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"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "MCTargetDesc/ARMAddressingModes.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/IntrinsicLowering.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/ADT/StringExtras.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
55 STATISTIC(NumTailCalls, "Number of tail calls");
56 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
58 // This option should go away when tail calls fully work.
60 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
61 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
65 EnableARMLongCalls("arm-long-calls", cl::Hidden,
66 cl::desc("Generate calls via indirect call instructions"),
70 ARMInterworking("arm-interworking", cl::Hidden,
71 cl::desc("Enable / disable ARM interworking (for debugging only)"),
75 class ARMCCState : public CCState {
77 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
78 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
79 LLVMContext &C, ParmContext PC)
80 : CCState(CC, isVarArg, MF, TM, locs, C) {
81 assert(((PC == Call) || (PC == Prologue)) &&
82 "ARMCCState users must specify whether their context is call"
83 "or prologue generation.");
89 // The APCS parameter registers.
90 static const unsigned GPRArgRegs[] = {
91 ARM::R0, ARM::R1, ARM::R2, ARM::R3
94 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
95 EVT PromotedBitwiseVT) {
96 if (VT != PromotedLdStVT) {
97 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
98 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
99 PromotedLdStVT.getSimpleVT());
101 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
102 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
103 PromotedLdStVT.getSimpleVT());
106 EVT ElemTy = VT.getVectorElementType();
107 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
108 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
109 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
111 if (ElemTy == MVT::i32) {
112 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
113 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
114 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
115 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
117 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
122 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
123 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
124 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
125 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
126 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
129 if (VT.isInteger()) {
130 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
131 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
132 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
135 // Promote all bit-wise operations.
136 if (VT.isInteger() && VT != PromotedBitwiseVT) {
137 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
138 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
139 PromotedBitwiseVT.getSimpleVT());
140 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
141 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
142 PromotedBitwiseVT.getSimpleVT());
143 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
144 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
145 PromotedBitwiseVT.getSimpleVT());
148 // Neon does not support vector divide/remainder operations.
149 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
150 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
154 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
157 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
158 addRegisterClass(VT, ARM::DPRRegisterClass);
159 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
162 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
163 addRegisterClass(VT, ARM::QPRRegisterClass);
164 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
167 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
168 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
169 return new TargetLoweringObjectFileMachO();
171 return new ARMElfTargetObjectFile();
174 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
175 : TargetLowering(TM, createTLOF(TM)) {
176 Subtarget = &TM.getSubtarget<ARMSubtarget>();
177 RegInfo = TM.getRegisterInfo();
178 Itins = TM.getInstrItineraryData();
180 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
182 if (Subtarget->isTargetDarwin()) {
183 // Uses VFP for Thumb libfuncs if available.
184 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
185 // Single-precision floating-point arithmetic.
186 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
187 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
188 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
189 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
191 // Double-precision floating-point arithmetic.
192 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
193 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
194 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
195 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
197 // Single-precision comparisons.
198 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
199 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
200 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
201 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
202 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
203 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
204 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
205 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
207 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
216 // Double-precision comparisons.
217 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
218 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
219 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
220 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
221 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
222 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
223 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
224 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
226 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
233 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
235 // Floating-point to integer conversions.
236 // i64 conversions are done via library routines even when generating VFP
237 // instructions, so use the same ones.
238 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
239 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
240 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
241 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
243 // Conversions between floating types.
244 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
245 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
247 // Integer to floating-point conversions.
248 // i64 conversions are done via library routines even when generating VFP
249 // instructions, so use the same ones.
250 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
251 // e.g., __floatunsidf vs. __floatunssidfvfp.
252 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
253 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
254 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
255 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
259 // These libcalls are not available in 32-bit.
260 setLibcallName(RTLIB::SHL_I128, 0);
261 setLibcallName(RTLIB::SRL_I128, 0);
262 setLibcallName(RTLIB::SRA_I128, 0);
264 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
265 // Double-precision floating-point arithmetic helper functions
266 // RTABI chapter 4.1.2, Table 2
267 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
268 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
269 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
270 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
271 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
276 // Double-precision floating-point comparison helper functions
277 // RTABI chapter 4.1.2, Table 3
278 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
279 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
280 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
281 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
282 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
283 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
284 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
285 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
286 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
287 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
288 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
289 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
290 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
291 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
292 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
293 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
294 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
301 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
303 // Single-precision floating-point arithmetic helper functions
304 // RTABI chapter 4.1.2, Table 4
305 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
306 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
307 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
308 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
309 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
314 // Single-precision floating-point comparison helper functions
315 // RTABI chapter 4.1.2, Table 5
316 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
317 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
318 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
319 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
320 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
321 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
322 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
323 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
324 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
325 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
326 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
327 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
328 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
329 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
330 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
331 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
332 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
339 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
341 // Floating-point to integer conversions.
342 // RTABI chapter 4.1.2, Table 6
343 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
344 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
345 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
346 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
347 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
348 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
349 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
350 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
351 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
360 // Conversions between floating types.
361 // RTABI chapter 4.1.2, Table 7
362 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
363 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
364 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
365 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
367 // Integer to floating-point conversions.
368 // RTABI chapter 4.1.2, Table 8
369 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
370 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
371 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
372 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
373 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
374 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
375 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
376 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
377 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
386 // Long long helper functions
387 // RTABI chapter 4.2, Table 9
388 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
389 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
390 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
391 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
392 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
399 // Integer division functions
400 // RTABI chapter 4.3.1
401 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
402 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
403 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
404 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
405 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
407 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
408 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
409 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
415 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
416 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
419 // RTABI chapter 4.3.4
420 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
421 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
422 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
423 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
424 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
425 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
428 // Use divmod compiler-rt calls for iOS 5.0 and later.
429 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
430 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
431 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
432 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
435 if (Subtarget->isThumb1Only())
436 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
438 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
439 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
440 !Subtarget->isThumb1Only()) {
441 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
442 if (!Subtarget->isFPOnlySP())
443 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
445 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
448 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
449 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
450 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
451 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
452 setTruncStoreAction((MVT::SimpleValueType)VT,
453 (MVT::SimpleValueType)InnerVT, Expand);
454 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
455 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
456 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
459 if (Subtarget->hasNEON()) {
460 addDRTypeForNEON(MVT::v2f32);
461 addDRTypeForNEON(MVT::v8i8);
462 addDRTypeForNEON(MVT::v4i16);
463 addDRTypeForNEON(MVT::v2i32);
464 addDRTypeForNEON(MVT::v1i64);
466 addQRTypeForNEON(MVT::v4f32);
467 addQRTypeForNEON(MVT::v2f64);
468 addQRTypeForNEON(MVT::v16i8);
469 addQRTypeForNEON(MVT::v8i16);
470 addQRTypeForNEON(MVT::v4i32);
471 addQRTypeForNEON(MVT::v2i64);
473 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
474 // neither Neon nor VFP support any arithmetic operations on it.
475 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
476 // supported for v4f32.
477 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
478 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
479 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
480 // FIXME: Code duplication: FDIV and FREM are expanded always, see
481 // ARMTargetLowering::addTypeForNEON method for details.
482 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
483 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
484 // FIXME: Create unittest.
485 // In another words, find a way when "copysign" appears in DAG with vector
487 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
488 // FIXME: Code duplication: SETCC has custom operation action, see
489 // ARMTargetLowering::addTypeForNEON method for details.
490 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
491 // FIXME: Create unittest for FNEG and for FABS.
492 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
493 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
495 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
496 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
497 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
498 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
499 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
500 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
501 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
502 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
503 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
504 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
505 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
506 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
507 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
508 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
509 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
511 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
512 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
513 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
514 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
515 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
518 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
519 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
520 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
522 // Neon does not support some operations on v1i64 and v2i64 types.
523 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
524 // Custom handling for some quad-vector types to detect VMULL.
525 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
526 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
527 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
528 // Custom handling for some vector types to avoid expensive expansions
529 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
532 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
533 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
534 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
535 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
536 // a destination type that is wider than the source, and nor does
537 // it have a FP_TO_[SU]INT instruction with a narrower destination than
539 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
540 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
541 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
542 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
544 setTargetDAGCombine(ISD::INTRINSIC_VOID);
545 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
546 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
547 setTargetDAGCombine(ISD::SHL);
548 setTargetDAGCombine(ISD::SRL);
549 setTargetDAGCombine(ISD::SRA);
550 setTargetDAGCombine(ISD::SIGN_EXTEND);
551 setTargetDAGCombine(ISD::ZERO_EXTEND);
552 setTargetDAGCombine(ISD::ANY_EXTEND);
553 setTargetDAGCombine(ISD::SELECT_CC);
554 setTargetDAGCombine(ISD::BUILD_VECTOR);
555 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
556 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
557 setTargetDAGCombine(ISD::STORE);
558 setTargetDAGCombine(ISD::FP_TO_SINT);
559 setTargetDAGCombine(ISD::FP_TO_UINT);
560 setTargetDAGCombine(ISD::FDIV);
562 // It is legal to extload from v4i8 to v4i16 or v4i32.
563 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
564 MVT::v4i16, MVT::v2i16,
566 for (unsigned i = 0; i < 6; ++i) {
567 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
568 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
569 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
573 computeRegisterProperties();
575 // ARM does not have f32 extending load.
576 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
578 // ARM does not have i1 sign extending load.
579 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
581 // ARM supports all 4 flavors of integer indexed load / store.
582 if (!Subtarget->isThumb1Only()) {
583 for (unsigned im = (unsigned)ISD::PRE_INC;
584 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
585 setIndexedLoadAction(im, MVT::i1, Legal);
586 setIndexedLoadAction(im, MVT::i8, Legal);
587 setIndexedLoadAction(im, MVT::i16, Legal);
588 setIndexedLoadAction(im, MVT::i32, Legal);
589 setIndexedStoreAction(im, MVT::i1, Legal);
590 setIndexedStoreAction(im, MVT::i8, Legal);
591 setIndexedStoreAction(im, MVT::i16, Legal);
592 setIndexedStoreAction(im, MVT::i32, Legal);
596 // i64 operation support.
597 setOperationAction(ISD::MUL, MVT::i64, Expand);
598 setOperationAction(ISD::MULHU, MVT::i32, Expand);
599 if (Subtarget->isThumb1Only()) {
600 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
601 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
603 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
604 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
605 setOperationAction(ISD::MULHS, MVT::i32, Expand);
607 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
608 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
609 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
610 setOperationAction(ISD::SRL, MVT::i64, Custom);
611 setOperationAction(ISD::SRA, MVT::i64, Custom);
613 if (!Subtarget->isThumb1Only()) {
614 // FIXME: We should do this for Thumb1 as well.
615 setOperationAction(ISD::ADDC, MVT::i32, Custom);
616 setOperationAction(ISD::ADDE, MVT::i32, Custom);
617 setOperationAction(ISD::SUBC, MVT::i32, Custom);
618 setOperationAction(ISD::SUBE, MVT::i32, Custom);
621 // ARM does not have ROTL.
622 setOperationAction(ISD::ROTL, MVT::i32, Expand);
623 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
624 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
625 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
626 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
628 // These just redirect to CTTZ and CTLZ on ARM.
629 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
630 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
632 // Only ARMv6 has BSWAP.
633 if (!Subtarget->hasV6Ops())
634 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
636 // These are expanded into libcalls.
637 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
638 // v7M has a hardware divider
639 setOperationAction(ISD::SDIV, MVT::i32, Expand);
640 setOperationAction(ISD::UDIV, MVT::i32, Expand);
642 setOperationAction(ISD::SREM, MVT::i32, Expand);
643 setOperationAction(ISD::UREM, MVT::i32, Expand);
644 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
645 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
647 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
648 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
649 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
650 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
651 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
653 setOperationAction(ISD::TRAP, MVT::Other, Legal);
655 // Use the default implementation.
656 setOperationAction(ISD::VASTART, MVT::Other, Custom);
657 setOperationAction(ISD::VAARG, MVT::Other, Expand);
658 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
659 setOperationAction(ISD::VAEND, MVT::Other, Expand);
660 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
661 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
663 if (!Subtarget->isTargetDarwin()) {
664 // Non-Darwin platforms may return values in these registers via the
665 // personality function.
666 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
667 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
668 setExceptionPointerRegister(ARM::R0);
669 setExceptionSelectorRegister(ARM::R1);
672 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
673 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
674 // the default expansion.
675 // FIXME: This should be checking for v6k, not just v6.
676 if (Subtarget->hasDataBarrier() ||
677 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
678 // membarrier needs custom lowering; the rest are legal and handled
680 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
681 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
682 // Custom lowering for 64-bit ops
683 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
684 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
685 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
686 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
690 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
691 setInsertFencesForAtomic(true);
693 // Set them all for expansion, which will force libcalls.
694 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
695 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
696 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
697 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
698 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
699 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
708 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
709 // Unordered/Monotonic case.
710 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
711 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
712 // Since the libcalls include locking, fold in the fences
713 setShouldFoldAtomicFences(true);
716 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
718 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
719 if (!Subtarget->hasV6Ops()) {
720 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
721 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
725 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
726 !Subtarget->isThumb1Only()) {
727 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
728 // iff target supports vfp2.
729 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
730 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
733 // We want to custom lower some of our intrinsics.
734 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
735 if (Subtarget->isTargetDarwin()) {
736 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
737 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
738 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
741 setOperationAction(ISD::SETCC, MVT::i32, Expand);
742 setOperationAction(ISD::SETCC, MVT::f32, Expand);
743 setOperationAction(ISD::SETCC, MVT::f64, Expand);
744 setOperationAction(ISD::SELECT, MVT::i32, Custom);
745 setOperationAction(ISD::SELECT, MVT::f32, Custom);
746 setOperationAction(ISD::SELECT, MVT::f64, Custom);
747 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
748 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
749 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
751 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
752 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
753 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
754 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
755 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
757 // We don't support sin/cos/fmod/copysign/pow
758 setOperationAction(ISD::FSIN, MVT::f64, Expand);
759 setOperationAction(ISD::FSIN, MVT::f32, Expand);
760 setOperationAction(ISD::FCOS, MVT::f32, Expand);
761 setOperationAction(ISD::FCOS, MVT::f64, Expand);
762 setOperationAction(ISD::FREM, MVT::f64, Expand);
763 setOperationAction(ISD::FREM, MVT::f32, Expand);
764 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
765 !Subtarget->isThumb1Only()) {
766 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
767 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
769 setOperationAction(ISD::FPOW, MVT::f64, Expand);
770 setOperationAction(ISD::FPOW, MVT::f32, Expand);
772 setOperationAction(ISD::FMA, MVT::f64, Expand);
773 setOperationAction(ISD::FMA, MVT::f32, Expand);
775 // Various VFP goodness
776 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
777 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
778 if (Subtarget->hasVFP2()) {
779 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
780 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
781 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
782 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
784 // Special handling for half-precision FP.
785 if (!Subtarget->hasFP16()) {
786 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
787 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
791 // We have target-specific dag combine patterns for the following nodes:
792 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
793 setTargetDAGCombine(ISD::ADD);
794 setTargetDAGCombine(ISD::SUB);
795 setTargetDAGCombine(ISD::MUL);
797 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
798 setTargetDAGCombine(ISD::AND);
799 setTargetDAGCombine(ISD::OR);
800 setTargetDAGCombine(ISD::XOR);
803 if (Subtarget->hasV6Ops())
804 setTargetDAGCombine(ISD::SRL);
806 setStackPointerRegisterToSaveRestore(ARM::SP);
808 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
809 !Subtarget->hasVFP2())
810 setSchedulingPreference(Sched::RegPressure);
812 setSchedulingPreference(Sched::Hybrid);
814 //// temporary - rewrite interface to use type
815 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
816 maxStoresPerMemset = 16;
817 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
819 // On ARM arguments smaller than 4 bytes are extended, so all arguments
820 // are at least 4 bytes aligned.
821 setMinStackArgumentAlignment(4);
823 benefitFromCodePlacementOpt = true;
825 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
828 // FIXME: It might make sense to define the representative register class as the
829 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
830 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
831 // SPR's representative would be DPR_VFP2. This should work well if register
832 // pressure tracking were modified such that a register use would increment the
833 // pressure of the register class's representative and all of it's super
834 // classes' representatives transitively. We have not implemented this because
835 // of the difficulty prior to coalescing of modeling operand register classes
836 // due to the common occurrence of cross class copies and subregister insertions
838 std::pair<const TargetRegisterClass*, uint8_t>
839 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
840 const TargetRegisterClass *RRC = 0;
842 switch (VT.getSimpleVT().SimpleTy) {
844 return TargetLowering::findRepresentativeClass(VT);
845 // Use DPR as representative register class for all floating point
846 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
847 // the cost is 1 for both f32 and f64.
848 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
849 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
850 RRC = ARM::DPRRegisterClass;
851 // When NEON is used for SP, only half of the register file is available
852 // because operations that define both SP and DP results will be constrained
853 // to the VFP2 class (D0-D15). We currently model this constraint prior to
854 // coalescing by double-counting the SP regs. See the FIXME above.
855 if (Subtarget->useNEONForSinglePrecisionFP())
858 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
859 case MVT::v4f32: case MVT::v2f64:
860 RRC = ARM::DPRRegisterClass;
864 RRC = ARM::DPRRegisterClass;
868 RRC = ARM::DPRRegisterClass;
872 return std::make_pair(RRC, Cost);
875 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
878 case ARMISD::Wrapper: return "ARMISD::Wrapper";
879 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
880 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
881 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
882 case ARMISD::CALL: return "ARMISD::CALL";
883 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
884 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
885 case ARMISD::tCALL: return "ARMISD::tCALL";
886 case ARMISD::BRCOND: return "ARMISD::BRCOND";
887 case ARMISD::BR_JT: return "ARMISD::BR_JT";
888 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
889 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
890 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
891 case ARMISD::CMP: return "ARMISD::CMP";
892 case ARMISD::CMPZ: return "ARMISD::CMPZ";
893 case ARMISD::CMPFP: return "ARMISD::CMPFP";
894 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
895 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
896 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
898 case ARMISD::CMOV: return "ARMISD::CMOV";
899 case ARMISD::CAND: return "ARMISD::CAND";
900 case ARMISD::COR: return "ARMISD::COR";
901 case ARMISD::CXOR: return "ARMISD::CXOR";
903 case ARMISD::RBIT: return "ARMISD::RBIT";
905 case ARMISD::FTOSI: return "ARMISD::FTOSI";
906 case ARMISD::FTOUI: return "ARMISD::FTOUI";
907 case ARMISD::SITOF: return "ARMISD::SITOF";
908 case ARMISD::UITOF: return "ARMISD::UITOF";
910 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
911 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
912 case ARMISD::RRX: return "ARMISD::RRX";
914 case ARMISD::ADDC: return "ARMISD::ADDC";
915 case ARMISD::ADDE: return "ARMISD::ADDE";
916 case ARMISD::SUBC: return "ARMISD::SUBC";
917 case ARMISD::SUBE: return "ARMISD::SUBE";
919 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
920 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
922 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
923 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
925 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
927 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
929 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
931 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
932 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
934 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
936 case ARMISD::VCEQ: return "ARMISD::VCEQ";
937 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
938 case ARMISD::VCGE: return "ARMISD::VCGE";
939 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
940 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
941 case ARMISD::VCGEU: return "ARMISD::VCGEU";
942 case ARMISD::VCGT: return "ARMISD::VCGT";
943 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
944 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
945 case ARMISD::VCGTU: return "ARMISD::VCGTU";
946 case ARMISD::VTST: return "ARMISD::VTST";
948 case ARMISD::VSHL: return "ARMISD::VSHL";
949 case ARMISD::VSHRs: return "ARMISD::VSHRs";
950 case ARMISD::VSHRu: return "ARMISD::VSHRu";
951 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
952 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
953 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
954 case ARMISD::VSHRN: return "ARMISD::VSHRN";
955 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
956 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
957 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
958 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
959 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
960 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
961 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
962 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
963 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
964 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
965 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
966 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
967 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
968 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
969 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
970 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
971 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
972 case ARMISD::VDUP: return "ARMISD::VDUP";
973 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
974 case ARMISD::VEXT: return "ARMISD::VEXT";
975 case ARMISD::VREV64: return "ARMISD::VREV64";
976 case ARMISD::VREV32: return "ARMISD::VREV32";
977 case ARMISD::VREV16: return "ARMISD::VREV16";
978 case ARMISD::VZIP: return "ARMISD::VZIP";
979 case ARMISD::VUZP: return "ARMISD::VUZP";
980 case ARMISD::VTRN: return "ARMISD::VTRN";
981 case ARMISD::VTBL1: return "ARMISD::VTBL1";
982 case ARMISD::VTBL2: return "ARMISD::VTBL2";
983 case ARMISD::VMULLs: return "ARMISD::VMULLs";
984 case ARMISD::VMULLu: return "ARMISD::VMULLu";
985 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
986 case ARMISD::FMAX: return "ARMISD::FMAX";
987 case ARMISD::FMIN: return "ARMISD::FMIN";
988 case ARMISD::BFI: return "ARMISD::BFI";
989 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
990 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
991 case ARMISD::VBSL: return "ARMISD::VBSL";
992 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
993 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
994 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
995 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
996 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
997 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
998 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
999 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1000 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1001 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1002 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1003 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1004 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1005 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1006 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1007 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1008 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1009 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1010 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1011 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1015 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1016 if (!VT.isVector()) return getPointerTy();
1017 return VT.changeVectorElementTypeToInteger();
1020 /// getRegClassFor - Return the register class that should be used for the
1021 /// specified value type.
1022 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1023 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1024 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1025 // load / store 4 to 8 consecutive D registers.
1026 if (Subtarget->hasNEON()) {
1027 if (VT == MVT::v4i64)
1028 return ARM::QQPRRegisterClass;
1029 else if (VT == MVT::v8i64)
1030 return ARM::QQQQPRRegisterClass;
1032 return TargetLowering::getRegClassFor(VT);
1035 // Create a fast isel object.
1037 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1038 return ARM::createFastISel(funcInfo);
1041 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1042 /// be used for loads / stores from the global.
1043 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1044 return (Subtarget->isThumb1Only() ? 127 : 4095);
1047 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1048 unsigned NumVals = N->getNumValues();
1050 return Sched::RegPressure;
1052 for (unsigned i = 0; i != NumVals; ++i) {
1053 EVT VT = N->getValueType(i);
1054 if (VT == MVT::Glue || VT == MVT::Other)
1056 if (VT.isFloatingPoint() || VT.isVector())
1060 if (!N->isMachineOpcode())
1061 return Sched::RegPressure;
1063 // Load are scheduled for latency even if there instruction itinerary
1064 // is not available.
1065 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1066 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1068 if (MCID.getNumDefs() == 0)
1069 return Sched::RegPressure;
1070 if (!Itins->isEmpty() &&
1071 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1074 return Sched::RegPressure;
1077 //===----------------------------------------------------------------------===//
1079 //===----------------------------------------------------------------------===//
1081 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1082 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1084 default: llvm_unreachable("Unknown condition code!");
1085 case ISD::SETNE: return ARMCC::NE;
1086 case ISD::SETEQ: return ARMCC::EQ;
1087 case ISD::SETGT: return ARMCC::GT;
1088 case ISD::SETGE: return ARMCC::GE;
1089 case ISD::SETLT: return ARMCC::LT;
1090 case ISD::SETLE: return ARMCC::LE;
1091 case ISD::SETUGT: return ARMCC::HI;
1092 case ISD::SETUGE: return ARMCC::HS;
1093 case ISD::SETULT: return ARMCC::LO;
1094 case ISD::SETULE: return ARMCC::LS;
1098 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1099 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1100 ARMCC::CondCodes &CondCode2) {
1101 CondCode2 = ARMCC::AL;
1103 default: llvm_unreachable("Unknown FP condition!");
1105 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1107 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1109 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1110 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1111 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1112 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1113 case ISD::SETO: CondCode = ARMCC::VC; break;
1114 case ISD::SETUO: CondCode = ARMCC::VS; break;
1115 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1116 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1117 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1119 case ISD::SETULT: CondCode = ARMCC::LT; break;
1121 case ISD::SETULE: CondCode = ARMCC::LE; break;
1123 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1127 //===----------------------------------------------------------------------===//
1128 // Calling Convention Implementation
1129 //===----------------------------------------------------------------------===//
1131 #include "ARMGenCallingConv.inc"
1133 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1134 /// given CallingConvention value.
1135 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1137 bool isVarArg) const {
1140 llvm_unreachable("Unsupported calling convention");
1141 case CallingConv::Fast:
1142 if (Subtarget->hasVFP2() && !isVarArg) {
1143 if (!Subtarget->isAAPCS_ABI())
1144 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1145 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1146 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1149 case CallingConv::C: {
1150 // Use target triple & subtarget features to do actual dispatch.
1151 if (!Subtarget->isAAPCS_ABI())
1152 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1153 else if (Subtarget->hasVFP2() &&
1154 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1156 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1157 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1159 case CallingConv::ARM_AAPCS_VFP:
1161 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 case CallingConv::ARM_AAPCS:
1164 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_APCS:
1166 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1170 /// LowerCallResult - Lower the result values of a call into the
1171 /// appropriate copies out of appropriate physical registers.
1173 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1174 CallingConv::ID CallConv, bool isVarArg,
1175 const SmallVectorImpl<ISD::InputArg> &Ins,
1176 DebugLoc dl, SelectionDAG &DAG,
1177 SmallVectorImpl<SDValue> &InVals) const {
1179 // Assign locations to each value returned by this call.
1180 SmallVector<CCValAssign, 16> RVLocs;
1181 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1182 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1183 CCInfo.AnalyzeCallResult(Ins,
1184 CCAssignFnForNode(CallConv, /* Return*/ true,
1187 // Copy all of the result registers out of their specified physreg.
1188 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1189 CCValAssign VA = RVLocs[i];
1192 if (VA.needsCustom()) {
1193 // Handle f64 or half of a v2f64.
1194 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1196 Chain = Lo.getValue(1);
1197 InFlag = Lo.getValue(2);
1198 VA = RVLocs[++i]; // skip ahead to next loc
1199 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1201 Chain = Hi.getValue(1);
1202 InFlag = Hi.getValue(2);
1203 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1205 if (VA.getLocVT() == MVT::v2f64) {
1206 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1207 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1208 DAG.getConstant(0, MVT::i32));
1210 VA = RVLocs[++i]; // skip ahead to next loc
1211 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1212 Chain = Lo.getValue(1);
1213 InFlag = Lo.getValue(2);
1214 VA = RVLocs[++i]; // skip ahead to next loc
1215 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1216 Chain = Hi.getValue(1);
1217 InFlag = Hi.getValue(2);
1218 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1219 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1220 DAG.getConstant(1, MVT::i32));
1223 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1225 Chain = Val.getValue(1);
1226 InFlag = Val.getValue(2);
1229 switch (VA.getLocInfo()) {
1230 default: llvm_unreachable("Unknown loc info!");
1231 case CCValAssign::Full: break;
1232 case CCValAssign::BCvt:
1233 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1237 InVals.push_back(Val);
1243 /// LowerMemOpCallTo - Store the argument to the stack.
1245 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1246 SDValue StackPtr, SDValue Arg,
1247 DebugLoc dl, SelectionDAG &DAG,
1248 const CCValAssign &VA,
1249 ISD::ArgFlagsTy Flags) const {
1250 unsigned LocMemOffset = VA.getLocMemOffset();
1251 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1252 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1253 return DAG.getStore(Chain, dl, Arg, PtrOff,
1254 MachinePointerInfo::getStack(LocMemOffset),
1258 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1259 SDValue Chain, SDValue &Arg,
1260 RegsToPassVector &RegsToPass,
1261 CCValAssign &VA, CCValAssign &NextVA,
1263 SmallVector<SDValue, 8> &MemOpChains,
1264 ISD::ArgFlagsTy Flags) const {
1266 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1267 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1268 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1270 if (NextVA.isRegLoc())
1271 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1273 assert(NextVA.isMemLoc());
1274 if (StackPtr.getNode() == 0)
1275 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1277 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1283 /// LowerCall - Lowering a call into a callseq_start <-
1284 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1287 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1288 CallingConv::ID CallConv, bool isVarArg,
1290 const SmallVectorImpl<ISD::OutputArg> &Outs,
1291 const SmallVectorImpl<SDValue> &OutVals,
1292 const SmallVectorImpl<ISD::InputArg> &Ins,
1293 DebugLoc dl, SelectionDAG &DAG,
1294 SmallVectorImpl<SDValue> &InVals) const {
1295 MachineFunction &MF = DAG.getMachineFunction();
1296 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1297 bool IsSibCall = false;
1298 // Disable tail calls if they're not supported.
1299 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1302 // Check if it's really possible to do a tail call.
1303 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1304 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1305 Outs, OutVals, Ins, DAG);
1306 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1307 // detected sibcalls.
1314 // Analyze operands of the call, assigning locations to each operand.
1315 SmallVector<CCValAssign, 16> ArgLocs;
1316 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1317 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1318 CCInfo.AnalyzeCallOperands(Outs,
1319 CCAssignFnForNode(CallConv, /* Return*/ false,
1322 // Get a count of how many bytes are to be pushed on the stack.
1323 unsigned NumBytes = CCInfo.getNextStackOffset();
1325 // For tail calls, memory operands are available in our caller's stack.
1329 // Adjust the stack pointer for the new arguments...
1330 // These operations are automatically eliminated by the prolog/epilog pass
1332 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1334 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1336 RegsToPassVector RegsToPass;
1337 SmallVector<SDValue, 8> MemOpChains;
1339 // Walk the register/memloc assignments, inserting copies/loads. In the case
1340 // of tail call optimization, arguments are handled later.
1341 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1343 ++i, ++realArgIdx) {
1344 CCValAssign &VA = ArgLocs[i];
1345 SDValue Arg = OutVals[realArgIdx];
1346 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1347 bool isByVal = Flags.isByVal();
1349 // Promote the value if needed.
1350 switch (VA.getLocInfo()) {
1351 default: llvm_unreachable("Unknown loc info!");
1352 case CCValAssign::Full: break;
1353 case CCValAssign::SExt:
1354 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1356 case CCValAssign::ZExt:
1357 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1359 case CCValAssign::AExt:
1360 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1362 case CCValAssign::BCvt:
1363 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1367 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1368 if (VA.needsCustom()) {
1369 if (VA.getLocVT() == MVT::v2f64) {
1370 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1371 DAG.getConstant(0, MVT::i32));
1372 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1373 DAG.getConstant(1, MVT::i32));
1375 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1376 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1378 VA = ArgLocs[++i]; // skip ahead to next loc
1379 if (VA.isRegLoc()) {
1380 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1381 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1383 assert(VA.isMemLoc());
1385 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1386 dl, DAG, VA, Flags));
1389 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1390 StackPtr, MemOpChains, Flags);
1392 } else if (VA.isRegLoc()) {
1393 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1394 } else if (isByVal) {
1395 assert(VA.isMemLoc());
1396 unsigned offset = 0;
1398 // True if this byval aggregate will be split between registers
1400 if (CCInfo.isFirstByValRegValid()) {
1401 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1403 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1404 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1405 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1406 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1407 MachinePointerInfo(),
1408 false, false, false, 0);
1409 MemOpChains.push_back(Load.getValue(1));
1410 RegsToPass.push_back(std::make_pair(j, Load));
1412 offset = ARM::R4 - CCInfo.getFirstByValReg();
1413 CCInfo.clearFirstByValReg();
1416 unsigned LocMemOffset = VA.getLocMemOffset();
1417 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1420 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1421 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1422 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1424 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1425 Flags.getByValAlign(),
1426 /*isVolatile=*/false,
1427 /*AlwaysInline=*/false,
1428 MachinePointerInfo(0),
1429 MachinePointerInfo(0)));
1431 } else if (!IsSibCall) {
1432 assert(VA.isMemLoc());
1434 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1435 dl, DAG, VA, Flags));
1439 if (!MemOpChains.empty())
1440 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1441 &MemOpChains[0], MemOpChains.size());
1443 // Build a sequence of copy-to-reg nodes chained together with token chain
1444 // and flag operands which copy the outgoing args into the appropriate regs.
1446 // Tail call byval lowering might overwrite argument registers so in case of
1447 // tail call optimization the copies to registers are lowered later.
1449 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1450 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1451 RegsToPass[i].second, InFlag);
1452 InFlag = Chain.getValue(1);
1455 // For tail calls lower the arguments to the 'real' stack slot.
1457 // Force all the incoming stack arguments to be loaded from the stack
1458 // before any new outgoing arguments are stored to the stack, because the
1459 // outgoing stack slots may alias the incoming argument stack slots, and
1460 // the alias isn't otherwise explicit. This is slightly more conservative
1461 // than necessary, because it means that each store effectively depends
1462 // on every argument instead of just those arguments it would clobber.
1464 // Do not flag preceding copytoreg stuff together with the following stuff.
1466 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1467 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1468 RegsToPass[i].second, InFlag);
1469 InFlag = Chain.getValue(1);
1474 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1475 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1476 // node so that legalize doesn't hack it.
1477 bool isDirect = false;
1478 bool isARMFunc = false;
1479 bool isLocalARMFunc = false;
1480 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1482 if (EnableARMLongCalls) {
1483 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1484 && "long-calls with non-static relocation model!");
1485 // Handle a global address or an external symbol. If it's not one of
1486 // those, the target's already in a register, so we don't need to do
1488 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1489 const GlobalValue *GV = G->getGlobal();
1490 // Create a constant pool entry for the callee address
1491 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1492 ARMConstantPoolValue *CPV =
1493 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1495 // Get the address of the callee into a register
1496 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1497 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1498 Callee = DAG.getLoad(getPointerTy(), dl,
1499 DAG.getEntryNode(), CPAddr,
1500 MachinePointerInfo::getConstantPool(),
1501 false, false, false, 0);
1502 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1503 const char *Sym = S->getSymbol();
1505 // Create a constant pool entry for the callee address
1506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507 ARMConstantPoolValue *CPV =
1508 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1509 ARMPCLabelIndex, 0);
1510 // Get the address of the callee into a register
1511 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513 Callee = DAG.getLoad(getPointerTy(), dl,
1514 DAG.getEntryNode(), CPAddr,
1515 MachinePointerInfo::getConstantPool(),
1516 false, false, false, 0);
1518 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1519 const GlobalValue *GV = G->getGlobal();
1521 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1522 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1523 getTargetMachine().getRelocationModel() != Reloc::Static;
1524 isARMFunc = !Subtarget->isThumb() || isStub;
1525 // ARM call to a local ARM function is predicable.
1526 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1527 // tBX takes a register source operand.
1528 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1529 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1530 ARMConstantPoolValue *CPV =
1531 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1532 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1533 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1534 Callee = DAG.getLoad(getPointerTy(), dl,
1535 DAG.getEntryNode(), CPAddr,
1536 MachinePointerInfo::getConstantPool(),
1537 false, false, false, 0);
1538 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1539 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1540 getPointerTy(), Callee, PICLabel);
1542 // On ELF targets for PIC code, direct calls should go through the PLT
1543 unsigned OpFlags = 0;
1544 if (Subtarget->isTargetELF() &&
1545 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1546 OpFlags = ARMII::MO_PLT;
1547 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1549 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1551 bool isStub = Subtarget->isTargetDarwin() &&
1552 getTargetMachine().getRelocationModel() != Reloc::Static;
1553 isARMFunc = !Subtarget->isThumb() || isStub;
1554 // tBX takes a register source operand.
1555 const char *Sym = S->getSymbol();
1556 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1557 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1558 ARMConstantPoolValue *CPV =
1559 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1560 ARMPCLabelIndex, 4);
1561 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1562 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1563 Callee = DAG.getLoad(getPointerTy(), dl,
1564 DAG.getEntryNode(), CPAddr,
1565 MachinePointerInfo::getConstantPool(),
1566 false, false, false, 0);
1567 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1568 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1569 getPointerTy(), Callee, PICLabel);
1571 unsigned OpFlags = 0;
1572 // On ELF targets for PIC code, direct calls should go through the PLT
1573 if (Subtarget->isTargetELF() &&
1574 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1575 OpFlags = ARMII::MO_PLT;
1576 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1580 // FIXME: handle tail calls differently.
1582 if (Subtarget->isThumb()) {
1583 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1584 CallOpc = ARMISD::CALL_NOLINK;
1586 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1588 CallOpc = (isDirect || Subtarget->hasV5TOps())
1589 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1590 : ARMISD::CALL_NOLINK;
1593 std::vector<SDValue> Ops;
1594 Ops.push_back(Chain);
1595 Ops.push_back(Callee);
1597 // Add argument registers to the end of the list so that they are known live
1599 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1600 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1601 RegsToPass[i].second.getValueType()));
1603 // Add a register mask operand representing the call-preserved registers.
1604 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1605 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1606 assert(Mask && "Missing call preserved mask for calling convention");
1607 Ops.push_back(DAG.getRegisterMask(Mask));
1609 if (InFlag.getNode())
1610 Ops.push_back(InFlag);
1612 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1614 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1616 // Returns a chain and a flag for retval copy to use.
1617 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1618 InFlag = Chain.getValue(1);
1620 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1621 DAG.getIntPtrConstant(0, true), InFlag);
1623 InFlag = Chain.getValue(1);
1625 // Handle result values, copying them out of physregs into vregs that we
1627 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1631 /// HandleByVal - Every parameter *after* a byval parameter is passed
1632 /// on the stack. Remember the next parameter register to allocate,
1633 /// and then confiscate the rest of the parameter registers to insure
1636 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1637 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1638 assert((State->getCallOrPrologue() == Prologue ||
1639 State->getCallOrPrologue() == Call) &&
1640 "unhandled ParmContext");
1641 if ((!State->isFirstByValRegValid()) &&
1642 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1643 State->setFirstByValReg(reg);
1644 // At a call site, a byval parameter that is split between
1645 // registers and memory needs its size truncated here. In a
1646 // function prologue, such byval parameters are reassembled in
1647 // memory, and are not truncated.
1648 if (State->getCallOrPrologue() == Call) {
1649 unsigned excess = 4 * (ARM::R4 - reg);
1650 assert(size >= excess && "expected larger existing stack allocation");
1654 // Confiscate any remaining parameter registers to preclude their
1655 // assignment to subsequent parameters.
1656 while (State->AllocateReg(GPRArgRegs, 4))
1660 /// MatchingStackOffset - Return true if the given stack call argument is
1661 /// already available in the same position (relatively) of the caller's
1662 /// incoming argument stack.
1664 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1665 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1666 const ARMInstrInfo *TII) {
1667 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1669 if (Arg.getOpcode() == ISD::CopyFromReg) {
1670 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1671 if (!TargetRegisterInfo::isVirtualRegister(VR))
1673 MachineInstr *Def = MRI->getVRegDef(VR);
1676 if (!Flags.isByVal()) {
1677 if (!TII->isLoadFromStackSlot(Def, FI))
1682 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1683 if (Flags.isByVal())
1684 // ByVal argument is passed in as a pointer but it's now being
1685 // dereferenced. e.g.
1686 // define @foo(%struct.X* %A) {
1687 // tail call @bar(%struct.X* byval %A)
1690 SDValue Ptr = Ld->getBasePtr();
1691 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1694 FI = FINode->getIndex();
1698 assert(FI != INT_MAX);
1699 if (!MFI->isFixedObjectIndex(FI))
1701 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1704 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1705 /// for tail call optimization. Targets which want to do tail call
1706 /// optimization should implement this function.
1708 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1709 CallingConv::ID CalleeCC,
1711 bool isCalleeStructRet,
1712 bool isCallerStructRet,
1713 const SmallVectorImpl<ISD::OutputArg> &Outs,
1714 const SmallVectorImpl<SDValue> &OutVals,
1715 const SmallVectorImpl<ISD::InputArg> &Ins,
1716 SelectionDAG& DAG) const {
1717 const Function *CallerF = DAG.getMachineFunction().getFunction();
1718 CallingConv::ID CallerCC = CallerF->getCallingConv();
1719 bool CCMatch = CallerCC == CalleeCC;
1721 // Look for obvious safe cases to perform tail call optimization that do not
1722 // require ABI changes. This is what gcc calls sibcall.
1724 // Do not sibcall optimize vararg calls unless the call site is not passing
1726 if (isVarArg && !Outs.empty())
1729 // Also avoid sibcall optimization if either caller or callee uses struct
1730 // return semantics.
1731 if (isCalleeStructRet || isCallerStructRet)
1734 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1735 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1736 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1737 // support in the assembler and linker to be used. This would need to be
1738 // fixed to fully support tail calls in Thumb1.
1740 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1741 // LR. This means if we need to reload LR, it takes an extra instructions,
1742 // which outweighs the value of the tail call; but here we don't know yet
1743 // whether LR is going to be used. Probably the right approach is to
1744 // generate the tail call here and turn it back into CALL/RET in
1745 // emitEpilogue if LR is used.
1747 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1748 // but we need to make sure there are enough registers; the only valid
1749 // registers are the 4 used for parameters. We don't currently do this
1751 if (Subtarget->isThumb1Only())
1754 // If the calling conventions do not match, then we'd better make sure the
1755 // results are returned in the same way as what the caller expects.
1757 SmallVector<CCValAssign, 16> RVLocs1;
1758 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1759 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1760 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1762 SmallVector<CCValAssign, 16> RVLocs2;
1763 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1764 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1765 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1767 if (RVLocs1.size() != RVLocs2.size())
1769 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1770 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1772 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1774 if (RVLocs1[i].isRegLoc()) {
1775 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1778 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1784 // If the callee takes no arguments then go on to check the results of the
1786 if (!Outs.empty()) {
1787 // Check if stack adjustment is needed. For now, do not do this if any
1788 // argument is passed on the stack.
1789 SmallVector<CCValAssign, 16> ArgLocs;
1790 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1791 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1792 CCInfo.AnalyzeCallOperands(Outs,
1793 CCAssignFnForNode(CalleeCC, false, isVarArg));
1794 if (CCInfo.getNextStackOffset()) {
1795 MachineFunction &MF = DAG.getMachineFunction();
1797 // Check if the arguments are already laid out in the right way as
1798 // the caller's fixed stack objects.
1799 MachineFrameInfo *MFI = MF.getFrameInfo();
1800 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1801 const ARMInstrInfo *TII =
1802 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1803 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1805 ++i, ++realArgIdx) {
1806 CCValAssign &VA = ArgLocs[i];
1807 EVT RegVT = VA.getLocVT();
1808 SDValue Arg = OutVals[realArgIdx];
1809 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1810 if (VA.getLocInfo() == CCValAssign::Indirect)
1812 if (VA.needsCustom()) {
1813 // f64 and vector types are split into multiple registers or
1814 // register/stack-slot combinations. The types will not match
1815 // the registers; give up on memory f64 refs until we figure
1816 // out what to do about this.
1819 if (!ArgLocs[++i].isRegLoc())
1821 if (RegVT == MVT::v2f64) {
1822 if (!ArgLocs[++i].isRegLoc())
1824 if (!ArgLocs[++i].isRegLoc())
1827 } else if (!VA.isRegLoc()) {
1828 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1840 ARMTargetLowering::LowerReturn(SDValue Chain,
1841 CallingConv::ID CallConv, bool isVarArg,
1842 const SmallVectorImpl<ISD::OutputArg> &Outs,
1843 const SmallVectorImpl<SDValue> &OutVals,
1844 DebugLoc dl, SelectionDAG &DAG) const {
1846 // CCValAssign - represent the assignment of the return value to a location.
1847 SmallVector<CCValAssign, 16> RVLocs;
1849 // CCState - Info about the registers and stack slots.
1850 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1851 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1853 // Analyze outgoing return values.
1854 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1857 // If this is the first return lowered for this function, add
1858 // the regs to the liveout set for the function.
1859 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1860 for (unsigned i = 0; i != RVLocs.size(); ++i)
1861 if (RVLocs[i].isRegLoc())
1862 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1867 // Copy the result values into the output registers.
1868 for (unsigned i = 0, realRVLocIdx = 0;
1870 ++i, ++realRVLocIdx) {
1871 CCValAssign &VA = RVLocs[i];
1872 assert(VA.isRegLoc() && "Can only return in registers!");
1874 SDValue Arg = OutVals[realRVLocIdx];
1876 switch (VA.getLocInfo()) {
1877 default: llvm_unreachable("Unknown loc info!");
1878 case CCValAssign::Full: break;
1879 case CCValAssign::BCvt:
1880 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1884 if (VA.needsCustom()) {
1885 if (VA.getLocVT() == MVT::v2f64) {
1886 // Extract the first half and return it in two registers.
1887 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1888 DAG.getConstant(0, MVT::i32));
1889 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1890 DAG.getVTList(MVT::i32, MVT::i32), Half);
1892 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1893 Flag = Chain.getValue(1);
1894 VA = RVLocs[++i]; // skip ahead to next loc
1895 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1896 HalfGPRs.getValue(1), Flag);
1897 Flag = Chain.getValue(1);
1898 VA = RVLocs[++i]; // skip ahead to next loc
1900 // Extract the 2nd half and fall through to handle it as an f64 value.
1901 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1902 DAG.getConstant(1, MVT::i32));
1904 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1906 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1907 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1908 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1909 Flag = Chain.getValue(1);
1910 VA = RVLocs[++i]; // skip ahead to next loc
1911 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1914 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1916 // Guarantee that all emitted copies are
1917 // stuck together, avoiding something bad.
1918 Flag = Chain.getValue(1);
1923 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1925 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1930 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1931 if (N->getNumValues() != 1)
1933 if (!N->hasNUsesOfValue(1, 0))
1936 unsigned NumCopies = 0;
1937 SDNode* Copies[2] = { 0, 0 };
1938 SDNode *Use = *N->use_begin();
1939 if (Use->getOpcode() == ISD::CopyToReg) {
1940 Copies[NumCopies++] = Use;
1941 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1942 // f64 returned in a pair of GPRs.
1943 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1945 if (UI->getOpcode() != ISD::CopyToReg)
1947 Copies[UI.getUse().getResNo()] = *UI;
1950 } else if (Use->getOpcode() == ISD::BITCAST) {
1951 // f32 returned in a single GPR.
1952 if (!Use->hasNUsesOfValue(1, 0))
1954 Use = *Use->use_begin();
1955 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1957 Copies[NumCopies++] = Use;
1962 if (NumCopies != 1 && NumCopies != 2)
1965 bool HasRet = false;
1966 for (unsigned i = 0; i < NumCopies; ++i) {
1967 SDNode *Copy = Copies[i];
1968 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1970 if (UI->getOpcode() == ISD::CopyToReg) {
1972 if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1976 if (UI->getOpcode() != ARMISD::RET_FLAG)
1985 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1986 if (!EnableARMTailCalls)
1989 if (!CI->isTailCall())
1992 return !Subtarget->isThumb1Only();
1995 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1996 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1997 // one of the above mentioned nodes. It has to be wrapped because otherwise
1998 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1999 // be used to form addressing mode. These wrapped nodes will be selected
2001 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2002 EVT PtrVT = Op.getValueType();
2003 // FIXME there is no actual debug info here
2004 DebugLoc dl = Op.getDebugLoc();
2005 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2007 if (CP->isMachineConstantPoolEntry())
2008 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2009 CP->getAlignment());
2011 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2012 CP->getAlignment());
2013 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2016 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2017 return MachineJumpTableInfo::EK_Inline;
2020 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2021 SelectionDAG &DAG) const {
2022 MachineFunction &MF = DAG.getMachineFunction();
2023 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2024 unsigned ARMPCLabelIndex = 0;
2025 DebugLoc DL = Op.getDebugLoc();
2026 EVT PtrVT = getPointerTy();
2027 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2028 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2030 if (RelocM == Reloc::Static) {
2031 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2033 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2034 ARMPCLabelIndex = AFI->createPICLabelUId();
2035 ARMConstantPoolValue *CPV =
2036 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2037 ARMCP::CPBlockAddress, PCAdj);
2038 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2040 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2041 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2042 MachinePointerInfo::getConstantPool(),
2043 false, false, false, 0);
2044 if (RelocM == Reloc::Static)
2046 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2047 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2050 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2052 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2053 SelectionDAG &DAG) const {
2054 DebugLoc dl = GA->getDebugLoc();
2055 EVT PtrVT = getPointerTy();
2056 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2057 MachineFunction &MF = DAG.getMachineFunction();
2058 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2059 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2060 ARMConstantPoolValue *CPV =
2061 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2062 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2063 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2064 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2065 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2066 MachinePointerInfo::getConstantPool(),
2067 false, false, false, 0);
2068 SDValue Chain = Argument.getValue(1);
2070 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2071 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2073 // call __tls_get_addr.
2076 Entry.Node = Argument;
2077 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2078 Args.push_back(Entry);
2079 // FIXME: is there useful debug info available here?
2080 std::pair<SDValue, SDValue> CallResult =
2081 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2082 false, false, false, false,
2083 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2084 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2085 return CallResult.first;
2088 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2089 // "local exec" model.
2091 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2092 SelectionDAG &DAG) const {
2093 const GlobalValue *GV = GA->getGlobal();
2094 DebugLoc dl = GA->getDebugLoc();
2096 SDValue Chain = DAG.getEntryNode();
2097 EVT PtrVT = getPointerTy();
2098 // Get the Thread Pointer
2099 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2101 if (GV->isDeclaration()) {
2102 MachineFunction &MF = DAG.getMachineFunction();
2103 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2104 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2105 // Initial exec model.
2106 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2107 ARMConstantPoolValue *CPV =
2108 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2109 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2111 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2112 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2113 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2114 MachinePointerInfo::getConstantPool(),
2115 false, false, false, 0);
2116 Chain = Offset.getValue(1);
2118 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2119 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2121 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2122 MachinePointerInfo::getConstantPool(),
2123 false, false, false, 0);
2126 ARMConstantPoolValue *CPV =
2127 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2128 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2129 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2130 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2131 MachinePointerInfo::getConstantPool(),
2132 false, false, false, 0);
2135 // The address of the thread local variable is the add of the thread
2136 // pointer with the offset of the variable.
2137 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2141 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2142 // TODO: implement the "local dynamic" model
2143 assert(Subtarget->isTargetELF() &&
2144 "TLS not implemented for non-ELF targets");
2145 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2146 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2147 // otherwise use the "Local Exec" TLS Model
2148 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2149 return LowerToTLSGeneralDynamicModel(GA, DAG);
2151 return LowerToTLSExecModels(GA, DAG);
2154 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2155 SelectionDAG &DAG) const {
2156 EVT PtrVT = getPointerTy();
2157 DebugLoc dl = Op.getDebugLoc();
2158 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2159 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2160 if (RelocM == Reloc::PIC_) {
2161 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2162 ARMConstantPoolValue *CPV =
2163 ARMConstantPoolConstant::Create(GV,
2164 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2165 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2166 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2167 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2169 MachinePointerInfo::getConstantPool(),
2170 false, false, false, 0);
2171 SDValue Chain = Result.getValue(1);
2172 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2173 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2175 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2176 MachinePointerInfo::getGOT(),
2177 false, false, false, 0);
2181 // If we have T2 ops, we can materialize the address directly via movt/movw
2182 // pair. This is always cheaper.
2183 if (Subtarget->useMovt()) {
2185 // FIXME: Once remat is capable of dealing with instructions with register
2186 // operands, expand this into two nodes.
2187 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2188 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2190 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2191 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2192 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2193 MachinePointerInfo::getConstantPool(),
2194 false, false, false, 0);
2198 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2199 SelectionDAG &DAG) const {
2200 EVT PtrVT = getPointerTy();
2201 DebugLoc dl = Op.getDebugLoc();
2202 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2203 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2204 MachineFunction &MF = DAG.getMachineFunction();
2205 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2207 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2208 // update ARMFastISel::ARMMaterializeGV.
2209 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2211 // FIXME: Once remat is capable of dealing with instructions with register
2212 // operands, expand this into two nodes.
2213 if (RelocM == Reloc::Static)
2214 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2215 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2217 unsigned Wrapper = (RelocM == Reloc::PIC_)
2218 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2219 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2220 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2221 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2222 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2223 MachinePointerInfo::getGOT(),
2224 false, false, false, 0);
2228 unsigned ARMPCLabelIndex = 0;
2230 if (RelocM == Reloc::Static) {
2231 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2233 ARMPCLabelIndex = AFI->createPICLabelUId();
2234 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2235 ARMConstantPoolValue *CPV =
2236 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2238 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2240 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2242 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2243 MachinePointerInfo::getConstantPool(),
2244 false, false, false, 0);
2245 SDValue Chain = Result.getValue(1);
2247 if (RelocM == Reloc::PIC_) {
2248 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2249 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2252 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2253 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2254 false, false, false, 0);
2259 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2260 SelectionDAG &DAG) const {
2261 assert(Subtarget->isTargetELF() &&
2262 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2263 MachineFunction &MF = DAG.getMachineFunction();
2264 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2265 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2266 EVT PtrVT = getPointerTy();
2267 DebugLoc dl = Op.getDebugLoc();
2268 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2269 ARMConstantPoolValue *CPV =
2270 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2271 ARMPCLabelIndex, PCAdj);
2272 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2273 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2274 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2275 MachinePointerInfo::getConstantPool(),
2276 false, false, false, 0);
2277 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2278 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2282 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2283 DebugLoc dl = Op.getDebugLoc();
2284 SDValue Val = DAG.getConstant(0, MVT::i32);
2285 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2286 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2287 Op.getOperand(1), Val);
2291 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2292 DebugLoc dl = Op.getDebugLoc();
2293 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2294 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2298 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2299 const ARMSubtarget *Subtarget) const {
2300 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2301 DebugLoc dl = Op.getDebugLoc();
2303 default: return SDValue(); // Don't custom lower most intrinsics.
2304 case Intrinsic::arm_thread_pointer: {
2305 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2306 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2308 case Intrinsic::eh_sjlj_lsda: {
2309 MachineFunction &MF = DAG.getMachineFunction();
2310 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2311 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2312 EVT PtrVT = getPointerTy();
2313 DebugLoc dl = Op.getDebugLoc();
2314 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2316 unsigned PCAdj = (RelocM != Reloc::PIC_)
2317 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2318 ARMConstantPoolValue *CPV =
2319 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2320 ARMCP::CPLSDA, PCAdj);
2321 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2322 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2324 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2325 MachinePointerInfo::getConstantPool(),
2326 false, false, false, 0);
2328 if (RelocM == Reloc::PIC_) {
2329 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2330 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2334 case Intrinsic::arm_neon_vmulls:
2335 case Intrinsic::arm_neon_vmullu: {
2336 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2337 ? ARMISD::VMULLs : ARMISD::VMULLu;
2338 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2339 Op.getOperand(1), Op.getOperand(2));
2344 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2345 const ARMSubtarget *Subtarget) {
2346 DebugLoc dl = Op.getDebugLoc();
2347 if (!Subtarget->hasDataBarrier()) {
2348 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2349 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2351 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2352 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2353 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2354 DAG.getConstant(0, MVT::i32));
2357 SDValue Op5 = Op.getOperand(5);
2358 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2359 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2360 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2361 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2363 ARM_MB::MemBOpt DMBOpt;
2364 if (isDeviceBarrier)
2365 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2367 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2368 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2369 DAG.getConstant(DMBOpt, MVT::i32));
2373 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2374 const ARMSubtarget *Subtarget) {
2375 // FIXME: handle "fence singlethread" more efficiently.
2376 DebugLoc dl = Op.getDebugLoc();
2377 if (!Subtarget->hasDataBarrier()) {
2378 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2379 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2381 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2382 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2383 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2384 DAG.getConstant(0, MVT::i32));
2387 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2388 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2391 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2392 const ARMSubtarget *Subtarget) {
2393 // ARM pre v5TE and Thumb1 does not have preload instructions.
2394 if (!(Subtarget->isThumb2() ||
2395 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2396 // Just preserve the chain.
2397 return Op.getOperand(0);
2399 DebugLoc dl = Op.getDebugLoc();
2400 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2402 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2403 // ARMv7 with MP extension has PLDW.
2404 return Op.getOperand(0);
2406 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2407 if (Subtarget->isThumb()) {
2409 isRead = ~isRead & 1;
2410 isData = ~isData & 1;
2413 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2414 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2415 DAG.getConstant(isData, MVT::i32));
2418 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2419 MachineFunction &MF = DAG.getMachineFunction();
2420 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2422 // vastart just stores the address of the VarArgsFrameIndex slot into the
2423 // memory location argument.
2424 DebugLoc dl = Op.getDebugLoc();
2425 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2426 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2427 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2428 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2429 MachinePointerInfo(SV), false, false, 0);
2433 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2434 SDValue &Root, SelectionDAG &DAG,
2435 DebugLoc dl) const {
2436 MachineFunction &MF = DAG.getMachineFunction();
2437 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2439 const TargetRegisterClass *RC;
2440 if (AFI->isThumb1OnlyFunction())
2441 RC = ARM::tGPRRegisterClass;
2443 RC = ARM::GPRRegisterClass;
2445 // Transform the arguments stored in physical registers into virtual ones.
2446 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2447 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2450 if (NextVA.isMemLoc()) {
2451 MachineFrameInfo *MFI = MF.getFrameInfo();
2452 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2454 // Create load node to retrieve arguments from the stack.
2455 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2456 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2457 MachinePointerInfo::getFixedStack(FI),
2458 false, false, false, 0);
2460 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2461 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2464 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2468 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2469 unsigned &VARegSize, unsigned &VARegSaveSize)
2472 if (CCInfo.isFirstByValRegValid())
2473 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2475 unsigned int firstUnalloced;
2476 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2477 sizeof(GPRArgRegs) /
2478 sizeof(GPRArgRegs[0]));
2479 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2482 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2483 VARegSize = NumGPRs * 4;
2484 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2487 // The remaining GPRs hold either the beginning of variable-argument
2488 // data, or the beginning of an aggregate passed by value (usuall
2489 // byval). Either way, we allocate stack slots adjacent to the data
2490 // provided by our caller, and store the unallocated registers there.
2491 // If this is a variadic function, the va_list pointer will begin with
2492 // these values; otherwise, this reassembles a (byval) structure that
2493 // was split between registers and memory.
2495 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2496 DebugLoc dl, SDValue &Chain,
2497 unsigned ArgOffset) const {
2498 MachineFunction &MF = DAG.getMachineFunction();
2499 MachineFrameInfo *MFI = MF.getFrameInfo();
2500 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2501 unsigned firstRegToSaveIndex;
2502 if (CCInfo.isFirstByValRegValid())
2503 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2505 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2506 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2509 unsigned VARegSize, VARegSaveSize;
2510 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2511 if (VARegSaveSize) {
2512 // If this function is vararg, store any remaining integer argument regs
2513 // to their spots on the stack so that they may be loaded by deferencing
2514 // the result of va_next.
2515 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2516 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2517 ArgOffset + VARegSaveSize
2520 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2523 SmallVector<SDValue, 4> MemOps;
2524 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2525 const TargetRegisterClass *RC;
2526 if (AFI->isThumb1OnlyFunction())
2527 RC = ARM::tGPRRegisterClass;
2529 RC = ARM::GPRRegisterClass;
2531 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2532 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2534 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2535 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2537 MemOps.push_back(Store);
2538 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2539 DAG.getConstant(4, getPointerTy()));
2541 if (!MemOps.empty())
2542 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2543 &MemOps[0], MemOps.size());
2545 // This will point to the next argument passed via stack.
2546 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2550 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2551 CallingConv::ID CallConv, bool isVarArg,
2552 const SmallVectorImpl<ISD::InputArg>
2554 DebugLoc dl, SelectionDAG &DAG,
2555 SmallVectorImpl<SDValue> &InVals)
2557 MachineFunction &MF = DAG.getMachineFunction();
2558 MachineFrameInfo *MFI = MF.getFrameInfo();
2560 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2562 // Assign locations to all of the incoming arguments.
2563 SmallVector<CCValAssign, 16> ArgLocs;
2564 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2565 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2566 CCInfo.AnalyzeFormalArguments(Ins,
2567 CCAssignFnForNode(CallConv, /* Return*/ false,
2570 SmallVector<SDValue, 16> ArgValues;
2571 int lastInsIndex = -1;
2574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2575 CCValAssign &VA = ArgLocs[i];
2577 // Arguments stored in registers.
2578 if (VA.isRegLoc()) {
2579 EVT RegVT = VA.getLocVT();
2581 if (VA.needsCustom()) {
2582 // f64 and vector types are split up into multiple registers or
2583 // combinations of registers and stack slots.
2584 if (VA.getLocVT() == MVT::v2f64) {
2585 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2587 VA = ArgLocs[++i]; // skip ahead to next loc
2589 if (VA.isMemLoc()) {
2590 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2591 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2592 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2593 MachinePointerInfo::getFixedStack(FI),
2594 false, false, false, 0);
2596 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2599 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2600 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2601 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2602 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2603 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2605 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2608 const TargetRegisterClass *RC;
2610 if (RegVT == MVT::f32)
2611 RC = ARM::SPRRegisterClass;
2612 else if (RegVT == MVT::f64)
2613 RC = ARM::DPRRegisterClass;
2614 else if (RegVT == MVT::v2f64)
2615 RC = ARM::QPRRegisterClass;
2616 else if (RegVT == MVT::i32)
2617 RC = (AFI->isThumb1OnlyFunction() ?
2618 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2620 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2622 // Transform the arguments in physical registers into virtual ones.
2623 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2624 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2627 // If this is an 8 or 16-bit value, it is really passed promoted
2628 // to 32 bits. Insert an assert[sz]ext to capture this, then
2629 // truncate to the right size.
2630 switch (VA.getLocInfo()) {
2631 default: llvm_unreachable("Unknown loc info!");
2632 case CCValAssign::Full: break;
2633 case CCValAssign::BCvt:
2634 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2636 case CCValAssign::SExt:
2637 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2638 DAG.getValueType(VA.getValVT()));
2639 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2641 case CCValAssign::ZExt:
2642 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2643 DAG.getValueType(VA.getValVT()));
2644 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2648 InVals.push_back(ArgValue);
2650 } else { // VA.isRegLoc()
2653 assert(VA.isMemLoc());
2654 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2656 int index = ArgLocs[i].getValNo();
2658 // Some Ins[] entries become multiple ArgLoc[] entries.
2659 // Process them only once.
2660 if (index != lastInsIndex)
2662 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2663 // FIXME: For now, all byval parameter objects are marked mutable.
2664 // This can be changed with more analysis.
2665 // In case of tail call optimization mark all arguments mutable.
2666 // Since they could be overwritten by lowering of arguments in case of
2668 if (Flags.isByVal()) {
2669 unsigned VARegSize, VARegSaveSize;
2670 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2671 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2672 unsigned Bytes = Flags.getByValSize() - VARegSize;
2673 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2674 int FI = MFI->CreateFixedObject(Bytes,
2675 VA.getLocMemOffset(), false);
2676 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2678 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2679 VA.getLocMemOffset(), true);
2681 // Create load nodes to retrieve arguments from the stack.
2682 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2683 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2684 MachinePointerInfo::getFixedStack(FI),
2685 false, false, false, 0));
2687 lastInsIndex = index;
2694 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2699 /// isFloatingPointZero - Return true if this is +0.0.
2700 static bool isFloatingPointZero(SDValue Op) {
2701 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2702 return CFP->getValueAPF().isPosZero();
2703 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2704 // Maybe this has already been legalized into the constant pool?
2705 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2706 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2707 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2708 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2709 return CFP->getValueAPF().isPosZero();
2715 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2716 /// the given operands.
2718 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2719 SDValue &ARMcc, SelectionDAG &DAG,
2720 DebugLoc dl) const {
2721 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2722 unsigned C = RHSC->getZExtValue();
2723 if (!isLegalICmpImmediate(C)) {
2724 // Constant does not fit, try adjusting it by one?
2729 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2730 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2731 RHS = DAG.getConstant(C-1, MVT::i32);
2736 if (C != 0 && isLegalICmpImmediate(C-1)) {
2737 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2738 RHS = DAG.getConstant(C-1, MVT::i32);
2743 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2744 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2745 RHS = DAG.getConstant(C+1, MVT::i32);
2750 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2751 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2752 RHS = DAG.getConstant(C+1, MVT::i32);
2759 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2760 ARMISD::NodeType CompareType;
2763 CompareType = ARMISD::CMP;
2768 CompareType = ARMISD::CMPZ;
2771 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2772 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2775 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2777 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2778 DebugLoc dl) const {
2780 if (!isFloatingPointZero(RHS))
2781 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2783 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2784 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2787 /// duplicateCmp - Glue values can have only one use, so this function
2788 /// duplicates a comparison node.
2790 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2791 unsigned Opc = Cmp.getOpcode();
2792 DebugLoc DL = Cmp.getDebugLoc();
2793 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2794 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2796 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2797 Cmp = Cmp.getOperand(0);
2798 Opc = Cmp.getOpcode();
2799 if (Opc == ARMISD::CMPFP)
2800 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2802 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2803 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2805 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2808 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2809 SDValue Cond = Op.getOperand(0);
2810 SDValue SelectTrue = Op.getOperand(1);
2811 SDValue SelectFalse = Op.getOperand(2);
2812 DebugLoc dl = Op.getDebugLoc();
2816 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2817 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2819 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2820 const ConstantSDNode *CMOVTrue =
2821 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2822 const ConstantSDNode *CMOVFalse =
2823 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2825 if (CMOVTrue && CMOVFalse) {
2826 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2827 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2831 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2833 False = SelectFalse;
2834 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2839 if (True.getNode() && False.getNode()) {
2840 EVT VT = Op.getValueType();
2841 SDValue ARMcc = Cond.getOperand(2);
2842 SDValue CCR = Cond.getOperand(3);
2843 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2844 assert(True.getValueType() == VT);
2845 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2850 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2851 // undefined bits before doing a full-word comparison with zero.
2852 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2853 DAG.getConstant(1, Cond.getValueType()));
2855 return DAG.getSelectCC(dl, Cond,
2856 DAG.getConstant(0, Cond.getValueType()),
2857 SelectTrue, SelectFalse, ISD::SETNE);
2860 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2861 EVT VT = Op.getValueType();
2862 SDValue LHS = Op.getOperand(0);
2863 SDValue RHS = Op.getOperand(1);
2864 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2865 SDValue TrueVal = Op.getOperand(2);
2866 SDValue FalseVal = Op.getOperand(3);
2867 DebugLoc dl = Op.getDebugLoc();
2869 if (LHS.getValueType() == MVT::i32) {
2871 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2872 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2873 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2876 ARMCC::CondCodes CondCode, CondCode2;
2877 FPCCToARMCC(CC, CondCode, CondCode2);
2879 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2880 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2881 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2882 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2884 if (CondCode2 != ARMCC::AL) {
2885 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2886 // FIXME: Needs another CMP because flag can have but one use.
2887 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2888 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2889 Result, TrueVal, ARMcc2, CCR, Cmp2);
2894 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2895 /// to morph to an integer compare sequence.
2896 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2897 const ARMSubtarget *Subtarget) {
2898 SDNode *N = Op.getNode();
2899 if (!N->hasOneUse())
2900 // Otherwise it requires moving the value from fp to integer registers.
2902 if (!N->getNumValues())
2904 EVT VT = Op.getValueType();
2905 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2906 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2907 // vmrs are very slow, e.g. cortex-a8.
2910 if (isFloatingPointZero(Op)) {
2914 return ISD::isNormalLoad(N);
2917 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2918 if (isFloatingPointZero(Op))
2919 return DAG.getConstant(0, MVT::i32);
2921 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2922 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2923 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2924 Ld->isVolatile(), Ld->isNonTemporal(),
2925 Ld->isInvariant(), Ld->getAlignment());
2927 llvm_unreachable("Unknown VFP cmp argument!");
2930 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2931 SDValue &RetVal1, SDValue &RetVal2) {
2932 if (isFloatingPointZero(Op)) {
2933 RetVal1 = DAG.getConstant(0, MVT::i32);
2934 RetVal2 = DAG.getConstant(0, MVT::i32);
2938 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2939 SDValue Ptr = Ld->getBasePtr();
2940 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2941 Ld->getChain(), Ptr,
2942 Ld->getPointerInfo(),
2943 Ld->isVolatile(), Ld->isNonTemporal(),
2944 Ld->isInvariant(), Ld->getAlignment());
2946 EVT PtrType = Ptr.getValueType();
2947 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2948 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2949 PtrType, Ptr, DAG.getConstant(4, PtrType));
2950 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2951 Ld->getChain(), NewPtr,
2952 Ld->getPointerInfo().getWithOffset(4),
2953 Ld->isVolatile(), Ld->isNonTemporal(),
2954 Ld->isInvariant(), NewAlign);
2958 llvm_unreachable("Unknown VFP cmp argument!");
2961 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2962 /// f32 and even f64 comparisons to integer ones.
2964 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2965 SDValue Chain = Op.getOperand(0);
2966 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2967 SDValue LHS = Op.getOperand(2);
2968 SDValue RHS = Op.getOperand(3);
2969 SDValue Dest = Op.getOperand(4);
2970 DebugLoc dl = Op.getDebugLoc();
2972 bool SeenZero = false;
2973 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2974 canChangeToInt(RHS, SeenZero, Subtarget) &&
2975 // If one of the operand is zero, it's safe to ignore the NaN case since
2976 // we only care about equality comparisons.
2977 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2978 // If unsafe fp math optimization is enabled and there are no other uses of
2979 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2980 // to an integer comparison.
2981 if (CC == ISD::SETOEQ)
2983 else if (CC == ISD::SETUNE)
2987 if (LHS.getValueType() == MVT::f32) {
2988 LHS = bitcastf32Toi32(LHS, DAG);
2989 RHS = bitcastf32Toi32(RHS, DAG);
2990 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2991 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2992 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2993 Chain, Dest, ARMcc, CCR, Cmp);
2998 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2999 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3000 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3001 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3002 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3003 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3004 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3010 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3011 SDValue Chain = Op.getOperand(0);
3012 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3013 SDValue LHS = Op.getOperand(2);
3014 SDValue RHS = Op.getOperand(3);
3015 SDValue Dest = Op.getOperand(4);
3016 DebugLoc dl = Op.getDebugLoc();
3018 if (LHS.getValueType() == MVT::i32) {
3020 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3021 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3022 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3023 Chain, Dest, ARMcc, CCR, Cmp);
3026 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3028 if (getTargetMachine().Options.UnsafeFPMath &&
3029 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3030 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3031 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3032 if (Result.getNode())
3036 ARMCC::CondCodes CondCode, CondCode2;
3037 FPCCToARMCC(CC, CondCode, CondCode2);
3039 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3040 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3041 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3042 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3043 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3044 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3045 if (CondCode2 != ARMCC::AL) {
3046 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3047 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3048 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3053 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3054 SDValue Chain = Op.getOperand(0);
3055 SDValue Table = Op.getOperand(1);
3056 SDValue Index = Op.getOperand(2);
3057 DebugLoc dl = Op.getDebugLoc();
3059 EVT PTy = getPointerTy();
3060 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3061 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3062 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3063 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3064 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3065 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3066 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3067 if (Subtarget->isThumb2()) {
3068 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3069 // which does another jump to the destination. This also makes it easier
3070 // to translate it to TBB / TBH later.
3071 // FIXME: This might not work if the function is extremely large.
3072 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3073 Addr, Op.getOperand(2), JTI, UId);
3075 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3076 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3077 MachinePointerInfo::getJumpTable(),
3078 false, false, false, 0);
3079 Chain = Addr.getValue(1);
3080 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3081 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3083 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3084 MachinePointerInfo::getJumpTable(),
3085 false, false, false, 0);
3086 Chain = Addr.getValue(1);
3087 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3091 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3092 EVT VT = Op.getValueType();
3093 DebugLoc dl = Op.getDebugLoc();
3095 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3096 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3098 return DAG.UnrollVectorOp(Op.getNode());
3101 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3102 "Invalid type for custom lowering!");
3103 if (VT != MVT::v4i16)
3104 return DAG.UnrollVectorOp(Op.getNode());
3106 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3107 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3110 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3111 EVT VT = Op.getValueType();
3113 return LowerVectorFP_TO_INT(Op, DAG);
3115 DebugLoc dl = Op.getDebugLoc();
3118 switch (Op.getOpcode()) {
3119 default: llvm_unreachable("Invalid opcode!");
3120 case ISD::FP_TO_SINT:
3121 Opc = ARMISD::FTOSI;
3123 case ISD::FP_TO_UINT:
3124 Opc = ARMISD::FTOUI;
3127 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3128 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3131 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3132 EVT VT = Op.getValueType();
3133 DebugLoc dl = Op.getDebugLoc();
3135 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3136 if (VT.getVectorElementType() == MVT::f32)
3138 return DAG.UnrollVectorOp(Op.getNode());
3141 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3142 "Invalid type for custom lowering!");
3143 if (VT != MVT::v4f32)
3144 return DAG.UnrollVectorOp(Op.getNode());
3148 switch (Op.getOpcode()) {
3149 default: llvm_unreachable("Invalid opcode!");
3150 case ISD::SINT_TO_FP:
3151 CastOpc = ISD::SIGN_EXTEND;
3152 Opc = ISD::SINT_TO_FP;
3154 case ISD::UINT_TO_FP:
3155 CastOpc = ISD::ZERO_EXTEND;
3156 Opc = ISD::UINT_TO_FP;
3160 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3161 return DAG.getNode(Opc, dl, VT, Op);
3164 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3165 EVT VT = Op.getValueType();
3167 return LowerVectorINT_TO_FP(Op, DAG);
3169 DebugLoc dl = Op.getDebugLoc();
3172 switch (Op.getOpcode()) {
3173 default: llvm_unreachable("Invalid opcode!");
3174 case ISD::SINT_TO_FP:
3175 Opc = ARMISD::SITOF;
3177 case ISD::UINT_TO_FP:
3178 Opc = ARMISD::UITOF;
3182 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3183 return DAG.getNode(Opc, dl, VT, Op);
3186 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3187 // Implement fcopysign with a fabs and a conditional fneg.
3188 SDValue Tmp0 = Op.getOperand(0);
3189 SDValue Tmp1 = Op.getOperand(1);
3190 DebugLoc dl = Op.getDebugLoc();
3191 EVT VT = Op.getValueType();
3192 EVT SrcVT = Tmp1.getValueType();
3193 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3194 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3195 bool UseNEON = !InGPR && Subtarget->hasNEON();
3198 // Use VBSL to copy the sign bit.
3199 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3200 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3201 DAG.getTargetConstant(EncodedVal, MVT::i32));
3202 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3204 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3205 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3206 DAG.getConstant(32, MVT::i32));
3207 else /*if (VT == MVT::f32)*/
3208 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3209 if (SrcVT == MVT::f32) {
3210 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3212 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3213 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3214 DAG.getConstant(32, MVT::i32));
3215 } else if (VT == MVT::f32)
3216 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3217 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3218 DAG.getConstant(32, MVT::i32));
3219 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3220 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3222 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3224 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3225 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3226 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3228 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3229 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3230 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3231 if (VT == MVT::f32) {
3232 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3233 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3234 DAG.getConstant(0, MVT::i32));
3236 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3242 // Bitcast operand 1 to i32.
3243 if (SrcVT == MVT::f64)
3244 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3245 &Tmp1, 1).getValue(1);
3246 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3248 // Or in the signbit with integer operations.
3249 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3250 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3251 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3252 if (VT == MVT::f32) {
3253 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3254 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3255 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3256 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3259 // f64: Or the high part with signbit and then combine two parts.
3260 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3262 SDValue Lo = Tmp0.getValue(0);
3263 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3264 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3265 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3268 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3269 MachineFunction &MF = DAG.getMachineFunction();
3270 MachineFrameInfo *MFI = MF.getFrameInfo();
3271 MFI->setReturnAddressIsTaken(true);
3273 EVT VT = Op.getValueType();
3274 DebugLoc dl = Op.getDebugLoc();
3275 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3277 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3278 SDValue Offset = DAG.getConstant(4, MVT::i32);
3279 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3280 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3281 MachinePointerInfo(), false, false, false, 0);
3284 // Return LR, which contains the return address. Mark it an implicit live-in.
3285 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3286 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3289 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3290 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3291 MFI->setFrameAddressIsTaken(true);
3293 EVT VT = Op.getValueType();
3294 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3295 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3296 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3297 ? ARM::R7 : ARM::R11;
3298 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3300 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3301 MachinePointerInfo(),
3302 false, false, false, 0);
3306 /// ExpandBITCAST - If the target supports VFP, this function is called to
3307 /// expand a bit convert where either the source or destination type is i64 to
3308 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3309 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3310 /// vectors), since the legalizer won't know what to do with that.
3311 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3312 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3313 DebugLoc dl = N->getDebugLoc();
3314 SDValue Op = N->getOperand(0);
3316 // This function is only supposed to be called for i64 types, either as the
3317 // source or destination of the bit convert.
3318 EVT SrcVT = Op.getValueType();
3319 EVT DstVT = N->getValueType(0);
3320 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3321 "ExpandBITCAST called for non-i64 type");
3323 // Turn i64->f64 into VMOVDRR.
3324 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3325 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3326 DAG.getConstant(0, MVT::i32));
3327 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3328 DAG.getConstant(1, MVT::i32));
3329 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3330 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3333 // Turn f64->i64 into VMOVRRD.
3334 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3335 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3336 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3337 // Merge the pieces into a single i64 value.
3338 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3344 /// getZeroVector - Returns a vector of specified type with all zero elements.
3345 /// Zero vectors are used to represent vector negation and in those cases
3346 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3347 /// not support i64 elements, so sometimes the zero vectors will need to be
3348 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3350 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3351 assert(VT.isVector() && "Expected a vector type");
3352 // The canonical modified immediate encoding of a zero vector is....0!
3353 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3354 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3355 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3356 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3359 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3360 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3361 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3362 SelectionDAG &DAG) const {
3363 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3364 EVT VT = Op.getValueType();
3365 unsigned VTBits = VT.getSizeInBits();
3366 DebugLoc dl = Op.getDebugLoc();
3367 SDValue ShOpLo = Op.getOperand(0);
3368 SDValue ShOpHi = Op.getOperand(1);
3369 SDValue ShAmt = Op.getOperand(2);
3371 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3373 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3375 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3376 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3377 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3378 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3379 DAG.getConstant(VTBits, MVT::i32));
3380 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3381 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3382 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3384 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3385 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3387 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3388 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3391 SDValue Ops[2] = { Lo, Hi };
3392 return DAG.getMergeValues(Ops, 2, dl);
3395 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3396 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3397 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3398 SelectionDAG &DAG) const {
3399 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3400 EVT VT = Op.getValueType();
3401 unsigned VTBits = VT.getSizeInBits();
3402 DebugLoc dl = Op.getDebugLoc();
3403 SDValue ShOpLo = Op.getOperand(0);
3404 SDValue ShOpHi = Op.getOperand(1);
3405 SDValue ShAmt = Op.getOperand(2);
3408 assert(Op.getOpcode() == ISD::SHL_PARTS);
3409 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3410 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3411 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3412 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3413 DAG.getConstant(VTBits, MVT::i32));
3414 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3415 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3417 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3418 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3419 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3421 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3422 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3425 SDValue Ops[2] = { Lo, Hi };
3426 return DAG.getMergeValues(Ops, 2, dl);
3429 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3430 SelectionDAG &DAG) const {
3431 // The rounding mode is in bits 23:22 of the FPSCR.
3432 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3433 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3434 // so that the shift + and get folded into a bitfield extract.
3435 DebugLoc dl = Op.getDebugLoc();
3436 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3437 DAG.getConstant(Intrinsic::arm_get_fpscr,
3439 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3440 DAG.getConstant(1U << 22, MVT::i32));
3441 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3442 DAG.getConstant(22, MVT::i32));
3443 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3444 DAG.getConstant(3, MVT::i32));
3447 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3448 const ARMSubtarget *ST) {
3449 EVT VT = N->getValueType(0);
3450 DebugLoc dl = N->getDebugLoc();
3452 if (!ST->hasV6T2Ops())
3455 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3456 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3459 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3460 const ARMSubtarget *ST) {
3461 EVT VT = N->getValueType(0);
3462 DebugLoc dl = N->getDebugLoc();
3467 // Lower vector shifts on NEON to use VSHL.
3468 assert(ST->hasNEON() && "unexpected vector shift");
3470 // Left shifts translate directly to the vshiftu intrinsic.
3471 if (N->getOpcode() == ISD::SHL)
3472 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3473 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3474 N->getOperand(0), N->getOperand(1));
3476 assert((N->getOpcode() == ISD::SRA ||
3477 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3479 // NEON uses the same intrinsics for both left and right shifts. For
3480 // right shifts, the shift amounts are negative, so negate the vector of
3482 EVT ShiftVT = N->getOperand(1).getValueType();
3483 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3484 getZeroVector(ShiftVT, DAG, dl),
3486 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3487 Intrinsic::arm_neon_vshifts :
3488 Intrinsic::arm_neon_vshiftu);
3489 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3490 DAG.getConstant(vshiftInt, MVT::i32),
3491 N->getOperand(0), NegatedCount);
3494 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3495 const ARMSubtarget *ST) {
3496 EVT VT = N->getValueType(0);
3497 DebugLoc dl = N->getDebugLoc();
3499 // We can get here for a node like i32 = ISD::SHL i32, i64
3503 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3504 "Unknown shift to lower!");
3506 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3507 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3508 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3511 // If we are in thumb mode, we don't have RRX.
3512 if (ST->isThumb1Only()) return SDValue();
3514 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3515 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3516 DAG.getConstant(0, MVT::i32));
3517 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3518 DAG.getConstant(1, MVT::i32));
3520 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3521 // captures the result into a carry flag.
3522 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3523 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3525 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3526 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3528 // Merge the pieces into a single i64 value.
3529 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3532 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3533 SDValue TmpOp0, TmpOp1;
3534 bool Invert = false;
3538 SDValue Op0 = Op.getOperand(0);
3539 SDValue Op1 = Op.getOperand(1);
3540 SDValue CC = Op.getOperand(2);
3541 EVT VT = Op.getValueType();
3542 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3543 DebugLoc dl = Op.getDebugLoc();
3545 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3546 switch (SetCCOpcode) {
3547 default: llvm_unreachable("Illegal FP comparison");
3549 case ISD::SETNE: Invert = true; // Fallthrough
3551 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3553 case ISD::SETLT: Swap = true; // Fallthrough
3555 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3557 case ISD::SETLE: Swap = true; // Fallthrough
3559 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3560 case ISD::SETUGE: Swap = true; // Fallthrough
3561 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3562 case ISD::SETUGT: Swap = true; // Fallthrough
3563 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3564 case ISD::SETUEQ: Invert = true; // Fallthrough
3566 // Expand this to (OLT | OGT).
3570 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3571 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3573 case ISD::SETUO: Invert = true; // Fallthrough
3575 // Expand this to (OLT | OGE).
3579 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3580 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3584 // Integer comparisons.
3585 switch (SetCCOpcode) {
3586 default: llvm_unreachable("Illegal integer comparison");
3587 case ISD::SETNE: Invert = true;
3588 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3589 case ISD::SETLT: Swap = true;
3590 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3591 case ISD::SETLE: Swap = true;
3592 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3593 case ISD::SETULT: Swap = true;
3594 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3595 case ISD::SETULE: Swap = true;
3596 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3599 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3600 if (Opc == ARMISD::VCEQ) {
3603 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3605 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3608 // Ignore bitconvert.
3609 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3610 AndOp = AndOp.getOperand(0);
3612 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3614 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3615 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3622 std::swap(Op0, Op1);
3624 // If one of the operands is a constant vector zero, attempt to fold the
3625 // comparison to a specialized compare-against-zero form.
3627 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3629 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3630 if (Opc == ARMISD::VCGE)
3631 Opc = ARMISD::VCLEZ;
3632 else if (Opc == ARMISD::VCGT)
3633 Opc = ARMISD::VCLTZ;
3638 if (SingleOp.getNode()) {
3641 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3643 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3645 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3647 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3649 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3651 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3654 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3658 Result = DAG.getNOT(dl, Result, VT);
3663 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3664 /// valid vector constant for a NEON instruction with a "modified immediate"
3665 /// operand (e.g., VMOV). If so, return the encoded value.
3666 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3667 unsigned SplatBitSize, SelectionDAG &DAG,
3668 EVT &VT, bool is128Bits, NEONModImmType type) {
3669 unsigned OpCmode, Imm;
3671 // SplatBitSize is set to the smallest size that splats the vector, so a
3672 // zero vector will always have SplatBitSize == 8. However, NEON modified
3673 // immediate instructions others than VMOV do not support the 8-bit encoding
3674 // of a zero vector, and the default encoding of zero is supposed to be the
3679 switch (SplatBitSize) {
3681 if (type != VMOVModImm)
3683 // Any 1-byte value is OK. Op=0, Cmode=1110.
3684 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3687 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3691 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3692 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3693 if ((SplatBits & ~0xff) == 0) {
3694 // Value = 0x00nn: Op=x, Cmode=100x.
3699 if ((SplatBits & ~0xff00) == 0) {
3700 // Value = 0xnn00: Op=x, Cmode=101x.
3702 Imm = SplatBits >> 8;
3708 // NEON's 32-bit VMOV supports splat values where:
3709 // * only one byte is nonzero, or
3710 // * the least significant byte is 0xff and the second byte is nonzero, or
3711 // * the least significant 2 bytes are 0xff and the third is nonzero.
3712 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3713 if ((SplatBits & ~0xff) == 0) {
3714 // Value = 0x000000nn: Op=x, Cmode=000x.
3719 if ((SplatBits & ~0xff00) == 0) {
3720 // Value = 0x0000nn00: Op=x, Cmode=001x.
3722 Imm = SplatBits >> 8;
3725 if ((SplatBits & ~0xff0000) == 0) {
3726 // Value = 0x00nn0000: Op=x, Cmode=010x.
3728 Imm = SplatBits >> 16;
3731 if ((SplatBits & ~0xff000000) == 0) {
3732 // Value = 0xnn000000: Op=x, Cmode=011x.
3734 Imm = SplatBits >> 24;
3738 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3739 if (type == OtherModImm) return SDValue();
3741 if ((SplatBits & ~0xffff) == 0 &&
3742 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3743 // Value = 0x0000nnff: Op=x, Cmode=1100.
3745 Imm = SplatBits >> 8;
3750 if ((SplatBits & ~0xffffff) == 0 &&
3751 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3752 // Value = 0x00nnffff: Op=x, Cmode=1101.
3754 Imm = SplatBits >> 16;
3755 SplatBits |= 0xffff;
3759 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3760 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3761 // VMOV.I32. A (very) minor optimization would be to replicate the value
3762 // and fall through here to test for a valid 64-bit splat. But, then the
3763 // caller would also need to check and handle the change in size.
3767 if (type != VMOVModImm)
3769 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3770 uint64_t BitMask = 0xff;
3772 unsigned ImmMask = 1;
3774 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3775 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3778 } else if ((SplatBits & BitMask) != 0) {
3784 // Op=1, Cmode=1110.
3787 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3792 llvm_unreachable("unexpected size for isNEONModifiedImm");
3795 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3796 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3799 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3800 bool &ReverseVEXT, unsigned &Imm) {
3801 unsigned NumElts = VT.getVectorNumElements();
3802 ReverseVEXT = false;
3804 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3810 // If this is a VEXT shuffle, the immediate value is the index of the first
3811 // element. The other shuffle indices must be the successive elements after
3813 unsigned ExpectedElt = Imm;
3814 for (unsigned i = 1; i < NumElts; ++i) {
3815 // Increment the expected index. If it wraps around, it may still be
3816 // a VEXT but the source vectors must be swapped.
3818 if (ExpectedElt == NumElts * 2) {
3823 if (M[i] < 0) continue; // ignore UNDEF indices
3824 if (ExpectedElt != static_cast<unsigned>(M[i]))
3828 // Adjust the index value if the source operands will be swapped.
3835 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3836 /// instruction with the specified blocksize. (The order of the elements
3837 /// within each block of the vector is reversed.)
3838 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3839 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3840 "Only possible block sizes for VREV are: 16, 32, 64");
3842 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3846 unsigned NumElts = VT.getVectorNumElements();
3847 unsigned BlockElts = M[0] + 1;
3848 // If the first shuffle index is UNDEF, be optimistic.
3850 BlockElts = BlockSize / EltSz;
3852 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3855 for (unsigned i = 0; i < NumElts; ++i) {
3856 if (M[i] < 0) continue; // ignore UNDEF indices
3857 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3864 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3865 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3866 // range, then 0 is placed into the resulting vector. So pretty much any mask
3867 // of 8 elements can work here.
3868 return VT == MVT::v8i8 && M.size() == 8;
3871 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3872 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3876 unsigned NumElts = VT.getVectorNumElements();
3877 WhichResult = (M[0] == 0 ? 0 : 1);
3878 for (unsigned i = 0; i < NumElts; i += 2) {
3879 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3880 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3886 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3887 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3888 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3889 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3890 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3894 unsigned NumElts = VT.getVectorNumElements();
3895 WhichResult = (M[0] == 0 ? 0 : 1);
3896 for (unsigned i = 0; i < NumElts; i += 2) {
3897 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3898 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3904 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3905 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3909 unsigned NumElts = VT.getVectorNumElements();
3910 WhichResult = (M[0] == 0 ? 0 : 1);
3911 for (unsigned i = 0; i != NumElts; ++i) {
3912 if (M[i] < 0) continue; // ignore UNDEF indices
3913 if ((unsigned) M[i] != 2 * i + WhichResult)
3917 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3918 if (VT.is64BitVector() && EltSz == 32)
3924 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3925 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3926 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3927 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3928 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3932 unsigned Half = VT.getVectorNumElements() / 2;
3933 WhichResult = (M[0] == 0 ? 0 : 1);
3934 for (unsigned j = 0; j != 2; ++j) {
3935 unsigned Idx = WhichResult;
3936 for (unsigned i = 0; i != Half; ++i) {
3937 int MIdx = M[i + j * Half];
3938 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3944 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3945 if (VT.is64BitVector() && EltSz == 32)
3951 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3952 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3956 unsigned NumElts = VT.getVectorNumElements();
3957 WhichResult = (M[0] == 0 ? 0 : 1);
3958 unsigned Idx = WhichResult * NumElts / 2;
3959 for (unsigned i = 0; i != NumElts; i += 2) {
3960 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3961 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3966 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3967 if (VT.is64BitVector() && EltSz == 32)
3973 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3974 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3975 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3976 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3977 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3981 unsigned NumElts = VT.getVectorNumElements();
3982 WhichResult = (M[0] == 0 ? 0 : 1);
3983 unsigned Idx = WhichResult * NumElts / 2;
3984 for (unsigned i = 0; i != NumElts; i += 2) {
3985 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3986 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3991 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3992 if (VT.is64BitVector() && EltSz == 32)
3998 // If N is an integer constant that can be moved into a register in one
3999 // instruction, return an SDValue of such a constant (will become a MOV
4000 // instruction). Otherwise return null.
4001 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4002 const ARMSubtarget *ST, DebugLoc dl) {
4004 if (!isa<ConstantSDNode>(N))
4006 Val = cast<ConstantSDNode>(N)->getZExtValue();
4008 if (ST->isThumb1Only()) {
4009 if (Val <= 255 || ~Val <= 255)
4010 return DAG.getConstant(Val, MVT::i32);
4012 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4013 return DAG.getConstant(Val, MVT::i32);
4018 // If this is a case we can't handle, return null and let the default
4019 // expansion code take care of it.
4020 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4021 const ARMSubtarget *ST) const {
4022 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4023 DebugLoc dl = Op.getDebugLoc();
4024 EVT VT = Op.getValueType();
4026 APInt SplatBits, SplatUndef;
4027 unsigned SplatBitSize;
4029 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4030 if (SplatBitSize <= 64) {
4031 // Check if an immediate VMOV works.
4033 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4034 SplatUndef.getZExtValue(), SplatBitSize,
4035 DAG, VmovVT, VT.is128BitVector(),
4037 if (Val.getNode()) {
4038 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4039 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4042 // Try an immediate VMVN.
4043 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4044 Val = isNEONModifiedImm(NegatedImm,
4045 SplatUndef.getZExtValue(), SplatBitSize,
4046 DAG, VmovVT, VT.is128BitVector(),
4048 if (Val.getNode()) {
4049 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4050 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4053 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4054 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4055 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4057 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4058 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4064 // Scan through the operands to see if only one value is used.
4065 unsigned NumElts = VT.getVectorNumElements();
4066 bool isOnlyLowElement = true;
4067 bool usesOnlyOneValue = true;
4068 bool isConstant = true;
4070 for (unsigned i = 0; i < NumElts; ++i) {
4071 SDValue V = Op.getOperand(i);
4072 if (V.getOpcode() == ISD::UNDEF)
4075 isOnlyLowElement = false;
4076 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4079 if (!Value.getNode())
4081 else if (V != Value)
4082 usesOnlyOneValue = false;
4085 if (!Value.getNode())
4086 return DAG.getUNDEF(VT);
4088 if (isOnlyLowElement)
4089 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4091 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4093 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4094 // i32 and try again.
4095 if (usesOnlyOneValue && EltSize <= 32) {
4097 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4098 if (VT.getVectorElementType().isFloatingPoint()) {
4099 SmallVector<SDValue, 8> Ops;
4100 for (unsigned i = 0; i < NumElts; ++i)
4101 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4103 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4104 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4105 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4107 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4109 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4111 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4114 // If all elements are constants and the case above didn't get hit, fall back
4115 // to the default expansion, which will generate a load from the constant
4120 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4122 SDValue shuffle = ReconstructShuffle(Op, DAG);
4123 if (shuffle != SDValue())
4127 // Vectors with 32- or 64-bit elements can be built by directly assigning
4128 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4129 // will be legalized.
4130 if (EltSize >= 32) {
4131 // Do the expansion with floating-point types, since that is what the VFP
4132 // registers are defined to use, and since i64 is not legal.
4133 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4134 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4135 SmallVector<SDValue, 8> Ops;
4136 for (unsigned i = 0; i < NumElts; ++i)
4137 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4138 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4139 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4145 // Gather data to see if the operation can be modelled as a
4146 // shuffle in combination with VEXTs.
4147 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4148 SelectionDAG &DAG) const {
4149 DebugLoc dl = Op.getDebugLoc();
4150 EVT VT = Op.getValueType();
4151 unsigned NumElts = VT.getVectorNumElements();
4153 SmallVector<SDValue, 2> SourceVecs;
4154 SmallVector<unsigned, 2> MinElts;
4155 SmallVector<unsigned, 2> MaxElts;
4157 for (unsigned i = 0; i < NumElts; ++i) {
4158 SDValue V = Op.getOperand(i);
4159 if (V.getOpcode() == ISD::UNDEF)
4161 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4162 // A shuffle can only come from building a vector from various
4163 // elements of other vectors.
4165 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4166 VT.getVectorElementType()) {
4167 // This code doesn't know how to handle shuffles where the vector
4168 // element types do not match (this happens because type legalization
4169 // promotes the return type of EXTRACT_VECTOR_ELT).
4170 // FIXME: It might be appropriate to extend this code to handle
4171 // mismatched types.
4175 // Record this extraction against the appropriate vector if possible...
4176 SDValue SourceVec = V.getOperand(0);
4177 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4178 bool FoundSource = false;
4179 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4180 if (SourceVecs[j] == SourceVec) {
4181 if (MinElts[j] > EltNo)
4183 if (MaxElts[j] < EltNo)
4190 // Or record a new source if not...
4192 SourceVecs.push_back(SourceVec);
4193 MinElts.push_back(EltNo);
4194 MaxElts.push_back(EltNo);
4198 // Currently only do something sane when at most two source vectors
4200 if (SourceVecs.size() > 2)
4203 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4204 int VEXTOffsets[2] = {0, 0};
4206 // This loop extracts the usage patterns of the source vectors
4207 // and prepares appropriate SDValues for a shuffle if possible.
4208 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4209 if (SourceVecs[i].getValueType() == VT) {
4210 // No VEXT necessary
4211 ShuffleSrcs[i] = SourceVecs[i];
4214 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4215 // It probably isn't worth padding out a smaller vector just to
4216 // break it down again in a shuffle.
4220 // Since only 64-bit and 128-bit vectors are legal on ARM and
4221 // we've eliminated the other cases...
4222 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4223 "unexpected vector sizes in ReconstructShuffle");
4225 if (MaxElts[i] - MinElts[i] >= NumElts) {
4226 // Span too large for a VEXT to cope
4230 if (MinElts[i] >= NumElts) {
4231 // The extraction can just take the second half
4232 VEXTOffsets[i] = NumElts;
4233 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4235 DAG.getIntPtrConstant(NumElts));
4236 } else if (MaxElts[i] < NumElts) {
4237 // The extraction can just take the first half
4239 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4241 DAG.getIntPtrConstant(0));
4243 // An actual VEXT is needed
4244 VEXTOffsets[i] = MinElts[i];
4245 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4247 DAG.getIntPtrConstant(0));
4248 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4250 DAG.getIntPtrConstant(NumElts));
4251 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4252 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4256 SmallVector<int, 8> Mask;
4258 for (unsigned i = 0; i < NumElts; ++i) {
4259 SDValue Entry = Op.getOperand(i);
4260 if (Entry.getOpcode() == ISD::UNDEF) {
4265 SDValue ExtractVec = Entry.getOperand(0);
4266 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4267 .getOperand(1))->getSExtValue();
4268 if (ExtractVec == SourceVecs[0]) {
4269 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4271 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4275 // Final check before we try to produce nonsense...
4276 if (isShuffleMaskLegal(Mask, VT))
4277 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4283 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4284 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4285 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4286 /// are assumed to be legal.
4288 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4290 if (VT.getVectorNumElements() == 4 &&
4291 (VT.is128BitVector() || VT.is64BitVector())) {
4292 unsigned PFIndexes[4];
4293 for (unsigned i = 0; i != 4; ++i) {
4297 PFIndexes[i] = M[i];
4300 // Compute the index in the perfect shuffle table.
4301 unsigned PFTableIndex =
4302 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4303 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4304 unsigned Cost = (PFEntry >> 30);
4311 unsigned Imm, WhichResult;
4313 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4314 return (EltSize >= 32 ||
4315 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4316 isVREVMask(M, VT, 64) ||
4317 isVREVMask(M, VT, 32) ||
4318 isVREVMask(M, VT, 16) ||
4319 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4320 isVTBLMask(M, VT) ||
4321 isVTRNMask(M, VT, WhichResult) ||
4322 isVUZPMask(M, VT, WhichResult) ||
4323 isVZIPMask(M, VT, WhichResult) ||
4324 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4325 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4326 isVZIP_v_undef_Mask(M, VT, WhichResult));
4329 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4330 /// the specified operations to build the shuffle.
4331 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4332 SDValue RHS, SelectionDAG &DAG,
4334 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4335 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4336 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4339 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4348 OP_VUZPL, // VUZP, left result
4349 OP_VUZPR, // VUZP, right result
4350 OP_VZIPL, // VZIP, left result
4351 OP_VZIPR, // VZIP, right result
4352 OP_VTRNL, // VTRN, left result
4353 OP_VTRNR // VTRN, right result
4356 if (OpNum == OP_COPY) {
4357 if (LHSID == (1*9+2)*9+3) return LHS;
4358 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4362 SDValue OpLHS, OpRHS;
4363 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4364 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4365 EVT VT = OpLHS.getValueType();
4368 default: llvm_unreachable("Unknown shuffle opcode!");
4370 // VREV divides the vector in half and swaps within the half.
4371 if (VT.getVectorElementType() == MVT::i32 ||
4372 VT.getVectorElementType() == MVT::f32)
4373 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4374 // vrev <4 x i16> -> VREV32
4375 if (VT.getVectorElementType() == MVT::i16)
4376 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4377 // vrev <4 x i8> -> VREV16
4378 assert(VT.getVectorElementType() == MVT::i8);
4379 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4384 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4385 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4389 return DAG.getNode(ARMISD::VEXT, dl, VT,
4391 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4394 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4395 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4398 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4399 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4402 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4403 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4407 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4408 ArrayRef<int> ShuffleMask,
4409 SelectionDAG &DAG) {
4410 // Check to see if we can use the VTBL instruction.
4411 SDValue V1 = Op.getOperand(0);
4412 SDValue V2 = Op.getOperand(1);
4413 DebugLoc DL = Op.getDebugLoc();
4415 SmallVector<SDValue, 8> VTBLMask;
4416 for (ArrayRef<int>::iterator
4417 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4418 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4420 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4421 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4422 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4425 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4426 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4430 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4431 SDValue V1 = Op.getOperand(0);
4432 SDValue V2 = Op.getOperand(1);
4433 DebugLoc dl = Op.getDebugLoc();
4434 EVT VT = Op.getValueType();
4435 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4437 // Convert shuffles that are directly supported on NEON to target-specific
4438 // DAG nodes, instead of keeping them as shuffles and matching them again
4439 // during code selection. This is more efficient and avoids the possibility
4440 // of inconsistencies between legalization and selection.
4441 // FIXME: floating-point vectors should be canonicalized to integer vectors
4442 // of the same time so that they get CSEd properly.
4443 ArrayRef<int> ShuffleMask = SVN->getMask();
4445 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4446 if (EltSize <= 32) {
4447 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4448 int Lane = SVN->getSplatIndex();
4449 // If this is undef splat, generate it via "just" vdup, if possible.
4450 if (Lane == -1) Lane = 0;
4452 // Test if V1 is a SCALAR_TO_VECTOR.
4453 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4454 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4456 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4457 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4459 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4460 !isa<ConstantSDNode>(V1.getOperand(0))) {
4461 bool IsScalarToVector = true;
4462 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4463 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4464 IsScalarToVector = false;
4467 if (IsScalarToVector)
4468 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4470 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4471 DAG.getConstant(Lane, MVT::i32));
4476 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4479 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4480 DAG.getConstant(Imm, MVT::i32));
4483 if (isVREVMask(ShuffleMask, VT, 64))
4484 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4485 if (isVREVMask(ShuffleMask, VT, 32))
4486 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4487 if (isVREVMask(ShuffleMask, VT, 16))
4488 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4490 // Check for Neon shuffles that modify both input vectors in place.
4491 // If both results are used, i.e., if there are two shuffles with the same
4492 // source operands and with masks corresponding to both results of one of
4493 // these operations, DAG memoization will ensure that a single node is
4494 // used for both shuffles.
4495 unsigned WhichResult;
4496 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4497 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4498 V1, V2).getValue(WhichResult);
4499 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4500 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4501 V1, V2).getValue(WhichResult);
4502 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4503 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4504 V1, V2).getValue(WhichResult);
4506 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4507 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4508 V1, V1).getValue(WhichResult);
4509 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4510 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4511 V1, V1).getValue(WhichResult);
4512 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4513 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4514 V1, V1).getValue(WhichResult);
4517 // If the shuffle is not directly supported and it has 4 elements, use
4518 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4519 unsigned NumElts = VT.getVectorNumElements();
4521 unsigned PFIndexes[4];
4522 for (unsigned i = 0; i != 4; ++i) {
4523 if (ShuffleMask[i] < 0)
4526 PFIndexes[i] = ShuffleMask[i];
4529 // Compute the index in the perfect shuffle table.
4530 unsigned PFTableIndex =
4531 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4532 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4533 unsigned Cost = (PFEntry >> 30);
4536 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4539 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4540 if (EltSize >= 32) {
4541 // Do the expansion with floating-point types, since that is what the VFP
4542 // registers are defined to use, and since i64 is not legal.
4543 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4544 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4545 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4546 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4547 SmallVector<SDValue, 8> Ops;
4548 for (unsigned i = 0; i < NumElts; ++i) {
4549 if (ShuffleMask[i] < 0)
4550 Ops.push_back(DAG.getUNDEF(EltVT));
4552 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4553 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4554 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4557 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4558 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4561 if (VT == MVT::v8i8) {
4562 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4563 if (NewOp.getNode())
4570 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4571 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4572 SDValue Lane = Op.getOperand(2);
4573 if (!isa<ConstantSDNode>(Lane))
4579 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4580 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4581 SDValue Lane = Op.getOperand(1);
4582 if (!isa<ConstantSDNode>(Lane))
4585 SDValue Vec = Op.getOperand(0);
4586 if (Op.getValueType() == MVT::i32 &&
4587 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4588 DebugLoc dl = Op.getDebugLoc();
4589 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4595 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4596 // The only time a CONCAT_VECTORS operation can have legal types is when
4597 // two 64-bit vectors are concatenated to a 128-bit vector.
4598 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4599 "unexpected CONCAT_VECTORS");
4600 DebugLoc dl = Op.getDebugLoc();
4601 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4602 SDValue Op0 = Op.getOperand(0);
4603 SDValue Op1 = Op.getOperand(1);
4604 if (Op0.getOpcode() != ISD::UNDEF)
4605 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4606 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4607 DAG.getIntPtrConstant(0));
4608 if (Op1.getOpcode() != ISD::UNDEF)
4609 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4610 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4611 DAG.getIntPtrConstant(1));
4612 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4615 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4616 /// element has been zero/sign-extended, depending on the isSigned parameter,
4617 /// from an integer type half its size.
4618 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4620 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4621 EVT VT = N->getValueType(0);
4622 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4623 SDNode *BVN = N->getOperand(0).getNode();
4624 if (BVN->getValueType(0) != MVT::v4i32 ||
4625 BVN->getOpcode() != ISD::BUILD_VECTOR)
4627 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4628 unsigned HiElt = 1 - LoElt;
4629 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4630 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4631 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4632 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4633 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4636 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4637 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4640 if (Hi0->isNullValue() && Hi1->isNullValue())
4646 if (N->getOpcode() != ISD::BUILD_VECTOR)
4649 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4650 SDNode *Elt = N->getOperand(i).getNode();
4651 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4652 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4653 unsigned HalfSize = EltSize / 2;
4655 if (!isIntN(HalfSize, C->getSExtValue()))
4658 if (!isUIntN(HalfSize, C->getZExtValue()))
4669 /// isSignExtended - Check if a node is a vector value that is sign-extended
4670 /// or a constant BUILD_VECTOR with sign-extended elements.
4671 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4672 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4674 if (isExtendedBUILD_VECTOR(N, DAG, true))
4679 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4680 /// or a constant BUILD_VECTOR with zero-extended elements.
4681 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4682 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4684 if (isExtendedBUILD_VECTOR(N, DAG, false))
4689 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4690 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4691 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4692 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4693 return N->getOperand(0);
4694 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4695 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4696 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4697 LD->isNonTemporal(), LD->isInvariant(),
4698 LD->getAlignment());
4699 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4700 // have been legalized as a BITCAST from v4i32.
4701 if (N->getOpcode() == ISD::BITCAST) {
4702 SDNode *BVN = N->getOperand(0).getNode();
4703 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4704 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4705 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4706 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4707 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4709 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4710 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4711 EVT VT = N->getValueType(0);
4712 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4713 unsigned NumElts = VT.getVectorNumElements();
4714 MVT TruncVT = MVT::getIntegerVT(EltSize);
4715 SmallVector<SDValue, 8> Ops;
4716 for (unsigned i = 0; i != NumElts; ++i) {
4717 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4718 const APInt &CInt = C->getAPIntValue();
4719 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4721 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4722 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4725 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4726 unsigned Opcode = N->getOpcode();
4727 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4728 SDNode *N0 = N->getOperand(0).getNode();
4729 SDNode *N1 = N->getOperand(1).getNode();
4730 return N0->hasOneUse() && N1->hasOneUse() &&
4731 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4736 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4737 unsigned Opcode = N->getOpcode();
4738 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4739 SDNode *N0 = N->getOperand(0).getNode();
4740 SDNode *N1 = N->getOperand(1).getNode();
4741 return N0->hasOneUse() && N1->hasOneUse() &&
4742 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4747 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4748 // Multiplications are only custom-lowered for 128-bit vectors so that
4749 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4750 EVT VT = Op.getValueType();
4751 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4752 SDNode *N0 = Op.getOperand(0).getNode();
4753 SDNode *N1 = Op.getOperand(1).getNode();
4754 unsigned NewOpc = 0;
4756 bool isN0SExt = isSignExtended(N0, DAG);
4757 bool isN1SExt = isSignExtended(N1, DAG);
4758 if (isN0SExt && isN1SExt)
4759 NewOpc = ARMISD::VMULLs;
4761 bool isN0ZExt = isZeroExtended(N0, DAG);
4762 bool isN1ZExt = isZeroExtended(N1, DAG);
4763 if (isN0ZExt && isN1ZExt)
4764 NewOpc = ARMISD::VMULLu;
4765 else if (isN1SExt || isN1ZExt) {
4766 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4767 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4768 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4769 NewOpc = ARMISD::VMULLs;
4771 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4772 NewOpc = ARMISD::VMULLu;
4774 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4776 NewOpc = ARMISD::VMULLu;
4782 if (VT == MVT::v2i64)
4783 // Fall through to expand this. It is not legal.
4786 // Other vector multiplications are legal.
4791 // Legalize to a VMULL instruction.
4792 DebugLoc DL = Op.getDebugLoc();
4794 SDValue Op1 = SkipExtension(N1, DAG);
4796 Op0 = SkipExtension(N0, DAG);
4797 assert(Op0.getValueType().is64BitVector() &&
4798 Op1.getValueType().is64BitVector() &&
4799 "unexpected types for extended operands to VMULL");
4800 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4803 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4804 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4811 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4812 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4813 EVT Op1VT = Op1.getValueType();
4814 return DAG.getNode(N0->getOpcode(), DL, VT,
4815 DAG.getNode(NewOpc, DL, VT,
4816 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4817 DAG.getNode(NewOpc, DL, VT,
4818 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4822 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4824 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4825 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4826 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4827 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4828 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4829 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4830 // Get reciprocal estimate.
4831 // float4 recip = vrecpeq_f32(yf);
4832 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4833 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4834 // Because char has a smaller range than uchar, we can actually get away
4835 // without any newton steps. This requires that we use a weird bias
4836 // of 0xb000, however (again, this has been exhaustively tested).
4837 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4838 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4839 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4840 Y = DAG.getConstant(0xb000, MVT::i32);
4841 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4842 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4843 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4844 // Convert back to short.
4845 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4846 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4851 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4853 // Convert to float.
4854 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4855 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4856 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4857 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4858 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4859 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4861 // Use reciprocal estimate and one refinement step.
4862 // float4 recip = vrecpeq_f32(yf);
4863 // recip *= vrecpsq_f32(yf, recip);
4864 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4865 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4866 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4867 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4869 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4870 // Because short has a smaller range than ushort, we can actually get away
4871 // with only a single newton step. This requires that we use a weird bias
4872 // of 89, however (again, this has been exhaustively tested).
4873 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4874 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4875 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4876 N1 = DAG.getConstant(0x89, MVT::i32);
4877 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4878 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4879 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4880 // Convert back to integer and return.
4881 // return vmovn_s32(vcvt_s32_f32(result));
4882 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4883 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4887 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4888 EVT VT = Op.getValueType();
4889 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4890 "unexpected type for custom-lowering ISD::SDIV");
4892 DebugLoc dl = Op.getDebugLoc();
4893 SDValue N0 = Op.getOperand(0);
4894 SDValue N1 = Op.getOperand(1);
4897 if (VT == MVT::v8i8) {
4898 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4899 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4901 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4902 DAG.getIntPtrConstant(4));
4903 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4904 DAG.getIntPtrConstant(4));
4905 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4906 DAG.getIntPtrConstant(0));
4907 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4908 DAG.getIntPtrConstant(0));
4910 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4911 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4913 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4914 N0 = LowerCONCAT_VECTORS(N0, DAG);
4916 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4919 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4922 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4923 EVT VT = Op.getValueType();
4924 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4925 "unexpected type for custom-lowering ISD::UDIV");
4927 DebugLoc dl = Op.getDebugLoc();
4928 SDValue N0 = Op.getOperand(0);
4929 SDValue N1 = Op.getOperand(1);
4932 if (VT == MVT::v8i8) {
4933 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4934 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4936 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4937 DAG.getIntPtrConstant(4));
4938 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4939 DAG.getIntPtrConstant(4));
4940 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4941 DAG.getIntPtrConstant(0));
4942 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4943 DAG.getIntPtrConstant(0));
4945 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4946 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4948 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4949 N0 = LowerCONCAT_VECTORS(N0, DAG);
4951 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4952 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4957 // v4i16 sdiv ... Convert to float.
4958 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4959 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4960 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4961 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4962 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4963 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4965 // Use reciprocal estimate and two refinement steps.
4966 // float4 recip = vrecpeq_f32(yf);
4967 // recip *= vrecpsq_f32(yf, recip);
4968 // recip *= vrecpsq_f32(yf, recip);
4969 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4970 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4971 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4972 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4974 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4975 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4976 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4978 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4979 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4980 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4981 // and that it will never cause us to return an answer too large).
4982 // float4 result = as_float4(as_int4(xf*recip) + 2);
4983 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4984 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4985 N1 = DAG.getConstant(2, MVT::i32);
4986 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4987 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4988 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4989 // Convert back to integer and return.
4990 // return vmovn_u32(vcvt_s32_f32(result));
4991 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4992 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4996 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4997 EVT VT = Op.getNode()->getValueType(0);
4998 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5001 bool ExtraOp = false;
5002 switch (Op.getOpcode()) {
5003 default: llvm_unreachable("Invalid code");
5004 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5005 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5006 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5007 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5011 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5013 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5014 Op.getOperand(1), Op.getOperand(2));
5017 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5018 // Monotonic load/store is legal for all targets
5019 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5022 // Aquire/Release load/store is not legal for targets without a
5023 // dmb or equivalent available.
5029 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5030 SelectionDAG &DAG, unsigned NewOp) {
5031 DebugLoc dl = Node->getDebugLoc();
5032 assert (Node->getValueType(0) == MVT::i64 &&
5033 "Only know how to expand i64 atomics");
5035 SmallVector<SDValue, 6> Ops;
5036 Ops.push_back(Node->getOperand(0)); // Chain
5037 Ops.push_back(Node->getOperand(1)); // Ptr
5039 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5040 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5041 // High part of Val1
5042 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5043 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5044 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5045 // High part of Val1
5046 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5047 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5048 // High part of Val2
5049 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5050 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5052 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5054 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5055 cast<MemSDNode>(Node)->getMemOperand());
5056 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5057 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5058 Results.push_back(Result.getValue(2));
5061 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5062 switch (Op.getOpcode()) {
5063 default: llvm_unreachable("Don't know how to custom lower this!");
5064 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5065 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5066 case ISD::GlobalAddress:
5067 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5068 LowerGlobalAddressELF(Op, DAG);
5069 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5070 case ISD::SELECT: return LowerSELECT(Op, DAG);
5071 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5072 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5073 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5074 case ISD::VASTART: return LowerVASTART(Op, DAG);
5075 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5076 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5077 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5078 case ISD::SINT_TO_FP:
5079 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5080 case ISD::FP_TO_SINT:
5081 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5082 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5083 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5084 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5085 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5086 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5087 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5088 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5090 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5093 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5094 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5095 case ISD::SRL_PARTS:
5096 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5097 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5098 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5099 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5100 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5101 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5102 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5103 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5104 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5105 case ISD::MUL: return LowerMUL(Op, DAG);
5106 case ISD::SDIV: return LowerSDIV(Op, DAG);
5107 case ISD::UDIV: return LowerUDIV(Op, DAG);
5111 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5112 case ISD::ATOMIC_LOAD:
5113 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5117 /// ReplaceNodeResults - Replace the results of node with an illegal result
5118 /// type with new values built out of custom code.
5119 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5120 SmallVectorImpl<SDValue>&Results,
5121 SelectionDAG &DAG) const {
5123 switch (N->getOpcode()) {
5125 llvm_unreachable("Don't know how to custom expand this!");
5127 Res = ExpandBITCAST(N, DAG);
5131 Res = Expand64BitShift(N, DAG, Subtarget);
5133 case ISD::ATOMIC_LOAD_ADD:
5134 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5136 case ISD::ATOMIC_LOAD_AND:
5137 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5139 case ISD::ATOMIC_LOAD_NAND:
5140 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5142 case ISD::ATOMIC_LOAD_OR:
5143 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5145 case ISD::ATOMIC_LOAD_SUB:
5146 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5148 case ISD::ATOMIC_LOAD_XOR:
5149 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5151 case ISD::ATOMIC_SWAP:
5152 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5154 case ISD::ATOMIC_CMP_SWAP:
5155 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5159 Results.push_back(Res);
5162 //===----------------------------------------------------------------------===//
5163 // ARM Scheduler Hooks
5164 //===----------------------------------------------------------------------===//
5167 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5168 MachineBasicBlock *BB,
5169 unsigned Size) const {
5170 unsigned dest = MI->getOperand(0).getReg();
5171 unsigned ptr = MI->getOperand(1).getReg();
5172 unsigned oldval = MI->getOperand(2).getReg();
5173 unsigned newval = MI->getOperand(3).getReg();
5174 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5175 DebugLoc dl = MI->getDebugLoc();
5176 bool isThumb2 = Subtarget->isThumb2();
5178 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5180 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5181 : ARM::GPRRegisterClass);
5184 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5185 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5186 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5189 unsigned ldrOpc, strOpc;
5191 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5193 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5194 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5197 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5198 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5201 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5202 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5206 MachineFunction *MF = BB->getParent();
5207 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5208 MachineFunction::iterator It = BB;
5209 ++It; // insert the new blocks after the current block
5211 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5212 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5213 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5214 MF->insert(It, loop1MBB);
5215 MF->insert(It, loop2MBB);
5216 MF->insert(It, exitMBB);
5218 // Transfer the remainder of BB and its successor edges to exitMBB.
5219 exitMBB->splice(exitMBB->begin(), BB,
5220 llvm::next(MachineBasicBlock::iterator(MI)),
5222 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5226 // fallthrough --> loop1MBB
5227 BB->addSuccessor(loop1MBB);
5230 // ldrex dest, [ptr]
5234 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5235 if (ldrOpc == ARM::t2LDREX)
5237 AddDefaultPred(MIB);
5238 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5239 .addReg(dest).addReg(oldval));
5240 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5241 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5242 BB->addSuccessor(loop2MBB);
5243 BB->addSuccessor(exitMBB);
5246 // strex scratch, newval, [ptr]
5250 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5251 if (strOpc == ARM::t2STREX)
5253 AddDefaultPred(MIB);
5254 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5255 .addReg(scratch).addImm(0));
5256 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5257 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5258 BB->addSuccessor(loop1MBB);
5259 BB->addSuccessor(exitMBB);
5265 MI->eraseFromParent(); // The instruction is gone now.
5271 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5272 unsigned Size, unsigned BinOpcode) const {
5273 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5274 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5276 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5277 MachineFunction *MF = BB->getParent();
5278 MachineFunction::iterator It = BB;
5281 unsigned dest = MI->getOperand(0).getReg();
5282 unsigned ptr = MI->getOperand(1).getReg();
5283 unsigned incr = MI->getOperand(2).getReg();
5284 DebugLoc dl = MI->getDebugLoc();
5285 bool isThumb2 = Subtarget->isThumb2();
5287 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5289 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5290 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5293 unsigned ldrOpc, strOpc;
5295 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5297 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5298 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5301 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5302 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5305 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5306 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5310 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5311 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5312 MF->insert(It, loopMBB);
5313 MF->insert(It, exitMBB);
5315 // Transfer the remainder of BB and its successor edges to exitMBB.
5316 exitMBB->splice(exitMBB->begin(), BB,
5317 llvm::next(MachineBasicBlock::iterator(MI)),
5319 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5321 const TargetRegisterClass *TRC =
5322 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5323 unsigned scratch = MRI.createVirtualRegister(TRC);
5324 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5328 // fallthrough --> loopMBB
5329 BB->addSuccessor(loopMBB);
5333 // <binop> scratch2, dest, incr
5334 // strex scratch, scratch2, ptr
5337 // fallthrough --> exitMBB
5339 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5340 if (ldrOpc == ARM::t2LDREX)
5342 AddDefaultPred(MIB);
5344 // operand order needs to go the other way for NAND
5345 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5346 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5347 addReg(incr).addReg(dest)).addReg(0);
5349 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5350 addReg(dest).addReg(incr)).addReg(0);
5353 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5354 if (strOpc == ARM::t2STREX)
5356 AddDefaultPred(MIB);
5357 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5358 .addReg(scratch).addImm(0));
5359 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5360 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5362 BB->addSuccessor(loopMBB);
5363 BB->addSuccessor(exitMBB);
5369 MI->eraseFromParent(); // The instruction is gone now.
5375 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5376 MachineBasicBlock *BB,
5379 ARMCC::CondCodes Cond) const {
5380 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5382 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5383 MachineFunction *MF = BB->getParent();
5384 MachineFunction::iterator It = BB;
5387 unsigned dest = MI->getOperand(0).getReg();
5388 unsigned ptr = MI->getOperand(1).getReg();
5389 unsigned incr = MI->getOperand(2).getReg();
5390 unsigned oldval = dest;
5391 DebugLoc dl = MI->getDebugLoc();
5392 bool isThumb2 = Subtarget->isThumb2();
5394 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5396 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5397 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5400 unsigned ldrOpc, strOpc, extendOpc;
5402 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5404 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5405 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5406 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5409 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5410 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5411 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5414 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5415 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5420 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5421 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5422 MF->insert(It, loopMBB);
5423 MF->insert(It, exitMBB);
5425 // Transfer the remainder of BB and its successor edges to exitMBB.
5426 exitMBB->splice(exitMBB->begin(), BB,
5427 llvm::next(MachineBasicBlock::iterator(MI)),
5429 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5431 const TargetRegisterClass *TRC =
5432 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5433 unsigned scratch = MRI.createVirtualRegister(TRC);
5434 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5438 // fallthrough --> loopMBB
5439 BB->addSuccessor(loopMBB);
5443 // (sign extend dest, if required)
5445 // cmov.cond scratch2, dest, incr
5446 // strex scratch, scratch2, ptr
5449 // fallthrough --> exitMBB
5451 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5452 if (ldrOpc == ARM::t2LDREX)
5454 AddDefaultPred(MIB);
5456 // Sign extend the value, if necessary.
5457 if (signExtend && extendOpc) {
5458 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5459 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5464 // Build compare and cmov instructions.
5465 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5466 .addReg(oldval).addReg(incr));
5467 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5468 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5470 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5471 if (strOpc == ARM::t2STREX)
5473 AddDefaultPred(MIB);
5474 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5475 .addReg(scratch).addImm(0));
5476 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5477 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5479 BB->addSuccessor(loopMBB);
5480 BB->addSuccessor(exitMBB);
5486 MI->eraseFromParent(); // The instruction is gone now.
5492 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5493 unsigned Op1, unsigned Op2,
5494 bool NeedsCarry, bool IsCmpxchg) const {
5495 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5496 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5498 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5499 MachineFunction *MF = BB->getParent();
5500 MachineFunction::iterator It = BB;
5503 unsigned destlo = MI->getOperand(0).getReg();
5504 unsigned desthi = MI->getOperand(1).getReg();
5505 unsigned ptr = MI->getOperand(2).getReg();
5506 unsigned vallo = MI->getOperand(3).getReg();
5507 unsigned valhi = MI->getOperand(4).getReg();
5508 DebugLoc dl = MI->getDebugLoc();
5509 bool isThumb2 = Subtarget->isThumb2();
5511 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5513 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5514 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5515 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5518 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5519 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5521 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5522 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5524 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5525 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5527 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5528 MF->insert(It, loopMBB);
5530 MF->insert(It, contBB);
5531 MF->insert(It, cont2BB);
5533 MF->insert(It, exitMBB);
5535 // Transfer the remainder of BB and its successor edges to exitMBB.
5536 exitMBB->splice(exitMBB->begin(), BB,
5537 llvm::next(MachineBasicBlock::iterator(MI)),
5539 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5541 const TargetRegisterClass *TRC =
5542 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5543 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5547 // fallthrough --> loopMBB
5548 BB->addSuccessor(loopMBB);
5551 // ldrexd r2, r3, ptr
5552 // <binopa> r0, r2, incr
5553 // <binopb> r1, r3, incr
5554 // strexd storesuccess, r0, r1, ptr
5555 // cmp storesuccess, #0
5557 // fallthrough --> exitMBB
5559 // Note that the registers are explicitly specified because there is not any
5560 // way to force the register allocator to allocate a register pair.
5562 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5563 // need to properly enforce the restriction that the two output registers
5564 // for ldrexd must be different.
5567 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5568 .addReg(ARM::R2, RegState::Define)
5569 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5570 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5571 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5572 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5576 for (unsigned i = 0; i < 2; i++) {
5577 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5579 .addReg(i == 0 ? destlo : desthi)
5580 .addReg(i == 0 ? vallo : valhi));
5581 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5582 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5583 BB->addSuccessor(exitMBB);
5584 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5585 BB = (i == 0 ? contBB : cont2BB);
5588 // Copy to physregs for strexd
5589 unsigned setlo = MI->getOperand(5).getReg();
5590 unsigned sethi = MI->getOperand(6).getReg();
5591 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5592 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5594 // Perform binary operation
5595 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5596 .addReg(destlo).addReg(vallo))
5597 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5598 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5599 .addReg(desthi).addReg(valhi)).addReg(0);
5601 // Copy to physregs for strexd
5602 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5603 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5607 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5608 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5610 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5611 .addReg(storesuccess).addImm(0));
5612 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5613 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5615 BB->addSuccessor(loopMBB);
5616 BB->addSuccessor(exitMBB);
5622 MI->eraseFromParent(); // The instruction is gone now.
5627 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5628 /// registers the function context.
5629 void ARMTargetLowering::
5630 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5631 MachineBasicBlock *DispatchBB, int FI) const {
5632 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5633 DebugLoc dl = MI->getDebugLoc();
5634 MachineFunction *MF = MBB->getParent();
5635 MachineRegisterInfo *MRI = &MF->getRegInfo();
5636 MachineConstantPool *MCP = MF->getConstantPool();
5637 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5638 const Function *F = MF->getFunction();
5640 bool isThumb = Subtarget->isThumb();
5641 bool isThumb2 = Subtarget->isThumb2();
5643 unsigned PCLabelId = AFI->createPICLabelUId();
5644 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5645 ARMConstantPoolValue *CPV =
5646 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5647 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5649 const TargetRegisterClass *TRC =
5650 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5652 // Grab constant pool and fixed stack memory operands.
5653 MachineMemOperand *CPMMO =
5654 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5655 MachineMemOperand::MOLoad, 4, 4);
5657 MachineMemOperand *FIMMOSt =
5658 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5659 MachineMemOperand::MOStore, 4, 4);
5661 // Load the address of the dispatch MBB into the jump buffer.
5663 // Incoming value: jbuf
5664 // ldr.n r5, LCPI1_1
5667 // str r5, [$jbuf, #+4] ; &jbuf[1]
5668 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5669 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5670 .addConstantPoolIndex(CPI)
5671 .addMemOperand(CPMMO));
5672 // Set the low bit because of thumb mode.
5673 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5675 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5676 .addReg(NewVReg1, RegState::Kill)
5678 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5679 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5680 .addReg(NewVReg2, RegState::Kill)
5682 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5683 .addReg(NewVReg3, RegState::Kill)
5685 .addImm(36) // &jbuf[1] :: pc
5686 .addMemOperand(FIMMOSt));
5687 } else if (isThumb) {
5688 // Incoming value: jbuf
5689 // ldr.n r1, LCPI1_4
5693 // add r2, $jbuf, #+4 ; &jbuf[1]
5695 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5696 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5697 .addConstantPoolIndex(CPI)
5698 .addMemOperand(CPMMO));
5699 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5700 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5701 .addReg(NewVReg1, RegState::Kill)
5703 // Set the low bit because of thumb mode.
5704 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5705 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5706 .addReg(ARM::CPSR, RegState::Define)
5708 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5709 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5710 .addReg(ARM::CPSR, RegState::Define)
5711 .addReg(NewVReg2, RegState::Kill)
5712 .addReg(NewVReg3, RegState::Kill));
5713 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5714 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5716 .addImm(36)); // &jbuf[1] :: pc
5717 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5718 .addReg(NewVReg4, RegState::Kill)
5719 .addReg(NewVReg5, RegState::Kill)
5721 .addMemOperand(FIMMOSt));
5723 // Incoming value: jbuf
5726 // str r1, [$jbuf, #+4] ; &jbuf[1]
5727 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5728 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5729 .addConstantPoolIndex(CPI)
5731 .addMemOperand(CPMMO));
5732 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5733 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5734 .addReg(NewVReg1, RegState::Kill)
5735 .addImm(PCLabelId));
5736 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5737 .addReg(NewVReg2, RegState::Kill)
5739 .addImm(36) // &jbuf[1] :: pc
5740 .addMemOperand(FIMMOSt));
5744 MachineBasicBlock *ARMTargetLowering::
5745 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5746 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5747 DebugLoc dl = MI->getDebugLoc();
5748 MachineFunction *MF = MBB->getParent();
5749 MachineRegisterInfo *MRI = &MF->getRegInfo();
5750 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5751 MachineFrameInfo *MFI = MF->getFrameInfo();
5752 int FI = MFI->getFunctionContextIndex();
5754 const TargetRegisterClass *TRC =
5755 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5757 // Get a mapping of the call site numbers to all of the landing pads they're
5759 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5760 unsigned MaxCSNum = 0;
5761 MachineModuleInfo &MMI = MF->getMMI();
5762 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5763 if (!BB->isLandingPad()) continue;
5765 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5767 for (MachineBasicBlock::iterator
5768 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5769 if (!II->isEHLabel()) continue;
5771 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5772 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5774 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5775 for (SmallVectorImpl<unsigned>::iterator
5776 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5777 CSI != CSE; ++CSI) {
5778 CallSiteNumToLPad[*CSI].push_back(BB);
5779 MaxCSNum = std::max(MaxCSNum, *CSI);
5785 // Get an ordered list of the machine basic blocks for the jump table.
5786 std::vector<MachineBasicBlock*> LPadList;
5787 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5788 LPadList.reserve(CallSiteNumToLPad.size());
5789 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5790 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5791 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5792 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5793 LPadList.push_back(*II);
5794 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5798 assert(!LPadList.empty() &&
5799 "No landing pad destinations for the dispatch jump table!");
5801 // Create the jump table and associated information.
5802 MachineJumpTableInfo *JTI =
5803 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5804 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5805 unsigned UId = AFI->createJumpTableUId();
5807 // Create the MBBs for the dispatch code.
5809 // Shove the dispatch's address into the return slot in the function context.
5810 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5811 DispatchBB->setIsLandingPad();
5813 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5814 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5815 DispatchBB->addSuccessor(TrapBB);
5817 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5818 DispatchBB->addSuccessor(DispContBB);
5821 MF->insert(MF->end(), DispatchBB);
5822 MF->insert(MF->end(), DispContBB);
5823 MF->insert(MF->end(), TrapBB);
5825 // Insert code into the entry block that creates and registers the function
5827 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5829 MachineMemOperand *FIMMOLd =
5830 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5831 MachineMemOperand::MOLoad |
5832 MachineMemOperand::MOVolatile, 4, 4);
5834 if (AFI->isThumb1OnlyFunction())
5835 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5836 else if (!Subtarget->hasVFP2())
5837 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5839 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5841 unsigned NumLPads = LPadList.size();
5842 if (Subtarget->isThumb2()) {
5843 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5844 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5847 .addMemOperand(FIMMOLd));
5849 if (NumLPads < 256) {
5850 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5852 .addImm(LPadList.size()));
5854 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5855 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5856 .addImm(NumLPads & 0xFFFF));
5858 unsigned VReg2 = VReg1;
5859 if ((NumLPads & 0xFFFF0000) != 0) {
5860 VReg2 = MRI->createVirtualRegister(TRC);
5861 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5863 .addImm(NumLPads >> 16));
5866 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5871 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5876 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5877 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5878 .addJumpTableIndex(MJTI)
5881 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5884 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5885 .addReg(NewVReg3, RegState::Kill)
5887 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5889 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5890 .addReg(NewVReg4, RegState::Kill)
5892 .addJumpTableIndex(MJTI)
5894 } else if (Subtarget->isThumb()) {
5895 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5896 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5899 .addMemOperand(FIMMOLd));
5901 if (NumLPads < 256) {
5902 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5906 MachineConstantPool *ConstantPool = MF->getConstantPool();
5907 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5908 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5910 // MachineConstantPool wants an explicit alignment.
5911 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5913 Align = getTargetData()->getTypeAllocSize(C->getType());
5914 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5916 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5917 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5918 .addReg(VReg1, RegState::Define)
5919 .addConstantPoolIndex(Idx));
5920 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5925 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5930 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5931 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5932 .addReg(ARM::CPSR, RegState::Define)
5936 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5937 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5938 .addJumpTableIndex(MJTI)
5941 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5942 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5943 .addReg(ARM::CPSR, RegState::Define)
5944 .addReg(NewVReg2, RegState::Kill)
5947 MachineMemOperand *JTMMOLd =
5948 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5949 MachineMemOperand::MOLoad, 4, 4);
5951 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5952 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5953 .addReg(NewVReg4, RegState::Kill)
5955 .addMemOperand(JTMMOLd));
5957 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5958 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5959 .addReg(ARM::CPSR, RegState::Define)
5960 .addReg(NewVReg5, RegState::Kill)
5963 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5964 .addReg(NewVReg6, RegState::Kill)
5965 .addJumpTableIndex(MJTI)
5968 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5969 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5972 .addMemOperand(FIMMOLd));
5974 if (NumLPads < 256) {
5975 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5978 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5979 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5980 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5981 .addImm(NumLPads & 0xFFFF));
5983 unsigned VReg2 = VReg1;
5984 if ((NumLPads & 0xFFFF0000) != 0) {
5985 VReg2 = MRI->createVirtualRegister(TRC);
5986 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5988 .addImm(NumLPads >> 16));
5991 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5995 MachineConstantPool *ConstantPool = MF->getConstantPool();
5996 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5997 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5999 // MachineConstantPool wants an explicit alignment.
6000 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6002 Align = getTargetData()->getTypeAllocSize(C->getType());
6003 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6005 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6006 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6007 .addReg(VReg1, RegState::Define)
6008 .addConstantPoolIndex(Idx)
6010 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6012 .addReg(VReg1, RegState::Kill));
6015 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6020 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6022 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6024 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6025 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6026 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6027 .addJumpTableIndex(MJTI)
6030 MachineMemOperand *JTMMOLd =
6031 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6032 MachineMemOperand::MOLoad, 4, 4);
6033 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6035 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6036 .addReg(NewVReg3, RegState::Kill)
6039 .addMemOperand(JTMMOLd));
6041 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6042 .addReg(NewVReg5, RegState::Kill)
6044 .addJumpTableIndex(MJTI)
6048 // Add the jump table entries as successors to the MBB.
6049 MachineBasicBlock *PrevMBB = 0;
6050 for (std::vector<MachineBasicBlock*>::iterator
6051 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6052 MachineBasicBlock *CurMBB = *I;
6053 if (PrevMBB != CurMBB)
6054 DispContBB->addSuccessor(CurMBB);
6058 // N.B. the order the invoke BBs are processed in doesn't matter here.
6059 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6060 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6061 const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
6062 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6063 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6064 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6065 MachineBasicBlock *BB = *I;
6067 // Remove the landing pad successor from the invoke block and replace it
6068 // with the new dispatch block.
6069 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6071 while (!Successors.empty()) {
6072 MachineBasicBlock *SMBB = Successors.pop_back_val();
6073 if (SMBB->isLandingPad()) {
6074 BB->removeSuccessor(SMBB);
6075 MBBLPads.push_back(SMBB);
6079 BB->addSuccessor(DispatchBB);
6081 // Find the invoke call and mark all of the callee-saved registers as
6082 // 'implicit defined' so that they're spilled. This prevents code from
6083 // moving instructions to before the EH block, where they will never be
6085 for (MachineBasicBlock::reverse_iterator
6086 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6087 if (!II->isCall()) continue;
6089 DenseMap<unsigned, bool> DefRegs;
6090 for (MachineInstr::mop_iterator
6091 OI = II->operands_begin(), OE = II->operands_end();
6093 if (!OI->isReg()) continue;
6094 DefRegs[OI->getReg()] = true;
6097 MachineInstrBuilder MIB(&*II);
6099 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6100 unsigned Reg = SavedRegs[i];
6101 if (Subtarget->isThumb2() &&
6102 !ARM::tGPRRegisterClass->contains(Reg) &&
6103 !ARM::hGPRRegisterClass->contains(Reg))
6105 else if (Subtarget->isThumb1Only() &&
6106 !ARM::tGPRRegisterClass->contains(Reg))
6108 else if (!Subtarget->isThumb() &&
6109 !ARM::GPRRegisterClass->contains(Reg))
6112 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6119 // Mark all former landing pads as non-landing pads. The dispatch is the only
6121 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6122 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6123 (*I)->setIsLandingPad(false);
6125 // The instruction is gone now.
6126 MI->eraseFromParent();
6132 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6133 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6134 E = MBB->succ_end(); I != E; ++I)
6137 llvm_unreachable("Expecting a BB with two successors!");
6141 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6142 MachineBasicBlock *BB) const {
6143 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6144 DebugLoc dl = MI->getDebugLoc();
6145 bool isThumb2 = Subtarget->isThumb2();
6146 switch (MI->getOpcode()) {
6149 llvm_unreachable("Unexpected instr type to insert");
6151 // The Thumb2 pre-indexed stores have the same MI operands, they just
6152 // define them differently in the .td files from the isel patterns, so
6153 // they need pseudos.
6154 case ARM::t2STR_preidx:
6155 MI->setDesc(TII->get(ARM::t2STR_PRE));
6157 case ARM::t2STRB_preidx:
6158 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6160 case ARM::t2STRH_preidx:
6161 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6164 case ARM::STRi_preidx:
6165 case ARM::STRBi_preidx: {
6166 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6167 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6168 // Decode the offset.
6169 unsigned Offset = MI->getOperand(4).getImm();
6170 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6171 Offset = ARM_AM::getAM2Offset(Offset);
6175 MachineMemOperand *MMO = *MI->memoperands_begin();
6176 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6177 .addOperand(MI->getOperand(0)) // Rn_wb
6178 .addOperand(MI->getOperand(1)) // Rt
6179 .addOperand(MI->getOperand(2)) // Rn
6180 .addImm(Offset) // offset (skip GPR==zero_reg)
6181 .addOperand(MI->getOperand(5)) // pred
6182 .addOperand(MI->getOperand(6))
6183 .addMemOperand(MMO);
6184 MI->eraseFromParent();
6187 case ARM::STRr_preidx:
6188 case ARM::STRBr_preidx:
6189 case ARM::STRH_preidx: {
6191 switch (MI->getOpcode()) {
6192 default: llvm_unreachable("unexpected opcode!");
6193 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6194 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6195 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6197 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6198 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6199 MIB.addOperand(MI->getOperand(i));
6200 MI->eraseFromParent();
6203 case ARM::ATOMIC_LOAD_ADD_I8:
6204 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6205 case ARM::ATOMIC_LOAD_ADD_I16:
6206 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6207 case ARM::ATOMIC_LOAD_ADD_I32:
6208 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6210 case ARM::ATOMIC_LOAD_AND_I8:
6211 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6212 case ARM::ATOMIC_LOAD_AND_I16:
6213 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6214 case ARM::ATOMIC_LOAD_AND_I32:
6215 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6217 case ARM::ATOMIC_LOAD_OR_I8:
6218 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6219 case ARM::ATOMIC_LOAD_OR_I16:
6220 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6221 case ARM::ATOMIC_LOAD_OR_I32:
6222 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6224 case ARM::ATOMIC_LOAD_XOR_I8:
6225 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6226 case ARM::ATOMIC_LOAD_XOR_I16:
6227 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6228 case ARM::ATOMIC_LOAD_XOR_I32:
6229 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6231 case ARM::ATOMIC_LOAD_NAND_I8:
6232 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6233 case ARM::ATOMIC_LOAD_NAND_I16:
6234 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6235 case ARM::ATOMIC_LOAD_NAND_I32:
6236 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6238 case ARM::ATOMIC_LOAD_SUB_I8:
6239 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6240 case ARM::ATOMIC_LOAD_SUB_I16:
6241 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6242 case ARM::ATOMIC_LOAD_SUB_I32:
6243 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6245 case ARM::ATOMIC_LOAD_MIN_I8:
6246 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6247 case ARM::ATOMIC_LOAD_MIN_I16:
6248 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6249 case ARM::ATOMIC_LOAD_MIN_I32:
6250 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6252 case ARM::ATOMIC_LOAD_MAX_I8:
6253 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6254 case ARM::ATOMIC_LOAD_MAX_I16:
6255 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6256 case ARM::ATOMIC_LOAD_MAX_I32:
6257 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6259 case ARM::ATOMIC_LOAD_UMIN_I8:
6260 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6261 case ARM::ATOMIC_LOAD_UMIN_I16:
6262 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6263 case ARM::ATOMIC_LOAD_UMIN_I32:
6264 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6266 case ARM::ATOMIC_LOAD_UMAX_I8:
6267 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6268 case ARM::ATOMIC_LOAD_UMAX_I16:
6269 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6270 case ARM::ATOMIC_LOAD_UMAX_I32:
6271 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6273 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6274 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6275 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6277 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6278 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6279 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6282 case ARM::ATOMADD6432:
6283 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6284 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6285 /*NeedsCarry*/ true);
6286 case ARM::ATOMSUB6432:
6287 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6288 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6289 /*NeedsCarry*/ true);
6290 case ARM::ATOMOR6432:
6291 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6292 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6293 case ARM::ATOMXOR6432:
6294 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6295 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6296 case ARM::ATOMAND6432:
6297 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6298 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6299 case ARM::ATOMSWAP6432:
6300 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6301 case ARM::ATOMCMPXCHG6432:
6302 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6303 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6304 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6306 case ARM::tMOVCCr_pseudo: {
6307 // To "insert" a SELECT_CC instruction, we actually have to insert the
6308 // diamond control-flow pattern. The incoming instruction knows the
6309 // destination vreg to set, the condition code register to branch on, the
6310 // true/false values to select between, and a branch opcode to use.
6311 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6312 MachineFunction::iterator It = BB;
6318 // cmpTY ccX, r1, r2
6320 // fallthrough --> copy0MBB
6321 MachineBasicBlock *thisMBB = BB;
6322 MachineFunction *F = BB->getParent();
6323 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6324 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6325 F->insert(It, copy0MBB);
6326 F->insert(It, sinkMBB);
6328 // Transfer the remainder of BB and its successor edges to sinkMBB.
6329 sinkMBB->splice(sinkMBB->begin(), BB,
6330 llvm::next(MachineBasicBlock::iterator(MI)),
6332 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6334 BB->addSuccessor(copy0MBB);
6335 BB->addSuccessor(sinkMBB);
6337 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6338 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6341 // %FalseValue = ...
6342 // # fallthrough to sinkMBB
6345 // Update machine-CFG edges
6346 BB->addSuccessor(sinkMBB);
6349 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6352 BuildMI(*BB, BB->begin(), dl,
6353 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6354 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6355 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6357 MI->eraseFromParent(); // The pseudo instruction is gone now.
6362 case ARM::BCCZi64: {
6363 // If there is an unconditional branch to the other successor, remove it.
6364 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6366 // Compare both parts that make up the double comparison separately for
6368 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6370 unsigned LHS1 = MI->getOperand(1).getReg();
6371 unsigned LHS2 = MI->getOperand(2).getReg();
6373 AddDefaultPred(BuildMI(BB, dl,
6374 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6375 .addReg(LHS1).addImm(0));
6376 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6377 .addReg(LHS2).addImm(0)
6378 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6380 unsigned RHS1 = MI->getOperand(3).getReg();
6381 unsigned RHS2 = MI->getOperand(4).getReg();
6382 AddDefaultPred(BuildMI(BB, dl,
6383 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6384 .addReg(LHS1).addReg(RHS1));
6385 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6386 .addReg(LHS2).addReg(RHS2)
6387 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6390 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6391 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6392 if (MI->getOperand(0).getImm() == ARMCC::NE)
6393 std::swap(destMBB, exitMBB);
6395 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6396 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6398 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6400 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6402 MI->eraseFromParent(); // The pseudo instruction is gone now.
6406 case ARM::Int_eh_sjlj_setjmp:
6407 case ARM::Int_eh_sjlj_setjmp_nofp:
6408 case ARM::tInt_eh_sjlj_setjmp:
6409 case ARM::t2Int_eh_sjlj_setjmp:
6410 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6411 EmitSjLjDispatchBlock(MI, BB);
6416 // To insert an ABS instruction, we have to insert the
6417 // diamond control-flow pattern. The incoming instruction knows the
6418 // source vreg to test against 0, the destination vreg to set,
6419 // the condition code register to branch on, the
6420 // true/false values to select between, and a branch opcode to use.
6425 // BCC (branch to SinkBB if V0 >= 0)
6426 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6427 // SinkBB: V1 = PHI(V2, V3)
6428 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6429 MachineFunction::iterator BBI = BB;
6431 MachineFunction *Fn = BB->getParent();
6432 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6433 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6434 Fn->insert(BBI, RSBBB);
6435 Fn->insert(BBI, SinkBB);
6437 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6438 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6439 bool isThumb2 = Subtarget->isThumb2();
6440 MachineRegisterInfo &MRI = Fn->getRegInfo();
6441 // In Thumb mode S must not be specified if source register is the SP or
6442 // PC and if destination register is the SP, so restrict register class
6443 unsigned NewMovDstReg = MRI.createVirtualRegister(
6444 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6445 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6446 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6448 // Transfer the remainder of BB and its successor edges to sinkMBB.
6449 SinkBB->splice(SinkBB->begin(), BB,
6450 llvm::next(MachineBasicBlock::iterator(MI)),
6452 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6454 BB->addSuccessor(RSBBB);
6455 BB->addSuccessor(SinkBB);
6457 // fall through to SinkMBB
6458 RSBBB->addSuccessor(SinkBB);
6460 // insert a movs at the end of BB
6461 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6463 .addReg(ABSSrcReg, RegState::Kill)
6464 .addImm((unsigned)ARMCC::AL).addReg(0)
6465 .addReg(ARM::CPSR, RegState::Define);
6467 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6469 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6470 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6472 // insert rsbri in RSBBB
6473 // Note: BCC and rsbri will be converted into predicated rsbmi
6474 // by if-conversion pass
6475 BuildMI(*RSBBB, RSBBB->begin(), dl,
6476 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6477 .addReg(NewMovDstReg, RegState::Kill)
6478 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6480 // insert PHI in SinkBB,
6481 // reuse ABSDstReg to not change uses of ABS instruction
6482 BuildMI(*SinkBB, SinkBB->begin(), dl,
6483 TII->get(ARM::PHI), ABSDstReg)
6484 .addReg(NewRsbDstReg).addMBB(RSBBB)
6485 .addReg(NewMovDstReg).addMBB(BB);
6487 // remove ABS instruction
6488 MI->eraseFromParent();
6490 // return last added BB
6496 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6497 SDNode *Node) const {
6498 if (!MI->hasPostISelHook()) {
6499 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6500 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6504 const MCInstrDesc *MCID = &MI->getDesc();
6505 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6506 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6507 // operand is still set to noreg. If needed, set the optional operand's
6508 // register to CPSR, and remove the redundant implicit def.
6510 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6512 // Rename pseudo opcodes.
6513 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6515 const ARMBaseInstrInfo *TII =
6516 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6517 MCID = &TII->get(NewOpc);
6519 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6520 "converted opcode should be the same except for cc_out");
6524 // Add the optional cc_out operand
6525 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6527 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6529 // Any ARM instruction that sets the 's' bit should specify an optional
6530 // "cc_out" operand in the last operand position.
6531 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6532 assert(!NewOpc && "Optional cc_out operand required");
6535 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6536 // since we already have an optional CPSR def.
6537 bool definesCPSR = false;
6538 bool deadCPSR = false;
6539 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6541 const MachineOperand &MO = MI->getOperand(i);
6542 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6546 MI->RemoveOperand(i);
6551 assert(!NewOpc && "Optional cc_out operand required");
6554 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6556 assert(!MI->getOperand(ccOutIdx).getReg() &&
6557 "expect uninitialized optional cc_out operand");
6561 // If this instruction was defined with an optional CPSR def and its dag node
6562 // had a live implicit CPSR def, then activate the optional CPSR def.
6563 MachineOperand &MO = MI->getOperand(ccOutIdx);
6564 MO.setReg(ARM::CPSR);
6568 //===----------------------------------------------------------------------===//
6569 // ARM Optimization Hooks
6570 //===----------------------------------------------------------------------===//
6573 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6574 TargetLowering::DAGCombinerInfo &DCI) {
6575 SelectionDAG &DAG = DCI.DAG;
6576 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6577 EVT VT = N->getValueType(0);
6578 unsigned Opc = N->getOpcode();
6579 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6580 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6581 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6582 ISD::CondCode CC = ISD::SETCC_INVALID;
6585 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6587 SDValue CCOp = Slct.getOperand(0);
6588 if (CCOp.getOpcode() == ISD::SETCC)
6589 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6592 bool DoXform = false;
6594 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6597 if (LHS.getOpcode() == ISD::Constant &&
6598 cast<ConstantSDNode>(LHS)->isNullValue()) {
6600 } else if (CC != ISD::SETCC_INVALID &&
6601 RHS.getOpcode() == ISD::Constant &&
6602 cast<ConstantSDNode>(RHS)->isNullValue()) {
6603 std::swap(LHS, RHS);
6604 SDValue Op0 = Slct.getOperand(0);
6605 EVT OpVT = isSlctCC ? Op0.getValueType() :
6606 Op0.getOperand(0).getValueType();
6607 bool isInt = OpVT.isInteger();
6608 CC = ISD::getSetCCInverse(CC, isInt);
6610 if (!TLI.isCondCodeLegal(CC, OpVT))
6611 return SDValue(); // Inverse operator isn't legal.
6618 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6620 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6621 Slct.getOperand(0), Slct.getOperand(1), CC);
6622 SDValue CCOp = Slct.getOperand(0);
6624 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6625 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6626 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6627 CCOp, OtherOp, Result);
6632 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6633 // (only after legalization).
6634 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6635 TargetLowering::DAGCombinerInfo &DCI,
6636 const ARMSubtarget *Subtarget) {
6638 // Only perform optimization if after legalize, and if NEON is available. We
6639 // also expected both operands to be BUILD_VECTORs.
6640 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6641 || N0.getOpcode() != ISD::BUILD_VECTOR
6642 || N1.getOpcode() != ISD::BUILD_VECTOR)
6645 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6646 EVT VT = N->getValueType(0);
6647 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6650 // Check that the vector operands are of the right form.
6651 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6652 // operands, where N is the size of the formed vector.
6653 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6654 // index such that we have a pair wise add pattern.
6656 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6657 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6659 SDValue Vec = N0->getOperand(0)->getOperand(0);
6660 SDNode *V = Vec.getNode();
6661 unsigned nextIndex = 0;
6663 // For each operands to the ADD which are BUILD_VECTORs,
6664 // check to see if each of their operands are an EXTRACT_VECTOR with
6665 // the same vector and appropriate index.
6666 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6667 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6668 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6670 SDValue ExtVec0 = N0->getOperand(i);
6671 SDValue ExtVec1 = N1->getOperand(i);
6673 // First operand is the vector, verify its the same.
6674 if (V != ExtVec0->getOperand(0).getNode() ||
6675 V != ExtVec1->getOperand(0).getNode())
6678 // Second is the constant, verify its correct.
6679 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6680 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6682 // For the constant, we want to see all the even or all the odd.
6683 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6684 || C1->getZExtValue() != nextIndex+1)
6693 // Create VPADDL node.
6694 SelectionDAG &DAG = DCI.DAG;
6695 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6697 // Build operand list.
6698 SmallVector<SDValue, 8> Ops;
6699 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6700 TLI.getPointerTy()));
6702 // Input is the vector.
6705 // Get widened type and narrowed type.
6707 unsigned numElem = VT.getVectorNumElements();
6708 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6709 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6710 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6711 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6713 llvm_unreachable("Invalid vector element type for padd optimization.");
6716 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6717 widenType, &Ops[0], Ops.size());
6718 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6721 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6722 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6723 /// called with the default operands, and if that fails, with commuted
6725 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6726 TargetLowering::DAGCombinerInfo &DCI,
6727 const ARMSubtarget *Subtarget){
6729 // Attempt to create vpaddl for this add.
6730 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6731 if (Result.getNode())
6734 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6735 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6736 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6737 if (Result.getNode()) return Result;
6742 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6744 static SDValue PerformADDCombine(SDNode *N,
6745 TargetLowering::DAGCombinerInfo &DCI,
6746 const ARMSubtarget *Subtarget) {
6747 SDValue N0 = N->getOperand(0);
6748 SDValue N1 = N->getOperand(1);
6750 // First try with the default operand order.
6751 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6752 if (Result.getNode())
6755 // If that didn't work, try again with the operands commuted.
6756 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6759 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6761 static SDValue PerformSUBCombine(SDNode *N,
6762 TargetLowering::DAGCombinerInfo &DCI) {
6763 SDValue N0 = N->getOperand(0);
6764 SDValue N1 = N->getOperand(1);
6766 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6767 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6768 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6769 if (Result.getNode()) return Result;
6775 /// PerformVMULCombine
6776 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6777 /// special multiplier accumulator forwarding.
6783 static SDValue PerformVMULCombine(SDNode *N,
6784 TargetLowering::DAGCombinerInfo &DCI,
6785 const ARMSubtarget *Subtarget) {
6786 if (!Subtarget->hasVMLxForwarding())
6789 SelectionDAG &DAG = DCI.DAG;
6790 SDValue N0 = N->getOperand(0);
6791 SDValue N1 = N->getOperand(1);
6792 unsigned Opcode = N0.getOpcode();
6793 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6794 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6795 Opcode = N1.getOpcode();
6796 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6797 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6802 EVT VT = N->getValueType(0);
6803 DebugLoc DL = N->getDebugLoc();
6804 SDValue N00 = N0->getOperand(0);
6805 SDValue N01 = N0->getOperand(1);
6806 return DAG.getNode(Opcode, DL, VT,
6807 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6808 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6811 static SDValue PerformMULCombine(SDNode *N,
6812 TargetLowering::DAGCombinerInfo &DCI,
6813 const ARMSubtarget *Subtarget) {
6814 SelectionDAG &DAG = DCI.DAG;
6816 if (Subtarget->isThumb1Only())
6819 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6822 EVT VT = N->getValueType(0);
6823 if (VT.is64BitVector() || VT.is128BitVector())
6824 return PerformVMULCombine(N, DCI, Subtarget);
6828 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6832 uint64_t MulAmt = C->getZExtValue();
6833 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6834 ShiftAmt = ShiftAmt & (32 - 1);
6835 SDValue V = N->getOperand(0);
6836 DebugLoc DL = N->getDebugLoc();
6839 MulAmt >>= ShiftAmt;
6840 if (isPowerOf2_32(MulAmt - 1)) {
6841 // (mul x, 2^N + 1) => (add (shl x, N), x)
6842 Res = DAG.getNode(ISD::ADD, DL, VT,
6843 V, DAG.getNode(ISD::SHL, DL, VT,
6844 V, DAG.getConstant(Log2_32(MulAmt-1),
6846 } else if (isPowerOf2_32(MulAmt + 1)) {
6847 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6848 Res = DAG.getNode(ISD::SUB, DL, VT,
6849 DAG.getNode(ISD::SHL, DL, VT,
6850 V, DAG.getConstant(Log2_32(MulAmt+1),
6857 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6858 DAG.getConstant(ShiftAmt, MVT::i32));
6860 // Do not add new nodes to DAG combiner worklist.
6861 DCI.CombineTo(N, Res, false);
6865 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6866 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6869 SDValue FalseVal = N.getOperand(0);
6870 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6874 return C->isAllOnesValue();
6875 return C->isNullValue();
6878 /// formConditionalOp - Combine an operation with a conditional move operand
6879 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
6880 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
6881 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
6883 SDValue N0 = N->getOperand(0);
6884 SDValue N1 = N->getOperand(1);
6886 bool isAND = N->getOpcode() == ISD::AND;
6887 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
6888 if (!isCand && Commutable) {
6889 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
6897 switch (N->getOpcode()) {
6898 default: llvm_unreachable("Unexpected node");
6899 case ISD::AND: Opc = ARMISD::CAND; break;
6900 case ISD::OR: Opc = ARMISD::COR; break;
6901 case ISD::XOR: Opc = ARMISD::CXOR; break;
6903 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
6904 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
6908 static SDValue PerformANDCombine(SDNode *N,
6909 TargetLowering::DAGCombinerInfo &DCI,
6910 const ARMSubtarget *Subtarget) {
6912 // Attempt to use immediate-form VBIC
6913 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6914 DebugLoc dl = N->getDebugLoc();
6915 EVT VT = N->getValueType(0);
6916 SelectionDAG &DAG = DCI.DAG;
6918 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6921 APInt SplatBits, SplatUndef;
6922 unsigned SplatBitSize;
6925 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6926 if (SplatBitSize <= 64) {
6928 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6929 SplatUndef.getZExtValue(), SplatBitSize,
6930 DAG, VbicVT, VT.is128BitVector(),
6932 if (Val.getNode()) {
6934 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6935 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6936 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6941 if (!Subtarget->isThumb1Only()) {
6942 // (and x, (cmov -1, y, cond)) => (and.cond x, y)
6943 SDValue CAND = formConditionalOp(N, DAG, true);
6951 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6952 static SDValue PerformORCombine(SDNode *N,
6953 TargetLowering::DAGCombinerInfo &DCI,
6954 const ARMSubtarget *Subtarget) {
6955 // Attempt to use immediate-form VORR
6956 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6957 DebugLoc dl = N->getDebugLoc();
6958 EVT VT = N->getValueType(0);
6959 SelectionDAG &DAG = DCI.DAG;
6961 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6964 APInt SplatBits, SplatUndef;
6965 unsigned SplatBitSize;
6967 if (BVN && Subtarget->hasNEON() &&
6968 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6969 if (SplatBitSize <= 64) {
6971 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6972 SplatUndef.getZExtValue(), SplatBitSize,
6973 DAG, VorrVT, VT.is128BitVector(),
6975 if (Val.getNode()) {
6977 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6978 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6979 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6984 if (!Subtarget->isThumb1Only()) {
6985 // (or x, (cmov 0, y, cond)) => (or.cond x, y)
6986 SDValue COR = formConditionalOp(N, DAG, true);
6991 SDValue N0 = N->getOperand(0);
6992 if (N0.getOpcode() != ISD::AND)
6994 SDValue N1 = N->getOperand(1);
6996 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6997 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6998 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7000 unsigned SplatBitSize;
7003 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7005 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7006 HasAnyUndefs) && !HasAnyUndefs) {
7007 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7009 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7010 HasAnyUndefs) && !HasAnyUndefs &&
7011 SplatBits0 == ~SplatBits1) {
7012 // Canonicalize the vector type to make instruction selection simpler.
7013 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7014 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7015 N0->getOperand(1), N0->getOperand(0),
7017 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7022 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7025 // BFI is only available on V6T2+
7026 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7029 DebugLoc DL = N->getDebugLoc();
7030 // 1) or (and A, mask), val => ARMbfi A, val, mask
7031 // iff (val & mask) == val
7033 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7034 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7035 // && mask == ~mask2
7036 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7037 // && ~mask == mask2
7038 // (i.e., copy a bitfield value into another bitfield of the same width)
7043 SDValue N00 = N0.getOperand(0);
7045 // The value and the mask need to be constants so we can verify this is
7046 // actually a bitfield set. If the mask is 0xffff, we can do better
7047 // via a movt instruction, so don't use BFI in that case.
7048 SDValue MaskOp = N0.getOperand(1);
7049 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7052 unsigned Mask = MaskC->getZExtValue();
7056 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7057 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7059 unsigned Val = N1C->getZExtValue();
7060 if ((Val & ~Mask) != Val)
7063 if (ARM::isBitFieldInvertedMask(Mask)) {
7064 Val >>= CountTrailingZeros_32(~Mask);
7066 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7067 DAG.getConstant(Val, MVT::i32),
7068 DAG.getConstant(Mask, MVT::i32));
7070 // Do not add new nodes to DAG combiner worklist.
7071 DCI.CombineTo(N, Res, false);
7074 } else if (N1.getOpcode() == ISD::AND) {
7075 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7076 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7079 unsigned Mask2 = N11C->getZExtValue();
7081 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7083 if (ARM::isBitFieldInvertedMask(Mask) &&
7085 // The pack halfword instruction works better for masks that fit it,
7086 // so use that when it's available.
7087 if (Subtarget->hasT2ExtractPack() &&
7088 (Mask == 0xffff || Mask == 0xffff0000))
7091 unsigned amt = CountTrailingZeros_32(Mask2);
7092 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7093 DAG.getConstant(amt, MVT::i32));
7094 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7095 DAG.getConstant(Mask, MVT::i32));
7096 // Do not add new nodes to DAG combiner worklist.
7097 DCI.CombineTo(N, Res, false);
7099 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7101 // The pack halfword instruction works better for masks that fit it,
7102 // so use that when it's available.
7103 if (Subtarget->hasT2ExtractPack() &&
7104 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7107 unsigned lsb = CountTrailingZeros_32(Mask);
7108 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7109 DAG.getConstant(lsb, MVT::i32));
7110 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7111 DAG.getConstant(Mask2, MVT::i32));
7112 // Do not add new nodes to DAG combiner worklist.
7113 DCI.CombineTo(N, Res, false);
7118 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7119 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7120 ARM::isBitFieldInvertedMask(~Mask)) {
7121 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7122 // where lsb(mask) == #shamt and masked bits of B are known zero.
7123 SDValue ShAmt = N00.getOperand(1);
7124 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7125 unsigned LSB = CountTrailingZeros_32(Mask);
7129 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7130 DAG.getConstant(~Mask, MVT::i32));
7132 // Do not add new nodes to DAG combiner worklist.
7133 DCI.CombineTo(N, Res, false);
7139 static SDValue PerformXORCombine(SDNode *N,
7140 TargetLowering::DAGCombinerInfo &DCI,
7141 const ARMSubtarget *Subtarget) {
7142 EVT VT = N->getValueType(0);
7143 SelectionDAG &DAG = DCI.DAG;
7145 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7148 if (!Subtarget->isThumb1Only()) {
7149 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7150 SDValue CXOR = formConditionalOp(N, DAG, true);
7158 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7159 /// the bits being cleared by the AND are not demanded by the BFI.
7160 static SDValue PerformBFICombine(SDNode *N,
7161 TargetLowering::DAGCombinerInfo &DCI) {
7162 SDValue N1 = N->getOperand(1);
7163 if (N1.getOpcode() == ISD::AND) {
7164 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7167 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7168 unsigned LSB = CountTrailingZeros_32(~InvMask);
7169 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7170 unsigned Mask = (1 << Width)-1;
7171 unsigned Mask2 = N11C->getZExtValue();
7172 if ((Mask & (~Mask2)) == 0)
7173 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7174 N->getOperand(0), N1.getOperand(0),
7180 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7181 /// ARMISD::VMOVRRD.
7182 static SDValue PerformVMOVRRDCombine(SDNode *N,
7183 TargetLowering::DAGCombinerInfo &DCI) {
7184 // vmovrrd(vmovdrr x, y) -> x,y
7185 SDValue InDouble = N->getOperand(0);
7186 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7187 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7189 // vmovrrd(load f64) -> (load i32), (load i32)
7190 SDNode *InNode = InDouble.getNode();
7191 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7192 InNode->getValueType(0) == MVT::f64 &&
7193 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7194 !cast<LoadSDNode>(InNode)->isVolatile()) {
7195 // TODO: Should this be done for non-FrameIndex operands?
7196 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7198 SelectionDAG &DAG = DCI.DAG;
7199 DebugLoc DL = LD->getDebugLoc();
7200 SDValue BasePtr = LD->getBasePtr();
7201 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7202 LD->getPointerInfo(), LD->isVolatile(),
7203 LD->isNonTemporal(), LD->isInvariant(),
7204 LD->getAlignment());
7206 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7207 DAG.getConstant(4, MVT::i32));
7208 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7209 LD->getPointerInfo(), LD->isVolatile(),
7210 LD->isNonTemporal(), LD->isInvariant(),
7211 std::min(4U, LD->getAlignment() / 2));
7213 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7214 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7215 DCI.RemoveFromWorklist(LD);
7223 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7224 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7225 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7226 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7227 SDValue Op0 = N->getOperand(0);
7228 SDValue Op1 = N->getOperand(1);
7229 if (Op0.getOpcode() == ISD::BITCAST)
7230 Op0 = Op0.getOperand(0);
7231 if (Op1.getOpcode() == ISD::BITCAST)
7232 Op1 = Op1.getOperand(0);
7233 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7234 Op0.getNode() == Op1.getNode() &&
7235 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7236 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7237 N->getValueType(0), Op0.getOperand(0));
7241 /// PerformSTORECombine - Target-specific dag combine xforms for
7243 static SDValue PerformSTORECombine(SDNode *N,
7244 TargetLowering::DAGCombinerInfo &DCI) {
7245 // Bitcast an i64 store extracted from a vector to f64.
7246 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7247 StoreSDNode *St = cast<StoreSDNode>(N);
7248 SDValue StVal = St->getValue();
7249 if (!ISD::isNormalStore(St) || St->isVolatile())
7252 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7253 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7254 SelectionDAG &DAG = DCI.DAG;
7255 DebugLoc DL = St->getDebugLoc();
7256 SDValue BasePtr = St->getBasePtr();
7257 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7258 StVal.getNode()->getOperand(0), BasePtr,
7259 St->getPointerInfo(), St->isVolatile(),
7260 St->isNonTemporal(), St->getAlignment());
7262 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7263 DAG.getConstant(4, MVT::i32));
7264 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7265 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7266 St->isNonTemporal(),
7267 std::min(4U, St->getAlignment() / 2));
7270 if (StVal.getValueType() != MVT::i64 ||
7271 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7274 SelectionDAG &DAG = DCI.DAG;
7275 DebugLoc dl = StVal.getDebugLoc();
7276 SDValue IntVec = StVal.getOperand(0);
7277 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7278 IntVec.getValueType().getVectorNumElements());
7279 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7280 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7281 Vec, StVal.getOperand(1));
7282 dl = N->getDebugLoc();
7283 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7284 // Make the DAGCombiner fold the bitcasts.
7285 DCI.AddToWorklist(Vec.getNode());
7286 DCI.AddToWorklist(ExtElt.getNode());
7287 DCI.AddToWorklist(V.getNode());
7288 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7289 St->getPointerInfo(), St->isVolatile(),
7290 St->isNonTemporal(), St->getAlignment(),
7294 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7295 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7296 /// i64 vector to have f64 elements, since the value can then be loaded
7297 /// directly into a VFP register.
7298 static bool hasNormalLoadOperand(SDNode *N) {
7299 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7300 for (unsigned i = 0; i < NumElts; ++i) {
7301 SDNode *Elt = N->getOperand(i).getNode();
7302 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7308 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7309 /// ISD::BUILD_VECTOR.
7310 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7311 TargetLowering::DAGCombinerInfo &DCI){
7312 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7313 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7314 // into a pair of GPRs, which is fine when the value is used as a scalar,
7315 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7316 SelectionDAG &DAG = DCI.DAG;
7317 if (N->getNumOperands() == 2) {
7318 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7323 // Load i64 elements as f64 values so that type legalization does not split
7324 // them up into i32 values.
7325 EVT VT = N->getValueType(0);
7326 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7328 DebugLoc dl = N->getDebugLoc();
7329 SmallVector<SDValue, 8> Ops;
7330 unsigned NumElts = VT.getVectorNumElements();
7331 for (unsigned i = 0; i < NumElts; ++i) {
7332 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7334 // Make the DAGCombiner fold the bitcast.
7335 DCI.AddToWorklist(V.getNode());
7337 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7338 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7339 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7342 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7343 /// ISD::INSERT_VECTOR_ELT.
7344 static SDValue PerformInsertEltCombine(SDNode *N,
7345 TargetLowering::DAGCombinerInfo &DCI) {
7346 // Bitcast an i64 load inserted into a vector to f64.
7347 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7348 EVT VT = N->getValueType(0);
7349 SDNode *Elt = N->getOperand(1).getNode();
7350 if (VT.getVectorElementType() != MVT::i64 ||
7351 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7354 SelectionDAG &DAG = DCI.DAG;
7355 DebugLoc dl = N->getDebugLoc();
7356 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7357 VT.getVectorNumElements());
7358 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7359 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7360 // Make the DAGCombiner fold the bitcasts.
7361 DCI.AddToWorklist(Vec.getNode());
7362 DCI.AddToWorklist(V.getNode());
7363 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7364 Vec, V, N->getOperand(2));
7365 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7368 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7369 /// ISD::VECTOR_SHUFFLE.
7370 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7371 // The LLVM shufflevector instruction does not require the shuffle mask
7372 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7373 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7374 // operands do not match the mask length, they are extended by concatenating
7375 // them with undef vectors. That is probably the right thing for other
7376 // targets, but for NEON it is better to concatenate two double-register
7377 // size vector operands into a single quad-register size vector. Do that
7378 // transformation here:
7379 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7380 // shuffle(concat(v1, v2), undef)
7381 SDValue Op0 = N->getOperand(0);
7382 SDValue Op1 = N->getOperand(1);
7383 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7384 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7385 Op0.getNumOperands() != 2 ||
7386 Op1.getNumOperands() != 2)
7388 SDValue Concat0Op1 = Op0.getOperand(1);
7389 SDValue Concat1Op1 = Op1.getOperand(1);
7390 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7391 Concat1Op1.getOpcode() != ISD::UNDEF)
7393 // Skip the transformation if any of the types are illegal.
7394 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7395 EVT VT = N->getValueType(0);
7396 if (!TLI.isTypeLegal(VT) ||
7397 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7398 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7401 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7402 Op0.getOperand(0), Op1.getOperand(0));
7403 // Translate the shuffle mask.
7404 SmallVector<int, 16> NewMask;
7405 unsigned NumElts = VT.getVectorNumElements();
7406 unsigned HalfElts = NumElts/2;
7407 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7408 for (unsigned n = 0; n < NumElts; ++n) {
7409 int MaskElt = SVN->getMaskElt(n);
7411 if (MaskElt < (int)HalfElts)
7413 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7414 NewElt = HalfElts + MaskElt - NumElts;
7415 NewMask.push_back(NewElt);
7417 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7418 DAG.getUNDEF(VT), NewMask.data());
7421 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7422 /// NEON load/store intrinsics to merge base address updates.
7423 static SDValue CombineBaseUpdate(SDNode *N,
7424 TargetLowering::DAGCombinerInfo &DCI) {
7425 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7428 SelectionDAG &DAG = DCI.DAG;
7429 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7430 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7431 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7432 SDValue Addr = N->getOperand(AddrOpIdx);
7434 // Search for a use of the address operand that is an increment.
7435 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7436 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7438 if (User->getOpcode() != ISD::ADD ||
7439 UI.getUse().getResNo() != Addr.getResNo())
7442 // Check that the add is independent of the load/store. Otherwise, folding
7443 // it would create a cycle.
7444 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7447 // Find the new opcode for the updating load/store.
7449 bool isLaneOp = false;
7450 unsigned NewOpc = 0;
7451 unsigned NumVecs = 0;
7453 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7455 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7456 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7458 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7460 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7462 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7464 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7465 NumVecs = 2; isLaneOp = true; break;
7466 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7467 NumVecs = 3; isLaneOp = true; break;
7468 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7469 NumVecs = 4; isLaneOp = true; break;
7470 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7471 NumVecs = 1; isLoad = false; break;
7472 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7473 NumVecs = 2; isLoad = false; break;
7474 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7475 NumVecs = 3; isLoad = false; break;
7476 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7477 NumVecs = 4; isLoad = false; break;
7478 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7479 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7480 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7481 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7482 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7483 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7487 switch (N->getOpcode()) {
7488 default: llvm_unreachable("unexpected opcode for Neon base update");
7489 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7490 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7491 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7495 // Find the size of memory referenced by the load/store.
7498 VecTy = N->getValueType(0);
7500 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7501 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7503 NumBytes /= VecTy.getVectorNumElements();
7505 // If the increment is a constant, it must match the memory ref size.
7506 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7507 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7508 uint64_t IncVal = CInc->getZExtValue();
7509 if (IncVal != NumBytes)
7511 } else if (NumBytes >= 3 * 16) {
7512 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7513 // separate instructions that make it harder to use a non-constant update.
7517 // Create the new updating load/store node.
7519 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7521 for (n = 0; n < NumResultVecs; ++n)
7523 Tys[n++] = MVT::i32;
7524 Tys[n] = MVT::Other;
7525 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7526 SmallVector<SDValue, 8> Ops;
7527 Ops.push_back(N->getOperand(0)); // incoming chain
7528 Ops.push_back(N->getOperand(AddrOpIdx));
7530 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7531 Ops.push_back(N->getOperand(i));
7533 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7534 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7535 Ops.data(), Ops.size(),
7536 MemInt->getMemoryVT(),
7537 MemInt->getMemOperand());
7540 std::vector<SDValue> NewResults;
7541 for (unsigned i = 0; i < NumResultVecs; ++i) {
7542 NewResults.push_back(SDValue(UpdN.getNode(), i));
7544 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7545 DCI.CombineTo(N, NewResults);
7546 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7553 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7554 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7555 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7557 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7558 SelectionDAG &DAG = DCI.DAG;
7559 EVT VT = N->getValueType(0);
7560 // vldN-dup instructions only support 64-bit vectors for N > 1.
7561 if (!VT.is64BitVector())
7564 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7565 SDNode *VLD = N->getOperand(0).getNode();
7566 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7568 unsigned NumVecs = 0;
7569 unsigned NewOpc = 0;
7570 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7571 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7573 NewOpc = ARMISD::VLD2DUP;
7574 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7576 NewOpc = ARMISD::VLD3DUP;
7577 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7579 NewOpc = ARMISD::VLD4DUP;
7584 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7585 // numbers match the load.
7586 unsigned VLDLaneNo =
7587 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7588 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7590 // Ignore uses of the chain result.
7591 if (UI.getUse().getResNo() == NumVecs)
7594 if (User->getOpcode() != ARMISD::VDUPLANE ||
7595 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7599 // Create the vldN-dup node.
7602 for (n = 0; n < NumVecs; ++n)
7604 Tys[n] = MVT::Other;
7605 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7606 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7607 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7608 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7609 Ops, 2, VLDMemInt->getMemoryVT(),
7610 VLDMemInt->getMemOperand());
7613 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7615 unsigned ResNo = UI.getUse().getResNo();
7616 // Ignore uses of the chain result.
7617 if (ResNo == NumVecs)
7620 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7623 // Now the vldN-lane intrinsic is dead except for its chain result.
7624 // Update uses of the chain.
7625 std::vector<SDValue> VLDDupResults;
7626 for (unsigned n = 0; n < NumVecs; ++n)
7627 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7628 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7629 DCI.CombineTo(VLD, VLDDupResults);
7634 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7635 /// ARMISD::VDUPLANE.
7636 static SDValue PerformVDUPLANECombine(SDNode *N,
7637 TargetLowering::DAGCombinerInfo &DCI) {
7638 SDValue Op = N->getOperand(0);
7640 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7641 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7642 if (CombineVLDDUP(N, DCI))
7643 return SDValue(N, 0);
7645 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7646 // redundant. Ignore bit_converts for now; element sizes are checked below.
7647 while (Op.getOpcode() == ISD::BITCAST)
7648 Op = Op.getOperand(0);
7649 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7652 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7653 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7654 // The canonical VMOV for a zero vector uses a 32-bit element size.
7655 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7657 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7659 EVT VT = N->getValueType(0);
7660 if (EltSize > VT.getVectorElementType().getSizeInBits())
7663 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7666 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7667 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7668 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7672 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7674 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7679 APFloat APF = C->getValueAPF();
7680 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7681 != APFloat::opOK || !isExact)
7684 c0 = (I == 0) ? cN : c0;
7685 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7692 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7693 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7694 /// when the VMUL has a constant operand that is a power of 2.
7696 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7697 /// vmul.f32 d16, d17, d16
7698 /// vcvt.s32.f32 d16, d16
7700 /// vcvt.s32.f32 d16, d16, #3
7701 static SDValue PerformVCVTCombine(SDNode *N,
7702 TargetLowering::DAGCombinerInfo &DCI,
7703 const ARMSubtarget *Subtarget) {
7704 SelectionDAG &DAG = DCI.DAG;
7705 SDValue Op = N->getOperand(0);
7707 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7708 Op.getOpcode() != ISD::FMUL)
7712 SDValue N0 = Op->getOperand(0);
7713 SDValue ConstVec = Op->getOperand(1);
7714 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7716 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7717 !isConstVecPow2(ConstVec, isSigned, C))
7720 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7721 Intrinsic::arm_neon_vcvtfp2fxu;
7722 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7724 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7725 DAG.getConstant(Log2_64(C), MVT::i32));
7728 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7729 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7730 /// when the VDIV has a constant operand that is a power of 2.
7732 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7733 /// vcvt.f32.s32 d16, d16
7734 /// vdiv.f32 d16, d17, d16
7736 /// vcvt.f32.s32 d16, d16, #3
7737 static SDValue PerformVDIVCombine(SDNode *N,
7738 TargetLowering::DAGCombinerInfo &DCI,
7739 const ARMSubtarget *Subtarget) {
7740 SelectionDAG &DAG = DCI.DAG;
7741 SDValue Op = N->getOperand(0);
7742 unsigned OpOpcode = Op.getNode()->getOpcode();
7744 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7745 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7749 SDValue ConstVec = N->getOperand(1);
7750 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7752 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7753 !isConstVecPow2(ConstVec, isSigned, C))
7756 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7757 Intrinsic::arm_neon_vcvtfxu2fp;
7758 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7760 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7761 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7764 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7765 /// operand of a vector shift operation, where all the elements of the
7766 /// build_vector must have the same constant integer value.
7767 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7768 // Ignore bit_converts.
7769 while (Op.getOpcode() == ISD::BITCAST)
7770 Op = Op.getOperand(0);
7771 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7772 APInt SplatBits, SplatUndef;
7773 unsigned SplatBitSize;
7775 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7776 HasAnyUndefs, ElementBits) ||
7777 SplatBitSize > ElementBits)
7779 Cnt = SplatBits.getSExtValue();
7783 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7784 /// operand of a vector shift left operation. That value must be in the range:
7785 /// 0 <= Value < ElementBits for a left shift; or
7786 /// 0 <= Value <= ElementBits for a long left shift.
7787 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7788 assert(VT.isVector() && "vector shift count is not a vector type");
7789 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7790 if (! getVShiftImm(Op, ElementBits, Cnt))
7792 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7795 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7796 /// operand of a vector shift right operation. For a shift opcode, the value
7797 /// is positive, but for an intrinsic the value count must be negative. The
7798 /// absolute value must be in the range:
7799 /// 1 <= |Value| <= ElementBits for a right shift; or
7800 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7801 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7803 assert(VT.isVector() && "vector shift count is not a vector type");
7804 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7805 if (! getVShiftImm(Op, ElementBits, Cnt))
7809 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7812 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7813 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7814 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7817 // Don't do anything for most intrinsics.
7820 // Vector shifts: check for immediate versions and lower them.
7821 // Note: This is done during DAG combining instead of DAG legalizing because
7822 // the build_vectors for 64-bit vector element shift counts are generally
7823 // not legal, and it is hard to see their values after they get legalized to
7824 // loads from a constant pool.
7825 case Intrinsic::arm_neon_vshifts:
7826 case Intrinsic::arm_neon_vshiftu:
7827 case Intrinsic::arm_neon_vshiftls:
7828 case Intrinsic::arm_neon_vshiftlu:
7829 case Intrinsic::arm_neon_vshiftn:
7830 case Intrinsic::arm_neon_vrshifts:
7831 case Intrinsic::arm_neon_vrshiftu:
7832 case Intrinsic::arm_neon_vrshiftn:
7833 case Intrinsic::arm_neon_vqshifts:
7834 case Intrinsic::arm_neon_vqshiftu:
7835 case Intrinsic::arm_neon_vqshiftsu:
7836 case Intrinsic::arm_neon_vqshiftns:
7837 case Intrinsic::arm_neon_vqshiftnu:
7838 case Intrinsic::arm_neon_vqshiftnsu:
7839 case Intrinsic::arm_neon_vqrshiftns:
7840 case Intrinsic::arm_neon_vqrshiftnu:
7841 case Intrinsic::arm_neon_vqrshiftnsu: {
7842 EVT VT = N->getOperand(1).getValueType();
7844 unsigned VShiftOpc = 0;
7847 case Intrinsic::arm_neon_vshifts:
7848 case Intrinsic::arm_neon_vshiftu:
7849 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7850 VShiftOpc = ARMISD::VSHL;
7853 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7854 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7855 ARMISD::VSHRs : ARMISD::VSHRu);
7860 case Intrinsic::arm_neon_vshiftls:
7861 case Intrinsic::arm_neon_vshiftlu:
7862 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7864 llvm_unreachable("invalid shift count for vshll intrinsic");
7866 case Intrinsic::arm_neon_vrshifts:
7867 case Intrinsic::arm_neon_vrshiftu:
7868 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7872 case Intrinsic::arm_neon_vqshifts:
7873 case Intrinsic::arm_neon_vqshiftu:
7874 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7878 case Intrinsic::arm_neon_vqshiftsu:
7879 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7881 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7883 case Intrinsic::arm_neon_vshiftn:
7884 case Intrinsic::arm_neon_vrshiftn:
7885 case Intrinsic::arm_neon_vqshiftns:
7886 case Intrinsic::arm_neon_vqshiftnu:
7887 case Intrinsic::arm_neon_vqshiftnsu:
7888 case Intrinsic::arm_neon_vqrshiftns:
7889 case Intrinsic::arm_neon_vqrshiftnu:
7890 case Intrinsic::arm_neon_vqrshiftnsu:
7891 // Narrowing shifts require an immediate right shift.
7892 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7894 llvm_unreachable("invalid shift count for narrowing vector shift "
7898 llvm_unreachable("unhandled vector shift");
7902 case Intrinsic::arm_neon_vshifts:
7903 case Intrinsic::arm_neon_vshiftu:
7904 // Opcode already set above.
7906 case Intrinsic::arm_neon_vshiftls:
7907 case Intrinsic::arm_neon_vshiftlu:
7908 if (Cnt == VT.getVectorElementType().getSizeInBits())
7909 VShiftOpc = ARMISD::VSHLLi;
7911 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7912 ARMISD::VSHLLs : ARMISD::VSHLLu);
7914 case Intrinsic::arm_neon_vshiftn:
7915 VShiftOpc = ARMISD::VSHRN; break;
7916 case Intrinsic::arm_neon_vrshifts:
7917 VShiftOpc = ARMISD::VRSHRs; break;
7918 case Intrinsic::arm_neon_vrshiftu:
7919 VShiftOpc = ARMISD::VRSHRu; break;
7920 case Intrinsic::arm_neon_vrshiftn:
7921 VShiftOpc = ARMISD::VRSHRN; break;
7922 case Intrinsic::arm_neon_vqshifts:
7923 VShiftOpc = ARMISD::VQSHLs; break;
7924 case Intrinsic::arm_neon_vqshiftu:
7925 VShiftOpc = ARMISD::VQSHLu; break;
7926 case Intrinsic::arm_neon_vqshiftsu:
7927 VShiftOpc = ARMISD::VQSHLsu; break;
7928 case Intrinsic::arm_neon_vqshiftns:
7929 VShiftOpc = ARMISD::VQSHRNs; break;
7930 case Intrinsic::arm_neon_vqshiftnu:
7931 VShiftOpc = ARMISD::VQSHRNu; break;
7932 case Intrinsic::arm_neon_vqshiftnsu:
7933 VShiftOpc = ARMISD::VQSHRNsu; break;
7934 case Intrinsic::arm_neon_vqrshiftns:
7935 VShiftOpc = ARMISD::VQRSHRNs; break;
7936 case Intrinsic::arm_neon_vqrshiftnu:
7937 VShiftOpc = ARMISD::VQRSHRNu; break;
7938 case Intrinsic::arm_neon_vqrshiftnsu:
7939 VShiftOpc = ARMISD::VQRSHRNsu; break;
7942 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7943 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7946 case Intrinsic::arm_neon_vshiftins: {
7947 EVT VT = N->getOperand(1).getValueType();
7949 unsigned VShiftOpc = 0;
7951 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7952 VShiftOpc = ARMISD::VSLI;
7953 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7954 VShiftOpc = ARMISD::VSRI;
7956 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7959 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7960 N->getOperand(1), N->getOperand(2),
7961 DAG.getConstant(Cnt, MVT::i32));
7964 case Intrinsic::arm_neon_vqrshifts:
7965 case Intrinsic::arm_neon_vqrshiftu:
7966 // No immediate versions of these to check for.
7973 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
7974 /// lowers them. As with the vector shift intrinsics, this is done during DAG
7975 /// combining instead of DAG legalizing because the build_vectors for 64-bit
7976 /// vector element shift counts are generally not legal, and it is hard to see
7977 /// their values after they get legalized to loads from a constant pool.
7978 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7979 const ARMSubtarget *ST) {
7980 EVT VT = N->getValueType(0);
7981 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
7982 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
7983 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
7984 SDValue N1 = N->getOperand(1);
7985 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
7986 SDValue N0 = N->getOperand(0);
7987 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
7988 DAG.MaskedValueIsZero(N0.getOperand(0),
7989 APInt::getHighBitsSet(32, 16)))
7990 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
7994 // Nothing to be done for scalar shifts.
7995 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7996 if (!VT.isVector() || !TLI.isTypeLegal(VT))
7999 assert(ST->hasNEON() && "unexpected vector shift");
8002 switch (N->getOpcode()) {
8003 default: llvm_unreachable("unexpected shift opcode");
8006 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8007 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8008 DAG.getConstant(Cnt, MVT::i32));
8013 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8014 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8015 ARMISD::VSHRs : ARMISD::VSHRu);
8016 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8017 DAG.getConstant(Cnt, MVT::i32));
8023 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8024 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8025 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8026 const ARMSubtarget *ST) {
8027 SDValue N0 = N->getOperand(0);
8029 // Check for sign- and zero-extensions of vector extract operations of 8-
8030 // and 16-bit vector elements. NEON supports these directly. They are
8031 // handled during DAG combining because type legalization will promote them
8032 // to 32-bit types and it is messy to recognize the operations after that.
8033 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8034 SDValue Vec = N0.getOperand(0);
8035 SDValue Lane = N0.getOperand(1);
8036 EVT VT = N->getValueType(0);
8037 EVT EltVT = N0.getValueType();
8038 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8040 if (VT == MVT::i32 &&
8041 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8042 TLI.isTypeLegal(Vec.getValueType()) &&
8043 isa<ConstantSDNode>(Lane)) {
8046 switch (N->getOpcode()) {
8047 default: llvm_unreachable("unexpected opcode");
8048 case ISD::SIGN_EXTEND:
8049 Opc = ARMISD::VGETLANEs;
8051 case ISD::ZERO_EXTEND:
8052 case ISD::ANY_EXTEND:
8053 Opc = ARMISD::VGETLANEu;
8056 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8063 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8064 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8065 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8066 const ARMSubtarget *ST) {
8067 // If the target supports NEON, try to use vmax/vmin instructions for f32
8068 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8069 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8070 // a NaN; only do the transformation when it matches that behavior.
8072 // For now only do this when using NEON for FP operations; if using VFP, it
8073 // is not obvious that the benefit outweighs the cost of switching to the
8075 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8076 N->getValueType(0) != MVT::f32)
8079 SDValue CondLHS = N->getOperand(0);
8080 SDValue CondRHS = N->getOperand(1);
8081 SDValue LHS = N->getOperand(2);
8082 SDValue RHS = N->getOperand(3);
8083 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8085 unsigned Opcode = 0;
8087 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8088 IsReversed = false; // x CC y ? x : y
8089 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8090 IsReversed = true ; // x CC y ? y : x
8104 // If LHS is NaN, an ordered comparison will be false and the result will
8105 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8106 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8107 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8108 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8110 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8111 // will return -0, so vmin can only be used for unsafe math or if one of
8112 // the operands is known to be nonzero.
8113 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8114 !DAG.getTarget().Options.UnsafeFPMath &&
8115 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8117 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8126 // If LHS is NaN, an ordered comparison will be false and the result will
8127 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8128 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8129 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8130 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8132 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8133 // will return +0, so vmax can only be used for unsafe math or if one of
8134 // the operands is known to be nonzero.
8135 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8136 !DAG.getTarget().Options.UnsafeFPMath &&
8137 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8139 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8145 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8148 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8150 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8151 SDValue Cmp = N->getOperand(4);
8152 if (Cmp.getOpcode() != ARMISD::CMPZ)
8153 // Only looking at EQ and NE cases.
8156 EVT VT = N->getValueType(0);
8157 DebugLoc dl = N->getDebugLoc();
8158 SDValue LHS = Cmp.getOperand(0);
8159 SDValue RHS = Cmp.getOperand(1);
8160 SDValue FalseVal = N->getOperand(0);
8161 SDValue TrueVal = N->getOperand(1);
8162 SDValue ARMcc = N->getOperand(2);
8163 ARMCC::CondCodes CC =
8164 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8182 /// FIXME: Turn this into a target neutral optimization?
8184 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8185 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8186 N->getOperand(3), Cmp);
8187 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8189 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8190 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8191 N->getOperand(3), NewCmp);
8194 if (Res.getNode()) {
8195 APInt KnownZero, KnownOne;
8196 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8197 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8198 // Capture demanded bits information that would be otherwise lost.
8199 if (KnownZero == 0xfffffffe)
8200 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8201 DAG.getValueType(MVT::i1));
8202 else if (KnownZero == 0xffffff00)
8203 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8204 DAG.getValueType(MVT::i8));
8205 else if (KnownZero == 0xffff0000)
8206 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8207 DAG.getValueType(MVT::i16));
8213 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8214 DAGCombinerInfo &DCI) const {
8215 switch (N->getOpcode()) {
8217 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8218 case ISD::SUB: return PerformSUBCombine(N, DCI);
8219 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8220 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8221 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8222 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8223 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8224 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8225 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8226 case ISD::STORE: return PerformSTORECombine(N, DCI);
8227 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8228 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8229 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8230 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8231 case ISD::FP_TO_SINT:
8232 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8233 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8234 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8237 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8238 case ISD::SIGN_EXTEND:
8239 case ISD::ZERO_EXTEND:
8240 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8241 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8242 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8243 case ARMISD::VLD2DUP:
8244 case ARMISD::VLD3DUP:
8245 case ARMISD::VLD4DUP:
8246 return CombineBaseUpdate(N, DCI);
8247 case ISD::INTRINSIC_VOID:
8248 case ISD::INTRINSIC_W_CHAIN:
8249 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8250 case Intrinsic::arm_neon_vld1:
8251 case Intrinsic::arm_neon_vld2:
8252 case Intrinsic::arm_neon_vld3:
8253 case Intrinsic::arm_neon_vld4:
8254 case Intrinsic::arm_neon_vld2lane:
8255 case Intrinsic::arm_neon_vld3lane:
8256 case Intrinsic::arm_neon_vld4lane:
8257 case Intrinsic::arm_neon_vst1:
8258 case Intrinsic::arm_neon_vst2:
8259 case Intrinsic::arm_neon_vst3:
8260 case Intrinsic::arm_neon_vst4:
8261 case Intrinsic::arm_neon_vst2lane:
8262 case Intrinsic::arm_neon_vst3lane:
8263 case Intrinsic::arm_neon_vst4lane:
8264 return CombineBaseUpdate(N, DCI);
8272 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8274 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8277 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8278 if (!Subtarget->allowsUnalignedMem())
8281 switch (VT.getSimpleVT().SimpleTy) {
8288 // FIXME: VLD1 etc with standard alignment is legal.
8292 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8293 unsigned AlignCheck) {
8294 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8295 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8298 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8299 unsigned DstAlign, unsigned SrcAlign,
8302 MachineFunction &MF) const {
8303 const Function *F = MF.getFunction();
8305 // See if we can use NEON instructions for this...
8307 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8308 Subtarget->hasNEON()) {
8309 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8311 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8316 // Lowering to i32/i16 if the size permits.
8319 } else if (Size >= 2) {
8323 // Let the target-independent logic figure it out.
8327 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8332 switch (VT.getSimpleVT().SimpleTy) {
8333 default: return false;
8348 if ((V & (Scale - 1)) != 0)
8351 return V == (V & ((1LL << 5) - 1));
8354 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8355 const ARMSubtarget *Subtarget) {
8362 switch (VT.getSimpleVT().SimpleTy) {
8363 default: return false;
8368 // + imm12 or - imm8
8370 return V == (V & ((1LL << 8) - 1));
8371 return V == (V & ((1LL << 12) - 1));
8374 // Same as ARM mode. FIXME: NEON?
8375 if (!Subtarget->hasVFP2())
8380 return V == (V & ((1LL << 8) - 1));
8384 /// isLegalAddressImmediate - Return true if the integer value can be used
8385 /// as the offset of the target addressing mode for load / store of the
8387 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8388 const ARMSubtarget *Subtarget) {
8395 if (Subtarget->isThumb1Only())
8396 return isLegalT1AddressImmediate(V, VT);
8397 else if (Subtarget->isThumb2())
8398 return isLegalT2AddressImmediate(V, VT, Subtarget);
8403 switch (VT.getSimpleVT().SimpleTy) {
8404 default: return false;
8409 return V == (V & ((1LL << 12) - 1));
8412 return V == (V & ((1LL << 8) - 1));
8415 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8420 return V == (V & ((1LL << 8) - 1));
8424 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8426 int Scale = AM.Scale;
8430 switch (VT.getSimpleVT().SimpleTy) {
8431 default: return false;
8440 return Scale == 2 || Scale == 4 || Scale == 8;
8443 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8447 // Note, we allow "void" uses (basically, uses that aren't loads or
8448 // stores), because arm allows folding a scale into many arithmetic
8449 // operations. This should be made more precise and revisited later.
8451 // Allow r << imm, but the imm has to be a multiple of two.
8452 if (Scale & 1) return false;
8453 return isPowerOf2_32(Scale);
8457 /// isLegalAddressingMode - Return true if the addressing mode represented
8458 /// by AM is legal for this target, for a load/store of the specified type.
8459 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8461 EVT VT = getValueType(Ty, true);
8462 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8465 // Can never fold addr of global into load/store.
8470 case 0: // no scale reg, must be "r+i" or "r", or "i".
8473 if (Subtarget->isThumb1Only())
8477 // ARM doesn't support any R+R*scale+imm addr modes.
8484 if (Subtarget->isThumb2())
8485 return isLegalT2ScaledAddressingMode(AM, VT);
8487 int Scale = AM.Scale;
8488 switch (VT.getSimpleVT().SimpleTy) {
8489 default: return false;
8493 if (Scale < 0) Scale = -Scale;
8497 return isPowerOf2_32(Scale & ~1);
8501 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8506 // Note, we allow "void" uses (basically, uses that aren't loads or
8507 // stores), because arm allows folding a scale into many arithmetic
8508 // operations. This should be made more precise and revisited later.
8510 // Allow r << imm, but the imm has to be a multiple of two.
8511 if (Scale & 1) return false;
8512 return isPowerOf2_32(Scale);
8518 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8519 /// icmp immediate, that is the target has icmp instructions which can compare
8520 /// a register against the immediate without having to materialize the
8521 /// immediate into a register.
8522 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8523 if (!Subtarget->isThumb())
8524 return ARM_AM::getSOImmVal(Imm) != -1;
8525 if (Subtarget->isThumb2())
8526 return ARM_AM::getT2SOImmVal(Imm) != -1;
8527 return Imm >= 0 && Imm <= 255;
8530 /// isLegalAddImmediate - Return true if the specified immediate is legal
8531 /// add immediate, that is the target has add instructions which can add
8532 /// a register with the immediate without having to materialize the
8533 /// immediate into a register.
8534 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8535 return ARM_AM::getSOImmVal(Imm) != -1;
8538 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8539 bool isSEXTLoad, SDValue &Base,
8540 SDValue &Offset, bool &isInc,
8541 SelectionDAG &DAG) {
8542 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8545 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8547 Base = Ptr->getOperand(0);
8548 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8549 int RHSC = (int)RHS->getZExtValue();
8550 if (RHSC < 0 && RHSC > -256) {
8551 assert(Ptr->getOpcode() == ISD::ADD);
8553 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8557 isInc = (Ptr->getOpcode() == ISD::ADD);
8558 Offset = Ptr->getOperand(1);
8560 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8562 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8563 int RHSC = (int)RHS->getZExtValue();
8564 if (RHSC < 0 && RHSC > -0x1000) {
8565 assert(Ptr->getOpcode() == ISD::ADD);
8567 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8568 Base = Ptr->getOperand(0);
8573 if (Ptr->getOpcode() == ISD::ADD) {
8575 ARM_AM::ShiftOpc ShOpcVal=
8576 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8577 if (ShOpcVal != ARM_AM::no_shift) {
8578 Base = Ptr->getOperand(1);
8579 Offset = Ptr->getOperand(0);
8581 Base = Ptr->getOperand(0);
8582 Offset = Ptr->getOperand(1);
8587 isInc = (Ptr->getOpcode() == ISD::ADD);
8588 Base = Ptr->getOperand(0);
8589 Offset = Ptr->getOperand(1);
8593 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8597 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8598 bool isSEXTLoad, SDValue &Base,
8599 SDValue &Offset, bool &isInc,
8600 SelectionDAG &DAG) {
8601 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8604 Base = Ptr->getOperand(0);
8605 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8606 int RHSC = (int)RHS->getZExtValue();
8607 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8608 assert(Ptr->getOpcode() == ISD::ADD);
8610 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8612 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8613 isInc = Ptr->getOpcode() == ISD::ADD;
8614 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8622 /// getPreIndexedAddressParts - returns true by value, base pointer and
8623 /// offset pointer and addressing mode by reference if the node's address
8624 /// can be legally represented as pre-indexed load / store address.
8626 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8628 ISD::MemIndexedMode &AM,
8629 SelectionDAG &DAG) const {
8630 if (Subtarget->isThumb1Only())
8635 bool isSEXTLoad = false;
8636 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8637 Ptr = LD->getBasePtr();
8638 VT = LD->getMemoryVT();
8639 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8640 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8641 Ptr = ST->getBasePtr();
8642 VT = ST->getMemoryVT();
8647 bool isLegal = false;
8648 if (Subtarget->isThumb2())
8649 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8650 Offset, isInc, DAG);
8652 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8653 Offset, isInc, DAG);
8657 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8661 /// getPostIndexedAddressParts - returns true by value, base pointer and
8662 /// offset pointer and addressing mode by reference if this node can be
8663 /// combined with a load / store to form a post-indexed load / store.
8664 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8667 ISD::MemIndexedMode &AM,
8668 SelectionDAG &DAG) const {
8669 if (Subtarget->isThumb1Only())
8674 bool isSEXTLoad = false;
8675 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8676 VT = LD->getMemoryVT();
8677 Ptr = LD->getBasePtr();
8678 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8679 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8680 VT = ST->getMemoryVT();
8681 Ptr = ST->getBasePtr();
8686 bool isLegal = false;
8687 if (Subtarget->isThumb2())
8688 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8691 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8697 // Swap base ptr and offset to catch more post-index load / store when
8698 // it's legal. In Thumb2 mode, offset must be an immediate.
8699 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8700 !Subtarget->isThumb2())
8701 std::swap(Base, Offset);
8703 // Post-indexed load / store update the base pointer.
8708 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8712 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8716 const SelectionDAG &DAG,
8717 unsigned Depth) const {
8718 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8719 switch (Op.getOpcode()) {
8721 case ARMISD::CMOV: {
8722 // Bits are known zero/one if known on the LHS and RHS.
8723 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8724 if (KnownZero == 0 && KnownOne == 0) return;
8726 APInt KnownZeroRHS, KnownOneRHS;
8727 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8728 KnownZeroRHS, KnownOneRHS, Depth+1);
8729 KnownZero &= KnownZeroRHS;
8730 KnownOne &= KnownOneRHS;
8736 //===----------------------------------------------------------------------===//
8737 // ARM Inline Assembly Support
8738 //===----------------------------------------------------------------------===//
8740 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8741 // Looking for "rev" which is V6+.
8742 if (!Subtarget->hasV6Ops())
8745 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8746 std::string AsmStr = IA->getAsmString();
8747 SmallVector<StringRef, 4> AsmPieces;
8748 SplitString(AsmStr, AsmPieces, ";\n");
8750 switch (AsmPieces.size()) {
8751 default: return false;
8753 AsmStr = AsmPieces[0];
8755 SplitString(AsmStr, AsmPieces, " \t,");
8758 if (AsmPieces.size() == 3 &&
8759 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8760 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8761 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8762 if (Ty && Ty->getBitWidth() == 32)
8763 return IntrinsicLowering::LowerToByteSwap(CI);
8771 /// getConstraintType - Given a constraint letter, return the type of
8772 /// constraint it is for this target.
8773 ARMTargetLowering::ConstraintType
8774 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8775 if (Constraint.size() == 1) {
8776 switch (Constraint[0]) {
8778 case 'l': return C_RegisterClass;
8779 case 'w': return C_RegisterClass;
8780 case 'h': return C_RegisterClass;
8781 case 'x': return C_RegisterClass;
8782 case 't': return C_RegisterClass;
8783 case 'j': return C_Other; // Constant for movw.
8784 // An address with a single base register. Due to the way we
8785 // currently handle addresses it is the same as an 'r' memory constraint.
8786 case 'Q': return C_Memory;
8788 } else if (Constraint.size() == 2) {
8789 switch (Constraint[0]) {
8791 // All 'U+' constraints are addresses.
8792 case 'U': return C_Memory;
8795 return TargetLowering::getConstraintType(Constraint);
8798 /// Examine constraint type and operand type and determine a weight value.
8799 /// This object must already have been set up with the operand type
8800 /// and the current alternative constraint selected.
8801 TargetLowering::ConstraintWeight
8802 ARMTargetLowering::getSingleConstraintMatchWeight(
8803 AsmOperandInfo &info, const char *constraint) const {
8804 ConstraintWeight weight = CW_Invalid;
8805 Value *CallOperandVal = info.CallOperandVal;
8806 // If we don't have a value, we can't do a match,
8807 // but allow it at the lowest weight.
8808 if (CallOperandVal == NULL)
8810 Type *type = CallOperandVal->getType();
8811 // Look at the constraint type.
8812 switch (*constraint) {
8814 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8817 if (type->isIntegerTy()) {
8818 if (Subtarget->isThumb())
8819 weight = CW_SpecificReg;
8821 weight = CW_Register;
8825 if (type->isFloatingPointTy())
8826 weight = CW_Register;
8832 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8834 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8836 if (Constraint.size() == 1) {
8837 // GCC ARM Constraint Letters
8838 switch (Constraint[0]) {
8839 case 'l': // Low regs or general regs.
8840 if (Subtarget->isThumb())
8841 return RCPair(0U, ARM::tGPRRegisterClass);
8843 return RCPair(0U, ARM::GPRRegisterClass);
8844 case 'h': // High regs or no regs.
8845 if (Subtarget->isThumb())
8846 return RCPair(0U, ARM::hGPRRegisterClass);
8849 return RCPair(0U, ARM::GPRRegisterClass);
8852 return RCPair(0U, ARM::SPRRegisterClass);
8853 if (VT.getSizeInBits() == 64)
8854 return RCPair(0U, ARM::DPRRegisterClass);
8855 if (VT.getSizeInBits() == 128)
8856 return RCPair(0U, ARM::QPRRegisterClass);
8860 return RCPair(0U, ARM::SPR_8RegisterClass);
8861 if (VT.getSizeInBits() == 64)
8862 return RCPair(0U, ARM::DPR_8RegisterClass);
8863 if (VT.getSizeInBits() == 128)
8864 return RCPair(0U, ARM::QPR_8RegisterClass);
8868 return RCPair(0U, ARM::SPRRegisterClass);
8872 if (StringRef("{cc}").equals_lower(Constraint))
8873 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8875 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8878 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8879 /// vector. If it is invalid, don't add anything to Ops.
8880 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8881 std::string &Constraint,
8882 std::vector<SDValue>&Ops,
8883 SelectionDAG &DAG) const {
8884 SDValue Result(0, 0);
8886 // Currently only support length 1 constraints.
8887 if (Constraint.length() != 1) return;
8889 char ConstraintLetter = Constraint[0];
8890 switch (ConstraintLetter) {
8893 case 'I': case 'J': case 'K': case 'L':
8894 case 'M': case 'N': case 'O':
8895 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8899 int64_t CVal64 = C->getSExtValue();
8900 int CVal = (int) CVal64;
8901 // None of these constraints allow values larger than 32 bits. Check
8902 // that the value fits in an int.
8906 switch (ConstraintLetter) {
8908 // Constant suitable for movw, must be between 0 and
8910 if (Subtarget->hasV6T2Ops())
8911 if (CVal >= 0 && CVal <= 65535)
8915 if (Subtarget->isThumb1Only()) {
8916 // This must be a constant between 0 and 255, for ADD
8918 if (CVal >= 0 && CVal <= 255)
8920 } else if (Subtarget->isThumb2()) {
8921 // A constant that can be used as an immediate value in a
8922 // data-processing instruction.
8923 if (ARM_AM::getT2SOImmVal(CVal) != -1)
8926 // A constant that can be used as an immediate value in a
8927 // data-processing instruction.
8928 if (ARM_AM::getSOImmVal(CVal) != -1)
8934 if (Subtarget->isThumb()) { // FIXME thumb2
8935 // This must be a constant between -255 and -1, for negated ADD
8936 // immediates. This can be used in GCC with an "n" modifier that
8937 // prints the negated value, for use with SUB instructions. It is
8938 // not useful otherwise but is implemented for compatibility.
8939 if (CVal >= -255 && CVal <= -1)
8942 // This must be a constant between -4095 and 4095. It is not clear
8943 // what this constraint is intended for. Implemented for
8944 // compatibility with GCC.
8945 if (CVal >= -4095 && CVal <= 4095)
8951 if (Subtarget->isThumb1Only()) {
8952 // A 32-bit value where only one byte has a nonzero value. Exclude
8953 // zero to match GCC. This constraint is used by GCC internally for
8954 // constants that can be loaded with a move/shift combination.
8955 // It is not useful otherwise but is implemented for compatibility.
8956 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8958 } else if (Subtarget->isThumb2()) {
8959 // A constant whose bitwise inverse can be used as an immediate
8960 // value in a data-processing instruction. This can be used in GCC
8961 // with a "B" modifier that prints the inverted value, for use with
8962 // BIC and MVN instructions. It is not useful otherwise but is
8963 // implemented for compatibility.
8964 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8967 // A constant whose bitwise inverse can be used as an immediate
8968 // value in a data-processing instruction. This can be used in GCC
8969 // with a "B" modifier that prints the inverted value, for use with
8970 // BIC and MVN instructions. It is not useful otherwise but is
8971 // implemented for compatibility.
8972 if (ARM_AM::getSOImmVal(~CVal) != -1)
8978 if (Subtarget->isThumb1Only()) {
8979 // This must be a constant between -7 and 7,
8980 // for 3-operand ADD/SUB immediate instructions.
8981 if (CVal >= -7 && CVal < 7)
8983 } else if (Subtarget->isThumb2()) {
8984 // A constant whose negation can be used as an immediate value in a
8985 // data-processing instruction. This can be used in GCC with an "n"
8986 // modifier that prints the negated value, for use with SUB
8987 // instructions. It is not useful otherwise but is implemented for
8989 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8992 // A constant whose negation can be used as an immediate value in a
8993 // data-processing instruction. This can be used in GCC with an "n"
8994 // modifier that prints the negated value, for use with SUB
8995 // instructions. It is not useful otherwise but is implemented for
8997 if (ARM_AM::getSOImmVal(-CVal) != -1)
9003 if (Subtarget->isThumb()) { // FIXME thumb2
9004 // This must be a multiple of 4 between 0 and 1020, for
9005 // ADD sp + immediate.
9006 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9009 // A power of two or a constant between 0 and 32. This is used in
9010 // GCC for the shift amount on shifted register operands, but it is
9011 // useful in general for any shift amounts.
9012 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9018 if (Subtarget->isThumb()) { // FIXME thumb2
9019 // This must be a constant between 0 and 31, for shift amounts.
9020 if (CVal >= 0 && CVal <= 31)
9026 if (Subtarget->isThumb()) { // FIXME thumb2
9027 // This must be a multiple of 4 between -508 and 508, for
9028 // ADD/SUB sp = sp + immediate.
9029 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9034 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9038 if (Result.getNode()) {
9039 Ops.push_back(Result);
9042 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9046 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9047 // The ARM target isn't yet aware of offsets.
9051 bool ARM::isBitFieldInvertedMask(unsigned v) {
9052 if (v == 0xffffffff)
9054 // there can be 1's on either or both "outsides", all the "inside"
9056 unsigned int lsb = 0, msb = 31;
9057 while (v & (1 << msb)) --msb;
9058 while (v & (1 << lsb)) ++lsb;
9059 for (unsigned int i = lsb; i <= msb; ++i) {
9066 /// isFPImmLegal - Returns true if the target can instruction select the
9067 /// specified FP immediate natively. If false, the legalizer will
9068 /// materialize the FP immediate as a load from a constant pool.
9069 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9070 if (!Subtarget->hasVFP3())
9073 return ARM_AM::getFP32Imm(Imm) != -1;
9075 return ARM_AM::getFP64Imm(Imm) != -1;
9079 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9080 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9081 /// specified in the intrinsic calls.
9082 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9084 unsigned Intrinsic) const {
9085 switch (Intrinsic) {
9086 case Intrinsic::arm_neon_vld1:
9087 case Intrinsic::arm_neon_vld2:
9088 case Intrinsic::arm_neon_vld3:
9089 case Intrinsic::arm_neon_vld4:
9090 case Intrinsic::arm_neon_vld2lane:
9091 case Intrinsic::arm_neon_vld3lane:
9092 case Intrinsic::arm_neon_vld4lane: {
9093 Info.opc = ISD::INTRINSIC_W_CHAIN;
9094 // Conservatively set memVT to the entire set of vectors loaded.
9095 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9096 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9097 Info.ptrVal = I.getArgOperand(0);
9099 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9100 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9101 Info.vol = false; // volatile loads with NEON intrinsics not supported
9102 Info.readMem = true;
9103 Info.writeMem = false;
9106 case Intrinsic::arm_neon_vst1:
9107 case Intrinsic::arm_neon_vst2:
9108 case Intrinsic::arm_neon_vst3:
9109 case Intrinsic::arm_neon_vst4:
9110 case Intrinsic::arm_neon_vst2lane:
9111 case Intrinsic::arm_neon_vst3lane:
9112 case Intrinsic::arm_neon_vst4lane: {
9113 Info.opc = ISD::INTRINSIC_VOID;
9114 // Conservatively set memVT to the entire set of vectors stored.
9115 unsigned NumElts = 0;
9116 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9117 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9118 if (!ArgTy->isVectorTy())
9120 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9122 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9123 Info.ptrVal = I.getArgOperand(0);
9125 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9126 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9127 Info.vol = false; // volatile stores with NEON intrinsics not supported
9128 Info.readMem = false;
9129 Info.writeMem = true;
9132 case Intrinsic::arm_strexd: {
9133 Info.opc = ISD::INTRINSIC_W_CHAIN;
9134 Info.memVT = MVT::i64;
9135 Info.ptrVal = I.getArgOperand(2);
9139 Info.readMem = false;
9140 Info.writeMem = true;
9143 case Intrinsic::arm_ldrexd: {
9144 Info.opc = ISD::INTRINSIC_W_CHAIN;
9145 Info.memVT = MVT::i64;
9146 Info.ptrVal = I.getArgOperand(0);
9150 Info.readMem = true;
9151 Info.writeMem = false;