1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "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"
54 STATISTIC(NumTailCalls, "Number of tail calls");
55 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
94 EVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
97 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
98 PromotedLdStVT.getSimpleVT());
100 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
101 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
102 PromotedLdStVT.getSimpleVT());
105 EVT ElemTy = VT.getVectorElementType();
106 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
107 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
109 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
110 if (ElemTy == MVT::i32) {
111 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
113 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
114 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
121 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
122 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
123 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
124 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
125 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
128 if (VT.isInteger()) {
129 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
130 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
131 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
134 // Promote all bit-wise operations.
135 if (VT.isInteger() && VT != PromotedBitwiseVT) {
136 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
137 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
138 PromotedBitwiseVT.getSimpleVT());
139 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
140 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
141 PromotedBitwiseVT.getSimpleVT());
142 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
143 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
144 PromotedBitwiseVT.getSimpleVT());
147 // Neon does not support vector divide/remainder operations.
148 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
149 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
150 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
156 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
157 addRegisterClass(VT, ARM::DPRRegisterClass);
158 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
161 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
162 addRegisterClass(VT, ARM::QPRRegisterClass);
163 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
166 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
167 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
168 return new TargetLoweringObjectFileMachO();
170 return new ARMElfTargetObjectFile();
173 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
174 : TargetLowering(TM, createTLOF(TM)) {
175 Subtarget = &TM.getSubtarget<ARMSubtarget>();
176 RegInfo = TM.getRegisterInfo();
177 Itins = TM.getInstrItineraryData();
179 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
181 if (Subtarget->isTargetDarwin()) {
182 // Uses VFP for Thumb libfuncs if available.
183 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
184 // Single-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
186 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
187 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
188 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
190 // Double-precision floating-point arithmetic.
191 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
192 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
193 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
194 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
196 // Single-precision comparisons.
197 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
198 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
199 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
200 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
201 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
202 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
203 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
204 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
206 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
215 // Double-precision comparisons.
216 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
217 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
218 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
219 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
220 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
221 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
222 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
223 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
225 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
234 // Floating-point to integer conversions.
235 // i64 conversions are done via library routines even when generating VFP
236 // instructions, so use the same ones.
237 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
238 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
239 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
240 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
242 // Conversions between floating types.
243 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
244 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
246 // Integer to floating-point conversions.
247 // i64 conversions are done via library routines even when generating VFP
248 // instructions, so use the same ones.
249 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
250 // e.g., __floatunsidf vs. __floatunssidfvfp.
251 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
252 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
253 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
254 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
258 // These libcalls are not available in 32-bit.
259 setLibcallName(RTLIB::SHL_I128, 0);
260 setLibcallName(RTLIB::SRL_I128, 0);
261 setLibcallName(RTLIB::SRA_I128, 0);
263 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
264 // Double-precision floating-point arithmetic helper functions
265 // RTABI chapter 4.1.2, Table 2
266 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
267 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
268 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
269 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
270 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
275 // Double-precision floating-point comparison helper functions
276 // RTABI chapter 4.1.2, Table 3
277 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
278 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
279 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
280 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
281 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
282 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
284 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
286 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
287 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
288 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
289 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
290 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
291 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
292 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
293 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
302 // Single-precision floating-point arithmetic helper functions
303 // RTABI chapter 4.1.2, Table 4
304 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
305 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
306 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
307 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
308 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
313 // Single-precision floating-point comparison helper functions
314 // RTABI chapter 4.1.2, Table 5
315 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
316 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
317 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
318 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
319 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
320 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
322 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
324 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
325 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
326 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
327 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
328 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
329 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
330 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
331 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
340 // Floating-point to integer conversions.
341 // RTABI chapter 4.1.2, Table 6
342 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
343 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
345 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
346 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
347 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
348 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
349 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
350 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
359 // Conversions between floating types.
360 // RTABI chapter 4.1.2, Table 7
361 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
362 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
363 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
366 // Integer to floating-point conversions.
367 // RTABI chapter 4.1.2, Table 8
368 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
369 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
370 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
371 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
372 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
373 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
374 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
375 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385 // Long long helper functions
386 // RTABI chapter 4.2, Table 9
387 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
388 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
389 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
390 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
391 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
398 // Integer division functions
399 // RTABI chapter 4.3.1
400 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
401 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
402 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
403 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
404 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
405 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
407 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
408 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
415 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
418 // RTABI chapter 4.3.4
419 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
420 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
421 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
422 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
423 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
424 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
427 // Use divmod compiler-rt calls for iOS 5.0 and later.
428 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
429 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
430 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
431 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
434 if (Subtarget->isThumb1Only())
435 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
437 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
438 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
439 !Subtarget->isThumb1Only()) {
440 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
441 if (!Subtarget->isFPOnlySP())
442 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
444 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
447 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
448 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
449 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
450 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
451 setTruncStoreAction((MVT::SimpleValueType)VT,
452 (MVT::SimpleValueType)InnerVT, Expand);
453 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
455 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
458 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
460 if (Subtarget->hasNEON()) {
461 addDRTypeForNEON(MVT::v2f32);
462 addDRTypeForNEON(MVT::v8i8);
463 addDRTypeForNEON(MVT::v4i16);
464 addDRTypeForNEON(MVT::v2i32);
465 addDRTypeForNEON(MVT::v1i64);
467 addQRTypeForNEON(MVT::v4f32);
468 addQRTypeForNEON(MVT::v2f64);
469 addQRTypeForNEON(MVT::v16i8);
470 addQRTypeForNEON(MVT::v8i16);
471 addQRTypeForNEON(MVT::v4i32);
472 addQRTypeForNEON(MVT::v2i64);
474 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
475 // neither Neon nor VFP support any arithmetic operations on it.
476 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
477 // supported for v4f32.
478 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
479 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
480 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
481 // FIXME: Code duplication: FDIV and FREM are expanded always, see
482 // ARMTargetLowering::addTypeForNEON method for details.
483 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
484 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
485 // FIXME: Create unittest.
486 // In another words, find a way when "copysign" appears in DAG with vector
488 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
489 // FIXME: Code duplication: SETCC has custom operation action, see
490 // ARMTargetLowering::addTypeForNEON method for details.
491 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
492 // FIXME: Create unittest for FNEG and for FABS.
493 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
494 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
495 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
496 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
497 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
498 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
499 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
500 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
501 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
502 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
503 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
504 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
505 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
506 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
508 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
509 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
510 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
512 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
513 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
514 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
515 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
516 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
518 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
520 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
521 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
523 // Neon does not support some operations on v1i64 and v2i64 types.
524 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
525 // Custom handling for some quad-vector types to detect VMULL.
526 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
527 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
528 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
529 // Custom handling for some vector types to avoid expensive expansions
530 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
531 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
532 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
533 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
534 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
535 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
536 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
537 // a destination type that is wider than the source, and nor does
538 // it have a FP_TO_[SU]INT instruction with a narrower destination than
540 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
541 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
542 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
543 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
545 setTargetDAGCombine(ISD::INTRINSIC_VOID);
546 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
547 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
548 setTargetDAGCombine(ISD::SHL);
549 setTargetDAGCombine(ISD::SRL);
550 setTargetDAGCombine(ISD::SRA);
551 setTargetDAGCombine(ISD::SIGN_EXTEND);
552 setTargetDAGCombine(ISD::ZERO_EXTEND);
553 setTargetDAGCombine(ISD::ANY_EXTEND);
554 setTargetDAGCombine(ISD::SELECT_CC);
555 setTargetDAGCombine(ISD::BUILD_VECTOR);
556 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
557 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
558 setTargetDAGCombine(ISD::STORE);
559 setTargetDAGCombine(ISD::FP_TO_SINT);
560 setTargetDAGCombine(ISD::FP_TO_UINT);
561 setTargetDAGCombine(ISD::FDIV);
563 // It is legal to extload from v4i8 to v4i16 or v4i32.
564 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
565 MVT::v4i16, MVT::v2i16,
567 for (unsigned i = 0; i < 6; ++i) {
568 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
569 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
570 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
574 computeRegisterProperties();
576 // ARM does not have f32 extending load.
577 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
579 // ARM does not have i1 sign extending load.
580 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
582 // ARM supports all 4 flavors of integer indexed load / store.
583 if (!Subtarget->isThumb1Only()) {
584 for (unsigned im = (unsigned)ISD::PRE_INC;
585 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
586 setIndexedLoadAction(im, MVT::i1, Legal);
587 setIndexedLoadAction(im, MVT::i8, Legal);
588 setIndexedLoadAction(im, MVT::i16, Legal);
589 setIndexedLoadAction(im, MVT::i32, Legal);
590 setIndexedStoreAction(im, MVT::i1, Legal);
591 setIndexedStoreAction(im, MVT::i8, Legal);
592 setIndexedStoreAction(im, MVT::i16, Legal);
593 setIndexedStoreAction(im, MVT::i32, Legal);
597 // i64 operation support.
598 setOperationAction(ISD::MUL, MVT::i64, Expand);
599 setOperationAction(ISD::MULHU, MVT::i32, Expand);
600 if (Subtarget->isThumb1Only()) {
601 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
602 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
604 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
605 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
606 setOperationAction(ISD::MULHS, MVT::i32, Expand);
608 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
609 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
610 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRL, MVT::i64, Custom);
612 setOperationAction(ISD::SRA, MVT::i64, Custom);
614 if (!Subtarget->isThumb1Only()) {
615 // FIXME: We should do this for Thumb1 as well.
616 setOperationAction(ISD::ADDC, MVT::i32, Custom);
617 setOperationAction(ISD::ADDE, MVT::i32, Custom);
618 setOperationAction(ISD::SUBC, MVT::i32, Custom);
619 setOperationAction(ISD::SUBE, MVT::i32, Custom);
622 // ARM does not have ROTL.
623 setOperationAction(ISD::ROTL, MVT::i32, Expand);
624 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
625 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
626 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
627 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
629 // These just redirect to CTTZ and CTLZ on ARM.
630 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
631 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
633 // Only ARMv6 has BSWAP.
634 if (!Subtarget->hasV6Ops())
635 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
637 // These are expanded into libcalls.
638 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
639 // v7M has a hardware divider
640 setOperationAction(ISD::SDIV, MVT::i32, Expand);
641 setOperationAction(ISD::UDIV, MVT::i32, Expand);
643 setOperationAction(ISD::SREM, MVT::i32, Expand);
644 setOperationAction(ISD::UREM, MVT::i32, Expand);
645 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
646 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
648 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
649 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
650 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
651 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
652 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
654 setOperationAction(ISD::TRAP, MVT::Other, Legal);
656 // Use the default implementation.
657 setOperationAction(ISD::VASTART, MVT::Other, Custom);
658 setOperationAction(ISD::VAARG, MVT::Other, Expand);
659 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
660 setOperationAction(ISD::VAEND, MVT::Other, Expand);
661 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
662 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
664 if (!Subtarget->isTargetDarwin()) {
665 // Non-Darwin platforms may return values in these registers via the
666 // personality function.
667 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
668 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
669 setExceptionPointerRegister(ARM::R0);
670 setExceptionSelectorRegister(ARM::R1);
673 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
674 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
675 // the default expansion.
676 // FIXME: This should be checking for v6k, not just v6.
677 if (Subtarget->hasDataBarrier() ||
678 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
679 // membarrier needs custom lowering; the rest are legal and handled
681 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
682 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
683 // Custom lowering for 64-bit ops
684 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
685 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
686 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
691 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
692 setInsertFencesForAtomic(true);
694 // Set them all for expansion, which will force libcalls.
695 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
696 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
697 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
698 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
699 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
709 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
710 // Unordered/Monotonic case.
711 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
712 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
713 // Since the libcalls include locking, fold in the fences
714 setShouldFoldAtomicFences(true);
717 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
719 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
720 if (!Subtarget->hasV6Ops()) {
721 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
722 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
726 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
727 !Subtarget->isThumb1Only()) {
728 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
729 // iff target supports vfp2.
730 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
731 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
734 // We want to custom lower some of our intrinsics.
735 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
736 if (Subtarget->isTargetDarwin()) {
737 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
738 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
739 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
742 setOperationAction(ISD::SETCC, MVT::i32, Expand);
743 setOperationAction(ISD::SETCC, MVT::f32, Expand);
744 setOperationAction(ISD::SETCC, MVT::f64, Expand);
745 setOperationAction(ISD::SELECT, MVT::i32, Custom);
746 setOperationAction(ISD::SELECT, MVT::f32, Custom);
747 setOperationAction(ISD::SELECT, MVT::f64, Custom);
748 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
749 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
752 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
753 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
754 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
755 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
756 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
758 // We don't support sin/cos/fmod/copysign/pow
759 setOperationAction(ISD::FSIN, MVT::f64, Expand);
760 setOperationAction(ISD::FSIN, MVT::f32, Expand);
761 setOperationAction(ISD::FCOS, MVT::f32, Expand);
762 setOperationAction(ISD::FCOS, MVT::f64, Expand);
763 setOperationAction(ISD::FREM, MVT::f64, Expand);
764 setOperationAction(ISD::FREM, MVT::f32, Expand);
765 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
766 !Subtarget->isThumb1Only()) {
767 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
768 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
770 setOperationAction(ISD::FPOW, MVT::f64, Expand);
771 setOperationAction(ISD::FPOW, MVT::f32, Expand);
773 setOperationAction(ISD::FMA, MVT::f64, Expand);
774 setOperationAction(ISD::FMA, MVT::f32, Expand);
776 // Various VFP goodness
777 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
778 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
779 if (Subtarget->hasVFP2()) {
780 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
781 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
782 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
783 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
785 // Special handling for half-precision FP.
786 if (!Subtarget->hasFP16()) {
787 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
788 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
792 // We have target-specific dag combine patterns for the following nodes:
793 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
794 setTargetDAGCombine(ISD::ADD);
795 setTargetDAGCombine(ISD::SUB);
796 setTargetDAGCombine(ISD::MUL);
798 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
799 setTargetDAGCombine(ISD::AND);
800 setTargetDAGCombine(ISD::OR);
801 setTargetDAGCombine(ISD::XOR);
804 if (Subtarget->hasV6Ops())
805 setTargetDAGCombine(ISD::SRL);
807 setStackPointerRegisterToSaveRestore(ARM::SP);
809 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
810 !Subtarget->hasVFP2())
811 setSchedulingPreference(Sched::RegPressure);
813 setSchedulingPreference(Sched::Hybrid);
815 //// temporary - rewrite interface to use type
816 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
817 maxStoresPerMemset = 16;
818 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
820 // On ARM arguments smaller than 4 bytes are extended, so all arguments
821 // are at least 4 bytes aligned.
822 setMinStackArgumentAlignment(4);
824 benefitFromCodePlacementOpt = true;
826 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
829 // FIXME: It might make sense to define the representative register class as the
830 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
831 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
832 // SPR's representative would be DPR_VFP2. This should work well if register
833 // pressure tracking were modified such that a register use would increment the
834 // pressure of the register class's representative and all of it's super
835 // classes' representatives transitively. We have not implemented this because
836 // of the difficulty prior to coalescing of modeling operand register classes
837 // due to the common occurrence of cross class copies and subregister insertions
839 std::pair<const TargetRegisterClass*, uint8_t>
840 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
841 const TargetRegisterClass *RRC = 0;
843 switch (VT.getSimpleVT().SimpleTy) {
845 return TargetLowering::findRepresentativeClass(VT);
846 // Use DPR as representative register class for all floating point
847 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
848 // the cost is 1 for both f32 and f64.
849 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
850 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
851 RRC = ARM::DPRRegisterClass;
852 // When NEON is used for SP, only half of the register file is available
853 // because operations that define both SP and DP results will be constrained
854 // to the VFP2 class (D0-D15). We currently model this constraint prior to
855 // coalescing by double-counting the SP regs. See the FIXME above.
856 if (Subtarget->useNEONForSinglePrecisionFP())
859 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
860 case MVT::v4f32: case MVT::v2f64:
861 RRC = ARM::DPRRegisterClass;
865 RRC = ARM::DPRRegisterClass;
869 RRC = ARM::DPRRegisterClass;
873 return std::make_pair(RRC, Cost);
876 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
879 case ARMISD::Wrapper: return "ARMISD::Wrapper";
880 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
881 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
882 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
883 case ARMISD::CALL: return "ARMISD::CALL";
884 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
885 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
886 case ARMISD::tCALL: return "ARMISD::tCALL";
887 case ARMISD::BRCOND: return "ARMISD::BRCOND";
888 case ARMISD::BR_JT: return "ARMISD::BR_JT";
889 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
890 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
891 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
892 case ARMISD::CMP: return "ARMISD::CMP";
893 case ARMISD::CMPZ: return "ARMISD::CMPZ";
894 case ARMISD::CMPFP: return "ARMISD::CMPFP";
895 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
896 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
897 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
899 case ARMISD::CMOV: return "ARMISD::CMOV";
900 case ARMISD::CAND: return "ARMISD::CAND";
901 case ARMISD::COR: return "ARMISD::COR";
902 case ARMISD::CXOR: return "ARMISD::CXOR";
904 case ARMISD::RBIT: return "ARMISD::RBIT";
906 case ARMISD::FTOSI: return "ARMISD::FTOSI";
907 case ARMISD::FTOUI: return "ARMISD::FTOUI";
908 case ARMISD::SITOF: return "ARMISD::SITOF";
909 case ARMISD::UITOF: return "ARMISD::UITOF";
911 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
912 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
913 case ARMISD::RRX: return "ARMISD::RRX";
915 case ARMISD::ADDC: return "ARMISD::ADDC";
916 case ARMISD::ADDE: return "ARMISD::ADDE";
917 case ARMISD::SUBC: return "ARMISD::SUBC";
918 case ARMISD::SUBE: return "ARMISD::SUBE";
920 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
921 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
923 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
924 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
926 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
928 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
930 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
932 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
933 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
935 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
937 case ARMISD::VCEQ: return "ARMISD::VCEQ";
938 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
939 case ARMISD::VCGE: return "ARMISD::VCGE";
940 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
941 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
942 case ARMISD::VCGEU: return "ARMISD::VCGEU";
943 case ARMISD::VCGT: return "ARMISD::VCGT";
944 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
945 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
946 case ARMISD::VCGTU: return "ARMISD::VCGTU";
947 case ARMISD::VTST: return "ARMISD::VTST";
949 case ARMISD::VSHL: return "ARMISD::VSHL";
950 case ARMISD::VSHRs: return "ARMISD::VSHRs";
951 case ARMISD::VSHRu: return "ARMISD::VSHRu";
952 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
953 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
954 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
955 case ARMISD::VSHRN: return "ARMISD::VSHRN";
956 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
957 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
958 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
959 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
960 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
961 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
962 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
963 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
964 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
965 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
966 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
967 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
968 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
969 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
970 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
971 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
972 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
973 case ARMISD::VDUP: return "ARMISD::VDUP";
974 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
975 case ARMISD::VEXT: return "ARMISD::VEXT";
976 case ARMISD::VREV64: return "ARMISD::VREV64";
977 case ARMISD::VREV32: return "ARMISD::VREV32";
978 case ARMISD::VREV16: return "ARMISD::VREV16";
979 case ARMISD::VZIP: return "ARMISD::VZIP";
980 case ARMISD::VUZP: return "ARMISD::VUZP";
981 case ARMISD::VTRN: return "ARMISD::VTRN";
982 case ARMISD::VTBL1: return "ARMISD::VTBL1";
983 case ARMISD::VTBL2: return "ARMISD::VTBL2";
984 case ARMISD::VMULLs: return "ARMISD::VMULLs";
985 case ARMISD::VMULLu: return "ARMISD::VMULLu";
986 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
987 case ARMISD::FMAX: return "ARMISD::FMAX";
988 case ARMISD::FMIN: return "ARMISD::FMIN";
989 case ARMISD::BFI: return "ARMISD::BFI";
990 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
991 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
992 case ARMISD::VBSL: return "ARMISD::VBSL";
993 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
994 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
995 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
996 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
997 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
998 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
999 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1000 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1001 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1002 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1003 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1004 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1005 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1006 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1007 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1008 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1009 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1010 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1011 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1012 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1016 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1017 if (!VT.isVector()) return getPointerTy();
1018 return VT.changeVectorElementTypeToInteger();
1021 /// getRegClassFor - Return the register class that should be used for the
1022 /// specified value type.
1023 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1024 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1025 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1026 // load / store 4 to 8 consecutive D registers.
1027 if (Subtarget->hasNEON()) {
1028 if (VT == MVT::v4i64)
1029 return ARM::QQPRRegisterClass;
1030 else if (VT == MVT::v8i64)
1031 return ARM::QQQQPRRegisterClass;
1033 return TargetLowering::getRegClassFor(VT);
1036 // Create a fast isel object.
1038 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1039 return ARM::createFastISel(funcInfo);
1042 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1043 /// be used for loads / stores from the global.
1044 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1045 return (Subtarget->isThumb1Only() ? 127 : 4095);
1048 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1049 unsigned NumVals = N->getNumValues();
1051 return Sched::RegPressure;
1053 for (unsigned i = 0; i != NumVals; ++i) {
1054 EVT VT = N->getValueType(i);
1055 if (VT == MVT::Glue || VT == MVT::Other)
1057 if (VT.isFloatingPoint() || VT.isVector())
1061 if (!N->isMachineOpcode())
1062 return Sched::RegPressure;
1064 // Load are scheduled for latency even if there instruction itinerary
1065 // is not available.
1066 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1067 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1069 if (MCID.getNumDefs() == 0)
1070 return Sched::RegPressure;
1071 if (!Itins->isEmpty() &&
1072 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1075 return Sched::RegPressure;
1078 //===----------------------------------------------------------------------===//
1080 //===----------------------------------------------------------------------===//
1082 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1083 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1085 default: llvm_unreachable("Unknown condition code!");
1086 case ISD::SETNE: return ARMCC::NE;
1087 case ISD::SETEQ: return ARMCC::EQ;
1088 case ISD::SETGT: return ARMCC::GT;
1089 case ISD::SETGE: return ARMCC::GE;
1090 case ISD::SETLT: return ARMCC::LT;
1091 case ISD::SETLE: return ARMCC::LE;
1092 case ISD::SETUGT: return ARMCC::HI;
1093 case ISD::SETUGE: return ARMCC::HS;
1094 case ISD::SETULT: return ARMCC::LO;
1095 case ISD::SETULE: return ARMCC::LS;
1099 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1100 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1101 ARMCC::CondCodes &CondCode2) {
1102 CondCode2 = ARMCC::AL;
1104 default: llvm_unreachable("Unknown FP condition!");
1106 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1108 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1110 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1111 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1112 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1113 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1114 case ISD::SETO: CondCode = ARMCC::VC; break;
1115 case ISD::SETUO: CondCode = ARMCC::VS; break;
1116 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1117 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1118 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1120 case ISD::SETULT: CondCode = ARMCC::LT; break;
1122 case ISD::SETULE: CondCode = ARMCC::LE; break;
1124 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1128 //===----------------------------------------------------------------------===//
1129 // Calling Convention Implementation
1130 //===----------------------------------------------------------------------===//
1132 #include "ARMGenCallingConv.inc"
1134 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1135 /// given CallingConvention value.
1136 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1138 bool isVarArg) const {
1141 llvm_unreachable("Unsupported calling convention");
1142 case CallingConv::Fast:
1143 if (Subtarget->hasVFP2() && !isVarArg) {
1144 if (!Subtarget->isAAPCS_ABI())
1145 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1146 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1147 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1150 case CallingConv::C: {
1151 // Use target triple & subtarget features to do actual dispatch.
1152 if (!Subtarget->isAAPCS_ABI())
1153 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1154 else if (Subtarget->hasVFP2() &&
1155 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1157 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1158 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1160 case CallingConv::ARM_AAPCS_VFP:
1162 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1164 case CallingConv::ARM_AAPCS:
1165 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1166 case CallingConv::ARM_APCS:
1167 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1171 /// LowerCallResult - Lower the result values of a call into the
1172 /// appropriate copies out of appropriate physical registers.
1174 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1175 CallingConv::ID CallConv, bool isVarArg,
1176 const SmallVectorImpl<ISD::InputArg> &Ins,
1177 DebugLoc dl, SelectionDAG &DAG,
1178 SmallVectorImpl<SDValue> &InVals) const {
1180 // Assign locations to each value returned by this call.
1181 SmallVector<CCValAssign, 16> RVLocs;
1182 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1183 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1184 CCInfo.AnalyzeCallResult(Ins,
1185 CCAssignFnForNode(CallConv, /* Return*/ true,
1188 // Copy all of the result registers out of their specified physreg.
1189 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1190 CCValAssign VA = RVLocs[i];
1193 if (VA.needsCustom()) {
1194 // Handle f64 or half of a v2f64.
1195 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1197 Chain = Lo.getValue(1);
1198 InFlag = Lo.getValue(2);
1199 VA = RVLocs[++i]; // skip ahead to next loc
1200 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1202 Chain = Hi.getValue(1);
1203 InFlag = Hi.getValue(2);
1204 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1206 if (VA.getLocVT() == MVT::v2f64) {
1207 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1208 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1209 DAG.getConstant(0, MVT::i32));
1211 VA = RVLocs[++i]; // skip ahead to next loc
1212 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1213 Chain = Lo.getValue(1);
1214 InFlag = Lo.getValue(2);
1215 VA = RVLocs[++i]; // skip ahead to next loc
1216 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1217 Chain = Hi.getValue(1);
1218 InFlag = Hi.getValue(2);
1219 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1220 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1221 DAG.getConstant(1, MVT::i32));
1224 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1226 Chain = Val.getValue(1);
1227 InFlag = Val.getValue(2);
1230 switch (VA.getLocInfo()) {
1231 default: llvm_unreachable("Unknown loc info!");
1232 case CCValAssign::Full: break;
1233 case CCValAssign::BCvt:
1234 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1238 InVals.push_back(Val);
1244 /// LowerMemOpCallTo - Store the argument to the stack.
1246 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1247 SDValue StackPtr, SDValue Arg,
1248 DebugLoc dl, SelectionDAG &DAG,
1249 const CCValAssign &VA,
1250 ISD::ArgFlagsTy Flags) const {
1251 unsigned LocMemOffset = VA.getLocMemOffset();
1252 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1253 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1254 return DAG.getStore(Chain, dl, Arg, PtrOff,
1255 MachinePointerInfo::getStack(LocMemOffset),
1259 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1260 SDValue Chain, SDValue &Arg,
1261 RegsToPassVector &RegsToPass,
1262 CCValAssign &VA, CCValAssign &NextVA,
1264 SmallVector<SDValue, 8> &MemOpChains,
1265 ISD::ArgFlagsTy Flags) const {
1267 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1268 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1269 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1271 if (NextVA.isRegLoc())
1272 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1274 assert(NextVA.isMemLoc());
1275 if (StackPtr.getNode() == 0)
1276 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1278 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1284 /// LowerCall - Lowering a call into a callseq_start <-
1285 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1288 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1289 CallingConv::ID CallConv, bool isVarArg,
1290 bool doesNotRet, bool &isTailCall,
1291 const SmallVectorImpl<ISD::OutputArg> &Outs,
1292 const SmallVectorImpl<SDValue> &OutVals,
1293 const SmallVectorImpl<ISD::InputArg> &Ins,
1294 DebugLoc dl, SelectionDAG &DAG,
1295 SmallVectorImpl<SDValue> &InVals) const {
1296 MachineFunction &MF = DAG.getMachineFunction();
1297 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1298 bool IsSibCall = false;
1299 // Disable tail calls if they're not supported.
1300 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1303 // Check if it's really possible to do a tail call.
1304 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1305 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1306 Outs, OutVals, Ins, DAG);
1307 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1308 // detected sibcalls.
1315 // Analyze operands of the call, assigning locations to each operand.
1316 SmallVector<CCValAssign, 16> ArgLocs;
1317 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1318 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1319 CCInfo.AnalyzeCallOperands(Outs,
1320 CCAssignFnForNode(CallConv, /* Return*/ false,
1323 // Get a count of how many bytes are to be pushed on the stack.
1324 unsigned NumBytes = CCInfo.getNextStackOffset();
1326 // For tail calls, memory operands are available in our caller's stack.
1330 // Adjust the stack pointer for the new arguments...
1331 // These operations are automatically eliminated by the prolog/epilog pass
1333 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1335 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1337 RegsToPassVector RegsToPass;
1338 SmallVector<SDValue, 8> MemOpChains;
1340 // Walk the register/memloc assignments, inserting copies/loads. In the case
1341 // of tail call optimization, arguments are handled later.
1342 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1344 ++i, ++realArgIdx) {
1345 CCValAssign &VA = ArgLocs[i];
1346 SDValue Arg = OutVals[realArgIdx];
1347 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1348 bool isByVal = Flags.isByVal();
1350 // Promote the value if needed.
1351 switch (VA.getLocInfo()) {
1352 default: llvm_unreachable("Unknown loc info!");
1353 case CCValAssign::Full: break;
1354 case CCValAssign::SExt:
1355 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1357 case CCValAssign::ZExt:
1358 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1360 case CCValAssign::AExt:
1361 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1363 case CCValAssign::BCvt:
1364 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1368 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1369 if (VA.needsCustom()) {
1370 if (VA.getLocVT() == MVT::v2f64) {
1371 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1372 DAG.getConstant(0, MVT::i32));
1373 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1374 DAG.getConstant(1, MVT::i32));
1376 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1377 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1379 VA = ArgLocs[++i]; // skip ahead to next loc
1380 if (VA.isRegLoc()) {
1381 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1382 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1384 assert(VA.isMemLoc());
1386 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1387 dl, DAG, VA, Flags));
1390 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1391 StackPtr, MemOpChains, Flags);
1393 } else if (VA.isRegLoc()) {
1394 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1395 } else if (isByVal) {
1396 assert(VA.isMemLoc());
1397 unsigned offset = 0;
1399 // True if this byval aggregate will be split between registers
1401 if (CCInfo.isFirstByValRegValid()) {
1402 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1404 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1405 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1406 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1407 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1408 MachinePointerInfo(),
1409 false, false, false, 0);
1410 MemOpChains.push_back(Load.getValue(1));
1411 RegsToPass.push_back(std::make_pair(j, Load));
1413 offset = ARM::R4 - CCInfo.getFirstByValReg();
1414 CCInfo.clearFirstByValReg();
1417 unsigned LocMemOffset = VA.getLocMemOffset();
1418 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1419 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1421 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1422 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1423 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1425 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1426 Flags.getByValAlign(),
1427 /*isVolatile=*/false,
1428 /*AlwaysInline=*/false,
1429 MachinePointerInfo(0),
1430 MachinePointerInfo(0)));
1432 } else if (!IsSibCall) {
1433 assert(VA.isMemLoc());
1435 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1436 dl, DAG, VA, Flags));
1440 if (!MemOpChains.empty())
1441 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1442 &MemOpChains[0], MemOpChains.size());
1444 // Build a sequence of copy-to-reg nodes chained together with token chain
1445 // and flag operands which copy the outgoing args into the appropriate regs.
1447 // Tail call byval lowering might overwrite argument registers so in case of
1448 // tail call optimization the copies to registers are lowered later.
1450 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1451 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1452 RegsToPass[i].second, InFlag);
1453 InFlag = Chain.getValue(1);
1456 // For tail calls lower the arguments to the 'real' stack slot.
1458 // Force all the incoming stack arguments to be loaded from the stack
1459 // before any new outgoing arguments are stored to the stack, because the
1460 // outgoing stack slots may alias the incoming argument stack slots, and
1461 // the alias isn't otherwise explicit. This is slightly more conservative
1462 // than necessary, because it means that each store effectively depends
1463 // on every argument instead of just those arguments it would clobber.
1465 // Do not flag preceding copytoreg stuff together with the following stuff.
1467 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1468 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1469 RegsToPass[i].second, InFlag);
1470 InFlag = Chain.getValue(1);
1475 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1476 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1477 // node so that legalize doesn't hack it.
1478 bool isDirect = false;
1479 bool isARMFunc = false;
1480 bool isLocalARMFunc = false;
1481 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1483 if (EnableARMLongCalls) {
1484 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1485 && "long-calls with non-static relocation model!");
1486 // Handle a global address or an external symbol. If it's not one of
1487 // those, the target's already in a register, so we don't need to do
1489 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1490 const GlobalValue *GV = G->getGlobal();
1491 // Create a constant pool entry for the callee address
1492 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1493 ARMConstantPoolValue *CPV =
1494 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1496 // Get the address of the callee into a register
1497 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1498 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1499 Callee = DAG.getLoad(getPointerTy(), dl,
1500 DAG.getEntryNode(), CPAddr,
1501 MachinePointerInfo::getConstantPool(),
1502 false, false, false, 0);
1503 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1504 const char *Sym = S->getSymbol();
1506 // Create a constant pool entry for the callee address
1507 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1508 ARMConstantPoolValue *CPV =
1509 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1510 ARMPCLabelIndex, 0);
1511 // Get the address of the callee into a register
1512 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1513 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1514 Callee = DAG.getLoad(getPointerTy(), dl,
1515 DAG.getEntryNode(), CPAddr,
1516 MachinePointerInfo::getConstantPool(),
1517 false, false, false, 0);
1519 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1520 const GlobalValue *GV = G->getGlobal();
1522 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1523 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1524 getTargetMachine().getRelocationModel() != Reloc::Static;
1525 isARMFunc = !Subtarget->isThumb() || isStub;
1526 // ARM call to a local ARM function is predicable.
1527 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1528 // tBX takes a register source operand.
1529 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1530 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1531 ARMConstantPoolValue *CPV =
1532 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1533 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1534 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1535 Callee = DAG.getLoad(getPointerTy(), dl,
1536 DAG.getEntryNode(), CPAddr,
1537 MachinePointerInfo::getConstantPool(),
1538 false, false, false, 0);
1539 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1540 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1541 getPointerTy(), Callee, PICLabel);
1543 // On ELF targets for PIC code, direct calls should go through the PLT
1544 unsigned OpFlags = 0;
1545 if (Subtarget->isTargetELF() &&
1546 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1547 OpFlags = ARMII::MO_PLT;
1548 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1550 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1552 bool isStub = Subtarget->isTargetDarwin() &&
1553 getTargetMachine().getRelocationModel() != Reloc::Static;
1554 isARMFunc = !Subtarget->isThumb() || isStub;
1555 // tBX takes a register source operand.
1556 const char *Sym = S->getSymbol();
1557 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1558 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1559 ARMConstantPoolValue *CPV =
1560 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1561 ARMPCLabelIndex, 4);
1562 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1563 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1564 Callee = DAG.getLoad(getPointerTy(), dl,
1565 DAG.getEntryNode(), CPAddr,
1566 MachinePointerInfo::getConstantPool(),
1567 false, false, false, 0);
1568 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1569 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1570 getPointerTy(), Callee, PICLabel);
1572 unsigned OpFlags = 0;
1573 // On ELF targets for PIC code, direct calls should go through the PLT
1574 if (Subtarget->isTargetELF() &&
1575 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1576 OpFlags = ARMII::MO_PLT;
1577 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1581 // FIXME: handle tail calls differently.
1583 if (Subtarget->isThumb()) {
1584 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1585 CallOpc = ARMISD::CALL_NOLINK;
1586 else if (doesNotRet && isDirect && !isARMFunc &&
1587 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1588 // "mov lr, pc; b _foo" to avoid confusing the RSP
1589 CallOpc = ARMISD::CALL_NOLINK;
1591 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1593 if (!isDirect && !Subtarget->hasV5TOps()) {
1594 CallOpc = ARMISD::CALL_NOLINK;
1595 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1596 // "mov lr, pc; b _foo" to avoid confusing the RSP
1597 CallOpc = ARMISD::CALL_NOLINK;
1599 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1602 std::vector<SDValue> Ops;
1603 Ops.push_back(Chain);
1604 Ops.push_back(Callee);
1606 // Add argument registers to the end of the list so that they are known live
1608 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1609 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1610 RegsToPass[i].second.getValueType()));
1612 // Add a register mask operand representing the call-preserved registers.
1613 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1614 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1615 assert(Mask && "Missing call preserved mask for calling convention");
1616 Ops.push_back(DAG.getRegisterMask(Mask));
1618 if (InFlag.getNode())
1619 Ops.push_back(InFlag);
1621 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1623 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1625 // Returns a chain and a flag for retval copy to use.
1626 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1627 InFlag = Chain.getValue(1);
1629 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1630 DAG.getIntPtrConstant(0, true), InFlag);
1632 InFlag = Chain.getValue(1);
1634 // Handle result values, copying them out of physregs into vregs that we
1636 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1640 /// HandleByVal - Every parameter *after* a byval parameter is passed
1641 /// on the stack. Remember the next parameter register to allocate,
1642 /// and then confiscate the rest of the parameter registers to insure
1645 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1646 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1647 assert((State->getCallOrPrologue() == Prologue ||
1648 State->getCallOrPrologue() == Call) &&
1649 "unhandled ParmContext");
1650 if ((!State->isFirstByValRegValid()) &&
1651 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1652 State->setFirstByValReg(reg);
1653 // At a call site, a byval parameter that is split between
1654 // registers and memory needs its size truncated here. In a
1655 // function prologue, such byval parameters are reassembled in
1656 // memory, and are not truncated.
1657 if (State->getCallOrPrologue() == Call) {
1658 unsigned excess = 4 * (ARM::R4 - reg);
1659 assert(size >= excess && "expected larger existing stack allocation");
1663 // Confiscate any remaining parameter registers to preclude their
1664 // assignment to subsequent parameters.
1665 while (State->AllocateReg(GPRArgRegs, 4))
1669 /// MatchingStackOffset - Return true if the given stack call argument is
1670 /// already available in the same position (relatively) of the caller's
1671 /// incoming argument stack.
1673 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1674 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1675 const TargetInstrInfo *TII) {
1676 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1678 if (Arg.getOpcode() == ISD::CopyFromReg) {
1679 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1680 if (!TargetRegisterInfo::isVirtualRegister(VR))
1682 MachineInstr *Def = MRI->getVRegDef(VR);
1685 if (!Flags.isByVal()) {
1686 if (!TII->isLoadFromStackSlot(Def, FI))
1691 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1692 if (Flags.isByVal())
1693 // ByVal argument is passed in as a pointer but it's now being
1694 // dereferenced. e.g.
1695 // define @foo(%struct.X* %A) {
1696 // tail call @bar(%struct.X* byval %A)
1699 SDValue Ptr = Ld->getBasePtr();
1700 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1703 FI = FINode->getIndex();
1707 assert(FI != INT_MAX);
1708 if (!MFI->isFixedObjectIndex(FI))
1710 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1713 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1714 /// for tail call optimization. Targets which want to do tail call
1715 /// optimization should implement this function.
1717 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1718 CallingConv::ID CalleeCC,
1720 bool isCalleeStructRet,
1721 bool isCallerStructRet,
1722 const SmallVectorImpl<ISD::OutputArg> &Outs,
1723 const SmallVectorImpl<SDValue> &OutVals,
1724 const SmallVectorImpl<ISD::InputArg> &Ins,
1725 SelectionDAG& DAG) const {
1726 const Function *CallerF = DAG.getMachineFunction().getFunction();
1727 CallingConv::ID CallerCC = CallerF->getCallingConv();
1728 bool CCMatch = CallerCC == CalleeCC;
1730 // Look for obvious safe cases to perform tail call optimization that do not
1731 // require ABI changes. This is what gcc calls sibcall.
1733 // Do not sibcall optimize vararg calls unless the call site is not passing
1735 if (isVarArg && !Outs.empty())
1738 // Also avoid sibcall optimization if either caller or callee uses struct
1739 // return semantics.
1740 if (isCalleeStructRet || isCallerStructRet)
1743 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1744 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1745 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1746 // support in the assembler and linker to be used. This would need to be
1747 // fixed to fully support tail calls in Thumb1.
1749 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1750 // LR. This means if we need to reload LR, it takes an extra instructions,
1751 // which outweighs the value of the tail call; but here we don't know yet
1752 // whether LR is going to be used. Probably the right approach is to
1753 // generate the tail call here and turn it back into CALL/RET in
1754 // emitEpilogue if LR is used.
1756 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1757 // but we need to make sure there are enough registers; the only valid
1758 // registers are the 4 used for parameters. We don't currently do this
1760 if (Subtarget->isThumb1Only())
1763 // If the calling conventions do not match, then we'd better make sure the
1764 // results are returned in the same way as what the caller expects.
1766 SmallVector<CCValAssign, 16> RVLocs1;
1767 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1768 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1769 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1771 SmallVector<CCValAssign, 16> RVLocs2;
1772 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1773 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1774 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1776 if (RVLocs1.size() != RVLocs2.size())
1778 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1779 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1781 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1783 if (RVLocs1[i].isRegLoc()) {
1784 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1787 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1793 // If the callee takes no arguments then go on to check the results of the
1795 if (!Outs.empty()) {
1796 // Check if stack adjustment is needed. For now, do not do this if any
1797 // argument is passed on the stack.
1798 SmallVector<CCValAssign, 16> ArgLocs;
1799 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1800 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1801 CCInfo.AnalyzeCallOperands(Outs,
1802 CCAssignFnForNode(CalleeCC, false, isVarArg));
1803 if (CCInfo.getNextStackOffset()) {
1804 MachineFunction &MF = DAG.getMachineFunction();
1806 // Check if the arguments are already laid out in the right way as
1807 // the caller's fixed stack objects.
1808 MachineFrameInfo *MFI = MF.getFrameInfo();
1809 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1810 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1811 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1813 ++i, ++realArgIdx) {
1814 CCValAssign &VA = ArgLocs[i];
1815 EVT RegVT = VA.getLocVT();
1816 SDValue Arg = OutVals[realArgIdx];
1817 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1818 if (VA.getLocInfo() == CCValAssign::Indirect)
1820 if (VA.needsCustom()) {
1821 // f64 and vector types are split into multiple registers or
1822 // register/stack-slot combinations. The types will not match
1823 // the registers; give up on memory f64 refs until we figure
1824 // out what to do about this.
1827 if (!ArgLocs[++i].isRegLoc())
1829 if (RegVT == MVT::v2f64) {
1830 if (!ArgLocs[++i].isRegLoc())
1832 if (!ArgLocs[++i].isRegLoc())
1835 } else if (!VA.isRegLoc()) {
1836 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1848 ARMTargetLowering::LowerReturn(SDValue Chain,
1849 CallingConv::ID CallConv, bool isVarArg,
1850 const SmallVectorImpl<ISD::OutputArg> &Outs,
1851 const SmallVectorImpl<SDValue> &OutVals,
1852 DebugLoc dl, SelectionDAG &DAG) const {
1854 // CCValAssign - represent the assignment of the return value to a location.
1855 SmallVector<CCValAssign, 16> RVLocs;
1857 // CCState - Info about the registers and stack slots.
1858 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1859 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1861 // Analyze outgoing return values.
1862 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1865 // If this is the first return lowered for this function, add
1866 // the regs to the liveout set for the function.
1867 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1868 for (unsigned i = 0; i != RVLocs.size(); ++i)
1869 if (RVLocs[i].isRegLoc())
1870 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1875 // Copy the result values into the output registers.
1876 for (unsigned i = 0, realRVLocIdx = 0;
1878 ++i, ++realRVLocIdx) {
1879 CCValAssign &VA = RVLocs[i];
1880 assert(VA.isRegLoc() && "Can only return in registers!");
1882 SDValue Arg = OutVals[realRVLocIdx];
1884 switch (VA.getLocInfo()) {
1885 default: llvm_unreachable("Unknown loc info!");
1886 case CCValAssign::Full: break;
1887 case CCValAssign::BCvt:
1888 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1892 if (VA.needsCustom()) {
1893 if (VA.getLocVT() == MVT::v2f64) {
1894 // Extract the first half and return it in two registers.
1895 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1896 DAG.getConstant(0, MVT::i32));
1897 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1898 DAG.getVTList(MVT::i32, MVT::i32), Half);
1900 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1901 Flag = Chain.getValue(1);
1902 VA = RVLocs[++i]; // skip ahead to next loc
1903 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1904 HalfGPRs.getValue(1), Flag);
1905 Flag = Chain.getValue(1);
1906 VA = RVLocs[++i]; // skip ahead to next loc
1908 // Extract the 2nd half and fall through to handle it as an f64 value.
1909 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1910 DAG.getConstant(1, MVT::i32));
1912 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1914 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1915 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1916 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1917 Flag = Chain.getValue(1);
1918 VA = RVLocs[++i]; // skip ahead to next loc
1919 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1922 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1924 // Guarantee that all emitted copies are
1925 // stuck together, avoiding something bad.
1926 Flag = Chain.getValue(1);
1931 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1933 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1938 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1939 if (N->getNumValues() != 1)
1941 if (!N->hasNUsesOfValue(1, 0))
1944 unsigned NumCopies = 0;
1945 SDNode* Copies[2] = { 0, 0 };
1946 SDNode *Use = *N->use_begin();
1947 if (Use->getOpcode() == ISD::CopyToReg) {
1948 Copies[NumCopies++] = Use;
1949 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1950 // f64 returned in a pair of GPRs.
1951 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1953 if (UI->getOpcode() != ISD::CopyToReg)
1955 Copies[UI.getUse().getResNo()] = *UI;
1958 } else if (Use->getOpcode() == ISD::BITCAST) {
1959 // f32 returned in a single GPR.
1960 if (!Use->hasNUsesOfValue(1, 0))
1962 Use = *Use->use_begin();
1963 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1965 Copies[NumCopies++] = Use;
1970 if (NumCopies != 1 && NumCopies != 2)
1973 bool HasRet = false;
1974 for (unsigned i = 0; i < NumCopies; ++i) {
1975 SDNode *Copy = Copies[i];
1976 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1978 if (UI->getOpcode() == ISD::CopyToReg) {
1980 if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1984 if (UI->getOpcode() != ARMISD::RET_FLAG)
1993 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1994 if (!EnableARMTailCalls)
1997 if (!CI->isTailCall())
2000 return !Subtarget->isThumb1Only();
2003 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2004 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2005 // one of the above mentioned nodes. It has to be wrapped because otherwise
2006 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2007 // be used to form addressing mode. These wrapped nodes will be selected
2009 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2010 EVT PtrVT = Op.getValueType();
2011 // FIXME there is no actual debug info here
2012 DebugLoc dl = Op.getDebugLoc();
2013 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2015 if (CP->isMachineConstantPoolEntry())
2016 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2017 CP->getAlignment());
2019 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2020 CP->getAlignment());
2021 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2024 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2025 return MachineJumpTableInfo::EK_Inline;
2028 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2029 SelectionDAG &DAG) const {
2030 MachineFunction &MF = DAG.getMachineFunction();
2031 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2032 unsigned ARMPCLabelIndex = 0;
2033 DebugLoc DL = Op.getDebugLoc();
2034 EVT PtrVT = getPointerTy();
2035 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2036 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2038 if (RelocM == Reloc::Static) {
2039 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2041 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2042 ARMPCLabelIndex = AFI->createPICLabelUId();
2043 ARMConstantPoolValue *CPV =
2044 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2045 ARMCP::CPBlockAddress, PCAdj);
2046 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2048 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2049 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2050 MachinePointerInfo::getConstantPool(),
2051 false, false, false, 0);
2052 if (RelocM == Reloc::Static)
2054 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2055 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2058 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2060 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2061 SelectionDAG &DAG) const {
2062 DebugLoc dl = GA->getDebugLoc();
2063 EVT PtrVT = getPointerTy();
2064 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2065 MachineFunction &MF = DAG.getMachineFunction();
2066 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2067 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2068 ARMConstantPoolValue *CPV =
2069 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2070 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2071 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2072 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2073 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2074 MachinePointerInfo::getConstantPool(),
2075 false, false, false, 0);
2076 SDValue Chain = Argument.getValue(1);
2078 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2079 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2081 // call __tls_get_addr.
2084 Entry.Node = Argument;
2085 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2086 Args.push_back(Entry);
2087 // FIXME: is there useful debug info available here?
2088 std::pair<SDValue, SDValue> CallResult =
2089 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2090 false, false, false, false,
2091 0, CallingConv::C, /*isTailCall=*/false,
2092 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2093 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2094 return CallResult.first;
2097 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2098 // "local exec" model.
2100 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2101 SelectionDAG &DAG) const {
2102 const GlobalValue *GV = GA->getGlobal();
2103 DebugLoc dl = GA->getDebugLoc();
2105 SDValue Chain = DAG.getEntryNode();
2106 EVT PtrVT = getPointerTy();
2107 // Get the Thread Pointer
2108 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2110 if (GV->isDeclaration()) {
2111 MachineFunction &MF = DAG.getMachineFunction();
2112 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2113 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2114 // Initial exec model.
2115 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2116 ARMConstantPoolValue *CPV =
2117 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2118 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2120 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2121 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2122 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2123 MachinePointerInfo::getConstantPool(),
2124 false, false, false, 0);
2125 Chain = Offset.getValue(1);
2127 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2128 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2130 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2131 MachinePointerInfo::getConstantPool(),
2132 false, false, false, 0);
2135 ARMConstantPoolValue *CPV =
2136 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2137 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2138 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2139 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2140 MachinePointerInfo::getConstantPool(),
2141 false, false, false, 0);
2144 // The address of the thread local variable is the add of the thread
2145 // pointer with the offset of the variable.
2146 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2150 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2151 // TODO: implement the "local dynamic" model
2152 assert(Subtarget->isTargetELF() &&
2153 "TLS not implemented for non-ELF targets");
2154 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2155 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2156 // otherwise use the "Local Exec" TLS Model
2157 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2158 return LowerToTLSGeneralDynamicModel(GA, DAG);
2160 return LowerToTLSExecModels(GA, DAG);
2163 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2164 SelectionDAG &DAG) const {
2165 EVT PtrVT = getPointerTy();
2166 DebugLoc dl = Op.getDebugLoc();
2167 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2168 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2169 if (RelocM == Reloc::PIC_) {
2170 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2171 ARMConstantPoolValue *CPV =
2172 ARMConstantPoolConstant::Create(GV,
2173 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2174 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2175 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2176 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2178 MachinePointerInfo::getConstantPool(),
2179 false, false, false, 0);
2180 SDValue Chain = Result.getValue(1);
2181 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2182 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2184 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2185 MachinePointerInfo::getGOT(),
2186 false, false, false, 0);
2190 // If we have T2 ops, we can materialize the address directly via movt/movw
2191 // pair. This is always cheaper.
2192 if (Subtarget->useMovt()) {
2194 // FIXME: Once remat is capable of dealing with instructions with register
2195 // operands, expand this into two nodes.
2196 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2197 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2199 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2200 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2201 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2202 MachinePointerInfo::getConstantPool(),
2203 false, false, false, 0);
2207 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2208 SelectionDAG &DAG) const {
2209 EVT PtrVT = getPointerTy();
2210 DebugLoc dl = Op.getDebugLoc();
2211 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2212 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2213 MachineFunction &MF = DAG.getMachineFunction();
2214 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2216 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2217 // update ARMFastISel::ARMMaterializeGV.
2218 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2220 // FIXME: Once remat is capable of dealing with instructions with register
2221 // operands, expand this into two nodes.
2222 if (RelocM == Reloc::Static)
2223 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2224 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2226 unsigned Wrapper = (RelocM == Reloc::PIC_)
2227 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2228 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2229 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2230 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2231 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2232 MachinePointerInfo::getGOT(),
2233 false, false, false, 0);
2237 unsigned ARMPCLabelIndex = 0;
2239 if (RelocM == Reloc::Static) {
2240 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2242 ARMPCLabelIndex = AFI->createPICLabelUId();
2243 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2244 ARMConstantPoolValue *CPV =
2245 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2247 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2249 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2251 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2252 MachinePointerInfo::getConstantPool(),
2253 false, false, false, 0);
2254 SDValue Chain = Result.getValue(1);
2256 if (RelocM == Reloc::PIC_) {
2257 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2258 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2261 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2262 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2263 false, false, false, 0);
2268 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2269 SelectionDAG &DAG) const {
2270 assert(Subtarget->isTargetELF() &&
2271 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2272 MachineFunction &MF = DAG.getMachineFunction();
2273 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2274 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2275 EVT PtrVT = getPointerTy();
2276 DebugLoc dl = Op.getDebugLoc();
2277 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2278 ARMConstantPoolValue *CPV =
2279 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2280 ARMPCLabelIndex, PCAdj);
2281 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2282 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2283 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2284 MachinePointerInfo::getConstantPool(),
2285 false, false, false, 0);
2286 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2287 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2291 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2292 DebugLoc dl = Op.getDebugLoc();
2293 SDValue Val = DAG.getConstant(0, MVT::i32);
2294 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2295 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2296 Op.getOperand(1), Val);
2300 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2301 DebugLoc dl = Op.getDebugLoc();
2302 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2303 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2307 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2308 const ARMSubtarget *Subtarget) const {
2309 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2310 DebugLoc dl = Op.getDebugLoc();
2312 default: return SDValue(); // Don't custom lower most intrinsics.
2313 case Intrinsic::arm_thread_pointer: {
2314 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2315 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2317 case Intrinsic::eh_sjlj_lsda: {
2318 MachineFunction &MF = DAG.getMachineFunction();
2319 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2320 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2321 EVT PtrVT = getPointerTy();
2322 DebugLoc dl = Op.getDebugLoc();
2323 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2325 unsigned PCAdj = (RelocM != Reloc::PIC_)
2326 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2327 ARMConstantPoolValue *CPV =
2328 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2329 ARMCP::CPLSDA, PCAdj);
2330 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2331 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2333 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2334 MachinePointerInfo::getConstantPool(),
2335 false, false, false, 0);
2337 if (RelocM == Reloc::PIC_) {
2338 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2339 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2343 case Intrinsic::arm_neon_vmulls:
2344 case Intrinsic::arm_neon_vmullu: {
2345 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2346 ? ARMISD::VMULLs : ARMISD::VMULLu;
2347 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2348 Op.getOperand(1), Op.getOperand(2));
2353 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2354 const ARMSubtarget *Subtarget) {
2355 DebugLoc dl = Op.getDebugLoc();
2356 if (!Subtarget->hasDataBarrier()) {
2357 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2358 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2360 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2361 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2362 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2363 DAG.getConstant(0, MVT::i32));
2366 SDValue Op5 = Op.getOperand(5);
2367 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2368 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2369 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2370 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2372 ARM_MB::MemBOpt DMBOpt;
2373 if (isDeviceBarrier)
2374 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2376 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2377 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2378 DAG.getConstant(DMBOpt, MVT::i32));
2382 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2383 const ARMSubtarget *Subtarget) {
2384 // FIXME: handle "fence singlethread" more efficiently.
2385 DebugLoc dl = Op.getDebugLoc();
2386 if (!Subtarget->hasDataBarrier()) {
2387 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2388 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2390 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2391 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2392 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2393 DAG.getConstant(0, MVT::i32));
2396 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2397 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2400 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2401 const ARMSubtarget *Subtarget) {
2402 // ARM pre v5TE and Thumb1 does not have preload instructions.
2403 if (!(Subtarget->isThumb2() ||
2404 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2405 // Just preserve the chain.
2406 return Op.getOperand(0);
2408 DebugLoc dl = Op.getDebugLoc();
2409 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2411 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2412 // ARMv7 with MP extension has PLDW.
2413 return Op.getOperand(0);
2415 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2416 if (Subtarget->isThumb()) {
2418 isRead = ~isRead & 1;
2419 isData = ~isData & 1;
2422 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2423 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2424 DAG.getConstant(isData, MVT::i32));
2427 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2428 MachineFunction &MF = DAG.getMachineFunction();
2429 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2431 // vastart just stores the address of the VarArgsFrameIndex slot into the
2432 // memory location argument.
2433 DebugLoc dl = Op.getDebugLoc();
2434 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2435 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2436 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2437 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2438 MachinePointerInfo(SV), false, false, 0);
2442 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2443 SDValue &Root, SelectionDAG &DAG,
2444 DebugLoc dl) const {
2445 MachineFunction &MF = DAG.getMachineFunction();
2446 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2448 const TargetRegisterClass *RC;
2449 if (AFI->isThumb1OnlyFunction())
2450 RC = ARM::tGPRRegisterClass;
2452 RC = ARM::GPRRegisterClass;
2454 // Transform the arguments stored in physical registers into virtual ones.
2455 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2456 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2459 if (NextVA.isMemLoc()) {
2460 MachineFrameInfo *MFI = MF.getFrameInfo();
2461 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2463 // Create load node to retrieve arguments from the stack.
2464 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2465 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2466 MachinePointerInfo::getFixedStack(FI),
2467 false, false, false, 0);
2469 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2470 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2473 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2477 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2478 unsigned &VARegSize, unsigned &VARegSaveSize)
2481 if (CCInfo.isFirstByValRegValid())
2482 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2484 unsigned int firstUnalloced;
2485 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2486 sizeof(GPRArgRegs) /
2487 sizeof(GPRArgRegs[0]));
2488 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2491 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2492 VARegSize = NumGPRs * 4;
2493 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2496 // The remaining GPRs hold either the beginning of variable-argument
2497 // data, or the beginning of an aggregate passed by value (usuall
2498 // byval). Either way, we allocate stack slots adjacent to the data
2499 // provided by our caller, and store the unallocated registers there.
2500 // If this is a variadic function, the va_list pointer will begin with
2501 // these values; otherwise, this reassembles a (byval) structure that
2502 // was split between registers and memory.
2504 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2505 DebugLoc dl, SDValue &Chain,
2506 unsigned ArgOffset) const {
2507 MachineFunction &MF = DAG.getMachineFunction();
2508 MachineFrameInfo *MFI = MF.getFrameInfo();
2509 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2510 unsigned firstRegToSaveIndex;
2511 if (CCInfo.isFirstByValRegValid())
2512 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2514 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2515 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2518 unsigned VARegSize, VARegSaveSize;
2519 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2520 if (VARegSaveSize) {
2521 // If this function is vararg, store any remaining integer argument regs
2522 // to their spots on the stack so that they may be loaded by deferencing
2523 // the result of va_next.
2524 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2525 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2526 ArgOffset + VARegSaveSize
2529 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2532 SmallVector<SDValue, 4> MemOps;
2533 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2534 const TargetRegisterClass *RC;
2535 if (AFI->isThumb1OnlyFunction())
2536 RC = ARM::tGPRRegisterClass;
2538 RC = ARM::GPRRegisterClass;
2540 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2541 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2543 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2544 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2546 MemOps.push_back(Store);
2547 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2548 DAG.getConstant(4, getPointerTy()));
2550 if (!MemOps.empty())
2551 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2552 &MemOps[0], MemOps.size());
2554 // This will point to the next argument passed via stack.
2555 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2559 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2560 CallingConv::ID CallConv, bool isVarArg,
2561 const SmallVectorImpl<ISD::InputArg>
2563 DebugLoc dl, SelectionDAG &DAG,
2564 SmallVectorImpl<SDValue> &InVals)
2566 MachineFunction &MF = DAG.getMachineFunction();
2567 MachineFrameInfo *MFI = MF.getFrameInfo();
2569 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2571 // Assign locations to all of the incoming arguments.
2572 SmallVector<CCValAssign, 16> ArgLocs;
2573 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2574 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2575 CCInfo.AnalyzeFormalArguments(Ins,
2576 CCAssignFnForNode(CallConv, /* Return*/ false,
2579 SmallVector<SDValue, 16> ArgValues;
2580 int lastInsIndex = -1;
2583 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2584 CCValAssign &VA = ArgLocs[i];
2586 // Arguments stored in registers.
2587 if (VA.isRegLoc()) {
2588 EVT RegVT = VA.getLocVT();
2590 if (VA.needsCustom()) {
2591 // f64 and vector types are split up into multiple registers or
2592 // combinations of registers and stack slots.
2593 if (VA.getLocVT() == MVT::v2f64) {
2594 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2596 VA = ArgLocs[++i]; // skip ahead to next loc
2598 if (VA.isMemLoc()) {
2599 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2600 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2601 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2602 MachinePointerInfo::getFixedStack(FI),
2603 false, false, false, 0);
2605 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2608 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2609 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2610 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2611 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2612 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2614 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2617 const TargetRegisterClass *RC;
2619 if (RegVT == MVT::f32)
2620 RC = ARM::SPRRegisterClass;
2621 else if (RegVT == MVT::f64)
2622 RC = ARM::DPRRegisterClass;
2623 else if (RegVT == MVT::v2f64)
2624 RC = ARM::QPRRegisterClass;
2625 else if (RegVT == MVT::i32)
2626 RC = (AFI->isThumb1OnlyFunction() ?
2627 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2629 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2631 // Transform the arguments in physical registers into virtual ones.
2632 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2633 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2636 // If this is an 8 or 16-bit value, it is really passed promoted
2637 // to 32 bits. Insert an assert[sz]ext to capture this, then
2638 // truncate to the right size.
2639 switch (VA.getLocInfo()) {
2640 default: llvm_unreachable("Unknown loc info!");
2641 case CCValAssign::Full: break;
2642 case CCValAssign::BCvt:
2643 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2645 case CCValAssign::SExt:
2646 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2647 DAG.getValueType(VA.getValVT()));
2648 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2650 case CCValAssign::ZExt:
2651 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2652 DAG.getValueType(VA.getValVT()));
2653 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2657 InVals.push_back(ArgValue);
2659 } else { // VA.isRegLoc()
2662 assert(VA.isMemLoc());
2663 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2665 int index = ArgLocs[i].getValNo();
2667 // Some Ins[] entries become multiple ArgLoc[] entries.
2668 // Process them only once.
2669 if (index != lastInsIndex)
2671 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2672 // FIXME: For now, all byval parameter objects are marked mutable.
2673 // This can be changed with more analysis.
2674 // In case of tail call optimization mark all arguments mutable.
2675 // Since they could be overwritten by lowering of arguments in case of
2677 if (Flags.isByVal()) {
2678 unsigned VARegSize, VARegSaveSize;
2679 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2680 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2681 unsigned Bytes = Flags.getByValSize() - VARegSize;
2682 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2683 int FI = MFI->CreateFixedObject(Bytes,
2684 VA.getLocMemOffset(), false);
2685 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2687 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2688 VA.getLocMemOffset(), true);
2690 // Create load nodes to retrieve arguments from the stack.
2691 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2692 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2693 MachinePointerInfo::getFixedStack(FI),
2694 false, false, false, 0));
2696 lastInsIndex = index;
2703 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2708 /// isFloatingPointZero - Return true if this is +0.0.
2709 static bool isFloatingPointZero(SDValue Op) {
2710 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2711 return CFP->getValueAPF().isPosZero();
2712 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2713 // Maybe this has already been legalized into the constant pool?
2714 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2715 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2716 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2717 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2718 return CFP->getValueAPF().isPosZero();
2724 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2725 /// the given operands.
2727 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2728 SDValue &ARMcc, SelectionDAG &DAG,
2729 DebugLoc dl) const {
2730 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2731 unsigned C = RHSC->getZExtValue();
2732 if (!isLegalICmpImmediate(C)) {
2733 // Constant does not fit, try adjusting it by one?
2738 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2739 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2740 RHS = DAG.getConstant(C-1, MVT::i32);
2745 if (C != 0 && isLegalICmpImmediate(C-1)) {
2746 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2747 RHS = DAG.getConstant(C-1, MVT::i32);
2752 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2753 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2754 RHS = DAG.getConstant(C+1, MVT::i32);
2759 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2760 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2761 RHS = DAG.getConstant(C+1, MVT::i32);
2768 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2769 ARMISD::NodeType CompareType;
2772 CompareType = ARMISD::CMP;
2777 CompareType = ARMISD::CMPZ;
2780 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2781 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2784 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2786 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2787 DebugLoc dl) const {
2789 if (!isFloatingPointZero(RHS))
2790 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2792 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2793 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2796 /// duplicateCmp - Glue values can have only one use, so this function
2797 /// duplicates a comparison node.
2799 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2800 unsigned Opc = Cmp.getOpcode();
2801 DebugLoc DL = Cmp.getDebugLoc();
2802 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2803 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2805 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2806 Cmp = Cmp.getOperand(0);
2807 Opc = Cmp.getOpcode();
2808 if (Opc == ARMISD::CMPFP)
2809 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2811 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2812 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2814 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2817 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2818 SDValue Cond = Op.getOperand(0);
2819 SDValue SelectTrue = Op.getOperand(1);
2820 SDValue SelectFalse = Op.getOperand(2);
2821 DebugLoc dl = Op.getDebugLoc();
2825 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2826 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2828 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2829 const ConstantSDNode *CMOVTrue =
2830 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2831 const ConstantSDNode *CMOVFalse =
2832 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2834 if (CMOVTrue && CMOVFalse) {
2835 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2836 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2840 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2842 False = SelectFalse;
2843 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2848 if (True.getNode() && False.getNode()) {
2849 EVT VT = Op.getValueType();
2850 SDValue ARMcc = Cond.getOperand(2);
2851 SDValue CCR = Cond.getOperand(3);
2852 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2853 assert(True.getValueType() == VT);
2854 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2859 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2860 // undefined bits before doing a full-word comparison with zero.
2861 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2862 DAG.getConstant(1, Cond.getValueType()));
2864 return DAG.getSelectCC(dl, Cond,
2865 DAG.getConstant(0, Cond.getValueType()),
2866 SelectTrue, SelectFalse, ISD::SETNE);
2869 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2870 EVT VT = Op.getValueType();
2871 SDValue LHS = Op.getOperand(0);
2872 SDValue RHS = Op.getOperand(1);
2873 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2874 SDValue TrueVal = Op.getOperand(2);
2875 SDValue FalseVal = Op.getOperand(3);
2876 DebugLoc dl = Op.getDebugLoc();
2878 if (LHS.getValueType() == MVT::i32) {
2880 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2881 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2882 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2885 ARMCC::CondCodes CondCode, CondCode2;
2886 FPCCToARMCC(CC, CondCode, CondCode2);
2888 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2889 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2890 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2891 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2893 if (CondCode2 != ARMCC::AL) {
2894 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2895 // FIXME: Needs another CMP because flag can have but one use.
2896 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2897 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2898 Result, TrueVal, ARMcc2, CCR, Cmp2);
2903 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2904 /// to morph to an integer compare sequence.
2905 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2906 const ARMSubtarget *Subtarget) {
2907 SDNode *N = Op.getNode();
2908 if (!N->hasOneUse())
2909 // Otherwise it requires moving the value from fp to integer registers.
2911 if (!N->getNumValues())
2913 EVT VT = Op.getValueType();
2914 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2915 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2916 // vmrs are very slow, e.g. cortex-a8.
2919 if (isFloatingPointZero(Op)) {
2923 return ISD::isNormalLoad(N);
2926 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2927 if (isFloatingPointZero(Op))
2928 return DAG.getConstant(0, MVT::i32);
2930 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2931 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2932 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2933 Ld->isVolatile(), Ld->isNonTemporal(),
2934 Ld->isInvariant(), Ld->getAlignment());
2936 llvm_unreachable("Unknown VFP cmp argument!");
2939 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2940 SDValue &RetVal1, SDValue &RetVal2) {
2941 if (isFloatingPointZero(Op)) {
2942 RetVal1 = DAG.getConstant(0, MVT::i32);
2943 RetVal2 = DAG.getConstant(0, MVT::i32);
2947 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2948 SDValue Ptr = Ld->getBasePtr();
2949 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2950 Ld->getChain(), Ptr,
2951 Ld->getPointerInfo(),
2952 Ld->isVolatile(), Ld->isNonTemporal(),
2953 Ld->isInvariant(), Ld->getAlignment());
2955 EVT PtrType = Ptr.getValueType();
2956 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2957 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2958 PtrType, Ptr, DAG.getConstant(4, PtrType));
2959 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2960 Ld->getChain(), NewPtr,
2961 Ld->getPointerInfo().getWithOffset(4),
2962 Ld->isVolatile(), Ld->isNonTemporal(),
2963 Ld->isInvariant(), NewAlign);
2967 llvm_unreachable("Unknown VFP cmp argument!");
2970 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2971 /// f32 and even f64 comparisons to integer ones.
2973 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2974 SDValue Chain = Op.getOperand(0);
2975 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2976 SDValue LHS = Op.getOperand(2);
2977 SDValue RHS = Op.getOperand(3);
2978 SDValue Dest = Op.getOperand(4);
2979 DebugLoc dl = Op.getDebugLoc();
2981 bool LHSSeenZero = false;
2982 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
2983 bool RHSSeenZero = false;
2984 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
2985 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
2986 // If unsafe fp math optimization is enabled and there are no other uses of
2987 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2988 // to an integer comparison.
2989 if (CC == ISD::SETOEQ)
2991 else if (CC == ISD::SETUNE)
2994 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
2996 if (LHS.getValueType() == MVT::f32) {
2997 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2998 bitcastf32Toi32(LHS, DAG), Mask);
2999 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3000 bitcastf32Toi32(RHS, DAG), Mask);
3001 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3002 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3003 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3004 Chain, Dest, ARMcc, CCR, Cmp);
3009 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3010 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3011 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3012 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3013 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3014 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3015 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3016 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3017 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3023 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3024 SDValue Chain = Op.getOperand(0);
3025 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3026 SDValue LHS = Op.getOperand(2);
3027 SDValue RHS = Op.getOperand(3);
3028 SDValue Dest = Op.getOperand(4);
3029 DebugLoc dl = Op.getDebugLoc();
3031 if (LHS.getValueType() == MVT::i32) {
3033 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3034 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3035 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3036 Chain, Dest, ARMcc, CCR, Cmp);
3039 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3041 if (getTargetMachine().Options.UnsafeFPMath &&
3042 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3043 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3044 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3045 if (Result.getNode())
3049 ARMCC::CondCodes CondCode, CondCode2;
3050 FPCCToARMCC(CC, CondCode, CondCode2);
3052 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3053 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3054 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3055 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3056 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3057 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3058 if (CondCode2 != ARMCC::AL) {
3059 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3060 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3061 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3066 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3067 SDValue Chain = Op.getOperand(0);
3068 SDValue Table = Op.getOperand(1);
3069 SDValue Index = Op.getOperand(2);
3070 DebugLoc dl = Op.getDebugLoc();
3072 EVT PTy = getPointerTy();
3073 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3074 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3075 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3076 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3077 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3078 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3079 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3080 if (Subtarget->isThumb2()) {
3081 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3082 // which does another jump to the destination. This also makes it easier
3083 // to translate it to TBB / TBH later.
3084 // FIXME: This might not work if the function is extremely large.
3085 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3086 Addr, Op.getOperand(2), JTI, UId);
3088 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3089 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3090 MachinePointerInfo::getJumpTable(),
3091 false, false, false, 0);
3092 Chain = Addr.getValue(1);
3093 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3094 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3096 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3097 MachinePointerInfo::getJumpTable(),
3098 false, false, false, 0);
3099 Chain = Addr.getValue(1);
3100 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3104 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3105 EVT VT = Op.getValueType();
3106 DebugLoc dl = Op.getDebugLoc();
3108 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3109 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3111 return DAG.UnrollVectorOp(Op.getNode());
3114 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3115 "Invalid type for custom lowering!");
3116 if (VT != MVT::v4i16)
3117 return DAG.UnrollVectorOp(Op.getNode());
3119 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3120 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3123 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3124 EVT VT = Op.getValueType();
3126 return LowerVectorFP_TO_INT(Op, DAG);
3128 DebugLoc dl = Op.getDebugLoc();
3131 switch (Op.getOpcode()) {
3132 default: llvm_unreachable("Invalid opcode!");
3133 case ISD::FP_TO_SINT:
3134 Opc = ARMISD::FTOSI;
3136 case ISD::FP_TO_UINT:
3137 Opc = ARMISD::FTOUI;
3140 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3141 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3144 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3145 EVT VT = Op.getValueType();
3146 DebugLoc dl = Op.getDebugLoc();
3148 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3149 if (VT.getVectorElementType() == MVT::f32)
3151 return DAG.UnrollVectorOp(Op.getNode());
3154 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3155 "Invalid type for custom lowering!");
3156 if (VT != MVT::v4f32)
3157 return DAG.UnrollVectorOp(Op.getNode());
3161 switch (Op.getOpcode()) {
3162 default: llvm_unreachable("Invalid opcode!");
3163 case ISD::SINT_TO_FP:
3164 CastOpc = ISD::SIGN_EXTEND;
3165 Opc = ISD::SINT_TO_FP;
3167 case ISD::UINT_TO_FP:
3168 CastOpc = ISD::ZERO_EXTEND;
3169 Opc = ISD::UINT_TO_FP;
3173 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3174 return DAG.getNode(Opc, dl, VT, Op);
3177 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3178 EVT VT = Op.getValueType();
3180 return LowerVectorINT_TO_FP(Op, DAG);
3182 DebugLoc dl = Op.getDebugLoc();
3185 switch (Op.getOpcode()) {
3186 default: llvm_unreachable("Invalid opcode!");
3187 case ISD::SINT_TO_FP:
3188 Opc = ARMISD::SITOF;
3190 case ISD::UINT_TO_FP:
3191 Opc = ARMISD::UITOF;
3195 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3196 return DAG.getNode(Opc, dl, VT, Op);
3199 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3200 // Implement fcopysign with a fabs and a conditional fneg.
3201 SDValue Tmp0 = Op.getOperand(0);
3202 SDValue Tmp1 = Op.getOperand(1);
3203 DebugLoc dl = Op.getDebugLoc();
3204 EVT VT = Op.getValueType();
3205 EVT SrcVT = Tmp1.getValueType();
3206 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3207 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3208 bool UseNEON = !InGPR && Subtarget->hasNEON();
3211 // Use VBSL to copy the sign bit.
3212 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3213 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3214 DAG.getTargetConstant(EncodedVal, MVT::i32));
3215 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3217 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3218 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3219 DAG.getConstant(32, MVT::i32));
3220 else /*if (VT == MVT::f32)*/
3221 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3222 if (SrcVT == MVT::f32) {
3223 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3225 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3226 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3227 DAG.getConstant(32, MVT::i32));
3228 } else if (VT == MVT::f32)
3229 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3230 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3231 DAG.getConstant(32, MVT::i32));
3232 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3233 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3235 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3237 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3238 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3239 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3241 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3242 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3243 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3244 if (VT == MVT::f32) {
3245 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3246 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3247 DAG.getConstant(0, MVT::i32));
3249 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3255 // Bitcast operand 1 to i32.
3256 if (SrcVT == MVT::f64)
3257 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3258 &Tmp1, 1).getValue(1);
3259 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3261 // Or in the signbit with integer operations.
3262 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3263 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3264 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3265 if (VT == MVT::f32) {
3266 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3267 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3268 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3269 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3272 // f64: Or the high part with signbit and then combine two parts.
3273 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3275 SDValue Lo = Tmp0.getValue(0);
3276 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3277 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3278 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3281 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3282 MachineFunction &MF = DAG.getMachineFunction();
3283 MachineFrameInfo *MFI = MF.getFrameInfo();
3284 MFI->setReturnAddressIsTaken(true);
3286 EVT VT = Op.getValueType();
3287 DebugLoc dl = Op.getDebugLoc();
3288 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3290 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3291 SDValue Offset = DAG.getConstant(4, MVT::i32);
3292 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3293 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3294 MachinePointerInfo(), false, false, false, 0);
3297 // Return LR, which contains the return address. Mark it an implicit live-in.
3298 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3299 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3302 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3303 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3304 MFI->setFrameAddressIsTaken(true);
3306 EVT VT = Op.getValueType();
3307 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3308 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3309 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3310 ? ARM::R7 : ARM::R11;
3311 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3313 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3314 MachinePointerInfo(),
3315 false, false, false, 0);
3319 /// ExpandBITCAST - If the target supports VFP, this function is called to
3320 /// expand a bit convert where either the source or destination type is i64 to
3321 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3322 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3323 /// vectors), since the legalizer won't know what to do with that.
3324 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3325 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3326 DebugLoc dl = N->getDebugLoc();
3327 SDValue Op = N->getOperand(0);
3329 // This function is only supposed to be called for i64 types, either as the
3330 // source or destination of the bit convert.
3331 EVT SrcVT = Op.getValueType();
3332 EVT DstVT = N->getValueType(0);
3333 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3334 "ExpandBITCAST called for non-i64 type");
3336 // Turn i64->f64 into VMOVDRR.
3337 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3338 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3339 DAG.getConstant(0, MVT::i32));
3340 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3341 DAG.getConstant(1, MVT::i32));
3342 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3343 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3346 // Turn f64->i64 into VMOVRRD.
3347 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3348 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3349 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3350 // Merge the pieces into a single i64 value.
3351 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3357 /// getZeroVector - Returns a vector of specified type with all zero elements.
3358 /// Zero vectors are used to represent vector negation and in those cases
3359 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3360 /// not support i64 elements, so sometimes the zero vectors will need to be
3361 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3363 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3364 assert(VT.isVector() && "Expected a vector type");
3365 // The canonical modified immediate encoding of a zero vector is....0!
3366 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3367 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3368 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3369 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3372 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3373 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3374 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3375 SelectionDAG &DAG) const {
3376 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3377 EVT VT = Op.getValueType();
3378 unsigned VTBits = VT.getSizeInBits();
3379 DebugLoc dl = Op.getDebugLoc();
3380 SDValue ShOpLo = Op.getOperand(0);
3381 SDValue ShOpHi = Op.getOperand(1);
3382 SDValue ShAmt = Op.getOperand(2);
3384 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3386 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3388 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3389 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3390 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3391 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3392 DAG.getConstant(VTBits, MVT::i32));
3393 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3394 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3395 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3397 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3398 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3400 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3401 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3404 SDValue Ops[2] = { Lo, Hi };
3405 return DAG.getMergeValues(Ops, 2, dl);
3408 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3409 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3410 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3411 SelectionDAG &DAG) const {
3412 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3413 EVT VT = Op.getValueType();
3414 unsigned VTBits = VT.getSizeInBits();
3415 DebugLoc dl = Op.getDebugLoc();
3416 SDValue ShOpLo = Op.getOperand(0);
3417 SDValue ShOpHi = Op.getOperand(1);
3418 SDValue ShAmt = Op.getOperand(2);
3421 assert(Op.getOpcode() == ISD::SHL_PARTS);
3422 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3423 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3424 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3425 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3426 DAG.getConstant(VTBits, MVT::i32));
3427 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3428 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3430 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3431 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3432 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3434 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3435 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3438 SDValue Ops[2] = { Lo, Hi };
3439 return DAG.getMergeValues(Ops, 2, dl);
3442 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3443 SelectionDAG &DAG) const {
3444 // The rounding mode is in bits 23:22 of the FPSCR.
3445 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3446 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3447 // so that the shift + and get folded into a bitfield extract.
3448 DebugLoc dl = Op.getDebugLoc();
3449 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3450 DAG.getConstant(Intrinsic::arm_get_fpscr,
3452 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3453 DAG.getConstant(1U << 22, MVT::i32));
3454 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3455 DAG.getConstant(22, MVT::i32));
3456 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3457 DAG.getConstant(3, MVT::i32));
3460 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3461 const ARMSubtarget *ST) {
3462 EVT VT = N->getValueType(0);
3463 DebugLoc dl = N->getDebugLoc();
3465 if (!ST->hasV6T2Ops())
3468 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3469 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3472 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3473 const ARMSubtarget *ST) {
3474 EVT VT = N->getValueType(0);
3475 DebugLoc dl = N->getDebugLoc();
3480 // Lower vector shifts on NEON to use VSHL.
3481 assert(ST->hasNEON() && "unexpected vector shift");
3483 // Left shifts translate directly to the vshiftu intrinsic.
3484 if (N->getOpcode() == ISD::SHL)
3485 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3486 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3487 N->getOperand(0), N->getOperand(1));
3489 assert((N->getOpcode() == ISD::SRA ||
3490 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3492 // NEON uses the same intrinsics for both left and right shifts. For
3493 // right shifts, the shift amounts are negative, so negate the vector of
3495 EVT ShiftVT = N->getOperand(1).getValueType();
3496 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3497 getZeroVector(ShiftVT, DAG, dl),
3499 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3500 Intrinsic::arm_neon_vshifts :
3501 Intrinsic::arm_neon_vshiftu);
3502 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3503 DAG.getConstant(vshiftInt, MVT::i32),
3504 N->getOperand(0), NegatedCount);
3507 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3508 const ARMSubtarget *ST) {
3509 EVT VT = N->getValueType(0);
3510 DebugLoc dl = N->getDebugLoc();
3512 // We can get here for a node like i32 = ISD::SHL i32, i64
3516 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3517 "Unknown shift to lower!");
3519 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3520 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3521 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3524 // If we are in thumb mode, we don't have RRX.
3525 if (ST->isThumb1Only()) return SDValue();
3527 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3528 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3529 DAG.getConstant(0, MVT::i32));
3530 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3531 DAG.getConstant(1, MVT::i32));
3533 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3534 // captures the result into a carry flag.
3535 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3536 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3538 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3539 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3541 // Merge the pieces into a single i64 value.
3542 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3545 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3546 SDValue TmpOp0, TmpOp1;
3547 bool Invert = false;
3551 SDValue Op0 = Op.getOperand(0);
3552 SDValue Op1 = Op.getOperand(1);
3553 SDValue CC = Op.getOperand(2);
3554 EVT VT = Op.getValueType();
3555 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3556 DebugLoc dl = Op.getDebugLoc();
3558 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3559 switch (SetCCOpcode) {
3560 default: llvm_unreachable("Illegal FP comparison");
3562 case ISD::SETNE: Invert = true; // Fallthrough
3564 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3566 case ISD::SETLT: Swap = true; // Fallthrough
3568 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3570 case ISD::SETLE: Swap = true; // Fallthrough
3572 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3573 case ISD::SETUGE: Swap = true; // Fallthrough
3574 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3575 case ISD::SETUGT: Swap = true; // Fallthrough
3576 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3577 case ISD::SETUEQ: Invert = true; // Fallthrough
3579 // Expand this to (OLT | OGT).
3583 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3584 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3586 case ISD::SETUO: Invert = true; // Fallthrough
3588 // Expand this to (OLT | OGE).
3592 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3593 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3597 // Integer comparisons.
3598 switch (SetCCOpcode) {
3599 default: llvm_unreachable("Illegal integer comparison");
3600 case ISD::SETNE: Invert = true;
3601 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3602 case ISD::SETLT: Swap = true;
3603 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3604 case ISD::SETLE: Swap = true;
3605 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3606 case ISD::SETULT: Swap = true;
3607 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3608 case ISD::SETULE: Swap = true;
3609 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3612 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3613 if (Opc == ARMISD::VCEQ) {
3616 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3618 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3621 // Ignore bitconvert.
3622 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3623 AndOp = AndOp.getOperand(0);
3625 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3627 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3628 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3635 std::swap(Op0, Op1);
3637 // If one of the operands is a constant vector zero, attempt to fold the
3638 // comparison to a specialized compare-against-zero form.
3640 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3642 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3643 if (Opc == ARMISD::VCGE)
3644 Opc = ARMISD::VCLEZ;
3645 else if (Opc == ARMISD::VCGT)
3646 Opc = ARMISD::VCLTZ;
3651 if (SingleOp.getNode()) {
3654 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3656 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3658 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3660 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3662 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3664 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3667 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3671 Result = DAG.getNOT(dl, Result, VT);
3676 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3677 const ARMSubtarget *ST) const {
3678 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3681 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3682 assert(Op.getValueType() == MVT::f32 &&
3683 "ConstantFP custom lowering should only occur for f32.");
3685 APFloat FPVal = CFP->getValueAPF();
3686 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3690 DebugLoc DL = Op.getDebugLoc();
3691 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3692 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, NewVal);
3693 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3694 DAG.getConstant(0, MVT::i32));
3697 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3698 /// valid vector constant for a NEON instruction with a "modified immediate"
3699 /// operand (e.g., VMOV). If so, return the encoded value.
3700 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3701 unsigned SplatBitSize, SelectionDAG &DAG,
3702 EVT &VT, bool is128Bits, NEONModImmType type) {
3703 unsigned OpCmode, Imm;
3705 // SplatBitSize is set to the smallest size that splats the vector, so a
3706 // zero vector will always have SplatBitSize == 8. However, NEON modified
3707 // immediate instructions others than VMOV do not support the 8-bit encoding
3708 // of a zero vector, and the default encoding of zero is supposed to be the
3713 switch (SplatBitSize) {
3715 if (type != VMOVModImm)
3717 // Any 1-byte value is OK. Op=0, Cmode=1110.
3718 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3721 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3725 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3726 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3727 if ((SplatBits & ~0xff) == 0) {
3728 // Value = 0x00nn: Op=x, Cmode=100x.
3733 if ((SplatBits & ~0xff00) == 0) {
3734 // Value = 0xnn00: Op=x, Cmode=101x.
3736 Imm = SplatBits >> 8;
3742 // NEON's 32-bit VMOV supports splat values where:
3743 // * only one byte is nonzero, or
3744 // * the least significant byte is 0xff and the second byte is nonzero, or
3745 // * the least significant 2 bytes are 0xff and the third is nonzero.
3746 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3747 if ((SplatBits & ~0xff) == 0) {
3748 // Value = 0x000000nn: Op=x, Cmode=000x.
3753 if ((SplatBits & ~0xff00) == 0) {
3754 // Value = 0x0000nn00: Op=x, Cmode=001x.
3756 Imm = SplatBits >> 8;
3759 if ((SplatBits & ~0xff0000) == 0) {
3760 // Value = 0x00nn0000: Op=x, Cmode=010x.
3762 Imm = SplatBits >> 16;
3765 if ((SplatBits & ~0xff000000) == 0) {
3766 // Value = 0xnn000000: Op=x, Cmode=011x.
3768 Imm = SplatBits >> 24;
3772 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3773 if (type == OtherModImm) return SDValue();
3775 if ((SplatBits & ~0xffff) == 0 &&
3776 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3777 // Value = 0x0000nnff: Op=x, Cmode=1100.
3779 Imm = SplatBits >> 8;
3784 if ((SplatBits & ~0xffffff) == 0 &&
3785 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3786 // Value = 0x00nnffff: Op=x, Cmode=1101.
3788 Imm = SplatBits >> 16;
3789 SplatBits |= 0xffff;
3793 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3794 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3795 // VMOV.I32. A (very) minor optimization would be to replicate the value
3796 // and fall through here to test for a valid 64-bit splat. But, then the
3797 // caller would also need to check and handle the change in size.
3801 if (type != VMOVModImm)
3803 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3804 uint64_t BitMask = 0xff;
3806 unsigned ImmMask = 1;
3808 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3809 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3812 } else if ((SplatBits & BitMask) != 0) {
3818 // Op=1, Cmode=1110.
3821 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3826 llvm_unreachable("unexpected size for isNEONModifiedImm");
3829 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3830 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3833 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3834 bool &ReverseVEXT, unsigned &Imm) {
3835 unsigned NumElts = VT.getVectorNumElements();
3836 ReverseVEXT = false;
3838 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3844 // If this is a VEXT shuffle, the immediate value is the index of the first
3845 // element. The other shuffle indices must be the successive elements after
3847 unsigned ExpectedElt = Imm;
3848 for (unsigned i = 1; i < NumElts; ++i) {
3849 // Increment the expected index. If it wraps around, it may still be
3850 // a VEXT but the source vectors must be swapped.
3852 if (ExpectedElt == NumElts * 2) {
3857 if (M[i] < 0) continue; // ignore UNDEF indices
3858 if (ExpectedElt != static_cast<unsigned>(M[i]))
3862 // Adjust the index value if the source operands will be swapped.
3869 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3870 /// instruction with the specified blocksize. (The order of the elements
3871 /// within each block of the vector is reversed.)
3872 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3873 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3874 "Only possible block sizes for VREV are: 16, 32, 64");
3876 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3880 unsigned NumElts = VT.getVectorNumElements();
3881 unsigned BlockElts = M[0] + 1;
3882 // If the first shuffle index is UNDEF, be optimistic.
3884 BlockElts = BlockSize / EltSz;
3886 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3889 for (unsigned i = 0; i < NumElts; ++i) {
3890 if (M[i] < 0) continue; // ignore UNDEF indices
3891 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3898 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3899 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3900 // range, then 0 is placed into the resulting vector. So pretty much any mask
3901 // of 8 elements can work here.
3902 return VT == MVT::v8i8 && M.size() == 8;
3905 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3906 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3910 unsigned NumElts = VT.getVectorNumElements();
3911 WhichResult = (M[0] == 0 ? 0 : 1);
3912 for (unsigned i = 0; i < NumElts; i += 2) {
3913 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3914 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3920 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3921 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3922 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3923 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3924 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3928 unsigned NumElts = VT.getVectorNumElements();
3929 WhichResult = (M[0] == 0 ? 0 : 1);
3930 for (unsigned i = 0; i < NumElts; i += 2) {
3931 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3932 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3938 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3939 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3943 unsigned NumElts = VT.getVectorNumElements();
3944 WhichResult = (M[0] == 0 ? 0 : 1);
3945 for (unsigned i = 0; i != NumElts; ++i) {
3946 if (M[i] < 0) continue; // ignore UNDEF indices
3947 if ((unsigned) M[i] != 2 * i + WhichResult)
3951 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3952 if (VT.is64BitVector() && EltSz == 32)
3958 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3959 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3960 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3961 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3962 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3966 unsigned Half = VT.getVectorNumElements() / 2;
3967 WhichResult = (M[0] == 0 ? 0 : 1);
3968 for (unsigned j = 0; j != 2; ++j) {
3969 unsigned Idx = WhichResult;
3970 for (unsigned i = 0; i != Half; ++i) {
3971 int MIdx = M[i + j * Half];
3972 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3978 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3979 if (VT.is64BitVector() && EltSz == 32)
3985 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3986 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3990 unsigned NumElts = VT.getVectorNumElements();
3991 WhichResult = (M[0] == 0 ? 0 : 1);
3992 unsigned Idx = WhichResult * NumElts / 2;
3993 for (unsigned i = 0; i != NumElts; i += 2) {
3994 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3995 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4000 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4001 if (VT.is64BitVector() && EltSz == 32)
4007 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4008 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4009 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4010 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4011 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4015 unsigned NumElts = VT.getVectorNumElements();
4016 WhichResult = (M[0] == 0 ? 0 : 1);
4017 unsigned Idx = WhichResult * NumElts / 2;
4018 for (unsigned i = 0; i != NumElts; i += 2) {
4019 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4020 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4025 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4026 if (VT.is64BitVector() && EltSz == 32)
4032 // If N is an integer constant that can be moved into a register in one
4033 // instruction, return an SDValue of such a constant (will become a MOV
4034 // instruction). Otherwise return null.
4035 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4036 const ARMSubtarget *ST, DebugLoc dl) {
4038 if (!isa<ConstantSDNode>(N))
4040 Val = cast<ConstantSDNode>(N)->getZExtValue();
4042 if (ST->isThumb1Only()) {
4043 if (Val <= 255 || ~Val <= 255)
4044 return DAG.getConstant(Val, MVT::i32);
4046 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4047 return DAG.getConstant(Val, MVT::i32);
4052 // If this is a case we can't handle, return null and let the default
4053 // expansion code take care of it.
4054 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4055 const ARMSubtarget *ST) const {
4056 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4057 DebugLoc dl = Op.getDebugLoc();
4058 EVT VT = Op.getValueType();
4060 APInt SplatBits, SplatUndef;
4061 unsigned SplatBitSize;
4063 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4064 if (SplatBitSize <= 64) {
4065 // Check if an immediate VMOV works.
4067 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4068 SplatUndef.getZExtValue(), SplatBitSize,
4069 DAG, VmovVT, VT.is128BitVector(),
4071 if (Val.getNode()) {
4072 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4073 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4076 // Try an immediate VMVN.
4077 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4078 Val = isNEONModifiedImm(NegatedImm,
4079 SplatUndef.getZExtValue(), SplatBitSize,
4080 DAG, VmovVT, VT.is128BitVector(),
4082 if (Val.getNode()) {
4083 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4084 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4087 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4088 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4089 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4091 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4092 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4098 // Scan through the operands to see if only one value is used.
4099 unsigned NumElts = VT.getVectorNumElements();
4100 bool isOnlyLowElement = true;
4101 bool usesOnlyOneValue = true;
4102 bool isConstant = true;
4104 for (unsigned i = 0; i < NumElts; ++i) {
4105 SDValue V = Op.getOperand(i);
4106 if (V.getOpcode() == ISD::UNDEF)
4109 isOnlyLowElement = false;
4110 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4113 if (!Value.getNode())
4115 else if (V != Value)
4116 usesOnlyOneValue = false;
4119 if (!Value.getNode())
4120 return DAG.getUNDEF(VT);
4122 if (isOnlyLowElement)
4123 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4125 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4127 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4128 // i32 and try again.
4129 if (usesOnlyOneValue && EltSize <= 32) {
4131 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4132 if (VT.getVectorElementType().isFloatingPoint()) {
4133 SmallVector<SDValue, 8> Ops;
4134 for (unsigned i = 0; i < NumElts; ++i)
4135 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4137 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4138 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4139 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4141 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4143 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4145 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4148 // If all elements are constants and the case above didn't get hit, fall back
4149 // to the default expansion, which will generate a load from the constant
4154 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4156 SDValue shuffle = ReconstructShuffle(Op, DAG);
4157 if (shuffle != SDValue())
4161 // Vectors with 32- or 64-bit elements can be built by directly assigning
4162 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4163 // will be legalized.
4164 if (EltSize >= 32) {
4165 // Do the expansion with floating-point types, since that is what the VFP
4166 // registers are defined to use, and since i64 is not legal.
4167 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4168 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4169 SmallVector<SDValue, 8> Ops;
4170 for (unsigned i = 0; i < NumElts; ++i)
4171 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4172 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4173 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4179 // Gather data to see if the operation can be modelled as a
4180 // shuffle in combination with VEXTs.
4181 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4182 SelectionDAG &DAG) const {
4183 DebugLoc dl = Op.getDebugLoc();
4184 EVT VT = Op.getValueType();
4185 unsigned NumElts = VT.getVectorNumElements();
4187 SmallVector<SDValue, 2> SourceVecs;
4188 SmallVector<unsigned, 2> MinElts;
4189 SmallVector<unsigned, 2> MaxElts;
4191 for (unsigned i = 0; i < NumElts; ++i) {
4192 SDValue V = Op.getOperand(i);
4193 if (V.getOpcode() == ISD::UNDEF)
4195 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4196 // A shuffle can only come from building a vector from various
4197 // elements of other vectors.
4199 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4200 VT.getVectorElementType()) {
4201 // This code doesn't know how to handle shuffles where the vector
4202 // element types do not match (this happens because type legalization
4203 // promotes the return type of EXTRACT_VECTOR_ELT).
4204 // FIXME: It might be appropriate to extend this code to handle
4205 // mismatched types.
4209 // Record this extraction against the appropriate vector if possible...
4210 SDValue SourceVec = V.getOperand(0);
4211 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4212 bool FoundSource = false;
4213 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4214 if (SourceVecs[j] == SourceVec) {
4215 if (MinElts[j] > EltNo)
4217 if (MaxElts[j] < EltNo)
4224 // Or record a new source if not...
4226 SourceVecs.push_back(SourceVec);
4227 MinElts.push_back(EltNo);
4228 MaxElts.push_back(EltNo);
4232 // Currently only do something sane when at most two source vectors
4234 if (SourceVecs.size() > 2)
4237 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4238 int VEXTOffsets[2] = {0, 0};
4240 // This loop extracts the usage patterns of the source vectors
4241 // and prepares appropriate SDValues for a shuffle if possible.
4242 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4243 if (SourceVecs[i].getValueType() == VT) {
4244 // No VEXT necessary
4245 ShuffleSrcs[i] = SourceVecs[i];
4248 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4249 // It probably isn't worth padding out a smaller vector just to
4250 // break it down again in a shuffle.
4254 // Since only 64-bit and 128-bit vectors are legal on ARM and
4255 // we've eliminated the other cases...
4256 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4257 "unexpected vector sizes in ReconstructShuffle");
4259 if (MaxElts[i] - MinElts[i] >= NumElts) {
4260 // Span too large for a VEXT to cope
4264 if (MinElts[i] >= NumElts) {
4265 // The extraction can just take the second half
4266 VEXTOffsets[i] = NumElts;
4267 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4269 DAG.getIntPtrConstant(NumElts));
4270 } else if (MaxElts[i] < NumElts) {
4271 // The extraction can just take the first half
4273 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4275 DAG.getIntPtrConstant(0));
4277 // An actual VEXT is needed
4278 VEXTOffsets[i] = MinElts[i];
4279 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4281 DAG.getIntPtrConstant(0));
4282 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4284 DAG.getIntPtrConstant(NumElts));
4285 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4286 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4290 SmallVector<int, 8> Mask;
4292 for (unsigned i = 0; i < NumElts; ++i) {
4293 SDValue Entry = Op.getOperand(i);
4294 if (Entry.getOpcode() == ISD::UNDEF) {
4299 SDValue ExtractVec = Entry.getOperand(0);
4300 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4301 .getOperand(1))->getSExtValue();
4302 if (ExtractVec == SourceVecs[0]) {
4303 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4305 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4309 // Final check before we try to produce nonsense...
4310 if (isShuffleMaskLegal(Mask, VT))
4311 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4317 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4318 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4319 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4320 /// are assumed to be legal.
4322 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4324 if (VT.getVectorNumElements() == 4 &&
4325 (VT.is128BitVector() || VT.is64BitVector())) {
4326 unsigned PFIndexes[4];
4327 for (unsigned i = 0; i != 4; ++i) {
4331 PFIndexes[i] = M[i];
4334 // Compute the index in the perfect shuffle table.
4335 unsigned PFTableIndex =
4336 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4337 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4338 unsigned Cost = (PFEntry >> 30);
4345 unsigned Imm, WhichResult;
4347 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4348 return (EltSize >= 32 ||
4349 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4350 isVREVMask(M, VT, 64) ||
4351 isVREVMask(M, VT, 32) ||
4352 isVREVMask(M, VT, 16) ||
4353 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4354 isVTBLMask(M, VT) ||
4355 isVTRNMask(M, VT, WhichResult) ||
4356 isVUZPMask(M, VT, WhichResult) ||
4357 isVZIPMask(M, VT, WhichResult) ||
4358 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4359 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4360 isVZIP_v_undef_Mask(M, VT, WhichResult));
4363 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4364 /// the specified operations to build the shuffle.
4365 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4366 SDValue RHS, SelectionDAG &DAG,
4368 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4369 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4370 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4373 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4382 OP_VUZPL, // VUZP, left result
4383 OP_VUZPR, // VUZP, right result
4384 OP_VZIPL, // VZIP, left result
4385 OP_VZIPR, // VZIP, right result
4386 OP_VTRNL, // VTRN, left result
4387 OP_VTRNR // VTRN, right result
4390 if (OpNum == OP_COPY) {
4391 if (LHSID == (1*9+2)*9+3) return LHS;
4392 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4396 SDValue OpLHS, OpRHS;
4397 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4398 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4399 EVT VT = OpLHS.getValueType();
4402 default: llvm_unreachable("Unknown shuffle opcode!");
4404 // VREV divides the vector in half and swaps within the half.
4405 if (VT.getVectorElementType() == MVT::i32 ||
4406 VT.getVectorElementType() == MVT::f32)
4407 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4408 // vrev <4 x i16> -> VREV32
4409 if (VT.getVectorElementType() == MVT::i16)
4410 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4411 // vrev <4 x i8> -> VREV16
4412 assert(VT.getVectorElementType() == MVT::i8);
4413 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4418 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4419 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4423 return DAG.getNode(ARMISD::VEXT, dl, VT,
4425 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4428 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4429 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4432 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4433 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4436 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4437 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4441 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4442 ArrayRef<int> ShuffleMask,
4443 SelectionDAG &DAG) {
4444 // Check to see if we can use the VTBL instruction.
4445 SDValue V1 = Op.getOperand(0);
4446 SDValue V2 = Op.getOperand(1);
4447 DebugLoc DL = Op.getDebugLoc();
4449 SmallVector<SDValue, 8> VTBLMask;
4450 for (ArrayRef<int>::iterator
4451 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4452 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4454 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4455 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4456 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4459 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4460 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4464 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4465 SDValue V1 = Op.getOperand(0);
4466 SDValue V2 = Op.getOperand(1);
4467 DebugLoc dl = Op.getDebugLoc();
4468 EVT VT = Op.getValueType();
4469 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4471 // Convert shuffles that are directly supported on NEON to target-specific
4472 // DAG nodes, instead of keeping them as shuffles and matching them again
4473 // during code selection. This is more efficient and avoids the possibility
4474 // of inconsistencies between legalization and selection.
4475 // FIXME: floating-point vectors should be canonicalized to integer vectors
4476 // of the same time so that they get CSEd properly.
4477 ArrayRef<int> ShuffleMask = SVN->getMask();
4479 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4480 if (EltSize <= 32) {
4481 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4482 int Lane = SVN->getSplatIndex();
4483 // If this is undef splat, generate it via "just" vdup, if possible.
4484 if (Lane == -1) Lane = 0;
4486 // Test if V1 is a SCALAR_TO_VECTOR.
4487 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4488 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4490 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4491 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4493 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4494 !isa<ConstantSDNode>(V1.getOperand(0))) {
4495 bool IsScalarToVector = true;
4496 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4497 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4498 IsScalarToVector = false;
4501 if (IsScalarToVector)
4502 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4504 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4505 DAG.getConstant(Lane, MVT::i32));
4510 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4513 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4514 DAG.getConstant(Imm, MVT::i32));
4517 if (isVREVMask(ShuffleMask, VT, 64))
4518 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4519 if (isVREVMask(ShuffleMask, VT, 32))
4520 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4521 if (isVREVMask(ShuffleMask, VT, 16))
4522 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4524 // Check for Neon shuffles that modify both input vectors in place.
4525 // If both results are used, i.e., if there are two shuffles with the same
4526 // source operands and with masks corresponding to both results of one of
4527 // these operations, DAG memoization will ensure that a single node is
4528 // used for both shuffles.
4529 unsigned WhichResult;
4530 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4531 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4532 V1, V2).getValue(WhichResult);
4533 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4534 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4535 V1, V2).getValue(WhichResult);
4536 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4537 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4538 V1, V2).getValue(WhichResult);
4540 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4541 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4542 V1, V1).getValue(WhichResult);
4543 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4544 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4545 V1, V1).getValue(WhichResult);
4546 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4547 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4548 V1, V1).getValue(WhichResult);
4551 // If the shuffle is not directly supported and it has 4 elements, use
4552 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4553 unsigned NumElts = VT.getVectorNumElements();
4555 unsigned PFIndexes[4];
4556 for (unsigned i = 0; i != 4; ++i) {
4557 if (ShuffleMask[i] < 0)
4560 PFIndexes[i] = ShuffleMask[i];
4563 // Compute the index in the perfect shuffle table.
4564 unsigned PFTableIndex =
4565 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4566 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4567 unsigned Cost = (PFEntry >> 30);
4570 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4573 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4574 if (EltSize >= 32) {
4575 // Do the expansion with floating-point types, since that is what the VFP
4576 // registers are defined to use, and since i64 is not legal.
4577 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4578 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4579 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4580 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4581 SmallVector<SDValue, 8> Ops;
4582 for (unsigned i = 0; i < NumElts; ++i) {
4583 if (ShuffleMask[i] < 0)
4584 Ops.push_back(DAG.getUNDEF(EltVT));
4586 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4587 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4588 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4591 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4592 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4595 if (VT == MVT::v8i8) {
4596 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4597 if (NewOp.getNode())
4604 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4605 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4606 SDValue Lane = Op.getOperand(2);
4607 if (!isa<ConstantSDNode>(Lane))
4613 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4614 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4615 SDValue Lane = Op.getOperand(1);
4616 if (!isa<ConstantSDNode>(Lane))
4619 SDValue Vec = Op.getOperand(0);
4620 if (Op.getValueType() == MVT::i32 &&
4621 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4622 DebugLoc dl = Op.getDebugLoc();
4623 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4629 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4630 // The only time a CONCAT_VECTORS operation can have legal types is when
4631 // two 64-bit vectors are concatenated to a 128-bit vector.
4632 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4633 "unexpected CONCAT_VECTORS");
4634 DebugLoc dl = Op.getDebugLoc();
4635 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4636 SDValue Op0 = Op.getOperand(0);
4637 SDValue Op1 = Op.getOperand(1);
4638 if (Op0.getOpcode() != ISD::UNDEF)
4639 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4640 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4641 DAG.getIntPtrConstant(0));
4642 if (Op1.getOpcode() != ISD::UNDEF)
4643 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4644 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4645 DAG.getIntPtrConstant(1));
4646 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4649 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4650 /// element has been zero/sign-extended, depending on the isSigned parameter,
4651 /// from an integer type half its size.
4652 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4654 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4655 EVT VT = N->getValueType(0);
4656 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4657 SDNode *BVN = N->getOperand(0).getNode();
4658 if (BVN->getValueType(0) != MVT::v4i32 ||
4659 BVN->getOpcode() != ISD::BUILD_VECTOR)
4661 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4662 unsigned HiElt = 1 - LoElt;
4663 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4664 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4665 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4666 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4667 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4670 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4671 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4674 if (Hi0->isNullValue() && Hi1->isNullValue())
4680 if (N->getOpcode() != ISD::BUILD_VECTOR)
4683 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4684 SDNode *Elt = N->getOperand(i).getNode();
4685 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4686 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4687 unsigned HalfSize = EltSize / 2;
4689 if (!isIntN(HalfSize, C->getSExtValue()))
4692 if (!isUIntN(HalfSize, C->getZExtValue()))
4703 /// isSignExtended - Check if a node is a vector value that is sign-extended
4704 /// or a constant BUILD_VECTOR with sign-extended elements.
4705 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4706 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4708 if (isExtendedBUILD_VECTOR(N, DAG, true))
4713 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4714 /// or a constant BUILD_VECTOR with zero-extended elements.
4715 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4716 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4718 if (isExtendedBUILD_VECTOR(N, DAG, false))
4723 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4724 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4725 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4726 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4727 return N->getOperand(0);
4728 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4729 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4730 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4731 LD->isNonTemporal(), LD->isInvariant(),
4732 LD->getAlignment());
4733 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4734 // have been legalized as a BITCAST from v4i32.
4735 if (N->getOpcode() == ISD::BITCAST) {
4736 SDNode *BVN = N->getOperand(0).getNode();
4737 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4738 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4739 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4740 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4741 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4743 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4744 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4745 EVT VT = N->getValueType(0);
4746 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4747 unsigned NumElts = VT.getVectorNumElements();
4748 MVT TruncVT = MVT::getIntegerVT(EltSize);
4749 SmallVector<SDValue, 8> Ops;
4750 for (unsigned i = 0; i != NumElts; ++i) {
4751 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4752 const APInt &CInt = C->getAPIntValue();
4753 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4755 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4756 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4759 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4760 unsigned Opcode = N->getOpcode();
4761 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4762 SDNode *N0 = N->getOperand(0).getNode();
4763 SDNode *N1 = N->getOperand(1).getNode();
4764 return N0->hasOneUse() && N1->hasOneUse() &&
4765 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4770 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4771 unsigned Opcode = N->getOpcode();
4772 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4773 SDNode *N0 = N->getOperand(0).getNode();
4774 SDNode *N1 = N->getOperand(1).getNode();
4775 return N0->hasOneUse() && N1->hasOneUse() &&
4776 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4781 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4782 // Multiplications are only custom-lowered for 128-bit vectors so that
4783 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4784 EVT VT = Op.getValueType();
4785 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4786 SDNode *N0 = Op.getOperand(0).getNode();
4787 SDNode *N1 = Op.getOperand(1).getNode();
4788 unsigned NewOpc = 0;
4790 bool isN0SExt = isSignExtended(N0, DAG);
4791 bool isN1SExt = isSignExtended(N1, DAG);
4792 if (isN0SExt && isN1SExt)
4793 NewOpc = ARMISD::VMULLs;
4795 bool isN0ZExt = isZeroExtended(N0, DAG);
4796 bool isN1ZExt = isZeroExtended(N1, DAG);
4797 if (isN0ZExt && isN1ZExt)
4798 NewOpc = ARMISD::VMULLu;
4799 else if (isN1SExt || isN1ZExt) {
4800 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4801 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4802 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4803 NewOpc = ARMISD::VMULLs;
4805 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4806 NewOpc = ARMISD::VMULLu;
4808 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4810 NewOpc = ARMISD::VMULLu;
4816 if (VT == MVT::v2i64)
4817 // Fall through to expand this. It is not legal.
4820 // Other vector multiplications are legal.
4825 // Legalize to a VMULL instruction.
4826 DebugLoc DL = Op.getDebugLoc();
4828 SDValue Op1 = SkipExtension(N1, DAG);
4830 Op0 = SkipExtension(N0, DAG);
4831 assert(Op0.getValueType().is64BitVector() &&
4832 Op1.getValueType().is64BitVector() &&
4833 "unexpected types for extended operands to VMULL");
4834 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4837 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4838 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4845 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4846 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4847 EVT Op1VT = Op1.getValueType();
4848 return DAG.getNode(N0->getOpcode(), DL, VT,
4849 DAG.getNode(NewOpc, DL, VT,
4850 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4851 DAG.getNode(NewOpc, DL, VT,
4852 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4856 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4858 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4859 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4860 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4861 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4862 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4863 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4864 // Get reciprocal estimate.
4865 // float4 recip = vrecpeq_f32(yf);
4866 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4867 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4868 // Because char has a smaller range than uchar, we can actually get away
4869 // without any newton steps. This requires that we use a weird bias
4870 // of 0xb000, however (again, this has been exhaustively tested).
4871 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4872 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4873 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4874 Y = DAG.getConstant(0xb000, MVT::i32);
4875 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4876 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4877 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4878 // Convert back to short.
4879 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4880 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4885 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4887 // Convert to float.
4888 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4889 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4890 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4891 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4892 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4893 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4895 // Use reciprocal estimate and one refinement step.
4896 // float4 recip = vrecpeq_f32(yf);
4897 // recip *= vrecpsq_f32(yf, recip);
4898 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4899 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4900 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4901 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4903 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4904 // Because short has a smaller range than ushort, we can actually get away
4905 // with only a single newton step. This requires that we use a weird bias
4906 // of 89, however (again, this has been exhaustively tested).
4907 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4908 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4909 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4910 N1 = DAG.getConstant(0x89, MVT::i32);
4911 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4912 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4913 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4914 // Convert back to integer and return.
4915 // return vmovn_s32(vcvt_s32_f32(result));
4916 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4917 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4921 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4922 EVT VT = Op.getValueType();
4923 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4924 "unexpected type for custom-lowering ISD::SDIV");
4926 DebugLoc dl = Op.getDebugLoc();
4927 SDValue N0 = Op.getOperand(0);
4928 SDValue N1 = Op.getOperand(1);
4931 if (VT == MVT::v8i8) {
4932 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4933 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4935 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4936 DAG.getIntPtrConstant(4));
4937 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4938 DAG.getIntPtrConstant(4));
4939 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4940 DAG.getIntPtrConstant(0));
4941 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4942 DAG.getIntPtrConstant(0));
4944 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4945 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4947 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4948 N0 = LowerCONCAT_VECTORS(N0, DAG);
4950 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4953 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4956 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4957 EVT VT = Op.getValueType();
4958 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4959 "unexpected type for custom-lowering ISD::UDIV");
4961 DebugLoc dl = Op.getDebugLoc();
4962 SDValue N0 = Op.getOperand(0);
4963 SDValue N1 = Op.getOperand(1);
4966 if (VT == MVT::v8i8) {
4967 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4968 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4970 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4971 DAG.getIntPtrConstant(4));
4972 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4973 DAG.getIntPtrConstant(4));
4974 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4975 DAG.getIntPtrConstant(0));
4976 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4977 DAG.getIntPtrConstant(0));
4979 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4980 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4982 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4983 N0 = LowerCONCAT_VECTORS(N0, DAG);
4985 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4986 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4991 // v4i16 sdiv ... Convert to float.
4992 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4993 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4994 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4995 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4996 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4997 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4999 // Use reciprocal estimate and two refinement steps.
5000 // float4 recip = vrecpeq_f32(yf);
5001 // recip *= vrecpsq_f32(yf, recip);
5002 // recip *= vrecpsq_f32(yf, recip);
5003 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5004 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5005 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5006 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5008 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5009 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5010 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5012 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5013 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5014 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5015 // and that it will never cause us to return an answer too large).
5016 // float4 result = as_float4(as_int4(xf*recip) + 2);
5017 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5018 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5019 N1 = DAG.getConstant(2, MVT::i32);
5020 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5021 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5022 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5023 // Convert back to integer and return.
5024 // return vmovn_u32(vcvt_s32_f32(result));
5025 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5026 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5030 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5031 EVT VT = Op.getNode()->getValueType(0);
5032 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5035 bool ExtraOp = false;
5036 switch (Op.getOpcode()) {
5037 default: llvm_unreachable("Invalid code");
5038 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5039 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5040 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5041 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5045 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5047 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5048 Op.getOperand(1), Op.getOperand(2));
5051 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5052 // Monotonic load/store is legal for all targets
5053 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5056 // Aquire/Release load/store is not legal for targets without a
5057 // dmb or equivalent available.
5063 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5064 SelectionDAG &DAG, unsigned NewOp) {
5065 DebugLoc dl = Node->getDebugLoc();
5066 assert (Node->getValueType(0) == MVT::i64 &&
5067 "Only know how to expand i64 atomics");
5069 SmallVector<SDValue, 6> Ops;
5070 Ops.push_back(Node->getOperand(0)); // Chain
5071 Ops.push_back(Node->getOperand(1)); // Ptr
5073 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5074 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5075 // High part of Val1
5076 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5077 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5078 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5079 // High part of Val1
5080 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5081 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5082 // High part of Val2
5083 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5084 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5086 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5088 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5089 cast<MemSDNode>(Node)->getMemOperand());
5090 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5091 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5092 Results.push_back(Result.getValue(2));
5095 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5096 switch (Op.getOpcode()) {
5097 default: llvm_unreachable("Don't know how to custom lower this!");
5098 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5099 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5100 case ISD::GlobalAddress:
5101 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5102 LowerGlobalAddressELF(Op, DAG);
5103 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5104 case ISD::SELECT: return LowerSELECT(Op, DAG);
5105 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5106 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5107 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5108 case ISD::VASTART: return LowerVASTART(Op, DAG);
5109 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5110 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5111 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5112 case ISD::SINT_TO_FP:
5113 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5114 case ISD::FP_TO_SINT:
5115 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5116 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5117 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5118 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5119 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5120 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5121 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5122 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5124 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5127 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5128 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5129 case ISD::SRL_PARTS:
5130 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5131 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5132 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5133 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5134 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5135 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5136 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5137 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5138 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5139 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5140 case ISD::MUL: return LowerMUL(Op, DAG);
5141 case ISD::SDIV: return LowerSDIV(Op, DAG);
5142 case ISD::UDIV: return LowerUDIV(Op, DAG);
5146 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5147 case ISD::ATOMIC_LOAD:
5148 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5152 /// ReplaceNodeResults - Replace the results of node with an illegal result
5153 /// type with new values built out of custom code.
5154 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5155 SmallVectorImpl<SDValue>&Results,
5156 SelectionDAG &DAG) const {
5158 switch (N->getOpcode()) {
5160 llvm_unreachable("Don't know how to custom expand this!");
5162 Res = ExpandBITCAST(N, DAG);
5166 Res = Expand64BitShift(N, DAG, Subtarget);
5168 case ISD::ATOMIC_LOAD_ADD:
5169 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5171 case ISD::ATOMIC_LOAD_AND:
5172 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5174 case ISD::ATOMIC_LOAD_NAND:
5175 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5177 case ISD::ATOMIC_LOAD_OR:
5178 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5180 case ISD::ATOMIC_LOAD_SUB:
5181 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5183 case ISD::ATOMIC_LOAD_XOR:
5184 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5186 case ISD::ATOMIC_SWAP:
5187 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5189 case ISD::ATOMIC_CMP_SWAP:
5190 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5194 Results.push_back(Res);
5197 //===----------------------------------------------------------------------===//
5198 // ARM Scheduler Hooks
5199 //===----------------------------------------------------------------------===//
5202 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5203 MachineBasicBlock *BB,
5204 unsigned Size) const {
5205 unsigned dest = MI->getOperand(0).getReg();
5206 unsigned ptr = MI->getOperand(1).getReg();
5207 unsigned oldval = MI->getOperand(2).getReg();
5208 unsigned newval = MI->getOperand(3).getReg();
5209 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5210 DebugLoc dl = MI->getDebugLoc();
5211 bool isThumb2 = Subtarget->isThumb2();
5213 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5215 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5216 : ARM::GPRRegisterClass);
5219 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5220 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5221 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5224 unsigned ldrOpc, strOpc;
5226 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5228 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5229 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5232 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5233 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5236 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5237 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5241 MachineFunction *MF = BB->getParent();
5242 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5243 MachineFunction::iterator It = BB;
5244 ++It; // insert the new blocks after the current block
5246 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5247 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5248 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5249 MF->insert(It, loop1MBB);
5250 MF->insert(It, loop2MBB);
5251 MF->insert(It, exitMBB);
5253 // Transfer the remainder of BB and its successor edges to exitMBB.
5254 exitMBB->splice(exitMBB->begin(), BB,
5255 llvm::next(MachineBasicBlock::iterator(MI)),
5257 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5261 // fallthrough --> loop1MBB
5262 BB->addSuccessor(loop1MBB);
5265 // ldrex dest, [ptr]
5269 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5270 if (ldrOpc == ARM::t2LDREX)
5272 AddDefaultPred(MIB);
5273 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5274 .addReg(dest).addReg(oldval));
5275 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5276 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5277 BB->addSuccessor(loop2MBB);
5278 BB->addSuccessor(exitMBB);
5281 // strex scratch, newval, [ptr]
5285 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5286 if (strOpc == ARM::t2STREX)
5288 AddDefaultPred(MIB);
5289 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5290 .addReg(scratch).addImm(0));
5291 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5292 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5293 BB->addSuccessor(loop1MBB);
5294 BB->addSuccessor(exitMBB);
5300 MI->eraseFromParent(); // The instruction is gone now.
5306 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5307 unsigned Size, unsigned BinOpcode) const {
5308 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5309 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5311 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5312 MachineFunction *MF = BB->getParent();
5313 MachineFunction::iterator It = BB;
5316 unsigned dest = MI->getOperand(0).getReg();
5317 unsigned ptr = MI->getOperand(1).getReg();
5318 unsigned incr = MI->getOperand(2).getReg();
5319 DebugLoc dl = MI->getDebugLoc();
5320 bool isThumb2 = Subtarget->isThumb2();
5322 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5324 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5325 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5328 unsigned ldrOpc, strOpc;
5330 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5332 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5333 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5336 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5337 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5340 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5341 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5345 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5346 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5347 MF->insert(It, loopMBB);
5348 MF->insert(It, exitMBB);
5350 // Transfer the remainder of BB and its successor edges to exitMBB.
5351 exitMBB->splice(exitMBB->begin(), BB,
5352 llvm::next(MachineBasicBlock::iterator(MI)),
5354 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5356 const TargetRegisterClass *TRC =
5357 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5358 unsigned scratch = MRI.createVirtualRegister(TRC);
5359 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5363 // fallthrough --> loopMBB
5364 BB->addSuccessor(loopMBB);
5368 // <binop> scratch2, dest, incr
5369 // strex scratch, scratch2, ptr
5372 // fallthrough --> exitMBB
5374 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5375 if (ldrOpc == ARM::t2LDREX)
5377 AddDefaultPred(MIB);
5379 // operand order needs to go the other way for NAND
5380 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5381 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5382 addReg(incr).addReg(dest)).addReg(0);
5384 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5385 addReg(dest).addReg(incr)).addReg(0);
5388 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5389 if (strOpc == ARM::t2STREX)
5391 AddDefaultPred(MIB);
5392 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5393 .addReg(scratch).addImm(0));
5394 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5395 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5397 BB->addSuccessor(loopMBB);
5398 BB->addSuccessor(exitMBB);
5404 MI->eraseFromParent(); // The instruction is gone now.
5410 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5411 MachineBasicBlock *BB,
5414 ARMCC::CondCodes Cond) const {
5415 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5417 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5418 MachineFunction *MF = BB->getParent();
5419 MachineFunction::iterator It = BB;
5422 unsigned dest = MI->getOperand(0).getReg();
5423 unsigned ptr = MI->getOperand(1).getReg();
5424 unsigned incr = MI->getOperand(2).getReg();
5425 unsigned oldval = dest;
5426 DebugLoc dl = MI->getDebugLoc();
5427 bool isThumb2 = Subtarget->isThumb2();
5429 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5431 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5432 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5435 unsigned ldrOpc, strOpc, extendOpc;
5437 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5439 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5440 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5441 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5444 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5445 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5446 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5449 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5450 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5455 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5456 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5457 MF->insert(It, loopMBB);
5458 MF->insert(It, exitMBB);
5460 // Transfer the remainder of BB and its successor edges to exitMBB.
5461 exitMBB->splice(exitMBB->begin(), BB,
5462 llvm::next(MachineBasicBlock::iterator(MI)),
5464 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5466 const TargetRegisterClass *TRC =
5467 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5468 unsigned scratch = MRI.createVirtualRegister(TRC);
5469 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5473 // fallthrough --> loopMBB
5474 BB->addSuccessor(loopMBB);
5478 // (sign extend dest, if required)
5480 // cmov.cond scratch2, dest, incr
5481 // strex scratch, scratch2, ptr
5484 // fallthrough --> exitMBB
5486 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5487 if (ldrOpc == ARM::t2LDREX)
5489 AddDefaultPred(MIB);
5491 // Sign extend the value, if necessary.
5492 if (signExtend && extendOpc) {
5493 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5494 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5499 // Build compare and cmov instructions.
5500 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5501 .addReg(oldval).addReg(incr));
5502 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5503 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5505 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5506 if (strOpc == ARM::t2STREX)
5508 AddDefaultPred(MIB);
5509 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5510 .addReg(scratch).addImm(0));
5511 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5512 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5514 BB->addSuccessor(loopMBB);
5515 BB->addSuccessor(exitMBB);
5521 MI->eraseFromParent(); // The instruction is gone now.
5527 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5528 unsigned Op1, unsigned Op2,
5529 bool NeedsCarry, bool IsCmpxchg) const {
5530 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5531 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5533 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5534 MachineFunction *MF = BB->getParent();
5535 MachineFunction::iterator It = BB;
5538 unsigned destlo = MI->getOperand(0).getReg();
5539 unsigned desthi = MI->getOperand(1).getReg();
5540 unsigned ptr = MI->getOperand(2).getReg();
5541 unsigned vallo = MI->getOperand(3).getReg();
5542 unsigned valhi = MI->getOperand(4).getReg();
5543 DebugLoc dl = MI->getDebugLoc();
5544 bool isThumb2 = Subtarget->isThumb2();
5546 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5548 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5549 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5550 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5553 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5554 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5556 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5557 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5559 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5560 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5562 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5563 MF->insert(It, loopMBB);
5565 MF->insert(It, contBB);
5566 MF->insert(It, cont2BB);
5568 MF->insert(It, exitMBB);
5570 // Transfer the remainder of BB and its successor edges to exitMBB.
5571 exitMBB->splice(exitMBB->begin(), BB,
5572 llvm::next(MachineBasicBlock::iterator(MI)),
5574 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5576 const TargetRegisterClass *TRC =
5577 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5578 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5582 // fallthrough --> loopMBB
5583 BB->addSuccessor(loopMBB);
5586 // ldrexd r2, r3, ptr
5587 // <binopa> r0, r2, incr
5588 // <binopb> r1, r3, incr
5589 // strexd storesuccess, r0, r1, ptr
5590 // cmp storesuccess, #0
5592 // fallthrough --> exitMBB
5594 // Note that the registers are explicitly specified because there is not any
5595 // way to force the register allocator to allocate a register pair.
5597 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5598 // need to properly enforce the restriction that the two output registers
5599 // for ldrexd must be different.
5602 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5603 .addReg(ARM::R2, RegState::Define)
5604 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5605 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5606 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5607 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5611 for (unsigned i = 0; i < 2; i++) {
5612 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5614 .addReg(i == 0 ? destlo : desthi)
5615 .addReg(i == 0 ? vallo : valhi));
5616 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5617 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5618 BB->addSuccessor(exitMBB);
5619 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5620 BB = (i == 0 ? contBB : cont2BB);
5623 // Copy to physregs for strexd
5624 unsigned setlo = MI->getOperand(5).getReg();
5625 unsigned sethi = MI->getOperand(6).getReg();
5626 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5627 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5629 // Perform binary operation
5630 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5631 .addReg(destlo).addReg(vallo))
5632 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5633 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5634 .addReg(desthi).addReg(valhi)).addReg(0);
5636 // Copy to physregs for strexd
5637 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5638 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5642 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5643 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5645 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5646 .addReg(storesuccess).addImm(0));
5647 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5648 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5650 BB->addSuccessor(loopMBB);
5651 BB->addSuccessor(exitMBB);
5657 MI->eraseFromParent(); // The instruction is gone now.
5662 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5663 /// registers the function context.
5664 void ARMTargetLowering::
5665 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5666 MachineBasicBlock *DispatchBB, int FI) const {
5667 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5668 DebugLoc dl = MI->getDebugLoc();
5669 MachineFunction *MF = MBB->getParent();
5670 MachineRegisterInfo *MRI = &MF->getRegInfo();
5671 MachineConstantPool *MCP = MF->getConstantPool();
5672 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5673 const Function *F = MF->getFunction();
5675 bool isThumb = Subtarget->isThumb();
5676 bool isThumb2 = Subtarget->isThumb2();
5678 unsigned PCLabelId = AFI->createPICLabelUId();
5679 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5680 ARMConstantPoolValue *CPV =
5681 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5682 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5684 const TargetRegisterClass *TRC =
5685 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5687 // Grab constant pool and fixed stack memory operands.
5688 MachineMemOperand *CPMMO =
5689 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5690 MachineMemOperand::MOLoad, 4, 4);
5692 MachineMemOperand *FIMMOSt =
5693 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5694 MachineMemOperand::MOStore, 4, 4);
5696 // Load the address of the dispatch MBB into the jump buffer.
5698 // Incoming value: jbuf
5699 // ldr.n r5, LCPI1_1
5702 // str r5, [$jbuf, #+4] ; &jbuf[1]
5703 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5704 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5705 .addConstantPoolIndex(CPI)
5706 .addMemOperand(CPMMO));
5707 // Set the low bit because of thumb mode.
5708 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5710 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5711 .addReg(NewVReg1, RegState::Kill)
5713 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5714 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5715 .addReg(NewVReg2, RegState::Kill)
5717 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5718 .addReg(NewVReg3, RegState::Kill)
5720 .addImm(36) // &jbuf[1] :: pc
5721 .addMemOperand(FIMMOSt));
5722 } else if (isThumb) {
5723 // Incoming value: jbuf
5724 // ldr.n r1, LCPI1_4
5728 // add r2, $jbuf, #+4 ; &jbuf[1]
5730 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5731 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5732 .addConstantPoolIndex(CPI)
5733 .addMemOperand(CPMMO));
5734 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5735 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5736 .addReg(NewVReg1, RegState::Kill)
5738 // Set the low bit because of thumb mode.
5739 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5740 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5741 .addReg(ARM::CPSR, RegState::Define)
5743 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5744 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5745 .addReg(ARM::CPSR, RegState::Define)
5746 .addReg(NewVReg2, RegState::Kill)
5747 .addReg(NewVReg3, RegState::Kill));
5748 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5749 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5751 .addImm(36)); // &jbuf[1] :: pc
5752 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5753 .addReg(NewVReg4, RegState::Kill)
5754 .addReg(NewVReg5, RegState::Kill)
5756 .addMemOperand(FIMMOSt));
5758 // Incoming value: jbuf
5761 // str r1, [$jbuf, #+4] ; &jbuf[1]
5762 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5763 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5764 .addConstantPoolIndex(CPI)
5766 .addMemOperand(CPMMO));
5767 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5768 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5769 .addReg(NewVReg1, RegState::Kill)
5770 .addImm(PCLabelId));
5771 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5772 .addReg(NewVReg2, RegState::Kill)
5774 .addImm(36) // &jbuf[1] :: pc
5775 .addMemOperand(FIMMOSt));
5779 MachineBasicBlock *ARMTargetLowering::
5780 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5781 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5782 DebugLoc dl = MI->getDebugLoc();
5783 MachineFunction *MF = MBB->getParent();
5784 MachineRegisterInfo *MRI = &MF->getRegInfo();
5785 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5786 MachineFrameInfo *MFI = MF->getFrameInfo();
5787 int FI = MFI->getFunctionContextIndex();
5789 const TargetRegisterClass *TRC =
5790 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5792 // Get a mapping of the call site numbers to all of the landing pads they're
5794 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5795 unsigned MaxCSNum = 0;
5796 MachineModuleInfo &MMI = MF->getMMI();
5797 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5798 if (!BB->isLandingPad()) continue;
5800 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5802 for (MachineBasicBlock::iterator
5803 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5804 if (!II->isEHLabel()) continue;
5806 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5807 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5809 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5810 for (SmallVectorImpl<unsigned>::iterator
5811 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5812 CSI != CSE; ++CSI) {
5813 CallSiteNumToLPad[*CSI].push_back(BB);
5814 MaxCSNum = std::max(MaxCSNum, *CSI);
5820 // Get an ordered list of the machine basic blocks for the jump table.
5821 std::vector<MachineBasicBlock*> LPadList;
5822 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5823 LPadList.reserve(CallSiteNumToLPad.size());
5824 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5825 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5826 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5827 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5828 LPadList.push_back(*II);
5829 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5833 assert(!LPadList.empty() &&
5834 "No landing pad destinations for the dispatch jump table!");
5836 // Create the jump table and associated information.
5837 MachineJumpTableInfo *JTI =
5838 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5839 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5840 unsigned UId = AFI->createJumpTableUId();
5842 // Create the MBBs for the dispatch code.
5844 // Shove the dispatch's address into the return slot in the function context.
5845 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5846 DispatchBB->setIsLandingPad();
5848 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5849 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5850 DispatchBB->addSuccessor(TrapBB);
5852 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5853 DispatchBB->addSuccessor(DispContBB);
5856 MF->insert(MF->end(), DispatchBB);
5857 MF->insert(MF->end(), DispContBB);
5858 MF->insert(MF->end(), TrapBB);
5860 // Insert code into the entry block that creates and registers the function
5862 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5864 MachineMemOperand *FIMMOLd =
5865 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5866 MachineMemOperand::MOLoad |
5867 MachineMemOperand::MOVolatile, 4, 4);
5869 if (AFI->isThumb1OnlyFunction())
5870 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5871 else if (!Subtarget->hasVFP2())
5872 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5874 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5876 unsigned NumLPads = LPadList.size();
5877 if (Subtarget->isThumb2()) {
5878 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5879 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5882 .addMemOperand(FIMMOLd));
5884 if (NumLPads < 256) {
5885 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5887 .addImm(LPadList.size()));
5889 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5890 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5891 .addImm(NumLPads & 0xFFFF));
5893 unsigned VReg2 = VReg1;
5894 if ((NumLPads & 0xFFFF0000) != 0) {
5895 VReg2 = MRI->createVirtualRegister(TRC);
5896 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5898 .addImm(NumLPads >> 16));
5901 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5906 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5911 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5912 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5913 .addJumpTableIndex(MJTI)
5916 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5919 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5920 .addReg(NewVReg3, RegState::Kill)
5922 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5924 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5925 .addReg(NewVReg4, RegState::Kill)
5927 .addJumpTableIndex(MJTI)
5929 } else if (Subtarget->isThumb()) {
5930 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5931 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5934 .addMemOperand(FIMMOLd));
5936 if (NumLPads < 256) {
5937 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5941 MachineConstantPool *ConstantPool = MF->getConstantPool();
5942 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5943 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5945 // MachineConstantPool wants an explicit alignment.
5946 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5948 Align = getTargetData()->getTypeAllocSize(C->getType());
5949 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5951 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5952 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5953 .addReg(VReg1, RegState::Define)
5954 .addConstantPoolIndex(Idx));
5955 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5960 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5965 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5966 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5967 .addReg(ARM::CPSR, RegState::Define)
5971 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5972 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5973 .addJumpTableIndex(MJTI)
5976 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5977 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5978 .addReg(ARM::CPSR, RegState::Define)
5979 .addReg(NewVReg2, RegState::Kill)
5982 MachineMemOperand *JTMMOLd =
5983 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5984 MachineMemOperand::MOLoad, 4, 4);
5986 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5987 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5988 .addReg(NewVReg4, RegState::Kill)
5990 .addMemOperand(JTMMOLd));
5992 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5993 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5994 .addReg(ARM::CPSR, RegState::Define)
5995 .addReg(NewVReg5, RegState::Kill)
5998 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5999 .addReg(NewVReg6, RegState::Kill)
6000 .addJumpTableIndex(MJTI)
6003 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6004 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6007 .addMemOperand(FIMMOLd));
6009 if (NumLPads < 256) {
6010 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6013 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6014 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6015 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6016 .addImm(NumLPads & 0xFFFF));
6018 unsigned VReg2 = VReg1;
6019 if ((NumLPads & 0xFFFF0000) != 0) {
6020 VReg2 = MRI->createVirtualRegister(TRC);
6021 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6023 .addImm(NumLPads >> 16));
6026 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6030 MachineConstantPool *ConstantPool = MF->getConstantPool();
6031 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6032 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6034 // MachineConstantPool wants an explicit alignment.
6035 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6037 Align = getTargetData()->getTypeAllocSize(C->getType());
6038 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6040 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6041 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6042 .addReg(VReg1, RegState::Define)
6043 .addConstantPoolIndex(Idx)
6045 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6047 .addReg(VReg1, RegState::Kill));
6050 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6055 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6057 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6059 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6060 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6061 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6062 .addJumpTableIndex(MJTI)
6065 MachineMemOperand *JTMMOLd =
6066 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6067 MachineMemOperand::MOLoad, 4, 4);
6068 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6070 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6071 .addReg(NewVReg3, RegState::Kill)
6074 .addMemOperand(JTMMOLd));
6076 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6077 .addReg(NewVReg5, RegState::Kill)
6079 .addJumpTableIndex(MJTI)
6083 // Add the jump table entries as successors to the MBB.
6084 MachineBasicBlock *PrevMBB = 0;
6085 for (std::vector<MachineBasicBlock*>::iterator
6086 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6087 MachineBasicBlock *CurMBB = *I;
6088 if (PrevMBB != CurMBB)
6089 DispContBB->addSuccessor(CurMBB);
6093 // N.B. the order the invoke BBs are processed in doesn't matter here.
6094 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6095 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6096 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6097 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6098 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6099 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6100 MachineBasicBlock *BB = *I;
6102 // Remove the landing pad successor from the invoke block and replace it
6103 // with the new dispatch block.
6104 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6106 while (!Successors.empty()) {
6107 MachineBasicBlock *SMBB = Successors.pop_back_val();
6108 if (SMBB->isLandingPad()) {
6109 BB->removeSuccessor(SMBB);
6110 MBBLPads.push_back(SMBB);
6114 BB->addSuccessor(DispatchBB);
6116 // Find the invoke call and mark all of the callee-saved registers as
6117 // 'implicit defined' so that they're spilled. This prevents code from
6118 // moving instructions to before the EH block, where they will never be
6120 for (MachineBasicBlock::reverse_iterator
6121 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6122 if (!II->isCall()) continue;
6124 DenseMap<unsigned, bool> DefRegs;
6125 for (MachineInstr::mop_iterator
6126 OI = II->operands_begin(), OE = II->operands_end();
6128 if (!OI->isReg()) continue;
6129 DefRegs[OI->getReg()] = true;
6132 MachineInstrBuilder MIB(&*II);
6134 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6135 unsigned Reg = SavedRegs[i];
6136 if (Subtarget->isThumb2() &&
6137 !ARM::tGPRRegisterClass->contains(Reg) &&
6138 !ARM::hGPRRegisterClass->contains(Reg))
6140 else if (Subtarget->isThumb1Only() &&
6141 !ARM::tGPRRegisterClass->contains(Reg))
6143 else if (!Subtarget->isThumb() &&
6144 !ARM::GPRRegisterClass->contains(Reg))
6147 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6154 // Mark all former landing pads as non-landing pads. The dispatch is the only
6156 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6157 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6158 (*I)->setIsLandingPad(false);
6160 // The instruction is gone now.
6161 MI->eraseFromParent();
6167 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6168 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6169 E = MBB->succ_end(); I != E; ++I)
6172 llvm_unreachable("Expecting a BB with two successors!");
6176 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6177 MachineBasicBlock *BB) const {
6178 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6179 DebugLoc dl = MI->getDebugLoc();
6180 bool isThumb2 = Subtarget->isThumb2();
6181 switch (MI->getOpcode()) {
6184 llvm_unreachable("Unexpected instr type to insert");
6186 // The Thumb2 pre-indexed stores have the same MI operands, they just
6187 // define them differently in the .td files from the isel patterns, so
6188 // they need pseudos.
6189 case ARM::t2STR_preidx:
6190 MI->setDesc(TII->get(ARM::t2STR_PRE));
6192 case ARM::t2STRB_preidx:
6193 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6195 case ARM::t2STRH_preidx:
6196 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6199 case ARM::STRi_preidx:
6200 case ARM::STRBi_preidx: {
6201 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6202 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6203 // Decode the offset.
6204 unsigned Offset = MI->getOperand(4).getImm();
6205 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6206 Offset = ARM_AM::getAM2Offset(Offset);
6210 MachineMemOperand *MMO = *MI->memoperands_begin();
6211 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6212 .addOperand(MI->getOperand(0)) // Rn_wb
6213 .addOperand(MI->getOperand(1)) // Rt
6214 .addOperand(MI->getOperand(2)) // Rn
6215 .addImm(Offset) // offset (skip GPR==zero_reg)
6216 .addOperand(MI->getOperand(5)) // pred
6217 .addOperand(MI->getOperand(6))
6218 .addMemOperand(MMO);
6219 MI->eraseFromParent();
6222 case ARM::STRr_preidx:
6223 case ARM::STRBr_preidx:
6224 case ARM::STRH_preidx: {
6226 switch (MI->getOpcode()) {
6227 default: llvm_unreachable("unexpected opcode!");
6228 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6229 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6230 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6232 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6233 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6234 MIB.addOperand(MI->getOperand(i));
6235 MI->eraseFromParent();
6238 case ARM::ATOMIC_LOAD_ADD_I8:
6239 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6240 case ARM::ATOMIC_LOAD_ADD_I16:
6241 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6242 case ARM::ATOMIC_LOAD_ADD_I32:
6243 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6245 case ARM::ATOMIC_LOAD_AND_I8:
6246 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6247 case ARM::ATOMIC_LOAD_AND_I16:
6248 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6249 case ARM::ATOMIC_LOAD_AND_I32:
6250 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6252 case ARM::ATOMIC_LOAD_OR_I8:
6253 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6254 case ARM::ATOMIC_LOAD_OR_I16:
6255 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6256 case ARM::ATOMIC_LOAD_OR_I32:
6257 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6259 case ARM::ATOMIC_LOAD_XOR_I8:
6260 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6261 case ARM::ATOMIC_LOAD_XOR_I16:
6262 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6263 case ARM::ATOMIC_LOAD_XOR_I32:
6264 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6266 case ARM::ATOMIC_LOAD_NAND_I8:
6267 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6268 case ARM::ATOMIC_LOAD_NAND_I16:
6269 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6270 case ARM::ATOMIC_LOAD_NAND_I32:
6271 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6273 case ARM::ATOMIC_LOAD_SUB_I8:
6274 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6275 case ARM::ATOMIC_LOAD_SUB_I16:
6276 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6277 case ARM::ATOMIC_LOAD_SUB_I32:
6278 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6280 case ARM::ATOMIC_LOAD_MIN_I8:
6281 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6282 case ARM::ATOMIC_LOAD_MIN_I16:
6283 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6284 case ARM::ATOMIC_LOAD_MIN_I32:
6285 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6287 case ARM::ATOMIC_LOAD_MAX_I8:
6288 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6289 case ARM::ATOMIC_LOAD_MAX_I16:
6290 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6291 case ARM::ATOMIC_LOAD_MAX_I32:
6292 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6294 case ARM::ATOMIC_LOAD_UMIN_I8:
6295 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6296 case ARM::ATOMIC_LOAD_UMIN_I16:
6297 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6298 case ARM::ATOMIC_LOAD_UMIN_I32:
6299 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6301 case ARM::ATOMIC_LOAD_UMAX_I8:
6302 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6303 case ARM::ATOMIC_LOAD_UMAX_I16:
6304 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6305 case ARM::ATOMIC_LOAD_UMAX_I32:
6306 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6308 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6309 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6310 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6312 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6313 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6314 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6317 case ARM::ATOMADD6432:
6318 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6319 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6320 /*NeedsCarry*/ true);
6321 case ARM::ATOMSUB6432:
6322 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6323 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6324 /*NeedsCarry*/ true);
6325 case ARM::ATOMOR6432:
6326 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6327 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6328 case ARM::ATOMXOR6432:
6329 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6330 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6331 case ARM::ATOMAND6432:
6332 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6333 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6334 case ARM::ATOMSWAP6432:
6335 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6336 case ARM::ATOMCMPXCHG6432:
6337 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6338 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6339 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6341 case ARM::tMOVCCr_pseudo: {
6342 // To "insert" a SELECT_CC instruction, we actually have to insert the
6343 // diamond control-flow pattern. The incoming instruction knows the
6344 // destination vreg to set, the condition code register to branch on, the
6345 // true/false values to select between, and a branch opcode to use.
6346 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6347 MachineFunction::iterator It = BB;
6353 // cmpTY ccX, r1, r2
6355 // fallthrough --> copy0MBB
6356 MachineBasicBlock *thisMBB = BB;
6357 MachineFunction *F = BB->getParent();
6358 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6359 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6360 F->insert(It, copy0MBB);
6361 F->insert(It, sinkMBB);
6363 // Transfer the remainder of BB and its successor edges to sinkMBB.
6364 sinkMBB->splice(sinkMBB->begin(), BB,
6365 llvm::next(MachineBasicBlock::iterator(MI)),
6367 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6369 BB->addSuccessor(copy0MBB);
6370 BB->addSuccessor(sinkMBB);
6372 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6373 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6376 // %FalseValue = ...
6377 // # fallthrough to sinkMBB
6380 // Update machine-CFG edges
6381 BB->addSuccessor(sinkMBB);
6384 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6387 BuildMI(*BB, BB->begin(), dl,
6388 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6389 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6390 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6392 MI->eraseFromParent(); // The pseudo instruction is gone now.
6397 case ARM::BCCZi64: {
6398 // If there is an unconditional branch to the other successor, remove it.
6399 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6401 // Compare both parts that make up the double comparison separately for
6403 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6405 unsigned LHS1 = MI->getOperand(1).getReg();
6406 unsigned LHS2 = MI->getOperand(2).getReg();
6408 AddDefaultPred(BuildMI(BB, dl,
6409 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6410 .addReg(LHS1).addImm(0));
6411 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6412 .addReg(LHS2).addImm(0)
6413 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6415 unsigned RHS1 = MI->getOperand(3).getReg();
6416 unsigned RHS2 = MI->getOperand(4).getReg();
6417 AddDefaultPred(BuildMI(BB, dl,
6418 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6419 .addReg(LHS1).addReg(RHS1));
6420 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6421 .addReg(LHS2).addReg(RHS2)
6422 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6425 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6426 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6427 if (MI->getOperand(0).getImm() == ARMCC::NE)
6428 std::swap(destMBB, exitMBB);
6430 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6431 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6433 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6435 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6437 MI->eraseFromParent(); // The pseudo instruction is gone now.
6441 case ARM::Int_eh_sjlj_setjmp:
6442 case ARM::Int_eh_sjlj_setjmp_nofp:
6443 case ARM::tInt_eh_sjlj_setjmp:
6444 case ARM::t2Int_eh_sjlj_setjmp:
6445 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6446 EmitSjLjDispatchBlock(MI, BB);
6451 // To insert an ABS instruction, we have to insert the
6452 // diamond control-flow pattern. The incoming instruction knows the
6453 // source vreg to test against 0, the destination vreg to set,
6454 // the condition code register to branch on, the
6455 // true/false values to select between, and a branch opcode to use.
6460 // BCC (branch to SinkBB if V0 >= 0)
6461 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6462 // SinkBB: V1 = PHI(V2, V3)
6463 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6464 MachineFunction::iterator BBI = BB;
6466 MachineFunction *Fn = BB->getParent();
6467 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6468 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6469 Fn->insert(BBI, RSBBB);
6470 Fn->insert(BBI, SinkBB);
6472 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6473 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6474 bool isThumb2 = Subtarget->isThumb2();
6475 MachineRegisterInfo &MRI = Fn->getRegInfo();
6476 // In Thumb mode S must not be specified if source register is the SP or
6477 // PC and if destination register is the SP, so restrict register class
6478 unsigned NewMovDstReg = MRI.createVirtualRegister(
6479 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6480 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6481 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6483 // Transfer the remainder of BB and its successor edges to sinkMBB.
6484 SinkBB->splice(SinkBB->begin(), BB,
6485 llvm::next(MachineBasicBlock::iterator(MI)),
6487 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6489 BB->addSuccessor(RSBBB);
6490 BB->addSuccessor(SinkBB);
6492 // fall through to SinkMBB
6493 RSBBB->addSuccessor(SinkBB);
6495 // insert a movs at the end of BB
6496 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6498 .addReg(ABSSrcReg, RegState::Kill)
6499 .addImm((unsigned)ARMCC::AL).addReg(0)
6500 .addReg(ARM::CPSR, RegState::Define);
6502 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6504 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6505 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6507 // insert rsbri in RSBBB
6508 // Note: BCC and rsbri will be converted into predicated rsbmi
6509 // by if-conversion pass
6510 BuildMI(*RSBBB, RSBBB->begin(), dl,
6511 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6512 .addReg(NewMovDstReg, RegState::Kill)
6513 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6515 // insert PHI in SinkBB,
6516 // reuse ABSDstReg to not change uses of ABS instruction
6517 BuildMI(*SinkBB, SinkBB->begin(), dl,
6518 TII->get(ARM::PHI), ABSDstReg)
6519 .addReg(NewRsbDstReg).addMBB(RSBBB)
6520 .addReg(NewMovDstReg).addMBB(BB);
6522 // remove ABS instruction
6523 MI->eraseFromParent();
6525 // return last added BB
6531 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6532 SDNode *Node) const {
6533 if (!MI->hasPostISelHook()) {
6534 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6535 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6539 const MCInstrDesc *MCID = &MI->getDesc();
6540 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6541 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6542 // operand is still set to noreg. If needed, set the optional operand's
6543 // register to CPSR, and remove the redundant implicit def.
6545 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6547 // Rename pseudo opcodes.
6548 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6550 const ARMBaseInstrInfo *TII =
6551 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6552 MCID = &TII->get(NewOpc);
6554 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6555 "converted opcode should be the same except for cc_out");
6559 // Add the optional cc_out operand
6560 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6562 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6564 // Any ARM instruction that sets the 's' bit should specify an optional
6565 // "cc_out" operand in the last operand position.
6566 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6567 assert(!NewOpc && "Optional cc_out operand required");
6570 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6571 // since we already have an optional CPSR def.
6572 bool definesCPSR = false;
6573 bool deadCPSR = false;
6574 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6576 const MachineOperand &MO = MI->getOperand(i);
6577 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6581 MI->RemoveOperand(i);
6586 assert(!NewOpc && "Optional cc_out operand required");
6589 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6591 assert(!MI->getOperand(ccOutIdx).getReg() &&
6592 "expect uninitialized optional cc_out operand");
6596 // If this instruction was defined with an optional CPSR def and its dag node
6597 // had a live implicit CPSR def, then activate the optional CPSR def.
6598 MachineOperand &MO = MI->getOperand(ccOutIdx);
6599 MO.setReg(ARM::CPSR);
6603 //===----------------------------------------------------------------------===//
6604 // ARM Optimization Hooks
6605 //===----------------------------------------------------------------------===//
6608 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6609 TargetLowering::DAGCombinerInfo &DCI) {
6610 SelectionDAG &DAG = DCI.DAG;
6611 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6612 EVT VT = N->getValueType(0);
6613 unsigned Opc = N->getOpcode();
6614 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6615 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6616 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6617 ISD::CondCode CC = ISD::SETCC_INVALID;
6620 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6622 SDValue CCOp = Slct.getOperand(0);
6623 if (CCOp.getOpcode() == ISD::SETCC)
6624 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6627 bool DoXform = false;
6629 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6632 if (LHS.getOpcode() == ISD::Constant &&
6633 cast<ConstantSDNode>(LHS)->isNullValue()) {
6635 } else if (CC != ISD::SETCC_INVALID &&
6636 RHS.getOpcode() == ISD::Constant &&
6637 cast<ConstantSDNode>(RHS)->isNullValue()) {
6638 std::swap(LHS, RHS);
6639 SDValue Op0 = Slct.getOperand(0);
6640 EVT OpVT = isSlctCC ? Op0.getValueType() :
6641 Op0.getOperand(0).getValueType();
6642 bool isInt = OpVT.isInteger();
6643 CC = ISD::getSetCCInverse(CC, isInt);
6645 if (!TLI.isCondCodeLegal(CC, OpVT))
6646 return SDValue(); // Inverse operator isn't legal.
6653 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6655 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6656 Slct.getOperand(0), Slct.getOperand(1), CC);
6657 SDValue CCOp = Slct.getOperand(0);
6659 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6660 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6661 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6662 CCOp, OtherOp, Result);
6667 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6668 // (only after legalization).
6669 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6670 TargetLowering::DAGCombinerInfo &DCI,
6671 const ARMSubtarget *Subtarget) {
6673 // Only perform optimization if after legalize, and if NEON is available. We
6674 // also expected both operands to be BUILD_VECTORs.
6675 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6676 || N0.getOpcode() != ISD::BUILD_VECTOR
6677 || N1.getOpcode() != ISD::BUILD_VECTOR)
6680 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6681 EVT VT = N->getValueType(0);
6682 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6685 // Check that the vector operands are of the right form.
6686 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6687 // operands, where N is the size of the formed vector.
6688 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6689 // index such that we have a pair wise add pattern.
6691 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6692 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6694 SDValue Vec = N0->getOperand(0)->getOperand(0);
6695 SDNode *V = Vec.getNode();
6696 unsigned nextIndex = 0;
6698 // For each operands to the ADD which are BUILD_VECTORs,
6699 // check to see if each of their operands are an EXTRACT_VECTOR with
6700 // the same vector and appropriate index.
6701 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6702 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6703 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6705 SDValue ExtVec0 = N0->getOperand(i);
6706 SDValue ExtVec1 = N1->getOperand(i);
6708 // First operand is the vector, verify its the same.
6709 if (V != ExtVec0->getOperand(0).getNode() ||
6710 V != ExtVec1->getOperand(0).getNode())
6713 // Second is the constant, verify its correct.
6714 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6715 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6717 // For the constant, we want to see all the even or all the odd.
6718 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6719 || C1->getZExtValue() != nextIndex+1)
6728 // Create VPADDL node.
6729 SelectionDAG &DAG = DCI.DAG;
6730 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6732 // Build operand list.
6733 SmallVector<SDValue, 8> Ops;
6734 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6735 TLI.getPointerTy()));
6737 // Input is the vector.
6740 // Get widened type and narrowed type.
6742 unsigned numElem = VT.getVectorNumElements();
6743 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6744 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6745 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6746 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6748 llvm_unreachable("Invalid vector element type for padd optimization.");
6751 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6752 widenType, &Ops[0], Ops.size());
6753 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6756 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6757 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6758 /// called with the default operands, and if that fails, with commuted
6760 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6761 TargetLowering::DAGCombinerInfo &DCI,
6762 const ARMSubtarget *Subtarget){
6764 // Attempt to create vpaddl for this add.
6765 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6766 if (Result.getNode())
6769 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6770 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6771 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6772 if (Result.getNode()) return Result;
6777 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6779 static SDValue PerformADDCombine(SDNode *N,
6780 TargetLowering::DAGCombinerInfo &DCI,
6781 const ARMSubtarget *Subtarget) {
6782 SDValue N0 = N->getOperand(0);
6783 SDValue N1 = N->getOperand(1);
6785 // First try with the default operand order.
6786 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6787 if (Result.getNode())
6790 // If that didn't work, try again with the operands commuted.
6791 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6794 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6796 static SDValue PerformSUBCombine(SDNode *N,
6797 TargetLowering::DAGCombinerInfo &DCI) {
6798 SDValue N0 = N->getOperand(0);
6799 SDValue N1 = N->getOperand(1);
6801 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6802 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6803 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6804 if (Result.getNode()) return Result;
6810 /// PerformVMULCombine
6811 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6812 /// special multiplier accumulator forwarding.
6818 static SDValue PerformVMULCombine(SDNode *N,
6819 TargetLowering::DAGCombinerInfo &DCI,
6820 const ARMSubtarget *Subtarget) {
6821 if (!Subtarget->hasVMLxForwarding())
6824 SelectionDAG &DAG = DCI.DAG;
6825 SDValue N0 = N->getOperand(0);
6826 SDValue N1 = N->getOperand(1);
6827 unsigned Opcode = N0.getOpcode();
6828 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6829 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6830 Opcode = N1.getOpcode();
6831 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6832 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6837 EVT VT = N->getValueType(0);
6838 DebugLoc DL = N->getDebugLoc();
6839 SDValue N00 = N0->getOperand(0);
6840 SDValue N01 = N0->getOperand(1);
6841 return DAG.getNode(Opcode, DL, VT,
6842 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6843 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6846 static SDValue PerformMULCombine(SDNode *N,
6847 TargetLowering::DAGCombinerInfo &DCI,
6848 const ARMSubtarget *Subtarget) {
6849 SelectionDAG &DAG = DCI.DAG;
6851 if (Subtarget->isThumb1Only())
6854 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6857 EVT VT = N->getValueType(0);
6858 if (VT.is64BitVector() || VT.is128BitVector())
6859 return PerformVMULCombine(N, DCI, Subtarget);
6863 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6867 int64_t MulAmt = C->getSExtValue();
6868 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6870 ShiftAmt = ShiftAmt & (32 - 1);
6871 SDValue V = N->getOperand(0);
6872 DebugLoc DL = N->getDebugLoc();
6875 MulAmt >>= ShiftAmt;
6878 if (isPowerOf2_32(MulAmt - 1)) {
6879 // (mul x, 2^N + 1) => (add (shl x, N), x)
6880 Res = DAG.getNode(ISD::ADD, DL, VT,
6882 DAG.getNode(ISD::SHL, DL, VT,
6884 DAG.getConstant(Log2_32(MulAmt - 1),
6886 } else if (isPowerOf2_32(MulAmt + 1)) {
6887 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6888 Res = DAG.getNode(ISD::SUB, DL, VT,
6889 DAG.getNode(ISD::SHL, DL, VT,
6891 DAG.getConstant(Log2_32(MulAmt + 1),
6897 uint64_t MulAmtAbs = -MulAmt;
6898 if (isPowerOf2_32(MulAmtAbs + 1)) {
6899 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6900 Res = DAG.getNode(ISD::SUB, DL, VT,
6902 DAG.getNode(ISD::SHL, DL, VT,
6904 DAG.getConstant(Log2_32(MulAmtAbs + 1),
6906 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
6907 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6908 Res = DAG.getNode(ISD::ADD, DL, VT,
6910 DAG.getNode(ISD::SHL, DL, VT,
6912 DAG.getConstant(Log2_32(MulAmtAbs-1),
6914 Res = DAG.getNode(ISD::SUB, DL, VT,
6915 DAG.getConstant(0, MVT::i32),Res);
6922 Res = DAG.getNode(ISD::SHL, DL, VT,
6923 Res, DAG.getConstant(ShiftAmt, MVT::i32));
6925 // Do not add new nodes to DAG combiner worklist.
6926 DCI.CombineTo(N, Res, false);
6930 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6931 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6934 SDValue FalseVal = N.getOperand(0);
6935 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6939 return C->isAllOnesValue();
6940 return C->isNullValue();
6943 /// formConditionalOp - Combine an operation with a conditional move operand
6944 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
6945 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
6946 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
6948 SDValue N0 = N->getOperand(0);
6949 SDValue N1 = N->getOperand(1);
6951 bool isAND = N->getOpcode() == ISD::AND;
6952 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
6953 if (!isCand && Commutable) {
6954 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
6962 switch (N->getOpcode()) {
6963 default: llvm_unreachable("Unexpected node");
6964 case ISD::AND: Opc = ARMISD::CAND; break;
6965 case ISD::OR: Opc = ARMISD::COR; break;
6966 case ISD::XOR: Opc = ARMISD::CXOR; break;
6968 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
6969 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
6973 static SDValue PerformANDCombine(SDNode *N,
6974 TargetLowering::DAGCombinerInfo &DCI,
6975 const ARMSubtarget *Subtarget) {
6977 // Attempt to use immediate-form VBIC
6978 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6979 DebugLoc dl = N->getDebugLoc();
6980 EVT VT = N->getValueType(0);
6981 SelectionDAG &DAG = DCI.DAG;
6983 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6986 APInt SplatBits, SplatUndef;
6987 unsigned SplatBitSize;
6990 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6991 if (SplatBitSize <= 64) {
6993 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6994 SplatUndef.getZExtValue(), SplatBitSize,
6995 DAG, VbicVT, VT.is128BitVector(),
6997 if (Val.getNode()) {
6999 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7000 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7001 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7006 if (!Subtarget->isThumb1Only()) {
7007 // (and x, (cmov -1, y, cond)) => (and.cond x, y)
7008 SDValue CAND = formConditionalOp(N, DAG, true);
7016 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7017 static SDValue PerformORCombine(SDNode *N,
7018 TargetLowering::DAGCombinerInfo &DCI,
7019 const ARMSubtarget *Subtarget) {
7020 // Attempt to use immediate-form VORR
7021 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7022 DebugLoc dl = N->getDebugLoc();
7023 EVT VT = N->getValueType(0);
7024 SelectionDAG &DAG = DCI.DAG;
7026 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7029 APInt SplatBits, SplatUndef;
7030 unsigned SplatBitSize;
7032 if (BVN && Subtarget->hasNEON() &&
7033 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7034 if (SplatBitSize <= 64) {
7036 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7037 SplatUndef.getZExtValue(), SplatBitSize,
7038 DAG, VorrVT, VT.is128BitVector(),
7040 if (Val.getNode()) {
7042 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7043 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7044 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7049 if (!Subtarget->isThumb1Only()) {
7050 // (or x, (cmov 0, y, cond)) => (or.cond x, y)
7051 SDValue COR = formConditionalOp(N, DAG, true);
7056 SDValue N0 = N->getOperand(0);
7057 if (N0.getOpcode() != ISD::AND)
7059 SDValue N1 = N->getOperand(1);
7061 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7062 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7063 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7065 unsigned SplatBitSize;
7068 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7070 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7071 HasAnyUndefs) && !HasAnyUndefs) {
7072 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7074 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7075 HasAnyUndefs) && !HasAnyUndefs &&
7076 SplatBits0 == ~SplatBits1) {
7077 // Canonicalize the vector type to make instruction selection simpler.
7078 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7079 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7080 N0->getOperand(1), N0->getOperand(0),
7082 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7087 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7090 // BFI is only available on V6T2+
7091 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7094 DebugLoc DL = N->getDebugLoc();
7095 // 1) or (and A, mask), val => ARMbfi A, val, mask
7096 // iff (val & mask) == val
7098 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7099 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7100 // && mask == ~mask2
7101 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7102 // && ~mask == mask2
7103 // (i.e., copy a bitfield value into another bitfield of the same width)
7108 SDValue N00 = N0.getOperand(0);
7110 // The value and the mask need to be constants so we can verify this is
7111 // actually a bitfield set. If the mask is 0xffff, we can do better
7112 // via a movt instruction, so don't use BFI in that case.
7113 SDValue MaskOp = N0.getOperand(1);
7114 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7117 unsigned Mask = MaskC->getZExtValue();
7121 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7122 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7124 unsigned Val = N1C->getZExtValue();
7125 if ((Val & ~Mask) != Val)
7128 if (ARM::isBitFieldInvertedMask(Mask)) {
7129 Val >>= CountTrailingZeros_32(~Mask);
7131 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7132 DAG.getConstant(Val, MVT::i32),
7133 DAG.getConstant(Mask, MVT::i32));
7135 // Do not add new nodes to DAG combiner worklist.
7136 DCI.CombineTo(N, Res, false);
7139 } else if (N1.getOpcode() == ISD::AND) {
7140 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7141 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7144 unsigned Mask2 = N11C->getZExtValue();
7146 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7148 if (ARM::isBitFieldInvertedMask(Mask) &&
7150 // The pack halfword instruction works better for masks that fit it,
7151 // so use that when it's available.
7152 if (Subtarget->hasT2ExtractPack() &&
7153 (Mask == 0xffff || Mask == 0xffff0000))
7156 unsigned amt = CountTrailingZeros_32(Mask2);
7157 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7158 DAG.getConstant(amt, MVT::i32));
7159 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7160 DAG.getConstant(Mask, MVT::i32));
7161 // Do not add new nodes to DAG combiner worklist.
7162 DCI.CombineTo(N, Res, false);
7164 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7166 // The pack halfword instruction works better for masks that fit it,
7167 // so use that when it's available.
7168 if (Subtarget->hasT2ExtractPack() &&
7169 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7172 unsigned lsb = CountTrailingZeros_32(Mask);
7173 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7174 DAG.getConstant(lsb, MVT::i32));
7175 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7176 DAG.getConstant(Mask2, MVT::i32));
7177 // Do not add new nodes to DAG combiner worklist.
7178 DCI.CombineTo(N, Res, false);
7183 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7184 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7185 ARM::isBitFieldInvertedMask(~Mask)) {
7186 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7187 // where lsb(mask) == #shamt and masked bits of B are known zero.
7188 SDValue ShAmt = N00.getOperand(1);
7189 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7190 unsigned LSB = CountTrailingZeros_32(Mask);
7194 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7195 DAG.getConstant(~Mask, MVT::i32));
7197 // Do not add new nodes to DAG combiner worklist.
7198 DCI.CombineTo(N, Res, false);
7204 static SDValue PerformXORCombine(SDNode *N,
7205 TargetLowering::DAGCombinerInfo &DCI,
7206 const ARMSubtarget *Subtarget) {
7207 EVT VT = N->getValueType(0);
7208 SelectionDAG &DAG = DCI.DAG;
7210 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7213 if (!Subtarget->isThumb1Only()) {
7214 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7215 SDValue CXOR = formConditionalOp(N, DAG, true);
7223 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7224 /// the bits being cleared by the AND are not demanded by the BFI.
7225 static SDValue PerformBFICombine(SDNode *N,
7226 TargetLowering::DAGCombinerInfo &DCI) {
7227 SDValue N1 = N->getOperand(1);
7228 if (N1.getOpcode() == ISD::AND) {
7229 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7232 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7233 unsigned LSB = CountTrailingZeros_32(~InvMask);
7234 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7235 unsigned Mask = (1 << Width)-1;
7236 unsigned Mask2 = N11C->getZExtValue();
7237 if ((Mask & (~Mask2)) == 0)
7238 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7239 N->getOperand(0), N1.getOperand(0),
7245 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7246 /// ARMISD::VMOVRRD.
7247 static SDValue PerformVMOVRRDCombine(SDNode *N,
7248 TargetLowering::DAGCombinerInfo &DCI) {
7249 // vmovrrd(vmovdrr x, y) -> x,y
7250 SDValue InDouble = N->getOperand(0);
7251 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7252 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7254 // vmovrrd(load f64) -> (load i32), (load i32)
7255 SDNode *InNode = InDouble.getNode();
7256 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7257 InNode->getValueType(0) == MVT::f64 &&
7258 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7259 !cast<LoadSDNode>(InNode)->isVolatile()) {
7260 // TODO: Should this be done for non-FrameIndex operands?
7261 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7263 SelectionDAG &DAG = DCI.DAG;
7264 DebugLoc DL = LD->getDebugLoc();
7265 SDValue BasePtr = LD->getBasePtr();
7266 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7267 LD->getPointerInfo(), LD->isVolatile(),
7268 LD->isNonTemporal(), LD->isInvariant(),
7269 LD->getAlignment());
7271 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7272 DAG.getConstant(4, MVT::i32));
7273 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7274 LD->getPointerInfo(), LD->isVolatile(),
7275 LD->isNonTemporal(), LD->isInvariant(),
7276 std::min(4U, LD->getAlignment() / 2));
7278 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7279 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7280 DCI.RemoveFromWorklist(LD);
7288 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7289 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7290 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7291 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7292 SDValue Op0 = N->getOperand(0);
7293 SDValue Op1 = N->getOperand(1);
7294 if (Op0.getOpcode() == ISD::BITCAST)
7295 Op0 = Op0.getOperand(0);
7296 if (Op1.getOpcode() == ISD::BITCAST)
7297 Op1 = Op1.getOperand(0);
7298 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7299 Op0.getNode() == Op1.getNode() &&
7300 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7301 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7302 N->getValueType(0), Op0.getOperand(0));
7306 /// PerformSTORECombine - Target-specific dag combine xforms for
7308 static SDValue PerformSTORECombine(SDNode *N,
7309 TargetLowering::DAGCombinerInfo &DCI) {
7310 // Bitcast an i64 store extracted from a vector to f64.
7311 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7312 StoreSDNode *St = cast<StoreSDNode>(N);
7313 SDValue StVal = St->getValue();
7314 if (!ISD::isNormalStore(St) || St->isVolatile())
7317 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7318 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7319 SelectionDAG &DAG = DCI.DAG;
7320 DebugLoc DL = St->getDebugLoc();
7321 SDValue BasePtr = St->getBasePtr();
7322 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7323 StVal.getNode()->getOperand(0), BasePtr,
7324 St->getPointerInfo(), St->isVolatile(),
7325 St->isNonTemporal(), St->getAlignment());
7327 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7328 DAG.getConstant(4, MVT::i32));
7329 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7330 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7331 St->isNonTemporal(),
7332 std::min(4U, St->getAlignment() / 2));
7335 if (StVal.getValueType() != MVT::i64 ||
7336 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7339 SelectionDAG &DAG = DCI.DAG;
7340 DebugLoc dl = StVal.getDebugLoc();
7341 SDValue IntVec = StVal.getOperand(0);
7342 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7343 IntVec.getValueType().getVectorNumElements());
7344 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7345 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7346 Vec, StVal.getOperand(1));
7347 dl = N->getDebugLoc();
7348 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7349 // Make the DAGCombiner fold the bitcasts.
7350 DCI.AddToWorklist(Vec.getNode());
7351 DCI.AddToWorklist(ExtElt.getNode());
7352 DCI.AddToWorklist(V.getNode());
7353 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7354 St->getPointerInfo(), St->isVolatile(),
7355 St->isNonTemporal(), St->getAlignment(),
7359 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7360 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7361 /// i64 vector to have f64 elements, since the value can then be loaded
7362 /// directly into a VFP register.
7363 static bool hasNormalLoadOperand(SDNode *N) {
7364 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7365 for (unsigned i = 0; i < NumElts; ++i) {
7366 SDNode *Elt = N->getOperand(i).getNode();
7367 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7373 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7374 /// ISD::BUILD_VECTOR.
7375 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7376 TargetLowering::DAGCombinerInfo &DCI){
7377 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7378 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7379 // into a pair of GPRs, which is fine when the value is used as a scalar,
7380 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7381 SelectionDAG &DAG = DCI.DAG;
7382 if (N->getNumOperands() == 2) {
7383 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7388 // Load i64 elements as f64 values so that type legalization does not split
7389 // them up into i32 values.
7390 EVT VT = N->getValueType(0);
7391 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7393 DebugLoc dl = N->getDebugLoc();
7394 SmallVector<SDValue, 8> Ops;
7395 unsigned NumElts = VT.getVectorNumElements();
7396 for (unsigned i = 0; i < NumElts; ++i) {
7397 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7399 // Make the DAGCombiner fold the bitcast.
7400 DCI.AddToWorklist(V.getNode());
7402 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7403 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7404 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7407 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7408 /// ISD::INSERT_VECTOR_ELT.
7409 static SDValue PerformInsertEltCombine(SDNode *N,
7410 TargetLowering::DAGCombinerInfo &DCI) {
7411 // Bitcast an i64 load inserted into a vector to f64.
7412 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7413 EVT VT = N->getValueType(0);
7414 SDNode *Elt = N->getOperand(1).getNode();
7415 if (VT.getVectorElementType() != MVT::i64 ||
7416 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7419 SelectionDAG &DAG = DCI.DAG;
7420 DebugLoc dl = N->getDebugLoc();
7421 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7422 VT.getVectorNumElements());
7423 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7424 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7425 // Make the DAGCombiner fold the bitcasts.
7426 DCI.AddToWorklist(Vec.getNode());
7427 DCI.AddToWorklist(V.getNode());
7428 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7429 Vec, V, N->getOperand(2));
7430 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7433 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7434 /// ISD::VECTOR_SHUFFLE.
7435 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7436 // The LLVM shufflevector instruction does not require the shuffle mask
7437 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7438 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7439 // operands do not match the mask length, they are extended by concatenating
7440 // them with undef vectors. That is probably the right thing for other
7441 // targets, but for NEON it is better to concatenate two double-register
7442 // size vector operands into a single quad-register size vector. Do that
7443 // transformation here:
7444 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7445 // shuffle(concat(v1, v2), undef)
7446 SDValue Op0 = N->getOperand(0);
7447 SDValue Op1 = N->getOperand(1);
7448 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7449 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7450 Op0.getNumOperands() != 2 ||
7451 Op1.getNumOperands() != 2)
7453 SDValue Concat0Op1 = Op0.getOperand(1);
7454 SDValue Concat1Op1 = Op1.getOperand(1);
7455 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7456 Concat1Op1.getOpcode() != ISD::UNDEF)
7458 // Skip the transformation if any of the types are illegal.
7459 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7460 EVT VT = N->getValueType(0);
7461 if (!TLI.isTypeLegal(VT) ||
7462 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7463 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7466 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7467 Op0.getOperand(0), Op1.getOperand(0));
7468 // Translate the shuffle mask.
7469 SmallVector<int, 16> NewMask;
7470 unsigned NumElts = VT.getVectorNumElements();
7471 unsigned HalfElts = NumElts/2;
7472 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7473 for (unsigned n = 0; n < NumElts; ++n) {
7474 int MaskElt = SVN->getMaskElt(n);
7476 if (MaskElt < (int)HalfElts)
7478 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7479 NewElt = HalfElts + MaskElt - NumElts;
7480 NewMask.push_back(NewElt);
7482 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7483 DAG.getUNDEF(VT), NewMask.data());
7486 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7487 /// NEON load/store intrinsics to merge base address updates.
7488 static SDValue CombineBaseUpdate(SDNode *N,
7489 TargetLowering::DAGCombinerInfo &DCI) {
7490 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7493 SelectionDAG &DAG = DCI.DAG;
7494 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7495 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7496 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7497 SDValue Addr = N->getOperand(AddrOpIdx);
7499 // Search for a use of the address operand that is an increment.
7500 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7501 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7503 if (User->getOpcode() != ISD::ADD ||
7504 UI.getUse().getResNo() != Addr.getResNo())
7507 // Check that the add is independent of the load/store. Otherwise, folding
7508 // it would create a cycle.
7509 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7512 // Find the new opcode for the updating load/store.
7514 bool isLaneOp = false;
7515 unsigned NewOpc = 0;
7516 unsigned NumVecs = 0;
7518 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7520 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7521 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7523 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7525 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7527 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7529 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7530 NumVecs = 2; isLaneOp = true; break;
7531 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7532 NumVecs = 3; isLaneOp = true; break;
7533 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7534 NumVecs = 4; isLaneOp = true; break;
7535 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7536 NumVecs = 1; isLoad = false; break;
7537 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7538 NumVecs = 2; isLoad = false; break;
7539 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7540 NumVecs = 3; isLoad = false; break;
7541 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7542 NumVecs = 4; isLoad = false; break;
7543 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7544 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7545 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7546 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7547 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7548 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7552 switch (N->getOpcode()) {
7553 default: llvm_unreachable("unexpected opcode for Neon base update");
7554 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7555 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7556 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7560 // Find the size of memory referenced by the load/store.
7563 VecTy = N->getValueType(0);
7565 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7566 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7568 NumBytes /= VecTy.getVectorNumElements();
7570 // If the increment is a constant, it must match the memory ref size.
7571 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7572 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7573 uint64_t IncVal = CInc->getZExtValue();
7574 if (IncVal != NumBytes)
7576 } else if (NumBytes >= 3 * 16) {
7577 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7578 // separate instructions that make it harder to use a non-constant update.
7582 // Create the new updating load/store node.
7584 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7586 for (n = 0; n < NumResultVecs; ++n)
7588 Tys[n++] = MVT::i32;
7589 Tys[n] = MVT::Other;
7590 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7591 SmallVector<SDValue, 8> Ops;
7592 Ops.push_back(N->getOperand(0)); // incoming chain
7593 Ops.push_back(N->getOperand(AddrOpIdx));
7595 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7596 Ops.push_back(N->getOperand(i));
7598 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7599 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7600 Ops.data(), Ops.size(),
7601 MemInt->getMemoryVT(),
7602 MemInt->getMemOperand());
7605 std::vector<SDValue> NewResults;
7606 for (unsigned i = 0; i < NumResultVecs; ++i) {
7607 NewResults.push_back(SDValue(UpdN.getNode(), i));
7609 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7610 DCI.CombineTo(N, NewResults);
7611 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7618 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7619 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7620 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7622 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7623 SelectionDAG &DAG = DCI.DAG;
7624 EVT VT = N->getValueType(0);
7625 // vldN-dup instructions only support 64-bit vectors for N > 1.
7626 if (!VT.is64BitVector())
7629 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7630 SDNode *VLD = N->getOperand(0).getNode();
7631 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7633 unsigned NumVecs = 0;
7634 unsigned NewOpc = 0;
7635 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7636 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7638 NewOpc = ARMISD::VLD2DUP;
7639 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7641 NewOpc = ARMISD::VLD3DUP;
7642 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7644 NewOpc = ARMISD::VLD4DUP;
7649 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7650 // numbers match the load.
7651 unsigned VLDLaneNo =
7652 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7653 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7655 // Ignore uses of the chain result.
7656 if (UI.getUse().getResNo() == NumVecs)
7659 if (User->getOpcode() != ARMISD::VDUPLANE ||
7660 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7664 // Create the vldN-dup node.
7667 for (n = 0; n < NumVecs; ++n)
7669 Tys[n] = MVT::Other;
7670 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7671 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7672 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7673 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7674 Ops, 2, VLDMemInt->getMemoryVT(),
7675 VLDMemInt->getMemOperand());
7678 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7680 unsigned ResNo = UI.getUse().getResNo();
7681 // Ignore uses of the chain result.
7682 if (ResNo == NumVecs)
7685 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7688 // Now the vldN-lane intrinsic is dead except for its chain result.
7689 // Update uses of the chain.
7690 std::vector<SDValue> VLDDupResults;
7691 for (unsigned n = 0; n < NumVecs; ++n)
7692 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7693 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7694 DCI.CombineTo(VLD, VLDDupResults);
7699 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7700 /// ARMISD::VDUPLANE.
7701 static SDValue PerformVDUPLANECombine(SDNode *N,
7702 TargetLowering::DAGCombinerInfo &DCI) {
7703 SDValue Op = N->getOperand(0);
7705 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7706 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7707 if (CombineVLDDUP(N, DCI))
7708 return SDValue(N, 0);
7710 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7711 // redundant. Ignore bit_converts for now; element sizes are checked below.
7712 while (Op.getOpcode() == ISD::BITCAST)
7713 Op = Op.getOperand(0);
7714 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7717 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7718 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7719 // The canonical VMOV for a zero vector uses a 32-bit element size.
7720 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7722 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7724 EVT VT = N->getValueType(0);
7725 if (EltSize > VT.getVectorElementType().getSizeInBits())
7728 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7731 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7732 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7733 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7737 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7739 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7744 APFloat APF = C->getValueAPF();
7745 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7746 != APFloat::opOK || !isExact)
7749 c0 = (I == 0) ? cN : c0;
7750 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7757 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7758 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7759 /// when the VMUL has a constant operand that is a power of 2.
7761 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7762 /// vmul.f32 d16, d17, d16
7763 /// vcvt.s32.f32 d16, d16
7765 /// vcvt.s32.f32 d16, d16, #3
7766 static SDValue PerformVCVTCombine(SDNode *N,
7767 TargetLowering::DAGCombinerInfo &DCI,
7768 const ARMSubtarget *Subtarget) {
7769 SelectionDAG &DAG = DCI.DAG;
7770 SDValue Op = N->getOperand(0);
7772 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7773 Op.getOpcode() != ISD::FMUL)
7777 SDValue N0 = Op->getOperand(0);
7778 SDValue ConstVec = Op->getOperand(1);
7779 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7781 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7782 !isConstVecPow2(ConstVec, isSigned, C))
7785 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7786 Intrinsic::arm_neon_vcvtfp2fxu;
7787 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7789 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7790 DAG.getConstant(Log2_64(C), MVT::i32));
7793 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7794 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7795 /// when the VDIV has a constant operand that is a power of 2.
7797 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7798 /// vcvt.f32.s32 d16, d16
7799 /// vdiv.f32 d16, d17, d16
7801 /// vcvt.f32.s32 d16, d16, #3
7802 static SDValue PerformVDIVCombine(SDNode *N,
7803 TargetLowering::DAGCombinerInfo &DCI,
7804 const ARMSubtarget *Subtarget) {
7805 SelectionDAG &DAG = DCI.DAG;
7806 SDValue Op = N->getOperand(0);
7807 unsigned OpOpcode = Op.getNode()->getOpcode();
7809 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7810 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7814 SDValue ConstVec = N->getOperand(1);
7815 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7817 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7818 !isConstVecPow2(ConstVec, isSigned, C))
7821 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7822 Intrinsic::arm_neon_vcvtfxu2fp;
7823 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7825 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7826 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7829 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7830 /// operand of a vector shift operation, where all the elements of the
7831 /// build_vector must have the same constant integer value.
7832 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7833 // Ignore bit_converts.
7834 while (Op.getOpcode() == ISD::BITCAST)
7835 Op = Op.getOperand(0);
7836 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7837 APInt SplatBits, SplatUndef;
7838 unsigned SplatBitSize;
7840 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7841 HasAnyUndefs, ElementBits) ||
7842 SplatBitSize > ElementBits)
7844 Cnt = SplatBits.getSExtValue();
7848 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7849 /// operand of a vector shift left operation. That value must be in the range:
7850 /// 0 <= Value < ElementBits for a left shift; or
7851 /// 0 <= Value <= ElementBits for a long left shift.
7852 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7853 assert(VT.isVector() && "vector shift count is not a vector type");
7854 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7855 if (! getVShiftImm(Op, ElementBits, Cnt))
7857 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7860 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7861 /// operand of a vector shift right operation. For a shift opcode, the value
7862 /// is positive, but for an intrinsic the value count must be negative. The
7863 /// absolute value must be in the range:
7864 /// 1 <= |Value| <= ElementBits for a right shift; or
7865 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7866 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7868 assert(VT.isVector() && "vector shift count is not a vector type");
7869 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7870 if (! getVShiftImm(Op, ElementBits, Cnt))
7874 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7877 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7878 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7879 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7882 // Don't do anything for most intrinsics.
7885 // Vector shifts: check for immediate versions and lower them.
7886 // Note: This is done during DAG combining instead of DAG legalizing because
7887 // the build_vectors for 64-bit vector element shift counts are generally
7888 // not legal, and it is hard to see their values after they get legalized to
7889 // loads from a constant pool.
7890 case Intrinsic::arm_neon_vshifts:
7891 case Intrinsic::arm_neon_vshiftu:
7892 case Intrinsic::arm_neon_vshiftls:
7893 case Intrinsic::arm_neon_vshiftlu:
7894 case Intrinsic::arm_neon_vshiftn:
7895 case Intrinsic::arm_neon_vrshifts:
7896 case Intrinsic::arm_neon_vrshiftu:
7897 case Intrinsic::arm_neon_vrshiftn:
7898 case Intrinsic::arm_neon_vqshifts:
7899 case Intrinsic::arm_neon_vqshiftu:
7900 case Intrinsic::arm_neon_vqshiftsu:
7901 case Intrinsic::arm_neon_vqshiftns:
7902 case Intrinsic::arm_neon_vqshiftnu:
7903 case Intrinsic::arm_neon_vqshiftnsu:
7904 case Intrinsic::arm_neon_vqrshiftns:
7905 case Intrinsic::arm_neon_vqrshiftnu:
7906 case Intrinsic::arm_neon_vqrshiftnsu: {
7907 EVT VT = N->getOperand(1).getValueType();
7909 unsigned VShiftOpc = 0;
7912 case Intrinsic::arm_neon_vshifts:
7913 case Intrinsic::arm_neon_vshiftu:
7914 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7915 VShiftOpc = ARMISD::VSHL;
7918 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7919 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7920 ARMISD::VSHRs : ARMISD::VSHRu);
7925 case Intrinsic::arm_neon_vshiftls:
7926 case Intrinsic::arm_neon_vshiftlu:
7927 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7929 llvm_unreachable("invalid shift count for vshll intrinsic");
7931 case Intrinsic::arm_neon_vrshifts:
7932 case Intrinsic::arm_neon_vrshiftu:
7933 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7937 case Intrinsic::arm_neon_vqshifts:
7938 case Intrinsic::arm_neon_vqshiftu:
7939 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7943 case Intrinsic::arm_neon_vqshiftsu:
7944 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7946 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7948 case Intrinsic::arm_neon_vshiftn:
7949 case Intrinsic::arm_neon_vrshiftn:
7950 case Intrinsic::arm_neon_vqshiftns:
7951 case Intrinsic::arm_neon_vqshiftnu:
7952 case Intrinsic::arm_neon_vqshiftnsu:
7953 case Intrinsic::arm_neon_vqrshiftns:
7954 case Intrinsic::arm_neon_vqrshiftnu:
7955 case Intrinsic::arm_neon_vqrshiftnsu:
7956 // Narrowing shifts require an immediate right shift.
7957 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7959 llvm_unreachable("invalid shift count for narrowing vector shift "
7963 llvm_unreachable("unhandled vector shift");
7967 case Intrinsic::arm_neon_vshifts:
7968 case Intrinsic::arm_neon_vshiftu:
7969 // Opcode already set above.
7971 case Intrinsic::arm_neon_vshiftls:
7972 case Intrinsic::arm_neon_vshiftlu:
7973 if (Cnt == VT.getVectorElementType().getSizeInBits())
7974 VShiftOpc = ARMISD::VSHLLi;
7976 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7977 ARMISD::VSHLLs : ARMISD::VSHLLu);
7979 case Intrinsic::arm_neon_vshiftn:
7980 VShiftOpc = ARMISD::VSHRN; break;
7981 case Intrinsic::arm_neon_vrshifts:
7982 VShiftOpc = ARMISD::VRSHRs; break;
7983 case Intrinsic::arm_neon_vrshiftu:
7984 VShiftOpc = ARMISD::VRSHRu; break;
7985 case Intrinsic::arm_neon_vrshiftn:
7986 VShiftOpc = ARMISD::VRSHRN; break;
7987 case Intrinsic::arm_neon_vqshifts:
7988 VShiftOpc = ARMISD::VQSHLs; break;
7989 case Intrinsic::arm_neon_vqshiftu:
7990 VShiftOpc = ARMISD::VQSHLu; break;
7991 case Intrinsic::arm_neon_vqshiftsu:
7992 VShiftOpc = ARMISD::VQSHLsu; break;
7993 case Intrinsic::arm_neon_vqshiftns:
7994 VShiftOpc = ARMISD::VQSHRNs; break;
7995 case Intrinsic::arm_neon_vqshiftnu:
7996 VShiftOpc = ARMISD::VQSHRNu; break;
7997 case Intrinsic::arm_neon_vqshiftnsu:
7998 VShiftOpc = ARMISD::VQSHRNsu; break;
7999 case Intrinsic::arm_neon_vqrshiftns:
8000 VShiftOpc = ARMISD::VQRSHRNs; break;
8001 case Intrinsic::arm_neon_vqrshiftnu:
8002 VShiftOpc = ARMISD::VQRSHRNu; break;
8003 case Intrinsic::arm_neon_vqrshiftnsu:
8004 VShiftOpc = ARMISD::VQRSHRNsu; break;
8007 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8008 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8011 case Intrinsic::arm_neon_vshiftins: {
8012 EVT VT = N->getOperand(1).getValueType();
8014 unsigned VShiftOpc = 0;
8016 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8017 VShiftOpc = ARMISD::VSLI;
8018 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8019 VShiftOpc = ARMISD::VSRI;
8021 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8024 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8025 N->getOperand(1), N->getOperand(2),
8026 DAG.getConstant(Cnt, MVT::i32));
8029 case Intrinsic::arm_neon_vqrshifts:
8030 case Intrinsic::arm_neon_vqrshiftu:
8031 // No immediate versions of these to check for.
8038 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8039 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8040 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8041 /// vector element shift counts are generally not legal, and it is hard to see
8042 /// their values after they get legalized to loads from a constant pool.
8043 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8044 const ARMSubtarget *ST) {
8045 EVT VT = N->getValueType(0);
8046 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8047 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8048 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8049 SDValue N1 = N->getOperand(1);
8050 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8051 SDValue N0 = N->getOperand(0);
8052 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8053 DAG.MaskedValueIsZero(N0.getOperand(0),
8054 APInt::getHighBitsSet(32, 16)))
8055 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8059 // Nothing to be done for scalar shifts.
8060 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8061 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8064 assert(ST->hasNEON() && "unexpected vector shift");
8067 switch (N->getOpcode()) {
8068 default: llvm_unreachable("unexpected shift opcode");
8071 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8072 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8073 DAG.getConstant(Cnt, MVT::i32));
8078 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8079 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8080 ARMISD::VSHRs : ARMISD::VSHRu);
8081 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8082 DAG.getConstant(Cnt, MVT::i32));
8088 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8089 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8090 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8091 const ARMSubtarget *ST) {
8092 SDValue N0 = N->getOperand(0);
8094 // Check for sign- and zero-extensions of vector extract operations of 8-
8095 // and 16-bit vector elements. NEON supports these directly. They are
8096 // handled during DAG combining because type legalization will promote them
8097 // to 32-bit types and it is messy to recognize the operations after that.
8098 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8099 SDValue Vec = N0.getOperand(0);
8100 SDValue Lane = N0.getOperand(1);
8101 EVT VT = N->getValueType(0);
8102 EVT EltVT = N0.getValueType();
8103 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8105 if (VT == MVT::i32 &&
8106 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8107 TLI.isTypeLegal(Vec.getValueType()) &&
8108 isa<ConstantSDNode>(Lane)) {
8111 switch (N->getOpcode()) {
8112 default: llvm_unreachable("unexpected opcode");
8113 case ISD::SIGN_EXTEND:
8114 Opc = ARMISD::VGETLANEs;
8116 case ISD::ZERO_EXTEND:
8117 case ISD::ANY_EXTEND:
8118 Opc = ARMISD::VGETLANEu;
8121 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8128 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8129 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8130 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8131 const ARMSubtarget *ST) {
8132 // If the target supports NEON, try to use vmax/vmin instructions for f32
8133 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8134 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8135 // a NaN; only do the transformation when it matches that behavior.
8137 // For now only do this when using NEON for FP operations; if using VFP, it
8138 // is not obvious that the benefit outweighs the cost of switching to the
8140 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8141 N->getValueType(0) != MVT::f32)
8144 SDValue CondLHS = N->getOperand(0);
8145 SDValue CondRHS = N->getOperand(1);
8146 SDValue LHS = N->getOperand(2);
8147 SDValue RHS = N->getOperand(3);
8148 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8150 unsigned Opcode = 0;
8152 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8153 IsReversed = false; // x CC y ? x : y
8154 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8155 IsReversed = true ; // x CC y ? y : x
8169 // If LHS is NaN, an ordered comparison will be false and the result will
8170 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8171 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8172 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8173 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8175 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8176 // will return -0, so vmin can only be used for unsafe math or if one of
8177 // the operands is known to be nonzero.
8178 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8179 !DAG.getTarget().Options.UnsafeFPMath &&
8180 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8182 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8191 // If LHS is NaN, an ordered comparison will be false and the result will
8192 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8193 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8194 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8195 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8197 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8198 // will return +0, so vmax can only be used for unsafe math or if one of
8199 // the operands is known to be nonzero.
8200 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8201 !DAG.getTarget().Options.UnsafeFPMath &&
8202 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8204 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8210 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8213 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8215 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8216 SDValue Cmp = N->getOperand(4);
8217 if (Cmp.getOpcode() != ARMISD::CMPZ)
8218 // Only looking at EQ and NE cases.
8221 EVT VT = N->getValueType(0);
8222 DebugLoc dl = N->getDebugLoc();
8223 SDValue LHS = Cmp.getOperand(0);
8224 SDValue RHS = Cmp.getOperand(1);
8225 SDValue FalseVal = N->getOperand(0);
8226 SDValue TrueVal = N->getOperand(1);
8227 SDValue ARMcc = N->getOperand(2);
8228 ARMCC::CondCodes CC =
8229 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8247 /// FIXME: Turn this into a target neutral optimization?
8249 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8250 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8251 N->getOperand(3), Cmp);
8252 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8254 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8255 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8256 N->getOperand(3), NewCmp);
8259 if (Res.getNode()) {
8260 APInt KnownZero, KnownOne;
8261 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8262 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8263 // Capture demanded bits information that would be otherwise lost.
8264 if (KnownZero == 0xfffffffe)
8265 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8266 DAG.getValueType(MVT::i1));
8267 else if (KnownZero == 0xffffff00)
8268 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8269 DAG.getValueType(MVT::i8));
8270 else if (KnownZero == 0xffff0000)
8271 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8272 DAG.getValueType(MVT::i16));
8278 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8279 DAGCombinerInfo &DCI) const {
8280 switch (N->getOpcode()) {
8282 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8283 case ISD::SUB: return PerformSUBCombine(N, DCI);
8284 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8285 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8286 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8287 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8288 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8289 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8290 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8291 case ISD::STORE: return PerformSTORECombine(N, DCI);
8292 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8293 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8294 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8295 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8296 case ISD::FP_TO_SINT:
8297 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8298 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8299 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8302 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8303 case ISD::SIGN_EXTEND:
8304 case ISD::ZERO_EXTEND:
8305 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8306 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8307 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8308 case ARMISD::VLD2DUP:
8309 case ARMISD::VLD3DUP:
8310 case ARMISD::VLD4DUP:
8311 return CombineBaseUpdate(N, DCI);
8312 case ISD::INTRINSIC_VOID:
8313 case ISD::INTRINSIC_W_CHAIN:
8314 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8315 case Intrinsic::arm_neon_vld1:
8316 case Intrinsic::arm_neon_vld2:
8317 case Intrinsic::arm_neon_vld3:
8318 case Intrinsic::arm_neon_vld4:
8319 case Intrinsic::arm_neon_vld2lane:
8320 case Intrinsic::arm_neon_vld3lane:
8321 case Intrinsic::arm_neon_vld4lane:
8322 case Intrinsic::arm_neon_vst1:
8323 case Intrinsic::arm_neon_vst2:
8324 case Intrinsic::arm_neon_vst3:
8325 case Intrinsic::arm_neon_vst4:
8326 case Intrinsic::arm_neon_vst2lane:
8327 case Intrinsic::arm_neon_vst3lane:
8328 case Intrinsic::arm_neon_vst4lane:
8329 return CombineBaseUpdate(N, DCI);
8337 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8339 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8342 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8343 if (!Subtarget->allowsUnalignedMem())
8346 switch (VT.getSimpleVT().SimpleTy) {
8353 // FIXME: VLD1 etc with standard alignment is legal.
8357 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8358 unsigned AlignCheck) {
8359 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8360 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8363 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8364 unsigned DstAlign, unsigned SrcAlign,
8367 MachineFunction &MF) const {
8368 const Function *F = MF.getFunction();
8370 // See if we can use NEON instructions for this...
8372 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8373 Subtarget->hasNEON()) {
8374 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8376 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8381 // Lowering to i32/i16 if the size permits.
8384 } else if (Size >= 2) {
8388 // Let the target-independent logic figure it out.
8392 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8397 switch (VT.getSimpleVT().SimpleTy) {
8398 default: return false;
8413 if ((V & (Scale - 1)) != 0)
8416 return V == (V & ((1LL << 5) - 1));
8419 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8420 const ARMSubtarget *Subtarget) {
8427 switch (VT.getSimpleVT().SimpleTy) {
8428 default: return false;
8433 // + imm12 or - imm8
8435 return V == (V & ((1LL << 8) - 1));
8436 return V == (V & ((1LL << 12) - 1));
8439 // Same as ARM mode. FIXME: NEON?
8440 if (!Subtarget->hasVFP2())
8445 return V == (V & ((1LL << 8) - 1));
8449 /// isLegalAddressImmediate - Return true if the integer value can be used
8450 /// as the offset of the target addressing mode for load / store of the
8452 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8453 const ARMSubtarget *Subtarget) {
8460 if (Subtarget->isThumb1Only())
8461 return isLegalT1AddressImmediate(V, VT);
8462 else if (Subtarget->isThumb2())
8463 return isLegalT2AddressImmediate(V, VT, Subtarget);
8468 switch (VT.getSimpleVT().SimpleTy) {
8469 default: return false;
8474 return V == (V & ((1LL << 12) - 1));
8477 return V == (V & ((1LL << 8) - 1));
8480 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8485 return V == (V & ((1LL << 8) - 1));
8489 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8491 int Scale = AM.Scale;
8495 switch (VT.getSimpleVT().SimpleTy) {
8496 default: return false;
8505 return Scale == 2 || Scale == 4 || Scale == 8;
8508 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8512 // Note, we allow "void" uses (basically, uses that aren't loads or
8513 // stores), because arm allows folding a scale into many arithmetic
8514 // operations. This should be made more precise and revisited later.
8516 // Allow r << imm, but the imm has to be a multiple of two.
8517 if (Scale & 1) return false;
8518 return isPowerOf2_32(Scale);
8522 /// isLegalAddressingMode - Return true if the addressing mode represented
8523 /// by AM is legal for this target, for a load/store of the specified type.
8524 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8526 EVT VT = getValueType(Ty, true);
8527 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8530 // Can never fold addr of global into load/store.
8535 case 0: // no scale reg, must be "r+i" or "r", or "i".
8538 if (Subtarget->isThumb1Only())
8542 // ARM doesn't support any R+R*scale+imm addr modes.
8549 if (Subtarget->isThumb2())
8550 return isLegalT2ScaledAddressingMode(AM, VT);
8552 int Scale = AM.Scale;
8553 switch (VT.getSimpleVT().SimpleTy) {
8554 default: return false;
8558 if (Scale < 0) Scale = -Scale;
8562 return isPowerOf2_32(Scale & ~1);
8566 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8571 // Note, we allow "void" uses (basically, uses that aren't loads or
8572 // stores), because arm allows folding a scale into many arithmetic
8573 // operations. This should be made more precise and revisited later.
8575 // Allow r << imm, but the imm has to be a multiple of two.
8576 if (Scale & 1) return false;
8577 return isPowerOf2_32(Scale);
8583 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8584 /// icmp immediate, that is the target has icmp instructions which can compare
8585 /// a register against the immediate without having to materialize the
8586 /// immediate into a register.
8587 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8588 if (!Subtarget->isThumb())
8589 return ARM_AM::getSOImmVal(Imm) != -1;
8590 if (Subtarget->isThumb2())
8591 return ARM_AM::getT2SOImmVal(Imm) != -1;
8592 return Imm >= 0 && Imm <= 255;
8595 /// isLegalAddImmediate - Return true if the specified immediate is legal
8596 /// add immediate, that is the target has add instructions which can add
8597 /// a register with the immediate without having to materialize the
8598 /// immediate into a register.
8599 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8600 return ARM_AM::getSOImmVal(Imm) != -1;
8603 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8604 bool isSEXTLoad, SDValue &Base,
8605 SDValue &Offset, bool &isInc,
8606 SelectionDAG &DAG) {
8607 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8610 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8612 Base = Ptr->getOperand(0);
8613 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8614 int RHSC = (int)RHS->getZExtValue();
8615 if (RHSC < 0 && RHSC > -256) {
8616 assert(Ptr->getOpcode() == ISD::ADD);
8618 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8622 isInc = (Ptr->getOpcode() == ISD::ADD);
8623 Offset = Ptr->getOperand(1);
8625 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8627 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8628 int RHSC = (int)RHS->getZExtValue();
8629 if (RHSC < 0 && RHSC > -0x1000) {
8630 assert(Ptr->getOpcode() == ISD::ADD);
8632 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8633 Base = Ptr->getOperand(0);
8638 if (Ptr->getOpcode() == ISD::ADD) {
8640 ARM_AM::ShiftOpc ShOpcVal=
8641 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8642 if (ShOpcVal != ARM_AM::no_shift) {
8643 Base = Ptr->getOperand(1);
8644 Offset = Ptr->getOperand(0);
8646 Base = Ptr->getOperand(0);
8647 Offset = Ptr->getOperand(1);
8652 isInc = (Ptr->getOpcode() == ISD::ADD);
8653 Base = Ptr->getOperand(0);
8654 Offset = Ptr->getOperand(1);
8658 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8662 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8663 bool isSEXTLoad, SDValue &Base,
8664 SDValue &Offset, bool &isInc,
8665 SelectionDAG &DAG) {
8666 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8669 Base = Ptr->getOperand(0);
8670 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8671 int RHSC = (int)RHS->getZExtValue();
8672 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8673 assert(Ptr->getOpcode() == ISD::ADD);
8675 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8677 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8678 isInc = Ptr->getOpcode() == ISD::ADD;
8679 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8687 /// getPreIndexedAddressParts - returns true by value, base pointer and
8688 /// offset pointer and addressing mode by reference if the node's address
8689 /// can be legally represented as pre-indexed load / store address.
8691 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8693 ISD::MemIndexedMode &AM,
8694 SelectionDAG &DAG) const {
8695 if (Subtarget->isThumb1Only())
8700 bool isSEXTLoad = false;
8701 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8702 Ptr = LD->getBasePtr();
8703 VT = LD->getMemoryVT();
8704 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8705 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8706 Ptr = ST->getBasePtr();
8707 VT = ST->getMemoryVT();
8712 bool isLegal = false;
8713 if (Subtarget->isThumb2())
8714 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8715 Offset, isInc, DAG);
8717 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8718 Offset, isInc, DAG);
8722 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8726 /// getPostIndexedAddressParts - returns true by value, base pointer and
8727 /// offset pointer and addressing mode by reference if this node can be
8728 /// combined with a load / store to form a post-indexed load / store.
8729 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8732 ISD::MemIndexedMode &AM,
8733 SelectionDAG &DAG) const {
8734 if (Subtarget->isThumb1Only())
8739 bool isSEXTLoad = false;
8740 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8741 VT = LD->getMemoryVT();
8742 Ptr = LD->getBasePtr();
8743 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8744 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8745 VT = ST->getMemoryVT();
8746 Ptr = ST->getBasePtr();
8751 bool isLegal = false;
8752 if (Subtarget->isThumb2())
8753 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8756 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8762 // Swap base ptr and offset to catch more post-index load / store when
8763 // it's legal. In Thumb2 mode, offset must be an immediate.
8764 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8765 !Subtarget->isThumb2())
8766 std::swap(Base, Offset);
8768 // Post-indexed load / store update the base pointer.
8773 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8777 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8781 const SelectionDAG &DAG,
8782 unsigned Depth) const {
8783 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8784 switch (Op.getOpcode()) {
8786 case ARMISD::CMOV: {
8787 // Bits are known zero/one if known on the LHS and RHS.
8788 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8789 if (KnownZero == 0 && KnownOne == 0) return;
8791 APInt KnownZeroRHS, KnownOneRHS;
8792 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8793 KnownZeroRHS, KnownOneRHS, Depth+1);
8794 KnownZero &= KnownZeroRHS;
8795 KnownOne &= KnownOneRHS;
8801 //===----------------------------------------------------------------------===//
8802 // ARM Inline Assembly Support
8803 //===----------------------------------------------------------------------===//
8805 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8806 // Looking for "rev" which is V6+.
8807 if (!Subtarget->hasV6Ops())
8810 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8811 std::string AsmStr = IA->getAsmString();
8812 SmallVector<StringRef, 4> AsmPieces;
8813 SplitString(AsmStr, AsmPieces, ";\n");
8815 switch (AsmPieces.size()) {
8816 default: return false;
8818 AsmStr = AsmPieces[0];
8820 SplitString(AsmStr, AsmPieces, " \t,");
8823 if (AsmPieces.size() == 3 &&
8824 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8825 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8826 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8827 if (Ty && Ty->getBitWidth() == 32)
8828 return IntrinsicLowering::LowerToByteSwap(CI);
8836 /// getConstraintType - Given a constraint letter, return the type of
8837 /// constraint it is for this target.
8838 ARMTargetLowering::ConstraintType
8839 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8840 if (Constraint.size() == 1) {
8841 switch (Constraint[0]) {
8843 case 'l': return C_RegisterClass;
8844 case 'w': return C_RegisterClass;
8845 case 'h': return C_RegisterClass;
8846 case 'x': return C_RegisterClass;
8847 case 't': return C_RegisterClass;
8848 case 'j': return C_Other; // Constant for movw.
8849 // An address with a single base register. Due to the way we
8850 // currently handle addresses it is the same as an 'r' memory constraint.
8851 case 'Q': return C_Memory;
8853 } else if (Constraint.size() == 2) {
8854 switch (Constraint[0]) {
8856 // All 'U+' constraints are addresses.
8857 case 'U': return C_Memory;
8860 return TargetLowering::getConstraintType(Constraint);
8863 /// Examine constraint type and operand type and determine a weight value.
8864 /// This object must already have been set up with the operand type
8865 /// and the current alternative constraint selected.
8866 TargetLowering::ConstraintWeight
8867 ARMTargetLowering::getSingleConstraintMatchWeight(
8868 AsmOperandInfo &info, const char *constraint) const {
8869 ConstraintWeight weight = CW_Invalid;
8870 Value *CallOperandVal = info.CallOperandVal;
8871 // If we don't have a value, we can't do a match,
8872 // but allow it at the lowest weight.
8873 if (CallOperandVal == NULL)
8875 Type *type = CallOperandVal->getType();
8876 // Look at the constraint type.
8877 switch (*constraint) {
8879 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8882 if (type->isIntegerTy()) {
8883 if (Subtarget->isThumb())
8884 weight = CW_SpecificReg;
8886 weight = CW_Register;
8890 if (type->isFloatingPointTy())
8891 weight = CW_Register;
8897 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8899 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8901 if (Constraint.size() == 1) {
8902 // GCC ARM Constraint Letters
8903 switch (Constraint[0]) {
8904 case 'l': // Low regs or general regs.
8905 if (Subtarget->isThumb())
8906 return RCPair(0U, ARM::tGPRRegisterClass);
8908 return RCPair(0U, ARM::GPRRegisterClass);
8909 case 'h': // High regs or no regs.
8910 if (Subtarget->isThumb())
8911 return RCPair(0U, ARM::hGPRRegisterClass);
8914 return RCPair(0U, ARM::GPRRegisterClass);
8917 return RCPair(0U, ARM::SPRRegisterClass);
8918 if (VT.getSizeInBits() == 64)
8919 return RCPair(0U, ARM::DPRRegisterClass);
8920 if (VT.getSizeInBits() == 128)
8921 return RCPair(0U, ARM::QPRRegisterClass);
8925 return RCPair(0U, ARM::SPR_8RegisterClass);
8926 if (VT.getSizeInBits() == 64)
8927 return RCPair(0U, ARM::DPR_8RegisterClass);
8928 if (VT.getSizeInBits() == 128)
8929 return RCPair(0U, ARM::QPR_8RegisterClass);
8933 return RCPair(0U, ARM::SPRRegisterClass);
8937 if (StringRef("{cc}").equals_lower(Constraint))
8938 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8940 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8943 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8944 /// vector. If it is invalid, don't add anything to Ops.
8945 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8946 std::string &Constraint,
8947 std::vector<SDValue>&Ops,
8948 SelectionDAG &DAG) const {
8949 SDValue Result(0, 0);
8951 // Currently only support length 1 constraints.
8952 if (Constraint.length() != 1) return;
8954 char ConstraintLetter = Constraint[0];
8955 switch (ConstraintLetter) {
8958 case 'I': case 'J': case 'K': case 'L':
8959 case 'M': case 'N': case 'O':
8960 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8964 int64_t CVal64 = C->getSExtValue();
8965 int CVal = (int) CVal64;
8966 // None of these constraints allow values larger than 32 bits. Check
8967 // that the value fits in an int.
8971 switch (ConstraintLetter) {
8973 // Constant suitable for movw, must be between 0 and
8975 if (Subtarget->hasV6T2Ops())
8976 if (CVal >= 0 && CVal <= 65535)
8980 if (Subtarget->isThumb1Only()) {
8981 // This must be a constant between 0 and 255, for ADD
8983 if (CVal >= 0 && CVal <= 255)
8985 } else if (Subtarget->isThumb2()) {
8986 // A constant that can be used as an immediate value in a
8987 // data-processing instruction.
8988 if (ARM_AM::getT2SOImmVal(CVal) != -1)
8991 // A constant that can be used as an immediate value in a
8992 // data-processing instruction.
8993 if (ARM_AM::getSOImmVal(CVal) != -1)
8999 if (Subtarget->isThumb()) { // FIXME thumb2
9000 // This must be a constant between -255 and -1, for negated ADD
9001 // immediates. This can be used in GCC with an "n" modifier that
9002 // prints the negated value, for use with SUB instructions. It is
9003 // not useful otherwise but is implemented for compatibility.
9004 if (CVal >= -255 && CVal <= -1)
9007 // This must be a constant between -4095 and 4095. It is not clear
9008 // what this constraint is intended for. Implemented for
9009 // compatibility with GCC.
9010 if (CVal >= -4095 && CVal <= 4095)
9016 if (Subtarget->isThumb1Only()) {
9017 // A 32-bit value where only one byte has a nonzero value. Exclude
9018 // zero to match GCC. This constraint is used by GCC internally for
9019 // constants that can be loaded with a move/shift combination.
9020 // It is not useful otherwise but is implemented for compatibility.
9021 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9023 } else if (Subtarget->isThumb2()) {
9024 // A constant whose bitwise inverse can be used as an immediate
9025 // value in a data-processing instruction. This can be used in GCC
9026 // with a "B" modifier that prints the inverted value, for use with
9027 // BIC and MVN instructions. It is not useful otherwise but is
9028 // implemented for compatibility.
9029 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9032 // A constant whose bitwise inverse can be used as an immediate
9033 // value in a data-processing instruction. This can be used in GCC
9034 // with a "B" modifier that prints the inverted value, for use with
9035 // BIC and MVN instructions. It is not useful otherwise but is
9036 // implemented for compatibility.
9037 if (ARM_AM::getSOImmVal(~CVal) != -1)
9043 if (Subtarget->isThumb1Only()) {
9044 // This must be a constant between -7 and 7,
9045 // for 3-operand ADD/SUB immediate instructions.
9046 if (CVal >= -7 && CVal < 7)
9048 } else if (Subtarget->isThumb2()) {
9049 // A constant whose negation can be used as an immediate value in a
9050 // data-processing instruction. This can be used in GCC with an "n"
9051 // modifier that prints the negated value, for use with SUB
9052 // instructions. It is not useful otherwise but is implemented for
9054 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9057 // A constant whose negation can be used as an immediate value in a
9058 // data-processing instruction. This can be used in GCC with an "n"
9059 // modifier that prints the negated value, for use with SUB
9060 // instructions. It is not useful otherwise but is implemented for
9062 if (ARM_AM::getSOImmVal(-CVal) != -1)
9068 if (Subtarget->isThumb()) { // FIXME thumb2
9069 // This must be a multiple of 4 between 0 and 1020, for
9070 // ADD sp + immediate.
9071 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9074 // A power of two or a constant between 0 and 32. This is used in
9075 // GCC for the shift amount on shifted register operands, but it is
9076 // useful in general for any shift amounts.
9077 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9083 if (Subtarget->isThumb()) { // FIXME thumb2
9084 // This must be a constant between 0 and 31, for shift amounts.
9085 if (CVal >= 0 && CVal <= 31)
9091 if (Subtarget->isThumb()) { // FIXME thumb2
9092 // This must be a multiple of 4 between -508 and 508, for
9093 // ADD/SUB sp = sp + immediate.
9094 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9099 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9103 if (Result.getNode()) {
9104 Ops.push_back(Result);
9107 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9111 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9112 // The ARM target isn't yet aware of offsets.
9116 bool ARM::isBitFieldInvertedMask(unsigned v) {
9117 if (v == 0xffffffff)
9119 // there can be 1's on either or both "outsides", all the "inside"
9121 unsigned int lsb = 0, msb = 31;
9122 while (v & (1 << msb)) --msb;
9123 while (v & (1 << lsb)) ++lsb;
9124 for (unsigned int i = lsb; i <= msb; ++i) {
9131 /// isFPImmLegal - Returns true if the target can instruction select the
9132 /// specified FP immediate natively. If false, the legalizer will
9133 /// materialize the FP immediate as a load from a constant pool.
9134 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9135 if (!Subtarget->hasVFP3())
9138 return ARM_AM::getFP32Imm(Imm) != -1;
9140 return ARM_AM::getFP64Imm(Imm) != -1;
9144 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9145 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9146 /// specified in the intrinsic calls.
9147 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9149 unsigned Intrinsic) const {
9150 switch (Intrinsic) {
9151 case Intrinsic::arm_neon_vld1:
9152 case Intrinsic::arm_neon_vld2:
9153 case Intrinsic::arm_neon_vld3:
9154 case Intrinsic::arm_neon_vld4:
9155 case Intrinsic::arm_neon_vld2lane:
9156 case Intrinsic::arm_neon_vld3lane:
9157 case Intrinsic::arm_neon_vld4lane: {
9158 Info.opc = ISD::INTRINSIC_W_CHAIN;
9159 // Conservatively set memVT to the entire set of vectors loaded.
9160 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9161 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9162 Info.ptrVal = I.getArgOperand(0);
9164 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9165 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9166 Info.vol = false; // volatile loads with NEON intrinsics not supported
9167 Info.readMem = true;
9168 Info.writeMem = false;
9171 case Intrinsic::arm_neon_vst1:
9172 case Intrinsic::arm_neon_vst2:
9173 case Intrinsic::arm_neon_vst3:
9174 case Intrinsic::arm_neon_vst4:
9175 case Intrinsic::arm_neon_vst2lane:
9176 case Intrinsic::arm_neon_vst3lane:
9177 case Intrinsic::arm_neon_vst4lane: {
9178 Info.opc = ISD::INTRINSIC_VOID;
9179 // Conservatively set memVT to the entire set of vectors stored.
9180 unsigned NumElts = 0;
9181 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9182 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9183 if (!ArgTy->isVectorTy())
9185 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9187 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9188 Info.ptrVal = I.getArgOperand(0);
9190 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9191 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9192 Info.vol = false; // volatile stores with NEON intrinsics not supported
9193 Info.readMem = false;
9194 Info.writeMem = true;
9197 case Intrinsic::arm_strexd: {
9198 Info.opc = ISD::INTRINSIC_W_CHAIN;
9199 Info.memVT = MVT::i64;
9200 Info.ptrVal = I.getArgOperand(2);
9204 Info.readMem = false;
9205 Info.writeMem = true;
9208 case Intrinsic::arm_ldrexd: {
9209 Info.opc = ISD::INTRINSIC_W_CHAIN;
9210 Info.memVT = MVT::i64;
9211 Info.ptrVal = I.getArgOperand(0);
9215 Info.readMem = true;
9216 Info.writeMem = false;