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::SIGN_EXTEND_INREG, VT, Expand);
126 if (VT.isInteger()) {
127 setOperationAction(ISD::SHL, VT, Custom);
128 setOperationAction(ISD::SRA, VT, Custom);
129 setOperationAction(ISD::SRL, VT, Custom);
132 // Promote all bit-wise operations.
133 if (VT.isInteger() && VT != PromotedBitwiseVT) {
134 setOperationAction(ISD::AND, VT, Promote);
135 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
136 setOperationAction(ISD::OR, VT, Promote);
137 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
138 setOperationAction(ISD::XOR, VT, Promote);
139 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
142 // Neon does not support vector divide/remainder operations.
143 setOperationAction(ISD::SDIV, VT, Expand);
144 setOperationAction(ISD::UDIV, VT, Expand);
145 setOperationAction(ISD::FDIV, VT, Expand);
146 setOperationAction(ISD::SREM, VT, Expand);
147 setOperationAction(ISD::UREM, VT, Expand);
148 setOperationAction(ISD::FREM, VT, Expand);
151 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
152 addRegisterClass(VT, &ARM::DPRRegClass);
153 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
156 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
157 addRegisterClass(VT, &ARM::QPRRegClass);
158 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
161 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
162 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
163 return new TargetLoweringObjectFileMachO();
165 return new ARMElfTargetObjectFile();
168 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
169 : TargetLowering(TM, createTLOF(TM)) {
170 Subtarget = &TM.getSubtarget<ARMSubtarget>();
171 RegInfo = TM.getRegisterInfo();
172 Itins = TM.getInstrItineraryData();
174 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
176 if (Subtarget->isTargetDarwin()) {
177 // Uses VFP for Thumb libfuncs if available.
178 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
179 // Single-precision floating-point arithmetic.
180 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
181 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
182 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
183 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
185 // Double-precision floating-point arithmetic.
186 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
187 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
188 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
189 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
191 // Single-precision comparisons.
192 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
193 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
194 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
195 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
196 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
197 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
198 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
199 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
201 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
210 // Double-precision comparisons.
211 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
212 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
213 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
214 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
215 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
216 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
217 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
218 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
220 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
229 // Floating-point to integer conversions.
230 // i64 conversions are done via library routines even when generating VFP
231 // instructions, so use the same ones.
232 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
233 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
234 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
235 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
237 // Conversions between floating types.
238 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
239 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
241 // Integer to floating-point conversions.
242 // i64 conversions are done via library routines even when generating VFP
243 // instructions, so use the same ones.
244 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
245 // e.g., __floatunsidf vs. __floatunssidfvfp.
246 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
247 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
248 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
249 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
253 // These libcalls are not available in 32-bit.
254 setLibcallName(RTLIB::SHL_I128, 0);
255 setLibcallName(RTLIB::SRL_I128, 0);
256 setLibcallName(RTLIB::SRA_I128, 0);
258 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
259 // Double-precision floating-point arithmetic helper functions
260 // RTABI chapter 4.1.2, Table 2
261 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
262 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
263 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
264 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
265 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
266 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
270 // Double-precision floating-point comparison helper functions
271 // RTABI chapter 4.1.2, Table 3
272 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
273 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
274 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
275 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
276 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
277 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
278 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
279 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
280 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
281 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
282 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
283 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
284 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
285 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
286 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
287 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
288 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
289 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
297 // Single-precision floating-point arithmetic helper functions
298 // RTABI chapter 4.1.2, Table 4
299 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
300 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
301 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
302 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
303 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
304 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
308 // Single-precision floating-point comparison helper functions
309 // RTABI chapter 4.1.2, Table 5
310 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
311 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
312 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
313 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
314 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
315 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
316 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
317 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
318 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
319 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
320 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
321 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
322 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
323 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
324 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
325 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
326 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
327 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
335 // Floating-point to integer conversions.
336 // RTABI chapter 4.1.2, Table 6
337 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
338 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
339 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
340 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
341 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
342 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
343 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
344 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
345 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
346 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
354 // Conversions between floating types.
355 // RTABI chapter 4.1.2, Table 7
356 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
357 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
358 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
361 // Integer to floating-point conversions.
362 // RTABI chapter 4.1.2, Table 8
363 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
364 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
365 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
366 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
367 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
368 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
369 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
370 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
371 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
372 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
380 // Long long helper functions
381 // RTABI chapter 4.2, Table 9
382 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
383 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
384 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
385 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
386 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
393 // Integer division functions
394 // RTABI chapter 4.3.1
395 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
396 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
399 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
400 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
403 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
404 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
413 // RTABI chapter 4.3.4
414 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
415 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
416 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
417 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
418 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
422 // Use divmod compiler-rt calls for iOS 5.0 and later.
423 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
424 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
425 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
426 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
429 if (Subtarget->isThumb1Only())
430 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
432 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
433 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
434 !Subtarget->isThumb1Only()) {
435 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
436 if (!Subtarget->isFPOnlySP())
437 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
439 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
442 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
443 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
444 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
445 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
446 setTruncStoreAction((MVT::SimpleValueType)VT,
447 (MVT::SimpleValueType)InnerVT, Expand);
448 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
449 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
453 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
455 if (Subtarget->hasNEON()) {
456 addDRTypeForNEON(MVT::v2f32);
457 addDRTypeForNEON(MVT::v8i8);
458 addDRTypeForNEON(MVT::v4i16);
459 addDRTypeForNEON(MVT::v2i32);
460 addDRTypeForNEON(MVT::v1i64);
462 addQRTypeForNEON(MVT::v4f32);
463 addQRTypeForNEON(MVT::v2f64);
464 addQRTypeForNEON(MVT::v16i8);
465 addQRTypeForNEON(MVT::v8i16);
466 addQRTypeForNEON(MVT::v4i32);
467 addQRTypeForNEON(MVT::v2i64);
469 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
470 // neither Neon nor VFP support any arithmetic operations on it.
471 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
472 // supported for v4f32.
473 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
474 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
475 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
476 // FIXME: Code duplication: FDIV and FREM are expanded always, see
477 // ARMTargetLowering::addTypeForNEON method for details.
478 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
479 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
480 // FIXME: Create unittest.
481 // In another words, find a way when "copysign" appears in DAG with vector
483 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
484 // FIXME: Code duplication: SETCC has custom operation action, see
485 // ARMTargetLowering::addTypeForNEON method for details.
486 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
487 // FIXME: Create unittest for FNEG and for FABS.
488 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
489 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
490 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
492 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
493 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
495 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
498 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
500 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
501 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
502 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
503 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
504 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
507 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
508 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
510 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
512 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
515 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
517 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
519 // Neon does not support some operations on v1i64 and v2i64 types.
520 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
521 // Custom handling for some quad-vector types to detect VMULL.
522 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
523 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
524 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
525 // Custom handling for some vector types to avoid expensive expansions
526 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
527 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
528 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
529 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
530 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
531 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
532 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
533 // a destination type that is wider than the source, and nor does
534 // it have a FP_TO_[SU]INT instruction with a narrower destination than
536 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
537 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
538 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
539 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
541 setTargetDAGCombine(ISD::INTRINSIC_VOID);
542 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
543 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
544 setTargetDAGCombine(ISD::SHL);
545 setTargetDAGCombine(ISD::SRL);
546 setTargetDAGCombine(ISD::SRA);
547 setTargetDAGCombine(ISD::SIGN_EXTEND);
548 setTargetDAGCombine(ISD::ZERO_EXTEND);
549 setTargetDAGCombine(ISD::ANY_EXTEND);
550 setTargetDAGCombine(ISD::SELECT_CC);
551 setTargetDAGCombine(ISD::BUILD_VECTOR);
552 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
553 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
554 setTargetDAGCombine(ISD::STORE);
555 setTargetDAGCombine(ISD::FP_TO_SINT);
556 setTargetDAGCombine(ISD::FP_TO_UINT);
557 setTargetDAGCombine(ISD::FDIV);
559 // It is legal to extload from v4i8 to v4i16 or v4i32.
560 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
561 MVT::v4i16, MVT::v2i16,
563 for (unsigned i = 0; i < 6; ++i) {
564 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
565 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
566 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
570 // ARM and Thumb2 support UMLAL/SMLAL.
571 if (!Subtarget->isThumb1Only())
572 setTargetDAGCombine(ISD::ADDC);
575 computeRegisterProperties();
577 // ARM does not have f32 extending load.
578 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
580 // ARM does not have i1 sign extending load.
581 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
583 // ARM supports all 4 flavors of integer indexed load / store.
584 if (!Subtarget->isThumb1Only()) {
585 for (unsigned im = (unsigned)ISD::PRE_INC;
586 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
587 setIndexedLoadAction(im, MVT::i1, Legal);
588 setIndexedLoadAction(im, MVT::i8, Legal);
589 setIndexedLoadAction(im, MVT::i16, Legal);
590 setIndexedLoadAction(im, MVT::i32, Legal);
591 setIndexedStoreAction(im, MVT::i1, Legal);
592 setIndexedStoreAction(im, MVT::i8, Legal);
593 setIndexedStoreAction(im, MVT::i16, Legal);
594 setIndexedStoreAction(im, MVT::i32, Legal);
598 // i64 operation support.
599 setOperationAction(ISD::MUL, MVT::i64, Expand);
600 setOperationAction(ISD::MULHU, MVT::i32, Expand);
601 if (Subtarget->isThumb1Only()) {
602 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
603 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
605 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
606 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
607 setOperationAction(ISD::MULHS, MVT::i32, Expand);
609 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
610 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
612 setOperationAction(ISD::SRL, MVT::i64, Custom);
613 setOperationAction(ISD::SRA, MVT::i64, Custom);
615 if (!Subtarget->isThumb1Only()) {
616 // FIXME: We should do this for Thumb1 as well.
617 setOperationAction(ISD::ADDC, MVT::i32, Custom);
618 setOperationAction(ISD::ADDE, MVT::i32, Custom);
619 setOperationAction(ISD::SUBC, MVT::i32, Custom);
620 setOperationAction(ISD::SUBE, MVT::i32, Custom);
623 // ARM does not have ROTL.
624 setOperationAction(ISD::ROTL, MVT::i32, Expand);
625 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
626 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
627 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
628 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
630 // These just redirect to CTTZ and CTLZ on ARM.
631 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
632 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
634 // Only ARMv6 has BSWAP.
635 if (!Subtarget->hasV6Ops())
636 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
638 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
639 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
640 // These are expanded into libcalls if the cpu doesn't have HW divider.
641 setOperationAction(ISD::SDIV, MVT::i32, Expand);
642 setOperationAction(ISD::UDIV, MVT::i32, Expand);
644 setOperationAction(ISD::SREM, MVT::i32, Expand);
645 setOperationAction(ISD::UREM, MVT::i32, Expand);
646 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
647 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
649 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
650 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
651 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
652 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
653 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
655 setOperationAction(ISD::TRAP, MVT::Other, Legal);
657 // Use the default implementation.
658 setOperationAction(ISD::VASTART, MVT::Other, Custom);
659 setOperationAction(ISD::VAARG, MVT::Other, Expand);
660 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
661 setOperationAction(ISD::VAEND, MVT::Other, Expand);
662 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
663 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
665 if (!Subtarget->isTargetDarwin()) {
666 // Non-Darwin platforms may return values in these registers via the
667 // personality function.
668 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
669 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
670 setExceptionPointerRegister(ARM::R0);
671 setExceptionSelectorRegister(ARM::R1);
674 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
675 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
676 // the default expansion.
677 // FIXME: This should be checking for v6k, not just v6.
678 if (Subtarget->hasDataBarrier() ||
679 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
680 // membarrier needs custom lowering; the rest are legal and handled
682 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
683 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
684 // Custom lowering for 64-bit ops
685 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
686 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
691 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
692 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
693 setInsertFencesForAtomic(true);
695 // Set them all for expansion, which will force libcalls.
696 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
697 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
698 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
699 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
710 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
711 // Unordered/Monotonic case.
712 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
713 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
714 // Since the libcalls include locking, fold in the fences
715 setShouldFoldAtomicFences(true);
718 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
720 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
721 if (!Subtarget->hasV6Ops()) {
722 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
725 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
727 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
728 !Subtarget->isThumb1Only()) {
729 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
730 // iff target supports vfp2.
731 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
732 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
735 // We want to custom lower some of our intrinsics.
736 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
737 if (Subtarget->isTargetDarwin()) {
738 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
739 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
740 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
743 setOperationAction(ISD::SETCC, MVT::i32, Expand);
744 setOperationAction(ISD::SETCC, MVT::f32, Expand);
745 setOperationAction(ISD::SETCC, MVT::f64, Expand);
746 setOperationAction(ISD::SELECT, MVT::i32, Custom);
747 setOperationAction(ISD::SELECT, MVT::f32, Custom);
748 setOperationAction(ISD::SELECT, MVT::f64, Custom);
749 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
751 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
753 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
754 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
755 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
756 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
757 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
759 // We don't support sin/cos/fmod/copysign/pow
760 setOperationAction(ISD::FSIN, MVT::f64, Expand);
761 setOperationAction(ISD::FSIN, MVT::f32, Expand);
762 setOperationAction(ISD::FCOS, MVT::f32, Expand);
763 setOperationAction(ISD::FCOS, MVT::f64, Expand);
764 setOperationAction(ISD::FREM, MVT::f64, Expand);
765 setOperationAction(ISD::FREM, MVT::f32, Expand);
766 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
767 !Subtarget->isThumb1Only()) {
768 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
769 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
771 setOperationAction(ISD::FPOW, MVT::f64, Expand);
772 setOperationAction(ISD::FPOW, MVT::f32, Expand);
774 if (!Subtarget->hasVFP4()) {
775 setOperationAction(ISD::FMA, MVT::f64, Expand);
776 setOperationAction(ISD::FMA, MVT::f32, Expand);
779 // Various VFP goodness
780 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
781 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
782 if (Subtarget->hasVFP2()) {
783 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
784 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
785 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
786 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
788 // Special handling for half-precision FP.
789 if (!Subtarget->hasFP16()) {
790 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
791 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
795 // We have target-specific dag combine patterns for the following nodes:
796 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
797 setTargetDAGCombine(ISD::ADD);
798 setTargetDAGCombine(ISD::SUB);
799 setTargetDAGCombine(ISD::MUL);
800 setTargetDAGCombine(ISD::AND);
801 setTargetDAGCombine(ISD::OR);
802 setTargetDAGCombine(ISD::XOR);
804 if (Subtarget->hasV6Ops())
805 setTargetDAGCombine(ISD::SRL);
807 setStackPointerRegisterToSaveRestore(ARM::SP);
809 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
810 !Subtarget->hasVFP2())
811 setSchedulingPreference(Sched::RegPressure);
813 setSchedulingPreference(Sched::Hybrid);
815 //// temporary - rewrite interface to use type
816 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
817 maxStoresPerMemset = 16;
818 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
820 // On ARM arguments smaller than 4 bytes are extended, so all arguments
821 // are at least 4 bytes aligned.
822 setMinStackArgumentAlignment(4);
824 benefitFromCodePlacementOpt = true;
826 // Prefer likely predicted branches to selects on out-of-order cores.
827 predictableSelectIsExpensive = Subtarget->isLikeA9();
829 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
832 // FIXME: It might make sense to define the representative register class as the
833 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
834 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
835 // SPR's representative would be DPR_VFP2. This should work well if register
836 // pressure tracking were modified such that a register use would increment the
837 // pressure of the register class's representative and all of it's super
838 // classes' representatives transitively. We have not implemented this because
839 // of the difficulty prior to coalescing of modeling operand register classes
840 // due to the common occurrence of cross class copies and subregister insertions
842 std::pair<const TargetRegisterClass*, uint8_t>
843 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
844 const TargetRegisterClass *RRC = 0;
846 switch (VT.getSimpleVT().SimpleTy) {
848 return TargetLowering::findRepresentativeClass(VT);
849 // Use DPR as representative register class for all floating point
850 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
851 // the cost is 1 for both f32 and f64.
852 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
853 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
854 RRC = &ARM::DPRRegClass;
855 // When NEON is used for SP, only half of the register file is available
856 // because operations that define both SP and DP results will be constrained
857 // to the VFP2 class (D0-D15). We currently model this constraint prior to
858 // coalescing by double-counting the SP regs. See the FIXME above.
859 if (Subtarget->useNEONForSinglePrecisionFP())
862 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
863 case MVT::v4f32: case MVT::v2f64:
864 RRC = &ARM::DPRRegClass;
868 RRC = &ARM::DPRRegClass;
872 RRC = &ARM::DPRRegClass;
876 return std::make_pair(RRC, Cost);
879 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
882 case ARMISD::Wrapper: return "ARMISD::Wrapper";
883 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
884 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
885 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
886 case ARMISD::CALL: return "ARMISD::CALL";
887 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
888 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
889 case ARMISD::tCALL: return "ARMISD::tCALL";
890 case ARMISD::BRCOND: return "ARMISD::BRCOND";
891 case ARMISD::BR_JT: return "ARMISD::BR_JT";
892 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
893 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
894 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
895 case ARMISD::CMP: return "ARMISD::CMP";
896 case ARMISD::CMN: return "ARMISD::CMN";
897 case ARMISD::CMPZ: return "ARMISD::CMPZ";
898 case ARMISD::CMPFP: return "ARMISD::CMPFP";
899 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
900 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
901 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
903 case ARMISD::CMOV: return "ARMISD::CMOV";
905 case ARMISD::RBIT: return "ARMISD::RBIT";
907 case ARMISD::FTOSI: return "ARMISD::FTOSI";
908 case ARMISD::FTOUI: return "ARMISD::FTOUI";
909 case ARMISD::SITOF: return "ARMISD::SITOF";
910 case ARMISD::UITOF: return "ARMISD::UITOF";
912 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
913 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
914 case ARMISD::RRX: return "ARMISD::RRX";
916 case ARMISD::ADDC: return "ARMISD::ADDC";
917 case ARMISD::ADDE: return "ARMISD::ADDE";
918 case ARMISD::SUBC: return "ARMISD::SUBC";
919 case ARMISD::SUBE: return "ARMISD::SUBE";
921 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
922 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
924 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
925 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
927 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
929 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
931 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
933 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
934 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
936 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
938 case ARMISD::VCEQ: return "ARMISD::VCEQ";
939 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
940 case ARMISD::VCGE: return "ARMISD::VCGE";
941 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
942 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
943 case ARMISD::VCGEU: return "ARMISD::VCGEU";
944 case ARMISD::VCGT: return "ARMISD::VCGT";
945 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
946 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
947 case ARMISD::VCGTU: return "ARMISD::VCGTU";
948 case ARMISD::VTST: return "ARMISD::VTST";
950 case ARMISD::VSHL: return "ARMISD::VSHL";
951 case ARMISD::VSHRs: return "ARMISD::VSHRs";
952 case ARMISD::VSHRu: return "ARMISD::VSHRu";
953 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
954 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
955 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
956 case ARMISD::VSHRN: return "ARMISD::VSHRN";
957 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
958 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
959 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
960 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
961 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
962 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
963 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
964 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
965 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
966 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
967 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
968 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
969 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
970 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
971 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
972 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
973 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
974 case ARMISD::VDUP: return "ARMISD::VDUP";
975 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
976 case ARMISD::VEXT: return "ARMISD::VEXT";
977 case ARMISD::VREV64: return "ARMISD::VREV64";
978 case ARMISD::VREV32: return "ARMISD::VREV32";
979 case ARMISD::VREV16: return "ARMISD::VREV16";
980 case ARMISD::VZIP: return "ARMISD::VZIP";
981 case ARMISD::VUZP: return "ARMISD::VUZP";
982 case ARMISD::VTRN: return "ARMISD::VTRN";
983 case ARMISD::VTBL1: return "ARMISD::VTBL1";
984 case ARMISD::VTBL2: return "ARMISD::VTBL2";
985 case ARMISD::VMULLs: return "ARMISD::VMULLs";
986 case ARMISD::VMULLu: return "ARMISD::VMULLu";
987 case ARMISD::UMLAL: return "ARMISD::UMLAL";
988 case ARMISD::SMLAL: return "ARMISD::SMLAL";
989 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
990 case ARMISD::FMAX: return "ARMISD::FMAX";
991 case ARMISD::FMIN: return "ARMISD::FMIN";
992 case ARMISD::BFI: return "ARMISD::BFI";
993 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
994 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
995 case ARMISD::VBSL: return "ARMISD::VBSL";
996 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
997 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
998 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
999 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1000 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1001 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1002 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1003 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1004 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1005 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1006 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1007 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1008 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1009 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1010 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1011 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1012 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1013 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1014 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1015 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1019 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1020 if (!VT.isVector()) return getPointerTy();
1021 return VT.changeVectorElementTypeToInteger();
1024 /// getRegClassFor - Return the register class that should be used for the
1025 /// specified value type.
1026 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1027 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1028 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1029 // load / store 4 to 8 consecutive D registers.
1030 if (Subtarget->hasNEON()) {
1031 if (VT == MVT::v4i64)
1032 return &ARM::QQPRRegClass;
1033 if (VT == MVT::v8i64)
1034 return &ARM::QQQQPRRegClass;
1036 return TargetLowering::getRegClassFor(VT);
1039 // Create a fast isel object.
1041 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1042 const TargetLibraryInfo *libInfo) const {
1043 return ARM::createFastISel(funcInfo, libInfo);
1046 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1047 /// be used for loads / stores from the global.
1048 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1049 return (Subtarget->isThumb1Only() ? 127 : 4095);
1052 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1053 unsigned NumVals = N->getNumValues();
1055 return Sched::RegPressure;
1057 for (unsigned i = 0; i != NumVals; ++i) {
1058 EVT VT = N->getValueType(i);
1059 if (VT == MVT::Glue || VT == MVT::Other)
1061 if (VT.isFloatingPoint() || VT.isVector())
1065 if (!N->isMachineOpcode())
1066 return Sched::RegPressure;
1068 // Load are scheduled for latency even if there instruction itinerary
1069 // is not available.
1070 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1071 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1073 if (MCID.getNumDefs() == 0)
1074 return Sched::RegPressure;
1075 if (!Itins->isEmpty() &&
1076 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1079 return Sched::RegPressure;
1082 //===----------------------------------------------------------------------===//
1084 //===----------------------------------------------------------------------===//
1086 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1087 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1089 default: llvm_unreachable("Unknown condition code!");
1090 case ISD::SETNE: return ARMCC::NE;
1091 case ISD::SETEQ: return ARMCC::EQ;
1092 case ISD::SETGT: return ARMCC::GT;
1093 case ISD::SETGE: return ARMCC::GE;
1094 case ISD::SETLT: return ARMCC::LT;
1095 case ISD::SETLE: return ARMCC::LE;
1096 case ISD::SETUGT: return ARMCC::HI;
1097 case ISD::SETUGE: return ARMCC::HS;
1098 case ISD::SETULT: return ARMCC::LO;
1099 case ISD::SETULE: return ARMCC::LS;
1103 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1104 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1105 ARMCC::CondCodes &CondCode2) {
1106 CondCode2 = ARMCC::AL;
1108 default: llvm_unreachable("Unknown FP condition!");
1110 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1112 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1114 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1115 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1116 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1117 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1118 case ISD::SETO: CondCode = ARMCC::VC; break;
1119 case ISD::SETUO: CondCode = ARMCC::VS; break;
1120 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1121 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1122 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1124 case ISD::SETULT: CondCode = ARMCC::LT; break;
1126 case ISD::SETULE: CondCode = ARMCC::LE; break;
1128 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1132 //===----------------------------------------------------------------------===//
1133 // Calling Convention Implementation
1134 //===----------------------------------------------------------------------===//
1136 #include "ARMGenCallingConv.inc"
1138 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1139 /// given CallingConvention value.
1140 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1142 bool isVarArg) const {
1145 llvm_unreachable("Unsupported calling convention");
1146 case CallingConv::Fast:
1147 if (Subtarget->hasVFP2() && !isVarArg) {
1148 if (!Subtarget->isAAPCS_ABI())
1149 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1150 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1151 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1154 case CallingConv::C: {
1155 // Use target triple & subtarget features to do actual dispatch.
1156 if (!Subtarget->isAAPCS_ABI())
1157 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1158 else if (Subtarget->hasVFP2() &&
1159 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1161 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1162 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1164 case CallingConv::ARM_AAPCS_VFP:
1166 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1168 case CallingConv::ARM_AAPCS:
1169 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1170 case CallingConv::ARM_APCS:
1171 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1172 case CallingConv::GHC:
1173 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1177 /// LowerCallResult - Lower the result values of a call into the
1178 /// appropriate copies out of appropriate physical registers.
1180 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1181 CallingConv::ID CallConv, bool isVarArg,
1182 const SmallVectorImpl<ISD::InputArg> &Ins,
1183 DebugLoc dl, SelectionDAG &DAG,
1184 SmallVectorImpl<SDValue> &InVals) const {
1186 // Assign locations to each value returned by this call.
1187 SmallVector<CCValAssign, 16> RVLocs;
1188 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1189 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1190 CCInfo.AnalyzeCallResult(Ins,
1191 CCAssignFnForNode(CallConv, /* Return*/ true,
1194 // Copy all of the result registers out of their specified physreg.
1195 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1196 CCValAssign VA = RVLocs[i];
1199 if (VA.needsCustom()) {
1200 // Handle f64 or half of a v2f64.
1201 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1203 Chain = Lo.getValue(1);
1204 InFlag = Lo.getValue(2);
1205 VA = RVLocs[++i]; // skip ahead to next loc
1206 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1208 Chain = Hi.getValue(1);
1209 InFlag = Hi.getValue(2);
1210 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1212 if (VA.getLocVT() == MVT::v2f64) {
1213 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1214 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1215 DAG.getConstant(0, MVT::i32));
1217 VA = RVLocs[++i]; // skip ahead to next loc
1218 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1219 Chain = Lo.getValue(1);
1220 InFlag = Lo.getValue(2);
1221 VA = RVLocs[++i]; // skip ahead to next loc
1222 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1223 Chain = Hi.getValue(1);
1224 InFlag = Hi.getValue(2);
1225 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1226 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1227 DAG.getConstant(1, MVT::i32));
1230 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1232 Chain = Val.getValue(1);
1233 InFlag = Val.getValue(2);
1236 switch (VA.getLocInfo()) {
1237 default: llvm_unreachable("Unknown loc info!");
1238 case CCValAssign::Full: break;
1239 case CCValAssign::BCvt:
1240 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1244 InVals.push_back(Val);
1250 /// LowerMemOpCallTo - Store the argument to the stack.
1252 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1253 SDValue StackPtr, SDValue Arg,
1254 DebugLoc dl, SelectionDAG &DAG,
1255 const CCValAssign &VA,
1256 ISD::ArgFlagsTy Flags) const {
1257 unsigned LocMemOffset = VA.getLocMemOffset();
1258 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1259 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1260 return DAG.getStore(Chain, dl, Arg, PtrOff,
1261 MachinePointerInfo::getStack(LocMemOffset),
1265 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1266 SDValue Chain, SDValue &Arg,
1267 RegsToPassVector &RegsToPass,
1268 CCValAssign &VA, CCValAssign &NextVA,
1270 SmallVector<SDValue, 8> &MemOpChains,
1271 ISD::ArgFlagsTy Flags) const {
1273 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1274 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1275 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1277 if (NextVA.isRegLoc())
1278 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1280 assert(NextVA.isMemLoc());
1281 if (StackPtr.getNode() == 0)
1282 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1284 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1290 /// LowerCall - Lowering a call into a callseq_start <-
1291 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1294 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1295 SmallVectorImpl<SDValue> &InVals) const {
1296 SelectionDAG &DAG = CLI.DAG;
1297 DebugLoc &dl = CLI.DL;
1298 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1299 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1300 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1301 SDValue Chain = CLI.Chain;
1302 SDValue Callee = CLI.Callee;
1303 bool &isTailCall = CLI.IsTailCall;
1304 CallingConv::ID CallConv = CLI.CallConv;
1305 bool doesNotRet = CLI.DoesNotReturn;
1306 bool isVarArg = CLI.IsVarArg;
1308 MachineFunction &MF = DAG.getMachineFunction();
1309 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1310 bool IsSibCall = false;
1311 // Disable tail calls if they're not supported.
1312 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1315 // Check if it's really possible to do a tail call.
1316 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1317 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1318 Outs, OutVals, Ins, DAG);
1319 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1320 // detected sibcalls.
1327 // Analyze operands of the call, assigning locations to each operand.
1328 SmallVector<CCValAssign, 16> ArgLocs;
1329 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1330 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1331 CCInfo.AnalyzeCallOperands(Outs,
1332 CCAssignFnForNode(CallConv, /* Return*/ false,
1335 // Get a count of how many bytes are to be pushed on the stack.
1336 unsigned NumBytes = CCInfo.getNextStackOffset();
1338 // For tail calls, memory operands are available in our caller's stack.
1342 // Adjust the stack pointer for the new arguments...
1343 // These operations are automatically eliminated by the prolog/epilog pass
1345 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1347 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1349 RegsToPassVector RegsToPass;
1350 SmallVector<SDValue, 8> MemOpChains;
1352 // Walk the register/memloc assignments, inserting copies/loads. In the case
1353 // of tail call optimization, arguments are handled later.
1354 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1356 ++i, ++realArgIdx) {
1357 CCValAssign &VA = ArgLocs[i];
1358 SDValue Arg = OutVals[realArgIdx];
1359 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1360 bool isByVal = Flags.isByVal();
1362 // Promote the value if needed.
1363 switch (VA.getLocInfo()) {
1364 default: llvm_unreachable("Unknown loc info!");
1365 case CCValAssign::Full: break;
1366 case CCValAssign::SExt:
1367 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1369 case CCValAssign::ZExt:
1370 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1372 case CCValAssign::AExt:
1373 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1375 case CCValAssign::BCvt:
1376 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1380 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1381 if (VA.needsCustom()) {
1382 if (VA.getLocVT() == MVT::v2f64) {
1383 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1384 DAG.getConstant(0, MVT::i32));
1385 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1386 DAG.getConstant(1, MVT::i32));
1388 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1389 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1391 VA = ArgLocs[++i]; // skip ahead to next loc
1392 if (VA.isRegLoc()) {
1393 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1394 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1396 assert(VA.isMemLoc());
1398 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1399 dl, DAG, VA, Flags));
1402 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1403 StackPtr, MemOpChains, Flags);
1405 } else if (VA.isRegLoc()) {
1406 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1407 } else if (isByVal) {
1408 assert(VA.isMemLoc());
1409 unsigned offset = 0;
1411 // True if this byval aggregate will be split between registers
1413 if (CCInfo.isFirstByValRegValid()) {
1414 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1416 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1417 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1418 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1419 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1420 MachinePointerInfo(),
1421 false, false, false, 0);
1422 MemOpChains.push_back(Load.getValue(1));
1423 RegsToPass.push_back(std::make_pair(j, Load));
1425 offset = ARM::R4 - CCInfo.getFirstByValReg();
1426 CCInfo.clearFirstByValReg();
1429 if (Flags.getByValSize() - 4*offset > 0) {
1430 unsigned LocMemOffset = VA.getLocMemOffset();
1431 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1432 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1434 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1435 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1436 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1438 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1440 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1441 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1442 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1443 Ops, array_lengthof(Ops)));
1445 } else if (!IsSibCall) {
1446 assert(VA.isMemLoc());
1448 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1449 dl, DAG, VA, Flags));
1453 if (!MemOpChains.empty())
1454 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1455 &MemOpChains[0], MemOpChains.size());
1457 // Build a sequence of copy-to-reg nodes chained together with token chain
1458 // and flag operands which copy the outgoing args into the appropriate regs.
1460 // Tail call byval lowering might overwrite argument registers so in case of
1461 // tail call optimization the copies to registers are lowered later.
1463 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1464 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1465 RegsToPass[i].second, InFlag);
1466 InFlag = Chain.getValue(1);
1469 // For tail calls lower the arguments to the 'real' stack slot.
1471 // Force all the incoming stack arguments to be loaded from the stack
1472 // before any new outgoing arguments are stored to the stack, because the
1473 // outgoing stack slots may alias the incoming argument stack slots, and
1474 // the alias isn't otherwise explicit. This is slightly more conservative
1475 // than necessary, because it means that each store effectively depends
1476 // on every argument instead of just those arguments it would clobber.
1478 // Do not flag preceding copytoreg stuff together with the following stuff.
1480 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1481 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1482 RegsToPass[i].second, InFlag);
1483 InFlag = Chain.getValue(1);
1488 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1489 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1490 // node so that legalize doesn't hack it.
1491 bool isDirect = false;
1492 bool isARMFunc = false;
1493 bool isLocalARMFunc = false;
1494 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1496 if (EnableARMLongCalls) {
1497 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1498 && "long-calls with non-static relocation model!");
1499 // Handle a global address or an external symbol. If it's not one of
1500 // those, the target's already in a register, so we don't need to do
1502 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1503 const GlobalValue *GV = G->getGlobal();
1504 // Create a constant pool entry for the callee address
1505 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1506 ARMConstantPoolValue *CPV =
1507 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1509 // Get the address of the callee into a register
1510 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1511 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1512 Callee = DAG.getLoad(getPointerTy(), dl,
1513 DAG.getEntryNode(), CPAddr,
1514 MachinePointerInfo::getConstantPool(),
1515 false, false, false, 0);
1516 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1517 const char *Sym = S->getSymbol();
1519 // Create a constant pool entry for the callee address
1520 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1521 ARMConstantPoolValue *CPV =
1522 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1523 ARMPCLabelIndex, 0);
1524 // Get the address of the callee into a register
1525 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1526 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1527 Callee = DAG.getLoad(getPointerTy(), dl,
1528 DAG.getEntryNode(), CPAddr,
1529 MachinePointerInfo::getConstantPool(),
1530 false, false, false, 0);
1532 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1533 const GlobalValue *GV = G->getGlobal();
1535 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1536 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1537 getTargetMachine().getRelocationModel() != Reloc::Static;
1538 isARMFunc = !Subtarget->isThumb() || isStub;
1539 // ARM call to a local ARM function is predicable.
1540 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1541 // tBX takes a register source operand.
1542 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1543 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1544 ARMConstantPoolValue *CPV =
1545 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1546 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1547 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1548 Callee = DAG.getLoad(getPointerTy(), dl,
1549 DAG.getEntryNode(), CPAddr,
1550 MachinePointerInfo::getConstantPool(),
1551 false, false, false, 0);
1552 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1553 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1554 getPointerTy(), Callee, PICLabel);
1556 // On ELF targets for PIC code, direct calls should go through the PLT
1557 unsigned OpFlags = 0;
1558 if (Subtarget->isTargetELF() &&
1559 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1560 OpFlags = ARMII::MO_PLT;
1561 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1563 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1565 bool isStub = Subtarget->isTargetDarwin() &&
1566 getTargetMachine().getRelocationModel() != Reloc::Static;
1567 isARMFunc = !Subtarget->isThumb() || isStub;
1568 // tBX takes a register source operand.
1569 const char *Sym = S->getSymbol();
1570 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1571 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1572 ARMConstantPoolValue *CPV =
1573 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1574 ARMPCLabelIndex, 4);
1575 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1576 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1577 Callee = DAG.getLoad(getPointerTy(), dl,
1578 DAG.getEntryNode(), CPAddr,
1579 MachinePointerInfo::getConstantPool(),
1580 false, false, false, 0);
1581 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1582 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1583 getPointerTy(), Callee, PICLabel);
1585 unsigned OpFlags = 0;
1586 // On ELF targets for PIC code, direct calls should go through the PLT
1587 if (Subtarget->isTargetELF() &&
1588 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1589 OpFlags = ARMII::MO_PLT;
1590 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1594 // FIXME: handle tail calls differently.
1596 if (Subtarget->isThumb()) {
1597 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1598 CallOpc = ARMISD::CALL_NOLINK;
1599 else if (doesNotRet && isDirect && !isARMFunc &&
1600 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1601 // "mov lr, pc; b _foo" to avoid confusing the RSP
1602 CallOpc = ARMISD::CALL_NOLINK;
1604 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1606 if (!isDirect && !Subtarget->hasV5TOps()) {
1607 CallOpc = ARMISD::CALL_NOLINK;
1608 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1609 // "mov lr, pc; b _foo" to avoid confusing the RSP
1610 CallOpc = ARMISD::CALL_NOLINK;
1612 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1615 std::vector<SDValue> Ops;
1616 Ops.push_back(Chain);
1617 Ops.push_back(Callee);
1619 // Add argument registers to the end of the list so that they are known live
1621 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1622 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1623 RegsToPass[i].second.getValueType()));
1625 // Add a register mask operand representing the call-preserved registers.
1626 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1627 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1628 assert(Mask && "Missing call preserved mask for calling convention");
1629 Ops.push_back(DAG.getRegisterMask(Mask));
1631 if (InFlag.getNode())
1632 Ops.push_back(InFlag);
1634 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1636 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1638 // Returns a chain and a flag for retval copy to use.
1639 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1640 InFlag = Chain.getValue(1);
1642 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1643 DAG.getIntPtrConstant(0, true), InFlag);
1645 InFlag = Chain.getValue(1);
1647 // Handle result values, copying them out of physregs into vregs that we
1649 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1653 /// HandleByVal - Every parameter *after* a byval parameter is passed
1654 /// on the stack. Remember the next parameter register to allocate,
1655 /// and then confiscate the rest of the parameter registers to insure
1658 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1659 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1660 assert((State->getCallOrPrologue() == Prologue ||
1661 State->getCallOrPrologue() == Call) &&
1662 "unhandled ParmContext");
1663 if ((!State->isFirstByValRegValid()) &&
1664 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1665 State->setFirstByValReg(reg);
1666 // At a call site, a byval parameter that is split between
1667 // registers and memory needs its size truncated here. In a
1668 // function prologue, such byval parameters are reassembled in
1669 // memory, and are not truncated.
1670 if (State->getCallOrPrologue() == Call) {
1671 unsigned excess = 4 * (ARM::R4 - reg);
1672 assert(size >= excess && "expected larger existing stack allocation");
1676 // Confiscate any remaining parameter registers to preclude their
1677 // assignment to subsequent parameters.
1678 while (State->AllocateReg(GPRArgRegs, 4))
1682 /// MatchingStackOffset - Return true if the given stack call argument is
1683 /// already available in the same position (relatively) of the caller's
1684 /// incoming argument stack.
1686 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1687 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1688 const TargetInstrInfo *TII) {
1689 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1691 if (Arg.getOpcode() == ISD::CopyFromReg) {
1692 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1693 if (!TargetRegisterInfo::isVirtualRegister(VR))
1695 MachineInstr *Def = MRI->getVRegDef(VR);
1698 if (!Flags.isByVal()) {
1699 if (!TII->isLoadFromStackSlot(Def, FI))
1704 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1705 if (Flags.isByVal())
1706 // ByVal argument is passed in as a pointer but it's now being
1707 // dereferenced. e.g.
1708 // define @foo(%struct.X* %A) {
1709 // tail call @bar(%struct.X* byval %A)
1712 SDValue Ptr = Ld->getBasePtr();
1713 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1716 FI = FINode->getIndex();
1720 assert(FI != INT_MAX);
1721 if (!MFI->isFixedObjectIndex(FI))
1723 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1726 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1727 /// for tail call optimization. Targets which want to do tail call
1728 /// optimization should implement this function.
1730 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1731 CallingConv::ID CalleeCC,
1733 bool isCalleeStructRet,
1734 bool isCallerStructRet,
1735 const SmallVectorImpl<ISD::OutputArg> &Outs,
1736 const SmallVectorImpl<SDValue> &OutVals,
1737 const SmallVectorImpl<ISD::InputArg> &Ins,
1738 SelectionDAG& DAG) const {
1739 const Function *CallerF = DAG.getMachineFunction().getFunction();
1740 CallingConv::ID CallerCC = CallerF->getCallingConv();
1741 bool CCMatch = CallerCC == CalleeCC;
1743 // Look for obvious safe cases to perform tail call optimization that do not
1744 // require ABI changes. This is what gcc calls sibcall.
1746 // Do not sibcall optimize vararg calls unless the call site is not passing
1748 if (isVarArg && !Outs.empty())
1751 // Also avoid sibcall optimization if either caller or callee uses struct
1752 // return semantics.
1753 if (isCalleeStructRet || isCallerStructRet)
1756 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1757 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1758 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1759 // support in the assembler and linker to be used. This would need to be
1760 // fixed to fully support tail calls in Thumb1.
1762 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1763 // LR. This means if we need to reload LR, it takes an extra instructions,
1764 // which outweighs the value of the tail call; but here we don't know yet
1765 // whether LR is going to be used. Probably the right approach is to
1766 // generate the tail call here and turn it back into CALL/RET in
1767 // emitEpilogue if LR is used.
1769 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1770 // but we need to make sure there are enough registers; the only valid
1771 // registers are the 4 used for parameters. We don't currently do this
1773 if (Subtarget->isThumb1Only())
1776 // If the calling conventions do not match, then we'd better make sure the
1777 // results are returned in the same way as what the caller expects.
1779 SmallVector<CCValAssign, 16> RVLocs1;
1780 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1781 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1782 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1784 SmallVector<CCValAssign, 16> RVLocs2;
1785 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1786 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1787 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1789 if (RVLocs1.size() != RVLocs2.size())
1791 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1792 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1794 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1796 if (RVLocs1[i].isRegLoc()) {
1797 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1800 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1806 // If the callee takes no arguments then go on to check the results of the
1808 if (!Outs.empty()) {
1809 // Check if stack adjustment is needed. For now, do not do this if any
1810 // argument is passed on the stack.
1811 SmallVector<CCValAssign, 16> ArgLocs;
1812 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1813 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1814 CCInfo.AnalyzeCallOperands(Outs,
1815 CCAssignFnForNode(CalleeCC, false, isVarArg));
1816 if (CCInfo.getNextStackOffset()) {
1817 MachineFunction &MF = DAG.getMachineFunction();
1819 // Check if the arguments are already laid out in the right way as
1820 // the caller's fixed stack objects.
1821 MachineFrameInfo *MFI = MF.getFrameInfo();
1822 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1823 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1824 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1826 ++i, ++realArgIdx) {
1827 CCValAssign &VA = ArgLocs[i];
1828 EVT RegVT = VA.getLocVT();
1829 SDValue Arg = OutVals[realArgIdx];
1830 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1831 if (VA.getLocInfo() == CCValAssign::Indirect)
1833 if (VA.needsCustom()) {
1834 // f64 and vector types are split into multiple registers or
1835 // register/stack-slot combinations. The types will not match
1836 // the registers; give up on memory f64 refs until we figure
1837 // out what to do about this.
1840 if (!ArgLocs[++i].isRegLoc())
1842 if (RegVT == MVT::v2f64) {
1843 if (!ArgLocs[++i].isRegLoc())
1845 if (!ArgLocs[++i].isRegLoc())
1848 } else if (!VA.isRegLoc()) {
1849 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1861 ARMTargetLowering::LowerReturn(SDValue Chain,
1862 CallingConv::ID CallConv, bool isVarArg,
1863 const SmallVectorImpl<ISD::OutputArg> &Outs,
1864 const SmallVectorImpl<SDValue> &OutVals,
1865 DebugLoc dl, SelectionDAG &DAG) const {
1867 // CCValAssign - represent the assignment of the return value to a location.
1868 SmallVector<CCValAssign, 16> RVLocs;
1870 // CCState - Info about the registers and stack slots.
1871 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1872 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1874 // Analyze outgoing return values.
1875 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1878 // If this is the first return lowered for this function, add
1879 // the regs to the liveout set for the function.
1880 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1881 for (unsigned i = 0; i != RVLocs.size(); ++i)
1882 if (RVLocs[i].isRegLoc())
1883 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1888 // Copy the result values into the output registers.
1889 for (unsigned i = 0, realRVLocIdx = 0;
1891 ++i, ++realRVLocIdx) {
1892 CCValAssign &VA = RVLocs[i];
1893 assert(VA.isRegLoc() && "Can only return in registers!");
1895 SDValue Arg = OutVals[realRVLocIdx];
1897 switch (VA.getLocInfo()) {
1898 default: llvm_unreachable("Unknown loc info!");
1899 case CCValAssign::Full: break;
1900 case CCValAssign::BCvt:
1901 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1905 if (VA.needsCustom()) {
1906 if (VA.getLocVT() == MVT::v2f64) {
1907 // Extract the first half and return it in two registers.
1908 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1909 DAG.getConstant(0, MVT::i32));
1910 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1911 DAG.getVTList(MVT::i32, MVT::i32), Half);
1913 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1914 Flag = Chain.getValue(1);
1915 VA = RVLocs[++i]; // skip ahead to next loc
1916 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1917 HalfGPRs.getValue(1), Flag);
1918 Flag = Chain.getValue(1);
1919 VA = RVLocs[++i]; // skip ahead to next loc
1921 // Extract the 2nd half and fall through to handle it as an f64 value.
1922 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1923 DAG.getConstant(1, MVT::i32));
1925 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1927 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1928 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1929 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1930 Flag = Chain.getValue(1);
1931 VA = RVLocs[++i]; // skip ahead to next loc
1932 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1935 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1937 // Guarantee that all emitted copies are
1938 // stuck together, avoiding something bad.
1939 Flag = Chain.getValue(1);
1944 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1946 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1951 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1952 if (N->getNumValues() != 1)
1954 if (!N->hasNUsesOfValue(1, 0))
1957 SDValue TCChain = Chain;
1958 SDNode *Copy = *N->use_begin();
1959 if (Copy->getOpcode() == ISD::CopyToReg) {
1960 // If the copy has a glue operand, we conservatively assume it isn't safe to
1961 // perform a tail call.
1962 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1964 TCChain = Copy->getOperand(0);
1965 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1966 SDNode *VMov = Copy;
1967 // f64 returned in a pair of GPRs.
1968 SmallPtrSet<SDNode*, 2> Copies;
1969 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1971 if (UI->getOpcode() != ISD::CopyToReg)
1975 if (Copies.size() > 2)
1978 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1980 SDValue UseChain = UI->getOperand(0);
1981 if (Copies.count(UseChain.getNode()))
1988 } else if (Copy->getOpcode() == ISD::BITCAST) {
1989 // f32 returned in a single GPR.
1990 if (!Copy->hasOneUse())
1992 Copy = *Copy->use_begin();
1993 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
1995 Chain = Copy->getOperand(0);
2000 bool HasRet = false;
2001 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2003 if (UI->getOpcode() != ARMISD::RET_FLAG)
2015 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2016 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2019 if (!CI->isTailCall())
2022 return !Subtarget->isThumb1Only();
2025 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2026 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2027 // one of the above mentioned nodes. It has to be wrapped because otherwise
2028 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2029 // be used to form addressing mode. These wrapped nodes will be selected
2031 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2032 EVT PtrVT = Op.getValueType();
2033 // FIXME there is no actual debug info here
2034 DebugLoc dl = Op.getDebugLoc();
2035 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2037 if (CP->isMachineConstantPoolEntry())
2038 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2039 CP->getAlignment());
2041 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2042 CP->getAlignment());
2043 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2046 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2047 return MachineJumpTableInfo::EK_Inline;
2050 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2051 SelectionDAG &DAG) const {
2052 MachineFunction &MF = DAG.getMachineFunction();
2053 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2054 unsigned ARMPCLabelIndex = 0;
2055 DebugLoc DL = Op.getDebugLoc();
2056 EVT PtrVT = getPointerTy();
2057 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2058 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2060 if (RelocM == Reloc::Static) {
2061 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2063 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2064 ARMPCLabelIndex = AFI->createPICLabelUId();
2065 ARMConstantPoolValue *CPV =
2066 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2067 ARMCP::CPBlockAddress, PCAdj);
2068 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2070 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2071 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2072 MachinePointerInfo::getConstantPool(),
2073 false, false, false, 0);
2074 if (RelocM == Reloc::Static)
2076 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2077 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2080 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2082 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2083 SelectionDAG &DAG) const {
2084 DebugLoc dl = GA->getDebugLoc();
2085 EVT PtrVT = getPointerTy();
2086 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2087 MachineFunction &MF = DAG.getMachineFunction();
2088 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2089 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2090 ARMConstantPoolValue *CPV =
2091 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2092 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2093 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2094 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2095 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2096 MachinePointerInfo::getConstantPool(),
2097 false, false, false, 0);
2098 SDValue Chain = Argument.getValue(1);
2100 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2101 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2103 // call __tls_get_addr.
2106 Entry.Node = Argument;
2107 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2108 Args.push_back(Entry);
2109 // FIXME: is there useful debug info available here?
2110 TargetLowering::CallLoweringInfo CLI(Chain,
2111 (Type *) Type::getInt32Ty(*DAG.getContext()),
2112 false, false, false, false,
2113 0, CallingConv::C, /*isTailCall=*/false,
2114 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2115 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2116 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2117 return CallResult.first;
2120 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2121 // "local exec" model.
2123 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2125 TLSModel::Model model) const {
2126 const GlobalValue *GV = GA->getGlobal();
2127 DebugLoc dl = GA->getDebugLoc();
2129 SDValue Chain = DAG.getEntryNode();
2130 EVT PtrVT = getPointerTy();
2131 // Get the Thread Pointer
2132 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2134 if (model == TLSModel::InitialExec) {
2135 MachineFunction &MF = DAG.getMachineFunction();
2136 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2137 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2138 // Initial exec model.
2139 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2140 ARMConstantPoolValue *CPV =
2141 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2142 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2144 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2145 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2146 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2147 MachinePointerInfo::getConstantPool(),
2148 false, false, false, 0);
2149 Chain = Offset.getValue(1);
2151 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2152 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2154 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2155 MachinePointerInfo::getConstantPool(),
2156 false, false, false, 0);
2159 assert(model == TLSModel::LocalExec);
2160 ARMConstantPoolValue *CPV =
2161 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2162 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2163 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2164 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2165 MachinePointerInfo::getConstantPool(),
2166 false, false, false, 0);
2169 // The address of the thread local variable is the add of the thread
2170 // pointer with the offset of the variable.
2171 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2175 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2176 // TODO: implement the "local dynamic" model
2177 assert(Subtarget->isTargetELF() &&
2178 "TLS not implemented for non-ELF targets");
2179 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2181 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2184 case TLSModel::GeneralDynamic:
2185 case TLSModel::LocalDynamic:
2186 return LowerToTLSGeneralDynamicModel(GA, DAG);
2187 case TLSModel::InitialExec:
2188 case TLSModel::LocalExec:
2189 return LowerToTLSExecModels(GA, DAG, model);
2191 llvm_unreachable("bogus TLS model");
2194 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2195 SelectionDAG &DAG) const {
2196 EVT PtrVT = getPointerTy();
2197 DebugLoc dl = Op.getDebugLoc();
2198 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2199 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2200 if (RelocM == Reloc::PIC_) {
2201 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2202 ARMConstantPoolValue *CPV =
2203 ARMConstantPoolConstant::Create(GV,
2204 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2205 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2206 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2207 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2209 MachinePointerInfo::getConstantPool(),
2210 false, false, false, 0);
2211 SDValue Chain = Result.getValue(1);
2212 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2213 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2215 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2216 MachinePointerInfo::getGOT(),
2217 false, false, false, 0);
2221 // If we have T2 ops, we can materialize the address directly via movt/movw
2222 // pair. This is always cheaper.
2223 if (Subtarget->useMovt()) {
2225 // FIXME: Once remat is capable of dealing with instructions with register
2226 // operands, expand this into two nodes.
2227 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2228 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2230 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2231 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2232 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2233 MachinePointerInfo::getConstantPool(),
2234 false, false, false, 0);
2238 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2239 SelectionDAG &DAG) const {
2240 EVT PtrVT = getPointerTy();
2241 DebugLoc dl = Op.getDebugLoc();
2242 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2243 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2244 MachineFunction &MF = DAG.getMachineFunction();
2245 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2247 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2248 // update ARMFastISel::ARMMaterializeGV.
2249 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2251 // FIXME: Once remat is capable of dealing with instructions with register
2252 // operands, expand this into two nodes.
2253 if (RelocM == Reloc::Static)
2254 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2255 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2257 unsigned Wrapper = (RelocM == Reloc::PIC_)
2258 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2259 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2260 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2261 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2262 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2263 MachinePointerInfo::getGOT(),
2264 false, false, false, 0);
2268 unsigned ARMPCLabelIndex = 0;
2270 if (RelocM == Reloc::Static) {
2271 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2273 ARMPCLabelIndex = AFI->createPICLabelUId();
2274 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2275 ARMConstantPoolValue *CPV =
2276 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2278 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2280 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2282 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2283 MachinePointerInfo::getConstantPool(),
2284 false, false, false, 0);
2285 SDValue Chain = Result.getValue(1);
2287 if (RelocM == Reloc::PIC_) {
2288 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2289 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2292 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2293 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2294 false, false, false, 0);
2299 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2300 SelectionDAG &DAG) const {
2301 assert(Subtarget->isTargetELF() &&
2302 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2303 MachineFunction &MF = DAG.getMachineFunction();
2304 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2305 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2306 EVT PtrVT = getPointerTy();
2307 DebugLoc dl = Op.getDebugLoc();
2308 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2309 ARMConstantPoolValue *CPV =
2310 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2311 ARMPCLabelIndex, PCAdj);
2312 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2313 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2314 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2315 MachinePointerInfo::getConstantPool(),
2316 false, false, false, 0);
2317 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2318 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2322 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2323 DebugLoc dl = Op.getDebugLoc();
2324 SDValue Val = DAG.getConstant(0, MVT::i32);
2325 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2326 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2327 Op.getOperand(1), Val);
2331 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2332 DebugLoc dl = Op.getDebugLoc();
2333 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2334 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2338 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2339 const ARMSubtarget *Subtarget) const {
2340 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2341 DebugLoc dl = Op.getDebugLoc();
2343 default: return SDValue(); // Don't custom lower most intrinsics.
2344 case Intrinsic::arm_thread_pointer: {
2345 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2346 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2348 case Intrinsic::eh_sjlj_lsda: {
2349 MachineFunction &MF = DAG.getMachineFunction();
2350 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2351 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2352 EVT PtrVT = getPointerTy();
2353 DebugLoc dl = Op.getDebugLoc();
2354 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2356 unsigned PCAdj = (RelocM != Reloc::PIC_)
2357 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2358 ARMConstantPoolValue *CPV =
2359 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2360 ARMCP::CPLSDA, PCAdj);
2361 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2362 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2364 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2365 MachinePointerInfo::getConstantPool(),
2366 false, false, false, 0);
2368 if (RelocM == Reloc::PIC_) {
2369 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2370 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2374 case Intrinsic::arm_neon_vmulls:
2375 case Intrinsic::arm_neon_vmullu: {
2376 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2377 ? ARMISD::VMULLs : ARMISD::VMULLu;
2378 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2379 Op.getOperand(1), Op.getOperand(2));
2384 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2385 const ARMSubtarget *Subtarget) {
2386 DebugLoc dl = Op.getDebugLoc();
2387 if (!Subtarget->hasDataBarrier()) {
2388 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2389 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2391 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2392 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2393 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2394 DAG.getConstant(0, MVT::i32));
2397 SDValue Op5 = Op.getOperand(5);
2398 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2399 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2400 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2401 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2403 ARM_MB::MemBOpt DMBOpt;
2404 if (isDeviceBarrier)
2405 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2407 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2408 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2409 DAG.getConstant(DMBOpt, MVT::i32));
2413 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2414 const ARMSubtarget *Subtarget) {
2415 // FIXME: handle "fence singlethread" more efficiently.
2416 DebugLoc dl = Op.getDebugLoc();
2417 if (!Subtarget->hasDataBarrier()) {
2418 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2419 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2421 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2422 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2423 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2424 DAG.getConstant(0, MVT::i32));
2427 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2428 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2431 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2432 const ARMSubtarget *Subtarget) {
2433 // ARM pre v5TE and Thumb1 does not have preload instructions.
2434 if (!(Subtarget->isThumb2() ||
2435 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2436 // Just preserve the chain.
2437 return Op.getOperand(0);
2439 DebugLoc dl = Op.getDebugLoc();
2440 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2442 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2443 // ARMv7 with MP extension has PLDW.
2444 return Op.getOperand(0);
2446 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2447 if (Subtarget->isThumb()) {
2449 isRead = ~isRead & 1;
2450 isData = ~isData & 1;
2453 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2454 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2455 DAG.getConstant(isData, MVT::i32));
2458 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2459 MachineFunction &MF = DAG.getMachineFunction();
2460 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2462 // vastart just stores the address of the VarArgsFrameIndex slot into the
2463 // memory location argument.
2464 DebugLoc dl = Op.getDebugLoc();
2465 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2466 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2467 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2468 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2469 MachinePointerInfo(SV), false, false, 0);
2473 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2474 SDValue &Root, SelectionDAG &DAG,
2475 DebugLoc dl) const {
2476 MachineFunction &MF = DAG.getMachineFunction();
2477 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2479 const TargetRegisterClass *RC;
2480 if (AFI->isThumb1OnlyFunction())
2481 RC = &ARM::tGPRRegClass;
2483 RC = &ARM::GPRRegClass;
2485 // Transform the arguments stored in physical registers into virtual ones.
2486 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2487 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2490 if (NextVA.isMemLoc()) {
2491 MachineFrameInfo *MFI = MF.getFrameInfo();
2492 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2494 // Create load node to retrieve arguments from the stack.
2495 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2496 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2497 MachinePointerInfo::getFixedStack(FI),
2498 false, false, false, 0);
2500 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2501 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2504 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2508 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2509 unsigned &VARegSize, unsigned &VARegSaveSize)
2512 if (CCInfo.isFirstByValRegValid())
2513 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2515 unsigned int firstUnalloced;
2516 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2517 sizeof(GPRArgRegs) /
2518 sizeof(GPRArgRegs[0]));
2519 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2522 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2523 VARegSize = NumGPRs * 4;
2524 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2527 // The remaining GPRs hold either the beginning of variable-argument
2528 // data, or the beginning of an aggregate passed by value (usuall
2529 // byval). Either way, we allocate stack slots adjacent to the data
2530 // provided by our caller, and store the unallocated registers there.
2531 // If this is a variadic function, the va_list pointer will begin with
2532 // these values; otherwise, this reassembles a (byval) structure that
2533 // was split between registers and memory.
2535 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2536 DebugLoc dl, SDValue &Chain,
2537 const Value *OrigArg,
2538 unsigned OffsetFromOrigArg,
2539 unsigned ArgOffset) const {
2540 MachineFunction &MF = DAG.getMachineFunction();
2541 MachineFrameInfo *MFI = MF.getFrameInfo();
2542 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2543 unsigned firstRegToSaveIndex;
2544 if (CCInfo.isFirstByValRegValid())
2545 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2547 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2548 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2551 unsigned VARegSize, VARegSaveSize;
2552 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2553 if (VARegSaveSize) {
2554 // If this function is vararg, store any remaining integer argument regs
2555 // to their spots on the stack so that they may be loaded by deferencing
2556 // the result of va_next.
2557 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2558 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2559 ArgOffset + VARegSaveSize
2562 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2565 SmallVector<SDValue, 4> MemOps;
2566 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2567 const TargetRegisterClass *RC;
2568 if (AFI->isThumb1OnlyFunction())
2569 RC = &ARM::tGPRRegClass;
2571 RC = &ARM::GPRRegClass;
2573 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2574 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2576 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2577 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2579 MemOps.push_back(Store);
2580 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2581 DAG.getConstant(4, getPointerTy()));
2583 if (!MemOps.empty())
2584 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2585 &MemOps[0], MemOps.size());
2587 // This will point to the next argument passed via stack.
2588 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2592 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2593 CallingConv::ID CallConv, bool isVarArg,
2594 const SmallVectorImpl<ISD::InputArg>
2596 DebugLoc dl, SelectionDAG &DAG,
2597 SmallVectorImpl<SDValue> &InVals)
2599 MachineFunction &MF = DAG.getMachineFunction();
2600 MachineFrameInfo *MFI = MF.getFrameInfo();
2602 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2604 // Assign locations to all of the incoming arguments.
2605 SmallVector<CCValAssign, 16> ArgLocs;
2606 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2607 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2608 CCInfo.AnalyzeFormalArguments(Ins,
2609 CCAssignFnForNode(CallConv, /* Return*/ false,
2612 SmallVector<SDValue, 16> ArgValues;
2613 int lastInsIndex = -1;
2615 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2616 unsigned CurArgIdx = 0;
2617 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2618 CCValAssign &VA = ArgLocs[i];
2619 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2620 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2621 // Arguments stored in registers.
2622 if (VA.isRegLoc()) {
2623 EVT RegVT = VA.getLocVT();
2625 if (VA.needsCustom()) {
2626 // f64 and vector types are split up into multiple registers or
2627 // combinations of registers and stack slots.
2628 if (VA.getLocVT() == MVT::v2f64) {
2629 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2631 VA = ArgLocs[++i]; // skip ahead to next loc
2633 if (VA.isMemLoc()) {
2634 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2635 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2636 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2637 MachinePointerInfo::getFixedStack(FI),
2638 false, false, false, 0);
2640 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2643 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2644 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2645 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2646 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2647 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2649 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2652 const TargetRegisterClass *RC;
2654 if (RegVT == MVT::f32)
2655 RC = &ARM::SPRRegClass;
2656 else if (RegVT == MVT::f64)
2657 RC = &ARM::DPRRegClass;
2658 else if (RegVT == MVT::v2f64)
2659 RC = &ARM::QPRRegClass;
2660 else if (RegVT == MVT::i32)
2661 RC = AFI->isThumb1OnlyFunction() ?
2662 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2663 (const TargetRegisterClass*)&ARM::GPRRegClass;
2665 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2667 // Transform the arguments in physical registers into virtual ones.
2668 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2669 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2672 // If this is an 8 or 16-bit value, it is really passed promoted
2673 // to 32 bits. Insert an assert[sz]ext to capture this, then
2674 // truncate to the right size.
2675 switch (VA.getLocInfo()) {
2676 default: llvm_unreachable("Unknown loc info!");
2677 case CCValAssign::Full: break;
2678 case CCValAssign::BCvt:
2679 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2681 case CCValAssign::SExt:
2682 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2683 DAG.getValueType(VA.getValVT()));
2684 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2686 case CCValAssign::ZExt:
2687 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2688 DAG.getValueType(VA.getValVT()));
2689 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2693 InVals.push_back(ArgValue);
2695 } else { // VA.isRegLoc()
2698 assert(VA.isMemLoc());
2699 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2701 int index = ArgLocs[i].getValNo();
2703 // Some Ins[] entries become multiple ArgLoc[] entries.
2704 // Process them only once.
2705 if (index != lastInsIndex)
2707 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2708 // FIXME: For now, all byval parameter objects are marked mutable.
2709 // This can be changed with more analysis.
2710 // In case of tail call optimization mark all arguments mutable.
2711 // Since they could be overwritten by lowering of arguments in case of
2713 if (Flags.isByVal()) {
2714 unsigned VARegSize, VARegSaveSize;
2715 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2716 VarArgStyleRegisters(CCInfo, DAG,
2717 dl, Chain, CurOrigArg, Ins[VA.getValNo()].PartOffset, 0);
2718 unsigned Bytes = Flags.getByValSize() - VARegSize;
2719 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2720 int FI = MFI->CreateFixedObject(Bytes,
2721 VA.getLocMemOffset(), false);
2722 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2724 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2725 VA.getLocMemOffset(), true);
2727 // Create load nodes to retrieve arguments from the stack.
2728 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2729 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2730 MachinePointerInfo::getFixedStack(FI),
2731 false, false, false, 0));
2733 lastInsIndex = index;
2740 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2741 CCInfo.getNextStackOffset());
2746 /// isFloatingPointZero - Return true if this is +0.0.
2747 static bool isFloatingPointZero(SDValue Op) {
2748 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2749 return CFP->getValueAPF().isPosZero();
2750 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2751 // Maybe this has already been legalized into the constant pool?
2752 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2753 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2754 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2755 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2756 return CFP->getValueAPF().isPosZero();
2762 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2763 /// the given operands.
2765 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2766 SDValue &ARMcc, SelectionDAG &DAG,
2767 DebugLoc dl) const {
2768 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2769 unsigned C = RHSC->getZExtValue();
2770 if (!isLegalICmpImmediate(C)) {
2771 // Constant does not fit, try adjusting it by one?
2776 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2777 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2778 RHS = DAG.getConstant(C-1, MVT::i32);
2783 if (C != 0 && isLegalICmpImmediate(C-1)) {
2784 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2785 RHS = DAG.getConstant(C-1, MVT::i32);
2790 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2791 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2792 RHS = DAG.getConstant(C+1, MVT::i32);
2797 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2798 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2799 RHS = DAG.getConstant(C+1, MVT::i32);
2806 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2807 ARMISD::NodeType CompareType;
2810 CompareType = ARMISD::CMP;
2815 CompareType = ARMISD::CMPZ;
2818 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2819 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2822 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2824 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2825 DebugLoc dl) const {
2827 if (!isFloatingPointZero(RHS))
2828 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2830 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2831 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2834 /// duplicateCmp - Glue values can have only one use, so this function
2835 /// duplicates a comparison node.
2837 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2838 unsigned Opc = Cmp.getOpcode();
2839 DebugLoc DL = Cmp.getDebugLoc();
2840 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2841 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2843 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2844 Cmp = Cmp.getOperand(0);
2845 Opc = Cmp.getOpcode();
2846 if (Opc == ARMISD::CMPFP)
2847 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2849 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2850 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2852 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2855 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2856 SDValue Cond = Op.getOperand(0);
2857 SDValue SelectTrue = Op.getOperand(1);
2858 SDValue SelectFalse = Op.getOperand(2);
2859 DebugLoc dl = Op.getDebugLoc();
2863 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2864 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2866 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2867 const ConstantSDNode *CMOVTrue =
2868 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2869 const ConstantSDNode *CMOVFalse =
2870 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2872 if (CMOVTrue && CMOVFalse) {
2873 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2874 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2878 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2880 False = SelectFalse;
2881 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2886 if (True.getNode() && False.getNode()) {
2887 EVT VT = Op.getValueType();
2888 SDValue ARMcc = Cond.getOperand(2);
2889 SDValue CCR = Cond.getOperand(3);
2890 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2891 assert(True.getValueType() == VT);
2892 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2897 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2898 // undefined bits before doing a full-word comparison with zero.
2899 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2900 DAG.getConstant(1, Cond.getValueType()));
2902 return DAG.getSelectCC(dl, Cond,
2903 DAG.getConstant(0, Cond.getValueType()),
2904 SelectTrue, SelectFalse, ISD::SETNE);
2907 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2908 EVT VT = Op.getValueType();
2909 SDValue LHS = Op.getOperand(0);
2910 SDValue RHS = Op.getOperand(1);
2911 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2912 SDValue TrueVal = Op.getOperand(2);
2913 SDValue FalseVal = Op.getOperand(3);
2914 DebugLoc dl = Op.getDebugLoc();
2916 if (LHS.getValueType() == MVT::i32) {
2918 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2919 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2920 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2923 ARMCC::CondCodes CondCode, CondCode2;
2924 FPCCToARMCC(CC, CondCode, CondCode2);
2926 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2927 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2928 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2929 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2931 if (CondCode2 != ARMCC::AL) {
2932 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2933 // FIXME: Needs another CMP because flag can have but one use.
2934 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2935 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2936 Result, TrueVal, ARMcc2, CCR, Cmp2);
2941 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2942 /// to morph to an integer compare sequence.
2943 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2944 const ARMSubtarget *Subtarget) {
2945 SDNode *N = Op.getNode();
2946 if (!N->hasOneUse())
2947 // Otherwise it requires moving the value from fp to integer registers.
2949 if (!N->getNumValues())
2951 EVT VT = Op.getValueType();
2952 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2953 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2954 // vmrs are very slow, e.g. cortex-a8.
2957 if (isFloatingPointZero(Op)) {
2961 return ISD::isNormalLoad(N);
2964 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2965 if (isFloatingPointZero(Op))
2966 return DAG.getConstant(0, MVT::i32);
2968 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2969 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2970 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2971 Ld->isVolatile(), Ld->isNonTemporal(),
2972 Ld->isInvariant(), Ld->getAlignment());
2974 llvm_unreachable("Unknown VFP cmp argument!");
2977 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2978 SDValue &RetVal1, SDValue &RetVal2) {
2979 if (isFloatingPointZero(Op)) {
2980 RetVal1 = DAG.getConstant(0, MVT::i32);
2981 RetVal2 = DAG.getConstant(0, MVT::i32);
2985 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2986 SDValue Ptr = Ld->getBasePtr();
2987 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2988 Ld->getChain(), Ptr,
2989 Ld->getPointerInfo(),
2990 Ld->isVolatile(), Ld->isNonTemporal(),
2991 Ld->isInvariant(), Ld->getAlignment());
2993 EVT PtrType = Ptr.getValueType();
2994 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2995 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2996 PtrType, Ptr, DAG.getConstant(4, PtrType));
2997 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2998 Ld->getChain(), NewPtr,
2999 Ld->getPointerInfo().getWithOffset(4),
3000 Ld->isVolatile(), Ld->isNonTemporal(),
3001 Ld->isInvariant(), NewAlign);
3005 llvm_unreachable("Unknown VFP cmp argument!");
3008 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3009 /// f32 and even f64 comparisons to integer ones.
3011 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3012 SDValue Chain = Op.getOperand(0);
3013 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3014 SDValue LHS = Op.getOperand(2);
3015 SDValue RHS = Op.getOperand(3);
3016 SDValue Dest = Op.getOperand(4);
3017 DebugLoc dl = Op.getDebugLoc();
3019 bool LHSSeenZero = false;
3020 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3021 bool RHSSeenZero = false;
3022 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3023 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3024 // If unsafe fp math optimization is enabled and there are no other uses of
3025 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3026 // to an integer comparison.
3027 if (CC == ISD::SETOEQ)
3029 else if (CC == ISD::SETUNE)
3032 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3034 if (LHS.getValueType() == MVT::f32) {
3035 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3036 bitcastf32Toi32(LHS, DAG), Mask);
3037 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3038 bitcastf32Toi32(RHS, DAG), Mask);
3039 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3040 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3041 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3042 Chain, Dest, ARMcc, CCR, Cmp);
3047 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3048 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3049 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3050 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3051 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3052 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3053 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3054 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3055 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3061 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3062 SDValue Chain = Op.getOperand(0);
3063 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3064 SDValue LHS = Op.getOperand(2);
3065 SDValue RHS = Op.getOperand(3);
3066 SDValue Dest = Op.getOperand(4);
3067 DebugLoc dl = Op.getDebugLoc();
3069 if (LHS.getValueType() == MVT::i32) {
3071 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3072 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3073 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3074 Chain, Dest, ARMcc, CCR, Cmp);
3077 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3079 if (getTargetMachine().Options.UnsafeFPMath &&
3080 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3081 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3082 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3083 if (Result.getNode())
3087 ARMCC::CondCodes CondCode, CondCode2;
3088 FPCCToARMCC(CC, CondCode, CondCode2);
3090 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3091 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3092 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3093 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3094 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3095 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3096 if (CondCode2 != ARMCC::AL) {
3097 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3098 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3099 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3104 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3105 SDValue Chain = Op.getOperand(0);
3106 SDValue Table = Op.getOperand(1);
3107 SDValue Index = Op.getOperand(2);
3108 DebugLoc dl = Op.getDebugLoc();
3110 EVT PTy = getPointerTy();
3111 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3112 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3113 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3114 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3115 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3116 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3117 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3118 if (Subtarget->isThumb2()) {
3119 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3120 // which does another jump to the destination. This also makes it easier
3121 // to translate it to TBB / TBH later.
3122 // FIXME: This might not work if the function is extremely large.
3123 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3124 Addr, Op.getOperand(2), JTI, UId);
3126 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3127 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3128 MachinePointerInfo::getJumpTable(),
3129 false, false, false, 0);
3130 Chain = Addr.getValue(1);
3131 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3132 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3134 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3135 MachinePointerInfo::getJumpTable(),
3136 false, false, false, 0);
3137 Chain = Addr.getValue(1);
3138 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3142 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3143 EVT VT = Op.getValueType();
3144 DebugLoc dl = Op.getDebugLoc();
3146 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3147 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3149 return DAG.UnrollVectorOp(Op.getNode());
3152 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3153 "Invalid type for custom lowering!");
3154 if (VT != MVT::v4i16)
3155 return DAG.UnrollVectorOp(Op.getNode());
3157 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3158 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3161 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3162 EVT VT = Op.getValueType();
3164 return LowerVectorFP_TO_INT(Op, DAG);
3166 DebugLoc dl = Op.getDebugLoc();
3169 switch (Op.getOpcode()) {
3170 default: llvm_unreachable("Invalid opcode!");
3171 case ISD::FP_TO_SINT:
3172 Opc = ARMISD::FTOSI;
3174 case ISD::FP_TO_UINT:
3175 Opc = ARMISD::FTOUI;
3178 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3179 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3182 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3183 EVT VT = Op.getValueType();
3184 DebugLoc dl = Op.getDebugLoc();
3186 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3187 if (VT.getVectorElementType() == MVT::f32)
3189 return DAG.UnrollVectorOp(Op.getNode());
3192 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3193 "Invalid type for custom lowering!");
3194 if (VT != MVT::v4f32)
3195 return DAG.UnrollVectorOp(Op.getNode());
3199 switch (Op.getOpcode()) {
3200 default: llvm_unreachable("Invalid opcode!");
3201 case ISD::SINT_TO_FP:
3202 CastOpc = ISD::SIGN_EXTEND;
3203 Opc = ISD::SINT_TO_FP;
3205 case ISD::UINT_TO_FP:
3206 CastOpc = ISD::ZERO_EXTEND;
3207 Opc = ISD::UINT_TO_FP;
3211 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3212 return DAG.getNode(Opc, dl, VT, Op);
3215 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3216 EVT VT = Op.getValueType();
3218 return LowerVectorINT_TO_FP(Op, DAG);
3220 DebugLoc dl = Op.getDebugLoc();
3223 switch (Op.getOpcode()) {
3224 default: llvm_unreachable("Invalid opcode!");
3225 case ISD::SINT_TO_FP:
3226 Opc = ARMISD::SITOF;
3228 case ISD::UINT_TO_FP:
3229 Opc = ARMISD::UITOF;
3233 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3234 return DAG.getNode(Opc, dl, VT, Op);
3237 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3238 // Implement fcopysign with a fabs and a conditional fneg.
3239 SDValue Tmp0 = Op.getOperand(0);
3240 SDValue Tmp1 = Op.getOperand(1);
3241 DebugLoc dl = Op.getDebugLoc();
3242 EVT VT = Op.getValueType();
3243 EVT SrcVT = Tmp1.getValueType();
3244 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3245 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3246 bool UseNEON = !InGPR && Subtarget->hasNEON();
3249 // Use VBSL to copy the sign bit.
3250 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3251 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3252 DAG.getTargetConstant(EncodedVal, MVT::i32));
3253 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3255 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3256 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3257 DAG.getConstant(32, MVT::i32));
3258 else /*if (VT == MVT::f32)*/
3259 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3260 if (SrcVT == MVT::f32) {
3261 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3263 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3264 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3265 DAG.getConstant(32, MVT::i32));
3266 } else if (VT == MVT::f32)
3267 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3268 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3269 DAG.getConstant(32, MVT::i32));
3270 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3271 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3273 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3275 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3276 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3277 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3279 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3280 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3281 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3282 if (VT == MVT::f32) {
3283 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3284 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3285 DAG.getConstant(0, MVT::i32));
3287 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3293 // Bitcast operand 1 to i32.
3294 if (SrcVT == MVT::f64)
3295 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3296 &Tmp1, 1).getValue(1);
3297 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3299 // Or in the signbit with integer operations.
3300 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3301 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3302 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3303 if (VT == MVT::f32) {
3304 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3305 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3306 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3307 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3310 // f64: Or the high part with signbit and then combine two parts.
3311 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3313 SDValue Lo = Tmp0.getValue(0);
3314 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3315 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3316 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3319 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3320 MachineFunction &MF = DAG.getMachineFunction();
3321 MachineFrameInfo *MFI = MF.getFrameInfo();
3322 MFI->setReturnAddressIsTaken(true);
3324 EVT VT = Op.getValueType();
3325 DebugLoc dl = Op.getDebugLoc();
3326 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3328 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3329 SDValue Offset = DAG.getConstant(4, MVT::i32);
3330 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3331 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3332 MachinePointerInfo(), false, false, false, 0);
3335 // Return LR, which contains the return address. Mark it an implicit live-in.
3336 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3337 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3340 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3341 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3342 MFI->setFrameAddressIsTaken(true);
3344 EVT VT = Op.getValueType();
3345 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3346 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3347 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3348 ? ARM::R7 : ARM::R11;
3349 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3351 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3352 MachinePointerInfo(),
3353 false, false, false, 0);
3357 /// ExpandBITCAST - If the target supports VFP, this function is called to
3358 /// expand a bit convert where either the source or destination type is i64 to
3359 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3360 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3361 /// vectors), since the legalizer won't know what to do with that.
3362 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3363 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3364 DebugLoc dl = N->getDebugLoc();
3365 SDValue Op = N->getOperand(0);
3367 // This function is only supposed to be called for i64 types, either as the
3368 // source or destination of the bit convert.
3369 EVT SrcVT = Op.getValueType();
3370 EVT DstVT = N->getValueType(0);
3371 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3372 "ExpandBITCAST called for non-i64 type");
3374 // Turn i64->f64 into VMOVDRR.
3375 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3376 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3377 DAG.getConstant(0, MVT::i32));
3378 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3379 DAG.getConstant(1, MVT::i32));
3380 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3381 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3384 // Turn f64->i64 into VMOVRRD.
3385 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3386 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3387 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3388 // Merge the pieces into a single i64 value.
3389 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3395 /// getZeroVector - Returns a vector of specified type with all zero elements.
3396 /// Zero vectors are used to represent vector negation and in those cases
3397 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3398 /// not support i64 elements, so sometimes the zero vectors will need to be
3399 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3401 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3402 assert(VT.isVector() && "Expected a vector type");
3403 // The canonical modified immediate encoding of a zero vector is....0!
3404 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3405 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3406 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3407 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3410 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3411 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3412 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3413 SelectionDAG &DAG) const {
3414 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3415 EVT VT = Op.getValueType();
3416 unsigned VTBits = VT.getSizeInBits();
3417 DebugLoc dl = Op.getDebugLoc();
3418 SDValue ShOpLo = Op.getOperand(0);
3419 SDValue ShOpHi = Op.getOperand(1);
3420 SDValue ShAmt = Op.getOperand(2);
3422 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3424 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3426 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3427 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3428 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3429 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3430 DAG.getConstant(VTBits, MVT::i32));
3431 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3432 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3433 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3435 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3436 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3438 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3439 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3442 SDValue Ops[2] = { Lo, Hi };
3443 return DAG.getMergeValues(Ops, 2, dl);
3446 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3447 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3448 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3449 SelectionDAG &DAG) const {
3450 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3451 EVT VT = Op.getValueType();
3452 unsigned VTBits = VT.getSizeInBits();
3453 DebugLoc dl = Op.getDebugLoc();
3454 SDValue ShOpLo = Op.getOperand(0);
3455 SDValue ShOpHi = Op.getOperand(1);
3456 SDValue ShAmt = Op.getOperand(2);
3459 assert(Op.getOpcode() == ISD::SHL_PARTS);
3460 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3461 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3462 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3463 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3464 DAG.getConstant(VTBits, MVT::i32));
3465 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3466 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3468 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3469 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3470 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3472 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3473 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3476 SDValue Ops[2] = { Lo, Hi };
3477 return DAG.getMergeValues(Ops, 2, dl);
3480 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3481 SelectionDAG &DAG) const {
3482 // The rounding mode is in bits 23:22 of the FPSCR.
3483 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3484 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3485 // so that the shift + and get folded into a bitfield extract.
3486 DebugLoc dl = Op.getDebugLoc();
3487 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3488 DAG.getConstant(Intrinsic::arm_get_fpscr,
3490 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3491 DAG.getConstant(1U << 22, MVT::i32));
3492 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3493 DAG.getConstant(22, MVT::i32));
3494 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3495 DAG.getConstant(3, MVT::i32));
3498 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3499 const ARMSubtarget *ST) {
3500 EVT VT = N->getValueType(0);
3501 DebugLoc dl = N->getDebugLoc();
3503 if (!ST->hasV6T2Ops())
3506 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3507 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3510 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3511 const ARMSubtarget *ST) {
3512 EVT VT = N->getValueType(0);
3513 DebugLoc dl = N->getDebugLoc();
3518 // Lower vector shifts on NEON to use VSHL.
3519 assert(ST->hasNEON() && "unexpected vector shift");
3521 // Left shifts translate directly to the vshiftu intrinsic.
3522 if (N->getOpcode() == ISD::SHL)
3523 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3524 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3525 N->getOperand(0), N->getOperand(1));
3527 assert((N->getOpcode() == ISD::SRA ||
3528 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3530 // NEON uses the same intrinsics for both left and right shifts. For
3531 // right shifts, the shift amounts are negative, so negate the vector of
3533 EVT ShiftVT = N->getOperand(1).getValueType();
3534 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3535 getZeroVector(ShiftVT, DAG, dl),
3537 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3538 Intrinsic::arm_neon_vshifts :
3539 Intrinsic::arm_neon_vshiftu);
3540 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3541 DAG.getConstant(vshiftInt, MVT::i32),
3542 N->getOperand(0), NegatedCount);
3545 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3546 const ARMSubtarget *ST) {
3547 EVT VT = N->getValueType(0);
3548 DebugLoc dl = N->getDebugLoc();
3550 // We can get here for a node like i32 = ISD::SHL i32, i64
3554 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3555 "Unknown shift to lower!");
3557 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3558 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3559 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3562 // If we are in thumb mode, we don't have RRX.
3563 if (ST->isThumb1Only()) return SDValue();
3565 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3566 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3567 DAG.getConstant(0, MVT::i32));
3568 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3569 DAG.getConstant(1, MVT::i32));
3571 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3572 // captures the result into a carry flag.
3573 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3574 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3576 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3577 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3579 // Merge the pieces into a single i64 value.
3580 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3583 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3584 SDValue TmpOp0, TmpOp1;
3585 bool Invert = false;
3589 SDValue Op0 = Op.getOperand(0);
3590 SDValue Op1 = Op.getOperand(1);
3591 SDValue CC = Op.getOperand(2);
3592 EVT VT = Op.getValueType();
3593 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3594 DebugLoc dl = Op.getDebugLoc();
3596 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3597 switch (SetCCOpcode) {
3598 default: llvm_unreachable("Illegal FP comparison");
3600 case ISD::SETNE: Invert = true; // Fallthrough
3602 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3604 case ISD::SETLT: Swap = true; // Fallthrough
3606 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3608 case ISD::SETLE: Swap = true; // Fallthrough
3610 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3611 case ISD::SETUGE: Swap = true; // Fallthrough
3612 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3613 case ISD::SETUGT: Swap = true; // Fallthrough
3614 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3615 case ISD::SETUEQ: Invert = true; // Fallthrough
3617 // Expand this to (OLT | OGT).
3621 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3622 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3624 case ISD::SETUO: Invert = true; // Fallthrough
3626 // Expand this to (OLT | OGE).
3630 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3631 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3635 // Integer comparisons.
3636 switch (SetCCOpcode) {
3637 default: llvm_unreachable("Illegal integer comparison");
3638 case ISD::SETNE: Invert = true;
3639 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3640 case ISD::SETLT: Swap = true;
3641 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3642 case ISD::SETLE: Swap = true;
3643 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3644 case ISD::SETULT: Swap = true;
3645 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3646 case ISD::SETULE: Swap = true;
3647 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3650 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3651 if (Opc == ARMISD::VCEQ) {
3654 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3656 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3659 // Ignore bitconvert.
3660 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3661 AndOp = AndOp.getOperand(0);
3663 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3665 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3666 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3673 std::swap(Op0, Op1);
3675 // If one of the operands is a constant vector zero, attempt to fold the
3676 // comparison to a specialized compare-against-zero form.
3678 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3680 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3681 if (Opc == ARMISD::VCGE)
3682 Opc = ARMISD::VCLEZ;
3683 else if (Opc == ARMISD::VCGT)
3684 Opc = ARMISD::VCLTZ;
3689 if (SingleOp.getNode()) {
3692 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3694 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3696 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3698 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3700 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3702 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3705 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3709 Result = DAG.getNOT(dl, Result, VT);
3714 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3715 /// valid vector constant for a NEON instruction with a "modified immediate"
3716 /// operand (e.g., VMOV). If so, return the encoded value.
3717 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3718 unsigned SplatBitSize, SelectionDAG &DAG,
3719 EVT &VT, bool is128Bits, NEONModImmType type) {
3720 unsigned OpCmode, Imm;
3722 // SplatBitSize is set to the smallest size that splats the vector, so a
3723 // zero vector will always have SplatBitSize == 8. However, NEON modified
3724 // immediate instructions others than VMOV do not support the 8-bit encoding
3725 // of a zero vector, and the default encoding of zero is supposed to be the
3730 switch (SplatBitSize) {
3732 if (type != VMOVModImm)
3734 // Any 1-byte value is OK. Op=0, Cmode=1110.
3735 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3738 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3742 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3743 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3744 if ((SplatBits & ~0xff) == 0) {
3745 // Value = 0x00nn: Op=x, Cmode=100x.
3750 if ((SplatBits & ~0xff00) == 0) {
3751 // Value = 0xnn00: Op=x, Cmode=101x.
3753 Imm = SplatBits >> 8;
3759 // NEON's 32-bit VMOV supports splat values where:
3760 // * only one byte is nonzero, or
3761 // * the least significant byte is 0xff and the second byte is nonzero, or
3762 // * the least significant 2 bytes are 0xff and the third is nonzero.
3763 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3764 if ((SplatBits & ~0xff) == 0) {
3765 // Value = 0x000000nn: Op=x, Cmode=000x.
3770 if ((SplatBits & ~0xff00) == 0) {
3771 // Value = 0x0000nn00: Op=x, Cmode=001x.
3773 Imm = SplatBits >> 8;
3776 if ((SplatBits & ~0xff0000) == 0) {
3777 // Value = 0x00nn0000: Op=x, Cmode=010x.
3779 Imm = SplatBits >> 16;
3782 if ((SplatBits & ~0xff000000) == 0) {
3783 // Value = 0xnn000000: Op=x, Cmode=011x.
3785 Imm = SplatBits >> 24;
3789 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3790 if (type == OtherModImm) return SDValue();
3792 if ((SplatBits & ~0xffff) == 0 &&
3793 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3794 // Value = 0x0000nnff: Op=x, Cmode=1100.
3796 Imm = SplatBits >> 8;
3801 if ((SplatBits & ~0xffffff) == 0 &&
3802 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3803 // Value = 0x00nnffff: Op=x, Cmode=1101.
3805 Imm = SplatBits >> 16;
3806 SplatBits |= 0xffff;
3810 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3811 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3812 // VMOV.I32. A (very) minor optimization would be to replicate the value
3813 // and fall through here to test for a valid 64-bit splat. But, then the
3814 // caller would also need to check and handle the change in size.
3818 if (type != VMOVModImm)
3820 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3821 uint64_t BitMask = 0xff;
3823 unsigned ImmMask = 1;
3825 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3826 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3829 } else if ((SplatBits & BitMask) != 0) {
3835 // Op=1, Cmode=1110.
3838 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3843 llvm_unreachable("unexpected size for isNEONModifiedImm");
3846 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3847 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3850 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3851 const ARMSubtarget *ST) const {
3852 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3855 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3856 assert(Op.getValueType() == MVT::f32 &&
3857 "ConstantFP custom lowering should only occur for f32.");
3859 // Try splatting with a VMOV.f32...
3860 APFloat FPVal = CFP->getValueAPF();
3861 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3863 DebugLoc DL = Op.getDebugLoc();
3864 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3865 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3867 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3868 DAG.getConstant(0, MVT::i32));
3871 // If that fails, try a VMOV.i32
3873 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3874 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3876 if (NewVal != SDValue()) {
3877 DebugLoc DL = Op.getDebugLoc();
3878 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3880 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3882 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3883 DAG.getConstant(0, MVT::i32));
3886 // Finally, try a VMVN.i32
3887 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3889 if (NewVal != SDValue()) {
3890 DebugLoc DL = Op.getDebugLoc();
3891 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3892 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3894 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3895 DAG.getConstant(0, MVT::i32));
3902 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3903 bool &ReverseVEXT, unsigned &Imm) {
3904 unsigned NumElts = VT.getVectorNumElements();
3905 ReverseVEXT = false;
3907 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3913 // If this is a VEXT shuffle, the immediate value is the index of the first
3914 // element. The other shuffle indices must be the successive elements after
3916 unsigned ExpectedElt = Imm;
3917 for (unsigned i = 1; i < NumElts; ++i) {
3918 // Increment the expected index. If it wraps around, it may still be
3919 // a VEXT but the source vectors must be swapped.
3921 if (ExpectedElt == NumElts * 2) {
3926 if (M[i] < 0) continue; // ignore UNDEF indices
3927 if (ExpectedElt != static_cast<unsigned>(M[i]))
3931 // Adjust the index value if the source operands will be swapped.
3938 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3939 /// instruction with the specified blocksize. (The order of the elements
3940 /// within each block of the vector is reversed.)
3941 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3942 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3943 "Only possible block sizes for VREV are: 16, 32, 64");
3945 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3949 unsigned NumElts = VT.getVectorNumElements();
3950 unsigned BlockElts = M[0] + 1;
3951 // If the first shuffle index is UNDEF, be optimistic.
3953 BlockElts = BlockSize / EltSz;
3955 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3958 for (unsigned i = 0; i < NumElts; ++i) {
3959 if (M[i] < 0) continue; // ignore UNDEF indices
3960 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3967 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3968 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3969 // range, then 0 is placed into the resulting vector. So pretty much any mask
3970 // of 8 elements can work here.
3971 return VT == MVT::v8i8 && M.size() == 8;
3974 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3975 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3979 unsigned NumElts = VT.getVectorNumElements();
3980 WhichResult = (M[0] == 0 ? 0 : 1);
3981 for (unsigned i = 0; i < NumElts; i += 2) {
3982 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3983 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3989 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3990 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3991 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3992 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3993 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3997 unsigned NumElts = VT.getVectorNumElements();
3998 WhichResult = (M[0] == 0 ? 0 : 1);
3999 for (unsigned i = 0; i < NumElts; i += 2) {
4000 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4001 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4007 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4008 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4012 unsigned NumElts = VT.getVectorNumElements();
4013 WhichResult = (M[0] == 0 ? 0 : 1);
4014 for (unsigned i = 0; i != NumElts; ++i) {
4015 if (M[i] < 0) continue; // ignore UNDEF indices
4016 if ((unsigned) M[i] != 2 * i + WhichResult)
4020 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4021 if (VT.is64BitVector() && EltSz == 32)
4027 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4028 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4029 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4030 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4031 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4035 unsigned Half = VT.getVectorNumElements() / 2;
4036 WhichResult = (M[0] == 0 ? 0 : 1);
4037 for (unsigned j = 0; j != 2; ++j) {
4038 unsigned Idx = WhichResult;
4039 for (unsigned i = 0; i != Half; ++i) {
4040 int MIdx = M[i + j * Half];
4041 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4047 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4048 if (VT.is64BitVector() && EltSz == 32)
4054 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4055 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4059 unsigned NumElts = VT.getVectorNumElements();
4060 WhichResult = (M[0] == 0 ? 0 : 1);
4061 unsigned Idx = WhichResult * NumElts / 2;
4062 for (unsigned i = 0; i != NumElts; i += 2) {
4063 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4064 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4069 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4070 if (VT.is64BitVector() && EltSz == 32)
4076 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4077 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4078 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4079 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4080 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4084 unsigned NumElts = VT.getVectorNumElements();
4085 WhichResult = (M[0] == 0 ? 0 : 1);
4086 unsigned Idx = WhichResult * NumElts / 2;
4087 for (unsigned i = 0; i != NumElts; i += 2) {
4088 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4089 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4094 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4095 if (VT.is64BitVector() && EltSz == 32)
4101 // If N is an integer constant that can be moved into a register in one
4102 // instruction, return an SDValue of such a constant (will become a MOV
4103 // instruction). Otherwise return null.
4104 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4105 const ARMSubtarget *ST, DebugLoc dl) {
4107 if (!isa<ConstantSDNode>(N))
4109 Val = cast<ConstantSDNode>(N)->getZExtValue();
4111 if (ST->isThumb1Only()) {
4112 if (Val <= 255 || ~Val <= 255)
4113 return DAG.getConstant(Val, MVT::i32);
4115 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4116 return DAG.getConstant(Val, MVT::i32);
4121 // If this is a case we can't handle, return null and let the default
4122 // expansion code take care of it.
4123 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4124 const ARMSubtarget *ST) const {
4125 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4126 DebugLoc dl = Op.getDebugLoc();
4127 EVT VT = Op.getValueType();
4129 APInt SplatBits, SplatUndef;
4130 unsigned SplatBitSize;
4132 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4133 if (SplatBitSize <= 64) {
4134 // Check if an immediate VMOV works.
4136 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4137 SplatUndef.getZExtValue(), SplatBitSize,
4138 DAG, VmovVT, VT.is128BitVector(),
4140 if (Val.getNode()) {
4141 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4142 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4145 // Try an immediate VMVN.
4146 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4147 Val = isNEONModifiedImm(NegatedImm,
4148 SplatUndef.getZExtValue(), SplatBitSize,
4149 DAG, VmovVT, VT.is128BitVector(),
4151 if (Val.getNode()) {
4152 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4153 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4156 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4157 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4158 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4160 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4161 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4167 // Scan through the operands to see if only one value is used.
4169 // As an optimisation, even if more than one value is used it may be more
4170 // profitable to splat with one value then change some lanes.
4172 // Heuristically we decide to do this if the vector has a "dominant" value,
4173 // defined as splatted to more than half of the lanes.
4174 unsigned NumElts = VT.getVectorNumElements();
4175 bool isOnlyLowElement = true;
4176 bool usesOnlyOneValue = true;
4177 bool hasDominantValue = false;
4178 bool isConstant = true;
4180 // Map of the number of times a particular SDValue appears in the
4182 DenseMap<SDValue, unsigned> ValueCounts;
4184 for (unsigned i = 0; i < NumElts; ++i) {
4185 SDValue V = Op.getOperand(i);
4186 if (V.getOpcode() == ISD::UNDEF)
4189 isOnlyLowElement = false;
4190 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4193 ValueCounts.insert(std::make_pair(V, 0));
4194 unsigned &Count = ValueCounts[V];
4196 // Is this value dominant? (takes up more than half of the lanes)
4197 if (++Count > (NumElts / 2)) {
4198 hasDominantValue = true;
4202 if (ValueCounts.size() != 1)
4203 usesOnlyOneValue = false;
4204 if (!Value.getNode() && ValueCounts.size() > 0)
4205 Value = ValueCounts.begin()->first;
4207 if (ValueCounts.size() == 0)
4208 return DAG.getUNDEF(VT);
4210 if (isOnlyLowElement)
4211 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4213 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4215 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4216 // i32 and try again.
4217 if (hasDominantValue && EltSize <= 32) {
4221 // If we are VDUPing a value that comes directly from a vector, that will
4222 // cause an unnecessary move to and from a GPR, where instead we could
4223 // just use VDUPLANE.
4224 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT)
4225 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4226 Value->getOperand(0), Value->getOperand(1));
4228 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4230 if (!usesOnlyOneValue) {
4231 // The dominant value was splatted as 'N', but we now have to insert
4232 // all differing elements.
4233 for (unsigned I = 0; I < NumElts; ++I) {
4234 if (Op.getOperand(I) == Value)
4236 SmallVector<SDValue, 3> Ops;
4238 Ops.push_back(Op.getOperand(I));
4239 Ops.push_back(DAG.getConstant(I, MVT::i32));
4240 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4245 if (VT.getVectorElementType().isFloatingPoint()) {
4246 SmallVector<SDValue, 8> Ops;
4247 for (unsigned i = 0; i < NumElts; ++i)
4248 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4250 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4251 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4252 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4254 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4256 if (usesOnlyOneValue) {
4257 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4258 if (isConstant && Val.getNode())
4259 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4263 // If all elements are constants and the case above didn't get hit, fall back
4264 // to the default expansion, which will generate a load from the constant
4269 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4271 SDValue shuffle = ReconstructShuffle(Op, DAG);
4272 if (shuffle != SDValue())
4276 // Vectors with 32- or 64-bit elements can be built by directly assigning
4277 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4278 // will be legalized.
4279 if (EltSize >= 32) {
4280 // Do the expansion with floating-point types, since that is what the VFP
4281 // registers are defined to use, and since i64 is not legal.
4282 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4283 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4284 SmallVector<SDValue, 8> Ops;
4285 for (unsigned i = 0; i < NumElts; ++i)
4286 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4287 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4288 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4294 // Gather data to see if the operation can be modelled as a
4295 // shuffle in combination with VEXTs.
4296 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4297 SelectionDAG &DAG) const {
4298 DebugLoc dl = Op.getDebugLoc();
4299 EVT VT = Op.getValueType();
4300 unsigned NumElts = VT.getVectorNumElements();
4302 SmallVector<SDValue, 2> SourceVecs;
4303 SmallVector<unsigned, 2> MinElts;
4304 SmallVector<unsigned, 2> MaxElts;
4306 for (unsigned i = 0; i < NumElts; ++i) {
4307 SDValue V = Op.getOperand(i);
4308 if (V.getOpcode() == ISD::UNDEF)
4310 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4311 // A shuffle can only come from building a vector from various
4312 // elements of other vectors.
4314 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4315 VT.getVectorElementType()) {
4316 // This code doesn't know how to handle shuffles where the vector
4317 // element types do not match (this happens because type legalization
4318 // promotes the return type of EXTRACT_VECTOR_ELT).
4319 // FIXME: It might be appropriate to extend this code to handle
4320 // mismatched types.
4324 // Record this extraction against the appropriate vector if possible...
4325 SDValue SourceVec = V.getOperand(0);
4326 // If the element number isn't a constant, we can't effectively
4327 // analyze what's going on.
4328 if (!isa<ConstantSDNode>(V.getOperand(1)))
4330 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4331 bool FoundSource = false;
4332 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4333 if (SourceVecs[j] == SourceVec) {
4334 if (MinElts[j] > EltNo)
4336 if (MaxElts[j] < EltNo)
4343 // Or record a new source if not...
4345 SourceVecs.push_back(SourceVec);
4346 MinElts.push_back(EltNo);
4347 MaxElts.push_back(EltNo);
4351 // Currently only do something sane when at most two source vectors
4353 if (SourceVecs.size() > 2)
4356 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4357 int VEXTOffsets[2] = {0, 0};
4359 // This loop extracts the usage patterns of the source vectors
4360 // and prepares appropriate SDValues for a shuffle if possible.
4361 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4362 if (SourceVecs[i].getValueType() == VT) {
4363 // No VEXT necessary
4364 ShuffleSrcs[i] = SourceVecs[i];
4367 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4368 // It probably isn't worth padding out a smaller vector just to
4369 // break it down again in a shuffle.
4373 // Since only 64-bit and 128-bit vectors are legal on ARM and
4374 // we've eliminated the other cases...
4375 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4376 "unexpected vector sizes in ReconstructShuffle");
4378 if (MaxElts[i] - MinElts[i] >= NumElts) {
4379 // Span too large for a VEXT to cope
4383 if (MinElts[i] >= NumElts) {
4384 // The extraction can just take the second half
4385 VEXTOffsets[i] = NumElts;
4386 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4388 DAG.getIntPtrConstant(NumElts));
4389 } else if (MaxElts[i] < NumElts) {
4390 // The extraction can just take the first half
4392 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4394 DAG.getIntPtrConstant(0));
4396 // An actual VEXT is needed
4397 VEXTOffsets[i] = MinElts[i];
4398 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4400 DAG.getIntPtrConstant(0));
4401 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4403 DAG.getIntPtrConstant(NumElts));
4404 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4405 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4409 SmallVector<int, 8> Mask;
4411 for (unsigned i = 0; i < NumElts; ++i) {
4412 SDValue Entry = Op.getOperand(i);
4413 if (Entry.getOpcode() == ISD::UNDEF) {
4418 SDValue ExtractVec = Entry.getOperand(0);
4419 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4420 .getOperand(1))->getSExtValue();
4421 if (ExtractVec == SourceVecs[0]) {
4422 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4424 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4428 // Final check before we try to produce nonsense...
4429 if (isShuffleMaskLegal(Mask, VT))
4430 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4436 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4437 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4438 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4439 /// are assumed to be legal.
4441 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4443 if (VT.getVectorNumElements() == 4 &&
4444 (VT.is128BitVector() || VT.is64BitVector())) {
4445 unsigned PFIndexes[4];
4446 for (unsigned i = 0; i != 4; ++i) {
4450 PFIndexes[i] = M[i];
4453 // Compute the index in the perfect shuffle table.
4454 unsigned PFTableIndex =
4455 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4456 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4457 unsigned Cost = (PFEntry >> 30);
4464 unsigned Imm, WhichResult;
4466 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4467 return (EltSize >= 32 ||
4468 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4469 isVREVMask(M, VT, 64) ||
4470 isVREVMask(M, VT, 32) ||
4471 isVREVMask(M, VT, 16) ||
4472 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4473 isVTBLMask(M, VT) ||
4474 isVTRNMask(M, VT, WhichResult) ||
4475 isVUZPMask(M, VT, WhichResult) ||
4476 isVZIPMask(M, VT, WhichResult) ||
4477 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4478 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4479 isVZIP_v_undef_Mask(M, VT, WhichResult));
4482 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4483 /// the specified operations to build the shuffle.
4484 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4485 SDValue RHS, SelectionDAG &DAG,
4487 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4488 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4489 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4492 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4501 OP_VUZPL, // VUZP, left result
4502 OP_VUZPR, // VUZP, right result
4503 OP_VZIPL, // VZIP, left result
4504 OP_VZIPR, // VZIP, right result
4505 OP_VTRNL, // VTRN, left result
4506 OP_VTRNR // VTRN, right result
4509 if (OpNum == OP_COPY) {
4510 if (LHSID == (1*9+2)*9+3) return LHS;
4511 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4515 SDValue OpLHS, OpRHS;
4516 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4517 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4518 EVT VT = OpLHS.getValueType();
4521 default: llvm_unreachable("Unknown shuffle opcode!");
4523 // VREV divides the vector in half and swaps within the half.
4524 if (VT.getVectorElementType() == MVT::i32 ||
4525 VT.getVectorElementType() == MVT::f32)
4526 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4527 // vrev <4 x i16> -> VREV32
4528 if (VT.getVectorElementType() == MVT::i16)
4529 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4530 // vrev <4 x i8> -> VREV16
4531 assert(VT.getVectorElementType() == MVT::i8);
4532 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4537 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4538 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4542 return DAG.getNode(ARMISD::VEXT, dl, VT,
4544 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4547 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4548 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4551 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4552 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4555 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4556 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4560 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4561 ArrayRef<int> ShuffleMask,
4562 SelectionDAG &DAG) {
4563 // Check to see if we can use the VTBL instruction.
4564 SDValue V1 = Op.getOperand(0);
4565 SDValue V2 = Op.getOperand(1);
4566 DebugLoc DL = Op.getDebugLoc();
4568 SmallVector<SDValue, 8> VTBLMask;
4569 for (ArrayRef<int>::iterator
4570 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4571 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4573 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4574 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4575 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4578 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4579 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4583 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4584 SDValue V1 = Op.getOperand(0);
4585 SDValue V2 = Op.getOperand(1);
4586 DebugLoc dl = Op.getDebugLoc();
4587 EVT VT = Op.getValueType();
4588 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4590 // Convert shuffles that are directly supported on NEON to target-specific
4591 // DAG nodes, instead of keeping them as shuffles and matching them again
4592 // during code selection. This is more efficient and avoids the possibility
4593 // of inconsistencies between legalization and selection.
4594 // FIXME: floating-point vectors should be canonicalized to integer vectors
4595 // of the same time so that they get CSEd properly.
4596 ArrayRef<int> ShuffleMask = SVN->getMask();
4598 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4599 if (EltSize <= 32) {
4600 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4601 int Lane = SVN->getSplatIndex();
4602 // If this is undef splat, generate it via "just" vdup, if possible.
4603 if (Lane == -1) Lane = 0;
4605 // Test if V1 is a SCALAR_TO_VECTOR.
4606 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4607 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4609 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4610 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4612 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4613 !isa<ConstantSDNode>(V1.getOperand(0))) {
4614 bool IsScalarToVector = true;
4615 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4616 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4617 IsScalarToVector = false;
4620 if (IsScalarToVector)
4621 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4623 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4624 DAG.getConstant(Lane, MVT::i32));
4629 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4632 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4633 DAG.getConstant(Imm, MVT::i32));
4636 if (isVREVMask(ShuffleMask, VT, 64))
4637 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4638 if (isVREVMask(ShuffleMask, VT, 32))
4639 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4640 if (isVREVMask(ShuffleMask, VT, 16))
4641 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4643 // Check for Neon shuffles that modify both input vectors in place.
4644 // If both results are used, i.e., if there are two shuffles with the same
4645 // source operands and with masks corresponding to both results of one of
4646 // these operations, DAG memoization will ensure that a single node is
4647 // used for both shuffles.
4648 unsigned WhichResult;
4649 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4650 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4651 V1, V2).getValue(WhichResult);
4652 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4653 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4654 V1, V2).getValue(WhichResult);
4655 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4656 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4657 V1, V2).getValue(WhichResult);
4659 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4660 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4661 V1, V1).getValue(WhichResult);
4662 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4663 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4664 V1, V1).getValue(WhichResult);
4665 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4666 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4667 V1, V1).getValue(WhichResult);
4670 // If the shuffle is not directly supported and it has 4 elements, use
4671 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4672 unsigned NumElts = VT.getVectorNumElements();
4674 unsigned PFIndexes[4];
4675 for (unsigned i = 0; i != 4; ++i) {
4676 if (ShuffleMask[i] < 0)
4679 PFIndexes[i] = ShuffleMask[i];
4682 // Compute the index in the perfect shuffle table.
4683 unsigned PFTableIndex =
4684 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4685 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4686 unsigned Cost = (PFEntry >> 30);
4689 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4692 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4693 if (EltSize >= 32) {
4694 // Do the expansion with floating-point types, since that is what the VFP
4695 // registers are defined to use, and since i64 is not legal.
4696 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4697 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4698 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4699 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4700 SmallVector<SDValue, 8> Ops;
4701 for (unsigned i = 0; i < NumElts; ++i) {
4702 if (ShuffleMask[i] < 0)
4703 Ops.push_back(DAG.getUNDEF(EltVT));
4705 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4706 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4707 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4710 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4711 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4714 if (VT == MVT::v8i8) {
4715 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4716 if (NewOp.getNode())
4723 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4724 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4725 SDValue Lane = Op.getOperand(2);
4726 if (!isa<ConstantSDNode>(Lane))
4732 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4733 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4734 SDValue Lane = Op.getOperand(1);
4735 if (!isa<ConstantSDNode>(Lane))
4738 SDValue Vec = Op.getOperand(0);
4739 if (Op.getValueType() == MVT::i32 &&
4740 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4741 DebugLoc dl = Op.getDebugLoc();
4742 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4748 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4749 // The only time a CONCAT_VECTORS operation can have legal types is when
4750 // two 64-bit vectors are concatenated to a 128-bit vector.
4751 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4752 "unexpected CONCAT_VECTORS");
4753 DebugLoc dl = Op.getDebugLoc();
4754 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4755 SDValue Op0 = Op.getOperand(0);
4756 SDValue Op1 = Op.getOperand(1);
4757 if (Op0.getOpcode() != ISD::UNDEF)
4758 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4759 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4760 DAG.getIntPtrConstant(0));
4761 if (Op1.getOpcode() != ISD::UNDEF)
4762 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4763 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4764 DAG.getIntPtrConstant(1));
4765 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4768 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4769 /// element has been zero/sign-extended, depending on the isSigned parameter,
4770 /// from an integer type half its size.
4771 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4773 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4774 EVT VT = N->getValueType(0);
4775 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4776 SDNode *BVN = N->getOperand(0).getNode();
4777 if (BVN->getValueType(0) != MVT::v4i32 ||
4778 BVN->getOpcode() != ISD::BUILD_VECTOR)
4780 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4781 unsigned HiElt = 1 - LoElt;
4782 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4783 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4784 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4785 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4786 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4789 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4790 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4793 if (Hi0->isNullValue() && Hi1->isNullValue())
4799 if (N->getOpcode() != ISD::BUILD_VECTOR)
4802 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4803 SDNode *Elt = N->getOperand(i).getNode();
4804 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4805 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4806 unsigned HalfSize = EltSize / 2;
4808 if (!isIntN(HalfSize, C->getSExtValue()))
4811 if (!isUIntN(HalfSize, C->getZExtValue()))
4822 /// isSignExtended - Check if a node is a vector value that is sign-extended
4823 /// or a constant BUILD_VECTOR with sign-extended elements.
4824 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4825 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4827 if (isExtendedBUILD_VECTOR(N, DAG, true))
4832 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4833 /// or a constant BUILD_VECTOR with zero-extended elements.
4834 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4835 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4837 if (isExtendedBUILD_VECTOR(N, DAG, false))
4842 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4843 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4844 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4845 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4846 return N->getOperand(0);
4847 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4848 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4849 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4850 LD->isNonTemporal(), LD->isInvariant(),
4851 LD->getAlignment());
4852 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4853 // have been legalized as a BITCAST from v4i32.
4854 if (N->getOpcode() == ISD::BITCAST) {
4855 SDNode *BVN = N->getOperand(0).getNode();
4856 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4857 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4858 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4859 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4860 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4862 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4863 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4864 EVT VT = N->getValueType(0);
4865 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4866 unsigned NumElts = VT.getVectorNumElements();
4867 MVT TruncVT = MVT::getIntegerVT(EltSize);
4868 SmallVector<SDValue, 8> Ops;
4869 for (unsigned i = 0; i != NumElts; ++i) {
4870 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4871 const APInt &CInt = C->getAPIntValue();
4872 // Element types smaller than 32 bits are not legal, so use i32 elements.
4873 // The values are implicitly truncated so sext vs. zext doesn't matter.
4874 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4876 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4877 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4880 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4881 unsigned Opcode = N->getOpcode();
4882 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4883 SDNode *N0 = N->getOperand(0).getNode();
4884 SDNode *N1 = N->getOperand(1).getNode();
4885 return N0->hasOneUse() && N1->hasOneUse() &&
4886 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4891 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4892 unsigned Opcode = N->getOpcode();
4893 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4894 SDNode *N0 = N->getOperand(0).getNode();
4895 SDNode *N1 = N->getOperand(1).getNode();
4896 return N0->hasOneUse() && N1->hasOneUse() &&
4897 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4902 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4903 // Multiplications are only custom-lowered for 128-bit vectors so that
4904 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4905 EVT VT = Op.getValueType();
4906 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4907 SDNode *N0 = Op.getOperand(0).getNode();
4908 SDNode *N1 = Op.getOperand(1).getNode();
4909 unsigned NewOpc = 0;
4911 bool isN0SExt = isSignExtended(N0, DAG);
4912 bool isN1SExt = isSignExtended(N1, DAG);
4913 if (isN0SExt && isN1SExt)
4914 NewOpc = ARMISD::VMULLs;
4916 bool isN0ZExt = isZeroExtended(N0, DAG);
4917 bool isN1ZExt = isZeroExtended(N1, DAG);
4918 if (isN0ZExt && isN1ZExt)
4919 NewOpc = ARMISD::VMULLu;
4920 else if (isN1SExt || isN1ZExt) {
4921 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4922 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4923 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4924 NewOpc = ARMISD::VMULLs;
4926 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4927 NewOpc = ARMISD::VMULLu;
4929 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4931 NewOpc = ARMISD::VMULLu;
4937 if (VT == MVT::v2i64)
4938 // Fall through to expand this. It is not legal.
4941 // Other vector multiplications are legal.
4946 // Legalize to a VMULL instruction.
4947 DebugLoc DL = Op.getDebugLoc();
4949 SDValue Op1 = SkipExtension(N1, DAG);
4951 Op0 = SkipExtension(N0, DAG);
4952 assert(Op0.getValueType().is64BitVector() &&
4953 Op1.getValueType().is64BitVector() &&
4954 "unexpected types for extended operands to VMULL");
4955 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4958 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4959 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4966 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4967 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4968 EVT Op1VT = Op1.getValueType();
4969 return DAG.getNode(N0->getOpcode(), DL, VT,
4970 DAG.getNode(NewOpc, DL, VT,
4971 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4972 DAG.getNode(NewOpc, DL, VT,
4973 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4977 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4979 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4980 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4981 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4982 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4983 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4984 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4985 // Get reciprocal estimate.
4986 // float4 recip = vrecpeq_f32(yf);
4987 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4988 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4989 // Because char has a smaller range than uchar, we can actually get away
4990 // without any newton steps. This requires that we use a weird bias
4991 // of 0xb000, however (again, this has been exhaustively tested).
4992 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4993 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4994 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4995 Y = DAG.getConstant(0xb000, MVT::i32);
4996 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4997 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4998 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4999 // Convert back to short.
5000 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5001 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5006 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5008 // Convert to float.
5009 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5010 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5011 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5012 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5013 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5014 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5016 // Use reciprocal estimate and one refinement step.
5017 // float4 recip = vrecpeq_f32(yf);
5018 // recip *= vrecpsq_f32(yf, recip);
5019 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5020 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5021 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5022 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5024 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5025 // Because short has a smaller range than ushort, we can actually get away
5026 // with only a single newton step. This requires that we use a weird bias
5027 // of 89, however (again, this has been exhaustively tested).
5028 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5029 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5030 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5031 N1 = DAG.getConstant(0x89, MVT::i32);
5032 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5033 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5034 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5035 // Convert back to integer and return.
5036 // return vmovn_s32(vcvt_s32_f32(result));
5037 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5038 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5042 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5043 EVT VT = Op.getValueType();
5044 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5045 "unexpected type for custom-lowering ISD::SDIV");
5047 DebugLoc dl = Op.getDebugLoc();
5048 SDValue N0 = Op.getOperand(0);
5049 SDValue N1 = Op.getOperand(1);
5052 if (VT == MVT::v8i8) {
5053 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5054 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5056 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5057 DAG.getIntPtrConstant(4));
5058 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5059 DAG.getIntPtrConstant(4));
5060 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5061 DAG.getIntPtrConstant(0));
5062 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5063 DAG.getIntPtrConstant(0));
5065 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5066 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5068 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5069 N0 = LowerCONCAT_VECTORS(N0, DAG);
5071 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5074 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5077 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5078 EVT VT = Op.getValueType();
5079 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5080 "unexpected type for custom-lowering ISD::UDIV");
5082 DebugLoc dl = Op.getDebugLoc();
5083 SDValue N0 = Op.getOperand(0);
5084 SDValue N1 = Op.getOperand(1);
5087 if (VT == MVT::v8i8) {
5088 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5089 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5091 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5092 DAG.getIntPtrConstant(4));
5093 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5094 DAG.getIntPtrConstant(4));
5095 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5096 DAG.getIntPtrConstant(0));
5097 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5098 DAG.getIntPtrConstant(0));
5100 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5101 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5103 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5104 N0 = LowerCONCAT_VECTORS(N0, DAG);
5106 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5107 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5112 // v4i16 sdiv ... Convert to float.
5113 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5114 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5115 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5116 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5117 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5118 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5120 // Use reciprocal estimate and two refinement steps.
5121 // float4 recip = vrecpeq_f32(yf);
5122 // recip *= vrecpsq_f32(yf, recip);
5123 // recip *= vrecpsq_f32(yf, recip);
5124 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5125 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5126 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5127 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5129 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5130 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5131 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5133 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5134 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5135 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5136 // and that it will never cause us to return an answer too large).
5137 // float4 result = as_float4(as_int4(xf*recip) + 2);
5138 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5139 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5140 N1 = DAG.getConstant(2, MVT::i32);
5141 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5142 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5143 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5144 // Convert back to integer and return.
5145 // return vmovn_u32(vcvt_s32_f32(result));
5146 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5147 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5151 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5152 EVT VT = Op.getNode()->getValueType(0);
5153 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5156 bool ExtraOp = false;
5157 switch (Op.getOpcode()) {
5158 default: llvm_unreachable("Invalid code");
5159 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5160 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5161 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5162 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5166 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5168 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5169 Op.getOperand(1), Op.getOperand(2));
5172 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5173 // Monotonic load/store is legal for all targets
5174 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5177 // Aquire/Release load/store is not legal for targets without a
5178 // dmb or equivalent available.
5184 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5185 SelectionDAG &DAG, unsigned NewOp) {
5186 DebugLoc dl = Node->getDebugLoc();
5187 assert (Node->getValueType(0) == MVT::i64 &&
5188 "Only know how to expand i64 atomics");
5190 SmallVector<SDValue, 6> Ops;
5191 Ops.push_back(Node->getOperand(0)); // Chain
5192 Ops.push_back(Node->getOperand(1)); // Ptr
5194 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5195 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5196 // High part of Val1
5197 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5198 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5199 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5200 // High part of Val1
5201 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5202 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5203 // High part of Val2
5204 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5205 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5207 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5209 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5210 cast<MemSDNode>(Node)->getMemOperand());
5211 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5212 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5213 Results.push_back(Result.getValue(2));
5216 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5217 switch (Op.getOpcode()) {
5218 default: llvm_unreachable("Don't know how to custom lower this!");
5219 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5220 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5221 case ISD::GlobalAddress:
5222 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5223 LowerGlobalAddressELF(Op, DAG);
5224 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5225 case ISD::SELECT: return LowerSELECT(Op, DAG);
5226 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5227 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5228 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5229 case ISD::VASTART: return LowerVASTART(Op, DAG);
5230 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5231 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5232 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5233 case ISD::SINT_TO_FP:
5234 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5235 case ISD::FP_TO_SINT:
5236 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5237 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5238 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5239 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5240 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5241 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5242 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5243 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5245 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5248 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5249 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5250 case ISD::SRL_PARTS:
5251 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5252 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5253 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5254 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5255 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5256 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5257 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5258 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5259 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5260 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5261 case ISD::MUL: return LowerMUL(Op, DAG);
5262 case ISD::SDIV: return LowerSDIV(Op, DAG);
5263 case ISD::UDIV: return LowerUDIV(Op, DAG);
5267 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5268 case ISD::ATOMIC_LOAD:
5269 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5273 /// ReplaceNodeResults - Replace the results of node with an illegal result
5274 /// type with new values built out of custom code.
5275 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5276 SmallVectorImpl<SDValue>&Results,
5277 SelectionDAG &DAG) const {
5279 switch (N->getOpcode()) {
5281 llvm_unreachable("Don't know how to custom expand this!");
5283 Res = ExpandBITCAST(N, DAG);
5287 Res = Expand64BitShift(N, DAG, Subtarget);
5289 case ISD::ATOMIC_LOAD_ADD:
5290 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5292 case ISD::ATOMIC_LOAD_AND:
5293 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5295 case ISD::ATOMIC_LOAD_NAND:
5296 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5298 case ISD::ATOMIC_LOAD_OR:
5299 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5301 case ISD::ATOMIC_LOAD_SUB:
5302 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5304 case ISD::ATOMIC_LOAD_XOR:
5305 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5307 case ISD::ATOMIC_SWAP:
5308 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5310 case ISD::ATOMIC_CMP_SWAP:
5311 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5315 Results.push_back(Res);
5318 //===----------------------------------------------------------------------===//
5319 // ARM Scheduler Hooks
5320 //===----------------------------------------------------------------------===//
5323 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5324 MachineBasicBlock *BB,
5325 unsigned Size) const {
5326 unsigned dest = MI->getOperand(0).getReg();
5327 unsigned ptr = MI->getOperand(1).getReg();
5328 unsigned oldval = MI->getOperand(2).getReg();
5329 unsigned newval = MI->getOperand(3).getReg();
5330 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5331 DebugLoc dl = MI->getDebugLoc();
5332 bool isThumb2 = Subtarget->isThumb2();
5334 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5335 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5336 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5337 (const TargetRegisterClass*)&ARM::GPRRegClass);
5340 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5341 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5342 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5345 unsigned ldrOpc, strOpc;
5347 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5349 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5350 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5353 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5354 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5357 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5358 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5362 MachineFunction *MF = BB->getParent();
5363 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5364 MachineFunction::iterator It = BB;
5365 ++It; // insert the new blocks after the current block
5367 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5368 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5369 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5370 MF->insert(It, loop1MBB);
5371 MF->insert(It, loop2MBB);
5372 MF->insert(It, exitMBB);
5374 // Transfer the remainder of BB and its successor edges to exitMBB.
5375 exitMBB->splice(exitMBB->begin(), BB,
5376 llvm::next(MachineBasicBlock::iterator(MI)),
5378 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5382 // fallthrough --> loop1MBB
5383 BB->addSuccessor(loop1MBB);
5386 // ldrex dest, [ptr]
5390 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5391 if (ldrOpc == ARM::t2LDREX)
5393 AddDefaultPred(MIB);
5394 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5395 .addReg(dest).addReg(oldval));
5396 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5397 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5398 BB->addSuccessor(loop2MBB);
5399 BB->addSuccessor(exitMBB);
5402 // strex scratch, newval, [ptr]
5406 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5407 if (strOpc == ARM::t2STREX)
5409 AddDefaultPred(MIB);
5410 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5411 .addReg(scratch).addImm(0));
5412 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5413 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5414 BB->addSuccessor(loop1MBB);
5415 BB->addSuccessor(exitMBB);
5421 MI->eraseFromParent(); // The instruction is gone now.
5427 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5428 unsigned Size, unsigned BinOpcode) const {
5429 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5430 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5432 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5433 MachineFunction *MF = BB->getParent();
5434 MachineFunction::iterator It = BB;
5437 unsigned dest = MI->getOperand(0).getReg();
5438 unsigned ptr = MI->getOperand(1).getReg();
5439 unsigned incr = MI->getOperand(2).getReg();
5440 DebugLoc dl = MI->getDebugLoc();
5441 bool isThumb2 = Subtarget->isThumb2();
5443 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5445 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5446 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5449 unsigned ldrOpc, strOpc;
5451 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5453 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5454 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5457 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5458 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5461 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5462 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5466 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5467 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5468 MF->insert(It, loopMBB);
5469 MF->insert(It, exitMBB);
5471 // Transfer the remainder of BB and its successor edges to exitMBB.
5472 exitMBB->splice(exitMBB->begin(), BB,
5473 llvm::next(MachineBasicBlock::iterator(MI)),
5475 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5477 const TargetRegisterClass *TRC = isThumb2 ?
5478 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5479 (const TargetRegisterClass*)&ARM::GPRRegClass;
5480 unsigned scratch = MRI.createVirtualRegister(TRC);
5481 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5485 // fallthrough --> loopMBB
5486 BB->addSuccessor(loopMBB);
5490 // <binop> scratch2, dest, incr
5491 // strex scratch, scratch2, ptr
5494 // fallthrough --> exitMBB
5496 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5497 if (ldrOpc == ARM::t2LDREX)
5499 AddDefaultPred(MIB);
5501 // operand order needs to go the other way for NAND
5502 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5503 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5504 addReg(incr).addReg(dest)).addReg(0);
5506 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5507 addReg(dest).addReg(incr)).addReg(0);
5510 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5511 if (strOpc == ARM::t2STREX)
5513 AddDefaultPred(MIB);
5514 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5515 .addReg(scratch).addImm(0));
5516 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5517 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5519 BB->addSuccessor(loopMBB);
5520 BB->addSuccessor(exitMBB);
5526 MI->eraseFromParent(); // The instruction is gone now.
5532 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5533 MachineBasicBlock *BB,
5536 ARMCC::CondCodes Cond) const {
5537 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5539 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5540 MachineFunction *MF = BB->getParent();
5541 MachineFunction::iterator It = BB;
5544 unsigned dest = MI->getOperand(0).getReg();
5545 unsigned ptr = MI->getOperand(1).getReg();
5546 unsigned incr = MI->getOperand(2).getReg();
5547 unsigned oldval = dest;
5548 DebugLoc dl = MI->getDebugLoc();
5549 bool isThumb2 = Subtarget->isThumb2();
5551 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5553 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5554 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5557 unsigned ldrOpc, strOpc, extendOpc;
5559 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5561 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5562 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5563 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5566 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5567 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5568 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5571 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5572 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5577 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5578 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5579 MF->insert(It, loopMBB);
5580 MF->insert(It, exitMBB);
5582 // Transfer the remainder of BB and its successor edges to exitMBB.
5583 exitMBB->splice(exitMBB->begin(), BB,
5584 llvm::next(MachineBasicBlock::iterator(MI)),
5586 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5588 const TargetRegisterClass *TRC = isThumb2 ?
5589 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5590 (const TargetRegisterClass*)&ARM::GPRRegClass;
5591 unsigned scratch = MRI.createVirtualRegister(TRC);
5592 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5596 // fallthrough --> loopMBB
5597 BB->addSuccessor(loopMBB);
5601 // (sign extend dest, if required)
5603 // cmov.cond scratch2, incr, dest
5604 // strex scratch, scratch2, ptr
5607 // fallthrough --> exitMBB
5609 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5610 if (ldrOpc == ARM::t2LDREX)
5612 AddDefaultPred(MIB);
5614 // Sign extend the value, if necessary.
5615 if (signExtend && extendOpc) {
5616 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5617 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5622 // Build compare and cmov instructions.
5623 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5624 .addReg(oldval).addReg(incr));
5625 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5626 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5628 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5629 if (strOpc == ARM::t2STREX)
5631 AddDefaultPred(MIB);
5632 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5633 .addReg(scratch).addImm(0));
5634 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5635 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5637 BB->addSuccessor(loopMBB);
5638 BB->addSuccessor(exitMBB);
5644 MI->eraseFromParent(); // The instruction is gone now.
5650 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5651 unsigned Op1, unsigned Op2,
5652 bool NeedsCarry, bool IsCmpxchg) const {
5653 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5654 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5656 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5657 MachineFunction *MF = BB->getParent();
5658 MachineFunction::iterator It = BB;
5661 unsigned destlo = MI->getOperand(0).getReg();
5662 unsigned desthi = MI->getOperand(1).getReg();
5663 unsigned ptr = MI->getOperand(2).getReg();
5664 unsigned vallo = MI->getOperand(3).getReg();
5665 unsigned valhi = MI->getOperand(4).getReg();
5666 DebugLoc dl = MI->getDebugLoc();
5667 bool isThumb2 = Subtarget->isThumb2();
5669 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5671 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5672 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5673 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5676 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5677 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5679 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5680 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5682 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5683 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5685 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5686 MF->insert(It, loopMBB);
5688 MF->insert(It, contBB);
5689 MF->insert(It, cont2BB);
5691 MF->insert(It, exitMBB);
5693 // Transfer the remainder of BB and its successor edges to exitMBB.
5694 exitMBB->splice(exitMBB->begin(), BB,
5695 llvm::next(MachineBasicBlock::iterator(MI)),
5697 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5699 const TargetRegisterClass *TRC = isThumb2 ?
5700 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5701 (const TargetRegisterClass*)&ARM::GPRRegClass;
5702 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5706 // fallthrough --> loopMBB
5707 BB->addSuccessor(loopMBB);
5710 // ldrexd r2, r3, ptr
5711 // <binopa> r0, r2, incr
5712 // <binopb> r1, r3, incr
5713 // strexd storesuccess, r0, r1, ptr
5714 // cmp storesuccess, #0
5716 // fallthrough --> exitMBB
5718 // Note that the registers are explicitly specified because there is not any
5719 // way to force the register allocator to allocate a register pair.
5721 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5722 // need to properly enforce the restriction that the two output registers
5723 // for ldrexd must be different.
5726 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5727 .addReg(ARM::R2, RegState::Define)
5728 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5729 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5730 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5731 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5735 for (unsigned i = 0; i < 2; i++) {
5736 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5738 .addReg(i == 0 ? destlo : desthi)
5739 .addReg(i == 0 ? vallo : valhi));
5740 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5741 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5742 BB->addSuccessor(exitMBB);
5743 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5744 BB = (i == 0 ? contBB : cont2BB);
5747 // Copy to physregs for strexd
5748 unsigned setlo = MI->getOperand(5).getReg();
5749 unsigned sethi = MI->getOperand(6).getReg();
5750 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5751 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5753 // Perform binary operation
5754 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5755 .addReg(destlo).addReg(vallo))
5756 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5757 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5758 .addReg(desthi).addReg(valhi)).addReg(0);
5760 // Copy to physregs for strexd
5761 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5762 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5766 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5767 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5769 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5770 .addReg(storesuccess).addImm(0));
5771 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5772 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5774 BB->addSuccessor(loopMBB);
5775 BB->addSuccessor(exitMBB);
5781 MI->eraseFromParent(); // The instruction is gone now.
5786 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5787 /// registers the function context.
5788 void ARMTargetLowering::
5789 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5790 MachineBasicBlock *DispatchBB, int FI) const {
5791 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5792 DebugLoc dl = MI->getDebugLoc();
5793 MachineFunction *MF = MBB->getParent();
5794 MachineRegisterInfo *MRI = &MF->getRegInfo();
5795 MachineConstantPool *MCP = MF->getConstantPool();
5796 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5797 const Function *F = MF->getFunction();
5799 bool isThumb = Subtarget->isThumb();
5800 bool isThumb2 = Subtarget->isThumb2();
5802 unsigned PCLabelId = AFI->createPICLabelUId();
5803 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5804 ARMConstantPoolValue *CPV =
5805 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5806 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5808 const TargetRegisterClass *TRC = isThumb ?
5809 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5810 (const TargetRegisterClass*)&ARM::GPRRegClass;
5812 // Grab constant pool and fixed stack memory operands.
5813 MachineMemOperand *CPMMO =
5814 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5815 MachineMemOperand::MOLoad, 4, 4);
5817 MachineMemOperand *FIMMOSt =
5818 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5819 MachineMemOperand::MOStore, 4, 4);
5821 // Load the address of the dispatch MBB into the jump buffer.
5823 // Incoming value: jbuf
5824 // ldr.n r5, LCPI1_1
5827 // str r5, [$jbuf, #+4] ; &jbuf[1]
5828 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5829 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5830 .addConstantPoolIndex(CPI)
5831 .addMemOperand(CPMMO));
5832 // Set the low bit because of thumb mode.
5833 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5835 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5836 .addReg(NewVReg1, RegState::Kill)
5838 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5839 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5840 .addReg(NewVReg2, RegState::Kill)
5842 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5843 .addReg(NewVReg3, RegState::Kill)
5845 .addImm(36) // &jbuf[1] :: pc
5846 .addMemOperand(FIMMOSt));
5847 } else if (isThumb) {
5848 // Incoming value: jbuf
5849 // ldr.n r1, LCPI1_4
5853 // add r2, $jbuf, #+4 ; &jbuf[1]
5855 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5856 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5857 .addConstantPoolIndex(CPI)
5858 .addMemOperand(CPMMO));
5859 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5860 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5861 .addReg(NewVReg1, RegState::Kill)
5863 // Set the low bit because of thumb mode.
5864 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5865 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5866 .addReg(ARM::CPSR, RegState::Define)
5868 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5869 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5870 .addReg(ARM::CPSR, RegState::Define)
5871 .addReg(NewVReg2, RegState::Kill)
5872 .addReg(NewVReg3, RegState::Kill));
5873 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5874 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5876 .addImm(36)); // &jbuf[1] :: pc
5877 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5878 .addReg(NewVReg4, RegState::Kill)
5879 .addReg(NewVReg5, RegState::Kill)
5881 .addMemOperand(FIMMOSt));
5883 // Incoming value: jbuf
5886 // str r1, [$jbuf, #+4] ; &jbuf[1]
5887 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5888 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5889 .addConstantPoolIndex(CPI)
5891 .addMemOperand(CPMMO));
5892 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5893 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5894 .addReg(NewVReg1, RegState::Kill)
5895 .addImm(PCLabelId));
5896 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5897 .addReg(NewVReg2, RegState::Kill)
5899 .addImm(36) // &jbuf[1] :: pc
5900 .addMemOperand(FIMMOSt));
5904 MachineBasicBlock *ARMTargetLowering::
5905 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5906 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5907 DebugLoc dl = MI->getDebugLoc();
5908 MachineFunction *MF = MBB->getParent();
5909 MachineRegisterInfo *MRI = &MF->getRegInfo();
5910 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5911 MachineFrameInfo *MFI = MF->getFrameInfo();
5912 int FI = MFI->getFunctionContextIndex();
5914 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5915 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5916 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
5918 // Get a mapping of the call site numbers to all of the landing pads they're
5920 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5921 unsigned MaxCSNum = 0;
5922 MachineModuleInfo &MMI = MF->getMMI();
5923 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5925 if (!BB->isLandingPad()) continue;
5927 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5929 for (MachineBasicBlock::iterator
5930 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5931 if (!II->isEHLabel()) continue;
5933 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5934 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5936 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5937 for (SmallVectorImpl<unsigned>::iterator
5938 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5939 CSI != CSE; ++CSI) {
5940 CallSiteNumToLPad[*CSI].push_back(BB);
5941 MaxCSNum = std::max(MaxCSNum, *CSI);
5947 // Get an ordered list of the machine basic blocks for the jump table.
5948 std::vector<MachineBasicBlock*> LPadList;
5949 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5950 LPadList.reserve(CallSiteNumToLPad.size());
5951 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5952 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5953 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5954 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5955 LPadList.push_back(*II);
5956 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5960 assert(!LPadList.empty() &&
5961 "No landing pad destinations for the dispatch jump table!");
5963 // Create the jump table and associated information.
5964 MachineJumpTableInfo *JTI =
5965 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5966 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5967 unsigned UId = AFI->createJumpTableUId();
5969 // Create the MBBs for the dispatch code.
5971 // Shove the dispatch's address into the return slot in the function context.
5972 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5973 DispatchBB->setIsLandingPad();
5975 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5976 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5977 DispatchBB->addSuccessor(TrapBB);
5979 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5980 DispatchBB->addSuccessor(DispContBB);
5983 MF->insert(MF->end(), DispatchBB);
5984 MF->insert(MF->end(), DispContBB);
5985 MF->insert(MF->end(), TrapBB);
5987 // Insert code into the entry block that creates and registers the function
5989 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5991 MachineMemOperand *FIMMOLd =
5992 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5993 MachineMemOperand::MOLoad |
5994 MachineMemOperand::MOVolatile, 4, 4);
5996 if (AFI->isThumb1OnlyFunction())
5997 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5998 else if (!Subtarget->hasVFP2())
5999 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
6001 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6003 unsigned NumLPads = LPadList.size();
6004 if (Subtarget->isThumb2()) {
6005 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6006 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6009 .addMemOperand(FIMMOLd));
6011 if (NumLPads < 256) {
6012 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6014 .addImm(LPadList.size()));
6016 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6017 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6018 .addImm(NumLPads & 0xFFFF));
6020 unsigned VReg2 = VReg1;
6021 if ((NumLPads & 0xFFFF0000) != 0) {
6022 VReg2 = MRI->createVirtualRegister(TRC);
6023 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6025 .addImm(NumLPads >> 16));
6028 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6033 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6038 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6039 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6040 .addJumpTableIndex(MJTI)
6043 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6046 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6047 .addReg(NewVReg3, RegState::Kill)
6049 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6051 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6052 .addReg(NewVReg4, RegState::Kill)
6054 .addJumpTableIndex(MJTI)
6056 } else if (Subtarget->isThumb()) {
6057 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6058 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6061 .addMemOperand(FIMMOLd));
6063 if (NumLPads < 256) {
6064 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6068 MachineConstantPool *ConstantPool = MF->getConstantPool();
6069 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6070 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6072 // MachineConstantPool wants an explicit alignment.
6073 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6075 Align = getDataLayout()->getTypeAllocSize(C->getType());
6076 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6078 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6079 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6080 .addReg(VReg1, RegState::Define)
6081 .addConstantPoolIndex(Idx));
6082 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6087 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6092 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6093 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6094 .addReg(ARM::CPSR, RegState::Define)
6098 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6099 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6100 .addJumpTableIndex(MJTI)
6103 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6104 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6105 .addReg(ARM::CPSR, RegState::Define)
6106 .addReg(NewVReg2, RegState::Kill)
6109 MachineMemOperand *JTMMOLd =
6110 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6111 MachineMemOperand::MOLoad, 4, 4);
6113 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6114 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6115 .addReg(NewVReg4, RegState::Kill)
6117 .addMemOperand(JTMMOLd));
6119 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6120 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6121 .addReg(ARM::CPSR, RegState::Define)
6122 .addReg(NewVReg5, RegState::Kill)
6125 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6126 .addReg(NewVReg6, RegState::Kill)
6127 .addJumpTableIndex(MJTI)
6130 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6131 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6134 .addMemOperand(FIMMOLd));
6136 if (NumLPads < 256) {
6137 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6140 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6141 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6142 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6143 .addImm(NumLPads & 0xFFFF));
6145 unsigned VReg2 = VReg1;
6146 if ((NumLPads & 0xFFFF0000) != 0) {
6147 VReg2 = MRI->createVirtualRegister(TRC);
6148 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6150 .addImm(NumLPads >> 16));
6153 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6157 MachineConstantPool *ConstantPool = MF->getConstantPool();
6158 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6159 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6161 // MachineConstantPool wants an explicit alignment.
6162 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6164 Align = getDataLayout()->getTypeAllocSize(C->getType());
6165 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6167 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6168 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6169 .addReg(VReg1, RegState::Define)
6170 .addConstantPoolIndex(Idx)
6172 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6174 .addReg(VReg1, RegState::Kill));
6177 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6182 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6184 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6186 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6187 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6188 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6189 .addJumpTableIndex(MJTI)
6192 MachineMemOperand *JTMMOLd =
6193 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6194 MachineMemOperand::MOLoad, 4, 4);
6195 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6197 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6198 .addReg(NewVReg3, RegState::Kill)
6201 .addMemOperand(JTMMOLd));
6203 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6204 .addReg(NewVReg5, RegState::Kill)
6206 .addJumpTableIndex(MJTI)
6210 // Add the jump table entries as successors to the MBB.
6211 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6212 for (std::vector<MachineBasicBlock*>::iterator
6213 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6214 MachineBasicBlock *CurMBB = *I;
6215 if (SeenMBBs.insert(CurMBB))
6216 DispContBB->addSuccessor(CurMBB);
6219 // N.B. the order the invoke BBs are processed in doesn't matter here.
6220 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6221 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6222 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6223 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6224 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6225 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6226 MachineBasicBlock *BB = *I;
6228 // Remove the landing pad successor from the invoke block and replace it
6229 // with the new dispatch block.
6230 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6232 while (!Successors.empty()) {
6233 MachineBasicBlock *SMBB = Successors.pop_back_val();
6234 if (SMBB->isLandingPad()) {
6235 BB->removeSuccessor(SMBB);
6236 MBBLPads.push_back(SMBB);
6240 BB->addSuccessor(DispatchBB);
6242 // Find the invoke call and mark all of the callee-saved registers as
6243 // 'implicit defined' so that they're spilled. This prevents code from
6244 // moving instructions to before the EH block, where they will never be
6246 for (MachineBasicBlock::reverse_iterator
6247 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6248 if (!II->isCall()) continue;
6250 DenseMap<unsigned, bool> DefRegs;
6251 for (MachineInstr::mop_iterator
6252 OI = II->operands_begin(), OE = II->operands_end();
6254 if (!OI->isReg()) continue;
6255 DefRegs[OI->getReg()] = true;
6258 MachineInstrBuilder MIB(&*II);
6260 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6261 unsigned Reg = SavedRegs[i];
6262 if (Subtarget->isThumb2() &&
6263 !ARM::tGPRRegClass.contains(Reg) &&
6264 !ARM::hGPRRegClass.contains(Reg))
6266 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6268 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6271 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6278 // Mark all former landing pads as non-landing pads. The dispatch is the only
6280 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6281 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6282 (*I)->setIsLandingPad(false);
6284 // The instruction is gone now.
6285 MI->eraseFromParent();
6291 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6292 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6293 E = MBB->succ_end(); I != E; ++I)
6296 llvm_unreachable("Expecting a BB with two successors!");
6299 MachineBasicBlock *ARMTargetLowering::
6300 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6301 // This pseudo instruction has 3 operands: dst, src, size
6302 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6303 // Otherwise, we will generate unrolled scalar copies.
6304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6305 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6306 MachineFunction::iterator It = BB;
6309 unsigned dest = MI->getOperand(0).getReg();
6310 unsigned src = MI->getOperand(1).getReg();
6311 unsigned SizeVal = MI->getOperand(2).getImm();
6312 unsigned Align = MI->getOperand(3).getImm();
6313 DebugLoc dl = MI->getDebugLoc();
6315 bool isThumb2 = Subtarget->isThumb2();
6316 MachineFunction *MF = BB->getParent();
6317 MachineRegisterInfo &MRI = MF->getRegInfo();
6318 unsigned ldrOpc, strOpc, UnitSize = 0;
6320 const TargetRegisterClass *TRC = isThumb2 ?
6321 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6322 (const TargetRegisterClass*)&ARM::GPRRegClass;
6323 const TargetRegisterClass *TRC_Vec = 0;
6326 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6327 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6329 } else if (Align & 2) {
6330 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6331 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6334 // Check whether we can use NEON instructions.
6335 if (!MF->getFunction()->getFnAttributes().
6336 hasAttribute(Attributes::NoImplicitFloat) &&
6337 Subtarget->hasNEON()) {
6338 if ((Align % 16 == 0) && SizeVal >= 16) {
6339 ldrOpc = ARM::VLD1q32wb_fixed;
6340 strOpc = ARM::VST1q32wb_fixed;
6342 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6344 else if ((Align % 8 == 0) && SizeVal >= 8) {
6345 ldrOpc = ARM::VLD1d32wb_fixed;
6346 strOpc = ARM::VST1d32wb_fixed;
6348 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6351 // Can't use NEON instructions.
6352 if (UnitSize == 0) {
6353 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6354 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6359 unsigned BytesLeft = SizeVal % UnitSize;
6360 unsigned LoopSize = SizeVal - BytesLeft;
6362 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6363 // Use LDR and STR to copy.
6364 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6365 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6366 unsigned srcIn = src;
6367 unsigned destIn = dest;
6368 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6369 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6370 unsigned srcOut = MRI.createVirtualRegister(TRC);
6371 unsigned destOut = MRI.createVirtualRegister(TRC);
6372 if (UnitSize >= 8) {
6373 AddDefaultPred(BuildMI(*BB, MI, dl,
6374 TII->get(ldrOpc), scratch)
6375 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6377 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6378 .addReg(destIn).addImm(0).addReg(scratch));
6379 } else if (isThumb2) {
6380 AddDefaultPred(BuildMI(*BB, MI, dl,
6381 TII->get(ldrOpc), scratch)
6382 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6384 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6385 .addReg(scratch).addReg(destIn)
6388 AddDefaultPred(BuildMI(*BB, MI, dl,
6389 TII->get(ldrOpc), scratch)
6390 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6393 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6394 .addReg(scratch).addReg(destIn)
6395 .addReg(0).addImm(UnitSize));
6401 // Handle the leftover bytes with LDRB and STRB.
6402 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6403 // [destOut] = STRB_POST(scratch, destIn, 1)
6404 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6405 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6406 for (unsigned i = 0; i < BytesLeft; i++) {
6407 unsigned scratch = MRI.createVirtualRegister(TRC);
6408 unsigned srcOut = MRI.createVirtualRegister(TRC);
6409 unsigned destOut = MRI.createVirtualRegister(TRC);
6411 AddDefaultPred(BuildMI(*BB, MI, dl,
6412 TII->get(ldrOpc),scratch)
6413 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6415 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6416 .addReg(scratch).addReg(destIn)
6417 .addReg(0).addImm(1));
6419 AddDefaultPred(BuildMI(*BB, MI, dl,
6420 TII->get(ldrOpc),scratch)
6421 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6423 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6424 .addReg(scratch).addReg(destIn)
6425 .addReg(0).addImm(1));
6430 MI->eraseFromParent(); // The instruction is gone now.
6434 // Expand the pseudo op to a loop.
6437 // movw varEnd, # --> with thumb2
6439 // ldrcp varEnd, idx --> without thumb2
6440 // fallthrough --> loopMBB
6442 // PHI varPhi, varEnd, varLoop
6443 // PHI srcPhi, src, srcLoop
6444 // PHI destPhi, dst, destLoop
6445 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6446 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6447 // subs varLoop, varPhi, #UnitSize
6449 // fallthrough --> exitMBB
6451 // epilogue to handle left-over bytes
6452 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6453 // [destOut] = STRB_POST(scratch, destLoop, 1)
6454 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6455 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6456 MF->insert(It, loopMBB);
6457 MF->insert(It, exitMBB);
6459 // Transfer the remainder of BB and its successor edges to exitMBB.
6460 exitMBB->splice(exitMBB->begin(), BB,
6461 llvm::next(MachineBasicBlock::iterator(MI)),
6463 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6465 // Load an immediate to varEnd.
6466 unsigned varEnd = MRI.createVirtualRegister(TRC);
6468 unsigned VReg1 = varEnd;
6469 if ((LoopSize & 0xFFFF0000) != 0)
6470 VReg1 = MRI.createVirtualRegister(TRC);
6471 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6472 .addImm(LoopSize & 0xFFFF));
6474 if ((LoopSize & 0xFFFF0000) != 0)
6475 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6477 .addImm(LoopSize >> 16));
6479 MachineConstantPool *ConstantPool = MF->getConstantPool();
6480 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6481 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6483 // MachineConstantPool wants an explicit alignment.
6484 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6486 Align = getDataLayout()->getTypeAllocSize(C->getType());
6487 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6489 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6490 .addReg(varEnd, RegState::Define)
6491 .addConstantPoolIndex(Idx)
6494 BB->addSuccessor(loopMBB);
6496 // Generate the loop body:
6497 // varPhi = PHI(varLoop, varEnd)
6498 // srcPhi = PHI(srcLoop, src)
6499 // destPhi = PHI(destLoop, dst)
6500 MachineBasicBlock *entryBB = BB;
6502 unsigned varLoop = MRI.createVirtualRegister(TRC);
6503 unsigned varPhi = MRI.createVirtualRegister(TRC);
6504 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6505 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6506 unsigned destLoop = MRI.createVirtualRegister(TRC);
6507 unsigned destPhi = MRI.createVirtualRegister(TRC);
6509 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6510 .addReg(varLoop).addMBB(loopMBB)
6511 .addReg(varEnd).addMBB(entryBB);
6512 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6513 .addReg(srcLoop).addMBB(loopMBB)
6514 .addReg(src).addMBB(entryBB);
6515 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6516 .addReg(destLoop).addMBB(loopMBB)
6517 .addReg(dest).addMBB(entryBB);
6519 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6520 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6521 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6522 if (UnitSize >= 8) {
6523 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6524 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6526 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6527 .addReg(destPhi).addImm(0).addReg(scratch));
6528 } else if (isThumb2) {
6529 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6530 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6532 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6533 .addReg(scratch).addReg(destPhi)
6536 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6537 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6540 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6541 .addReg(scratch).addReg(destPhi)
6542 .addReg(0).addImm(UnitSize));
6545 // Decrement loop variable by UnitSize.
6546 MachineInstrBuilder MIB = BuildMI(BB, dl,
6547 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6548 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6549 MIB->getOperand(5).setReg(ARM::CPSR);
6550 MIB->getOperand(5).setIsDef(true);
6552 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6553 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6555 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6556 BB->addSuccessor(loopMBB);
6557 BB->addSuccessor(exitMBB);
6559 // Add epilogue to handle BytesLeft.
6561 MachineInstr *StartOfExit = exitMBB->begin();
6562 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6563 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6565 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6566 // [destOut] = STRB_POST(scratch, destLoop, 1)
6567 unsigned srcIn = srcLoop;
6568 unsigned destIn = destLoop;
6569 for (unsigned i = 0; i < BytesLeft; i++) {
6570 unsigned scratch = MRI.createVirtualRegister(TRC);
6571 unsigned srcOut = MRI.createVirtualRegister(TRC);
6572 unsigned destOut = MRI.createVirtualRegister(TRC);
6574 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6575 TII->get(ldrOpc),scratch)
6576 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6578 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6579 .addReg(scratch).addReg(destIn)
6582 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6583 TII->get(ldrOpc),scratch)
6584 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6586 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6587 .addReg(scratch).addReg(destIn)
6588 .addReg(0).addImm(1));
6594 MI->eraseFromParent(); // The instruction is gone now.
6599 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6600 MachineBasicBlock *BB) const {
6601 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6602 DebugLoc dl = MI->getDebugLoc();
6603 bool isThumb2 = Subtarget->isThumb2();
6604 switch (MI->getOpcode()) {
6607 llvm_unreachable("Unexpected instr type to insert");
6609 // The Thumb2 pre-indexed stores have the same MI operands, they just
6610 // define them differently in the .td files from the isel patterns, so
6611 // they need pseudos.
6612 case ARM::t2STR_preidx:
6613 MI->setDesc(TII->get(ARM::t2STR_PRE));
6615 case ARM::t2STRB_preidx:
6616 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6618 case ARM::t2STRH_preidx:
6619 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6622 case ARM::STRi_preidx:
6623 case ARM::STRBi_preidx: {
6624 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6625 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6626 // Decode the offset.
6627 unsigned Offset = MI->getOperand(4).getImm();
6628 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6629 Offset = ARM_AM::getAM2Offset(Offset);
6633 MachineMemOperand *MMO = *MI->memoperands_begin();
6634 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6635 .addOperand(MI->getOperand(0)) // Rn_wb
6636 .addOperand(MI->getOperand(1)) // Rt
6637 .addOperand(MI->getOperand(2)) // Rn
6638 .addImm(Offset) // offset (skip GPR==zero_reg)
6639 .addOperand(MI->getOperand(5)) // pred
6640 .addOperand(MI->getOperand(6))
6641 .addMemOperand(MMO);
6642 MI->eraseFromParent();
6645 case ARM::STRr_preidx:
6646 case ARM::STRBr_preidx:
6647 case ARM::STRH_preidx: {
6649 switch (MI->getOpcode()) {
6650 default: llvm_unreachable("unexpected opcode!");
6651 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6652 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6653 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6655 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6656 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6657 MIB.addOperand(MI->getOperand(i));
6658 MI->eraseFromParent();
6661 case ARM::ATOMIC_LOAD_ADD_I8:
6662 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6663 case ARM::ATOMIC_LOAD_ADD_I16:
6664 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6665 case ARM::ATOMIC_LOAD_ADD_I32:
6666 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6668 case ARM::ATOMIC_LOAD_AND_I8:
6669 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6670 case ARM::ATOMIC_LOAD_AND_I16:
6671 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6672 case ARM::ATOMIC_LOAD_AND_I32:
6673 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6675 case ARM::ATOMIC_LOAD_OR_I8:
6676 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6677 case ARM::ATOMIC_LOAD_OR_I16:
6678 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6679 case ARM::ATOMIC_LOAD_OR_I32:
6680 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6682 case ARM::ATOMIC_LOAD_XOR_I8:
6683 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6684 case ARM::ATOMIC_LOAD_XOR_I16:
6685 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6686 case ARM::ATOMIC_LOAD_XOR_I32:
6687 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6689 case ARM::ATOMIC_LOAD_NAND_I8:
6690 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6691 case ARM::ATOMIC_LOAD_NAND_I16:
6692 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6693 case ARM::ATOMIC_LOAD_NAND_I32:
6694 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6696 case ARM::ATOMIC_LOAD_SUB_I8:
6697 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6698 case ARM::ATOMIC_LOAD_SUB_I16:
6699 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6700 case ARM::ATOMIC_LOAD_SUB_I32:
6701 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6703 case ARM::ATOMIC_LOAD_MIN_I8:
6704 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6705 case ARM::ATOMIC_LOAD_MIN_I16:
6706 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6707 case ARM::ATOMIC_LOAD_MIN_I32:
6708 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6710 case ARM::ATOMIC_LOAD_MAX_I8:
6711 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6712 case ARM::ATOMIC_LOAD_MAX_I16:
6713 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6714 case ARM::ATOMIC_LOAD_MAX_I32:
6715 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6717 case ARM::ATOMIC_LOAD_UMIN_I8:
6718 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6719 case ARM::ATOMIC_LOAD_UMIN_I16:
6720 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6721 case ARM::ATOMIC_LOAD_UMIN_I32:
6722 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6724 case ARM::ATOMIC_LOAD_UMAX_I8:
6725 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6726 case ARM::ATOMIC_LOAD_UMAX_I16:
6727 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6728 case ARM::ATOMIC_LOAD_UMAX_I32:
6729 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6731 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6732 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6733 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6735 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6736 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6737 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6740 case ARM::ATOMADD6432:
6741 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6742 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6743 /*NeedsCarry*/ true);
6744 case ARM::ATOMSUB6432:
6745 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6746 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6747 /*NeedsCarry*/ true);
6748 case ARM::ATOMOR6432:
6749 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6750 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6751 case ARM::ATOMXOR6432:
6752 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6753 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6754 case ARM::ATOMAND6432:
6755 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6756 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6757 case ARM::ATOMSWAP6432:
6758 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6759 case ARM::ATOMCMPXCHG6432:
6760 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6761 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6762 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6764 case ARM::tMOVCCr_pseudo: {
6765 // To "insert" a SELECT_CC instruction, we actually have to insert the
6766 // diamond control-flow pattern. The incoming instruction knows the
6767 // destination vreg to set, the condition code register to branch on, the
6768 // true/false values to select between, and a branch opcode to use.
6769 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6770 MachineFunction::iterator It = BB;
6776 // cmpTY ccX, r1, r2
6778 // fallthrough --> copy0MBB
6779 MachineBasicBlock *thisMBB = BB;
6780 MachineFunction *F = BB->getParent();
6781 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6782 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6783 F->insert(It, copy0MBB);
6784 F->insert(It, sinkMBB);
6786 // Transfer the remainder of BB and its successor edges to sinkMBB.
6787 sinkMBB->splice(sinkMBB->begin(), BB,
6788 llvm::next(MachineBasicBlock::iterator(MI)),
6790 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6792 BB->addSuccessor(copy0MBB);
6793 BB->addSuccessor(sinkMBB);
6795 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6796 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6799 // %FalseValue = ...
6800 // # fallthrough to sinkMBB
6803 // Update machine-CFG edges
6804 BB->addSuccessor(sinkMBB);
6807 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6810 BuildMI(*BB, BB->begin(), dl,
6811 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6812 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6813 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6815 MI->eraseFromParent(); // The pseudo instruction is gone now.
6820 case ARM::BCCZi64: {
6821 // If there is an unconditional branch to the other successor, remove it.
6822 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6824 // Compare both parts that make up the double comparison separately for
6826 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6828 unsigned LHS1 = MI->getOperand(1).getReg();
6829 unsigned LHS2 = MI->getOperand(2).getReg();
6831 AddDefaultPred(BuildMI(BB, dl,
6832 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6833 .addReg(LHS1).addImm(0));
6834 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6835 .addReg(LHS2).addImm(0)
6836 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6838 unsigned RHS1 = MI->getOperand(3).getReg();
6839 unsigned RHS2 = MI->getOperand(4).getReg();
6840 AddDefaultPred(BuildMI(BB, dl,
6841 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6842 .addReg(LHS1).addReg(RHS1));
6843 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6844 .addReg(LHS2).addReg(RHS2)
6845 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6848 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6849 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6850 if (MI->getOperand(0).getImm() == ARMCC::NE)
6851 std::swap(destMBB, exitMBB);
6853 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6854 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6856 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6858 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6860 MI->eraseFromParent(); // The pseudo instruction is gone now.
6864 case ARM::Int_eh_sjlj_setjmp:
6865 case ARM::Int_eh_sjlj_setjmp_nofp:
6866 case ARM::tInt_eh_sjlj_setjmp:
6867 case ARM::t2Int_eh_sjlj_setjmp:
6868 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6869 EmitSjLjDispatchBlock(MI, BB);
6874 // To insert an ABS instruction, we have to insert the
6875 // diamond control-flow pattern. The incoming instruction knows the
6876 // source vreg to test against 0, the destination vreg to set,
6877 // the condition code register to branch on, the
6878 // true/false values to select between, and a branch opcode to use.
6883 // BCC (branch to SinkBB if V0 >= 0)
6884 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6885 // SinkBB: V1 = PHI(V2, V3)
6886 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6887 MachineFunction::iterator BBI = BB;
6889 MachineFunction *Fn = BB->getParent();
6890 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6891 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6892 Fn->insert(BBI, RSBBB);
6893 Fn->insert(BBI, SinkBB);
6895 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6896 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6897 bool isThumb2 = Subtarget->isThumb2();
6898 MachineRegisterInfo &MRI = Fn->getRegInfo();
6899 // In Thumb mode S must not be specified if source register is the SP or
6900 // PC and if destination register is the SP, so restrict register class
6901 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6902 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6903 (const TargetRegisterClass*)&ARM::GPRRegClass);
6905 // Transfer the remainder of BB and its successor edges to sinkMBB.
6906 SinkBB->splice(SinkBB->begin(), BB,
6907 llvm::next(MachineBasicBlock::iterator(MI)),
6909 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6911 BB->addSuccessor(RSBBB);
6912 BB->addSuccessor(SinkBB);
6914 // fall through to SinkMBB
6915 RSBBB->addSuccessor(SinkBB);
6917 // insert a cmp at the end of BB
6918 AddDefaultPred(BuildMI(BB, dl,
6919 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6920 .addReg(ABSSrcReg).addImm(0));
6922 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6924 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6925 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6927 // insert rsbri in RSBBB
6928 // Note: BCC and rsbri will be converted into predicated rsbmi
6929 // by if-conversion pass
6930 BuildMI(*RSBBB, RSBBB->begin(), dl,
6931 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6932 .addReg(ABSSrcReg, RegState::Kill)
6933 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6935 // insert PHI in SinkBB,
6936 // reuse ABSDstReg to not change uses of ABS instruction
6937 BuildMI(*SinkBB, SinkBB->begin(), dl,
6938 TII->get(ARM::PHI), ABSDstReg)
6939 .addReg(NewRsbDstReg).addMBB(RSBBB)
6940 .addReg(ABSSrcReg).addMBB(BB);
6942 // remove ABS instruction
6943 MI->eraseFromParent();
6945 // return last added BB
6948 case ARM::COPY_STRUCT_BYVAL_I32:
6950 return EmitStructByval(MI, BB);
6954 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6955 SDNode *Node) const {
6956 if (!MI->hasPostISelHook()) {
6957 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6958 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6962 const MCInstrDesc *MCID = &MI->getDesc();
6963 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6964 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6965 // operand is still set to noreg. If needed, set the optional operand's
6966 // register to CPSR, and remove the redundant implicit def.
6968 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6970 // Rename pseudo opcodes.
6971 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6973 const ARMBaseInstrInfo *TII =
6974 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6975 MCID = &TII->get(NewOpc);
6977 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6978 "converted opcode should be the same except for cc_out");
6982 // Add the optional cc_out operand
6983 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6985 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6987 // Any ARM instruction that sets the 's' bit should specify an optional
6988 // "cc_out" operand in the last operand position.
6989 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6990 assert(!NewOpc && "Optional cc_out operand required");
6993 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6994 // since we already have an optional CPSR def.
6995 bool definesCPSR = false;
6996 bool deadCPSR = false;
6997 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6999 const MachineOperand &MO = MI->getOperand(i);
7000 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7004 MI->RemoveOperand(i);
7009 assert(!NewOpc && "Optional cc_out operand required");
7012 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7014 assert(!MI->getOperand(ccOutIdx).getReg() &&
7015 "expect uninitialized optional cc_out operand");
7019 // If this instruction was defined with an optional CPSR def and its dag node
7020 // had a live implicit CPSR def, then activate the optional CPSR def.
7021 MachineOperand &MO = MI->getOperand(ccOutIdx);
7022 MO.setReg(ARM::CPSR);
7026 //===----------------------------------------------------------------------===//
7027 // ARM Optimization Hooks
7028 //===----------------------------------------------------------------------===//
7030 // Helper function that checks if N is a null or all ones constant.
7031 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7032 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7035 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7038 // Return true if N is conditionally 0 or all ones.
7039 // Detects these expressions where cc is an i1 value:
7041 // (select cc 0, y) [AllOnes=0]
7042 // (select cc y, 0) [AllOnes=0]
7043 // (zext cc) [AllOnes=0]
7044 // (sext cc) [AllOnes=0/1]
7045 // (select cc -1, y) [AllOnes=1]
7046 // (select cc y, -1) [AllOnes=1]
7048 // Invert is set when N is the null/all ones constant when CC is false.
7049 // OtherOp is set to the alternative value of N.
7050 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7051 SDValue &CC, bool &Invert,
7053 SelectionDAG &DAG) {
7054 switch (N->getOpcode()) {
7055 default: return false;
7057 CC = N->getOperand(0);
7058 SDValue N1 = N->getOperand(1);
7059 SDValue N2 = N->getOperand(2);
7060 if (isZeroOrAllOnes(N1, AllOnes)) {
7065 if (isZeroOrAllOnes(N2, AllOnes)) {
7072 case ISD::ZERO_EXTEND:
7073 // (zext cc) can never be the all ones value.
7077 case ISD::SIGN_EXTEND: {
7078 EVT VT = N->getValueType(0);
7079 CC = N->getOperand(0);
7080 if (CC.getValueType() != MVT::i1)
7084 // When looking for an AllOnes constant, N is an sext, and the 'other'
7086 OtherOp = DAG.getConstant(0, VT);
7087 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7088 // When looking for a 0 constant, N can be zext or sext.
7089 OtherOp = DAG.getConstant(1, VT);
7091 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7097 // Combine a constant select operand into its use:
7099 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7100 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7101 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7102 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7103 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7105 // The transform is rejected if the select doesn't have a constant operand that
7106 // is null, or all ones when AllOnes is set.
7108 // Also recognize sext/zext from i1:
7110 // (add (zext cc), x) -> (select cc (add x, 1), x)
7111 // (add (sext cc), x) -> (select cc (add x, -1), x)
7113 // These transformations eventually create predicated instructions.
7115 // @param N The node to transform.
7116 // @param Slct The N operand that is a select.
7117 // @param OtherOp The other N operand (x above).
7118 // @param DCI Context.
7119 // @param AllOnes Require the select constant to be all ones instead of null.
7120 // @returns The new node, or SDValue() on failure.
7122 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7123 TargetLowering::DAGCombinerInfo &DCI,
7124 bool AllOnes = false) {
7125 SelectionDAG &DAG = DCI.DAG;
7126 EVT VT = N->getValueType(0);
7127 SDValue NonConstantVal;
7130 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7131 NonConstantVal, DAG))
7134 // Slct is now know to be the desired identity constant when CC is true.
7135 SDValue TrueVal = OtherOp;
7136 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7137 OtherOp, NonConstantVal);
7138 // Unless SwapSelectOps says CC should be false.
7140 std::swap(TrueVal, FalseVal);
7142 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7143 CCOp, TrueVal, FalseVal);
7146 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7148 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7149 TargetLowering::DAGCombinerInfo &DCI) {
7150 SDValue N0 = N->getOperand(0);
7151 SDValue N1 = N->getOperand(1);
7152 if (N0.getNode()->hasOneUse()) {
7153 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7154 if (Result.getNode())
7157 if (N1.getNode()->hasOneUse()) {
7158 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7159 if (Result.getNode())
7165 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7166 // (only after legalization).
7167 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7168 TargetLowering::DAGCombinerInfo &DCI,
7169 const ARMSubtarget *Subtarget) {
7171 // Only perform optimization if after legalize, and if NEON is available. We
7172 // also expected both operands to be BUILD_VECTORs.
7173 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7174 || N0.getOpcode() != ISD::BUILD_VECTOR
7175 || N1.getOpcode() != ISD::BUILD_VECTOR)
7178 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7179 EVT VT = N->getValueType(0);
7180 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7183 // Check that the vector operands are of the right form.
7184 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7185 // operands, where N is the size of the formed vector.
7186 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7187 // index such that we have a pair wise add pattern.
7189 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7190 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7192 SDValue Vec = N0->getOperand(0)->getOperand(0);
7193 SDNode *V = Vec.getNode();
7194 unsigned nextIndex = 0;
7196 // For each operands to the ADD which are BUILD_VECTORs,
7197 // check to see if each of their operands are an EXTRACT_VECTOR with
7198 // the same vector and appropriate index.
7199 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7200 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7201 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7203 SDValue ExtVec0 = N0->getOperand(i);
7204 SDValue ExtVec1 = N1->getOperand(i);
7206 // First operand is the vector, verify its the same.
7207 if (V != ExtVec0->getOperand(0).getNode() ||
7208 V != ExtVec1->getOperand(0).getNode())
7211 // Second is the constant, verify its correct.
7212 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7213 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7215 // For the constant, we want to see all the even or all the odd.
7216 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7217 || C1->getZExtValue() != nextIndex+1)
7226 // Create VPADDL node.
7227 SelectionDAG &DAG = DCI.DAG;
7228 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7230 // Build operand list.
7231 SmallVector<SDValue, 8> Ops;
7232 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7233 TLI.getPointerTy()));
7235 // Input is the vector.
7238 // Get widened type and narrowed type.
7240 unsigned numElem = VT.getVectorNumElements();
7241 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7242 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7243 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7244 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7246 llvm_unreachable("Invalid vector element type for padd optimization.");
7249 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7250 widenType, &Ops[0], Ops.size());
7251 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7254 static SDValue findMUL_LOHI(SDValue V) {
7255 if (V->getOpcode() == ISD::UMUL_LOHI ||
7256 V->getOpcode() == ISD::SMUL_LOHI)
7261 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7262 TargetLowering::DAGCombinerInfo &DCI,
7263 const ARMSubtarget *Subtarget) {
7265 if (Subtarget->isThumb1Only()) return SDValue();
7267 // Only perform the checks after legalize when the pattern is available.
7268 if (DCI.isBeforeLegalize()) return SDValue();
7270 // Look for multiply add opportunities.
7271 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7272 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7273 // a glue link from the first add to the second add.
7274 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7275 // a S/UMLAL instruction.
7278 // \ / \ [no multiline comment]
7284 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7285 SDValue AddcOp0 = AddcNode->getOperand(0);
7286 SDValue AddcOp1 = AddcNode->getOperand(1);
7288 // Check if the two operands are from the same mul_lohi node.
7289 if (AddcOp0.getNode() == AddcOp1.getNode())
7292 assert(AddcNode->getNumValues() == 2 &&
7293 AddcNode->getValueType(0) == MVT::i32 &&
7294 AddcNode->getValueType(1) == MVT::Glue &&
7295 "Expect ADDC with two result values: i32, glue");
7297 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7298 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7299 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7300 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7301 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7304 // Look for the glued ADDE.
7305 SDNode* AddeNode = AddcNode->getGluedUser();
7306 if (AddeNode == NULL)
7309 // Make sure it is really an ADDE.
7310 if (AddeNode->getOpcode() != ISD::ADDE)
7313 assert(AddeNode->getNumOperands() == 3 &&
7314 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7315 "ADDE node has the wrong inputs");
7317 // Check for the triangle shape.
7318 SDValue AddeOp0 = AddeNode->getOperand(0);
7319 SDValue AddeOp1 = AddeNode->getOperand(1);
7321 // Make sure that the ADDE operands are not coming from the same node.
7322 if (AddeOp0.getNode() == AddeOp1.getNode())
7325 // Find the MUL_LOHI node walking up ADDE's operands.
7326 bool IsLeftOperandMUL = false;
7327 SDValue MULOp = findMUL_LOHI(AddeOp0);
7328 if (MULOp == SDValue())
7329 MULOp = findMUL_LOHI(AddeOp1);
7331 IsLeftOperandMUL = true;
7332 if (MULOp == SDValue())
7335 // Figure out the right opcode.
7336 unsigned Opc = MULOp->getOpcode();
7337 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7339 // Figure out the high and low input values to the MLAL node.
7340 SDValue* HiMul = &MULOp;
7341 SDValue* HiAdd = NULL;
7342 SDValue* LoMul = NULL;
7343 SDValue* LowAdd = NULL;
7345 if (IsLeftOperandMUL)
7351 if (AddcOp0->getOpcode() == Opc) {
7355 if (AddcOp1->getOpcode() == Opc) {
7363 if (LoMul->getNode() != HiMul->getNode())
7366 // Create the merged node.
7367 SelectionDAG &DAG = DCI.DAG;
7369 // Build operand list.
7370 SmallVector<SDValue, 8> Ops;
7371 Ops.push_back(LoMul->getOperand(0));
7372 Ops.push_back(LoMul->getOperand(1));
7373 Ops.push_back(*LowAdd);
7374 Ops.push_back(*HiAdd);
7376 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7377 DAG.getVTList(MVT::i32, MVT::i32),
7378 &Ops[0], Ops.size());
7380 // Replace the ADDs' nodes uses by the MLA node's values.
7381 SDValue HiMLALResult(MLALNode.getNode(), 1);
7382 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7384 SDValue LoMLALResult(MLALNode.getNode(), 0);
7385 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7387 // Return original node to notify the driver to stop replacing.
7388 SDValue resNode(AddcNode, 0);
7392 /// PerformADDCCombine - Target-specific dag combine transform from
7393 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7394 static SDValue PerformADDCCombine(SDNode *N,
7395 TargetLowering::DAGCombinerInfo &DCI,
7396 const ARMSubtarget *Subtarget) {
7398 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7402 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7403 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7404 /// called with the default operands, and if that fails, with commuted
7406 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7407 TargetLowering::DAGCombinerInfo &DCI,
7408 const ARMSubtarget *Subtarget){
7410 // Attempt to create vpaddl for this add.
7411 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7412 if (Result.getNode())
7415 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7416 if (N0.getNode()->hasOneUse()) {
7417 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7418 if (Result.getNode()) return Result;
7423 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7425 static SDValue PerformADDCombine(SDNode *N,
7426 TargetLowering::DAGCombinerInfo &DCI,
7427 const ARMSubtarget *Subtarget) {
7428 SDValue N0 = N->getOperand(0);
7429 SDValue N1 = N->getOperand(1);
7431 // First try with the default operand order.
7432 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7433 if (Result.getNode())
7436 // If that didn't work, try again with the operands commuted.
7437 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7440 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7442 static SDValue PerformSUBCombine(SDNode *N,
7443 TargetLowering::DAGCombinerInfo &DCI) {
7444 SDValue N0 = N->getOperand(0);
7445 SDValue N1 = N->getOperand(1);
7447 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7448 if (N1.getNode()->hasOneUse()) {
7449 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7450 if (Result.getNode()) return Result;
7456 /// PerformVMULCombine
7457 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7458 /// special multiplier accumulator forwarding.
7464 static SDValue PerformVMULCombine(SDNode *N,
7465 TargetLowering::DAGCombinerInfo &DCI,
7466 const ARMSubtarget *Subtarget) {
7467 if (!Subtarget->hasVMLxForwarding())
7470 SelectionDAG &DAG = DCI.DAG;
7471 SDValue N0 = N->getOperand(0);
7472 SDValue N1 = N->getOperand(1);
7473 unsigned Opcode = N0.getOpcode();
7474 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7475 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7476 Opcode = N1.getOpcode();
7477 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7478 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7483 EVT VT = N->getValueType(0);
7484 DebugLoc DL = N->getDebugLoc();
7485 SDValue N00 = N0->getOperand(0);
7486 SDValue N01 = N0->getOperand(1);
7487 return DAG.getNode(Opcode, DL, VT,
7488 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7489 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7492 static SDValue PerformMULCombine(SDNode *N,
7493 TargetLowering::DAGCombinerInfo &DCI,
7494 const ARMSubtarget *Subtarget) {
7495 SelectionDAG &DAG = DCI.DAG;
7497 if (Subtarget->isThumb1Only())
7500 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7503 EVT VT = N->getValueType(0);
7504 if (VT.is64BitVector() || VT.is128BitVector())
7505 return PerformVMULCombine(N, DCI, Subtarget);
7509 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7513 int64_t MulAmt = C->getSExtValue();
7514 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7516 ShiftAmt = ShiftAmt & (32 - 1);
7517 SDValue V = N->getOperand(0);
7518 DebugLoc DL = N->getDebugLoc();
7521 MulAmt >>= ShiftAmt;
7524 if (isPowerOf2_32(MulAmt - 1)) {
7525 // (mul x, 2^N + 1) => (add (shl x, N), x)
7526 Res = DAG.getNode(ISD::ADD, DL, VT,
7528 DAG.getNode(ISD::SHL, DL, VT,
7530 DAG.getConstant(Log2_32(MulAmt - 1),
7532 } else if (isPowerOf2_32(MulAmt + 1)) {
7533 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7534 Res = DAG.getNode(ISD::SUB, DL, VT,
7535 DAG.getNode(ISD::SHL, DL, VT,
7537 DAG.getConstant(Log2_32(MulAmt + 1),
7543 uint64_t MulAmtAbs = -MulAmt;
7544 if (isPowerOf2_32(MulAmtAbs + 1)) {
7545 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7546 Res = DAG.getNode(ISD::SUB, DL, VT,
7548 DAG.getNode(ISD::SHL, DL, VT,
7550 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7552 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7553 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7554 Res = DAG.getNode(ISD::ADD, DL, VT,
7556 DAG.getNode(ISD::SHL, DL, VT,
7558 DAG.getConstant(Log2_32(MulAmtAbs-1),
7560 Res = DAG.getNode(ISD::SUB, DL, VT,
7561 DAG.getConstant(0, MVT::i32),Res);
7568 Res = DAG.getNode(ISD::SHL, DL, VT,
7569 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7571 // Do not add new nodes to DAG combiner worklist.
7572 DCI.CombineTo(N, Res, false);
7576 static SDValue PerformANDCombine(SDNode *N,
7577 TargetLowering::DAGCombinerInfo &DCI,
7578 const ARMSubtarget *Subtarget) {
7580 // Attempt to use immediate-form VBIC
7581 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7582 DebugLoc dl = N->getDebugLoc();
7583 EVT VT = N->getValueType(0);
7584 SelectionDAG &DAG = DCI.DAG;
7586 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7589 APInt SplatBits, SplatUndef;
7590 unsigned SplatBitSize;
7593 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7594 if (SplatBitSize <= 64) {
7596 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7597 SplatUndef.getZExtValue(), SplatBitSize,
7598 DAG, VbicVT, VT.is128BitVector(),
7600 if (Val.getNode()) {
7602 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7603 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7604 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7609 if (!Subtarget->isThumb1Only()) {
7610 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7611 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7612 if (Result.getNode())
7619 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7620 static SDValue PerformORCombine(SDNode *N,
7621 TargetLowering::DAGCombinerInfo &DCI,
7622 const ARMSubtarget *Subtarget) {
7623 // Attempt to use immediate-form VORR
7624 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7625 DebugLoc dl = N->getDebugLoc();
7626 EVT VT = N->getValueType(0);
7627 SelectionDAG &DAG = DCI.DAG;
7629 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7632 APInt SplatBits, SplatUndef;
7633 unsigned SplatBitSize;
7635 if (BVN && Subtarget->hasNEON() &&
7636 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7637 if (SplatBitSize <= 64) {
7639 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7640 SplatUndef.getZExtValue(), SplatBitSize,
7641 DAG, VorrVT, VT.is128BitVector(),
7643 if (Val.getNode()) {
7645 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7646 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7647 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7652 if (!Subtarget->isThumb1Only()) {
7653 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7654 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7655 if (Result.getNode())
7659 // The code below optimizes (or (and X, Y), Z).
7660 // The AND operand needs to have a single user to make these optimizations
7662 SDValue N0 = N->getOperand(0);
7663 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7665 SDValue N1 = N->getOperand(1);
7667 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7668 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7669 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7671 unsigned SplatBitSize;
7674 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7676 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7677 HasAnyUndefs) && !HasAnyUndefs) {
7678 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7680 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7681 HasAnyUndefs) && !HasAnyUndefs &&
7682 SplatBits0 == ~SplatBits1) {
7683 // Canonicalize the vector type to make instruction selection simpler.
7684 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7685 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7686 N0->getOperand(1), N0->getOperand(0),
7688 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7693 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7696 // BFI is only available on V6T2+
7697 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7700 DebugLoc DL = N->getDebugLoc();
7701 // 1) or (and A, mask), val => ARMbfi A, val, mask
7702 // iff (val & mask) == val
7704 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7705 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7706 // && mask == ~mask2
7707 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7708 // && ~mask == mask2
7709 // (i.e., copy a bitfield value into another bitfield of the same width)
7714 SDValue N00 = N0.getOperand(0);
7716 // The value and the mask need to be constants so we can verify this is
7717 // actually a bitfield set. If the mask is 0xffff, we can do better
7718 // via a movt instruction, so don't use BFI in that case.
7719 SDValue MaskOp = N0.getOperand(1);
7720 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7723 unsigned Mask = MaskC->getZExtValue();
7727 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7728 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7730 unsigned Val = N1C->getZExtValue();
7731 if ((Val & ~Mask) != Val)
7734 if (ARM::isBitFieldInvertedMask(Mask)) {
7735 Val >>= CountTrailingZeros_32(~Mask);
7737 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7738 DAG.getConstant(Val, MVT::i32),
7739 DAG.getConstant(Mask, MVT::i32));
7741 // Do not add new nodes to DAG combiner worklist.
7742 DCI.CombineTo(N, Res, false);
7745 } else if (N1.getOpcode() == ISD::AND) {
7746 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7747 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7750 unsigned Mask2 = N11C->getZExtValue();
7752 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7754 if (ARM::isBitFieldInvertedMask(Mask) &&
7756 // The pack halfword instruction works better for masks that fit it,
7757 // so use that when it's available.
7758 if (Subtarget->hasT2ExtractPack() &&
7759 (Mask == 0xffff || Mask == 0xffff0000))
7762 unsigned amt = CountTrailingZeros_32(Mask2);
7763 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7764 DAG.getConstant(amt, MVT::i32));
7765 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7766 DAG.getConstant(Mask, MVT::i32));
7767 // Do not add new nodes to DAG combiner worklist.
7768 DCI.CombineTo(N, Res, false);
7770 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7772 // The pack halfword instruction works better for masks that fit it,
7773 // so use that when it's available.
7774 if (Subtarget->hasT2ExtractPack() &&
7775 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7778 unsigned lsb = CountTrailingZeros_32(Mask);
7779 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7780 DAG.getConstant(lsb, MVT::i32));
7781 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7782 DAG.getConstant(Mask2, MVT::i32));
7783 // Do not add new nodes to DAG combiner worklist.
7784 DCI.CombineTo(N, Res, false);
7789 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7790 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7791 ARM::isBitFieldInvertedMask(~Mask)) {
7792 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7793 // where lsb(mask) == #shamt and masked bits of B are known zero.
7794 SDValue ShAmt = N00.getOperand(1);
7795 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7796 unsigned LSB = CountTrailingZeros_32(Mask);
7800 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7801 DAG.getConstant(~Mask, MVT::i32));
7803 // Do not add new nodes to DAG combiner worklist.
7804 DCI.CombineTo(N, Res, false);
7810 static SDValue PerformXORCombine(SDNode *N,
7811 TargetLowering::DAGCombinerInfo &DCI,
7812 const ARMSubtarget *Subtarget) {
7813 EVT VT = N->getValueType(0);
7814 SelectionDAG &DAG = DCI.DAG;
7816 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7819 if (!Subtarget->isThumb1Only()) {
7820 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7821 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7822 if (Result.getNode())
7829 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7830 /// the bits being cleared by the AND are not demanded by the BFI.
7831 static SDValue PerformBFICombine(SDNode *N,
7832 TargetLowering::DAGCombinerInfo &DCI) {
7833 SDValue N1 = N->getOperand(1);
7834 if (N1.getOpcode() == ISD::AND) {
7835 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7838 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7839 unsigned LSB = CountTrailingZeros_32(~InvMask);
7840 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7841 unsigned Mask = (1 << Width)-1;
7842 unsigned Mask2 = N11C->getZExtValue();
7843 if ((Mask & (~Mask2)) == 0)
7844 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7845 N->getOperand(0), N1.getOperand(0),
7851 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7852 /// ARMISD::VMOVRRD.
7853 static SDValue PerformVMOVRRDCombine(SDNode *N,
7854 TargetLowering::DAGCombinerInfo &DCI) {
7855 // vmovrrd(vmovdrr x, y) -> x,y
7856 SDValue InDouble = N->getOperand(0);
7857 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7858 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7860 // vmovrrd(load f64) -> (load i32), (load i32)
7861 SDNode *InNode = InDouble.getNode();
7862 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7863 InNode->getValueType(0) == MVT::f64 &&
7864 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7865 !cast<LoadSDNode>(InNode)->isVolatile()) {
7866 // TODO: Should this be done for non-FrameIndex operands?
7867 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7869 SelectionDAG &DAG = DCI.DAG;
7870 DebugLoc DL = LD->getDebugLoc();
7871 SDValue BasePtr = LD->getBasePtr();
7872 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7873 LD->getPointerInfo(), LD->isVolatile(),
7874 LD->isNonTemporal(), LD->isInvariant(),
7875 LD->getAlignment());
7877 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7878 DAG.getConstant(4, MVT::i32));
7879 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7880 LD->getPointerInfo(), LD->isVolatile(),
7881 LD->isNonTemporal(), LD->isInvariant(),
7882 std::min(4U, LD->getAlignment() / 2));
7884 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7885 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7886 DCI.RemoveFromWorklist(LD);
7894 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7895 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7896 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7897 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7898 SDValue Op0 = N->getOperand(0);
7899 SDValue Op1 = N->getOperand(1);
7900 if (Op0.getOpcode() == ISD::BITCAST)
7901 Op0 = Op0.getOperand(0);
7902 if (Op1.getOpcode() == ISD::BITCAST)
7903 Op1 = Op1.getOperand(0);
7904 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7905 Op0.getNode() == Op1.getNode() &&
7906 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7907 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7908 N->getValueType(0), Op0.getOperand(0));
7912 /// PerformSTORECombine - Target-specific dag combine xforms for
7914 static SDValue PerformSTORECombine(SDNode *N,
7915 TargetLowering::DAGCombinerInfo &DCI) {
7916 StoreSDNode *St = cast<StoreSDNode>(N);
7917 if (St->isVolatile())
7920 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7921 // pack all of the elements in one place. Next, store to memory in fewer
7923 SDValue StVal = St->getValue();
7924 EVT VT = StVal.getValueType();
7925 if (St->isTruncatingStore() && VT.isVector()) {
7926 SelectionDAG &DAG = DCI.DAG;
7927 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7928 EVT StVT = St->getMemoryVT();
7929 unsigned NumElems = VT.getVectorNumElements();
7930 assert(StVT != VT && "Cannot truncate to the same type");
7931 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7932 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7934 // From, To sizes and ElemCount must be pow of two
7935 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7937 // We are going to use the original vector elt for storing.
7938 // Accumulated smaller vector elements must be a multiple of the store size.
7939 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7941 unsigned SizeRatio = FromEltSz / ToEltSz;
7942 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7944 // Create a type on which we perform the shuffle.
7945 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7946 NumElems*SizeRatio);
7947 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7949 DebugLoc DL = St->getDebugLoc();
7950 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
7951 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
7952 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
7954 // Can't shuffle using an illegal type.
7955 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
7957 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
7958 DAG.getUNDEF(WideVec.getValueType()),
7960 // At this point all of the data is stored at the bottom of the
7961 // register. We now need to save it to mem.
7963 // Find the largest store unit
7964 MVT StoreType = MVT::i8;
7965 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
7966 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
7967 MVT Tp = (MVT::SimpleValueType)tp;
7968 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
7971 // Didn't find a legal store type.
7972 if (!TLI.isTypeLegal(StoreType))
7975 // Bitcast the original vector into a vector of store-size units
7976 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
7977 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
7978 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
7979 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
7980 SmallVector<SDValue, 8> Chains;
7981 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
7982 TLI.getPointerTy());
7983 SDValue BasePtr = St->getBasePtr();
7985 // Perform one or more big stores into memory.
7986 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
7987 for (unsigned I = 0; I < E; I++) {
7988 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
7989 StoreType, ShuffWide,
7990 DAG.getIntPtrConstant(I));
7991 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
7992 St->getPointerInfo(), St->isVolatile(),
7993 St->isNonTemporal(), St->getAlignment());
7994 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
7996 Chains.push_back(Ch);
7998 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8002 if (!ISD::isNormalStore(St))
8005 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8006 // ARM stores of arguments in the same cache line.
8007 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8008 StVal.getNode()->hasOneUse()) {
8009 SelectionDAG &DAG = DCI.DAG;
8010 DebugLoc DL = St->getDebugLoc();
8011 SDValue BasePtr = St->getBasePtr();
8012 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8013 StVal.getNode()->getOperand(0), BasePtr,
8014 St->getPointerInfo(), St->isVolatile(),
8015 St->isNonTemporal(), St->getAlignment());
8017 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8018 DAG.getConstant(4, MVT::i32));
8019 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8020 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8021 St->isNonTemporal(),
8022 std::min(4U, St->getAlignment() / 2));
8025 if (StVal.getValueType() != MVT::i64 ||
8026 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8029 // Bitcast an i64 store extracted from a vector to f64.
8030 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8031 SelectionDAG &DAG = DCI.DAG;
8032 DebugLoc dl = StVal.getDebugLoc();
8033 SDValue IntVec = StVal.getOperand(0);
8034 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8035 IntVec.getValueType().getVectorNumElements());
8036 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8037 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8038 Vec, StVal.getOperand(1));
8039 dl = N->getDebugLoc();
8040 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8041 // Make the DAGCombiner fold the bitcasts.
8042 DCI.AddToWorklist(Vec.getNode());
8043 DCI.AddToWorklist(ExtElt.getNode());
8044 DCI.AddToWorklist(V.getNode());
8045 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8046 St->getPointerInfo(), St->isVolatile(),
8047 St->isNonTemporal(), St->getAlignment(),
8051 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8052 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8053 /// i64 vector to have f64 elements, since the value can then be loaded
8054 /// directly into a VFP register.
8055 static bool hasNormalLoadOperand(SDNode *N) {
8056 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8057 for (unsigned i = 0; i < NumElts; ++i) {
8058 SDNode *Elt = N->getOperand(i).getNode();
8059 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8065 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8066 /// ISD::BUILD_VECTOR.
8067 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8068 TargetLowering::DAGCombinerInfo &DCI){
8069 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8070 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8071 // into a pair of GPRs, which is fine when the value is used as a scalar,
8072 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8073 SelectionDAG &DAG = DCI.DAG;
8074 if (N->getNumOperands() == 2) {
8075 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8080 // Load i64 elements as f64 values so that type legalization does not split
8081 // them up into i32 values.
8082 EVT VT = N->getValueType(0);
8083 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8085 DebugLoc dl = N->getDebugLoc();
8086 SmallVector<SDValue, 8> Ops;
8087 unsigned NumElts = VT.getVectorNumElements();
8088 for (unsigned i = 0; i < NumElts; ++i) {
8089 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8091 // Make the DAGCombiner fold the bitcast.
8092 DCI.AddToWorklist(V.getNode());
8094 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8095 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8096 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8099 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8100 /// ISD::INSERT_VECTOR_ELT.
8101 static SDValue PerformInsertEltCombine(SDNode *N,
8102 TargetLowering::DAGCombinerInfo &DCI) {
8103 // Bitcast an i64 load inserted into a vector to f64.
8104 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8105 EVT VT = N->getValueType(0);
8106 SDNode *Elt = N->getOperand(1).getNode();
8107 if (VT.getVectorElementType() != MVT::i64 ||
8108 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8111 SelectionDAG &DAG = DCI.DAG;
8112 DebugLoc dl = N->getDebugLoc();
8113 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8114 VT.getVectorNumElements());
8115 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8116 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8117 // Make the DAGCombiner fold the bitcasts.
8118 DCI.AddToWorklist(Vec.getNode());
8119 DCI.AddToWorklist(V.getNode());
8120 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8121 Vec, V, N->getOperand(2));
8122 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8125 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8126 /// ISD::VECTOR_SHUFFLE.
8127 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8128 // The LLVM shufflevector instruction does not require the shuffle mask
8129 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8130 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8131 // operands do not match the mask length, they are extended by concatenating
8132 // them with undef vectors. That is probably the right thing for other
8133 // targets, but for NEON it is better to concatenate two double-register
8134 // size vector operands into a single quad-register size vector. Do that
8135 // transformation here:
8136 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8137 // shuffle(concat(v1, v2), undef)
8138 SDValue Op0 = N->getOperand(0);
8139 SDValue Op1 = N->getOperand(1);
8140 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8141 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8142 Op0.getNumOperands() != 2 ||
8143 Op1.getNumOperands() != 2)
8145 SDValue Concat0Op1 = Op0.getOperand(1);
8146 SDValue Concat1Op1 = Op1.getOperand(1);
8147 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8148 Concat1Op1.getOpcode() != ISD::UNDEF)
8150 // Skip the transformation if any of the types are illegal.
8151 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8152 EVT VT = N->getValueType(0);
8153 if (!TLI.isTypeLegal(VT) ||
8154 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8155 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8158 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8159 Op0.getOperand(0), Op1.getOperand(0));
8160 // Translate the shuffle mask.
8161 SmallVector<int, 16> NewMask;
8162 unsigned NumElts = VT.getVectorNumElements();
8163 unsigned HalfElts = NumElts/2;
8164 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8165 for (unsigned n = 0; n < NumElts; ++n) {
8166 int MaskElt = SVN->getMaskElt(n);
8168 if (MaskElt < (int)HalfElts)
8170 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8171 NewElt = HalfElts + MaskElt - NumElts;
8172 NewMask.push_back(NewElt);
8174 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8175 DAG.getUNDEF(VT), NewMask.data());
8178 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8179 /// NEON load/store intrinsics to merge base address updates.
8180 static SDValue CombineBaseUpdate(SDNode *N,
8181 TargetLowering::DAGCombinerInfo &DCI) {
8182 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8185 SelectionDAG &DAG = DCI.DAG;
8186 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8187 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8188 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8189 SDValue Addr = N->getOperand(AddrOpIdx);
8191 // Search for a use of the address operand that is an increment.
8192 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8193 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8195 if (User->getOpcode() != ISD::ADD ||
8196 UI.getUse().getResNo() != Addr.getResNo())
8199 // Check that the add is independent of the load/store. Otherwise, folding
8200 // it would create a cycle.
8201 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8204 // Find the new opcode for the updating load/store.
8206 bool isLaneOp = false;
8207 unsigned NewOpc = 0;
8208 unsigned NumVecs = 0;
8210 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8212 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8213 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8215 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8217 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8219 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8221 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8222 NumVecs = 2; isLaneOp = true; break;
8223 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8224 NumVecs = 3; isLaneOp = true; break;
8225 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8226 NumVecs = 4; isLaneOp = true; break;
8227 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8228 NumVecs = 1; isLoad = false; break;
8229 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8230 NumVecs = 2; isLoad = false; break;
8231 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8232 NumVecs = 3; isLoad = false; break;
8233 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8234 NumVecs = 4; isLoad = false; break;
8235 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8236 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8237 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8238 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8239 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8240 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8244 switch (N->getOpcode()) {
8245 default: llvm_unreachable("unexpected opcode for Neon base update");
8246 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8247 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8248 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8252 // Find the size of memory referenced by the load/store.
8255 VecTy = N->getValueType(0);
8257 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8258 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8260 NumBytes /= VecTy.getVectorNumElements();
8262 // If the increment is a constant, it must match the memory ref size.
8263 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8264 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8265 uint64_t IncVal = CInc->getZExtValue();
8266 if (IncVal != NumBytes)
8268 } else if (NumBytes >= 3 * 16) {
8269 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8270 // separate instructions that make it harder to use a non-constant update.
8274 // Create the new updating load/store node.
8276 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8278 for (n = 0; n < NumResultVecs; ++n)
8280 Tys[n++] = MVT::i32;
8281 Tys[n] = MVT::Other;
8282 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8283 SmallVector<SDValue, 8> Ops;
8284 Ops.push_back(N->getOperand(0)); // incoming chain
8285 Ops.push_back(N->getOperand(AddrOpIdx));
8287 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8288 Ops.push_back(N->getOperand(i));
8290 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8291 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8292 Ops.data(), Ops.size(),
8293 MemInt->getMemoryVT(),
8294 MemInt->getMemOperand());
8297 std::vector<SDValue> NewResults;
8298 for (unsigned i = 0; i < NumResultVecs; ++i) {
8299 NewResults.push_back(SDValue(UpdN.getNode(), i));
8301 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8302 DCI.CombineTo(N, NewResults);
8303 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8310 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8311 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8312 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8314 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8315 SelectionDAG &DAG = DCI.DAG;
8316 EVT VT = N->getValueType(0);
8317 // vldN-dup instructions only support 64-bit vectors for N > 1.
8318 if (!VT.is64BitVector())
8321 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8322 SDNode *VLD = N->getOperand(0).getNode();
8323 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8325 unsigned NumVecs = 0;
8326 unsigned NewOpc = 0;
8327 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8328 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8330 NewOpc = ARMISD::VLD2DUP;
8331 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8333 NewOpc = ARMISD::VLD3DUP;
8334 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8336 NewOpc = ARMISD::VLD4DUP;
8341 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8342 // numbers match the load.
8343 unsigned VLDLaneNo =
8344 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8345 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8347 // Ignore uses of the chain result.
8348 if (UI.getUse().getResNo() == NumVecs)
8351 if (User->getOpcode() != ARMISD::VDUPLANE ||
8352 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8356 // Create the vldN-dup node.
8359 for (n = 0; n < NumVecs; ++n)
8361 Tys[n] = MVT::Other;
8362 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8363 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8364 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8365 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8366 Ops, 2, VLDMemInt->getMemoryVT(),
8367 VLDMemInt->getMemOperand());
8370 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8372 unsigned ResNo = UI.getUse().getResNo();
8373 // Ignore uses of the chain result.
8374 if (ResNo == NumVecs)
8377 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8380 // Now the vldN-lane intrinsic is dead except for its chain result.
8381 // Update uses of the chain.
8382 std::vector<SDValue> VLDDupResults;
8383 for (unsigned n = 0; n < NumVecs; ++n)
8384 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8385 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8386 DCI.CombineTo(VLD, VLDDupResults);
8391 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8392 /// ARMISD::VDUPLANE.
8393 static SDValue PerformVDUPLANECombine(SDNode *N,
8394 TargetLowering::DAGCombinerInfo &DCI) {
8395 SDValue Op = N->getOperand(0);
8397 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8398 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8399 if (CombineVLDDUP(N, DCI))
8400 return SDValue(N, 0);
8402 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8403 // redundant. Ignore bit_converts for now; element sizes are checked below.
8404 while (Op.getOpcode() == ISD::BITCAST)
8405 Op = Op.getOperand(0);
8406 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8409 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8410 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8411 // The canonical VMOV for a zero vector uses a 32-bit element size.
8412 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8414 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8416 EVT VT = N->getValueType(0);
8417 if (EltSize > VT.getVectorElementType().getSizeInBits())
8420 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8423 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8424 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8425 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8429 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8431 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8436 APFloat APF = C->getValueAPF();
8437 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8438 != APFloat::opOK || !isExact)
8441 c0 = (I == 0) ? cN : c0;
8442 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8449 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8450 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8451 /// when the VMUL has a constant operand that is a power of 2.
8453 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8454 /// vmul.f32 d16, d17, d16
8455 /// vcvt.s32.f32 d16, d16
8457 /// vcvt.s32.f32 d16, d16, #3
8458 static SDValue PerformVCVTCombine(SDNode *N,
8459 TargetLowering::DAGCombinerInfo &DCI,
8460 const ARMSubtarget *Subtarget) {
8461 SelectionDAG &DAG = DCI.DAG;
8462 SDValue Op = N->getOperand(0);
8464 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8465 Op.getOpcode() != ISD::FMUL)
8469 SDValue N0 = Op->getOperand(0);
8470 SDValue ConstVec = Op->getOperand(1);
8471 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8473 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8474 !isConstVecPow2(ConstVec, isSigned, C))
8477 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8478 Intrinsic::arm_neon_vcvtfp2fxu;
8479 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8481 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8482 DAG.getConstant(Log2_64(C), MVT::i32));
8485 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8486 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8487 /// when the VDIV has a constant operand that is a power of 2.
8489 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8490 /// vcvt.f32.s32 d16, d16
8491 /// vdiv.f32 d16, d17, d16
8493 /// vcvt.f32.s32 d16, d16, #3
8494 static SDValue PerformVDIVCombine(SDNode *N,
8495 TargetLowering::DAGCombinerInfo &DCI,
8496 const ARMSubtarget *Subtarget) {
8497 SelectionDAG &DAG = DCI.DAG;
8498 SDValue Op = N->getOperand(0);
8499 unsigned OpOpcode = Op.getNode()->getOpcode();
8501 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8502 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8506 SDValue ConstVec = N->getOperand(1);
8507 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8509 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8510 !isConstVecPow2(ConstVec, isSigned, C))
8513 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8514 Intrinsic::arm_neon_vcvtfxu2fp;
8515 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8517 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8518 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8521 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8522 /// operand of a vector shift operation, where all the elements of the
8523 /// build_vector must have the same constant integer value.
8524 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8525 // Ignore bit_converts.
8526 while (Op.getOpcode() == ISD::BITCAST)
8527 Op = Op.getOperand(0);
8528 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8529 APInt SplatBits, SplatUndef;
8530 unsigned SplatBitSize;
8532 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8533 HasAnyUndefs, ElementBits) ||
8534 SplatBitSize > ElementBits)
8536 Cnt = SplatBits.getSExtValue();
8540 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8541 /// operand of a vector shift left operation. That value must be in the range:
8542 /// 0 <= Value < ElementBits for a left shift; or
8543 /// 0 <= Value <= ElementBits for a long left shift.
8544 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8545 assert(VT.isVector() && "vector shift count is not a vector type");
8546 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8547 if (! getVShiftImm(Op, ElementBits, Cnt))
8549 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8552 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8553 /// operand of a vector shift right operation. For a shift opcode, the value
8554 /// is positive, but for an intrinsic the value count must be negative. The
8555 /// absolute value must be in the range:
8556 /// 1 <= |Value| <= ElementBits for a right shift; or
8557 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8558 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8560 assert(VT.isVector() && "vector shift count is not a vector type");
8561 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8562 if (! getVShiftImm(Op, ElementBits, Cnt))
8566 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8569 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8570 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8571 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8574 // Don't do anything for most intrinsics.
8577 // Vector shifts: check for immediate versions and lower them.
8578 // Note: This is done during DAG combining instead of DAG legalizing because
8579 // the build_vectors for 64-bit vector element shift counts are generally
8580 // not legal, and it is hard to see their values after they get legalized to
8581 // loads from a constant pool.
8582 case Intrinsic::arm_neon_vshifts:
8583 case Intrinsic::arm_neon_vshiftu:
8584 case Intrinsic::arm_neon_vshiftls:
8585 case Intrinsic::arm_neon_vshiftlu:
8586 case Intrinsic::arm_neon_vshiftn:
8587 case Intrinsic::arm_neon_vrshifts:
8588 case Intrinsic::arm_neon_vrshiftu:
8589 case Intrinsic::arm_neon_vrshiftn:
8590 case Intrinsic::arm_neon_vqshifts:
8591 case Intrinsic::arm_neon_vqshiftu:
8592 case Intrinsic::arm_neon_vqshiftsu:
8593 case Intrinsic::arm_neon_vqshiftns:
8594 case Intrinsic::arm_neon_vqshiftnu:
8595 case Intrinsic::arm_neon_vqshiftnsu:
8596 case Intrinsic::arm_neon_vqrshiftns:
8597 case Intrinsic::arm_neon_vqrshiftnu:
8598 case Intrinsic::arm_neon_vqrshiftnsu: {
8599 EVT VT = N->getOperand(1).getValueType();
8601 unsigned VShiftOpc = 0;
8604 case Intrinsic::arm_neon_vshifts:
8605 case Intrinsic::arm_neon_vshiftu:
8606 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8607 VShiftOpc = ARMISD::VSHL;
8610 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8611 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8612 ARMISD::VSHRs : ARMISD::VSHRu);
8617 case Intrinsic::arm_neon_vshiftls:
8618 case Intrinsic::arm_neon_vshiftlu:
8619 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8621 llvm_unreachable("invalid shift count for vshll intrinsic");
8623 case Intrinsic::arm_neon_vrshifts:
8624 case Intrinsic::arm_neon_vrshiftu:
8625 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8629 case Intrinsic::arm_neon_vqshifts:
8630 case Intrinsic::arm_neon_vqshiftu:
8631 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8635 case Intrinsic::arm_neon_vqshiftsu:
8636 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8638 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8640 case Intrinsic::arm_neon_vshiftn:
8641 case Intrinsic::arm_neon_vrshiftn:
8642 case Intrinsic::arm_neon_vqshiftns:
8643 case Intrinsic::arm_neon_vqshiftnu:
8644 case Intrinsic::arm_neon_vqshiftnsu:
8645 case Intrinsic::arm_neon_vqrshiftns:
8646 case Intrinsic::arm_neon_vqrshiftnu:
8647 case Intrinsic::arm_neon_vqrshiftnsu:
8648 // Narrowing shifts require an immediate right shift.
8649 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8651 llvm_unreachable("invalid shift count for narrowing vector shift "
8655 llvm_unreachable("unhandled vector shift");
8659 case Intrinsic::arm_neon_vshifts:
8660 case Intrinsic::arm_neon_vshiftu:
8661 // Opcode already set above.
8663 case Intrinsic::arm_neon_vshiftls:
8664 case Intrinsic::arm_neon_vshiftlu:
8665 if (Cnt == VT.getVectorElementType().getSizeInBits())
8666 VShiftOpc = ARMISD::VSHLLi;
8668 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8669 ARMISD::VSHLLs : ARMISD::VSHLLu);
8671 case Intrinsic::arm_neon_vshiftn:
8672 VShiftOpc = ARMISD::VSHRN; break;
8673 case Intrinsic::arm_neon_vrshifts:
8674 VShiftOpc = ARMISD::VRSHRs; break;
8675 case Intrinsic::arm_neon_vrshiftu:
8676 VShiftOpc = ARMISD::VRSHRu; break;
8677 case Intrinsic::arm_neon_vrshiftn:
8678 VShiftOpc = ARMISD::VRSHRN; break;
8679 case Intrinsic::arm_neon_vqshifts:
8680 VShiftOpc = ARMISD::VQSHLs; break;
8681 case Intrinsic::arm_neon_vqshiftu:
8682 VShiftOpc = ARMISD::VQSHLu; break;
8683 case Intrinsic::arm_neon_vqshiftsu:
8684 VShiftOpc = ARMISD::VQSHLsu; break;
8685 case Intrinsic::arm_neon_vqshiftns:
8686 VShiftOpc = ARMISD::VQSHRNs; break;
8687 case Intrinsic::arm_neon_vqshiftnu:
8688 VShiftOpc = ARMISD::VQSHRNu; break;
8689 case Intrinsic::arm_neon_vqshiftnsu:
8690 VShiftOpc = ARMISD::VQSHRNsu; break;
8691 case Intrinsic::arm_neon_vqrshiftns:
8692 VShiftOpc = ARMISD::VQRSHRNs; break;
8693 case Intrinsic::arm_neon_vqrshiftnu:
8694 VShiftOpc = ARMISD::VQRSHRNu; break;
8695 case Intrinsic::arm_neon_vqrshiftnsu:
8696 VShiftOpc = ARMISD::VQRSHRNsu; break;
8699 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8700 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8703 case Intrinsic::arm_neon_vshiftins: {
8704 EVT VT = N->getOperand(1).getValueType();
8706 unsigned VShiftOpc = 0;
8708 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8709 VShiftOpc = ARMISD::VSLI;
8710 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8711 VShiftOpc = ARMISD::VSRI;
8713 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8716 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8717 N->getOperand(1), N->getOperand(2),
8718 DAG.getConstant(Cnt, MVT::i32));
8721 case Intrinsic::arm_neon_vqrshifts:
8722 case Intrinsic::arm_neon_vqrshiftu:
8723 // No immediate versions of these to check for.
8730 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8731 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8732 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8733 /// vector element shift counts are generally not legal, and it is hard to see
8734 /// their values after they get legalized to loads from a constant pool.
8735 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8736 const ARMSubtarget *ST) {
8737 EVT VT = N->getValueType(0);
8738 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8739 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8740 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8741 SDValue N1 = N->getOperand(1);
8742 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8743 SDValue N0 = N->getOperand(0);
8744 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8745 DAG.MaskedValueIsZero(N0.getOperand(0),
8746 APInt::getHighBitsSet(32, 16)))
8747 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8751 // Nothing to be done for scalar shifts.
8752 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8753 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8756 assert(ST->hasNEON() && "unexpected vector shift");
8759 switch (N->getOpcode()) {
8760 default: llvm_unreachable("unexpected shift opcode");
8763 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8764 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8765 DAG.getConstant(Cnt, MVT::i32));
8770 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8771 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8772 ARMISD::VSHRs : ARMISD::VSHRu);
8773 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8774 DAG.getConstant(Cnt, MVT::i32));
8780 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8781 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8782 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8783 const ARMSubtarget *ST) {
8784 SDValue N0 = N->getOperand(0);
8786 // Check for sign- and zero-extensions of vector extract operations of 8-
8787 // and 16-bit vector elements. NEON supports these directly. They are
8788 // handled during DAG combining because type legalization will promote them
8789 // to 32-bit types and it is messy to recognize the operations after that.
8790 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8791 SDValue Vec = N0.getOperand(0);
8792 SDValue Lane = N0.getOperand(1);
8793 EVT VT = N->getValueType(0);
8794 EVT EltVT = N0.getValueType();
8795 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8797 if (VT == MVT::i32 &&
8798 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8799 TLI.isTypeLegal(Vec.getValueType()) &&
8800 isa<ConstantSDNode>(Lane)) {
8803 switch (N->getOpcode()) {
8804 default: llvm_unreachable("unexpected opcode");
8805 case ISD::SIGN_EXTEND:
8806 Opc = ARMISD::VGETLANEs;
8808 case ISD::ZERO_EXTEND:
8809 case ISD::ANY_EXTEND:
8810 Opc = ARMISD::VGETLANEu;
8813 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8820 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8821 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8822 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8823 const ARMSubtarget *ST) {
8824 // If the target supports NEON, try to use vmax/vmin instructions for f32
8825 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8826 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8827 // a NaN; only do the transformation when it matches that behavior.
8829 // For now only do this when using NEON for FP operations; if using VFP, it
8830 // is not obvious that the benefit outweighs the cost of switching to the
8832 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8833 N->getValueType(0) != MVT::f32)
8836 SDValue CondLHS = N->getOperand(0);
8837 SDValue CondRHS = N->getOperand(1);
8838 SDValue LHS = N->getOperand(2);
8839 SDValue RHS = N->getOperand(3);
8840 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8842 unsigned Opcode = 0;
8844 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8845 IsReversed = false; // x CC y ? x : y
8846 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8847 IsReversed = true ; // x CC y ? y : x
8861 // If LHS is NaN, an ordered comparison will be false and the result will
8862 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8863 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8864 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8865 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8867 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8868 // will return -0, so vmin can only be used for unsafe math or if one of
8869 // the operands is known to be nonzero.
8870 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8871 !DAG.getTarget().Options.UnsafeFPMath &&
8872 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8874 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8883 // If LHS is NaN, an ordered comparison will be false and the result will
8884 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8885 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8886 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8887 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8889 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8890 // will return +0, so vmax can only be used for unsafe math or if one of
8891 // the operands is known to be nonzero.
8892 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8893 !DAG.getTarget().Options.UnsafeFPMath &&
8894 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8896 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8902 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8905 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8907 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8908 SDValue Cmp = N->getOperand(4);
8909 if (Cmp.getOpcode() != ARMISD::CMPZ)
8910 // Only looking at EQ and NE cases.
8913 EVT VT = N->getValueType(0);
8914 DebugLoc dl = N->getDebugLoc();
8915 SDValue LHS = Cmp.getOperand(0);
8916 SDValue RHS = Cmp.getOperand(1);
8917 SDValue FalseVal = N->getOperand(0);
8918 SDValue TrueVal = N->getOperand(1);
8919 SDValue ARMcc = N->getOperand(2);
8920 ARMCC::CondCodes CC =
8921 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8939 /// FIXME: Turn this into a target neutral optimization?
8941 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8942 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8943 N->getOperand(3), Cmp);
8944 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8946 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8947 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8948 N->getOperand(3), NewCmp);
8951 if (Res.getNode()) {
8952 APInt KnownZero, KnownOne;
8953 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8954 // Capture demanded bits information that would be otherwise lost.
8955 if (KnownZero == 0xfffffffe)
8956 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8957 DAG.getValueType(MVT::i1));
8958 else if (KnownZero == 0xffffff00)
8959 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8960 DAG.getValueType(MVT::i8));
8961 else if (KnownZero == 0xffff0000)
8962 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8963 DAG.getValueType(MVT::i16));
8969 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8970 DAGCombinerInfo &DCI) const {
8971 switch (N->getOpcode()) {
8973 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
8974 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8975 case ISD::SUB: return PerformSUBCombine(N, DCI);
8976 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8977 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8978 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8979 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8980 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8981 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8982 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8983 case ISD::STORE: return PerformSTORECombine(N, DCI);
8984 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8985 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8986 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8987 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8988 case ISD::FP_TO_SINT:
8989 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8990 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8991 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8994 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8995 case ISD::SIGN_EXTEND:
8996 case ISD::ZERO_EXTEND:
8997 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8998 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8999 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9000 case ARMISD::VLD2DUP:
9001 case ARMISD::VLD3DUP:
9002 case ARMISD::VLD4DUP:
9003 return CombineBaseUpdate(N, DCI);
9004 case ISD::INTRINSIC_VOID:
9005 case ISD::INTRINSIC_W_CHAIN:
9006 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9007 case Intrinsic::arm_neon_vld1:
9008 case Intrinsic::arm_neon_vld2:
9009 case Intrinsic::arm_neon_vld3:
9010 case Intrinsic::arm_neon_vld4:
9011 case Intrinsic::arm_neon_vld2lane:
9012 case Intrinsic::arm_neon_vld3lane:
9013 case Intrinsic::arm_neon_vld4lane:
9014 case Intrinsic::arm_neon_vst1:
9015 case Intrinsic::arm_neon_vst2:
9016 case Intrinsic::arm_neon_vst3:
9017 case Intrinsic::arm_neon_vst4:
9018 case Intrinsic::arm_neon_vst2lane:
9019 case Intrinsic::arm_neon_vst3lane:
9020 case Intrinsic::arm_neon_vst4lane:
9021 return CombineBaseUpdate(N, DCI);
9029 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9031 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9034 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9035 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9036 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9038 switch (VT.getSimpleVT().SimpleTy) {
9044 // Unaligned access can use (for example) LRDB, LRDH, LDR
9045 return AllowsUnaligned;
9048 // For any little-endian targets with neon, we can support unaligned ld/st
9049 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9050 // A big-endian target may also explictly support unaligned accesses
9051 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9055 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9056 unsigned AlignCheck) {
9057 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9058 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9061 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9062 unsigned DstAlign, unsigned SrcAlign,
9065 MachineFunction &MF) const {
9066 const Function *F = MF.getFunction();
9068 // See if we can use NEON instructions for this...
9070 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9071 Subtarget->hasNEON()) {
9072 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9074 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9079 // Lowering to i32/i16 if the size permits.
9082 } else if (Size >= 2) {
9086 // Let the target-independent logic figure it out.
9090 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9095 switch (VT.getSimpleVT().SimpleTy) {
9096 default: return false;
9111 if ((V & (Scale - 1)) != 0)
9114 return V == (V & ((1LL << 5) - 1));
9117 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9118 const ARMSubtarget *Subtarget) {
9125 switch (VT.getSimpleVT().SimpleTy) {
9126 default: return false;
9131 // + imm12 or - imm8
9133 return V == (V & ((1LL << 8) - 1));
9134 return V == (V & ((1LL << 12) - 1));
9137 // Same as ARM mode. FIXME: NEON?
9138 if (!Subtarget->hasVFP2())
9143 return V == (V & ((1LL << 8) - 1));
9147 /// isLegalAddressImmediate - Return true if the integer value can be used
9148 /// as the offset of the target addressing mode for load / store of the
9150 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9151 const ARMSubtarget *Subtarget) {
9158 if (Subtarget->isThumb1Only())
9159 return isLegalT1AddressImmediate(V, VT);
9160 else if (Subtarget->isThumb2())
9161 return isLegalT2AddressImmediate(V, VT, Subtarget);
9166 switch (VT.getSimpleVT().SimpleTy) {
9167 default: return false;
9172 return V == (V & ((1LL << 12) - 1));
9175 return V == (V & ((1LL << 8) - 1));
9178 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9183 return V == (V & ((1LL << 8) - 1));
9187 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9189 int Scale = AM.Scale;
9193 switch (VT.getSimpleVT().SimpleTy) {
9194 default: return false;
9203 return Scale == 2 || Scale == 4 || Scale == 8;
9206 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9210 // Note, we allow "void" uses (basically, uses that aren't loads or
9211 // stores), because arm allows folding a scale into many arithmetic
9212 // operations. This should be made more precise and revisited later.
9214 // Allow r << imm, but the imm has to be a multiple of two.
9215 if (Scale & 1) return false;
9216 return isPowerOf2_32(Scale);
9220 /// isLegalAddressingMode - Return true if the addressing mode represented
9221 /// by AM is legal for this target, for a load/store of the specified type.
9222 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9224 EVT VT = getValueType(Ty, true);
9225 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9228 // Can never fold addr of global into load/store.
9233 case 0: // no scale reg, must be "r+i" or "r", or "i".
9236 if (Subtarget->isThumb1Only())
9240 // ARM doesn't support any R+R*scale+imm addr modes.
9247 if (Subtarget->isThumb2())
9248 return isLegalT2ScaledAddressingMode(AM, VT);
9250 int Scale = AM.Scale;
9251 switch (VT.getSimpleVT().SimpleTy) {
9252 default: return false;
9256 if (Scale < 0) Scale = -Scale;
9260 return isPowerOf2_32(Scale & ~1);
9264 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9269 // Note, we allow "void" uses (basically, uses that aren't loads or
9270 // stores), because arm allows folding a scale into many arithmetic
9271 // operations. This should be made more precise and revisited later.
9273 // Allow r << imm, but the imm has to be a multiple of two.
9274 if (Scale & 1) return false;
9275 return isPowerOf2_32(Scale);
9281 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9282 /// icmp immediate, that is the target has icmp instructions which can compare
9283 /// a register against the immediate without having to materialize the
9284 /// immediate into a register.
9285 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9286 // Thumb2 and ARM modes can use cmn for negative immediates.
9287 if (!Subtarget->isThumb())
9288 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9289 if (Subtarget->isThumb2())
9290 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9291 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9292 return Imm >= 0 && Imm <= 255;
9295 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9296 /// *or sub* immediate, that is the target has add or sub instructions which can
9297 /// add a register with the immediate without having to materialize the
9298 /// immediate into a register.
9299 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9300 // Same encoding for add/sub, just flip the sign.
9301 int64_t AbsImm = llvm::abs64(Imm);
9302 if (!Subtarget->isThumb())
9303 return ARM_AM::getSOImmVal(AbsImm) != -1;
9304 if (Subtarget->isThumb2())
9305 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9306 // Thumb1 only has 8-bit unsigned immediate.
9307 return AbsImm >= 0 && AbsImm <= 255;
9310 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9311 bool isSEXTLoad, SDValue &Base,
9312 SDValue &Offset, bool &isInc,
9313 SelectionDAG &DAG) {
9314 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9317 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9319 Base = Ptr->getOperand(0);
9320 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9321 int RHSC = (int)RHS->getZExtValue();
9322 if (RHSC < 0 && RHSC > -256) {
9323 assert(Ptr->getOpcode() == ISD::ADD);
9325 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9329 isInc = (Ptr->getOpcode() == ISD::ADD);
9330 Offset = Ptr->getOperand(1);
9332 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9334 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9335 int RHSC = (int)RHS->getZExtValue();
9336 if (RHSC < 0 && RHSC > -0x1000) {
9337 assert(Ptr->getOpcode() == ISD::ADD);
9339 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9340 Base = Ptr->getOperand(0);
9345 if (Ptr->getOpcode() == ISD::ADD) {
9347 ARM_AM::ShiftOpc ShOpcVal=
9348 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9349 if (ShOpcVal != ARM_AM::no_shift) {
9350 Base = Ptr->getOperand(1);
9351 Offset = Ptr->getOperand(0);
9353 Base = Ptr->getOperand(0);
9354 Offset = Ptr->getOperand(1);
9359 isInc = (Ptr->getOpcode() == ISD::ADD);
9360 Base = Ptr->getOperand(0);
9361 Offset = Ptr->getOperand(1);
9365 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9369 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9370 bool isSEXTLoad, SDValue &Base,
9371 SDValue &Offset, bool &isInc,
9372 SelectionDAG &DAG) {
9373 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9376 Base = Ptr->getOperand(0);
9377 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9378 int RHSC = (int)RHS->getZExtValue();
9379 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9380 assert(Ptr->getOpcode() == ISD::ADD);
9382 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9384 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9385 isInc = Ptr->getOpcode() == ISD::ADD;
9386 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9394 /// getPreIndexedAddressParts - returns true by value, base pointer and
9395 /// offset pointer and addressing mode by reference if the node's address
9396 /// can be legally represented as pre-indexed load / store address.
9398 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9400 ISD::MemIndexedMode &AM,
9401 SelectionDAG &DAG) const {
9402 if (Subtarget->isThumb1Only())
9407 bool isSEXTLoad = false;
9408 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9409 Ptr = LD->getBasePtr();
9410 VT = LD->getMemoryVT();
9411 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9412 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9413 Ptr = ST->getBasePtr();
9414 VT = ST->getMemoryVT();
9419 bool isLegal = false;
9420 if (Subtarget->isThumb2())
9421 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9422 Offset, isInc, DAG);
9424 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9425 Offset, isInc, DAG);
9429 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9433 /// getPostIndexedAddressParts - returns true by value, base pointer and
9434 /// offset pointer and addressing mode by reference if this node can be
9435 /// combined with a load / store to form a post-indexed load / store.
9436 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9439 ISD::MemIndexedMode &AM,
9440 SelectionDAG &DAG) const {
9441 if (Subtarget->isThumb1Only())
9446 bool isSEXTLoad = false;
9447 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9448 VT = LD->getMemoryVT();
9449 Ptr = LD->getBasePtr();
9450 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9451 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9452 VT = ST->getMemoryVT();
9453 Ptr = ST->getBasePtr();
9458 bool isLegal = false;
9459 if (Subtarget->isThumb2())
9460 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9463 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9469 // Swap base ptr and offset to catch more post-index load / store when
9470 // it's legal. In Thumb2 mode, offset must be an immediate.
9471 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9472 !Subtarget->isThumb2())
9473 std::swap(Base, Offset);
9475 // Post-indexed load / store update the base pointer.
9480 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9484 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9487 const SelectionDAG &DAG,
9488 unsigned Depth) const {
9489 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9490 switch (Op.getOpcode()) {
9492 case ARMISD::CMOV: {
9493 // Bits are known zero/one if known on the LHS and RHS.
9494 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9495 if (KnownZero == 0 && KnownOne == 0) return;
9497 APInt KnownZeroRHS, KnownOneRHS;
9498 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9499 KnownZero &= KnownZeroRHS;
9500 KnownOne &= KnownOneRHS;
9506 //===----------------------------------------------------------------------===//
9507 // ARM Inline Assembly Support
9508 //===----------------------------------------------------------------------===//
9510 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9511 // Looking for "rev" which is V6+.
9512 if (!Subtarget->hasV6Ops())
9515 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9516 std::string AsmStr = IA->getAsmString();
9517 SmallVector<StringRef, 4> AsmPieces;
9518 SplitString(AsmStr, AsmPieces, ";\n");
9520 switch (AsmPieces.size()) {
9521 default: return false;
9523 AsmStr = AsmPieces[0];
9525 SplitString(AsmStr, AsmPieces, " \t,");
9528 if (AsmPieces.size() == 3 &&
9529 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9530 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9531 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9532 if (Ty && Ty->getBitWidth() == 32)
9533 return IntrinsicLowering::LowerToByteSwap(CI);
9541 /// getConstraintType - Given a constraint letter, return the type of
9542 /// constraint it is for this target.
9543 ARMTargetLowering::ConstraintType
9544 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9545 if (Constraint.size() == 1) {
9546 switch (Constraint[0]) {
9548 case 'l': return C_RegisterClass;
9549 case 'w': return C_RegisterClass;
9550 case 'h': return C_RegisterClass;
9551 case 'x': return C_RegisterClass;
9552 case 't': return C_RegisterClass;
9553 case 'j': return C_Other; // Constant for movw.
9554 // An address with a single base register. Due to the way we
9555 // currently handle addresses it is the same as an 'r' memory constraint.
9556 case 'Q': return C_Memory;
9558 } else if (Constraint.size() == 2) {
9559 switch (Constraint[0]) {
9561 // All 'U+' constraints are addresses.
9562 case 'U': return C_Memory;
9565 return TargetLowering::getConstraintType(Constraint);
9568 /// Examine constraint type and operand type and determine a weight value.
9569 /// This object must already have been set up with the operand type
9570 /// and the current alternative constraint selected.
9571 TargetLowering::ConstraintWeight
9572 ARMTargetLowering::getSingleConstraintMatchWeight(
9573 AsmOperandInfo &info, const char *constraint) const {
9574 ConstraintWeight weight = CW_Invalid;
9575 Value *CallOperandVal = info.CallOperandVal;
9576 // If we don't have a value, we can't do a match,
9577 // but allow it at the lowest weight.
9578 if (CallOperandVal == NULL)
9580 Type *type = CallOperandVal->getType();
9581 // Look at the constraint type.
9582 switch (*constraint) {
9584 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9587 if (type->isIntegerTy()) {
9588 if (Subtarget->isThumb())
9589 weight = CW_SpecificReg;
9591 weight = CW_Register;
9595 if (type->isFloatingPointTy())
9596 weight = CW_Register;
9602 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9604 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9606 if (Constraint.size() == 1) {
9607 // GCC ARM Constraint Letters
9608 switch (Constraint[0]) {
9609 case 'l': // Low regs or general regs.
9610 if (Subtarget->isThumb())
9611 return RCPair(0U, &ARM::tGPRRegClass);
9612 return RCPair(0U, &ARM::GPRRegClass);
9613 case 'h': // High regs or no regs.
9614 if (Subtarget->isThumb())
9615 return RCPair(0U, &ARM::hGPRRegClass);
9618 return RCPair(0U, &ARM::GPRRegClass);
9621 return RCPair(0U, &ARM::SPRRegClass);
9622 if (VT.getSizeInBits() == 64)
9623 return RCPair(0U, &ARM::DPRRegClass);
9624 if (VT.getSizeInBits() == 128)
9625 return RCPair(0U, &ARM::QPRRegClass);
9629 return RCPair(0U, &ARM::SPR_8RegClass);
9630 if (VT.getSizeInBits() == 64)
9631 return RCPair(0U, &ARM::DPR_8RegClass);
9632 if (VT.getSizeInBits() == 128)
9633 return RCPair(0U, &ARM::QPR_8RegClass);
9637 return RCPair(0U, &ARM::SPRRegClass);
9641 if (StringRef("{cc}").equals_lower(Constraint))
9642 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9644 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9647 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9648 /// vector. If it is invalid, don't add anything to Ops.
9649 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9650 std::string &Constraint,
9651 std::vector<SDValue>&Ops,
9652 SelectionDAG &DAG) const {
9653 SDValue Result(0, 0);
9655 // Currently only support length 1 constraints.
9656 if (Constraint.length() != 1) return;
9658 char ConstraintLetter = Constraint[0];
9659 switch (ConstraintLetter) {
9662 case 'I': case 'J': case 'K': case 'L':
9663 case 'M': case 'N': case 'O':
9664 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9668 int64_t CVal64 = C->getSExtValue();
9669 int CVal = (int) CVal64;
9670 // None of these constraints allow values larger than 32 bits. Check
9671 // that the value fits in an int.
9675 switch (ConstraintLetter) {
9677 // Constant suitable for movw, must be between 0 and
9679 if (Subtarget->hasV6T2Ops())
9680 if (CVal >= 0 && CVal <= 65535)
9684 if (Subtarget->isThumb1Only()) {
9685 // This must be a constant between 0 and 255, for ADD
9687 if (CVal >= 0 && CVal <= 255)
9689 } else if (Subtarget->isThumb2()) {
9690 // A constant that can be used as an immediate value in a
9691 // data-processing instruction.
9692 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9695 // A constant that can be used as an immediate value in a
9696 // data-processing instruction.
9697 if (ARM_AM::getSOImmVal(CVal) != -1)
9703 if (Subtarget->isThumb()) { // FIXME thumb2
9704 // This must be a constant between -255 and -1, for negated ADD
9705 // immediates. This can be used in GCC with an "n" modifier that
9706 // prints the negated value, for use with SUB instructions. It is
9707 // not useful otherwise but is implemented for compatibility.
9708 if (CVal >= -255 && CVal <= -1)
9711 // This must be a constant between -4095 and 4095. It is not clear
9712 // what this constraint is intended for. Implemented for
9713 // compatibility with GCC.
9714 if (CVal >= -4095 && CVal <= 4095)
9720 if (Subtarget->isThumb1Only()) {
9721 // A 32-bit value where only one byte has a nonzero value. Exclude
9722 // zero to match GCC. This constraint is used by GCC internally for
9723 // constants that can be loaded with a move/shift combination.
9724 // It is not useful otherwise but is implemented for compatibility.
9725 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9727 } else if (Subtarget->isThumb2()) {
9728 // A constant whose bitwise inverse can be used as an immediate
9729 // value in a data-processing instruction. This can be used in GCC
9730 // with a "B" modifier that prints the inverted value, for use with
9731 // BIC and MVN instructions. It is not useful otherwise but is
9732 // implemented for compatibility.
9733 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9736 // A constant whose bitwise inverse can be used as an immediate
9737 // value in a data-processing instruction. This can be used in GCC
9738 // with a "B" modifier that prints the inverted value, for use with
9739 // BIC and MVN instructions. It is not useful otherwise but is
9740 // implemented for compatibility.
9741 if (ARM_AM::getSOImmVal(~CVal) != -1)
9747 if (Subtarget->isThumb1Only()) {
9748 // This must be a constant between -7 and 7,
9749 // for 3-operand ADD/SUB immediate instructions.
9750 if (CVal >= -7 && CVal < 7)
9752 } else if (Subtarget->isThumb2()) {
9753 // A constant whose negation can be used as an immediate value in a
9754 // data-processing instruction. This can be used in GCC with an "n"
9755 // modifier that prints the negated value, for use with SUB
9756 // instructions. It is not useful otherwise but is implemented for
9758 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9761 // A constant whose negation can be used as an immediate value in a
9762 // data-processing instruction. This can be used in GCC with an "n"
9763 // modifier that prints the negated value, for use with SUB
9764 // instructions. It is not useful otherwise but is implemented for
9766 if (ARM_AM::getSOImmVal(-CVal) != -1)
9772 if (Subtarget->isThumb()) { // FIXME thumb2
9773 // This must be a multiple of 4 between 0 and 1020, for
9774 // ADD sp + immediate.
9775 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9778 // A power of two or a constant between 0 and 32. This is used in
9779 // GCC for the shift amount on shifted register operands, but it is
9780 // useful in general for any shift amounts.
9781 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9787 if (Subtarget->isThumb()) { // FIXME thumb2
9788 // This must be a constant between 0 and 31, for shift amounts.
9789 if (CVal >= 0 && CVal <= 31)
9795 if (Subtarget->isThumb()) { // FIXME thumb2
9796 // This must be a multiple of 4 between -508 and 508, for
9797 // ADD/SUB sp = sp + immediate.
9798 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9803 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9807 if (Result.getNode()) {
9808 Ops.push_back(Result);
9811 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9815 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9816 // The ARM target isn't yet aware of offsets.
9820 bool ARM::isBitFieldInvertedMask(unsigned v) {
9821 if (v == 0xffffffff)
9823 // there can be 1's on either or both "outsides", all the "inside"
9825 unsigned int lsb = 0, msb = 31;
9826 while (v & (1 << msb)) --msb;
9827 while (v & (1 << lsb)) ++lsb;
9828 for (unsigned int i = lsb; i <= msb; ++i) {
9835 /// isFPImmLegal - Returns true if the target can instruction select the
9836 /// specified FP immediate natively. If false, the legalizer will
9837 /// materialize the FP immediate as a load from a constant pool.
9838 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9839 if (!Subtarget->hasVFP3())
9842 return ARM_AM::getFP32Imm(Imm) != -1;
9844 return ARM_AM::getFP64Imm(Imm) != -1;
9848 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9849 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9850 /// specified in the intrinsic calls.
9851 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9853 unsigned Intrinsic) const {
9854 switch (Intrinsic) {
9855 case Intrinsic::arm_neon_vld1:
9856 case Intrinsic::arm_neon_vld2:
9857 case Intrinsic::arm_neon_vld3:
9858 case Intrinsic::arm_neon_vld4:
9859 case Intrinsic::arm_neon_vld2lane:
9860 case Intrinsic::arm_neon_vld3lane:
9861 case Intrinsic::arm_neon_vld4lane: {
9862 Info.opc = ISD::INTRINSIC_W_CHAIN;
9863 // Conservatively set memVT to the entire set of vectors loaded.
9864 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9865 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9866 Info.ptrVal = I.getArgOperand(0);
9868 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9869 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9870 Info.vol = false; // volatile loads with NEON intrinsics not supported
9871 Info.readMem = true;
9872 Info.writeMem = false;
9875 case Intrinsic::arm_neon_vst1:
9876 case Intrinsic::arm_neon_vst2:
9877 case Intrinsic::arm_neon_vst3:
9878 case Intrinsic::arm_neon_vst4:
9879 case Intrinsic::arm_neon_vst2lane:
9880 case Intrinsic::arm_neon_vst3lane:
9881 case Intrinsic::arm_neon_vst4lane: {
9882 Info.opc = ISD::INTRINSIC_VOID;
9883 // Conservatively set memVT to the entire set of vectors stored.
9884 unsigned NumElts = 0;
9885 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9886 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9887 if (!ArgTy->isVectorTy())
9889 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9891 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9892 Info.ptrVal = I.getArgOperand(0);
9894 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9895 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9896 Info.vol = false; // volatile stores with NEON intrinsics not supported
9897 Info.readMem = false;
9898 Info.writeMem = true;
9901 case Intrinsic::arm_strexd: {
9902 Info.opc = ISD::INTRINSIC_W_CHAIN;
9903 Info.memVT = MVT::i64;
9904 Info.ptrVal = I.getArgOperand(2);
9908 Info.readMem = false;
9909 Info.writeMem = true;
9912 case Intrinsic::arm_ldrexd: {
9913 Info.opc = ISD::INTRINSIC_W_CHAIN;
9914 Info.memVT = MVT::i64;
9915 Info.ptrVal = I.getArgOperand(0);
9919 Info.readMem = true;
9920 Info.writeMem = false;