Res, false);
7773 static SDValue PerformANDCombine(SDNode *N,
7774 TargetLowering::DAGCombinerInfo &DCI,
7775 const ARMSubtarget *Subtarget) {
7777 // Attempt to use immediate-form VBIC
7778 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7779 DebugLoc dl = N->getDebugLoc();
7780 EVT VT = N->getValueType(0);
7781 SelectionDAG &DAG = DCI.DAG;
7783 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7786 APInt SplatBits, SplatUndef;
7787 unsigned SplatBitSize;
7790 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7791 if (SplatBitSize <= 64) {
7793 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7794 SplatUndef.getZExtValue(), SplatBitSize,
7795 DAG, VbicVT, VT.is128BitVector(),
7797 if (Val.getNode()) {
7799 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7800 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7801 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7806 if (!Subtarget->isThumb1Only()) {
7807 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7808 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7809 if (Result.getNode())
7816 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7817 static SDValue PerformORCombine(SDNode *N,
7818 TargetLowering::DAGCombinerInfo &DCI,
7819 const ARMSubtarget *Subtarget) {
7820 // Attempt to use immediate-form VORR
7821 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7822 DebugLoc dl = N->getDebugLoc();
7823 EVT VT = N->getValueType(0);
7824 SelectionDAG &DAG = DCI.DAG;
7826 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7829 APInt SplatBits, SplatUndef;
7830 unsigned SplatBitSize;
7832 if (BVN && Subtarget->hasNEON() &&
7833 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7834 if (SplatBitSize <= 64) {
7836 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7837 SplatUndef.getZExtValue(), SplatBitSize,
7838 DAG, VorrVT, VT.is128BitVector(),
7840 if (Val.getNode()) {
7842 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7843 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7844 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7849 if (!Subtarget->isThumb1Only()) {
7850 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7851 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7852 if (Result.getNode())
7856 // The code below optimizes (or (and X, Y), Z).
7857 // The AND operand needs to have a single user to make these optimizations
7859 SDValue N0 = N->getOperand(0);
7860 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7862 SDValue N1 = N->getOperand(1);
7864 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7865 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7866 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7868 unsigned SplatBitSize;
7871 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7873 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7874 HasAnyUndefs) && !HasAnyUndefs) {
7875 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7877 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7878 HasAnyUndefs) && !HasAnyUndefs &&
7879 SplatBits0 == ~SplatBits1) {
7880 // Canonicalize the vector type to make instruction selection simpler.
7881 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7882 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7883 N0->getOperand(1), N0->getOperand(0),
7885 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7890 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7893 // BFI is only available on V6T2+
7894 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7897 DebugLoc DL = N->getDebugLoc();
7898 // 1) or (and A, mask), val => ARMbfi A, val, mask
7899 // iff (val & mask) == val
7901 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7902 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7903 // && mask == ~mask2
7904 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7905 // && ~mask == mask2
7906 // (i.e., copy a bitfield value into another bitfield of the same width)
7911 SDValue N00 = N0.getOperand(0);
7913 // The value and the mask need to be constants so we can verify this is
7914 // actually a bitfield set. If the mask is 0xffff, we can do better
7915 // via a movt instruction, so don't use BFI in that case.
7916 SDValue MaskOp = N0.getOperand(1);
7917 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7920 unsigned Mask = MaskC->getZExtValue();
7924 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7925 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7927 unsigned Val = N1C->getZExtValue();
7928 if ((Val & ~Mask) != Val)
7931 if (ARM::isBitFieldInvertedMask(Mask)) {
7932 Val >>= CountTrailingZeros_32(~Mask);
7934 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7935 DAG.getConstant(Val, MVT::i32),
7936 DAG.getConstant(Mask, MVT::i32));
7938 // Do not add new nodes to DAG combiner worklist.
7939 DCI.CombineTo(N, Res, false);
7942 } else if (N1.getOpcode() == ISD::AND) {
7943 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7944 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7947 unsigned Mask2 = N11C->getZExtValue();
7949 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7951 if (ARM::isBitFieldInvertedMask(Mask) &&
7953 // The pack halfword instruction works better for masks that fit it,
7954 // so use that when it's available.
7955 if (Subtarget->hasT2ExtractPack() &&
7956 (Mask == 0xffff || Mask == 0xffff0000))
7959 unsigned amt = CountTrailingZeros_32(Mask2);
7960 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7961 DAG.getConstant(amt, MVT::i32));
7962 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7963 DAG.getConstant(Mask, MVT::i32));
7964 // Do not add new nodes to DAG combiner worklist.
7965 DCI.CombineTo(N, Res, false);
7967 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7969 // The pack halfword instruction works better for masks that fit it,
7970 // so use that when it's available.
7971 if (Subtarget->hasT2ExtractPack() &&
7972 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7975 unsigned lsb = CountTrailingZeros_32(Mask);
7976 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7977 DAG.getConstant(lsb, MVT::i32));
7978 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7979 DAG.getConstant(Mask2, MVT::i32));
7980 // Do not add new nodes to DAG combiner worklist.
7981 DCI.CombineTo(N, Res, false);
7986 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7987 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7988 ARM::isBitFieldInvertedMask(~Mask)) {
7989 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7990 // where lsb(mask) == #shamt and masked bits of B are known zero.
7991 SDValue ShAmt = N00.getOperand(1);
7992 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7993 unsigned LSB = CountTrailingZeros_32(Mask);
7997 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7998 DAG.getConstant(~Mask, MVT::i32));
8000 // Do not add new nodes to DAG combiner worklist.
8001 DCI.CombineTo(N, Res, false);
8007 static SDValue PerformXORCombine(SDNode *N,
8008 TargetLowering::DAGCombinerInfo &DCI,
8009 const ARMSubtarget *Subtarget) {
8010 EVT VT = N->getValueType(0);
8011 SelectionDAG &DAG = DCI.DAG;
8013 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8016 if (!Subtarget->isThumb1Only()) {
8017 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8018 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8019 if (Result.getNode())
8026 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8027 /// the bits being cleared by the AND are not demanded by the BFI.
8028 static SDValue PerformBFICombine(SDNode *N,
8029 TargetLowering::DAGCombinerInfo &DCI) {
8030 SDValue N1 = N->getOperand(1);
8031 if (N1.getOpcode() == ISD::AND) {
8032 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8035 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8036 unsigned LSB = CountTrailingZeros_32(~InvMask);
8037 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
8038 unsigned Mask = (1 << Width)-1;
8039 unsigned Mask2 = N11C->getZExtValue();
8040 if ((Mask & (~Mask2)) == 0)
8041 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
8042 N->getOperand(0), N1.getOperand(0),
8048 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8049 /// ARMISD::VMOVRRD.
8050 static SDValue PerformVMOVRRDCombine(SDNode *N,
8051 TargetLowering::DAGCombinerInfo &DCI) {
8052 // vmovrrd(vmovdrr x, y) -> x,y
8053 SDValue InDouble = N->getOperand(0);
8054 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8055 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8057 // vmovrrd(load f64) -> (load i32), (load i32)
8058 SDNode *InNode = InDouble.getNode();
8059 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8060 InNode->getValueType(0) == MVT::f64 &&
8061 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8062 !cast<LoadSDNode>(InNode)->isVolatile()) {
8063 // TODO: Should this be done for non-FrameIndex operands?
8064 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8066 SelectionDAG &DAG = DCI.DAG;
8067 DebugLoc DL = LD->getDebugLoc();
8068 SDValue BasePtr = LD->getBasePtr();
8069 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8070 LD->getPointerInfo(), LD->isVolatile(),
8071 LD->isNonTemporal(), LD->isInvariant(),
8072 LD->getAlignment());
8074 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8075 DAG.getConstant(4, MVT::i32));
8076 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8077 LD->getPointerInfo(), LD->isVolatile(),
8078 LD->isNonTemporal(), LD->isInvariant(),
8079 std::min(4U, LD->getAlignment() / 2));
8081 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8082 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8083 DCI.RemoveFromWorklist(LD);
8091 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8092 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8093 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8094 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8095 SDValue Op0 = N->getOperand(0);
8096 SDValue Op1 = N->getOperand(1);
8097 if (Op0.getOpcode() == ISD::BITCAST)
8098 Op0 = Op0.getOperand(0);
8099 if (Op1.getOpcode() == ISD::BITCAST)
8100 Op1 = Op1.getOperand(0);
8101 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8102 Op0.getNode() == Op1.getNode() &&
8103 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8104 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
8105 N->getValueType(0), Op0.getOperand(0));
8109 /// PerformSTORECombine - Target-specific dag combine xforms for
8111 static SDValue PerformSTORECombine(SDNode *N,
8112 TargetLowering::DAGCombinerInfo &DCI) {
8113 StoreSDNode *St = cast<StoreSDNode>(N);
8114 if (St->isVolatile())
8117 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8118 // pack all of the elements in one place. Next, store to memory in fewer
8120 SDValue StVal = St->getValue();
8121 EVT VT = StVal.getValueType();
8122 if (St->isTruncatingStore() && VT.isVector()) {
8123 SelectionDAG &DAG = DCI.DAG;
8124 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8125 EVT StVT = St->getMemoryVT();
8126 unsigned NumElems = VT.getVectorNumElements();
8127 assert(StVT != VT && "Cannot truncate to the same type");
8128 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8129 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8131 // From, To sizes and ElemCount must be pow of two
8132 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8134 // We are going to use the original vector elt for storing.
8135 // Accumulated smaller vector elements must be a multiple of the store size.
8136 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8138 unsigned SizeRatio = FromEltSz / ToEltSz;
8139 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8141 // Create a type on which we perform the shuffle.
8142 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8143 NumElems*SizeRatio);
8144 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8146 DebugLoc DL = St->getDebugLoc();
8147 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8148 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8149 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8151 // Can't shuffle using an illegal type.
8152 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8154 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8155 DAG.getUNDEF(WideVec.getValueType()),
8157 // At this point all of the data is stored at the bottom of the
8158 // register. We now need to save it to mem.
8160 // Find the largest store unit
8161 MVT StoreType = MVT::i8;
8162 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8163 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8164 MVT Tp = (MVT::SimpleValueType)tp;
8165 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8168 // Didn't find a legal store type.
8169 if (!TLI.isTypeLegal(StoreType))
8172 // Bitcast the original vector into a vector of store-size units
8173 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8174 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8175 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8176 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8177 SmallVector<SDValue, 8> Chains;
8178 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8179 TLI.getPointerTy());
8180 SDValue BasePtr = St->getBasePtr();
8182 // Perform one or more big stores into memory.
8183 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8184 for (unsigned I = 0; I < E; I++) {
8185 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8186 StoreType, ShuffWide,
8187 DAG.getIntPtrConstant(I));
8188 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8189 St->getPointerInfo(), St->isVolatile(),
8190 St->isNonTemporal(), St->getAlignment());
8191 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8193 Chains.push_back(Ch);
8195 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8199 if (!ISD::isNormalStore(St))
8202 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8203 // ARM stores of arguments in the same cache line.
8204 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8205 StVal.getNode()->hasOneUse()) {
8206 SelectionDAG &DAG = DCI.DAG;
8207 DebugLoc DL = St->getDebugLoc();
8208 SDValue BasePtr = St->getBasePtr();
8209 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8210 StVal.getNode()->getOperand(0), BasePtr,
8211 St->getPointerInfo(), St->isVolatile(),
8212 St->isNonTemporal(), St->getAlignment());
8214 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8215 DAG.getConstant(4, MVT::i32));
8216 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8217 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8218 St->isNonTemporal(),
8219 std::min(4U, St->getAlignment() / 2));
8222 if (StVal.getValueType() != MVT::i64 ||
8223 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8226 // Bitcast an i64 store extracted from a vector to f64.
8227 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8228 SelectionDAG &DAG = DCI.DAG;
8229 DebugLoc dl = StVal.getDebugLoc();
8230 SDValue IntVec = StVal.getOperand(0);
8231 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8232 IntVec.getValueType().getVectorNumElements());
8233 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8234 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8235 Vec, StVal.getOperand(1));
8236 dl = N->getDebugLoc();
8237 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8238 // Make the DAGCombiner fold the bitcasts.
8239 DCI.AddToWorklist(Vec.getNode());
8240 DCI.AddToWorklist(ExtElt.getNode());
8241 DCI.AddToWorklist(V.getNode());
8242 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8243 St->getPointerInfo(), St->isVolatile(),
8244 St->isNonTemporal(), St->getAlignment(),
8248 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8249 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8250 /// i64 vector to have f64 elements, since the value can then be loaded
8251 /// directly into a VFP register.
8252 static bool hasNormalLoadOperand(SDNode *N) {
8253 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8254 for (unsigned i = 0; i < NumElts; ++i) {
8255 SDNode *Elt = N->getOperand(i).getNode();
8256 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8262 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8263 /// ISD::BUILD_VECTOR.
8264 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8265 TargetLowering::DAGCombinerInfo &DCI){
8266 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8267 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8268 // into a pair of GPRs, which is fine when the value is used as a scalar,
8269 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8270 SelectionDAG &DAG = DCI.DAG;
8271 if (N->getNumOperands() == 2) {
8272 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8277 // Load i64 elements as f64 values so that type legalization does not split
8278 // them up into i32 values.
8279 EVT VT = N->getValueType(0);
8280 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8282 DebugLoc dl = N->getDebugLoc();
8283 SmallVector<SDValue, 8> Ops;
8284 unsigned NumElts = VT.getVectorNumElements();
8285 for (unsigned i = 0; i < NumElts; ++i) {
8286 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8288 // Make the DAGCombiner fold the bitcast.
8289 DCI.AddToWorklist(V.getNode());
8291 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8292 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8293 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8296 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8297 /// ISD::INSERT_VECTOR_ELT.
8298 static SDValue PerformInsertEltCombine(SDNode *N,
8299 TargetLowering::DAGCombinerInfo &DCI) {
8300 // Bitcast an i64 load inserted into a vector to f64.
8301 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8302 EVT VT = N->getValueType(0);
8303 SDNode *Elt = N->getOperand(1).getNode();
8304 if (VT.getVectorElementType() != MVT::i64 ||
8305 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8308 SelectionDAG &DAG = DCI.DAG;
8309 DebugLoc dl = N->getDebugLoc();
8310 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8311 VT.getVectorNumElements());
8312 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8313 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8314 // Make the DAGCombiner fold the bitcasts.
8315 DCI.AddToWorklist(Vec.getNode());
8316 DCI.AddToWorklist(V.getNode());
8317 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8318 Vec, V, N->getOperand(2));
8319 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8322 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8323 /// ISD::VECTOR_SHUFFLE.
8324 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8325 // The LLVM shufflevector instruction does not require the shuffle mask
8326 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8327 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8328 // operands do not match the mask length, they are extended by concatenating
8329 // them with undef vectors. That is probably the right thing for other
8330 // targets, but for NEON it is better to concatenate two double-register
8331 // size vector operands into a single quad-register size vector. Do that
8332 // transformation here:
8333 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8334 // shuffle(concat(v1, v2), undef)
8335 SDValue Op0 = N->getOperand(0);
8336 SDValue Op1 = N->getOperand(1);
8337 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8338 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8339 Op0.getNumOperands() != 2 ||
8340 Op1.getNumOperands() != 2)
8342 SDValue Concat0Op1 = Op0.getOperand(1);
8343 SDValue Concat1Op1 = Op1.getOperand(1);
8344 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8345 Concat1Op1.getOpcode() != ISD::UNDEF)
8347 // Skip the transformation if any of the types are illegal.
8348 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8349 EVT VT = N->getValueType(0);
8350 if (!TLI.isTypeLegal(VT) ||
8351 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8352 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8355 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8356 Op0.getOperand(0), Op1.getOperand(0));
8357 // Translate the shuffle mask.
8358 SmallVector<int, 16> NewMask;
8359 unsigned NumElts = VT.getVectorNumElements();
8360 unsigned HalfElts = NumElts/2;
8361 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8362 for (unsigned n = 0; n < NumElts; ++n) {
8363 int MaskElt = SVN->getMaskElt(n);
8365 if (MaskElt < (int)HalfElts)
8367 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8368 NewElt = HalfElts + MaskElt - NumElts;
8369 NewMask.push_back(NewElt);
8371 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8372 DAG.getUNDEF(VT), NewMask.data());
8375 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8376 /// NEON load/store intrinsics to merge base address updates.
8377 static SDValue CombineBaseUpdate(SDNode *N,
8378 TargetLowering::DAGCombinerInfo &DCI) {
8379 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8382 SelectionDAG &DAG = DCI.DAG;
8383 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8384 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8385 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8386 SDValue Addr = N->getOperand(AddrOpIdx);
8388 // Search for a use of the address operand that is an increment.
8389 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8390 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8392 if (User->getOpcode() != ISD::ADD ||
8393 UI.getUse().getResNo() != Addr.getResNo())
8396 // Check that the add is independent of the load/store. Otherwise, folding
8397 // it would create a cycle.
8398 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8401 // Find the new opcode for the updating load/store.
8403 bool isLaneOp = false;
8404 unsigned NewOpc = 0;
8405 unsigned NumVecs = 0;
8407 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8409 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8410 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8412 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8414 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8416 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8418 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8419 NumVecs = 2; isLaneOp = true; break;
8420 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8421 NumVecs = 3; isLaneOp = true; break;
8422 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8423 NumVecs = 4; isLaneOp = true; break;
8424 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8425 NumVecs = 1; isLoad = false; break;
8426 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8427 NumVecs = 2; isLoad = false; break;
8428 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8429 NumVecs = 3; isLoad = false; break;
8430 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8431 NumVecs = 4; isLoad = false; break;
8432 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8433 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8434 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8435 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8436 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8437 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8441 switch (N->getOpcode()) {
8442 default: llvm_unreachable("unexpected opcode for Neon base update");
8443 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8444 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8445 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8449 // Find the size of memory referenced by the load/store.
8452 VecTy = N->getValueType(0);
8454 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8455 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8457 NumBytes /= VecTy.getVectorNumElements();
8459 // If the increment is a constant, it must match the memory ref size.
8460 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8461 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8462 uint64_t IncVal = CInc->getZExtValue();
8463 if (IncVal != NumBytes)
8465 } else if (NumBytes >= 3 * 16) {
8466 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8467 // separate instructions that make it harder to use a non-constant update.
8471 // Create the new updating load/store node.
8473 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8475 for (n = 0; n < NumResultVecs; ++n)
8477 Tys[n++] = MVT::i32;
8478 Tys[n] = MVT::Other;
8479 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8480 SmallVector<SDValue, 8> Ops;
8481 Ops.push_back(N->getOperand(0)); // incoming chain
8482 Ops.push_back(N->getOperand(AddrOpIdx));
8484 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8485 Ops.push_back(N->getOperand(i));
8487 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8488 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8489 Ops.data(), Ops.size(),
8490 MemInt->getMemoryVT(),
8491 MemInt->getMemOperand());
8494 std::vector<SDValue> NewResults;
8495 for (unsigned i = 0; i < NumResultVecs; ++i) {
8496 NewResults.push_back(SDValue(UpdN.getNode(), i));
8498 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8499 DCI.CombineTo(N, NewResults);
8500 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8507 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8508 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8509 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8511 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8512 SelectionDAG &DAG = DCI.DAG;
8513 EVT VT = N->getValueType(0);
8514 // vldN-dup instructions only support 64-bit vectors for N > 1.
8515 if (!VT.is64BitVector())
8518 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8519 SDNode *VLD = N->getOperand(0).getNode();
8520 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8522 unsigned NumVecs = 0;
8523 unsigned NewOpc = 0;
8524 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8525 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8527 NewOpc = ARMISD::VLD2DUP;
8528 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8530 NewOpc = ARMISD::VLD3DUP;
8531 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8533 NewOpc = ARMISD::VLD4DUP;
8538 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8539 // numbers match the load.
8540 unsigned VLDLaneNo =
8541 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8542 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8544 // Ignore uses of the chain result.
8545 if (UI.getUse().getResNo() == NumVecs)
8548 if (User->getOpcode() != ARMISD::VDUPLANE ||
8549 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8553 // Create the vldN-dup node.
8556 for (n = 0; n < NumVecs; ++n)
8558 Tys[n] = MVT::Other;
8559 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8560 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8561 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8562 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8563 Ops, 2, VLDMemInt->getMemoryVT(),
8564 VLDMemInt->getMemOperand());
8567 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8569 unsigned ResNo = UI.getUse().getResNo();
8570 // Ignore uses of the chain result.
8571 if (ResNo == NumVecs)
8574 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8577 // Now the vldN-lane intrinsic is dead except for its chain result.
8578 // Update uses of the chain.
8579 std::vector<SDValue> VLDDupResults;
8580 for (unsigned n = 0; n < NumVecs; ++n)
8581 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8582 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8583 DCI.CombineTo(VLD, VLDDupResults);
8588 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8589 /// ARMISD::VDUPLANE.
8590 static SDValue PerformVDUPLANECombine(SDNode *N,
8591 TargetLowering::DAGCombinerInfo &DCI) {
8592 SDValue Op = N->getOperand(0);
8594 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8595 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8596 if (CombineVLDDUP(N, DCI))
8597 return SDValue(N, 0);
8599 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8600 // redundant. Ignore bit_converts for now; element sizes are checked below.
8601 while (Op.getOpcode() == ISD::BITCAST)
8602 Op = Op.getOperand(0);
8603 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8606 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8607 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8608 // The canonical VMOV for a zero vector uses a 32-bit element size.
8609 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8611 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8613 EVT VT = N->getValueType(0);
8614 if (EltSize > VT.getVectorElementType().getSizeInBits())
8617 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8620 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8621 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8622 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8626 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8628 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8633 APFloat APF = C->getValueAPF();
8634 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8635 != APFloat::opOK || !isExact)
8638 c0 = (I == 0) ? cN : c0;
8639 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8646 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8647 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8648 /// when the VMUL has a constant operand that is a power of 2.
8650 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8651 /// vmul.f32 d16, d17, d16
8652 /// vcvt.s32.f32 d16, d16
8654 /// vcvt.s32.f32 d16, d16, #3
8655 static SDValue PerformVCVTCombine(SDNode *N,
8656 TargetLowering::DAGCombinerInfo &DCI,
8657 const ARMSubtarget *Subtarget) {
8658 SelectionDAG &DAG = DCI.DAG;
8659 SDValue Op = N->getOperand(0);
8661 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8662 Op.getOpcode() != ISD::FMUL)
8666 SDValue N0 = Op->getOperand(0);
8667 SDValue ConstVec = Op->getOperand(1);
8668 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8670 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8671 !isConstVecPow2(ConstVec, isSigned, C))
8674 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8675 Intrinsic::arm_neon_vcvtfp2fxu;
8676 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8678 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8679 DAG.getConstant(Log2_64(C), MVT::i32));
8682 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8683 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8684 /// when the VDIV has a constant operand that is a power of 2.
8686 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8687 /// vcvt.f32.s32 d16, d16
8688 /// vdiv.f32 d16, d17, d16
8690 /// vcvt.f32.s32 d16, d16, #3
8691 static SDValue PerformVDIVCombine(SDNode *N,
8692 TargetLowering::DAGCombinerInfo &DCI,
8693 const ARMSubtarget *Subtarget) {
8694 SelectionDAG &DAG = DCI.DAG;
8695 SDValue Op = N->getOperand(0);
8696 unsigned OpOpcode = Op.getNode()->getOpcode();
8698 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8699 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8703 SDValue ConstVec = N->getOperand(1);
8704 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8706 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8707 !isConstVecPow2(ConstVec, isSigned, C))
8710 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8711 Intrinsic::arm_neon_vcvtfxu2fp;
8712 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8714 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8715 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8718 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8719 /// operand of a vector shift operation, where all the elements of the
8720 /// build_vector must have the same constant integer value.
8721 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8722 // Ignore bit_converts.
8723 while (Op.getOpcode() == ISD::BITCAST)
8724 Op = Op.getOperand(0);
8725 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8726 APInt SplatBits, SplatUndef;
8727 unsigned SplatBitSize;
8729 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8730 HasAnyUndefs, ElementBits) ||
8731 SplatBitSize > ElementBits)
8733 Cnt = SplatBits.getSExtValue();
8737 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8738 /// operand of a vector shift left operation. That value must be in the range:
8739 /// 0 <= Value < ElementBits for a left shift; or
8740 /// 0 <= Value <= ElementBits for a long left shift.
8741 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8742 assert(VT.isVector() && "vector shift count is not a vector type");
8743 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8744 if (! getVShiftImm(Op, ElementBits, Cnt))
8746 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8749 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8750 /// operand of a vector shift right operation. For a shift opcode, the value
8751 /// is positive, but for an intrinsic the value count must be negative. The
8752 /// absolute value must be in the range:
8753 /// 1 <= |Value| <= ElementBits for a right shift; or
8754 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8755 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8757 assert(VT.isVector() && "vector shift count is not a vector type");
8758 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8759 if (! getVShiftImm(Op, ElementBits, Cnt))
8763 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8766 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8767 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8768 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8771 // Don't do anything for most intrinsics.
8774 // Vector shifts: check for immediate versions and lower them.
8775 // Note: This is done during DAG combining instead of DAG legalizing because
8776 // the build_vectors for 64-bit vector element shift counts are generally
8777 // not legal, and it is hard to see their values after they get legalized to
8778 // loads from a constant pool.
8779 case Intrinsic::arm_neon_vshifts:
8780 case Intrinsic::arm_neon_vshiftu:
8781 case Intrinsic::arm_neon_vshiftls:
8782 case Intrinsic::arm_neon_vshiftlu:
8783 case Intrinsic::arm_neon_vshiftn:
8784 case Intrinsic::arm_neon_vrshifts:
8785 case Intrinsic::arm_neon_vrshiftu:
8786 case Intrinsic::arm_neon_vrshiftn:
8787 case Intrinsic::arm_neon_vqshifts:
8788 case Intrinsic::arm_neon_vqshiftu:
8789 case Intrinsic::arm_neon_vqshiftsu:
8790 case Intrinsic::arm_neon_vqshiftns:
8791 case Intrinsic::arm_neon_vqshiftnu:
8792 case Intrinsic::arm_neon_vqshiftnsu:
8793 case Intrinsic::arm_neon_vqrshiftns:
8794 case Intrinsic::arm_neon_vqrshiftnu:
8795 case Intrinsic::arm_neon_vqrshiftnsu: {
8796 EVT VT = N->getOperand(1).getValueType();
8798 unsigned VShiftOpc = 0;
8801 case Intrinsic::arm_neon_vshifts:
8802 case Intrinsic::arm_neon_vshiftu:
8803 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8804 VShiftOpc = ARMISD::VSHL;
8807 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8808 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8809 ARMISD::VSHRs : ARMISD::VSHRu);
8814 case Intrinsic::arm_neon_vshiftls:
8815 case Intrinsic::arm_neon_vshiftlu:
8816 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8818 llvm_unreachable("invalid shift count for vshll intrinsic");
8820 case Intrinsic::arm_neon_vrshifts:
8821 case Intrinsic::arm_neon_vrshiftu:
8822 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8826 case Intrinsic::arm_neon_vqshifts:
8827 case Intrinsic::arm_neon_vqshiftu:
8828 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8832 case Intrinsic::arm_neon_vqshiftsu:
8833 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8835 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8837 case Intrinsic::arm_neon_vshiftn:
8838 case Intrinsic::arm_neon_vrshiftn:
8839 case Intrinsic::arm_neon_vqshiftns:
8840 case Intrinsic::arm_neon_vqshiftnu:
8841 case Intrinsic::arm_neon_vqshiftnsu:
8842 case Intrinsic::arm_neon_vqrshiftns:
8843 case Intrinsic::arm_neon_vqrshiftnu:
8844 case Intrinsic::arm_neon_vqrshiftnsu:
8845 // Narrowing shifts require an immediate right shift.
8846 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8848 llvm_unreachable("invalid shift count for narrowing vector shift "
8852 llvm_unreachable("unhandled vector shift");
8856 case Intrinsic::arm_neon_vshifts:
8857 case Intrinsic::arm_neon_vshiftu:
8858 // Opcode already set above.
8860 case Intrinsic::arm_neon_vshiftls:
8861 case Intrinsic::arm_neon_vshiftlu:
8862 if (Cnt == VT.getVectorElementType().getSizeInBits())
8863 VShiftOpc = ARMISD::VSHLLi;
8865 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8866 ARMISD::VSHLLs : ARMISD::VSHLLu);
8868 case Intrinsic::arm_neon_vshiftn:
8869 VShiftOpc = ARMISD::VSHRN; break;
8870 case Intrinsic::arm_neon_vrshifts:
8871 VShiftOpc = ARMISD::VRSHRs; break;
8872 case Intrinsic::arm_neon_vrshiftu:
8873 VShiftOpc = ARMISD::VRSHRu; break;
8874 case Intrinsic::arm_neon_vrshiftn:
8875 VShiftOpc = ARMISD::VRSHRN; break;
8876 case Intrinsic::arm_neon_vqshifts:
8877 VShiftOpc = ARMISD::VQSHLs; break;
8878 case Intrinsic::arm_neon_vqshiftu:
8879 VShiftOpc = ARMISD::VQSHLu; break;
8880 case Intrinsic::arm_neon_vqshiftsu:
8881 VShiftOpc = ARMISD::VQSHLsu; break;
8882 case Intrinsic::arm_neon_vqshiftns:
8883 VShiftOpc = ARMISD::VQSHRNs; break;
8884 case Intrinsic::arm_neon_vqshiftnu:
8885 VShiftOpc = ARMISD::VQSHRNu; break;
8886 case Intrinsic::arm_neon_vqshiftnsu:
8887 VShiftOpc = ARMISD::VQSHRNsu; break;
8888 case Intrinsic::arm_neon_vqrshiftns:
8889 VShiftOpc = ARMISD::VQRSHRNs; break;
8890 case Intrinsic::arm_neon_vqrshiftnu:
8891 VShiftOpc = ARMISD::VQRSHRNu; break;
8892 case Intrinsic::arm_neon_vqrshiftnsu:
8893 VShiftOpc = ARMISD::VQRSHRNsu; break;
8896 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8897 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8900 case Intrinsic::arm_neon_vshiftins: {
8901 EVT VT = N->getOperand(1).getValueType();
8903 unsigned VShiftOpc = 0;
8905 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8906 VShiftOpc = ARMISD::VSLI;
8907 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8908 VShiftOpc = ARMISD::VSRI;
8910 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8913 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8914 N->getOperand(1), N->getOperand(2),
8915 DAG.getConstant(Cnt, MVT::i32));
8918 case Intrinsic::arm_neon_vqrshifts:
8919 case Intrinsic::arm_neon_vqrshiftu:
8920 // No immediate versions of these to check for.
8927 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8928 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8929 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8930 /// vector element shift counts are generally not legal, and it is hard to see
8931 /// their values after they get legalized to loads from a constant pool.
8932 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8933 const ARMSubtarget *ST) {
8934 EVT VT = N->getValueType(0);
8935 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8936 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8937 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8938 SDValue N1 = N->getOperand(1);
8939 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8940 SDValue N0 = N->getOperand(0);
8941 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8942 DAG.MaskedValueIsZero(N0.getOperand(0),
8943 APInt::getHighBitsSet(32, 16)))
8944 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8948 // Nothing to be done for scalar shifts.
8949 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8950 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8953 assert(ST->hasNEON() && "unexpected vector shift");
8956 switch (N->getOpcode()) {
8957 default: llvm_unreachable("unexpected shift opcode");
8960 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8961 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8962 DAG.getConstant(Cnt, MVT::i32));
8967 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8968 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8969 ARMISD::VSHRs : ARMISD::VSHRu);
8970 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8971 DAG.getConstant(Cnt, MVT::i32));
8977 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8978 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8979 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8980 const ARMSubtarget *ST) {
8981 SDValue N0 = N->getOperand(0);
8983 // Check for sign- and zero-extensions of vector extract operations of 8-
8984 // and 16-bit vector elements. NEON supports these directly. They are
8985 // handled during DAG combining because type legalization will promote them
8986 // to 32-bit types and it is messy to recognize the operations after that.
8987 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8988 SDValue Vec = N0.getOperand(0);
8989 SDValue Lane = N0.getOperand(1);
8990 EVT VT = N->getValueType(0);
8991 EVT EltVT = N0.getValueType();
8992 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8994 if (VT == MVT::i32 &&
8995 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8996 TLI.isTypeLegal(Vec.getValueType()) &&
8997 isa<ConstantSDNode>(Lane)) {
9000 switch (N->getOpcode()) {
9001 default: llvm_unreachable("unexpected opcode");
9002 case ISD::SIGN_EXTEND:
9003 Opc = ARMISD::VGETLANEs;
9005 case ISD::ZERO_EXTEND:
9006 case ISD::ANY_EXTEND:
9007 Opc = ARMISD::VGETLANEu;
9010 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
9017 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9018 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9019 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9020 const ARMSubtarget *ST) {
9021 // If the target supports NEON, try to use vmax/vmin instructions for f32
9022 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9023 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9024 // a NaN; only do the transformation when it matches that behavior.
9026 // For now only do this when using NEON for FP operations; if using VFP, it
9027 // is not obvious that the benefit outweighs the cost of switching to the
9029 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9030 N->getValueType(0) != MVT::f32)
9033 SDValue CondLHS = N->getOperand(0);
9034 SDValue CondRHS = N->getOperand(1);
9035 SDValue LHS = N->getOperand(2);
9036 SDValue RHS = N->getOperand(3);
9037 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9039 unsigned Opcode = 0;
9041 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9042 IsReversed = false; // x CC y ? x : y
9043 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9044 IsReversed = true ; // x CC y ? y : x
9058 // If LHS is NaN, an ordered comparison will be false and the result will
9059 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9060 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9061 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9062 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9064 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9065 // will return -0, so vmin can only be used for unsafe math or if one of
9066 // the operands is known to be nonzero.
9067 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9068 !DAG.getTarget().Options.UnsafeFPMath &&
9069 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9071 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9080 // If LHS is NaN, an ordered comparison will be false and the result will
9081 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9082 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9083 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9084 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9086 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9087 // will return +0, so vmax can only be used for unsafe math or if one of
9088 // the operands is known to be nonzero.
9089 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9090 !DAG.getTarget().Options.UnsafeFPMath &&
9091 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9093 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9099 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
9102 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9104 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9105 SDValue Cmp = N->getOperand(4);
9106 if (Cmp.getOpcode() != ARMISD::CMPZ)
9107 // Only looking at EQ and NE cases.
9110 EVT VT = N->getValueType(0);
9111 DebugLoc dl = N->getDebugLoc();
9112 SDValue LHS = Cmp.getOperand(0);
9113 SDValue RHS = Cmp.getOperand(1);
9114 SDValue FalseVal = N->getOperand(0);
9115 SDValue TrueVal = N->getOperand(1);
9116 SDValue ARMcc = N->getOperand(2);
9117 ARMCC::CondCodes CC =
9118 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9136 /// FIXME: Turn this into a target neutral optimization?
9138 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9139 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9140 N->getOperand(3), Cmp);
9141 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9143 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9144 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9145 N->getOperand(3), NewCmp);
9148 if (Res.getNode()) {
9149 APInt KnownZero, KnownOne;
9150 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9151 // Capture demanded bits information that would be otherwise lost.
9152 if (KnownZero == 0xfffffffe)
9153 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9154 DAG.getValueType(MVT::i1));
9155 else if (KnownZero == 0xffffff00)
9156 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9157 DAG.getValueType(MVT::i8));
9158 else if (KnownZero == 0xffff0000)
9159 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9160 DAG.getValueType(MVT::i16));
9166 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9167 DAGCombinerInfo &DCI) const {
9168 switch (N->getOpcode()) {
9170 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9171 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9172 case ISD::SUB: return PerformSUBCombine(N, DCI);
9173 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9174 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9175 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9176 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9177 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9178 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9179 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9180 case ISD::STORE: return PerformSTORECombine(N, DCI);
9181 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9182 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9183 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9184 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9185 case ISD::FP_TO_SINT:
9186 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9187 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9188 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9191 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9192 case ISD::SIGN_EXTEND:
9193 case ISD::ZERO_EXTEND:
9194 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9195 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9196 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9197 case ARMISD::VLD2DUP:
9198 case ARMISD::VLD3DUP:
9199 case ARMISD::VLD4DUP:
9200 return CombineBaseUpdate(N, DCI);
9201 case ISD::INTRINSIC_VOID:
9202 case ISD::INTRINSIC_W_CHAIN:
9203 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9204 case Intrinsic::arm_neon_vld1:
9205 case Intrinsic::arm_neon_vld2:
9206 case Intrinsic::arm_neon_vld3:
9207 case Intrinsic::arm_neon_vld4:
9208 case Intrinsic::arm_neon_vld2lane:
9209 case Intrinsic::arm_neon_vld3lane:
9210 case Intrinsic::arm_neon_vld4lane:
9211 case Intrinsic::arm_neon_vst1:
9212 case Intrinsic::arm_neon_vst2:
9213 case Intrinsic::arm_neon_vst3:
9214 case Intrinsic::arm_neon_vst4:
9215 case Intrinsic::arm_neon_vst2lane:
9216 case Intrinsic::arm_neon_vst3lane:
9217 case Intrinsic::arm_neon_vst4lane:
9218 return CombineBaseUpdate(N, DCI);
9226 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9228 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9231 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9232 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9233 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9235 switch (VT.getSimpleVT().SimpleTy) {
9241 // Unaligned access can use (for example) LRDB, LRDH, LDR
9242 return AllowsUnaligned;
9245 // For any little-endian targets with neon, we can support unaligned ld/st
9246 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9247 // A big-endian target may also explictly support unaligned accesses
9248 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9252 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9253 unsigned AlignCheck) {
9254 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9255 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9258 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9259 unsigned DstAlign, unsigned SrcAlign,
9262 MachineFunction &MF) const {
9263 const Function *F = MF.getFunction();
9265 // See if we can use NEON instructions for this...
9267 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9268 Subtarget->hasNEON()) {
9269 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9271 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9276 // Lowering to i32/i16 if the size permits.
9279 } else if (Size >= 2) {
9283 // Let the target-independent logic figure it out.
9287 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9292 switch (VT.getSimpleVT().SimpleTy) {
9293 default: return false;
9308 if ((V & (Scale - 1)) != 0)
9311 return V == (V & ((1LL << 5) - 1));
9314 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9315 const ARMSubtarget *Subtarget) {
9322 switch (VT.getSimpleVT().SimpleTy) {
9323 default: return false;
9328 // + imm12 or - imm8
9330 return V == (V & ((1LL << 8) - 1));
9331 return V == (V & ((1LL << 12) - 1));
9334 // Same as ARM mode. FIXME: NEON?
9335 if (!Subtarget->hasVFP2())
9340 return V == (V & ((1LL << 8) - 1));
9344 /// isLegalAddressImmediate - Return true if the integer value can be used
9345 /// as the offset of the target addressing mode for load / store of the
9347 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9348 const ARMSubtarget *Subtarget) {
9355 if (Subtarget->isThumb1Only())
9356 return isLegalT1AddressImmediate(V, VT);
9357 else if (Subtarget->isThumb2())
9358 return isLegalT2AddressImmediate(V, VT, Subtarget);
9363 switch (VT.getSimpleVT().SimpleTy) {
9364 default: return false;
9369 return V == (V & ((1LL << 12) - 1));
9372 return V == (V & ((1LL << 8) - 1));
9375 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9380 return V == (V & ((1LL << 8) - 1));
9384 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9386 int Scale = AM.Scale;
9390 switch (VT.getSimpleVT().SimpleTy) {
9391 default: return false;
9400 return Scale == 2 || Scale == 4 || Scale == 8;
9403 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9407 // Note, we allow "void" uses (basically, uses that aren't loads or
9408 // stores), because arm allows folding a scale into many arithmetic
9409 // operations. This should be made more precise and revisited later.
9411 // Allow r << imm, but the imm has to be a multiple of two.
9412 if (Scale & 1) return false;
9413 return isPowerOf2_32(Scale);
9417 /// isLegalAddressingMode - Return true if the addressing mode represented
9418 /// by AM is legal for this target, for a load/store of the specified type.
9419 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9421 EVT VT = getValueType(Ty, true);
9422 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9425 // Can never fold addr of global into load/store.
9430 case 0: // no scale reg, must be "r+i" or "r", or "i".
9433 if (Subtarget->isThumb1Only())
9437 // ARM doesn't support any R+R*scale+imm addr modes.
9444 if (Subtarget->isThumb2())
9445 return isLegalT2ScaledAddressingMode(AM, VT);
9447 int Scale = AM.Scale;
9448 switch (VT.getSimpleVT().SimpleTy) {
9449 default: return false;
9453 if (Scale < 0) Scale = -Scale;
9457 return isPowerOf2_32(Scale & ~1);
9461 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9466 // Note, we allow "void" uses (basically, uses that aren't loads or
9467 // stores), because arm allows folding a scale into many arithmetic
9468 // operations. This should be made more precise and revisited later.
9470 // Allow r << imm, but the imm has to be a multiple of two.
9471 if (Scale & 1) return false;
9472 return isPowerOf2_32(Scale);
9478 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9479 /// icmp immediate, that is the target has icmp instructions which can compare
9480 /// a register against the immediate without having to materialize the
9481 /// immediate into a register.
9482 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9483 // Thumb2 and ARM modes can use cmn for negative immediates.
9484 if (!Subtarget->isThumb())
9485 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9486 if (Subtarget->isThumb2())
9487 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9488 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9489 return Imm >= 0 && Imm <= 255;
9492 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9493 /// *or sub* immediate, that is the target has add or sub instructions which can
9494 /// add a register with the immediate without having to materialize the
9495 /// immediate into a register.
9496 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9497 // Same encoding for add/sub, just flip the sign.
9498 int64_t AbsImm = llvm::abs64(Imm);
9499 if (!Subtarget->isThumb())
9500 return ARM_AM::getSOImmVal(AbsImm) != -1;
9501 if (Subtarget->isThumb2())
9502 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9503 // Thumb1 only has 8-bit unsigned immediate.
9504 return AbsImm >= 0 && AbsImm <= 255;
9507 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9508 bool isSEXTLoad, SDValue &Base,
9509 SDValue &Offset, bool &isInc,
9510 SelectionDAG &DAG) {
9511 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9514 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9516 Base = Ptr->getOperand(0);
9517 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9518 int RHSC = (int)RHS->getZExtValue();
9519 if (RHSC < 0 && RHSC > -256) {
9520 assert(Ptr->getOpcode() == ISD::ADD);
9522 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9526 isInc = (Ptr->getOpcode() == ISD::ADD);
9527 Offset = Ptr->getOperand(1);
9529 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9531 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9532 int RHSC = (int)RHS->getZExtValue();
9533 if (RHSC < 0 && RHSC > -0x1000) {
9534 assert(Ptr->getOpcode() == ISD::ADD);
9536 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9537 Base = Ptr->getOperand(0);
9542 if (Ptr->getOpcode() == ISD::ADD) {
9544 ARM_AM::ShiftOpc ShOpcVal=
9545 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9546 if (ShOpcVal != ARM_AM::no_shift) {
9547 Base = Ptr->getOperand(1);
9548 Offset = Ptr->getOperand(0);
9550 Base = Ptr->getOperand(0);
9551 Offset = Ptr->getOperand(1);
9556 isInc = (Ptr->getOpcode() == ISD::ADD);
9557 Base = Ptr->getOperand(0);
9558 Offset = Ptr->getOperand(1);
9562 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9566 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9567 bool isSEXTLoad, SDValue &Base,
9568 SDValue &Offset, bool &isInc,
9569 SelectionDAG &DAG) {
9570 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9573 Base = Ptr->getOperand(0);
9574 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9575 int RHSC = (int)RHS->getZExtValue();
9576 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9577 assert(Ptr->getOpcode() == ISD::ADD);
9579 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9581 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9582 isInc = Ptr->getOpcode() == ISD::ADD;
9583 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9591 /// getPreIndexedAddressParts - returns true by value, base pointer and
9592 /// offset pointer and addressing mode by reference if the node's address
9593 /// can be legally represented as pre-indexed load / store address.
9595 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9597 ISD::MemIndexedMode &AM,
9598 SelectionDAG &DAG) const {
9599 if (Subtarget->isThumb1Only())
9604 bool isSEXTLoad = false;
9605 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9606 Ptr = LD->getBasePtr();
9607 VT = LD->getMemoryVT();
9608 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9609 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9610 Ptr = ST->getBasePtr();
9611 VT = ST->getMemoryVT();
9616 bool isLegal = false;
9617 if (Subtarget->isThumb2())
9618 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9619 Offset, isInc, DAG);
9621 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9622 Offset, isInc, DAG);
9626 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9630 /// getPostIndexedAddressParts - returns true by value, base pointer and
9631 /// offset pointer and addressing mode by reference if this node can be
9632 /// combined with a load / store to form a post-indexed load / store.
9633 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9636 ISD::MemIndexedMode &AM,
9637 SelectionDAG &DAG) const {
9638 if (Subtarget->isThumb1Only())
9643 bool isSEXTLoad = false;
9644 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9645 VT = LD->getMemoryVT();
9646 Ptr = LD->getBasePtr();
9647 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9648 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9649 VT = ST->getMemoryVT();
9650 Ptr = ST->getBasePtr();
9655 bool isLegal = false;
9656 if (Subtarget->isThumb2())
9657 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9660 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9666 // Swap base ptr and offset to catch more post-index load / store when
9667 // it's legal. In Thumb2 mode, offset must be an immediate.
9668 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9669 !Subtarget->isThumb2())
9670 std::swap(Base, Offset);
9672 // Post-indexed load / store update the base pointer.
9677 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9681 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9684 const SelectionDAG &DAG,
9685 unsigned Depth) const {
9686 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9687 switch (Op.getOpcode()) {
9689 case ARMISD::CMOV: {
9690 // Bits are known zero/one if known on the LHS and RHS.
9691 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9692 if (KnownZero == 0 && KnownOne == 0) return;
9694 APInt KnownZeroRHS, KnownOneRHS;
9695 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9696 KnownZero &= KnownZeroRHS;
9697 KnownOne &= KnownOneRHS;
9703 //===----------------------------------------------------------------------===//
9704 // ARM Inline Assembly Support
9705 //===----------------------------------------------------------------------===//
9707 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9708 // Looking for "rev" which is V6+.
9709 if (!Subtarget->hasV6Ops())
9712 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9713 std::string AsmStr = IA->getAsmString();
9714 SmallVector<StringRef, 4> AsmPieces;
9715 SplitString(AsmStr, AsmPieces, ";\n");
9717 switch (AsmPieces.size()) {
9718 default: return false;
9720 AsmStr = AsmPieces[0];
9722 SplitString(AsmStr, AsmPieces, " \t,");
9725 if (AsmPieces.size() == 3 &&
9726 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9727 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9728 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9729 if (Ty && Ty->getBitWidth() == 32)
9730 return IntrinsicLowering::LowerToByteSwap(CI);
9738 /// getConstraintType - Given a constraint letter, return the type of
9739 /// constraint it is for this target.
9740 ARMTargetLowering::ConstraintType
9741 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9742 if (Constraint.size() == 1) {
9743 switch (Constraint[0]) {
9745 case 'l': return C_RegisterClass;
9746 case 'w': return C_RegisterClass;
9747 case 'h': return C_RegisterClass;
9748 case 'x': return C_RegisterClass;
9749 case 't': return C_RegisterClass;
9750 case 'j': return C_Other; // Constant for movw.
9751 // An address with a single base register. Due to the way we
9752 // currently handle addresses it is the same as an 'r' memory constraint.
9753 case 'Q': return C_Memory;
9755 } else if (Constraint.size() == 2) {
9756 switch (Constraint[0]) {
9758 // All 'U+' constraints are addresses.
9759 case 'U': return C_Memory;
9762 return TargetLowering::getConstraintType(Constraint);
9765 /// Examine constraint type and operand type and determine a weight value.
9766 /// This object must already have been set up with the operand type
9767 /// and the current alternative constraint selected.
9768 TargetLowering::ConstraintWeight
9769 ARMTargetLowering::getSingleConstraintMatchWeight(
9770 AsmOperandInfo &info, const char *constraint) const {
9771 ConstraintWeight weight = CW_Invalid;
9772 Value *CallOperandVal = info.CallOperandVal;
9773 // If we don't have a value, we can't do a match,
9774 // but allow it at the lowest weight.
9775 if (CallOperandVal == NULL)
9777 Type *type = CallOperandVal->getType();
9778 // Look at the constraint type.
9779 switch (*constraint) {
9781 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9784 if (type->isIntegerTy()) {
9785 if (Subtarget->isThumb())
9786 weight = CW_SpecificReg;
9788 weight = CW_Register;
9792 if (type->isFloatingPointTy())
9793 weight = CW_Register;
9799 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9801 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9803 if (Constraint.size() == 1) {
9804 // GCC ARM Constraint Letters
9805 switch (Constraint[0]) {
9806 case 'l': // Low regs or general regs.
9807 if (Subtarget->isThumb())
9808 return RCPair(0U, &ARM::tGPRRegClass);
9809 return RCPair(0U, &ARM::GPRRegClass);
9810 case 'h': // High regs or no regs.
9811 if (Subtarget->isThumb())
9812 return RCPair(0U, &ARM::hGPRRegClass);
9815 return RCPair(0U, &ARM::GPRRegClass);
9818 return RCPair(0U, &ARM::SPRRegClass);
9819 if (VT.getSizeInBits() == 64)
9820 return RCPair(0U, &ARM::DPRRegClass);
9821 if (VT.getSizeInBits() == 128)
9822 return RCPair(0U, &ARM::QPRRegClass);
9826 return RCPair(0U, &ARM::SPR_8RegClass);
9827 if (VT.getSizeInBits() == 64)
9828 return RCPair(0U, &ARM::DPR_8RegClass);
9829 if (VT.getSizeInBits() == 128)
9830 return RCPair(0U, &ARM::QPR_8RegClass);
9834 return RCPair(0U, &ARM::SPRRegClass);
9838 if (StringRef("{cc}").equals_lower(Constraint))
9839 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9841 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9844 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9845 /// vector. If it is invalid, don't add anything to Ops.
9846 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9847 std::string &Constraint,
9848 std::vector<SDValue>&Ops,
9849 SelectionDAG &DAG) const {
9850 SDValue Result(0, 0);
9852 // Currently only support length 1 constraints.
9853 if (Constraint.length() != 1) return;
9855 char ConstraintLetter = Constraint[0];
9856 switch (ConstraintLetter) {
9859 case 'I': case 'J': case 'K': case 'L':
9860 case 'M': case 'N': case 'O':
9861 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9865 int64_t CVal64 = C->getSExtValue();
9866 int CVal = (int) CVal64;
9867 // None of these constraints allow values larger than 32 bits. Check
9868 // that the value fits in an int.
9872 switch (ConstraintLetter) {
9874 // Constant suitable for movw, must be between 0 and
9876 if (Subtarget->hasV6T2Ops())
9877 if (CVal >= 0 && CVal <= 65535)
9881 if (Subtarget->isThumb1Only()) {
9882 // This must be a constant between 0 and 255, for ADD
9884 if (CVal >= 0 && CVal <= 255)
9886 } else if (Subtarget->isThumb2()) {
9887 // A constant that can be used as an immediate value in a
9888 // data-processing instruction.
9889 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9892 // A constant that can be used as an immediate value in a
9893 // data-processing instruction.
9894 if (ARM_AM::getSOImmVal(CVal) != -1)
9900 if (Subtarget->isThumb()) { // FIXME thumb2
9901 // This must be a constant between -255 and -1, for negated ADD
9902 // immediates. This can be used in GCC with an "n" modifier that
9903 // prints the negated value, for use with SUB instructions. It is
9904 // not useful otherwise but is implemented for compatibility.
9905 if (CVal >= -255 && CVal <= -1)
9908 // This must be a constant between -4095 and 4095. It is not clear
9909 // what this constraint is intended for. Implemented for
9910 // compatibility with GCC.
9911 if (CVal >= -4095 && CVal <= 4095)
9917 if (Subtarget->isThumb1Only()) {
9918 // A 32-bit value where only one byte has a nonzero value. Exclude
9919 // zero to match GCC. This constraint is used by GCC internally for
9920 // constants that can be loaded with a move/shift combination.
9921 // It is not useful otherwise but is implemented for compatibility.
9922 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9924 } else if (Subtarget->isThumb2()) {
9925 // A constant whose bitwise inverse can be used as an immediate
9926 // value in a data-processing instruction. This can be used in GCC
9927 // with a "B" modifier that prints the inverted value, for use with
9928 // BIC and MVN instructions. It is not useful otherwise but is
9929 // implemented for compatibility.
9930 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9933 // A constant whose bitwise inverse can be used as an immediate
9934 // value in a data-processing instruction. This can be used in GCC
9935 // with a "B" modifier that prints the inverted value, for use with
9936 // BIC and MVN instructions. It is not useful otherwise but is
9937 // implemented for compatibility.
9938 if (ARM_AM::getSOImmVal(~CVal) != -1)
9944 if (Subtarget->isThumb1Only()) {
9945 // This must be a constant between -7 and 7,
9946 // for 3-operand ADD/SUB immediate instructions.
9947 if (CVal >= -7 && CVal < 7)
9949 } else if (Subtarget->isThumb2()) {
9950 // A constant whose negation can be used as an immediate value in a
9951 // data-processing instruction. This can be used in GCC with an "n"
9952 // modifier that prints the negated value, for use with SUB
9953 // instructions. It is not useful otherwise but is implemented for
9955 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9958 // A constant whose negation can be used as an immediate value in a
9959 // data-processing instruction. This can be used in GCC with an "n"
9960 // modifier that prints the negated value, for use with SUB
9961 // instructions. It is not useful otherwise but is implemented for
9963 if (ARM_AM::getSOImmVal(-CVal) != -1)
9969 if (Subtarget->isThumb()) { // FIXME thumb2
9970 // This must be a multiple of 4 between 0 and 1020, for
9971 // ADD sp + immediate.
9972 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9975 // A power of two or a constant between 0 and 32. This is used in
9976 // GCC for the shift amount on shifted register operands, but it is
9977 // useful in general for any shift amounts.
9978 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9984 if (Subtarget->isThumb()) { // FIXME thumb2
9985 // This must be a constant between 0 and 31, for shift amounts.
9986 if (CVal >= 0 && CVal <= 31)
9992 if (Subtarget->isThumb()) { // FIXME thumb2
9993 // This must be a multiple of 4 between -508 and 508, for
9994 // ADD/SUB sp = sp + immediate.
9995 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10000 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10004 if (Result.getNode()) {
10005 Ops.push_back(Result);
10008 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10012 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10013 // The ARM target isn't yet aware of offsets.
10017 bool ARM::isBitFieldInvertedMask(unsigned v) {
10018 if (v == 0xffffffff)
10020 // there can be 1's on either or both "outsides", all the "inside"
10021 // bits must be 0's
10022 unsigned int lsb = 0, msb = 31;
10023 while (v & (1 << msb)) --msb;
10024 while (v & (1 << lsb)) ++lsb;
10025 for (unsigned int i = lsb; i <= msb; ++i) {
10032 /// isFPImmLegal - Returns true if the target can instruction select the
10033 /// specified FP immediate natively. If false, the legalizer will
10034 /// materialize the FP immediate as a load from a constant pool.
10035 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10036 if (!Subtarget->hasVFP3())
10038 if (VT == MVT::f32)
10039 return ARM_AM::getFP32Imm(Imm) != -1;
10040 if (VT == MVT::f64)
10041 return ARM_AM::getFP64Imm(Imm) != -1;
10045 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10046 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10047 /// specified in the intrinsic calls.
10048 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10050 unsigned Intrinsic) const {
10051 switch (Intrinsic) {
10052 case Intrinsic::arm_neon_vld1:
10053 case Intrinsic::arm_neon_vld2:
10054 case Intrinsic::arm_neon_vld3:
10055 case Intrinsic::arm_neon_vld4:
10056 case Intrinsic::arm_neon_vld2lane:
10057 case Intrinsic::arm_neon_vld3lane:
10058 case Intrinsic::arm_neon_vld4lane: {
10059 Info.opc = ISD::INTRINSIC_W_CHAIN;
10060 // Conservatively set memVT to the entire set of vectors loaded.
10061 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10062 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10063 Info.ptrVal = I.getArgOperand(0);
10065 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10066 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10067 Info.vol = false; // volatile loads with NEON intrinsics not supported
10068 Info.readMem = true;
10069 Info.writeMem = false;
10072 case Intrinsic::arm_neon_vst1:
10073 case Intrinsic::arm_neon_vst2:
10074 case Intrinsic::arm_neon_vst3:
10075 case Intrinsic::arm_neon_vst4:
10076 case Intrinsic::arm_neon_vst2lane:
10077 case Intrinsic::arm_neon_vst3lane:
10078 case Intrinsic::arm_neon_vst4lane: {
10079 Info.opc = ISD::INTRINSIC_VOID;
10080 // Conservatively set memVT to the entire set of vectors stored.
10081 unsigned NumElts = 0;
10082 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10083 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10084 if (!ArgTy->isVectorTy())
10086 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10088 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10089 Info.ptrVal = I.getArgOperand(0);
10091 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10092 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10093 Info.vol = false; // volatile stores with NEON intrinsics not supported
10094 Info.readMem = false;
10095 Info.writeMem = true;
10098 case Intrinsic::arm_strexd: {
10099 Info.opc = ISD::INTRINSIC_W_CHAIN;
10100 Info.memVT = MVT::i64;
10101 Info.ptrVal = I.getArgOperand(2);
10105 Info.readMem = false;
10106 Info.writeMem = true;
10109 case Intrinsic::arm_ldrexd: {
10110 Info.opc = ISD::INTRINSIC_W_CHAIN;
10111 Info.memVT = MVT::i64;
10112 Info.ptrVal = I.getArgOperand(0);
10116 Info.readMem = true;
10117 Info.writeMem = false;