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()) {
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.
537 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
538 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
540 setTargetDAGCombine(ISD::INTRINSIC_VOID);
541 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
542 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
543 setTargetDAGCombine(ISD::SHL);
544 setTargetDAGCombine(ISD::SRL);
545 setTargetDAGCombine(ISD::SRA);
546 setTargetDAGCombine(ISD::SIGN_EXTEND);
547 setTargetDAGCombine(ISD::ZERO_EXTEND);
548 setTargetDAGCombine(ISD::ANY_EXTEND);
549 setTargetDAGCombine(ISD::SELECT_CC);
550 setTargetDAGCombine(ISD::BUILD_VECTOR);
551 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
552 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
553 setTargetDAGCombine(ISD::STORE);
554 setTargetDAGCombine(ISD::FP_TO_SINT);
555 setTargetDAGCombine(ISD::FP_TO_UINT);
556 setTargetDAGCombine(ISD::FDIV);
558 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
561 computeRegisterProperties();
563 // ARM does not have f32 extending load.
564 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
566 // ARM does not have i1 sign extending load.
567 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
569 // ARM supports all 4 flavors of integer indexed load / store.
570 if (!Subtarget->isThumb1Only()) {
571 for (unsigned im = (unsigned)ISD::PRE_INC;
572 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
573 setIndexedLoadAction(im, MVT::i1, Legal);
574 setIndexedLoadAction(im, MVT::i8, Legal);
575 setIndexedLoadAction(im, MVT::i16, Legal);
576 setIndexedLoadAction(im, MVT::i32, Legal);
577 setIndexedStoreAction(im, MVT::i1, Legal);
578 setIndexedStoreAction(im, MVT::i8, Legal);
579 setIndexedStoreAction(im, MVT::i16, Legal);
580 setIndexedStoreAction(im, MVT::i32, Legal);
584 // i64 operation support.
585 setOperationAction(ISD::MUL, MVT::i64, Expand);
586 setOperationAction(ISD::MULHU, MVT::i32, Expand);
587 if (Subtarget->isThumb1Only()) {
588 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
589 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
591 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
592 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
593 setOperationAction(ISD::MULHS, MVT::i32, Expand);
595 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
596 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
597 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
598 setOperationAction(ISD::SRL, MVT::i64, Custom);
599 setOperationAction(ISD::SRA, MVT::i64, Custom);
601 if (!Subtarget->isThumb1Only()) {
602 // FIXME: We should do this for Thumb1 as well.
603 setOperationAction(ISD::ADDC, MVT::i32, Custom);
604 setOperationAction(ISD::ADDE, MVT::i32, Custom);
605 setOperationAction(ISD::SUBC, MVT::i32, Custom);
606 setOperationAction(ISD::SUBE, MVT::i32, Custom);
609 // ARM does not have ROTL.
610 setOperationAction(ISD::ROTL, MVT::i32, Expand);
611 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
612 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
613 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
614 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
616 // These just redirect to CTTZ and CTLZ on ARM.
617 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
618 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
620 // Only ARMv6 has BSWAP.
621 if (!Subtarget->hasV6Ops())
622 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
624 // These are expanded into libcalls.
625 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
626 // v7M has a hardware divider
627 setOperationAction(ISD::SDIV, MVT::i32, Expand);
628 setOperationAction(ISD::UDIV, MVT::i32, Expand);
630 setOperationAction(ISD::SREM, MVT::i32, Expand);
631 setOperationAction(ISD::UREM, MVT::i32, Expand);
632 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
633 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
635 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
636 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
637 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
638 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
639 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
641 setOperationAction(ISD::TRAP, MVT::Other, Legal);
643 // Use the default implementation.
644 setOperationAction(ISD::VASTART, MVT::Other, Custom);
645 setOperationAction(ISD::VAARG, MVT::Other, Expand);
646 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
647 setOperationAction(ISD::VAEND, MVT::Other, Expand);
648 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
649 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
650 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
651 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
652 setExceptionPointerRegister(ARM::R0);
653 setExceptionSelectorRegister(ARM::R1);
655 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
656 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
657 // the default expansion.
658 // FIXME: This should be checking for v6k, not just v6.
659 if (Subtarget->hasDataBarrier() ||
660 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
661 // membarrier needs custom lowering; the rest are legal and handled
663 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
664 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
665 // Custom lowering for 64-bit ops
666 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
667 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
668 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
669 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
670 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
671 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
672 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
673 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
674 setInsertFencesForAtomic(true);
676 // Set them all for expansion, which will force libcalls.
677 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
678 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
679 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
680 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
681 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
682 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
683 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
684 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
685 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
686 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
687 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
688 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
689 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
690 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
691 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
692 // Unordered/Monotonic case.
693 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
694 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
695 // Since the libcalls include locking, fold in the fences
696 setShouldFoldAtomicFences(true);
699 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
701 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
702 if (!Subtarget->hasV6Ops()) {
703 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
704 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
706 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
708 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
709 !Subtarget->isThumb1Only()) {
710 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
711 // iff target supports vfp2.
712 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
713 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
716 // We want to custom lower some of our intrinsics.
717 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
718 if (Subtarget->isTargetDarwin()) {
719 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
720 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
721 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
724 setOperationAction(ISD::SETCC, MVT::i32, Expand);
725 setOperationAction(ISD::SETCC, MVT::f32, Expand);
726 setOperationAction(ISD::SETCC, MVT::f64, Expand);
727 setOperationAction(ISD::SELECT, MVT::i32, Custom);
728 setOperationAction(ISD::SELECT, MVT::f32, Custom);
729 setOperationAction(ISD::SELECT, MVT::f64, Custom);
730 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
731 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
732 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
734 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
735 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
736 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
737 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
738 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
740 // We don't support sin/cos/fmod/copysign/pow
741 setOperationAction(ISD::FSIN, MVT::f64, Expand);
742 setOperationAction(ISD::FSIN, MVT::f32, Expand);
743 setOperationAction(ISD::FCOS, MVT::f32, Expand);
744 setOperationAction(ISD::FCOS, MVT::f64, Expand);
745 setOperationAction(ISD::FREM, MVT::f64, Expand);
746 setOperationAction(ISD::FREM, MVT::f32, Expand);
747 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
748 !Subtarget->isThumb1Only()) {
749 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
750 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
752 setOperationAction(ISD::FPOW, MVT::f64, Expand);
753 setOperationAction(ISD::FPOW, MVT::f32, Expand);
755 setOperationAction(ISD::FMA, MVT::f64, Expand);
756 setOperationAction(ISD::FMA, MVT::f32, Expand);
758 // Various VFP goodness
759 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
760 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
761 if (Subtarget->hasVFP2()) {
762 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
763 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
764 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
765 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
767 // Special handling for half-precision FP.
768 if (!Subtarget->hasFP16()) {
769 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
770 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
774 // We have target-specific dag combine patterns for the following nodes:
775 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
776 setTargetDAGCombine(ISD::ADD);
777 setTargetDAGCombine(ISD::SUB);
778 setTargetDAGCombine(ISD::MUL);
780 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
781 setTargetDAGCombine(ISD::OR);
782 if (Subtarget->hasNEON())
783 setTargetDAGCombine(ISD::AND);
785 setStackPointerRegisterToSaveRestore(ARM::SP);
787 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
788 !Subtarget->hasVFP2())
789 setSchedulingPreference(Sched::RegPressure);
791 setSchedulingPreference(Sched::Hybrid);
793 //// temporary - rewrite interface to use type
794 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
795 maxStoresPerMemset = 16;
796 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
798 // On ARM arguments smaller than 4 bytes are extended, so all arguments
799 // are at least 4 bytes aligned.
800 setMinStackArgumentAlignment(4);
802 benefitFromCodePlacementOpt = true;
804 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
807 // FIXME: It might make sense to define the representative register class as the
808 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
809 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
810 // SPR's representative would be DPR_VFP2. This should work well if register
811 // pressure tracking were modified such that a register use would increment the
812 // pressure of the register class's representative and all of it's super
813 // classes' representatives transitively. We have not implemented this because
814 // of the difficulty prior to coalescing of modeling operand register classes
815 // due to the common occurrence of cross class copies and subregister insertions
817 std::pair<const TargetRegisterClass*, uint8_t>
818 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
819 const TargetRegisterClass *RRC = 0;
821 switch (VT.getSimpleVT().SimpleTy) {
823 return TargetLowering::findRepresentativeClass(VT);
824 // Use DPR as representative register class for all floating point
825 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
826 // the cost is 1 for both f32 and f64.
827 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
828 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
829 RRC = ARM::DPRRegisterClass;
830 // When NEON is used for SP, only half of the register file is available
831 // because operations that define both SP and DP results will be constrained
832 // to the VFP2 class (D0-D15). We currently model this constraint prior to
833 // coalescing by double-counting the SP regs. See the FIXME above.
834 if (Subtarget->useNEONForSinglePrecisionFP())
837 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
838 case MVT::v4f32: case MVT::v2f64:
839 RRC = ARM::DPRRegisterClass;
843 RRC = ARM::DPRRegisterClass;
847 RRC = ARM::DPRRegisterClass;
851 return std::make_pair(RRC, Cost);
854 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
857 case ARMISD::Wrapper: return "ARMISD::Wrapper";
858 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
859 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
860 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
861 case ARMISD::CALL: return "ARMISD::CALL";
862 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
863 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
864 case ARMISD::tCALL: return "ARMISD::tCALL";
865 case ARMISD::BRCOND: return "ARMISD::BRCOND";
866 case ARMISD::BR_JT: return "ARMISD::BR_JT";
867 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
868 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
869 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
870 case ARMISD::CMP: return "ARMISD::CMP";
871 case ARMISD::CMPZ: return "ARMISD::CMPZ";
872 case ARMISD::CMPFP: return "ARMISD::CMPFP";
873 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
874 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
875 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
876 case ARMISD::CMOV: return "ARMISD::CMOV";
878 case ARMISD::RBIT: return "ARMISD::RBIT";
880 case ARMISD::FTOSI: return "ARMISD::FTOSI";
881 case ARMISD::FTOUI: return "ARMISD::FTOUI";
882 case ARMISD::SITOF: return "ARMISD::SITOF";
883 case ARMISD::UITOF: return "ARMISD::UITOF";
885 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
886 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
887 case ARMISD::RRX: return "ARMISD::RRX";
889 case ARMISD::ADDC: return "ARMISD::ADDC";
890 case ARMISD::ADDE: return "ARMISD::ADDE";
891 case ARMISD::SUBC: return "ARMISD::SUBC";
892 case ARMISD::SUBE: return "ARMISD::SUBE";
894 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
895 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
897 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
898 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
900 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
902 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
904 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
906 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
907 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
909 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
911 case ARMISD::VCEQ: return "ARMISD::VCEQ";
912 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
913 case ARMISD::VCGE: return "ARMISD::VCGE";
914 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
915 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
916 case ARMISD::VCGEU: return "ARMISD::VCGEU";
917 case ARMISD::VCGT: return "ARMISD::VCGT";
918 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
919 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
920 case ARMISD::VCGTU: return "ARMISD::VCGTU";
921 case ARMISD::VTST: return "ARMISD::VTST";
923 case ARMISD::VSHL: return "ARMISD::VSHL";
924 case ARMISD::VSHRs: return "ARMISD::VSHRs";
925 case ARMISD::VSHRu: return "ARMISD::VSHRu";
926 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
927 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
928 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
929 case ARMISD::VSHRN: return "ARMISD::VSHRN";
930 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
931 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
932 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
933 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
934 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
935 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
936 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
937 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
938 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
939 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
940 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
941 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
942 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
943 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
944 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
945 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
946 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
947 case ARMISD::VDUP: return "ARMISD::VDUP";
948 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
949 case ARMISD::VEXT: return "ARMISD::VEXT";
950 case ARMISD::VREV64: return "ARMISD::VREV64";
951 case ARMISD::VREV32: return "ARMISD::VREV32";
952 case ARMISD::VREV16: return "ARMISD::VREV16";
953 case ARMISD::VZIP: return "ARMISD::VZIP";
954 case ARMISD::VUZP: return "ARMISD::VUZP";
955 case ARMISD::VTRN: return "ARMISD::VTRN";
956 case ARMISD::VTBL1: return "ARMISD::VTBL1";
957 case ARMISD::VTBL2: return "ARMISD::VTBL2";
958 case ARMISD::VMULLs: return "ARMISD::VMULLs";
959 case ARMISD::VMULLu: return "ARMISD::VMULLu";
960 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
961 case ARMISD::FMAX: return "ARMISD::FMAX";
962 case ARMISD::FMIN: return "ARMISD::FMIN";
963 case ARMISD::BFI: return "ARMISD::BFI";
964 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
965 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
966 case ARMISD::VBSL: return "ARMISD::VBSL";
967 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
968 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
969 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
970 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
971 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
972 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
973 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
974 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
975 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
976 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
977 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
978 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
979 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
980 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
981 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
982 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
983 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
984 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
985 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
986 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
990 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
991 if (!VT.isVector()) return getPointerTy();
992 return VT.changeVectorElementTypeToInteger();
995 /// getRegClassFor - Return the register class that should be used for the
996 /// specified value type.
997 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
998 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
999 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1000 // load / store 4 to 8 consecutive D registers.
1001 if (Subtarget->hasNEON()) {
1002 if (VT == MVT::v4i64)
1003 return ARM::QQPRRegisterClass;
1004 else if (VT == MVT::v8i64)
1005 return ARM::QQQQPRRegisterClass;
1007 return TargetLowering::getRegClassFor(VT);
1010 // Create a fast isel object.
1012 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1013 return ARM::createFastISel(funcInfo);
1016 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1017 /// be used for loads / stores from the global.
1018 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1019 return (Subtarget->isThumb1Only() ? 127 : 4095);
1022 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1023 unsigned NumVals = N->getNumValues();
1025 return Sched::RegPressure;
1027 for (unsigned i = 0; i != NumVals; ++i) {
1028 EVT VT = N->getValueType(i);
1029 if (VT == MVT::Glue || VT == MVT::Other)
1031 if (VT.isFloatingPoint() || VT.isVector())
1035 if (!N->isMachineOpcode())
1036 return Sched::RegPressure;
1038 // Load are scheduled for latency even if there instruction itinerary
1039 // is not available.
1040 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1041 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1043 if (MCID.getNumDefs() == 0)
1044 return Sched::RegPressure;
1045 if (!Itins->isEmpty() &&
1046 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1049 return Sched::RegPressure;
1052 //===----------------------------------------------------------------------===//
1054 //===----------------------------------------------------------------------===//
1056 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1057 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1059 default: llvm_unreachable("Unknown condition code!");
1060 case ISD::SETNE: return ARMCC::NE;
1061 case ISD::SETEQ: return ARMCC::EQ;
1062 case ISD::SETGT: return ARMCC::GT;
1063 case ISD::SETGE: return ARMCC::GE;
1064 case ISD::SETLT: return ARMCC::LT;
1065 case ISD::SETLE: return ARMCC::LE;
1066 case ISD::SETUGT: return ARMCC::HI;
1067 case ISD::SETUGE: return ARMCC::HS;
1068 case ISD::SETULT: return ARMCC::LO;
1069 case ISD::SETULE: return ARMCC::LS;
1073 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1074 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1075 ARMCC::CondCodes &CondCode2) {
1076 CondCode2 = ARMCC::AL;
1078 default: llvm_unreachable("Unknown FP condition!");
1080 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1082 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1084 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1085 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1086 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1087 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1088 case ISD::SETO: CondCode = ARMCC::VC; break;
1089 case ISD::SETUO: CondCode = ARMCC::VS; break;
1090 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1091 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1092 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1094 case ISD::SETULT: CondCode = ARMCC::LT; break;
1096 case ISD::SETULE: CondCode = ARMCC::LE; break;
1098 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1102 //===----------------------------------------------------------------------===//
1103 // Calling Convention Implementation
1104 //===----------------------------------------------------------------------===//
1106 #include "ARMGenCallingConv.inc"
1108 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1109 /// given CallingConvention value.
1110 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1112 bool isVarArg) const {
1115 llvm_unreachable("Unsupported calling convention");
1116 case CallingConv::Fast:
1117 if (Subtarget->hasVFP2() && !isVarArg) {
1118 if (!Subtarget->isAAPCS_ABI())
1119 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1120 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1121 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1124 case CallingConv::C: {
1125 // Use target triple & subtarget features to do actual dispatch.
1126 if (!Subtarget->isAAPCS_ABI())
1127 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1128 else if (Subtarget->hasVFP2() &&
1129 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1131 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1132 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1134 case CallingConv::ARM_AAPCS_VFP:
1136 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1138 case CallingConv::ARM_AAPCS:
1139 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1140 case CallingConv::ARM_APCS:
1141 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1145 /// LowerCallResult - Lower the result values of a call into the
1146 /// appropriate copies out of appropriate physical registers.
1148 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1149 CallingConv::ID CallConv, bool isVarArg,
1150 const SmallVectorImpl<ISD::InputArg> &Ins,
1151 DebugLoc dl, SelectionDAG &DAG,
1152 SmallVectorImpl<SDValue> &InVals) const {
1154 // Assign locations to each value returned by this call.
1155 SmallVector<CCValAssign, 16> RVLocs;
1156 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1157 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1158 CCInfo.AnalyzeCallResult(Ins,
1159 CCAssignFnForNode(CallConv, /* Return*/ true,
1162 // Copy all of the result registers out of their specified physreg.
1163 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1164 CCValAssign VA = RVLocs[i];
1167 if (VA.needsCustom()) {
1168 // Handle f64 or half of a v2f64.
1169 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1171 Chain = Lo.getValue(1);
1172 InFlag = Lo.getValue(2);
1173 VA = RVLocs[++i]; // skip ahead to next loc
1174 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1176 Chain = Hi.getValue(1);
1177 InFlag = Hi.getValue(2);
1178 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1180 if (VA.getLocVT() == MVT::v2f64) {
1181 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1182 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1183 DAG.getConstant(0, MVT::i32));
1185 VA = RVLocs[++i]; // skip ahead to next loc
1186 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1187 Chain = Lo.getValue(1);
1188 InFlag = Lo.getValue(2);
1189 VA = RVLocs[++i]; // skip ahead to next loc
1190 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1191 Chain = Hi.getValue(1);
1192 InFlag = Hi.getValue(2);
1193 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1194 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1195 DAG.getConstant(1, MVT::i32));
1198 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1200 Chain = Val.getValue(1);
1201 InFlag = Val.getValue(2);
1204 switch (VA.getLocInfo()) {
1205 default: llvm_unreachable("Unknown loc info!");
1206 case CCValAssign::Full: break;
1207 case CCValAssign::BCvt:
1208 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1212 InVals.push_back(Val);
1218 /// LowerMemOpCallTo - Store the argument to the stack.
1220 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1221 SDValue StackPtr, SDValue Arg,
1222 DebugLoc dl, SelectionDAG &DAG,
1223 const CCValAssign &VA,
1224 ISD::ArgFlagsTy Flags) const {
1225 unsigned LocMemOffset = VA.getLocMemOffset();
1226 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1227 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1228 return DAG.getStore(Chain, dl, Arg, PtrOff,
1229 MachinePointerInfo::getStack(LocMemOffset),
1233 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1234 SDValue Chain, SDValue &Arg,
1235 RegsToPassVector &RegsToPass,
1236 CCValAssign &VA, CCValAssign &NextVA,
1238 SmallVector<SDValue, 8> &MemOpChains,
1239 ISD::ArgFlagsTy Flags) const {
1241 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1242 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1243 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1245 if (NextVA.isRegLoc())
1246 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1248 assert(NextVA.isMemLoc());
1249 if (StackPtr.getNode() == 0)
1250 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1252 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1258 /// LowerCall - Lowering a call into a callseq_start <-
1259 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1262 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1263 CallingConv::ID CallConv, bool isVarArg,
1265 const SmallVectorImpl<ISD::OutputArg> &Outs,
1266 const SmallVectorImpl<SDValue> &OutVals,
1267 const SmallVectorImpl<ISD::InputArg> &Ins,
1268 DebugLoc dl, SelectionDAG &DAG,
1269 SmallVectorImpl<SDValue> &InVals) const {
1270 MachineFunction &MF = DAG.getMachineFunction();
1271 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1272 bool IsSibCall = false;
1273 // Disable tail calls if they're not supported.
1274 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1277 // Check if it's really possible to do a tail call.
1278 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1279 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1280 Outs, OutVals, Ins, DAG);
1281 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1282 // detected sibcalls.
1289 // Analyze operands of the call, assigning locations to each operand.
1290 SmallVector<CCValAssign, 16> ArgLocs;
1291 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1292 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1293 CCInfo.AnalyzeCallOperands(Outs,
1294 CCAssignFnForNode(CallConv, /* Return*/ false,
1297 // Get a count of how many bytes are to be pushed on the stack.
1298 unsigned NumBytes = CCInfo.getNextStackOffset();
1300 // For tail calls, memory operands are available in our caller's stack.
1304 // Adjust the stack pointer for the new arguments...
1305 // These operations are automatically eliminated by the prolog/epilog pass
1307 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1309 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1311 RegsToPassVector RegsToPass;
1312 SmallVector<SDValue, 8> MemOpChains;
1314 // Walk the register/memloc assignments, inserting copies/loads. In the case
1315 // of tail call optimization, arguments are handled later.
1316 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1318 ++i, ++realArgIdx) {
1319 CCValAssign &VA = ArgLocs[i];
1320 SDValue Arg = OutVals[realArgIdx];
1321 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1322 bool isByVal = Flags.isByVal();
1324 // Promote the value if needed.
1325 switch (VA.getLocInfo()) {
1326 default: llvm_unreachable("Unknown loc info!");
1327 case CCValAssign::Full: break;
1328 case CCValAssign::SExt:
1329 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1331 case CCValAssign::ZExt:
1332 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1334 case CCValAssign::AExt:
1335 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1337 case CCValAssign::BCvt:
1338 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1342 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1343 if (VA.needsCustom()) {
1344 if (VA.getLocVT() == MVT::v2f64) {
1345 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1346 DAG.getConstant(0, MVT::i32));
1347 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1348 DAG.getConstant(1, MVT::i32));
1350 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1351 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1353 VA = ArgLocs[++i]; // skip ahead to next loc
1354 if (VA.isRegLoc()) {
1355 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1356 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1358 assert(VA.isMemLoc());
1360 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1361 dl, DAG, VA, Flags));
1364 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1365 StackPtr, MemOpChains, Flags);
1367 } else if (VA.isRegLoc()) {
1368 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1369 } else if (isByVal) {
1370 assert(VA.isMemLoc());
1371 unsigned offset = 0;
1373 // True if this byval aggregate will be split between registers
1375 if (CCInfo.isFirstByValRegValid()) {
1376 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1378 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1379 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1380 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1381 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1382 MachinePointerInfo(),
1383 false, false, false, 0);
1384 MemOpChains.push_back(Load.getValue(1));
1385 RegsToPass.push_back(std::make_pair(j, Load));
1387 offset = ARM::R4 - CCInfo.getFirstByValReg();
1388 CCInfo.clearFirstByValReg();
1391 unsigned LocMemOffset = VA.getLocMemOffset();
1392 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1393 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1395 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1396 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1397 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1399 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1400 Flags.getByValAlign(),
1401 /*isVolatile=*/false,
1402 /*AlwaysInline=*/false,
1403 MachinePointerInfo(0),
1404 MachinePointerInfo(0)));
1406 } else if (!IsSibCall) {
1407 assert(VA.isMemLoc());
1409 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1410 dl, DAG, VA, Flags));
1414 if (!MemOpChains.empty())
1415 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1416 &MemOpChains[0], MemOpChains.size());
1418 // Build a sequence of copy-to-reg nodes chained together with token chain
1419 // and flag operands which copy the outgoing args into the appropriate regs.
1421 // Tail call byval lowering might overwrite argument registers so in case of
1422 // tail call optimization the copies to registers are lowered later.
1424 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1425 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1426 RegsToPass[i].second, InFlag);
1427 InFlag = Chain.getValue(1);
1430 // For tail calls lower the arguments to the 'real' stack slot.
1432 // Force all the incoming stack arguments to be loaded from the stack
1433 // before any new outgoing arguments are stored to the stack, because the
1434 // outgoing stack slots may alias the incoming argument stack slots, and
1435 // the alias isn't otherwise explicit. This is slightly more conservative
1436 // than necessary, because it means that each store effectively depends
1437 // on every argument instead of just those arguments it would clobber.
1439 // Do not flag preceding copytoreg stuff together with the following stuff.
1441 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1442 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1443 RegsToPass[i].second, InFlag);
1444 InFlag = Chain.getValue(1);
1449 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1450 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1451 // node so that legalize doesn't hack it.
1452 bool isDirect = false;
1453 bool isARMFunc = false;
1454 bool isLocalARMFunc = false;
1455 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1457 if (EnableARMLongCalls) {
1458 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1459 && "long-calls with non-static relocation model!");
1460 // Handle a global address or an external symbol. If it's not one of
1461 // those, the target's already in a register, so we don't need to do
1463 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1464 const GlobalValue *GV = G->getGlobal();
1465 // Create a constant pool entry for the callee address
1466 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1467 ARMConstantPoolValue *CPV =
1468 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1470 // Get the address of the callee into a register
1471 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1472 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1473 Callee = DAG.getLoad(getPointerTy(), dl,
1474 DAG.getEntryNode(), CPAddr,
1475 MachinePointerInfo::getConstantPool(),
1476 false, false, false, 0);
1477 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1478 const char *Sym = S->getSymbol();
1480 // Create a constant pool entry for the callee address
1481 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1482 ARMConstantPoolValue *CPV =
1483 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1484 ARMPCLabelIndex, 0);
1485 // Get the address of the callee into a register
1486 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1487 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1488 Callee = DAG.getLoad(getPointerTy(), dl,
1489 DAG.getEntryNode(), CPAddr,
1490 MachinePointerInfo::getConstantPool(),
1491 false, false, false, 0);
1493 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1494 const GlobalValue *GV = G->getGlobal();
1496 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1497 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1498 getTargetMachine().getRelocationModel() != Reloc::Static;
1499 isARMFunc = !Subtarget->isThumb() || isStub;
1500 // ARM call to a local ARM function is predicable.
1501 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1502 // tBX takes a register source operand.
1503 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1504 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1505 ARMConstantPoolValue *CPV =
1506 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1507 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1508 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1509 Callee = DAG.getLoad(getPointerTy(), dl,
1510 DAG.getEntryNode(), CPAddr,
1511 MachinePointerInfo::getConstantPool(),
1512 false, false, false, 0);
1513 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1514 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1515 getPointerTy(), Callee, PICLabel);
1517 // On ELF targets for PIC code, direct calls should go through the PLT
1518 unsigned OpFlags = 0;
1519 if (Subtarget->isTargetELF() &&
1520 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1521 OpFlags = ARMII::MO_PLT;
1522 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1524 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1526 bool isStub = Subtarget->isTargetDarwin() &&
1527 getTargetMachine().getRelocationModel() != Reloc::Static;
1528 isARMFunc = !Subtarget->isThumb() || isStub;
1529 // tBX takes a register source operand.
1530 const char *Sym = S->getSymbol();
1531 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1532 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1533 ARMConstantPoolValue *CPV =
1534 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1535 ARMPCLabelIndex, 4);
1536 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1537 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1538 Callee = DAG.getLoad(getPointerTy(), dl,
1539 DAG.getEntryNode(), CPAddr,
1540 MachinePointerInfo::getConstantPool(),
1541 false, false, false, 0);
1542 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1543 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1544 getPointerTy(), Callee, PICLabel);
1546 unsigned OpFlags = 0;
1547 // On ELF targets for PIC code, direct calls should go through the PLT
1548 if (Subtarget->isTargetELF() &&
1549 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1550 OpFlags = ARMII::MO_PLT;
1551 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1555 // FIXME: handle tail calls differently.
1557 if (Subtarget->isThumb()) {
1558 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1559 CallOpc = ARMISD::CALL_NOLINK;
1561 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1563 CallOpc = (isDirect || Subtarget->hasV5TOps())
1564 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1565 : ARMISD::CALL_NOLINK;
1568 std::vector<SDValue> Ops;
1569 Ops.push_back(Chain);
1570 Ops.push_back(Callee);
1572 // Add argument registers to the end of the list so that they are known live
1574 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1575 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1576 RegsToPass[i].second.getValueType()));
1578 if (InFlag.getNode())
1579 Ops.push_back(InFlag);
1581 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1583 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1585 // Returns a chain and a flag for retval copy to use.
1586 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1587 InFlag = Chain.getValue(1);
1589 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1590 DAG.getIntPtrConstant(0, true), InFlag);
1592 InFlag = Chain.getValue(1);
1594 // Handle result values, copying them out of physregs into vregs that we
1596 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1600 /// HandleByVal - Every parameter *after* a byval parameter is passed
1601 /// on the stack. Remember the next parameter register to allocate,
1602 /// and then confiscate the rest of the parameter registers to insure
1605 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1606 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1607 assert((State->getCallOrPrologue() == Prologue ||
1608 State->getCallOrPrologue() == Call) &&
1609 "unhandled ParmContext");
1610 if ((!State->isFirstByValRegValid()) &&
1611 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1612 State->setFirstByValReg(reg);
1613 // At a call site, a byval parameter that is split between
1614 // registers and memory needs its size truncated here. In a
1615 // function prologue, such byval parameters are reassembled in
1616 // memory, and are not truncated.
1617 if (State->getCallOrPrologue() == Call) {
1618 unsigned excess = 4 * (ARM::R4 - reg);
1619 assert(size >= excess && "expected larger existing stack allocation");
1623 // Confiscate any remaining parameter registers to preclude their
1624 // assignment to subsequent parameters.
1625 while (State->AllocateReg(GPRArgRegs, 4))
1629 /// MatchingStackOffset - Return true if the given stack call argument is
1630 /// already available in the same position (relatively) of the caller's
1631 /// incoming argument stack.
1633 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1634 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1635 const ARMInstrInfo *TII) {
1636 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1638 if (Arg.getOpcode() == ISD::CopyFromReg) {
1639 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1640 if (!TargetRegisterInfo::isVirtualRegister(VR))
1642 MachineInstr *Def = MRI->getVRegDef(VR);
1645 if (!Flags.isByVal()) {
1646 if (!TII->isLoadFromStackSlot(Def, FI))
1651 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1652 if (Flags.isByVal())
1653 // ByVal argument is passed in as a pointer but it's now being
1654 // dereferenced. e.g.
1655 // define @foo(%struct.X* %A) {
1656 // tail call @bar(%struct.X* byval %A)
1659 SDValue Ptr = Ld->getBasePtr();
1660 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1663 FI = FINode->getIndex();
1667 assert(FI != INT_MAX);
1668 if (!MFI->isFixedObjectIndex(FI))
1670 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1673 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1674 /// for tail call optimization. Targets which want to do tail call
1675 /// optimization should implement this function.
1677 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1678 CallingConv::ID CalleeCC,
1680 bool isCalleeStructRet,
1681 bool isCallerStructRet,
1682 const SmallVectorImpl<ISD::OutputArg> &Outs,
1683 const SmallVectorImpl<SDValue> &OutVals,
1684 const SmallVectorImpl<ISD::InputArg> &Ins,
1685 SelectionDAG& DAG) const {
1686 const Function *CallerF = DAG.getMachineFunction().getFunction();
1687 CallingConv::ID CallerCC = CallerF->getCallingConv();
1688 bool CCMatch = CallerCC == CalleeCC;
1690 // Look for obvious safe cases to perform tail call optimization that do not
1691 // require ABI changes. This is what gcc calls sibcall.
1693 // Do not sibcall optimize vararg calls unless the call site is not passing
1695 if (isVarArg && !Outs.empty())
1698 // Also avoid sibcall optimization if either caller or callee uses struct
1699 // return semantics.
1700 if (isCalleeStructRet || isCallerStructRet)
1703 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1704 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1705 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1706 // support in the assembler and linker to be used. This would need to be
1707 // fixed to fully support tail calls in Thumb1.
1709 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1710 // LR. This means if we need to reload LR, it takes an extra instructions,
1711 // which outweighs the value of the tail call; but here we don't know yet
1712 // whether LR is going to be used. Probably the right approach is to
1713 // generate the tail call here and turn it back into CALL/RET in
1714 // emitEpilogue if LR is used.
1716 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1717 // but we need to make sure there are enough registers; the only valid
1718 // registers are the 4 used for parameters. We don't currently do this
1720 if (Subtarget->isThumb1Only())
1723 // If the calling conventions do not match, then we'd better make sure the
1724 // results are returned in the same way as what the caller expects.
1726 SmallVector<CCValAssign, 16> RVLocs1;
1727 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1728 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1729 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1731 SmallVector<CCValAssign, 16> RVLocs2;
1732 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1733 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1734 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1736 if (RVLocs1.size() != RVLocs2.size())
1738 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1739 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1741 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1743 if (RVLocs1[i].isRegLoc()) {
1744 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1747 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1753 // If the callee takes no arguments then go on to check the results of the
1755 if (!Outs.empty()) {
1756 // Check if stack adjustment is needed. For now, do not do this if any
1757 // argument is passed on the stack.
1758 SmallVector<CCValAssign, 16> ArgLocs;
1759 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1760 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1761 CCInfo.AnalyzeCallOperands(Outs,
1762 CCAssignFnForNode(CalleeCC, false, isVarArg));
1763 if (CCInfo.getNextStackOffset()) {
1764 MachineFunction &MF = DAG.getMachineFunction();
1766 // Check if the arguments are already laid out in the right way as
1767 // the caller's fixed stack objects.
1768 MachineFrameInfo *MFI = MF.getFrameInfo();
1769 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1770 const ARMInstrInfo *TII =
1771 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1772 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1774 ++i, ++realArgIdx) {
1775 CCValAssign &VA = ArgLocs[i];
1776 EVT RegVT = VA.getLocVT();
1777 SDValue Arg = OutVals[realArgIdx];
1778 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1779 if (VA.getLocInfo() == CCValAssign::Indirect)
1781 if (VA.needsCustom()) {
1782 // f64 and vector types are split into multiple registers or
1783 // register/stack-slot combinations. The types will not match
1784 // the registers; give up on memory f64 refs until we figure
1785 // out what to do about this.
1788 if (!ArgLocs[++i].isRegLoc())
1790 if (RegVT == MVT::v2f64) {
1791 if (!ArgLocs[++i].isRegLoc())
1793 if (!ArgLocs[++i].isRegLoc())
1796 } else if (!VA.isRegLoc()) {
1797 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1809 ARMTargetLowering::LowerReturn(SDValue Chain,
1810 CallingConv::ID CallConv, bool isVarArg,
1811 const SmallVectorImpl<ISD::OutputArg> &Outs,
1812 const SmallVectorImpl<SDValue> &OutVals,
1813 DebugLoc dl, SelectionDAG &DAG) const {
1815 // CCValAssign - represent the assignment of the return value to a location.
1816 SmallVector<CCValAssign, 16> RVLocs;
1818 // CCState - Info about the registers and stack slots.
1819 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1820 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1822 // Analyze outgoing return values.
1823 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1826 // If this is the first return lowered for this function, add
1827 // the regs to the liveout set for the function.
1828 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1829 for (unsigned i = 0; i != RVLocs.size(); ++i)
1830 if (RVLocs[i].isRegLoc())
1831 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1836 // Copy the result values into the output registers.
1837 for (unsigned i = 0, realRVLocIdx = 0;
1839 ++i, ++realRVLocIdx) {
1840 CCValAssign &VA = RVLocs[i];
1841 assert(VA.isRegLoc() && "Can only return in registers!");
1843 SDValue Arg = OutVals[realRVLocIdx];
1845 switch (VA.getLocInfo()) {
1846 default: llvm_unreachable("Unknown loc info!");
1847 case CCValAssign::Full: break;
1848 case CCValAssign::BCvt:
1849 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1853 if (VA.needsCustom()) {
1854 if (VA.getLocVT() == MVT::v2f64) {
1855 // Extract the first half and return it in two registers.
1856 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1857 DAG.getConstant(0, MVT::i32));
1858 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1859 DAG.getVTList(MVT::i32, MVT::i32), Half);
1861 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1862 Flag = Chain.getValue(1);
1863 VA = RVLocs[++i]; // skip ahead to next loc
1864 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1865 HalfGPRs.getValue(1), Flag);
1866 Flag = Chain.getValue(1);
1867 VA = RVLocs[++i]; // skip ahead to next loc
1869 // Extract the 2nd half and fall through to handle it as an f64 value.
1870 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1871 DAG.getConstant(1, MVT::i32));
1873 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1875 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1876 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1877 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1878 Flag = Chain.getValue(1);
1879 VA = RVLocs[++i]; // skip ahead to next loc
1880 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1883 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1885 // Guarantee that all emitted copies are
1886 // stuck together, avoiding something bad.
1887 Flag = Chain.getValue(1);
1892 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1894 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1899 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1900 if (N->getNumValues() != 1)
1902 if (!N->hasNUsesOfValue(1, 0))
1905 unsigned NumCopies = 0;
1907 SDNode *Use = *N->use_begin();
1908 if (Use->getOpcode() == ISD::CopyToReg) {
1909 Copies[NumCopies++] = Use;
1910 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1911 // f64 returned in a pair of GPRs.
1912 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1914 if (UI->getOpcode() != ISD::CopyToReg)
1916 Copies[UI.getUse().getResNo()] = *UI;
1919 } else if (Use->getOpcode() == ISD::BITCAST) {
1920 // f32 returned in a single GPR.
1921 if (!Use->hasNUsesOfValue(1, 0))
1923 Use = *Use->use_begin();
1924 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1926 Copies[NumCopies++] = Use;
1931 if (NumCopies != 1 && NumCopies != 2)
1934 bool HasRet = false;
1935 for (unsigned i = 0; i < NumCopies; ++i) {
1936 SDNode *Copy = Copies[i];
1937 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1939 if (UI->getOpcode() == ISD::CopyToReg) {
1941 if (Use == Copies[0] || Use == Copies[1])
1945 if (UI->getOpcode() != ARMISD::RET_FLAG)
1954 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1955 if (!EnableARMTailCalls)
1958 if (!CI->isTailCall())
1961 return !Subtarget->isThumb1Only();
1964 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1965 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1966 // one of the above mentioned nodes. It has to be wrapped because otherwise
1967 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1968 // be used to form addressing mode. These wrapped nodes will be selected
1970 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1971 EVT PtrVT = Op.getValueType();
1972 // FIXME there is no actual debug info here
1973 DebugLoc dl = Op.getDebugLoc();
1974 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1976 if (CP->isMachineConstantPoolEntry())
1977 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1978 CP->getAlignment());
1980 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1981 CP->getAlignment());
1982 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1985 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1986 return MachineJumpTableInfo::EK_Inline;
1989 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1990 SelectionDAG &DAG) const {
1991 MachineFunction &MF = DAG.getMachineFunction();
1992 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1993 unsigned ARMPCLabelIndex = 0;
1994 DebugLoc DL = Op.getDebugLoc();
1995 EVT PtrVT = getPointerTy();
1996 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1997 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1999 if (RelocM == Reloc::Static) {
2000 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2002 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2003 ARMPCLabelIndex = AFI->createPICLabelUId();
2004 ARMConstantPoolValue *CPV =
2005 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2006 ARMCP::CPBlockAddress, PCAdj);
2007 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2009 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2010 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2011 MachinePointerInfo::getConstantPool(),
2012 false, false, false, 0);
2013 if (RelocM == Reloc::Static)
2015 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2016 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2019 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2021 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2022 SelectionDAG &DAG) const {
2023 DebugLoc dl = GA->getDebugLoc();
2024 EVT PtrVT = getPointerTy();
2025 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2026 MachineFunction &MF = DAG.getMachineFunction();
2027 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2028 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2029 ARMConstantPoolValue *CPV =
2030 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2031 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2032 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2033 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2034 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2035 MachinePointerInfo::getConstantPool(),
2036 false, false, false, 0);
2037 SDValue Chain = Argument.getValue(1);
2039 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2040 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2042 // call __tls_get_addr.
2045 Entry.Node = Argument;
2046 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2047 Args.push_back(Entry);
2048 // FIXME: is there useful debug info available here?
2049 std::pair<SDValue, SDValue> CallResult =
2050 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2051 false, false, false, false,
2052 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2053 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2054 return CallResult.first;
2057 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2058 // "local exec" model.
2060 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2061 SelectionDAG &DAG) const {
2062 const GlobalValue *GV = GA->getGlobal();
2063 DebugLoc dl = GA->getDebugLoc();
2065 SDValue Chain = DAG.getEntryNode();
2066 EVT PtrVT = getPointerTy();
2067 // Get the Thread Pointer
2068 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2070 if (GV->isDeclaration()) {
2071 MachineFunction &MF = DAG.getMachineFunction();
2072 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2073 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2074 // Initial exec model.
2075 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2076 ARMConstantPoolValue *CPV =
2077 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2078 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2080 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2081 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2082 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2083 MachinePointerInfo::getConstantPool(),
2084 false, false, false, 0);
2085 Chain = Offset.getValue(1);
2087 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2088 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2090 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2091 MachinePointerInfo::getConstantPool(),
2092 false, false, false, 0);
2095 ARMConstantPoolValue *CPV =
2096 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2097 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2098 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2099 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2100 MachinePointerInfo::getConstantPool(),
2101 false, false, false, 0);
2104 // The address of the thread local variable is the add of the thread
2105 // pointer with the offset of the variable.
2106 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2110 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2111 // TODO: implement the "local dynamic" model
2112 assert(Subtarget->isTargetELF() &&
2113 "TLS not implemented for non-ELF targets");
2114 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2115 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2116 // otherwise use the "Local Exec" TLS Model
2117 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2118 return LowerToTLSGeneralDynamicModel(GA, DAG);
2120 return LowerToTLSExecModels(GA, DAG);
2123 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2124 SelectionDAG &DAG) const {
2125 EVT PtrVT = getPointerTy();
2126 DebugLoc dl = Op.getDebugLoc();
2127 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2128 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2129 if (RelocM == Reloc::PIC_) {
2130 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2131 ARMConstantPoolValue *CPV =
2132 ARMConstantPoolConstant::Create(GV,
2133 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2134 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2135 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2136 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2138 MachinePointerInfo::getConstantPool(),
2139 false, false, false, 0);
2140 SDValue Chain = Result.getValue(1);
2141 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2142 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2144 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2145 MachinePointerInfo::getGOT(),
2146 false, false, false, 0);
2150 // If we have T2 ops, we can materialize the address directly via movt/movw
2151 // pair. This is always cheaper.
2152 if (Subtarget->useMovt()) {
2154 // FIXME: Once remat is capable of dealing with instructions with register
2155 // operands, expand this into two nodes.
2156 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2157 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2159 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2160 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2161 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2162 MachinePointerInfo::getConstantPool(),
2163 false, false, false, 0);
2167 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2168 SelectionDAG &DAG) const {
2169 EVT PtrVT = getPointerTy();
2170 DebugLoc dl = Op.getDebugLoc();
2171 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2172 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2173 MachineFunction &MF = DAG.getMachineFunction();
2174 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2176 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2177 // update ARMFastISel::ARMMaterializeGV.
2178 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2180 // FIXME: Once remat is capable of dealing with instructions with register
2181 // operands, expand this into two nodes.
2182 if (RelocM == Reloc::Static)
2183 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2184 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2186 unsigned Wrapper = (RelocM == Reloc::PIC_)
2187 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2188 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2189 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2190 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2191 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2192 MachinePointerInfo::getGOT(),
2193 false, false, false, 0);
2197 unsigned ARMPCLabelIndex = 0;
2199 if (RelocM == Reloc::Static) {
2200 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2202 ARMPCLabelIndex = AFI->createPICLabelUId();
2203 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2204 ARMConstantPoolValue *CPV =
2205 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2207 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2209 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2211 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2212 MachinePointerInfo::getConstantPool(),
2213 false, false, false, 0);
2214 SDValue Chain = Result.getValue(1);
2216 if (RelocM == Reloc::PIC_) {
2217 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2218 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2221 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2222 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2223 false, false, false, 0);
2228 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2229 SelectionDAG &DAG) const {
2230 assert(Subtarget->isTargetELF() &&
2231 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2232 MachineFunction &MF = DAG.getMachineFunction();
2233 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2234 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2235 EVT PtrVT = getPointerTy();
2236 DebugLoc dl = Op.getDebugLoc();
2237 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2238 ARMConstantPoolValue *CPV =
2239 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2240 ARMPCLabelIndex, PCAdj);
2241 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2242 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2243 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2244 MachinePointerInfo::getConstantPool(),
2245 false, false, false, 0);
2246 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2247 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2251 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2252 DebugLoc dl = Op.getDebugLoc();
2253 SDValue Val = DAG.getConstant(0, MVT::i32);
2254 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2255 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2256 Op.getOperand(1), Val);
2260 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2261 DebugLoc dl = Op.getDebugLoc();
2262 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2263 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2267 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2268 const ARMSubtarget *Subtarget) const {
2269 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2270 DebugLoc dl = Op.getDebugLoc();
2272 default: return SDValue(); // Don't custom lower most intrinsics.
2273 case Intrinsic::arm_thread_pointer: {
2274 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2275 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2277 case Intrinsic::eh_sjlj_lsda: {
2278 MachineFunction &MF = DAG.getMachineFunction();
2279 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2280 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2281 EVT PtrVT = getPointerTy();
2282 DebugLoc dl = Op.getDebugLoc();
2283 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2285 unsigned PCAdj = (RelocM != Reloc::PIC_)
2286 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2287 ARMConstantPoolValue *CPV =
2288 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2289 ARMCP::CPLSDA, PCAdj);
2290 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2291 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2293 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2294 MachinePointerInfo::getConstantPool(),
2295 false, false, false, 0);
2297 if (RelocM == Reloc::PIC_) {
2298 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2299 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2303 case Intrinsic::arm_neon_vmulls:
2304 case Intrinsic::arm_neon_vmullu: {
2305 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2306 ? ARMISD::VMULLs : ARMISD::VMULLu;
2307 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2308 Op.getOperand(1), Op.getOperand(2));
2313 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2314 const ARMSubtarget *Subtarget) {
2315 DebugLoc dl = Op.getDebugLoc();
2316 if (!Subtarget->hasDataBarrier()) {
2317 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2318 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2320 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2321 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2322 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2323 DAG.getConstant(0, MVT::i32));
2326 SDValue Op5 = Op.getOperand(5);
2327 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2328 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2329 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2330 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2332 ARM_MB::MemBOpt DMBOpt;
2333 if (isDeviceBarrier)
2334 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2336 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2337 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2338 DAG.getConstant(DMBOpt, MVT::i32));
2342 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2343 const ARMSubtarget *Subtarget) {
2344 // FIXME: handle "fence singlethread" more efficiently.
2345 DebugLoc dl = Op.getDebugLoc();
2346 if (!Subtarget->hasDataBarrier()) {
2347 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2348 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2350 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2351 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2352 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2353 DAG.getConstant(0, MVT::i32));
2356 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2357 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2360 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2361 const ARMSubtarget *Subtarget) {
2362 // ARM pre v5TE and Thumb1 does not have preload instructions.
2363 if (!(Subtarget->isThumb2() ||
2364 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2365 // Just preserve the chain.
2366 return Op.getOperand(0);
2368 DebugLoc dl = Op.getDebugLoc();
2369 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2371 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2372 // ARMv7 with MP extension has PLDW.
2373 return Op.getOperand(0);
2375 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2376 if (Subtarget->isThumb()) {
2378 isRead = ~isRead & 1;
2379 isData = ~isData & 1;
2382 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2383 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2384 DAG.getConstant(isData, MVT::i32));
2387 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2388 MachineFunction &MF = DAG.getMachineFunction();
2389 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2391 // vastart just stores the address of the VarArgsFrameIndex slot into the
2392 // memory location argument.
2393 DebugLoc dl = Op.getDebugLoc();
2394 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2395 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2396 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2397 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2398 MachinePointerInfo(SV), false, false, 0);
2402 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2403 SDValue &Root, SelectionDAG &DAG,
2404 DebugLoc dl) const {
2405 MachineFunction &MF = DAG.getMachineFunction();
2406 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2408 TargetRegisterClass *RC;
2409 if (AFI->isThumb1OnlyFunction())
2410 RC = ARM::tGPRRegisterClass;
2412 RC = ARM::GPRRegisterClass;
2414 // Transform the arguments stored in physical registers into virtual ones.
2415 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2416 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2419 if (NextVA.isMemLoc()) {
2420 MachineFrameInfo *MFI = MF.getFrameInfo();
2421 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2423 // Create load node to retrieve arguments from the stack.
2424 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2425 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2426 MachinePointerInfo::getFixedStack(FI),
2427 false, false, false, 0);
2429 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2430 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2433 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2437 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2438 unsigned &VARegSize, unsigned &VARegSaveSize)
2441 if (CCInfo.isFirstByValRegValid())
2442 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2444 unsigned int firstUnalloced;
2445 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2446 sizeof(GPRArgRegs) /
2447 sizeof(GPRArgRegs[0]));
2448 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2451 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2452 VARegSize = NumGPRs * 4;
2453 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2456 // The remaining GPRs hold either the beginning of variable-argument
2457 // data, or the beginning of an aggregate passed by value (usuall
2458 // byval). Either way, we allocate stack slots adjacent to the data
2459 // provided by our caller, and store the unallocated registers there.
2460 // If this is a variadic function, the va_list pointer will begin with
2461 // these values; otherwise, this reassembles a (byval) structure that
2462 // was split between registers and memory.
2464 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2465 DebugLoc dl, SDValue &Chain,
2466 unsigned ArgOffset) const {
2467 MachineFunction &MF = DAG.getMachineFunction();
2468 MachineFrameInfo *MFI = MF.getFrameInfo();
2469 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2470 unsigned firstRegToSaveIndex;
2471 if (CCInfo.isFirstByValRegValid())
2472 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2474 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2475 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2478 unsigned VARegSize, VARegSaveSize;
2479 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2480 if (VARegSaveSize) {
2481 // If this function is vararg, store any remaining integer argument regs
2482 // to their spots on the stack so that they may be loaded by deferencing
2483 // the result of va_next.
2484 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2485 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2486 ArgOffset + VARegSaveSize
2489 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2492 SmallVector<SDValue, 4> MemOps;
2493 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2494 TargetRegisterClass *RC;
2495 if (AFI->isThumb1OnlyFunction())
2496 RC = ARM::tGPRRegisterClass;
2498 RC = ARM::GPRRegisterClass;
2500 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2501 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2503 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2504 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2506 MemOps.push_back(Store);
2507 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2508 DAG.getConstant(4, getPointerTy()));
2510 if (!MemOps.empty())
2511 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2512 &MemOps[0], MemOps.size());
2514 // This will point to the next argument passed via stack.
2515 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2519 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2520 CallingConv::ID CallConv, bool isVarArg,
2521 const SmallVectorImpl<ISD::InputArg>
2523 DebugLoc dl, SelectionDAG &DAG,
2524 SmallVectorImpl<SDValue> &InVals)
2526 MachineFunction &MF = DAG.getMachineFunction();
2527 MachineFrameInfo *MFI = MF.getFrameInfo();
2529 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2531 // Assign locations to all of the incoming arguments.
2532 SmallVector<CCValAssign, 16> ArgLocs;
2533 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2534 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2535 CCInfo.AnalyzeFormalArguments(Ins,
2536 CCAssignFnForNode(CallConv, /* Return*/ false,
2539 SmallVector<SDValue, 16> ArgValues;
2540 int lastInsIndex = -1;
2543 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2544 CCValAssign &VA = ArgLocs[i];
2546 // Arguments stored in registers.
2547 if (VA.isRegLoc()) {
2548 EVT RegVT = VA.getLocVT();
2550 if (VA.needsCustom()) {
2551 // f64 and vector types are split up into multiple registers or
2552 // combinations of registers and stack slots.
2553 if (VA.getLocVT() == MVT::v2f64) {
2554 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2556 VA = ArgLocs[++i]; // skip ahead to next loc
2558 if (VA.isMemLoc()) {
2559 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2560 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2561 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2562 MachinePointerInfo::getFixedStack(FI),
2563 false, false, false, 0);
2565 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2568 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2569 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2570 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2571 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2572 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2574 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2577 TargetRegisterClass *RC;
2579 if (RegVT == MVT::f32)
2580 RC = ARM::SPRRegisterClass;
2581 else if (RegVT == MVT::f64)
2582 RC = ARM::DPRRegisterClass;
2583 else if (RegVT == MVT::v2f64)
2584 RC = ARM::QPRRegisterClass;
2585 else if (RegVT == MVT::i32)
2586 RC = (AFI->isThumb1OnlyFunction() ?
2587 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2589 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2591 // Transform the arguments in physical registers into virtual ones.
2592 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2593 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2596 // If this is an 8 or 16-bit value, it is really passed promoted
2597 // to 32 bits. Insert an assert[sz]ext to capture this, then
2598 // truncate to the right size.
2599 switch (VA.getLocInfo()) {
2600 default: llvm_unreachable("Unknown loc info!");
2601 case CCValAssign::Full: break;
2602 case CCValAssign::BCvt:
2603 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2605 case CCValAssign::SExt:
2606 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2607 DAG.getValueType(VA.getValVT()));
2608 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2610 case CCValAssign::ZExt:
2611 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2612 DAG.getValueType(VA.getValVT()));
2613 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2617 InVals.push_back(ArgValue);
2619 } else { // VA.isRegLoc()
2622 assert(VA.isMemLoc());
2623 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2625 int index = ArgLocs[i].getValNo();
2627 // Some Ins[] entries become multiple ArgLoc[] entries.
2628 // Process them only once.
2629 if (index != lastInsIndex)
2631 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2632 // FIXME: For now, all byval parameter objects are marked mutable.
2633 // This can be changed with more analysis.
2634 // In case of tail call optimization mark all arguments mutable.
2635 // Since they could be overwritten by lowering of arguments in case of
2637 if (Flags.isByVal()) {
2638 unsigned VARegSize, VARegSaveSize;
2639 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2640 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2641 unsigned Bytes = Flags.getByValSize() - VARegSize;
2642 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2643 int FI = MFI->CreateFixedObject(Bytes,
2644 VA.getLocMemOffset(), false);
2645 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2647 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2648 VA.getLocMemOffset(), true);
2650 // Create load nodes to retrieve arguments from the stack.
2651 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2652 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2653 MachinePointerInfo::getFixedStack(FI),
2654 false, false, false, 0));
2656 lastInsIndex = index;
2663 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2668 /// isFloatingPointZero - Return true if this is +0.0.
2669 static bool isFloatingPointZero(SDValue Op) {
2670 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2671 return CFP->getValueAPF().isPosZero();
2672 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2673 // Maybe this has already been legalized into the constant pool?
2674 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2675 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2676 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2677 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2678 return CFP->getValueAPF().isPosZero();
2684 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2685 /// the given operands.
2687 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2688 SDValue &ARMcc, SelectionDAG &DAG,
2689 DebugLoc dl) const {
2690 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2691 unsigned C = RHSC->getZExtValue();
2692 if (!isLegalICmpImmediate(C)) {
2693 // Constant does not fit, try adjusting it by one?
2698 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2699 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2700 RHS = DAG.getConstant(C-1, MVT::i32);
2705 if (C != 0 && isLegalICmpImmediate(C-1)) {
2706 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2707 RHS = DAG.getConstant(C-1, MVT::i32);
2712 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2713 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2714 RHS = DAG.getConstant(C+1, MVT::i32);
2719 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2720 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2721 RHS = DAG.getConstant(C+1, MVT::i32);
2728 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2729 ARMISD::NodeType CompareType;
2732 CompareType = ARMISD::CMP;
2737 CompareType = ARMISD::CMPZ;
2740 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2741 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2744 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2746 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2747 DebugLoc dl) const {
2749 if (!isFloatingPointZero(RHS))
2750 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2752 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2753 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2756 /// duplicateCmp - Glue values can have only one use, so this function
2757 /// duplicates a comparison node.
2759 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2760 unsigned Opc = Cmp.getOpcode();
2761 DebugLoc DL = Cmp.getDebugLoc();
2762 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2763 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2765 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2766 Cmp = Cmp.getOperand(0);
2767 Opc = Cmp.getOpcode();
2768 if (Opc == ARMISD::CMPFP)
2769 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2771 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2772 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2774 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2777 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2778 SDValue Cond = Op.getOperand(0);
2779 SDValue SelectTrue = Op.getOperand(1);
2780 SDValue SelectFalse = Op.getOperand(2);
2781 DebugLoc dl = Op.getDebugLoc();
2785 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2786 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2788 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2789 const ConstantSDNode *CMOVTrue =
2790 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2791 const ConstantSDNode *CMOVFalse =
2792 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2794 if (CMOVTrue && CMOVFalse) {
2795 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2796 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2800 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2802 False = SelectFalse;
2803 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2808 if (True.getNode() && False.getNode()) {
2809 EVT VT = Op.getValueType();
2810 SDValue ARMcc = Cond.getOperand(2);
2811 SDValue CCR = Cond.getOperand(3);
2812 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2813 assert(True.getValueType() == VT);
2814 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2819 return DAG.getSelectCC(dl, Cond,
2820 DAG.getConstant(0, Cond.getValueType()),
2821 SelectTrue, SelectFalse, ISD::SETNE);
2824 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2825 EVT VT = Op.getValueType();
2826 SDValue LHS = Op.getOperand(0);
2827 SDValue RHS = Op.getOperand(1);
2828 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2829 SDValue TrueVal = Op.getOperand(2);
2830 SDValue FalseVal = Op.getOperand(3);
2831 DebugLoc dl = Op.getDebugLoc();
2833 if (LHS.getValueType() == MVT::i32) {
2835 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2836 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2837 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2840 ARMCC::CondCodes CondCode, CondCode2;
2841 FPCCToARMCC(CC, CondCode, CondCode2);
2843 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2844 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2845 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2846 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2848 if (CondCode2 != ARMCC::AL) {
2849 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2850 // FIXME: Needs another CMP because flag can have but one use.
2851 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2852 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2853 Result, TrueVal, ARMcc2, CCR, Cmp2);
2858 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2859 /// to morph to an integer compare sequence.
2860 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2861 const ARMSubtarget *Subtarget) {
2862 SDNode *N = Op.getNode();
2863 if (!N->hasOneUse())
2864 // Otherwise it requires moving the value from fp to integer registers.
2866 if (!N->getNumValues())
2868 EVT VT = Op.getValueType();
2869 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2870 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2871 // vmrs are very slow, e.g. cortex-a8.
2874 if (isFloatingPointZero(Op)) {
2878 return ISD::isNormalLoad(N);
2881 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2882 if (isFloatingPointZero(Op))
2883 return DAG.getConstant(0, MVT::i32);
2885 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2886 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2887 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2888 Ld->isVolatile(), Ld->isNonTemporal(),
2889 Ld->isInvariant(), Ld->getAlignment());
2891 llvm_unreachable("Unknown VFP cmp argument!");
2894 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2895 SDValue &RetVal1, SDValue &RetVal2) {
2896 if (isFloatingPointZero(Op)) {
2897 RetVal1 = DAG.getConstant(0, MVT::i32);
2898 RetVal2 = DAG.getConstant(0, MVT::i32);
2902 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2903 SDValue Ptr = Ld->getBasePtr();
2904 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2905 Ld->getChain(), Ptr,
2906 Ld->getPointerInfo(),
2907 Ld->isVolatile(), Ld->isNonTemporal(),
2908 Ld->isInvariant(), Ld->getAlignment());
2910 EVT PtrType = Ptr.getValueType();
2911 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2912 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2913 PtrType, Ptr, DAG.getConstant(4, PtrType));
2914 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2915 Ld->getChain(), NewPtr,
2916 Ld->getPointerInfo().getWithOffset(4),
2917 Ld->isVolatile(), Ld->isNonTemporal(),
2918 Ld->isInvariant(), NewAlign);
2922 llvm_unreachable("Unknown VFP cmp argument!");
2925 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2926 /// f32 and even f64 comparisons to integer ones.
2928 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2929 SDValue Chain = Op.getOperand(0);
2930 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2931 SDValue LHS = Op.getOperand(2);
2932 SDValue RHS = Op.getOperand(3);
2933 SDValue Dest = Op.getOperand(4);
2934 DebugLoc dl = Op.getDebugLoc();
2936 bool SeenZero = false;
2937 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2938 canChangeToInt(RHS, SeenZero, Subtarget) &&
2939 // If one of the operand is zero, it's safe to ignore the NaN case since
2940 // we only care about equality comparisons.
2941 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2942 // If unsafe fp math optimization is enabled and there are no other uses of
2943 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2944 // to an integer comparison.
2945 if (CC == ISD::SETOEQ)
2947 else if (CC == ISD::SETUNE)
2951 if (LHS.getValueType() == MVT::f32) {
2952 LHS = bitcastf32Toi32(LHS, DAG);
2953 RHS = bitcastf32Toi32(RHS, DAG);
2954 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2955 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2956 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2957 Chain, Dest, ARMcc, CCR, Cmp);
2962 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2963 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2964 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2965 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2966 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2967 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2968 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2974 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2975 SDValue Chain = Op.getOperand(0);
2976 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2977 SDValue LHS = Op.getOperand(2);
2978 SDValue RHS = Op.getOperand(3);
2979 SDValue Dest = Op.getOperand(4);
2980 DebugLoc dl = Op.getDebugLoc();
2982 if (LHS.getValueType() == MVT::i32) {
2984 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2985 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2986 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2987 Chain, Dest, ARMcc, CCR, Cmp);
2990 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2992 if (getTargetMachine().Options.UnsafeFPMath &&
2993 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2994 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2995 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2996 if (Result.getNode())
3000 ARMCC::CondCodes CondCode, CondCode2;
3001 FPCCToARMCC(CC, CondCode, CondCode2);
3003 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3004 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3005 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3006 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3007 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3008 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3009 if (CondCode2 != ARMCC::AL) {
3010 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3011 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3012 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3017 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3018 SDValue Chain = Op.getOperand(0);
3019 SDValue Table = Op.getOperand(1);
3020 SDValue Index = Op.getOperand(2);
3021 DebugLoc dl = Op.getDebugLoc();
3023 EVT PTy = getPointerTy();
3024 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3025 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3026 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3027 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3028 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3029 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3030 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3031 if (Subtarget->isThumb2()) {
3032 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3033 // which does another jump to the destination. This also makes it easier
3034 // to translate it to TBB / TBH later.
3035 // FIXME: This might not work if the function is extremely large.
3036 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3037 Addr, Op.getOperand(2), JTI, UId);
3039 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3040 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3041 MachinePointerInfo::getJumpTable(),
3042 false, false, false, 0);
3043 Chain = Addr.getValue(1);
3044 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3045 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3047 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3048 MachinePointerInfo::getJumpTable(),
3049 false, false, false, 0);
3050 Chain = Addr.getValue(1);
3051 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3055 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3056 assert(Op.getValueType().getVectorElementType() == MVT::i32
3057 && "Unexpected custom lowering");
3059 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3061 return DAG.UnrollVectorOp(Op.getNode());
3064 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3065 EVT VT = Op.getValueType();
3067 return LowerVectorFP_TO_INT(Op, DAG);
3069 DebugLoc dl = Op.getDebugLoc();
3072 switch (Op.getOpcode()) {
3074 assert(0 && "Invalid opcode!");
3075 case ISD::FP_TO_SINT:
3076 Opc = ARMISD::FTOSI;
3078 case ISD::FP_TO_UINT:
3079 Opc = ARMISD::FTOUI;
3082 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3083 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3086 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3087 EVT VT = Op.getValueType();
3088 DebugLoc dl = Op.getDebugLoc();
3090 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3091 if (VT.getVectorElementType() == MVT::f32)
3093 return DAG.UnrollVectorOp(Op.getNode());
3096 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3097 "Invalid type for custom lowering!");
3098 if (VT != MVT::v4f32)
3099 return DAG.UnrollVectorOp(Op.getNode());
3103 switch (Op.getOpcode()) {
3105 assert(0 && "Invalid opcode!");
3106 case ISD::SINT_TO_FP:
3107 CastOpc = ISD::SIGN_EXTEND;
3108 Opc = ISD::SINT_TO_FP;
3110 case ISD::UINT_TO_FP:
3111 CastOpc = ISD::ZERO_EXTEND;
3112 Opc = ISD::UINT_TO_FP;
3116 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3117 return DAG.getNode(Opc, dl, VT, Op);
3120 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3121 EVT VT = Op.getValueType();
3123 return LowerVectorINT_TO_FP(Op, DAG);
3125 DebugLoc dl = Op.getDebugLoc();
3128 switch (Op.getOpcode()) {
3130 assert(0 && "Invalid opcode!");
3131 case ISD::SINT_TO_FP:
3132 Opc = ARMISD::SITOF;
3134 case ISD::UINT_TO_FP:
3135 Opc = ARMISD::UITOF;
3139 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3140 return DAG.getNode(Opc, dl, VT, Op);
3143 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3144 // Implement fcopysign with a fabs and a conditional fneg.
3145 SDValue Tmp0 = Op.getOperand(0);
3146 SDValue Tmp1 = Op.getOperand(1);
3147 DebugLoc dl = Op.getDebugLoc();
3148 EVT VT = Op.getValueType();
3149 EVT SrcVT = Tmp1.getValueType();
3150 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3151 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3152 bool UseNEON = !InGPR && Subtarget->hasNEON();
3155 // Use VBSL to copy the sign bit.
3156 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3157 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3158 DAG.getTargetConstant(EncodedVal, MVT::i32));
3159 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3161 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3162 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3163 DAG.getConstant(32, MVT::i32));
3164 else /*if (VT == MVT::f32)*/
3165 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3166 if (SrcVT == MVT::f32) {
3167 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3169 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3170 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3171 DAG.getConstant(32, MVT::i32));
3172 } else if (VT == MVT::f32)
3173 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3174 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3175 DAG.getConstant(32, MVT::i32));
3176 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3177 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3179 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3181 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3182 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3183 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3185 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3186 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3187 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3188 if (VT == MVT::f32) {
3189 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3190 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3191 DAG.getConstant(0, MVT::i32));
3193 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3199 // Bitcast operand 1 to i32.
3200 if (SrcVT == MVT::f64)
3201 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3202 &Tmp1, 1).getValue(1);
3203 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3205 // Or in the signbit with integer operations.
3206 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3207 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3208 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3209 if (VT == MVT::f32) {
3210 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3211 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3212 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3213 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3216 // f64: Or the high part with signbit and then combine two parts.
3217 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3219 SDValue Lo = Tmp0.getValue(0);
3220 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3221 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3222 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3225 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3226 MachineFunction &MF = DAG.getMachineFunction();
3227 MachineFrameInfo *MFI = MF.getFrameInfo();
3228 MFI->setReturnAddressIsTaken(true);
3230 EVT VT = Op.getValueType();
3231 DebugLoc dl = Op.getDebugLoc();
3232 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3234 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3235 SDValue Offset = DAG.getConstant(4, MVT::i32);
3236 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3237 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3238 MachinePointerInfo(), false, false, false, 0);
3241 // Return LR, which contains the return address. Mark it an implicit live-in.
3242 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3243 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3246 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3247 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3248 MFI->setFrameAddressIsTaken(true);
3250 EVT VT = Op.getValueType();
3251 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3252 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3253 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3254 ? ARM::R7 : ARM::R11;
3255 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3257 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3258 MachinePointerInfo(),
3259 false, false, false, 0);
3263 /// ExpandBITCAST - If the target supports VFP, this function is called to
3264 /// expand a bit convert where either the source or destination type is i64 to
3265 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3266 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3267 /// vectors), since the legalizer won't know what to do with that.
3268 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3269 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3270 DebugLoc dl = N->getDebugLoc();
3271 SDValue Op = N->getOperand(0);
3273 // This function is only supposed to be called for i64 types, either as the
3274 // source or destination of the bit convert.
3275 EVT SrcVT = Op.getValueType();
3276 EVT DstVT = N->getValueType(0);
3277 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3278 "ExpandBITCAST called for non-i64 type");
3280 // Turn i64->f64 into VMOVDRR.
3281 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3282 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3283 DAG.getConstant(0, MVT::i32));
3284 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3285 DAG.getConstant(1, MVT::i32));
3286 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3287 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3290 // Turn f64->i64 into VMOVRRD.
3291 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3292 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3293 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3294 // Merge the pieces into a single i64 value.
3295 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3301 /// getZeroVector - Returns a vector of specified type with all zero elements.
3302 /// Zero vectors are used to represent vector negation and in those cases
3303 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3304 /// not support i64 elements, so sometimes the zero vectors will need to be
3305 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3307 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3308 assert(VT.isVector() && "Expected a vector type");
3309 // The canonical modified immediate encoding of a zero vector is....0!
3310 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3311 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3312 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3313 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3316 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3317 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3318 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3319 SelectionDAG &DAG) const {
3320 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3321 EVT VT = Op.getValueType();
3322 unsigned VTBits = VT.getSizeInBits();
3323 DebugLoc dl = Op.getDebugLoc();
3324 SDValue ShOpLo = Op.getOperand(0);
3325 SDValue ShOpHi = Op.getOperand(1);
3326 SDValue ShAmt = Op.getOperand(2);
3328 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3330 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3332 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3333 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3334 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3335 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3336 DAG.getConstant(VTBits, MVT::i32));
3337 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3338 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3339 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3341 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3342 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3344 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3345 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3348 SDValue Ops[2] = { Lo, Hi };
3349 return DAG.getMergeValues(Ops, 2, dl);
3352 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3353 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3354 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3355 SelectionDAG &DAG) const {
3356 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3357 EVT VT = Op.getValueType();
3358 unsigned VTBits = VT.getSizeInBits();
3359 DebugLoc dl = Op.getDebugLoc();
3360 SDValue ShOpLo = Op.getOperand(0);
3361 SDValue ShOpHi = Op.getOperand(1);
3362 SDValue ShAmt = Op.getOperand(2);
3365 assert(Op.getOpcode() == ISD::SHL_PARTS);
3366 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3367 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3368 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3369 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3370 DAG.getConstant(VTBits, MVT::i32));
3371 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3372 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3374 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3375 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3376 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3378 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3379 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3382 SDValue Ops[2] = { Lo, Hi };
3383 return DAG.getMergeValues(Ops, 2, dl);
3386 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3387 SelectionDAG &DAG) const {
3388 // The rounding mode is in bits 23:22 of the FPSCR.
3389 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3390 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3391 // so that the shift + and get folded into a bitfield extract.
3392 DebugLoc dl = Op.getDebugLoc();
3393 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3394 DAG.getConstant(Intrinsic::arm_get_fpscr,
3396 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3397 DAG.getConstant(1U << 22, MVT::i32));
3398 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3399 DAG.getConstant(22, MVT::i32));
3400 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3401 DAG.getConstant(3, MVT::i32));
3404 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3405 const ARMSubtarget *ST) {
3406 EVT VT = N->getValueType(0);
3407 DebugLoc dl = N->getDebugLoc();
3409 if (!ST->hasV6T2Ops())
3412 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3413 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3416 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3417 const ARMSubtarget *ST) {
3418 EVT VT = N->getValueType(0);
3419 DebugLoc dl = N->getDebugLoc();
3424 // Lower vector shifts on NEON to use VSHL.
3425 assert(ST->hasNEON() && "unexpected vector shift");
3427 // Left shifts translate directly to the vshiftu intrinsic.
3428 if (N->getOpcode() == ISD::SHL)
3429 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3430 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3431 N->getOperand(0), N->getOperand(1));
3433 assert((N->getOpcode() == ISD::SRA ||
3434 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3436 // NEON uses the same intrinsics for both left and right shifts. For
3437 // right shifts, the shift amounts are negative, so negate the vector of
3439 EVT ShiftVT = N->getOperand(1).getValueType();
3440 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3441 getZeroVector(ShiftVT, DAG, dl),
3443 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3444 Intrinsic::arm_neon_vshifts :
3445 Intrinsic::arm_neon_vshiftu);
3446 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3447 DAG.getConstant(vshiftInt, MVT::i32),
3448 N->getOperand(0), NegatedCount);
3451 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3452 const ARMSubtarget *ST) {
3453 EVT VT = N->getValueType(0);
3454 DebugLoc dl = N->getDebugLoc();
3456 // We can get here for a node like i32 = ISD::SHL i32, i64
3460 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3461 "Unknown shift to lower!");
3463 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3464 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3465 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3468 // If we are in thumb mode, we don't have RRX.
3469 if (ST->isThumb1Only()) return SDValue();
3471 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3472 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3473 DAG.getConstant(0, MVT::i32));
3474 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3475 DAG.getConstant(1, MVT::i32));
3477 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3478 // captures the result into a carry flag.
3479 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3480 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3482 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3483 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3485 // Merge the pieces into a single i64 value.
3486 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3489 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3490 SDValue TmpOp0, TmpOp1;
3491 bool Invert = false;
3495 SDValue Op0 = Op.getOperand(0);
3496 SDValue Op1 = Op.getOperand(1);
3497 SDValue CC = Op.getOperand(2);
3498 EVT VT = Op.getValueType();
3499 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3500 DebugLoc dl = Op.getDebugLoc();
3502 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3503 switch (SetCCOpcode) {
3504 default: llvm_unreachable("Illegal FP comparison");
3506 case ISD::SETNE: Invert = true; // Fallthrough
3508 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3510 case ISD::SETLT: Swap = true; // Fallthrough
3512 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3514 case ISD::SETLE: Swap = true; // Fallthrough
3516 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3517 case ISD::SETUGE: Swap = true; // Fallthrough
3518 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3519 case ISD::SETUGT: Swap = true; // Fallthrough
3520 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3521 case ISD::SETUEQ: Invert = true; // Fallthrough
3523 // Expand this to (OLT | OGT).
3527 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3528 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3530 case ISD::SETUO: Invert = true; // Fallthrough
3532 // Expand this to (OLT | OGE).
3536 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3537 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3541 // Integer comparisons.
3542 switch (SetCCOpcode) {
3543 default: llvm_unreachable("Illegal integer comparison");
3544 case ISD::SETNE: Invert = true;
3545 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3546 case ISD::SETLT: Swap = true;
3547 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3548 case ISD::SETLE: Swap = true;
3549 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3550 case ISD::SETULT: Swap = true;
3551 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3552 case ISD::SETULE: Swap = true;
3553 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3556 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3557 if (Opc == ARMISD::VCEQ) {
3560 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3562 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3565 // Ignore bitconvert.
3566 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3567 AndOp = AndOp.getOperand(0);
3569 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3571 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3572 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3579 std::swap(Op0, Op1);
3581 // If one of the operands is a constant vector zero, attempt to fold the
3582 // comparison to a specialized compare-against-zero form.
3584 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3586 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3587 if (Opc == ARMISD::VCGE)
3588 Opc = ARMISD::VCLEZ;
3589 else if (Opc == ARMISD::VCGT)
3590 Opc = ARMISD::VCLTZ;
3595 if (SingleOp.getNode()) {
3598 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3600 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3602 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3604 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3606 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3608 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3611 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3615 Result = DAG.getNOT(dl, Result, VT);
3620 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3621 /// valid vector constant for a NEON instruction with a "modified immediate"
3622 /// operand (e.g., VMOV). If so, return the encoded value.
3623 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3624 unsigned SplatBitSize, SelectionDAG &DAG,
3625 EVT &VT, bool is128Bits, NEONModImmType type) {
3626 unsigned OpCmode, Imm;
3628 // SplatBitSize is set to the smallest size that splats the vector, so a
3629 // zero vector will always have SplatBitSize == 8. However, NEON modified
3630 // immediate instructions others than VMOV do not support the 8-bit encoding
3631 // of a zero vector, and the default encoding of zero is supposed to be the
3636 switch (SplatBitSize) {
3638 if (type != VMOVModImm)
3640 // Any 1-byte value is OK. Op=0, Cmode=1110.
3641 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3644 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3648 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3649 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3650 if ((SplatBits & ~0xff) == 0) {
3651 // Value = 0x00nn: Op=x, Cmode=100x.
3656 if ((SplatBits & ~0xff00) == 0) {
3657 // Value = 0xnn00: Op=x, Cmode=101x.
3659 Imm = SplatBits >> 8;
3665 // NEON's 32-bit VMOV supports splat values where:
3666 // * only one byte is nonzero, or
3667 // * the least significant byte is 0xff and the second byte is nonzero, or
3668 // * the least significant 2 bytes are 0xff and the third is nonzero.
3669 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3670 if ((SplatBits & ~0xff) == 0) {
3671 // Value = 0x000000nn: Op=x, Cmode=000x.
3676 if ((SplatBits & ~0xff00) == 0) {
3677 // Value = 0x0000nn00: Op=x, Cmode=001x.
3679 Imm = SplatBits >> 8;
3682 if ((SplatBits & ~0xff0000) == 0) {
3683 // Value = 0x00nn0000: Op=x, Cmode=010x.
3685 Imm = SplatBits >> 16;
3688 if ((SplatBits & ~0xff000000) == 0) {
3689 // Value = 0xnn000000: Op=x, Cmode=011x.
3691 Imm = SplatBits >> 24;
3695 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3696 if (type == OtherModImm) return SDValue();
3698 if ((SplatBits & ~0xffff) == 0 &&
3699 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3700 // Value = 0x0000nnff: Op=x, Cmode=1100.
3702 Imm = SplatBits >> 8;
3707 if ((SplatBits & ~0xffffff) == 0 &&
3708 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3709 // Value = 0x00nnffff: Op=x, Cmode=1101.
3711 Imm = SplatBits >> 16;
3712 SplatBits |= 0xffff;
3716 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3717 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3718 // VMOV.I32. A (very) minor optimization would be to replicate the value
3719 // and fall through here to test for a valid 64-bit splat. But, then the
3720 // caller would also need to check and handle the change in size.
3724 if (type != VMOVModImm)
3726 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3727 uint64_t BitMask = 0xff;
3729 unsigned ImmMask = 1;
3731 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3732 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3735 } else if ((SplatBits & BitMask) != 0) {
3741 // Op=1, Cmode=1110.
3744 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3749 llvm_unreachable("unexpected size for isNEONModifiedImm");
3752 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3753 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3756 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3757 bool &ReverseVEXT, unsigned &Imm) {
3758 unsigned NumElts = VT.getVectorNumElements();
3759 ReverseVEXT = false;
3761 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3767 // If this is a VEXT shuffle, the immediate value is the index of the first
3768 // element. The other shuffle indices must be the successive elements after
3770 unsigned ExpectedElt = Imm;
3771 for (unsigned i = 1; i < NumElts; ++i) {
3772 // Increment the expected index. If it wraps around, it may still be
3773 // a VEXT but the source vectors must be swapped.
3775 if (ExpectedElt == NumElts * 2) {
3780 if (M[i] < 0) continue; // ignore UNDEF indices
3781 if (ExpectedElt != static_cast<unsigned>(M[i]))
3785 // Adjust the index value if the source operands will be swapped.
3792 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3793 /// instruction with the specified blocksize. (The order of the elements
3794 /// within each block of the vector is reversed.)
3795 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3796 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3797 "Only possible block sizes for VREV are: 16, 32, 64");
3799 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3803 unsigned NumElts = VT.getVectorNumElements();
3804 unsigned BlockElts = M[0] + 1;
3805 // If the first shuffle index is UNDEF, be optimistic.
3807 BlockElts = BlockSize / EltSz;
3809 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3812 for (unsigned i = 0; i < NumElts; ++i) {
3813 if (M[i] < 0) continue; // ignore UNDEF indices
3814 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3821 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3822 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3823 // range, then 0 is placed into the resulting vector. So pretty much any mask
3824 // of 8 elements can work here.
3825 return VT == MVT::v8i8 && M.size() == 8;
3828 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3829 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3833 unsigned NumElts = VT.getVectorNumElements();
3834 WhichResult = (M[0] == 0 ? 0 : 1);
3835 for (unsigned i = 0; i < NumElts; i += 2) {
3836 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3837 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3843 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3844 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3845 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3846 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3847 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3851 unsigned NumElts = VT.getVectorNumElements();
3852 WhichResult = (M[0] == 0 ? 0 : 1);
3853 for (unsigned i = 0; i < NumElts; i += 2) {
3854 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3855 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3861 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3862 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3866 unsigned NumElts = VT.getVectorNumElements();
3867 WhichResult = (M[0] == 0 ? 0 : 1);
3868 for (unsigned i = 0; i != NumElts; ++i) {
3869 if (M[i] < 0) continue; // ignore UNDEF indices
3870 if ((unsigned) M[i] != 2 * i + WhichResult)
3874 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3875 if (VT.is64BitVector() && EltSz == 32)
3881 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3882 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3883 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3884 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3885 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3889 unsigned Half = VT.getVectorNumElements() / 2;
3890 WhichResult = (M[0] == 0 ? 0 : 1);
3891 for (unsigned j = 0; j != 2; ++j) {
3892 unsigned Idx = WhichResult;
3893 for (unsigned i = 0; i != Half; ++i) {
3894 int MIdx = M[i + j * Half];
3895 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3901 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3902 if (VT.is64BitVector() && EltSz == 32)
3908 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3909 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3913 unsigned NumElts = VT.getVectorNumElements();
3914 WhichResult = (M[0] == 0 ? 0 : 1);
3915 unsigned Idx = WhichResult * NumElts / 2;
3916 for (unsigned i = 0; i != NumElts; i += 2) {
3917 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3918 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3923 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3924 if (VT.is64BitVector() && EltSz == 32)
3930 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3931 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3932 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3933 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3934 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3938 unsigned NumElts = VT.getVectorNumElements();
3939 WhichResult = (M[0] == 0 ? 0 : 1);
3940 unsigned Idx = WhichResult * NumElts / 2;
3941 for (unsigned i = 0; i != NumElts; i += 2) {
3942 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3943 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3948 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3949 if (VT.is64BitVector() && EltSz == 32)
3955 // If N is an integer constant that can be moved into a register in one
3956 // instruction, return an SDValue of such a constant (will become a MOV
3957 // instruction). Otherwise return null.
3958 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3959 const ARMSubtarget *ST, DebugLoc dl) {
3961 if (!isa<ConstantSDNode>(N))
3963 Val = cast<ConstantSDNode>(N)->getZExtValue();
3965 if (ST->isThumb1Only()) {
3966 if (Val <= 255 || ~Val <= 255)
3967 return DAG.getConstant(Val, MVT::i32);
3969 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3970 return DAG.getConstant(Val, MVT::i32);
3975 // If this is a case we can't handle, return null and let the default
3976 // expansion code take care of it.
3977 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3978 const ARMSubtarget *ST) const {
3979 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3980 DebugLoc dl = Op.getDebugLoc();
3981 EVT VT = Op.getValueType();
3983 APInt SplatBits, SplatUndef;
3984 unsigned SplatBitSize;
3986 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3987 if (SplatBitSize <= 64) {
3988 // Check if an immediate VMOV works.
3990 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3991 SplatUndef.getZExtValue(), SplatBitSize,
3992 DAG, VmovVT, VT.is128BitVector(),
3994 if (Val.getNode()) {
3995 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3996 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3999 // Try an immediate VMVN.
4000 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4001 Val = isNEONModifiedImm(NegatedImm,
4002 SplatUndef.getZExtValue(), SplatBitSize,
4003 DAG, VmovVT, VT.is128BitVector(),
4005 if (Val.getNode()) {
4006 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4007 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4010 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4011 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4012 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4014 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4015 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4021 // Scan through the operands to see if only one value is used.
4022 unsigned NumElts = VT.getVectorNumElements();
4023 bool isOnlyLowElement = true;
4024 bool usesOnlyOneValue = true;
4025 bool isConstant = true;
4027 for (unsigned i = 0; i < NumElts; ++i) {
4028 SDValue V = Op.getOperand(i);
4029 if (V.getOpcode() == ISD::UNDEF)
4032 isOnlyLowElement = false;
4033 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4036 if (!Value.getNode())
4038 else if (V != Value)
4039 usesOnlyOneValue = false;
4042 if (!Value.getNode())
4043 return DAG.getUNDEF(VT);
4045 if (isOnlyLowElement)
4046 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4048 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4050 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4051 // i32 and try again.
4052 if (usesOnlyOneValue && EltSize <= 32) {
4054 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4055 if (VT.getVectorElementType().isFloatingPoint()) {
4056 SmallVector<SDValue, 8> Ops;
4057 for (unsigned i = 0; i < NumElts; ++i)
4058 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4060 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4061 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4062 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4064 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4066 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4068 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4071 // If all elements are constants and the case above didn't get hit, fall back
4072 // to the default expansion, which will generate a load from the constant
4077 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4079 SDValue shuffle = ReconstructShuffle(Op, DAG);
4080 if (shuffle != SDValue())
4084 // Vectors with 32- or 64-bit elements can be built by directly assigning
4085 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4086 // will be legalized.
4087 if (EltSize >= 32) {
4088 // Do the expansion with floating-point types, since that is what the VFP
4089 // registers are defined to use, and since i64 is not legal.
4090 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4091 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4092 SmallVector<SDValue, 8> Ops;
4093 for (unsigned i = 0; i < NumElts; ++i)
4094 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4095 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4096 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4102 // Gather data to see if the operation can be modelled as a
4103 // shuffle in combination with VEXTs.
4104 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4105 SelectionDAG &DAG) const {
4106 DebugLoc dl = Op.getDebugLoc();
4107 EVT VT = Op.getValueType();
4108 unsigned NumElts = VT.getVectorNumElements();
4110 SmallVector<SDValue, 2> SourceVecs;
4111 SmallVector<unsigned, 2> MinElts;
4112 SmallVector<unsigned, 2> MaxElts;
4114 for (unsigned i = 0; i < NumElts; ++i) {
4115 SDValue V = Op.getOperand(i);
4116 if (V.getOpcode() == ISD::UNDEF)
4118 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4119 // A shuffle can only come from building a vector from various
4120 // elements of other vectors.
4122 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4123 VT.getVectorElementType()) {
4124 // This code doesn't know how to handle shuffles where the vector
4125 // element types do not match (this happens because type legalization
4126 // promotes the return type of EXTRACT_VECTOR_ELT).
4127 // FIXME: It might be appropriate to extend this code to handle
4128 // mismatched types.
4132 // Record this extraction against the appropriate vector if possible...
4133 SDValue SourceVec = V.getOperand(0);
4134 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4135 bool FoundSource = false;
4136 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4137 if (SourceVecs[j] == SourceVec) {
4138 if (MinElts[j] > EltNo)
4140 if (MaxElts[j] < EltNo)
4147 // Or record a new source if not...
4149 SourceVecs.push_back(SourceVec);
4150 MinElts.push_back(EltNo);
4151 MaxElts.push_back(EltNo);
4155 // Currently only do something sane when at most two source vectors
4157 if (SourceVecs.size() > 2)
4160 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4161 int VEXTOffsets[2] = {0, 0};
4163 // This loop extracts the usage patterns of the source vectors
4164 // and prepares appropriate SDValues for a shuffle if possible.
4165 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4166 if (SourceVecs[i].getValueType() == VT) {
4167 // No VEXT necessary
4168 ShuffleSrcs[i] = SourceVecs[i];
4171 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4172 // It probably isn't worth padding out a smaller vector just to
4173 // break it down again in a shuffle.
4177 // Since only 64-bit and 128-bit vectors are legal on ARM and
4178 // we've eliminated the other cases...
4179 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4180 "unexpected vector sizes in ReconstructShuffle");
4182 if (MaxElts[i] - MinElts[i] >= NumElts) {
4183 // Span too large for a VEXT to cope
4187 if (MinElts[i] >= NumElts) {
4188 // The extraction can just take the second half
4189 VEXTOffsets[i] = NumElts;
4190 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4192 DAG.getIntPtrConstant(NumElts));
4193 } else if (MaxElts[i] < NumElts) {
4194 // The extraction can just take the first half
4196 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4198 DAG.getIntPtrConstant(0));
4200 // An actual VEXT is needed
4201 VEXTOffsets[i] = MinElts[i];
4202 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4204 DAG.getIntPtrConstant(0));
4205 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4207 DAG.getIntPtrConstant(NumElts));
4208 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4209 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4213 SmallVector<int, 8> Mask;
4215 for (unsigned i = 0; i < NumElts; ++i) {
4216 SDValue Entry = Op.getOperand(i);
4217 if (Entry.getOpcode() == ISD::UNDEF) {
4222 SDValue ExtractVec = Entry.getOperand(0);
4223 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4224 .getOperand(1))->getSExtValue();
4225 if (ExtractVec == SourceVecs[0]) {
4226 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4228 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4232 // Final check before we try to produce nonsense...
4233 if (isShuffleMaskLegal(Mask, VT))
4234 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4240 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4241 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4242 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4243 /// are assumed to be legal.
4245 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4247 if (VT.getVectorNumElements() == 4 &&
4248 (VT.is128BitVector() || VT.is64BitVector())) {
4249 unsigned PFIndexes[4];
4250 for (unsigned i = 0; i != 4; ++i) {
4254 PFIndexes[i] = M[i];
4257 // Compute the index in the perfect shuffle table.
4258 unsigned PFTableIndex =
4259 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4260 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4261 unsigned Cost = (PFEntry >> 30);
4268 unsigned Imm, WhichResult;
4270 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4271 return (EltSize >= 32 ||
4272 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4273 isVREVMask(M, VT, 64) ||
4274 isVREVMask(M, VT, 32) ||
4275 isVREVMask(M, VT, 16) ||
4276 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4277 isVTBLMask(M, VT) ||
4278 isVTRNMask(M, VT, WhichResult) ||
4279 isVUZPMask(M, VT, WhichResult) ||
4280 isVZIPMask(M, VT, WhichResult) ||
4281 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4282 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4283 isVZIP_v_undef_Mask(M, VT, WhichResult));
4286 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4287 /// the specified operations to build the shuffle.
4288 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4289 SDValue RHS, SelectionDAG &DAG,
4291 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4292 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4293 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4296 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4305 OP_VUZPL, // VUZP, left result
4306 OP_VUZPR, // VUZP, right result
4307 OP_VZIPL, // VZIP, left result
4308 OP_VZIPR, // VZIP, right result
4309 OP_VTRNL, // VTRN, left result
4310 OP_VTRNR // VTRN, right result
4313 if (OpNum == OP_COPY) {
4314 if (LHSID == (1*9+2)*9+3) return LHS;
4315 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4319 SDValue OpLHS, OpRHS;
4320 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4321 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4322 EVT VT = OpLHS.getValueType();
4325 default: llvm_unreachable("Unknown shuffle opcode!");
4327 // VREV divides the vector in half and swaps within the half.
4328 if (VT.getVectorElementType() == MVT::i32 ||
4329 VT.getVectorElementType() == MVT::f32)
4330 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4331 // vrev <4 x i16> -> VREV32
4332 if (VT.getVectorElementType() == MVT::i16)
4333 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4334 // vrev <4 x i8> -> VREV16
4335 assert(VT.getVectorElementType() == MVT::i8);
4336 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4341 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4342 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4346 return DAG.getNode(ARMISD::VEXT, dl, VT,
4348 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4351 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4352 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4355 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4356 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4359 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4360 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4364 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4365 ArrayRef<int> ShuffleMask,
4366 SelectionDAG &DAG) {
4367 // Check to see if we can use the VTBL instruction.
4368 SDValue V1 = Op.getOperand(0);
4369 SDValue V2 = Op.getOperand(1);
4370 DebugLoc DL = Op.getDebugLoc();
4372 SmallVector<SDValue, 8> VTBLMask;
4373 for (ArrayRef<int>::iterator
4374 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4375 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4377 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4378 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4379 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4382 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4383 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4387 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4388 SDValue V1 = Op.getOperand(0);
4389 SDValue V2 = Op.getOperand(1);
4390 DebugLoc dl = Op.getDebugLoc();
4391 EVT VT = Op.getValueType();
4392 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4394 // Convert shuffles that are directly supported on NEON to target-specific
4395 // DAG nodes, instead of keeping them as shuffles and matching them again
4396 // during code selection. This is more efficient and avoids the possibility
4397 // of inconsistencies between legalization and selection.
4398 // FIXME: floating-point vectors should be canonicalized to integer vectors
4399 // of the same time so that they get CSEd properly.
4400 ArrayRef<int> ShuffleMask = SVN->getMask();
4402 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4403 if (EltSize <= 32) {
4404 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4405 int Lane = SVN->getSplatIndex();
4406 // If this is undef splat, generate it via "just" vdup, if possible.
4407 if (Lane == -1) Lane = 0;
4409 // Test if V1 is a SCALAR_TO_VECTOR.
4410 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4411 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4413 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4414 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4416 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4417 !isa<ConstantSDNode>(V1.getOperand(0))) {
4418 bool IsScalarToVector = true;
4419 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4420 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4421 IsScalarToVector = false;
4424 if (IsScalarToVector)
4425 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4427 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4428 DAG.getConstant(Lane, MVT::i32));
4433 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4436 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4437 DAG.getConstant(Imm, MVT::i32));
4440 if (isVREVMask(ShuffleMask, VT, 64))
4441 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4442 if (isVREVMask(ShuffleMask, VT, 32))
4443 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4444 if (isVREVMask(ShuffleMask, VT, 16))
4445 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4447 // Check for Neon shuffles that modify both input vectors in place.
4448 // If both results are used, i.e., if there are two shuffles with the same
4449 // source operands and with masks corresponding to both results of one of
4450 // these operations, DAG memoization will ensure that a single node is
4451 // used for both shuffles.
4452 unsigned WhichResult;
4453 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4454 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4455 V1, V2).getValue(WhichResult);
4456 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4457 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4458 V1, V2).getValue(WhichResult);
4459 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4460 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4461 V1, V2).getValue(WhichResult);
4463 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4464 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4465 V1, V1).getValue(WhichResult);
4466 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4467 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4468 V1, V1).getValue(WhichResult);
4469 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4470 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4471 V1, V1).getValue(WhichResult);
4474 // If the shuffle is not directly supported and it has 4 elements, use
4475 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4476 unsigned NumElts = VT.getVectorNumElements();
4478 unsigned PFIndexes[4];
4479 for (unsigned i = 0; i != 4; ++i) {
4480 if (ShuffleMask[i] < 0)
4483 PFIndexes[i] = ShuffleMask[i];
4486 // Compute the index in the perfect shuffle table.
4487 unsigned PFTableIndex =
4488 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4489 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4490 unsigned Cost = (PFEntry >> 30);
4493 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4496 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4497 if (EltSize >= 32) {
4498 // Do the expansion with floating-point types, since that is what the VFP
4499 // registers are defined to use, and since i64 is not legal.
4500 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4501 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4502 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4503 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4504 SmallVector<SDValue, 8> Ops;
4505 for (unsigned i = 0; i < NumElts; ++i) {
4506 if (ShuffleMask[i] < 0)
4507 Ops.push_back(DAG.getUNDEF(EltVT));
4509 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4510 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4511 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4514 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4515 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4518 if (VT == MVT::v8i8) {
4519 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4520 if (NewOp.getNode())
4527 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4528 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4529 SDValue Lane = Op.getOperand(2);
4530 if (!isa<ConstantSDNode>(Lane))
4536 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4537 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4538 SDValue Lane = Op.getOperand(1);
4539 if (!isa<ConstantSDNode>(Lane))
4542 SDValue Vec = Op.getOperand(0);
4543 if (Op.getValueType() == MVT::i32 &&
4544 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4545 DebugLoc dl = Op.getDebugLoc();
4546 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4552 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4553 // The only time a CONCAT_VECTORS operation can have legal types is when
4554 // two 64-bit vectors are concatenated to a 128-bit vector.
4555 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4556 "unexpected CONCAT_VECTORS");
4557 DebugLoc dl = Op.getDebugLoc();
4558 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4559 SDValue Op0 = Op.getOperand(0);
4560 SDValue Op1 = Op.getOperand(1);
4561 if (Op0.getOpcode() != ISD::UNDEF)
4562 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4563 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4564 DAG.getIntPtrConstant(0));
4565 if (Op1.getOpcode() != ISD::UNDEF)
4566 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4567 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4568 DAG.getIntPtrConstant(1));
4569 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4572 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4573 /// element has been zero/sign-extended, depending on the isSigned parameter,
4574 /// from an integer type half its size.
4575 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4577 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4578 EVT VT = N->getValueType(0);
4579 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4580 SDNode *BVN = N->getOperand(0).getNode();
4581 if (BVN->getValueType(0) != MVT::v4i32 ||
4582 BVN->getOpcode() != ISD::BUILD_VECTOR)
4584 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4585 unsigned HiElt = 1 - LoElt;
4586 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4587 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4588 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4589 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4590 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4593 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4594 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4597 if (Hi0->isNullValue() && Hi1->isNullValue())
4603 if (N->getOpcode() != ISD::BUILD_VECTOR)
4606 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4607 SDNode *Elt = N->getOperand(i).getNode();
4608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4609 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4610 unsigned HalfSize = EltSize / 2;
4612 if (!isIntN(HalfSize, C->getSExtValue()))
4615 if (!isUIntN(HalfSize, C->getZExtValue()))
4626 /// isSignExtended - Check if a node is a vector value that is sign-extended
4627 /// or a constant BUILD_VECTOR with sign-extended elements.
4628 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4629 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4631 if (isExtendedBUILD_VECTOR(N, DAG, true))
4636 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4637 /// or a constant BUILD_VECTOR with zero-extended elements.
4638 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4639 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4641 if (isExtendedBUILD_VECTOR(N, DAG, false))
4646 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4647 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4648 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4649 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4650 return N->getOperand(0);
4651 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4652 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4653 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4654 LD->isNonTemporal(), LD->isInvariant(),
4655 LD->getAlignment());
4656 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4657 // have been legalized as a BITCAST from v4i32.
4658 if (N->getOpcode() == ISD::BITCAST) {
4659 SDNode *BVN = N->getOperand(0).getNode();
4660 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4661 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4662 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4663 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4664 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4666 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4667 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4668 EVT VT = N->getValueType(0);
4669 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4670 unsigned NumElts = VT.getVectorNumElements();
4671 MVT TruncVT = MVT::getIntegerVT(EltSize);
4672 SmallVector<SDValue, 8> Ops;
4673 for (unsigned i = 0; i != NumElts; ++i) {
4674 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4675 const APInt &CInt = C->getAPIntValue();
4676 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4678 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4679 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4682 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4683 unsigned Opcode = N->getOpcode();
4684 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4685 SDNode *N0 = N->getOperand(0).getNode();
4686 SDNode *N1 = N->getOperand(1).getNode();
4687 return N0->hasOneUse() && N1->hasOneUse() &&
4688 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4693 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4694 unsigned Opcode = N->getOpcode();
4695 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4696 SDNode *N0 = N->getOperand(0).getNode();
4697 SDNode *N1 = N->getOperand(1).getNode();
4698 return N0->hasOneUse() && N1->hasOneUse() &&
4699 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4704 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4705 // Multiplications are only custom-lowered for 128-bit vectors so that
4706 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4707 EVT VT = Op.getValueType();
4708 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4709 SDNode *N0 = Op.getOperand(0).getNode();
4710 SDNode *N1 = Op.getOperand(1).getNode();
4711 unsigned NewOpc = 0;
4713 bool isN0SExt = isSignExtended(N0, DAG);
4714 bool isN1SExt = isSignExtended(N1, DAG);
4715 if (isN0SExt && isN1SExt)
4716 NewOpc = ARMISD::VMULLs;
4718 bool isN0ZExt = isZeroExtended(N0, DAG);
4719 bool isN1ZExt = isZeroExtended(N1, DAG);
4720 if (isN0ZExt && isN1ZExt)
4721 NewOpc = ARMISD::VMULLu;
4722 else if (isN1SExt || isN1ZExt) {
4723 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4724 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4725 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4726 NewOpc = ARMISD::VMULLs;
4728 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4729 NewOpc = ARMISD::VMULLu;
4731 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4733 NewOpc = ARMISD::VMULLu;
4739 if (VT == MVT::v2i64)
4740 // Fall through to expand this. It is not legal.
4743 // Other vector multiplications are legal.
4748 // Legalize to a VMULL instruction.
4749 DebugLoc DL = Op.getDebugLoc();
4751 SDValue Op1 = SkipExtension(N1, DAG);
4753 Op0 = SkipExtension(N0, DAG);
4754 assert(Op0.getValueType().is64BitVector() &&
4755 Op1.getValueType().is64BitVector() &&
4756 "unexpected types for extended operands to VMULL");
4757 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4760 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4761 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4768 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4769 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4770 EVT Op1VT = Op1.getValueType();
4771 return DAG.getNode(N0->getOpcode(), DL, VT,
4772 DAG.getNode(NewOpc, DL, VT,
4773 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4774 DAG.getNode(NewOpc, DL, VT,
4775 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4779 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4781 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4782 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4783 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4784 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4785 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4786 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4787 // Get reciprocal estimate.
4788 // float4 recip = vrecpeq_f32(yf);
4789 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4790 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4791 // Because char has a smaller range than uchar, we can actually get away
4792 // without any newton steps. This requires that we use a weird bias
4793 // of 0xb000, however (again, this has been exhaustively tested).
4794 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4795 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4796 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4797 Y = DAG.getConstant(0xb000, MVT::i32);
4798 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4799 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4800 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4801 // Convert back to short.
4802 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4803 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4808 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4810 // Convert to float.
4811 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4812 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4813 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4814 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4815 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4816 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4818 // Use reciprocal estimate and one refinement step.
4819 // float4 recip = vrecpeq_f32(yf);
4820 // recip *= vrecpsq_f32(yf, recip);
4821 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4822 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4823 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4824 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4826 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4827 // Because short has a smaller range than ushort, we can actually get away
4828 // with only a single newton step. This requires that we use a weird bias
4829 // of 89, however (again, this has been exhaustively tested).
4830 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4831 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4832 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4833 N1 = DAG.getConstant(0x89, MVT::i32);
4834 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4835 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4836 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4837 // Convert back to integer and return.
4838 // return vmovn_s32(vcvt_s32_f32(result));
4839 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4840 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4844 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4845 EVT VT = Op.getValueType();
4846 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4847 "unexpected type for custom-lowering ISD::SDIV");
4849 DebugLoc dl = Op.getDebugLoc();
4850 SDValue N0 = Op.getOperand(0);
4851 SDValue N1 = Op.getOperand(1);
4854 if (VT == MVT::v8i8) {
4855 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4856 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4858 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4859 DAG.getIntPtrConstant(4));
4860 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4861 DAG.getIntPtrConstant(4));
4862 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4863 DAG.getIntPtrConstant(0));
4864 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4865 DAG.getIntPtrConstant(0));
4867 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4868 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4870 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4871 N0 = LowerCONCAT_VECTORS(N0, DAG);
4873 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4876 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4879 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4880 EVT VT = Op.getValueType();
4881 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4882 "unexpected type for custom-lowering ISD::UDIV");
4884 DebugLoc dl = Op.getDebugLoc();
4885 SDValue N0 = Op.getOperand(0);
4886 SDValue N1 = Op.getOperand(1);
4889 if (VT == MVT::v8i8) {
4890 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4891 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4893 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4894 DAG.getIntPtrConstant(4));
4895 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4896 DAG.getIntPtrConstant(4));
4897 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4898 DAG.getIntPtrConstant(0));
4899 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4900 DAG.getIntPtrConstant(0));
4902 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4903 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4905 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4906 N0 = LowerCONCAT_VECTORS(N0, DAG);
4908 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4909 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4914 // v4i16 sdiv ... Convert to float.
4915 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4916 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4917 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4918 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4919 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4920 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4922 // Use reciprocal estimate and two refinement steps.
4923 // float4 recip = vrecpeq_f32(yf);
4924 // recip *= vrecpsq_f32(yf, recip);
4925 // recip *= vrecpsq_f32(yf, recip);
4926 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4927 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4928 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4929 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4931 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4932 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4933 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4935 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4936 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4937 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4938 // and that it will never cause us to return an answer too large).
4939 // float4 result = as_float4(as_int4(xf*recip) + 2);
4940 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4941 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4942 N1 = DAG.getConstant(2, MVT::i32);
4943 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4944 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4945 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4946 // Convert back to integer and return.
4947 // return vmovn_u32(vcvt_s32_f32(result));
4948 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4949 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4953 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4954 EVT VT = Op.getNode()->getValueType(0);
4955 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4958 bool ExtraOp = false;
4959 switch (Op.getOpcode()) {
4960 default: assert(0 && "Invalid code");
4961 case ISD::ADDC: Opc = ARMISD::ADDC; break;
4962 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4963 case ISD::SUBC: Opc = ARMISD::SUBC; break;
4964 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4968 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4970 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4971 Op.getOperand(1), Op.getOperand(2));
4974 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
4975 // Monotonic load/store is legal for all targets
4976 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
4979 // Aquire/Release load/store is not legal for targets without a
4980 // dmb or equivalent available.
4986 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
4987 SelectionDAG &DAG, unsigned NewOp) {
4988 DebugLoc dl = Node->getDebugLoc();
4989 assert (Node->getValueType(0) == MVT::i64 &&
4990 "Only know how to expand i64 atomics");
4992 SmallVector<SDValue, 6> Ops;
4993 Ops.push_back(Node->getOperand(0)); // Chain
4994 Ops.push_back(Node->getOperand(1)); // Ptr
4996 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4997 Node->getOperand(2), DAG.getIntPtrConstant(0)));
4998 // High part of Val1
4999 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5000 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5001 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5002 // High part of Val1
5003 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5004 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5005 // High part of Val2
5006 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5007 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5009 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5011 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5012 cast<MemSDNode>(Node)->getMemOperand());
5013 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5014 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5015 Results.push_back(Result.getValue(2));
5018 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5019 switch (Op.getOpcode()) {
5020 default: llvm_unreachable("Don't know how to custom lower this!");
5021 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5022 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5023 case ISD::GlobalAddress:
5024 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5025 LowerGlobalAddressELF(Op, DAG);
5026 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5027 case ISD::SELECT: return LowerSELECT(Op, DAG);
5028 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5029 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5030 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5031 case ISD::VASTART: return LowerVASTART(Op, DAG);
5032 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5033 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5034 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5035 case ISD::SINT_TO_FP:
5036 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5037 case ISD::FP_TO_SINT:
5038 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5039 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5040 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5041 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5042 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5043 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5044 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5045 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5047 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5050 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5051 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5052 case ISD::SRL_PARTS:
5053 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5054 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5055 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5056 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5057 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5058 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5059 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5060 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5061 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5062 case ISD::MUL: return LowerMUL(Op, DAG);
5063 case ISD::SDIV: return LowerSDIV(Op, DAG);
5064 case ISD::UDIV: return LowerUDIV(Op, DAG);
5068 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5069 case ISD::ATOMIC_LOAD:
5070 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5074 /// ReplaceNodeResults - Replace the results of node with an illegal result
5075 /// type with new values built out of custom code.
5076 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5077 SmallVectorImpl<SDValue>&Results,
5078 SelectionDAG &DAG) const {
5080 switch (N->getOpcode()) {
5082 llvm_unreachable("Don't know how to custom expand this!");
5084 Res = ExpandBITCAST(N, DAG);
5088 Res = Expand64BitShift(N, DAG, Subtarget);
5090 case ISD::ATOMIC_LOAD_ADD:
5091 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5093 case ISD::ATOMIC_LOAD_AND:
5094 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5096 case ISD::ATOMIC_LOAD_NAND:
5097 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5099 case ISD::ATOMIC_LOAD_OR:
5100 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5102 case ISD::ATOMIC_LOAD_SUB:
5103 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5105 case ISD::ATOMIC_LOAD_XOR:
5106 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5108 case ISD::ATOMIC_SWAP:
5109 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5111 case ISD::ATOMIC_CMP_SWAP:
5112 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5116 Results.push_back(Res);
5119 //===----------------------------------------------------------------------===//
5120 // ARM Scheduler Hooks
5121 //===----------------------------------------------------------------------===//
5124 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5125 MachineBasicBlock *BB,
5126 unsigned Size) const {
5127 unsigned dest = MI->getOperand(0).getReg();
5128 unsigned ptr = MI->getOperand(1).getReg();
5129 unsigned oldval = MI->getOperand(2).getReg();
5130 unsigned newval = MI->getOperand(3).getReg();
5131 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5132 DebugLoc dl = MI->getDebugLoc();
5133 bool isThumb2 = Subtarget->isThumb2();
5135 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5137 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5138 : ARM::GPRRegisterClass);
5141 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5142 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5143 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5146 unsigned ldrOpc, strOpc;
5148 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5150 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5151 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5154 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5155 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5158 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5159 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5163 MachineFunction *MF = BB->getParent();
5164 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5165 MachineFunction::iterator It = BB;
5166 ++It; // insert the new blocks after the current block
5168 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5169 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5170 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5171 MF->insert(It, loop1MBB);
5172 MF->insert(It, loop2MBB);
5173 MF->insert(It, exitMBB);
5175 // Transfer the remainder of BB and its successor edges to exitMBB.
5176 exitMBB->splice(exitMBB->begin(), BB,
5177 llvm::next(MachineBasicBlock::iterator(MI)),
5179 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5183 // fallthrough --> loop1MBB
5184 BB->addSuccessor(loop1MBB);
5187 // ldrex dest, [ptr]
5191 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5192 if (ldrOpc == ARM::t2LDREX)
5194 AddDefaultPred(MIB);
5195 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5196 .addReg(dest).addReg(oldval));
5197 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5198 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5199 BB->addSuccessor(loop2MBB);
5200 BB->addSuccessor(exitMBB);
5203 // strex scratch, newval, [ptr]
5207 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5208 if (strOpc == ARM::t2STREX)
5210 AddDefaultPred(MIB);
5211 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5212 .addReg(scratch).addImm(0));
5213 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5214 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5215 BB->addSuccessor(loop1MBB);
5216 BB->addSuccessor(exitMBB);
5222 MI->eraseFromParent(); // The instruction is gone now.
5228 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5229 unsigned Size, unsigned BinOpcode) const {
5230 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5231 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5233 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5234 MachineFunction *MF = BB->getParent();
5235 MachineFunction::iterator It = BB;
5238 unsigned dest = MI->getOperand(0).getReg();
5239 unsigned ptr = MI->getOperand(1).getReg();
5240 unsigned incr = MI->getOperand(2).getReg();
5241 DebugLoc dl = MI->getDebugLoc();
5242 bool isThumb2 = Subtarget->isThumb2();
5244 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5246 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5247 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5250 unsigned ldrOpc, strOpc;
5252 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5254 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5255 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5258 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5259 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5262 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5263 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5267 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5268 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5269 MF->insert(It, loopMBB);
5270 MF->insert(It, exitMBB);
5272 // Transfer the remainder of BB and its successor edges to exitMBB.
5273 exitMBB->splice(exitMBB->begin(), BB,
5274 llvm::next(MachineBasicBlock::iterator(MI)),
5276 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5278 TargetRegisterClass *TRC =
5279 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5280 unsigned scratch = MRI.createVirtualRegister(TRC);
5281 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5285 // fallthrough --> loopMBB
5286 BB->addSuccessor(loopMBB);
5290 // <binop> scratch2, dest, incr
5291 // strex scratch, scratch2, ptr
5294 // fallthrough --> exitMBB
5296 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5297 if (ldrOpc == ARM::t2LDREX)
5299 AddDefaultPred(MIB);
5301 // operand order needs to go the other way for NAND
5302 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5303 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5304 addReg(incr).addReg(dest)).addReg(0);
5306 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5307 addReg(dest).addReg(incr)).addReg(0);
5310 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5311 if (strOpc == ARM::t2STREX)
5313 AddDefaultPred(MIB);
5314 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5315 .addReg(scratch).addImm(0));
5316 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5317 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5319 BB->addSuccessor(loopMBB);
5320 BB->addSuccessor(exitMBB);
5326 MI->eraseFromParent(); // The instruction is gone now.
5332 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5333 MachineBasicBlock *BB,
5336 ARMCC::CondCodes Cond) const {
5337 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5339 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5340 MachineFunction *MF = BB->getParent();
5341 MachineFunction::iterator It = BB;
5344 unsigned dest = MI->getOperand(0).getReg();
5345 unsigned ptr = MI->getOperand(1).getReg();
5346 unsigned incr = MI->getOperand(2).getReg();
5347 unsigned oldval = dest;
5348 DebugLoc dl = MI->getDebugLoc();
5349 bool isThumb2 = Subtarget->isThumb2();
5351 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5353 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5354 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5357 unsigned ldrOpc, strOpc, extendOpc;
5359 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5361 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5362 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5363 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5366 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5367 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5368 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5371 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5372 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5377 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5378 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5379 MF->insert(It, loopMBB);
5380 MF->insert(It, exitMBB);
5382 // Transfer the remainder of BB and its successor edges to exitMBB.
5383 exitMBB->splice(exitMBB->begin(), BB,
5384 llvm::next(MachineBasicBlock::iterator(MI)),
5386 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5388 TargetRegisterClass *TRC =
5389 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5390 unsigned scratch = MRI.createVirtualRegister(TRC);
5391 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5395 // fallthrough --> loopMBB
5396 BB->addSuccessor(loopMBB);
5400 // (sign extend dest, if required)
5402 // cmov.cond scratch2, dest, incr
5403 // strex scratch, scratch2, ptr
5406 // fallthrough --> exitMBB
5408 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5409 if (ldrOpc == ARM::t2LDREX)
5411 AddDefaultPred(MIB);
5413 // Sign extend the value, if necessary.
5414 if (signExtend && extendOpc) {
5415 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5416 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5421 // Build compare and cmov instructions.
5422 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5423 .addReg(oldval).addReg(incr));
5424 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5425 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5427 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5428 if (strOpc == ARM::t2STREX)
5430 AddDefaultPred(MIB);
5431 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5432 .addReg(scratch).addImm(0));
5433 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5434 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5436 BB->addSuccessor(loopMBB);
5437 BB->addSuccessor(exitMBB);
5443 MI->eraseFromParent(); // The instruction is gone now.
5449 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5450 unsigned Op1, unsigned Op2,
5451 bool NeedsCarry, bool IsCmpxchg) const {
5452 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5453 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5455 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5456 MachineFunction *MF = BB->getParent();
5457 MachineFunction::iterator It = BB;
5460 unsigned destlo = MI->getOperand(0).getReg();
5461 unsigned desthi = MI->getOperand(1).getReg();
5462 unsigned ptr = MI->getOperand(2).getReg();
5463 unsigned vallo = MI->getOperand(3).getReg();
5464 unsigned valhi = MI->getOperand(4).getReg();
5465 DebugLoc dl = MI->getDebugLoc();
5466 bool isThumb2 = Subtarget->isThumb2();
5468 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5470 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5471 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5472 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5475 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5476 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5478 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5479 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5481 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5482 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5484 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5485 MF->insert(It, loopMBB);
5487 MF->insert(It, contBB);
5488 MF->insert(It, cont2BB);
5490 MF->insert(It, exitMBB);
5492 // Transfer the remainder of BB and its successor edges to exitMBB.
5493 exitMBB->splice(exitMBB->begin(), BB,
5494 llvm::next(MachineBasicBlock::iterator(MI)),
5496 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5498 TargetRegisterClass *TRC =
5499 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5500 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5504 // fallthrough --> loopMBB
5505 BB->addSuccessor(loopMBB);
5508 // ldrexd r2, r3, ptr
5509 // <binopa> r0, r2, incr
5510 // <binopb> r1, r3, incr
5511 // strexd storesuccess, r0, r1, ptr
5512 // cmp storesuccess, #0
5514 // fallthrough --> exitMBB
5516 // Note that the registers are explicitly specified because there is not any
5517 // way to force the register allocator to allocate a register pair.
5519 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5520 // need to properly enforce the restriction that the two output registers
5521 // for ldrexd must be different.
5524 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5525 .addReg(ARM::R2, RegState::Define)
5526 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5527 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5528 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5529 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5533 for (unsigned i = 0; i < 2; i++) {
5534 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5536 .addReg(i == 0 ? destlo : desthi)
5537 .addReg(i == 0 ? vallo : valhi));
5538 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5539 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5540 BB->addSuccessor(exitMBB);
5541 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5542 BB = (i == 0 ? contBB : cont2BB);
5545 // Copy to physregs for strexd
5546 unsigned setlo = MI->getOperand(5).getReg();
5547 unsigned sethi = MI->getOperand(6).getReg();
5548 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5549 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5551 // Perform binary operation
5552 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5553 .addReg(destlo).addReg(vallo))
5554 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5555 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5556 .addReg(desthi).addReg(valhi)).addReg(0);
5558 // Copy to physregs for strexd
5559 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5560 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5564 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5565 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5567 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5568 .addReg(storesuccess).addImm(0));
5569 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5570 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5572 BB->addSuccessor(loopMBB);
5573 BB->addSuccessor(exitMBB);
5579 MI->eraseFromParent(); // The instruction is gone now.
5584 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5585 /// registers the function context.
5586 void ARMTargetLowering::
5587 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5588 MachineBasicBlock *DispatchBB, int FI) const {
5589 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5590 DebugLoc dl = MI->getDebugLoc();
5591 MachineFunction *MF = MBB->getParent();
5592 MachineRegisterInfo *MRI = &MF->getRegInfo();
5593 MachineConstantPool *MCP = MF->getConstantPool();
5594 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5595 const Function *F = MF->getFunction();
5597 bool isThumb = Subtarget->isThumb();
5598 bool isThumb2 = Subtarget->isThumb2();
5600 unsigned PCLabelId = AFI->createPICLabelUId();
5601 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5602 ARMConstantPoolValue *CPV =
5603 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5604 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5606 const TargetRegisterClass *TRC =
5607 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5609 // Grab constant pool and fixed stack memory operands.
5610 MachineMemOperand *CPMMO =
5611 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5612 MachineMemOperand::MOLoad, 4, 4);
5614 MachineMemOperand *FIMMOSt =
5615 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5616 MachineMemOperand::MOStore, 4, 4);
5618 // Load the address of the dispatch MBB into the jump buffer.
5620 // Incoming value: jbuf
5621 // ldr.n r5, LCPI1_1
5624 // str r5, [$jbuf, #+4] ; &jbuf[1]
5625 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5626 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5627 .addConstantPoolIndex(CPI)
5628 .addMemOperand(CPMMO));
5629 // Set the low bit because of thumb mode.
5630 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5632 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5633 .addReg(NewVReg1, RegState::Kill)
5635 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5636 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5637 .addReg(NewVReg2, RegState::Kill)
5639 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5640 .addReg(NewVReg3, RegState::Kill)
5642 .addImm(36) // &jbuf[1] :: pc
5643 .addMemOperand(FIMMOSt));
5644 } else if (isThumb) {
5645 // Incoming value: jbuf
5646 // ldr.n r1, LCPI1_4
5650 // add r2, $jbuf, #+4 ; &jbuf[1]
5652 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5653 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5654 .addConstantPoolIndex(CPI)
5655 .addMemOperand(CPMMO));
5656 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5657 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5658 .addReg(NewVReg1, RegState::Kill)
5660 // Set the low bit because of thumb mode.
5661 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5662 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5663 .addReg(ARM::CPSR, RegState::Define)
5665 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5666 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5667 .addReg(ARM::CPSR, RegState::Define)
5668 .addReg(NewVReg2, RegState::Kill)
5669 .addReg(NewVReg3, RegState::Kill));
5670 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5671 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5673 .addImm(36)); // &jbuf[1] :: pc
5674 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5675 .addReg(NewVReg4, RegState::Kill)
5676 .addReg(NewVReg5, RegState::Kill)
5678 .addMemOperand(FIMMOSt));
5680 // Incoming value: jbuf
5683 // str r1, [$jbuf, #+4] ; &jbuf[1]
5684 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5685 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5686 .addConstantPoolIndex(CPI)
5688 .addMemOperand(CPMMO));
5689 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5690 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5691 .addReg(NewVReg1, RegState::Kill)
5692 .addImm(PCLabelId));
5693 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5694 .addReg(NewVReg2, RegState::Kill)
5696 .addImm(36) // &jbuf[1] :: pc
5697 .addMemOperand(FIMMOSt));
5701 MachineBasicBlock *ARMTargetLowering::
5702 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5703 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5704 DebugLoc dl = MI->getDebugLoc();
5705 MachineFunction *MF = MBB->getParent();
5706 MachineRegisterInfo *MRI = &MF->getRegInfo();
5707 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5708 MachineFrameInfo *MFI = MF->getFrameInfo();
5709 int FI = MFI->getFunctionContextIndex();
5711 const TargetRegisterClass *TRC =
5712 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5714 // Get a mapping of the call site numbers to all of the landing pads they're
5716 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5717 unsigned MaxCSNum = 0;
5718 MachineModuleInfo &MMI = MF->getMMI();
5719 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5720 if (!BB->isLandingPad()) continue;
5722 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5724 for (MachineBasicBlock::iterator
5725 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5726 if (!II->isEHLabel()) continue;
5728 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5729 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5731 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5732 for (SmallVectorImpl<unsigned>::iterator
5733 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5734 CSI != CSE; ++CSI) {
5735 CallSiteNumToLPad[*CSI].push_back(BB);
5736 MaxCSNum = std::max(MaxCSNum, *CSI);
5742 // Get an ordered list of the machine basic blocks for the jump table.
5743 std::vector<MachineBasicBlock*> LPadList;
5744 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5745 LPadList.reserve(CallSiteNumToLPad.size());
5746 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5747 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5748 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5749 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5750 LPadList.push_back(*II);
5751 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5755 assert(!LPadList.empty() &&
5756 "No landing pad destinations for the dispatch jump table!");
5758 // Create the jump table and associated information.
5759 MachineJumpTableInfo *JTI =
5760 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5761 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5762 unsigned UId = AFI->createJumpTableUId();
5764 // Create the MBBs for the dispatch code.
5766 // Shove the dispatch's address into the return slot in the function context.
5767 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5768 DispatchBB->setIsLandingPad();
5770 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5771 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5772 DispatchBB->addSuccessor(TrapBB);
5774 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5775 DispatchBB->addSuccessor(DispContBB);
5778 MF->insert(MF->end(), DispatchBB);
5779 MF->insert(MF->end(), DispContBB);
5780 MF->insert(MF->end(), TrapBB);
5782 // Insert code into the entry block that creates and registers the function
5784 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5786 MachineMemOperand *FIMMOLd =
5787 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5788 MachineMemOperand::MOLoad |
5789 MachineMemOperand::MOVolatile, 4, 4);
5791 if (AFI->isThumb1OnlyFunction())
5792 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5793 else if (!Subtarget->hasVFP2())
5794 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5796 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5798 unsigned NumLPads = LPadList.size();
5799 if (Subtarget->isThumb2()) {
5800 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5801 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5804 .addMemOperand(FIMMOLd));
5806 if (NumLPads < 256) {
5807 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5809 .addImm(LPadList.size()));
5811 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5812 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5813 .addImm(NumLPads & 0xFFFF));
5815 unsigned VReg2 = VReg1;
5816 if ((NumLPads & 0xFFFF0000) != 0) {
5817 VReg2 = MRI->createVirtualRegister(TRC);
5818 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5820 .addImm(NumLPads >> 16));
5823 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5828 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5833 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5834 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5835 .addJumpTableIndex(MJTI)
5838 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5841 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5842 .addReg(NewVReg3, RegState::Kill)
5844 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5846 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5847 .addReg(NewVReg4, RegState::Kill)
5849 .addJumpTableIndex(MJTI)
5851 } else if (Subtarget->isThumb()) {
5852 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5853 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5856 .addMemOperand(FIMMOLd));
5858 if (NumLPads < 256) {
5859 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5863 MachineConstantPool *ConstantPool = MF->getConstantPool();
5864 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5865 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5867 // MachineConstantPool wants an explicit alignment.
5868 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5870 Align = getTargetData()->getTypeAllocSize(C->getType());
5871 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5873 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5874 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5875 .addReg(VReg1, RegState::Define)
5876 .addConstantPoolIndex(Idx));
5877 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5882 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5887 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5888 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5889 .addReg(ARM::CPSR, RegState::Define)
5893 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5894 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5895 .addJumpTableIndex(MJTI)
5898 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5899 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5900 .addReg(ARM::CPSR, RegState::Define)
5901 .addReg(NewVReg2, RegState::Kill)
5904 MachineMemOperand *JTMMOLd =
5905 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5906 MachineMemOperand::MOLoad, 4, 4);
5908 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5909 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5910 .addReg(NewVReg4, RegState::Kill)
5912 .addMemOperand(JTMMOLd));
5914 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5915 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5916 .addReg(ARM::CPSR, RegState::Define)
5917 .addReg(NewVReg5, RegState::Kill)
5920 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5921 .addReg(NewVReg6, RegState::Kill)
5922 .addJumpTableIndex(MJTI)
5925 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5926 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5929 .addMemOperand(FIMMOLd));
5931 if (NumLPads < 256) {
5932 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5935 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5936 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5937 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5938 .addImm(NumLPads & 0xFFFF));
5940 unsigned VReg2 = VReg1;
5941 if ((NumLPads & 0xFFFF0000) != 0) {
5942 VReg2 = MRI->createVirtualRegister(TRC);
5943 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5945 .addImm(NumLPads >> 16));
5948 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5952 MachineConstantPool *ConstantPool = MF->getConstantPool();
5953 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5954 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5956 // MachineConstantPool wants an explicit alignment.
5957 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5959 Align = getTargetData()->getTypeAllocSize(C->getType());
5960 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5962 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5963 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
5964 .addReg(VReg1, RegState::Define)
5965 .addConstantPoolIndex(Idx)
5967 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5969 .addReg(VReg1, RegState::Kill));
5972 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
5977 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5979 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
5981 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5982 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5983 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
5984 .addJumpTableIndex(MJTI)
5987 MachineMemOperand *JTMMOLd =
5988 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5989 MachineMemOperand::MOLoad, 4, 4);
5990 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5992 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
5993 .addReg(NewVReg3, RegState::Kill)
5996 .addMemOperand(JTMMOLd));
5998 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
5999 .addReg(NewVReg5, RegState::Kill)
6001 .addJumpTableIndex(MJTI)
6005 // Add the jump table entries as successors to the MBB.
6006 MachineBasicBlock *PrevMBB = 0;
6007 for (std::vector<MachineBasicBlock*>::iterator
6008 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6009 MachineBasicBlock *CurMBB = *I;
6010 if (PrevMBB != CurMBB)
6011 DispContBB->addSuccessor(CurMBB);
6015 // N.B. the order the invoke BBs are processed in doesn't matter here.
6016 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6017 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6018 const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
6019 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6020 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6021 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6022 MachineBasicBlock *BB = *I;
6024 // Remove the landing pad successor from the invoke block and replace it
6025 // with the new dispatch block.
6026 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6028 while (!Successors.empty()) {
6029 MachineBasicBlock *SMBB = Successors.pop_back_val();
6030 if (SMBB->isLandingPad()) {
6031 BB->removeSuccessor(SMBB);
6032 MBBLPads.push_back(SMBB);
6036 BB->addSuccessor(DispatchBB);
6038 // Find the invoke call and mark all of the callee-saved registers as
6039 // 'implicit defined' so that they're spilled. This prevents code from
6040 // moving instructions to before the EH block, where they will never be
6042 for (MachineBasicBlock::reverse_iterator
6043 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6044 if (!II->isCall()) continue;
6046 DenseMap<unsigned, bool> DefRegs;
6047 for (MachineInstr::mop_iterator
6048 OI = II->operands_begin(), OE = II->operands_end();
6050 if (!OI->isReg()) continue;
6051 DefRegs[OI->getReg()] = true;
6054 MachineInstrBuilder MIB(&*II);
6056 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6057 unsigned Reg = SavedRegs[i];
6058 if (Subtarget->isThumb2() &&
6059 !ARM::tGPRRegisterClass->contains(Reg) &&
6060 !ARM::hGPRRegisterClass->contains(Reg))
6062 else if (Subtarget->isThumb1Only() &&
6063 !ARM::tGPRRegisterClass->contains(Reg))
6065 else if (!Subtarget->isThumb() &&
6066 !ARM::GPRRegisterClass->contains(Reg))
6069 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6076 // Mark all former landing pads as non-landing pads. The dispatch is the only
6078 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6079 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6080 (*I)->setIsLandingPad(false);
6082 // The instruction is gone now.
6083 MI->eraseFromParent();
6089 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6090 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6091 E = MBB->succ_end(); I != E; ++I)
6094 llvm_unreachable("Expecting a BB with two successors!");
6098 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6099 MachineBasicBlock *BB) const {
6100 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6101 DebugLoc dl = MI->getDebugLoc();
6102 bool isThumb2 = Subtarget->isThumb2();
6103 switch (MI->getOpcode()) {
6106 llvm_unreachable("Unexpected instr type to insert");
6108 // The Thumb2 pre-indexed stores have the same MI operands, they just
6109 // define them differently in the .td files from the isel patterns, so
6110 // they need pseudos.
6111 case ARM::t2STR_preidx:
6112 MI->setDesc(TII->get(ARM::t2STR_PRE));
6114 case ARM::t2STRB_preidx:
6115 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6117 case ARM::t2STRH_preidx:
6118 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6121 case ARM::STRi_preidx:
6122 case ARM::STRBi_preidx: {
6123 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6124 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6125 // Decode the offset.
6126 unsigned Offset = MI->getOperand(4).getImm();
6127 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6128 Offset = ARM_AM::getAM2Offset(Offset);
6132 MachineMemOperand *MMO = *MI->memoperands_begin();
6133 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6134 .addOperand(MI->getOperand(0)) // Rn_wb
6135 .addOperand(MI->getOperand(1)) // Rt
6136 .addOperand(MI->getOperand(2)) // Rn
6137 .addImm(Offset) // offset (skip GPR==zero_reg)
6138 .addOperand(MI->getOperand(5)) // pred
6139 .addOperand(MI->getOperand(6))
6140 .addMemOperand(MMO);
6141 MI->eraseFromParent();
6144 case ARM::STRr_preidx:
6145 case ARM::STRBr_preidx:
6146 case ARM::STRH_preidx: {
6148 switch (MI->getOpcode()) {
6149 default: llvm_unreachable("unexpected opcode!");
6150 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6151 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6152 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6154 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6155 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6156 MIB.addOperand(MI->getOperand(i));
6157 MI->eraseFromParent();
6160 case ARM::ATOMIC_LOAD_ADD_I8:
6161 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6162 case ARM::ATOMIC_LOAD_ADD_I16:
6163 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6164 case ARM::ATOMIC_LOAD_ADD_I32:
6165 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6167 case ARM::ATOMIC_LOAD_AND_I8:
6168 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6169 case ARM::ATOMIC_LOAD_AND_I16:
6170 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6171 case ARM::ATOMIC_LOAD_AND_I32:
6172 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6174 case ARM::ATOMIC_LOAD_OR_I8:
6175 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6176 case ARM::ATOMIC_LOAD_OR_I16:
6177 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6178 case ARM::ATOMIC_LOAD_OR_I32:
6179 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6181 case ARM::ATOMIC_LOAD_XOR_I8:
6182 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6183 case ARM::ATOMIC_LOAD_XOR_I16:
6184 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6185 case ARM::ATOMIC_LOAD_XOR_I32:
6186 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6188 case ARM::ATOMIC_LOAD_NAND_I8:
6189 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6190 case ARM::ATOMIC_LOAD_NAND_I16:
6191 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6192 case ARM::ATOMIC_LOAD_NAND_I32:
6193 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6195 case ARM::ATOMIC_LOAD_SUB_I8:
6196 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6197 case ARM::ATOMIC_LOAD_SUB_I16:
6198 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6199 case ARM::ATOMIC_LOAD_SUB_I32:
6200 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6202 case ARM::ATOMIC_LOAD_MIN_I8:
6203 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6204 case ARM::ATOMIC_LOAD_MIN_I16:
6205 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6206 case ARM::ATOMIC_LOAD_MIN_I32:
6207 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6209 case ARM::ATOMIC_LOAD_MAX_I8:
6210 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6211 case ARM::ATOMIC_LOAD_MAX_I16:
6212 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6213 case ARM::ATOMIC_LOAD_MAX_I32:
6214 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6216 case ARM::ATOMIC_LOAD_UMIN_I8:
6217 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6218 case ARM::ATOMIC_LOAD_UMIN_I16:
6219 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6220 case ARM::ATOMIC_LOAD_UMIN_I32:
6221 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6223 case ARM::ATOMIC_LOAD_UMAX_I8:
6224 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6225 case ARM::ATOMIC_LOAD_UMAX_I16:
6226 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6227 case ARM::ATOMIC_LOAD_UMAX_I32:
6228 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6230 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6231 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6232 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6234 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6235 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6236 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6239 case ARM::ATOMADD6432:
6240 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6241 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6242 /*NeedsCarry*/ true);
6243 case ARM::ATOMSUB6432:
6244 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6245 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6246 /*NeedsCarry*/ true);
6247 case ARM::ATOMOR6432:
6248 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6249 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6250 case ARM::ATOMXOR6432:
6251 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6252 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6253 case ARM::ATOMAND6432:
6254 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6255 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6256 case ARM::ATOMSWAP6432:
6257 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6258 case ARM::ATOMCMPXCHG6432:
6259 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6260 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6261 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6263 case ARM::tMOVCCr_pseudo: {
6264 // To "insert" a SELECT_CC instruction, we actually have to insert the
6265 // diamond control-flow pattern. The incoming instruction knows the
6266 // destination vreg to set, the condition code register to branch on, the
6267 // true/false values to select between, and a branch opcode to use.
6268 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6269 MachineFunction::iterator It = BB;
6275 // cmpTY ccX, r1, r2
6277 // fallthrough --> copy0MBB
6278 MachineBasicBlock *thisMBB = BB;
6279 MachineFunction *F = BB->getParent();
6280 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6281 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6282 F->insert(It, copy0MBB);
6283 F->insert(It, sinkMBB);
6285 // Transfer the remainder of BB and its successor edges to sinkMBB.
6286 sinkMBB->splice(sinkMBB->begin(), BB,
6287 llvm::next(MachineBasicBlock::iterator(MI)),
6289 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6291 BB->addSuccessor(copy0MBB);
6292 BB->addSuccessor(sinkMBB);
6294 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6295 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6298 // %FalseValue = ...
6299 // # fallthrough to sinkMBB
6302 // Update machine-CFG edges
6303 BB->addSuccessor(sinkMBB);
6306 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6309 BuildMI(*BB, BB->begin(), dl,
6310 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6311 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6312 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6314 MI->eraseFromParent(); // The pseudo instruction is gone now.
6319 case ARM::BCCZi64: {
6320 // If there is an unconditional branch to the other successor, remove it.
6321 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6323 // Compare both parts that make up the double comparison separately for
6325 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6327 unsigned LHS1 = MI->getOperand(1).getReg();
6328 unsigned LHS2 = MI->getOperand(2).getReg();
6330 AddDefaultPred(BuildMI(BB, dl,
6331 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6332 .addReg(LHS1).addImm(0));
6333 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6334 .addReg(LHS2).addImm(0)
6335 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6337 unsigned RHS1 = MI->getOperand(3).getReg();
6338 unsigned RHS2 = MI->getOperand(4).getReg();
6339 AddDefaultPred(BuildMI(BB, dl,
6340 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6341 .addReg(LHS1).addReg(RHS1));
6342 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6343 .addReg(LHS2).addReg(RHS2)
6344 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6347 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6348 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6349 if (MI->getOperand(0).getImm() == ARMCC::NE)
6350 std::swap(destMBB, exitMBB);
6352 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6353 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6355 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6357 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6359 MI->eraseFromParent(); // The pseudo instruction is gone now.
6363 case ARM::Int_eh_sjlj_setjmp:
6364 case ARM::Int_eh_sjlj_setjmp_nofp:
6365 case ARM::tInt_eh_sjlj_setjmp:
6366 case ARM::t2Int_eh_sjlj_setjmp:
6367 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6368 EmitSjLjDispatchBlock(MI, BB);
6373 // To insert an ABS instruction, we have to insert the
6374 // diamond control-flow pattern. The incoming instruction knows the
6375 // source vreg to test against 0, the destination vreg to set,
6376 // the condition code register to branch on, the
6377 // true/false values to select between, and a branch opcode to use.
6382 // BCC (branch to SinkBB if V0 >= 0)
6383 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6384 // SinkBB: V1 = PHI(V2, V3)
6385 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6386 MachineFunction::iterator BBI = BB;
6388 MachineFunction *Fn = BB->getParent();
6389 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6390 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6391 Fn->insert(BBI, RSBBB);
6392 Fn->insert(BBI, SinkBB);
6394 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6395 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6396 bool isThumb2 = Subtarget->isThumb2();
6397 MachineRegisterInfo &MRI = Fn->getRegInfo();
6398 // In Thumb mode S must not be specified if source register is the SP or
6399 // PC and if destination register is the SP, so restrict register class
6400 unsigned NewMovDstReg = MRI.createVirtualRegister(
6401 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6402 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6403 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6405 // Transfer the remainder of BB and its successor edges to sinkMBB.
6406 SinkBB->splice(SinkBB->begin(), BB,
6407 llvm::next(MachineBasicBlock::iterator(MI)),
6409 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6411 BB->addSuccessor(RSBBB);
6412 BB->addSuccessor(SinkBB);
6414 // fall through to SinkMBB
6415 RSBBB->addSuccessor(SinkBB);
6417 // insert a movs at the end of BB
6418 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6420 .addReg(ABSSrcReg, RegState::Kill)
6421 .addImm((unsigned)ARMCC::AL).addReg(0)
6422 .addReg(ARM::CPSR, RegState::Define);
6424 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6426 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6427 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6429 // insert rsbri in RSBBB
6430 // Note: BCC and rsbri will be converted into predicated rsbmi
6431 // by if-conversion pass
6432 BuildMI(*RSBBB, RSBBB->begin(), dl,
6433 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6434 .addReg(NewMovDstReg, RegState::Kill)
6435 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6437 // insert PHI in SinkBB,
6438 // reuse ABSDstReg to not change uses of ABS instruction
6439 BuildMI(*SinkBB, SinkBB->begin(), dl,
6440 TII->get(ARM::PHI), ABSDstReg)
6441 .addReg(NewRsbDstReg).addMBB(RSBBB)
6442 .addReg(NewMovDstReg).addMBB(BB);
6444 // remove ABS instruction
6445 MI->eraseFromParent();
6447 // return last added BB
6453 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6454 SDNode *Node) const {
6455 if (!MI->hasPostISelHook()) {
6456 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6457 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6461 const MCInstrDesc *MCID = &MI->getDesc();
6462 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6463 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6464 // operand is still set to noreg. If needed, set the optional operand's
6465 // register to CPSR, and remove the redundant implicit def.
6467 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6469 // Rename pseudo opcodes.
6470 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6472 const ARMBaseInstrInfo *TII =
6473 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6474 MCID = &TII->get(NewOpc);
6476 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6477 "converted opcode should be the same except for cc_out");
6481 // Add the optional cc_out operand
6482 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6484 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6486 // Any ARM instruction that sets the 's' bit should specify an optional
6487 // "cc_out" operand in the last operand position.
6488 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6489 assert(!NewOpc && "Optional cc_out operand required");
6492 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6493 // since we already have an optional CPSR def.
6494 bool definesCPSR = false;
6495 bool deadCPSR = false;
6496 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6498 const MachineOperand &MO = MI->getOperand(i);
6499 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6503 MI->RemoveOperand(i);
6508 assert(!NewOpc && "Optional cc_out operand required");
6511 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6513 assert(!MI->getOperand(ccOutIdx).getReg() &&
6514 "expect uninitialized optional cc_out operand");
6518 // If this instruction was defined with an optional CPSR def and its dag node
6519 // had a live implicit CPSR def, then activate the optional CPSR def.
6520 MachineOperand &MO = MI->getOperand(ccOutIdx);
6521 MO.setReg(ARM::CPSR);
6525 //===----------------------------------------------------------------------===//
6526 // ARM Optimization Hooks
6527 //===----------------------------------------------------------------------===//
6530 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6531 TargetLowering::DAGCombinerInfo &DCI) {
6532 SelectionDAG &DAG = DCI.DAG;
6533 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6534 EVT VT = N->getValueType(0);
6535 unsigned Opc = N->getOpcode();
6536 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6537 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6538 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6539 ISD::CondCode CC = ISD::SETCC_INVALID;
6542 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6544 SDValue CCOp = Slct.getOperand(0);
6545 if (CCOp.getOpcode() == ISD::SETCC)
6546 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6549 bool DoXform = false;
6551 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6554 if (LHS.getOpcode() == ISD::Constant &&
6555 cast<ConstantSDNode>(LHS)->isNullValue()) {
6557 } else if (CC != ISD::SETCC_INVALID &&
6558 RHS.getOpcode() == ISD::Constant &&
6559 cast<ConstantSDNode>(RHS)->isNullValue()) {
6560 std::swap(LHS, RHS);
6561 SDValue Op0 = Slct.getOperand(0);
6562 EVT OpVT = isSlctCC ? Op0.getValueType() :
6563 Op0.getOperand(0).getValueType();
6564 bool isInt = OpVT.isInteger();
6565 CC = ISD::getSetCCInverse(CC, isInt);
6567 if (!TLI.isCondCodeLegal(CC, OpVT))
6568 return SDValue(); // Inverse operator isn't legal.
6575 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6577 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6578 Slct.getOperand(0), Slct.getOperand(1), CC);
6579 SDValue CCOp = Slct.getOperand(0);
6581 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6582 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6583 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6584 CCOp, OtherOp, Result);
6589 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6590 // (only after legalization).
6591 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6592 TargetLowering::DAGCombinerInfo &DCI,
6593 const ARMSubtarget *Subtarget) {
6595 // Only perform optimization if after legalize, and if NEON is available. We
6596 // also expected both operands to be BUILD_VECTORs.
6597 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6598 || N0.getOpcode() != ISD::BUILD_VECTOR
6599 || N1.getOpcode() != ISD::BUILD_VECTOR)
6602 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6603 EVT VT = N->getValueType(0);
6604 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6607 // Check that the vector operands are of the right form.
6608 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6609 // operands, where N is the size of the formed vector.
6610 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6611 // index such that we have a pair wise add pattern.
6613 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6614 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6616 SDValue Vec = N0->getOperand(0)->getOperand(0);
6617 SDNode *V = Vec.getNode();
6618 unsigned nextIndex = 0;
6620 // For each operands to the ADD which are BUILD_VECTORs,
6621 // check to see if each of their operands are an EXTRACT_VECTOR with
6622 // the same vector and appropriate index.
6623 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6624 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6625 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6627 SDValue ExtVec0 = N0->getOperand(i);
6628 SDValue ExtVec1 = N1->getOperand(i);
6630 // First operand is the vector, verify its the same.
6631 if (V != ExtVec0->getOperand(0).getNode() ||
6632 V != ExtVec1->getOperand(0).getNode())
6635 // Second is the constant, verify its correct.
6636 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6637 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6639 // For the constant, we want to see all the even or all the odd.
6640 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6641 || C1->getZExtValue() != nextIndex+1)
6650 // Create VPADDL node.
6651 SelectionDAG &DAG = DCI.DAG;
6652 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6654 // Build operand list.
6655 SmallVector<SDValue, 8> Ops;
6656 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6657 TLI.getPointerTy()));
6659 // Input is the vector.
6662 // Get widened type and narrowed type.
6664 unsigned numElem = VT.getVectorNumElements();
6665 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6666 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6667 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6668 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6670 assert(0 && "Invalid vector element type for padd optimization.");
6673 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6674 widenType, &Ops[0], Ops.size());
6675 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6678 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6679 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6680 /// called with the default operands, and if that fails, with commuted
6682 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6683 TargetLowering::DAGCombinerInfo &DCI,
6684 const ARMSubtarget *Subtarget){
6686 // Attempt to create vpaddl for this add.
6687 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6688 if (Result.getNode())
6691 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6692 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6693 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6694 if (Result.getNode()) return Result;
6699 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6701 static SDValue PerformADDCombine(SDNode *N,
6702 TargetLowering::DAGCombinerInfo &DCI,
6703 const ARMSubtarget *Subtarget) {
6704 SDValue N0 = N->getOperand(0);
6705 SDValue N1 = N->getOperand(1);
6707 // First try with the default operand order.
6708 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6709 if (Result.getNode())
6712 // If that didn't work, try again with the operands commuted.
6713 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6716 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6718 static SDValue PerformSUBCombine(SDNode *N,
6719 TargetLowering::DAGCombinerInfo &DCI) {
6720 SDValue N0 = N->getOperand(0);
6721 SDValue N1 = N->getOperand(1);
6723 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6724 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6725 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6726 if (Result.getNode()) return Result;
6732 /// PerformVMULCombine
6733 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6734 /// special multiplier accumulator forwarding.
6740 static SDValue PerformVMULCombine(SDNode *N,
6741 TargetLowering::DAGCombinerInfo &DCI,
6742 const ARMSubtarget *Subtarget) {
6743 if (!Subtarget->hasVMLxForwarding())
6746 SelectionDAG &DAG = DCI.DAG;
6747 SDValue N0 = N->getOperand(0);
6748 SDValue N1 = N->getOperand(1);
6749 unsigned Opcode = N0.getOpcode();
6750 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6751 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6752 Opcode = N1.getOpcode();
6753 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6754 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6759 EVT VT = N->getValueType(0);
6760 DebugLoc DL = N->getDebugLoc();
6761 SDValue N00 = N0->getOperand(0);
6762 SDValue N01 = N0->getOperand(1);
6763 return DAG.getNode(Opcode, DL, VT,
6764 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6765 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6768 static SDValue PerformMULCombine(SDNode *N,
6769 TargetLowering::DAGCombinerInfo &DCI,
6770 const ARMSubtarget *Subtarget) {
6771 SelectionDAG &DAG = DCI.DAG;
6773 if (Subtarget->isThumb1Only())
6776 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6779 EVT VT = N->getValueType(0);
6780 if (VT.is64BitVector() || VT.is128BitVector())
6781 return PerformVMULCombine(N, DCI, Subtarget);
6785 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6789 uint64_t MulAmt = C->getZExtValue();
6790 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6791 ShiftAmt = ShiftAmt & (32 - 1);
6792 SDValue V = N->getOperand(0);
6793 DebugLoc DL = N->getDebugLoc();
6796 MulAmt >>= ShiftAmt;
6797 if (isPowerOf2_32(MulAmt - 1)) {
6798 // (mul x, 2^N + 1) => (add (shl x, N), x)
6799 Res = DAG.getNode(ISD::ADD, DL, VT,
6800 V, DAG.getNode(ISD::SHL, DL, VT,
6801 V, DAG.getConstant(Log2_32(MulAmt-1),
6803 } else if (isPowerOf2_32(MulAmt + 1)) {
6804 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6805 Res = DAG.getNode(ISD::SUB, DL, VT,
6806 DAG.getNode(ISD::SHL, DL, VT,
6807 V, DAG.getConstant(Log2_32(MulAmt+1),
6814 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6815 DAG.getConstant(ShiftAmt, MVT::i32));
6817 // Do not add new nodes to DAG combiner worklist.
6818 DCI.CombineTo(N, Res, false);
6822 static SDValue PerformANDCombine(SDNode *N,
6823 TargetLowering::DAGCombinerInfo &DCI) {
6825 // Attempt to use immediate-form VBIC
6826 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6827 DebugLoc dl = N->getDebugLoc();
6828 EVT VT = N->getValueType(0);
6829 SelectionDAG &DAG = DCI.DAG;
6831 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6834 APInt SplatBits, SplatUndef;
6835 unsigned SplatBitSize;
6838 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6839 if (SplatBitSize <= 64) {
6841 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6842 SplatUndef.getZExtValue(), SplatBitSize,
6843 DAG, VbicVT, VT.is128BitVector(),
6845 if (Val.getNode()) {
6847 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6848 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6849 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6857 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6858 static SDValue PerformORCombine(SDNode *N,
6859 TargetLowering::DAGCombinerInfo &DCI,
6860 const ARMSubtarget *Subtarget) {
6861 // Attempt to use immediate-form VORR
6862 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6863 DebugLoc dl = N->getDebugLoc();
6864 EVT VT = N->getValueType(0);
6865 SelectionDAG &DAG = DCI.DAG;
6867 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6870 APInt SplatBits, SplatUndef;
6871 unsigned SplatBitSize;
6873 if (BVN && Subtarget->hasNEON() &&
6874 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6875 if (SplatBitSize <= 64) {
6877 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6878 SplatUndef.getZExtValue(), SplatBitSize,
6879 DAG, VorrVT, VT.is128BitVector(),
6881 if (Val.getNode()) {
6883 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6884 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6885 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6890 SDValue N0 = N->getOperand(0);
6891 if (N0.getOpcode() != ISD::AND)
6893 SDValue N1 = N->getOperand(1);
6895 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6896 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6897 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6899 unsigned SplatBitSize;
6902 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6904 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6905 HasAnyUndefs) && !HasAnyUndefs) {
6906 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6908 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6909 HasAnyUndefs) && !HasAnyUndefs &&
6910 SplatBits0 == ~SplatBits1) {
6911 // Canonicalize the vector type to make instruction selection simpler.
6912 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6913 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6914 N0->getOperand(1), N0->getOperand(0),
6916 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6921 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6924 // BFI is only available on V6T2+
6925 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6928 DebugLoc DL = N->getDebugLoc();
6929 // 1) or (and A, mask), val => ARMbfi A, val, mask
6930 // iff (val & mask) == val
6932 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6933 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6934 // && mask == ~mask2
6935 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6936 // && ~mask == mask2
6937 // (i.e., copy a bitfield value into another bitfield of the same width)
6942 SDValue N00 = N0.getOperand(0);
6944 // The value and the mask need to be constants so we can verify this is
6945 // actually a bitfield set. If the mask is 0xffff, we can do better
6946 // via a movt instruction, so don't use BFI in that case.
6947 SDValue MaskOp = N0.getOperand(1);
6948 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6951 unsigned Mask = MaskC->getZExtValue();
6955 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6956 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6958 unsigned Val = N1C->getZExtValue();
6959 if ((Val & ~Mask) != Val)
6962 if (ARM::isBitFieldInvertedMask(Mask)) {
6963 Val >>= CountTrailingZeros_32(~Mask);
6965 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6966 DAG.getConstant(Val, MVT::i32),
6967 DAG.getConstant(Mask, MVT::i32));
6969 // Do not add new nodes to DAG combiner worklist.
6970 DCI.CombineTo(N, Res, false);
6973 } else if (N1.getOpcode() == ISD::AND) {
6974 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6975 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6978 unsigned Mask2 = N11C->getZExtValue();
6980 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
6982 if (ARM::isBitFieldInvertedMask(Mask) &&
6984 // The pack halfword instruction works better for masks that fit it,
6985 // so use that when it's available.
6986 if (Subtarget->hasT2ExtractPack() &&
6987 (Mask == 0xffff || Mask == 0xffff0000))
6990 unsigned amt = CountTrailingZeros_32(Mask2);
6991 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6992 DAG.getConstant(amt, MVT::i32));
6993 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6994 DAG.getConstant(Mask, MVT::i32));
6995 // Do not add new nodes to DAG combiner worklist.
6996 DCI.CombineTo(N, Res, false);
6998 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7000 // The pack halfword instruction works better for masks that fit it,
7001 // so use that when it's available.
7002 if (Subtarget->hasT2ExtractPack() &&
7003 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7006 unsigned lsb = CountTrailingZeros_32(Mask);
7007 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7008 DAG.getConstant(lsb, MVT::i32));
7009 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7010 DAG.getConstant(Mask2, MVT::i32));
7011 // Do not add new nodes to DAG combiner worklist.
7012 DCI.CombineTo(N, Res, false);
7017 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7018 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7019 ARM::isBitFieldInvertedMask(~Mask)) {
7020 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7021 // where lsb(mask) == #shamt and masked bits of B are known zero.
7022 SDValue ShAmt = N00.getOperand(1);
7023 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7024 unsigned LSB = CountTrailingZeros_32(Mask);
7028 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7029 DAG.getConstant(~Mask, MVT::i32));
7031 // Do not add new nodes to DAG combiner worklist.
7032 DCI.CombineTo(N, Res, false);
7038 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7039 /// the bits being cleared by the AND are not demanded by the BFI.
7040 static SDValue PerformBFICombine(SDNode *N,
7041 TargetLowering::DAGCombinerInfo &DCI) {
7042 SDValue N1 = N->getOperand(1);
7043 if (N1.getOpcode() == ISD::AND) {
7044 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7047 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7048 unsigned LSB = CountTrailingZeros_32(~InvMask);
7049 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7050 unsigned Mask = (1 << Width)-1;
7051 unsigned Mask2 = N11C->getZExtValue();
7052 if ((Mask & (~Mask2)) == 0)
7053 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7054 N->getOperand(0), N1.getOperand(0),
7060 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7061 /// ARMISD::VMOVRRD.
7062 static SDValue PerformVMOVRRDCombine(SDNode *N,
7063 TargetLowering::DAGCombinerInfo &DCI) {
7064 // vmovrrd(vmovdrr x, y) -> x,y
7065 SDValue InDouble = N->getOperand(0);
7066 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7067 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7069 // vmovrrd(load f64) -> (load i32), (load i32)
7070 SDNode *InNode = InDouble.getNode();
7071 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7072 InNode->getValueType(0) == MVT::f64 &&
7073 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7074 !cast<LoadSDNode>(InNode)->isVolatile()) {
7075 // TODO: Should this be done for non-FrameIndex operands?
7076 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7078 SelectionDAG &DAG = DCI.DAG;
7079 DebugLoc DL = LD->getDebugLoc();
7080 SDValue BasePtr = LD->getBasePtr();
7081 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7082 LD->getPointerInfo(), LD->isVolatile(),
7083 LD->isNonTemporal(), LD->isInvariant(),
7084 LD->getAlignment());
7086 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7087 DAG.getConstant(4, MVT::i32));
7088 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7089 LD->getPointerInfo(), LD->isVolatile(),
7090 LD->isNonTemporal(), LD->isInvariant(),
7091 std::min(4U, LD->getAlignment() / 2));
7093 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7094 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7095 DCI.RemoveFromWorklist(LD);
7103 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7104 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7105 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7106 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7107 SDValue Op0 = N->getOperand(0);
7108 SDValue Op1 = N->getOperand(1);
7109 if (Op0.getOpcode() == ISD::BITCAST)
7110 Op0 = Op0.getOperand(0);
7111 if (Op1.getOpcode() == ISD::BITCAST)
7112 Op1 = Op1.getOperand(0);
7113 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7114 Op0.getNode() == Op1.getNode() &&
7115 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7116 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7117 N->getValueType(0), Op0.getOperand(0));
7121 /// PerformSTORECombine - Target-specific dag combine xforms for
7123 static SDValue PerformSTORECombine(SDNode *N,
7124 TargetLowering::DAGCombinerInfo &DCI) {
7125 // Bitcast an i64 store extracted from a vector to f64.
7126 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7127 StoreSDNode *St = cast<StoreSDNode>(N);
7128 SDValue StVal = St->getValue();
7129 if (!ISD::isNormalStore(St) || St->isVolatile())
7132 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7133 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7134 SelectionDAG &DAG = DCI.DAG;
7135 DebugLoc DL = St->getDebugLoc();
7136 SDValue BasePtr = St->getBasePtr();
7137 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7138 StVal.getNode()->getOperand(0), BasePtr,
7139 St->getPointerInfo(), St->isVolatile(),
7140 St->isNonTemporal(), St->getAlignment());
7142 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7143 DAG.getConstant(4, MVT::i32));
7144 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7145 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7146 St->isNonTemporal(),
7147 std::min(4U, St->getAlignment() / 2));
7150 if (StVal.getValueType() != MVT::i64 ||
7151 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7154 SelectionDAG &DAG = DCI.DAG;
7155 DebugLoc dl = StVal.getDebugLoc();
7156 SDValue IntVec = StVal.getOperand(0);
7157 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7158 IntVec.getValueType().getVectorNumElements());
7159 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7160 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7161 Vec, StVal.getOperand(1));
7162 dl = N->getDebugLoc();
7163 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7164 // Make the DAGCombiner fold the bitcasts.
7165 DCI.AddToWorklist(Vec.getNode());
7166 DCI.AddToWorklist(ExtElt.getNode());
7167 DCI.AddToWorklist(V.getNode());
7168 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7169 St->getPointerInfo(), St->isVolatile(),
7170 St->isNonTemporal(), St->getAlignment(),
7174 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7175 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7176 /// i64 vector to have f64 elements, since the value can then be loaded
7177 /// directly into a VFP register.
7178 static bool hasNormalLoadOperand(SDNode *N) {
7179 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7180 for (unsigned i = 0; i < NumElts; ++i) {
7181 SDNode *Elt = N->getOperand(i).getNode();
7182 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7188 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7189 /// ISD::BUILD_VECTOR.
7190 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7191 TargetLowering::DAGCombinerInfo &DCI){
7192 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7193 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7194 // into a pair of GPRs, which is fine when the value is used as a scalar,
7195 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7196 SelectionDAG &DAG = DCI.DAG;
7197 if (N->getNumOperands() == 2) {
7198 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7203 // Load i64 elements as f64 values so that type legalization does not split
7204 // them up into i32 values.
7205 EVT VT = N->getValueType(0);
7206 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7208 DebugLoc dl = N->getDebugLoc();
7209 SmallVector<SDValue, 8> Ops;
7210 unsigned NumElts = VT.getVectorNumElements();
7211 for (unsigned i = 0; i < NumElts; ++i) {
7212 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7214 // Make the DAGCombiner fold the bitcast.
7215 DCI.AddToWorklist(V.getNode());
7217 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7218 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7219 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7222 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7223 /// ISD::INSERT_VECTOR_ELT.
7224 static SDValue PerformInsertEltCombine(SDNode *N,
7225 TargetLowering::DAGCombinerInfo &DCI) {
7226 // Bitcast an i64 load inserted into a vector to f64.
7227 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7228 EVT VT = N->getValueType(0);
7229 SDNode *Elt = N->getOperand(1).getNode();
7230 if (VT.getVectorElementType() != MVT::i64 ||
7231 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7234 SelectionDAG &DAG = DCI.DAG;
7235 DebugLoc dl = N->getDebugLoc();
7236 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7237 VT.getVectorNumElements());
7238 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7239 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7240 // Make the DAGCombiner fold the bitcasts.
7241 DCI.AddToWorklist(Vec.getNode());
7242 DCI.AddToWorklist(V.getNode());
7243 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7244 Vec, V, N->getOperand(2));
7245 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7248 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7249 /// ISD::VECTOR_SHUFFLE.
7250 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7251 // The LLVM shufflevector instruction does not require the shuffle mask
7252 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7253 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7254 // operands do not match the mask length, they are extended by concatenating
7255 // them with undef vectors. That is probably the right thing for other
7256 // targets, but for NEON it is better to concatenate two double-register
7257 // size vector operands into a single quad-register size vector. Do that
7258 // transformation here:
7259 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7260 // shuffle(concat(v1, v2), undef)
7261 SDValue Op0 = N->getOperand(0);
7262 SDValue Op1 = N->getOperand(1);
7263 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7264 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7265 Op0.getNumOperands() != 2 ||
7266 Op1.getNumOperands() != 2)
7268 SDValue Concat0Op1 = Op0.getOperand(1);
7269 SDValue Concat1Op1 = Op1.getOperand(1);
7270 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7271 Concat1Op1.getOpcode() != ISD::UNDEF)
7273 // Skip the transformation if any of the types are illegal.
7274 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7275 EVT VT = N->getValueType(0);
7276 if (!TLI.isTypeLegal(VT) ||
7277 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7278 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7281 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7282 Op0.getOperand(0), Op1.getOperand(0));
7283 // Translate the shuffle mask.
7284 SmallVector<int, 16> NewMask;
7285 unsigned NumElts = VT.getVectorNumElements();
7286 unsigned HalfElts = NumElts/2;
7287 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7288 for (unsigned n = 0; n < NumElts; ++n) {
7289 int MaskElt = SVN->getMaskElt(n);
7291 if (MaskElt < (int)HalfElts)
7293 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7294 NewElt = HalfElts + MaskElt - NumElts;
7295 NewMask.push_back(NewElt);
7297 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7298 DAG.getUNDEF(VT), NewMask.data());
7301 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7302 /// NEON load/store intrinsics to merge base address updates.
7303 static SDValue CombineBaseUpdate(SDNode *N,
7304 TargetLowering::DAGCombinerInfo &DCI) {
7305 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7308 SelectionDAG &DAG = DCI.DAG;
7309 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7310 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7311 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7312 SDValue Addr = N->getOperand(AddrOpIdx);
7314 // Search for a use of the address operand that is an increment.
7315 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7316 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7318 if (User->getOpcode() != ISD::ADD ||
7319 UI.getUse().getResNo() != Addr.getResNo())
7322 // Check that the add is independent of the load/store. Otherwise, folding
7323 // it would create a cycle.
7324 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7327 // Find the new opcode for the updating load/store.
7329 bool isLaneOp = false;
7330 unsigned NewOpc = 0;
7331 unsigned NumVecs = 0;
7333 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7335 default: assert(0 && "unexpected intrinsic for Neon base update");
7336 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7338 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7340 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7342 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7344 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7345 NumVecs = 2; isLaneOp = true; break;
7346 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7347 NumVecs = 3; isLaneOp = true; break;
7348 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7349 NumVecs = 4; isLaneOp = true; break;
7350 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7351 NumVecs = 1; isLoad = false; break;
7352 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7353 NumVecs = 2; isLoad = false; break;
7354 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7355 NumVecs = 3; isLoad = false; break;
7356 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7357 NumVecs = 4; isLoad = false; break;
7358 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7359 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7360 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7361 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7362 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7363 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7367 switch (N->getOpcode()) {
7368 default: assert(0 && "unexpected opcode for Neon base update");
7369 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7370 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7371 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7375 // Find the size of memory referenced by the load/store.
7378 VecTy = N->getValueType(0);
7380 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7381 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7383 NumBytes /= VecTy.getVectorNumElements();
7385 // If the increment is a constant, it must match the memory ref size.
7386 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7387 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7388 uint64_t IncVal = CInc->getZExtValue();
7389 if (IncVal != NumBytes)
7391 } else if (NumBytes >= 3 * 16) {
7392 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7393 // separate instructions that make it harder to use a non-constant update.
7397 // Create the new updating load/store node.
7399 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7401 for (n = 0; n < NumResultVecs; ++n)
7403 Tys[n++] = MVT::i32;
7404 Tys[n] = MVT::Other;
7405 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7406 SmallVector<SDValue, 8> Ops;
7407 Ops.push_back(N->getOperand(0)); // incoming chain
7408 Ops.push_back(N->getOperand(AddrOpIdx));
7410 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7411 Ops.push_back(N->getOperand(i));
7413 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7414 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7415 Ops.data(), Ops.size(),
7416 MemInt->getMemoryVT(),
7417 MemInt->getMemOperand());
7420 std::vector<SDValue> NewResults;
7421 for (unsigned i = 0; i < NumResultVecs; ++i) {
7422 NewResults.push_back(SDValue(UpdN.getNode(), i));
7424 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7425 DCI.CombineTo(N, NewResults);
7426 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7433 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7434 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7435 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7437 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7438 SelectionDAG &DAG = DCI.DAG;
7439 EVT VT = N->getValueType(0);
7440 // vldN-dup instructions only support 64-bit vectors for N > 1.
7441 if (!VT.is64BitVector())
7444 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7445 SDNode *VLD = N->getOperand(0).getNode();
7446 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7448 unsigned NumVecs = 0;
7449 unsigned NewOpc = 0;
7450 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7451 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7453 NewOpc = ARMISD::VLD2DUP;
7454 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7456 NewOpc = ARMISD::VLD3DUP;
7457 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7459 NewOpc = ARMISD::VLD4DUP;
7464 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7465 // numbers match the load.
7466 unsigned VLDLaneNo =
7467 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7468 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7470 // Ignore uses of the chain result.
7471 if (UI.getUse().getResNo() == NumVecs)
7474 if (User->getOpcode() != ARMISD::VDUPLANE ||
7475 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7479 // Create the vldN-dup node.
7482 for (n = 0; n < NumVecs; ++n)
7484 Tys[n] = MVT::Other;
7485 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7486 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7487 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7488 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7489 Ops, 2, VLDMemInt->getMemoryVT(),
7490 VLDMemInt->getMemOperand());
7493 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7495 unsigned ResNo = UI.getUse().getResNo();
7496 // Ignore uses of the chain result.
7497 if (ResNo == NumVecs)
7500 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7503 // Now the vldN-lane intrinsic is dead except for its chain result.
7504 // Update uses of the chain.
7505 std::vector<SDValue> VLDDupResults;
7506 for (unsigned n = 0; n < NumVecs; ++n)
7507 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7508 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7509 DCI.CombineTo(VLD, VLDDupResults);
7514 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7515 /// ARMISD::VDUPLANE.
7516 static SDValue PerformVDUPLANECombine(SDNode *N,
7517 TargetLowering::DAGCombinerInfo &DCI) {
7518 SDValue Op = N->getOperand(0);
7520 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7521 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7522 if (CombineVLDDUP(N, DCI))
7523 return SDValue(N, 0);
7525 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7526 // redundant. Ignore bit_converts for now; element sizes are checked below.
7527 while (Op.getOpcode() == ISD::BITCAST)
7528 Op = Op.getOperand(0);
7529 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7532 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7533 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7534 // The canonical VMOV for a zero vector uses a 32-bit element size.
7535 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7537 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7539 EVT VT = N->getValueType(0);
7540 if (EltSize > VT.getVectorElementType().getSizeInBits())
7543 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7546 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7547 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7548 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7552 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7554 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7559 APFloat APF = C->getValueAPF();
7560 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7561 != APFloat::opOK || !isExact)
7564 c0 = (I == 0) ? cN : c0;
7565 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7572 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7573 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7574 /// when the VMUL has a constant operand that is a power of 2.
7576 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7577 /// vmul.f32 d16, d17, d16
7578 /// vcvt.s32.f32 d16, d16
7580 /// vcvt.s32.f32 d16, d16, #3
7581 static SDValue PerformVCVTCombine(SDNode *N,
7582 TargetLowering::DAGCombinerInfo &DCI,
7583 const ARMSubtarget *Subtarget) {
7584 SelectionDAG &DAG = DCI.DAG;
7585 SDValue Op = N->getOperand(0);
7587 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7588 Op.getOpcode() != ISD::FMUL)
7592 SDValue N0 = Op->getOperand(0);
7593 SDValue ConstVec = Op->getOperand(1);
7594 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7596 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7597 !isConstVecPow2(ConstVec, isSigned, C))
7600 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7601 Intrinsic::arm_neon_vcvtfp2fxu;
7602 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7604 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7605 DAG.getConstant(Log2_64(C), MVT::i32));
7608 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7609 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7610 /// when the VDIV has a constant operand that is a power of 2.
7612 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7613 /// vcvt.f32.s32 d16, d16
7614 /// vdiv.f32 d16, d17, d16
7616 /// vcvt.f32.s32 d16, d16, #3
7617 static SDValue PerformVDIVCombine(SDNode *N,
7618 TargetLowering::DAGCombinerInfo &DCI,
7619 const ARMSubtarget *Subtarget) {
7620 SelectionDAG &DAG = DCI.DAG;
7621 SDValue Op = N->getOperand(0);
7622 unsigned OpOpcode = Op.getNode()->getOpcode();
7624 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7625 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7629 SDValue ConstVec = N->getOperand(1);
7630 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7632 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7633 !isConstVecPow2(ConstVec, isSigned, C))
7636 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7637 Intrinsic::arm_neon_vcvtfxu2fp;
7638 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7640 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7641 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7644 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7645 /// operand of a vector shift operation, where all the elements of the
7646 /// build_vector must have the same constant integer value.
7647 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7648 // Ignore bit_converts.
7649 while (Op.getOpcode() == ISD::BITCAST)
7650 Op = Op.getOperand(0);
7651 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7652 APInt SplatBits, SplatUndef;
7653 unsigned SplatBitSize;
7655 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7656 HasAnyUndefs, ElementBits) ||
7657 SplatBitSize > ElementBits)
7659 Cnt = SplatBits.getSExtValue();
7663 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7664 /// operand of a vector shift left operation. That value must be in the range:
7665 /// 0 <= Value < ElementBits for a left shift; or
7666 /// 0 <= Value <= ElementBits for a long left shift.
7667 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7668 assert(VT.isVector() && "vector shift count is not a vector type");
7669 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7670 if (! getVShiftImm(Op, ElementBits, Cnt))
7672 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7675 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7676 /// operand of a vector shift right operation. For a shift opcode, the value
7677 /// is positive, but for an intrinsic the value count must be negative. The
7678 /// absolute value must be in the range:
7679 /// 1 <= |Value| <= ElementBits for a right shift; or
7680 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7681 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7683 assert(VT.isVector() && "vector shift count is not a vector type");
7684 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7685 if (! getVShiftImm(Op, ElementBits, Cnt))
7689 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7692 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7693 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7694 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7697 // Don't do anything for most intrinsics.
7700 // Vector shifts: check for immediate versions and lower them.
7701 // Note: This is done during DAG combining instead of DAG legalizing because
7702 // the build_vectors for 64-bit vector element shift counts are generally
7703 // not legal, and it is hard to see their values after they get legalized to
7704 // loads from a constant pool.
7705 case Intrinsic::arm_neon_vshifts:
7706 case Intrinsic::arm_neon_vshiftu:
7707 case Intrinsic::arm_neon_vshiftls:
7708 case Intrinsic::arm_neon_vshiftlu:
7709 case Intrinsic::arm_neon_vshiftn:
7710 case Intrinsic::arm_neon_vrshifts:
7711 case Intrinsic::arm_neon_vrshiftu:
7712 case Intrinsic::arm_neon_vrshiftn:
7713 case Intrinsic::arm_neon_vqshifts:
7714 case Intrinsic::arm_neon_vqshiftu:
7715 case Intrinsic::arm_neon_vqshiftsu:
7716 case Intrinsic::arm_neon_vqshiftns:
7717 case Intrinsic::arm_neon_vqshiftnu:
7718 case Intrinsic::arm_neon_vqshiftnsu:
7719 case Intrinsic::arm_neon_vqrshiftns:
7720 case Intrinsic::arm_neon_vqrshiftnu:
7721 case Intrinsic::arm_neon_vqrshiftnsu: {
7722 EVT VT = N->getOperand(1).getValueType();
7724 unsigned VShiftOpc = 0;
7727 case Intrinsic::arm_neon_vshifts:
7728 case Intrinsic::arm_neon_vshiftu:
7729 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7730 VShiftOpc = ARMISD::VSHL;
7733 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7734 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7735 ARMISD::VSHRs : ARMISD::VSHRu);
7740 case Intrinsic::arm_neon_vshiftls:
7741 case Intrinsic::arm_neon_vshiftlu:
7742 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7744 llvm_unreachable("invalid shift count for vshll intrinsic");
7746 case Intrinsic::arm_neon_vrshifts:
7747 case Intrinsic::arm_neon_vrshiftu:
7748 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7752 case Intrinsic::arm_neon_vqshifts:
7753 case Intrinsic::arm_neon_vqshiftu:
7754 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7758 case Intrinsic::arm_neon_vqshiftsu:
7759 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7761 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7763 case Intrinsic::arm_neon_vshiftn:
7764 case Intrinsic::arm_neon_vrshiftn:
7765 case Intrinsic::arm_neon_vqshiftns:
7766 case Intrinsic::arm_neon_vqshiftnu:
7767 case Intrinsic::arm_neon_vqshiftnsu:
7768 case Intrinsic::arm_neon_vqrshiftns:
7769 case Intrinsic::arm_neon_vqrshiftnu:
7770 case Intrinsic::arm_neon_vqrshiftnsu:
7771 // Narrowing shifts require an immediate right shift.
7772 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7774 llvm_unreachable("invalid shift count for narrowing vector shift "
7778 llvm_unreachable("unhandled vector shift");
7782 case Intrinsic::arm_neon_vshifts:
7783 case Intrinsic::arm_neon_vshiftu:
7784 // Opcode already set above.
7786 case Intrinsic::arm_neon_vshiftls:
7787 case Intrinsic::arm_neon_vshiftlu:
7788 if (Cnt == VT.getVectorElementType().getSizeInBits())
7789 VShiftOpc = ARMISD::VSHLLi;
7791 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7792 ARMISD::VSHLLs : ARMISD::VSHLLu);
7794 case Intrinsic::arm_neon_vshiftn:
7795 VShiftOpc = ARMISD::VSHRN; break;
7796 case Intrinsic::arm_neon_vrshifts:
7797 VShiftOpc = ARMISD::VRSHRs; break;
7798 case Intrinsic::arm_neon_vrshiftu:
7799 VShiftOpc = ARMISD::VRSHRu; break;
7800 case Intrinsic::arm_neon_vrshiftn:
7801 VShiftOpc = ARMISD::VRSHRN; break;
7802 case Intrinsic::arm_neon_vqshifts:
7803 VShiftOpc = ARMISD::VQSHLs; break;
7804 case Intrinsic::arm_neon_vqshiftu:
7805 VShiftOpc = ARMISD::VQSHLu; break;
7806 case Intrinsic::arm_neon_vqshiftsu:
7807 VShiftOpc = ARMISD::VQSHLsu; break;
7808 case Intrinsic::arm_neon_vqshiftns:
7809 VShiftOpc = ARMISD::VQSHRNs; break;
7810 case Intrinsic::arm_neon_vqshiftnu:
7811 VShiftOpc = ARMISD::VQSHRNu; break;
7812 case Intrinsic::arm_neon_vqshiftnsu:
7813 VShiftOpc = ARMISD::VQSHRNsu; break;
7814 case Intrinsic::arm_neon_vqrshiftns:
7815 VShiftOpc = ARMISD::VQRSHRNs; break;
7816 case Intrinsic::arm_neon_vqrshiftnu:
7817 VShiftOpc = ARMISD::VQRSHRNu; break;
7818 case Intrinsic::arm_neon_vqrshiftnsu:
7819 VShiftOpc = ARMISD::VQRSHRNsu; break;
7822 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7823 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7826 case Intrinsic::arm_neon_vshiftins: {
7827 EVT VT = N->getOperand(1).getValueType();
7829 unsigned VShiftOpc = 0;
7831 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7832 VShiftOpc = ARMISD::VSLI;
7833 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7834 VShiftOpc = ARMISD::VSRI;
7836 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7839 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7840 N->getOperand(1), N->getOperand(2),
7841 DAG.getConstant(Cnt, MVT::i32));
7844 case Intrinsic::arm_neon_vqrshifts:
7845 case Intrinsic::arm_neon_vqrshiftu:
7846 // No immediate versions of these to check for.
7853 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
7854 /// lowers them. As with the vector shift intrinsics, this is done during DAG
7855 /// combining instead of DAG legalizing because the build_vectors for 64-bit
7856 /// vector element shift counts are generally not legal, and it is hard to see
7857 /// their values after they get legalized to loads from a constant pool.
7858 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7859 const ARMSubtarget *ST) {
7860 EVT VT = N->getValueType(0);
7862 // Nothing to be done for scalar shifts.
7863 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7864 if (!VT.isVector() || !TLI.isTypeLegal(VT))
7867 assert(ST->hasNEON() && "unexpected vector shift");
7870 switch (N->getOpcode()) {
7871 default: llvm_unreachable("unexpected shift opcode");
7874 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7875 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7876 DAG.getConstant(Cnt, MVT::i32));
7881 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7882 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7883 ARMISD::VSHRs : ARMISD::VSHRu);
7884 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7885 DAG.getConstant(Cnt, MVT::i32));
7891 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7892 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7893 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7894 const ARMSubtarget *ST) {
7895 SDValue N0 = N->getOperand(0);
7897 // Check for sign- and zero-extensions of vector extract operations of 8-
7898 // and 16-bit vector elements. NEON supports these directly. They are
7899 // handled during DAG combining because type legalization will promote them
7900 // to 32-bit types and it is messy to recognize the operations after that.
7901 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7902 SDValue Vec = N0.getOperand(0);
7903 SDValue Lane = N0.getOperand(1);
7904 EVT VT = N->getValueType(0);
7905 EVT EltVT = N0.getValueType();
7906 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7908 if (VT == MVT::i32 &&
7909 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7910 TLI.isTypeLegal(Vec.getValueType()) &&
7911 isa<ConstantSDNode>(Lane)) {
7914 switch (N->getOpcode()) {
7915 default: llvm_unreachable("unexpected opcode");
7916 case ISD::SIGN_EXTEND:
7917 Opc = ARMISD::VGETLANEs;
7919 case ISD::ZERO_EXTEND:
7920 case ISD::ANY_EXTEND:
7921 Opc = ARMISD::VGETLANEu;
7924 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7931 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7932 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7933 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7934 const ARMSubtarget *ST) {
7935 // If the target supports NEON, try to use vmax/vmin instructions for f32
7936 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
7937 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
7938 // a NaN; only do the transformation when it matches that behavior.
7940 // For now only do this when using NEON for FP operations; if using VFP, it
7941 // is not obvious that the benefit outweighs the cost of switching to the
7943 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7944 N->getValueType(0) != MVT::f32)
7947 SDValue CondLHS = N->getOperand(0);
7948 SDValue CondRHS = N->getOperand(1);
7949 SDValue LHS = N->getOperand(2);
7950 SDValue RHS = N->getOperand(3);
7951 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7953 unsigned Opcode = 0;
7955 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7956 IsReversed = false; // x CC y ? x : y
7957 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7958 IsReversed = true ; // x CC y ? y : x
7972 // If LHS is NaN, an ordered comparison will be false and the result will
7973 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
7974 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7975 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
7976 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7978 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
7979 // will return -0, so vmin can only be used for unsafe math or if one of
7980 // the operands is known to be nonzero.
7981 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
7982 !DAG.getTarget().Options.UnsafeFPMath &&
7983 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7985 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7994 // If LHS is NaN, an ordered comparison will be false and the result will
7995 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
7996 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7997 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7998 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8000 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8001 // will return +0, so vmax can only be used for unsafe math or if one of
8002 // the operands is known to be nonzero.
8003 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8004 !DAG.getTarget().Options.UnsafeFPMath &&
8005 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8007 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8013 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8016 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8018 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8019 SDValue Cmp = N->getOperand(4);
8020 if (Cmp.getOpcode() != ARMISD::CMPZ)
8021 // Only looking at EQ and NE cases.
8024 EVT VT = N->getValueType(0);
8025 DebugLoc dl = N->getDebugLoc();
8026 SDValue LHS = Cmp.getOperand(0);
8027 SDValue RHS = Cmp.getOperand(1);
8028 SDValue FalseVal = N->getOperand(0);
8029 SDValue TrueVal = N->getOperand(1);
8030 SDValue ARMcc = N->getOperand(2);
8031 ARMCC::CondCodes CC =
8032 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8050 /// FIXME: Turn this into a target neutral optimization?
8052 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8053 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8054 N->getOperand(3), Cmp);
8055 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8057 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8058 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8059 N->getOperand(3), NewCmp);
8062 if (Res.getNode()) {
8063 APInt KnownZero, KnownOne;
8064 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8065 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8066 // Capture demanded bits information that would be otherwise lost.
8067 if (KnownZero == 0xfffffffe)
8068 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8069 DAG.getValueType(MVT::i1));
8070 else if (KnownZero == 0xffffff00)
8071 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8072 DAG.getValueType(MVT::i8));
8073 else if (KnownZero == 0xffff0000)
8074 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8075 DAG.getValueType(MVT::i16));
8081 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8082 DAGCombinerInfo &DCI) const {
8083 switch (N->getOpcode()) {
8085 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8086 case ISD::SUB: return PerformSUBCombine(N, DCI);
8087 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8088 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8089 case ISD::AND: return PerformANDCombine(N, DCI);
8090 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8091 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8092 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8093 case ISD::STORE: return PerformSTORECombine(N, DCI);
8094 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8095 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8096 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8097 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8098 case ISD::FP_TO_SINT:
8099 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8100 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8101 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8104 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8105 case ISD::SIGN_EXTEND:
8106 case ISD::ZERO_EXTEND:
8107 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8108 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8109 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8110 case ARMISD::VLD2DUP:
8111 case ARMISD::VLD3DUP:
8112 case ARMISD::VLD4DUP:
8113 return CombineBaseUpdate(N, DCI);
8114 case ISD::INTRINSIC_VOID:
8115 case ISD::INTRINSIC_W_CHAIN:
8116 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8117 case Intrinsic::arm_neon_vld1:
8118 case Intrinsic::arm_neon_vld2:
8119 case Intrinsic::arm_neon_vld3:
8120 case Intrinsic::arm_neon_vld4:
8121 case Intrinsic::arm_neon_vld2lane:
8122 case Intrinsic::arm_neon_vld3lane:
8123 case Intrinsic::arm_neon_vld4lane:
8124 case Intrinsic::arm_neon_vst1:
8125 case Intrinsic::arm_neon_vst2:
8126 case Intrinsic::arm_neon_vst3:
8127 case Intrinsic::arm_neon_vst4:
8128 case Intrinsic::arm_neon_vst2lane:
8129 case Intrinsic::arm_neon_vst3lane:
8130 case Intrinsic::arm_neon_vst4lane:
8131 return CombineBaseUpdate(N, DCI);
8139 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8141 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8144 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8145 if (!Subtarget->allowsUnalignedMem())
8148 switch (VT.getSimpleVT().SimpleTy) {
8155 // FIXME: VLD1 etc with standard alignment is legal.
8159 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8160 unsigned AlignCheck) {
8161 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8162 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8165 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8166 unsigned DstAlign, unsigned SrcAlign,
8169 MachineFunction &MF) const {
8170 const Function *F = MF.getFunction();
8172 // See if we can use NEON instructions for this...
8174 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8175 Subtarget->hasNEON()) {
8176 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8178 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8183 // Lowering to i32/i16 if the size permits.
8186 } else if (Size >= 2) {
8190 // Let the target-independent logic figure it out.
8194 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8199 switch (VT.getSimpleVT().SimpleTy) {
8200 default: return false;
8215 if ((V & (Scale - 1)) != 0)
8218 return V == (V & ((1LL << 5) - 1));
8221 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8222 const ARMSubtarget *Subtarget) {
8229 switch (VT.getSimpleVT().SimpleTy) {
8230 default: return false;
8235 // + imm12 or - imm8
8237 return V == (V & ((1LL << 8) - 1));
8238 return V == (V & ((1LL << 12) - 1));
8241 // Same as ARM mode. FIXME: NEON?
8242 if (!Subtarget->hasVFP2())
8247 return V == (V & ((1LL << 8) - 1));
8251 /// isLegalAddressImmediate - Return true if the integer value can be used
8252 /// as the offset of the target addressing mode for load / store of the
8254 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8255 const ARMSubtarget *Subtarget) {
8262 if (Subtarget->isThumb1Only())
8263 return isLegalT1AddressImmediate(V, VT);
8264 else if (Subtarget->isThumb2())
8265 return isLegalT2AddressImmediate(V, VT, Subtarget);
8270 switch (VT.getSimpleVT().SimpleTy) {
8271 default: return false;
8276 return V == (V & ((1LL << 12) - 1));
8279 return V == (V & ((1LL << 8) - 1));
8282 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8287 return V == (V & ((1LL << 8) - 1));
8291 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8293 int Scale = AM.Scale;
8297 switch (VT.getSimpleVT().SimpleTy) {
8298 default: return false;
8307 return Scale == 2 || Scale == 4 || Scale == 8;
8310 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8314 // Note, we allow "void" uses (basically, uses that aren't loads or
8315 // stores), because arm allows folding a scale into many arithmetic
8316 // operations. This should be made more precise and revisited later.
8318 // Allow r << imm, but the imm has to be a multiple of two.
8319 if (Scale & 1) return false;
8320 return isPowerOf2_32(Scale);
8324 /// isLegalAddressingMode - Return true if the addressing mode represented
8325 /// by AM is legal for this target, for a load/store of the specified type.
8326 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8328 EVT VT = getValueType(Ty, true);
8329 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8332 // Can never fold addr of global into load/store.
8337 case 0: // no scale reg, must be "r+i" or "r", or "i".
8340 if (Subtarget->isThumb1Only())
8344 // ARM doesn't support any R+R*scale+imm addr modes.
8351 if (Subtarget->isThumb2())
8352 return isLegalT2ScaledAddressingMode(AM, VT);
8354 int Scale = AM.Scale;
8355 switch (VT.getSimpleVT().SimpleTy) {
8356 default: return false;
8360 if (Scale < 0) Scale = -Scale;
8364 return isPowerOf2_32(Scale & ~1);
8368 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8373 // Note, we allow "void" uses (basically, uses that aren't loads or
8374 // stores), because arm allows folding a scale into many arithmetic
8375 // operations. This should be made more precise and revisited later.
8377 // Allow r << imm, but the imm has to be a multiple of two.
8378 if (Scale & 1) return false;
8379 return isPowerOf2_32(Scale);
8385 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8386 /// icmp immediate, that is the target has icmp instructions which can compare
8387 /// a register against the immediate without having to materialize the
8388 /// immediate into a register.
8389 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8390 if (!Subtarget->isThumb())
8391 return ARM_AM::getSOImmVal(Imm) != -1;
8392 if (Subtarget->isThumb2())
8393 return ARM_AM::getT2SOImmVal(Imm) != -1;
8394 return Imm >= 0 && Imm <= 255;
8397 /// isLegalAddImmediate - Return true if the specified immediate is legal
8398 /// add immediate, that is the target has add instructions which can add
8399 /// a register with the immediate without having to materialize the
8400 /// immediate into a register.
8401 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8402 return ARM_AM::getSOImmVal(Imm) != -1;
8405 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8406 bool isSEXTLoad, SDValue &Base,
8407 SDValue &Offset, bool &isInc,
8408 SelectionDAG &DAG) {
8409 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8412 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8414 Base = Ptr->getOperand(0);
8415 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8416 int RHSC = (int)RHS->getZExtValue();
8417 if (RHSC < 0 && RHSC > -256) {
8418 assert(Ptr->getOpcode() == ISD::ADD);
8420 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8424 isInc = (Ptr->getOpcode() == ISD::ADD);
8425 Offset = Ptr->getOperand(1);
8427 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8429 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8430 int RHSC = (int)RHS->getZExtValue();
8431 if (RHSC < 0 && RHSC > -0x1000) {
8432 assert(Ptr->getOpcode() == ISD::ADD);
8434 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8435 Base = Ptr->getOperand(0);
8440 if (Ptr->getOpcode() == ISD::ADD) {
8442 ARM_AM::ShiftOpc ShOpcVal=
8443 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8444 if (ShOpcVal != ARM_AM::no_shift) {
8445 Base = Ptr->getOperand(1);
8446 Offset = Ptr->getOperand(0);
8448 Base = Ptr->getOperand(0);
8449 Offset = Ptr->getOperand(1);
8454 isInc = (Ptr->getOpcode() == ISD::ADD);
8455 Base = Ptr->getOperand(0);
8456 Offset = Ptr->getOperand(1);
8460 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8464 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8465 bool isSEXTLoad, SDValue &Base,
8466 SDValue &Offset, bool &isInc,
8467 SelectionDAG &DAG) {
8468 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8471 Base = Ptr->getOperand(0);
8472 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8473 int RHSC = (int)RHS->getZExtValue();
8474 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8475 assert(Ptr->getOpcode() == ISD::ADD);
8477 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8479 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8480 isInc = Ptr->getOpcode() == ISD::ADD;
8481 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8489 /// getPreIndexedAddressParts - returns true by value, base pointer and
8490 /// offset pointer and addressing mode by reference if the node's address
8491 /// can be legally represented as pre-indexed load / store address.
8493 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8495 ISD::MemIndexedMode &AM,
8496 SelectionDAG &DAG) const {
8497 if (Subtarget->isThumb1Only())
8502 bool isSEXTLoad = false;
8503 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8504 Ptr = LD->getBasePtr();
8505 VT = LD->getMemoryVT();
8506 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8507 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8508 Ptr = ST->getBasePtr();
8509 VT = ST->getMemoryVT();
8514 bool isLegal = false;
8515 if (Subtarget->isThumb2())
8516 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8517 Offset, isInc, DAG);
8519 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8520 Offset, isInc, DAG);
8524 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8528 /// getPostIndexedAddressParts - returns true by value, base pointer and
8529 /// offset pointer and addressing mode by reference if this node can be
8530 /// combined with a load / store to form a post-indexed load / store.
8531 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8534 ISD::MemIndexedMode &AM,
8535 SelectionDAG &DAG) const {
8536 if (Subtarget->isThumb1Only())
8541 bool isSEXTLoad = false;
8542 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8543 VT = LD->getMemoryVT();
8544 Ptr = LD->getBasePtr();
8545 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8546 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8547 VT = ST->getMemoryVT();
8548 Ptr = ST->getBasePtr();
8553 bool isLegal = false;
8554 if (Subtarget->isThumb2())
8555 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8558 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8564 // Swap base ptr and offset to catch more post-index load / store when
8565 // it's legal. In Thumb2 mode, offset must be an immediate.
8566 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8567 !Subtarget->isThumb2())
8568 std::swap(Base, Offset);
8570 // Post-indexed load / store update the base pointer.
8575 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8579 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8583 const SelectionDAG &DAG,
8584 unsigned Depth) const {
8585 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8586 switch (Op.getOpcode()) {
8588 case ARMISD::CMOV: {
8589 // Bits are known zero/one if known on the LHS and RHS.
8590 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8591 if (KnownZero == 0 && KnownOne == 0) return;
8593 APInt KnownZeroRHS, KnownOneRHS;
8594 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8595 KnownZeroRHS, KnownOneRHS, Depth+1);
8596 KnownZero &= KnownZeroRHS;
8597 KnownOne &= KnownOneRHS;
8603 //===----------------------------------------------------------------------===//
8604 // ARM Inline Assembly Support
8605 //===----------------------------------------------------------------------===//
8607 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8608 // Looking for "rev" which is V6+.
8609 if (!Subtarget->hasV6Ops())
8612 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8613 std::string AsmStr = IA->getAsmString();
8614 SmallVector<StringRef, 4> AsmPieces;
8615 SplitString(AsmStr, AsmPieces, ";\n");
8617 switch (AsmPieces.size()) {
8618 default: return false;
8620 AsmStr = AsmPieces[0];
8622 SplitString(AsmStr, AsmPieces, " \t,");
8625 if (AsmPieces.size() == 3 &&
8626 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8627 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8628 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8629 if (Ty && Ty->getBitWidth() == 32)
8630 return IntrinsicLowering::LowerToByteSwap(CI);
8638 /// getConstraintType - Given a constraint letter, return the type of
8639 /// constraint it is for this target.
8640 ARMTargetLowering::ConstraintType
8641 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8642 if (Constraint.size() == 1) {
8643 switch (Constraint[0]) {
8645 case 'l': return C_RegisterClass;
8646 case 'w': return C_RegisterClass;
8647 case 'h': return C_RegisterClass;
8648 case 'x': return C_RegisterClass;
8649 case 't': return C_RegisterClass;
8650 case 'j': return C_Other; // Constant for movw.
8651 // An address with a single base register. Due to the way we
8652 // currently handle addresses it is the same as an 'r' memory constraint.
8653 case 'Q': return C_Memory;
8655 } else if (Constraint.size() == 2) {
8656 switch (Constraint[0]) {
8658 // All 'U+' constraints are addresses.
8659 case 'U': return C_Memory;
8662 return TargetLowering::getConstraintType(Constraint);
8665 /// Examine constraint type and operand type and determine a weight value.
8666 /// This object must already have been set up with the operand type
8667 /// and the current alternative constraint selected.
8668 TargetLowering::ConstraintWeight
8669 ARMTargetLowering::getSingleConstraintMatchWeight(
8670 AsmOperandInfo &info, const char *constraint) const {
8671 ConstraintWeight weight = CW_Invalid;
8672 Value *CallOperandVal = info.CallOperandVal;
8673 // If we don't have a value, we can't do a match,
8674 // but allow it at the lowest weight.
8675 if (CallOperandVal == NULL)
8677 Type *type = CallOperandVal->getType();
8678 // Look at the constraint type.
8679 switch (*constraint) {
8681 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8684 if (type->isIntegerTy()) {
8685 if (Subtarget->isThumb())
8686 weight = CW_SpecificReg;
8688 weight = CW_Register;
8692 if (type->isFloatingPointTy())
8693 weight = CW_Register;
8699 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8701 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8703 if (Constraint.size() == 1) {
8704 // GCC ARM Constraint Letters
8705 switch (Constraint[0]) {
8706 case 'l': // Low regs or general regs.
8707 if (Subtarget->isThumb())
8708 return RCPair(0U, ARM::tGPRRegisterClass);
8710 return RCPair(0U, ARM::GPRRegisterClass);
8711 case 'h': // High regs or no regs.
8712 if (Subtarget->isThumb())
8713 return RCPair(0U, ARM::hGPRRegisterClass);
8716 return RCPair(0U, ARM::GPRRegisterClass);
8719 return RCPair(0U, ARM::SPRRegisterClass);
8720 if (VT.getSizeInBits() == 64)
8721 return RCPair(0U, ARM::DPRRegisterClass);
8722 if (VT.getSizeInBits() == 128)
8723 return RCPair(0U, ARM::QPRRegisterClass);
8727 return RCPair(0U, ARM::SPR_8RegisterClass);
8728 if (VT.getSizeInBits() == 64)
8729 return RCPair(0U, ARM::DPR_8RegisterClass);
8730 if (VT.getSizeInBits() == 128)
8731 return RCPair(0U, ARM::QPR_8RegisterClass);
8735 return RCPair(0U, ARM::SPRRegisterClass);
8739 if (StringRef("{cc}").equals_lower(Constraint))
8740 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8742 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8745 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8746 /// vector. If it is invalid, don't add anything to Ops.
8747 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8748 std::string &Constraint,
8749 std::vector<SDValue>&Ops,
8750 SelectionDAG &DAG) const {
8751 SDValue Result(0, 0);
8753 // Currently only support length 1 constraints.
8754 if (Constraint.length() != 1) return;
8756 char ConstraintLetter = Constraint[0];
8757 switch (ConstraintLetter) {
8760 case 'I': case 'J': case 'K': case 'L':
8761 case 'M': case 'N': case 'O':
8762 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8766 int64_t CVal64 = C->getSExtValue();
8767 int CVal = (int) CVal64;
8768 // None of these constraints allow values larger than 32 bits. Check
8769 // that the value fits in an int.
8773 switch (ConstraintLetter) {
8775 // Constant suitable for movw, must be between 0 and
8777 if (Subtarget->hasV6T2Ops())
8778 if (CVal >= 0 && CVal <= 65535)
8782 if (Subtarget->isThumb1Only()) {
8783 // This must be a constant between 0 and 255, for ADD
8785 if (CVal >= 0 && CVal <= 255)
8787 } else if (Subtarget->isThumb2()) {
8788 // A constant that can be used as an immediate value in a
8789 // data-processing instruction.
8790 if (ARM_AM::getT2SOImmVal(CVal) != -1)
8793 // A constant that can be used as an immediate value in a
8794 // data-processing instruction.
8795 if (ARM_AM::getSOImmVal(CVal) != -1)
8801 if (Subtarget->isThumb()) { // FIXME thumb2
8802 // This must be a constant between -255 and -1, for negated ADD
8803 // immediates. This can be used in GCC with an "n" modifier that
8804 // prints the negated value, for use with SUB instructions. It is
8805 // not useful otherwise but is implemented for compatibility.
8806 if (CVal >= -255 && CVal <= -1)
8809 // This must be a constant between -4095 and 4095. It is not clear
8810 // what this constraint is intended for. Implemented for
8811 // compatibility with GCC.
8812 if (CVal >= -4095 && CVal <= 4095)
8818 if (Subtarget->isThumb1Only()) {
8819 // A 32-bit value where only one byte has a nonzero value. Exclude
8820 // zero to match GCC. This constraint is used by GCC internally for
8821 // constants that can be loaded with a move/shift combination.
8822 // It is not useful otherwise but is implemented for compatibility.
8823 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8825 } else if (Subtarget->isThumb2()) {
8826 // A constant whose bitwise inverse can be used as an immediate
8827 // value in a data-processing instruction. This can be used in GCC
8828 // with a "B" modifier that prints the inverted value, for use with
8829 // BIC and MVN instructions. It is not useful otherwise but is
8830 // implemented for compatibility.
8831 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8834 // A constant whose bitwise inverse can be used as an immediate
8835 // value in a data-processing instruction. This can be used in GCC
8836 // with a "B" modifier that prints the inverted value, for use with
8837 // BIC and MVN instructions. It is not useful otherwise but is
8838 // implemented for compatibility.
8839 if (ARM_AM::getSOImmVal(~CVal) != -1)
8845 if (Subtarget->isThumb1Only()) {
8846 // This must be a constant between -7 and 7,
8847 // for 3-operand ADD/SUB immediate instructions.
8848 if (CVal >= -7 && CVal < 7)
8850 } else if (Subtarget->isThumb2()) {
8851 // A constant whose negation can be used as an immediate value in a
8852 // data-processing instruction. This can be used in GCC with an "n"
8853 // modifier that prints the negated value, for use with SUB
8854 // instructions. It is not useful otherwise but is implemented for
8856 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8859 // A constant whose negation can be used as an immediate value in a
8860 // data-processing instruction. This can be used in GCC with an "n"
8861 // modifier that prints the negated value, for use with SUB
8862 // instructions. It is not useful otherwise but is implemented for
8864 if (ARM_AM::getSOImmVal(-CVal) != -1)
8870 if (Subtarget->isThumb()) { // FIXME thumb2
8871 // This must be a multiple of 4 between 0 and 1020, for
8872 // ADD sp + immediate.
8873 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
8876 // A power of two or a constant between 0 and 32. This is used in
8877 // GCC for the shift amount on shifted register operands, but it is
8878 // useful in general for any shift amounts.
8879 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8885 if (Subtarget->isThumb()) { // FIXME thumb2
8886 // This must be a constant between 0 and 31, for shift amounts.
8887 if (CVal >= 0 && CVal <= 31)
8893 if (Subtarget->isThumb()) { // FIXME thumb2
8894 // This must be a multiple of 4 between -508 and 508, for
8895 // ADD/SUB sp = sp + immediate.
8896 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8901 Result = DAG.getTargetConstant(CVal, Op.getValueType());
8905 if (Result.getNode()) {
8906 Ops.push_back(Result);
8909 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8913 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8914 // The ARM target isn't yet aware of offsets.
8918 bool ARM::isBitFieldInvertedMask(unsigned v) {
8919 if (v == 0xffffffff)
8921 // there can be 1's on either or both "outsides", all the "inside"
8923 unsigned int lsb = 0, msb = 31;
8924 while (v & (1 << msb)) --msb;
8925 while (v & (1 << lsb)) ++lsb;
8926 for (unsigned int i = lsb; i <= msb; ++i) {
8933 /// isFPImmLegal - Returns true if the target can instruction select the
8934 /// specified FP immediate natively. If false, the legalizer will
8935 /// materialize the FP immediate as a load from a constant pool.
8936 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8937 if (!Subtarget->hasVFP3())
8940 return ARM_AM::getFP32Imm(Imm) != -1;
8942 return ARM_AM::getFP64Imm(Imm) != -1;
8946 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8947 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
8948 /// specified in the intrinsic calls.
8949 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8951 unsigned Intrinsic) const {
8952 switch (Intrinsic) {
8953 case Intrinsic::arm_neon_vld1:
8954 case Intrinsic::arm_neon_vld2:
8955 case Intrinsic::arm_neon_vld3:
8956 case Intrinsic::arm_neon_vld4:
8957 case Intrinsic::arm_neon_vld2lane:
8958 case Intrinsic::arm_neon_vld3lane:
8959 case Intrinsic::arm_neon_vld4lane: {
8960 Info.opc = ISD::INTRINSIC_W_CHAIN;
8961 // Conservatively set memVT to the entire set of vectors loaded.
8962 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8963 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8964 Info.ptrVal = I.getArgOperand(0);
8966 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8967 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8968 Info.vol = false; // volatile loads with NEON intrinsics not supported
8969 Info.readMem = true;
8970 Info.writeMem = false;
8973 case Intrinsic::arm_neon_vst1:
8974 case Intrinsic::arm_neon_vst2:
8975 case Intrinsic::arm_neon_vst3:
8976 case Intrinsic::arm_neon_vst4:
8977 case Intrinsic::arm_neon_vst2lane:
8978 case Intrinsic::arm_neon_vst3lane:
8979 case Intrinsic::arm_neon_vst4lane: {
8980 Info.opc = ISD::INTRINSIC_VOID;
8981 // Conservatively set memVT to the entire set of vectors stored.
8982 unsigned NumElts = 0;
8983 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8984 Type *ArgTy = I.getArgOperand(ArgI)->getType();
8985 if (!ArgTy->isVectorTy())
8987 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8989 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8990 Info.ptrVal = I.getArgOperand(0);
8992 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8993 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8994 Info.vol = false; // volatile stores with NEON intrinsics not supported
8995 Info.readMem = false;
8996 Info.writeMem = true;
8999 case Intrinsic::arm_strexd: {
9000 Info.opc = ISD::INTRINSIC_W_CHAIN;
9001 Info.memVT = MVT::i64;
9002 Info.ptrVal = I.getArgOperand(2);
9006 Info.readMem = false;
9007 Info.writeMem = true;
9010 case Intrinsic::arm_ldrexd: {
9011 Info.opc = ISD::INTRINSIC_W_CHAIN;
9012 Info.memVT = MVT::i64;
9013 Info.ptrVal = I.getArgOperand(0);
9017 Info.readMem = true;
9018 Info.writeMem = false;