1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "MCTargetDesc/ARMAddressingModes.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/IntrinsicLowering.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
56 // This option should go away when tail calls fully work.
58 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
59 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
63 EnableARMLongCalls("arm-long-calls", cl::Hidden,
64 cl::desc("Generate calls via indirect call instructions"),
68 ARMInterworking("arm-interworking", cl::Hidden,
69 cl::desc("Enable / disable ARM interworking (for debugging only)"),
73 class ARMCCState : public CCState {
75 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
76 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
77 LLVMContext &C, ParmContext PC)
78 : CCState(CC, isVarArg, MF, TM, locs, C) {
79 assert(((PC == Call) || (PC == Prologue)) &&
80 "ARMCCState users must specify whether their context is call"
81 "or prologue generation.");
87 // The APCS parameter registers.
88 static const uint16_t GPRArgRegs[] = {
89 ARM::R0, ARM::R1, ARM::R2, ARM::R3
92 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
93 EVT PromotedBitwiseVT) {
94 if (VT != PromotedLdStVT) {
95 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
96 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
97 PromotedLdStVT.getSimpleVT());
99 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
100 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
101 PromotedLdStVT.getSimpleVT());
104 EVT ElemTy = VT.getVectorElementType();
105 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
106 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
107 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
109 if (ElemTy == MVT::i32) {
110 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
113 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
115 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
121 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
122 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
123 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
124 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
125 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
129 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
130 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
136 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
137 PromotedBitwiseVT.getSimpleVT());
138 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
139 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
140 PromotedBitwiseVT.getSimpleVT());
141 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
142 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
143 PromotedBitwiseVT.getSimpleVT());
146 // Neon does not support vector divide/remainder operations.
147 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
148 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
149 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
150 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
155 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
156 addRegisterClass(VT, &ARM::DPRRegClass);
157 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
160 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
161 addRegisterClass(VT, &ARM::QPRRegClass);
162 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
165 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
166 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
167 return new TargetLoweringObjectFileMachO();
169 return new ARMElfTargetObjectFile();
172 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
173 : TargetLowering(TM, createTLOF(TM)) {
174 Subtarget = &TM.getSubtarget<ARMSubtarget>();
175 RegInfo = TM.getRegisterInfo();
176 Itins = TM.getInstrItineraryData();
178 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
180 if (Subtarget->isTargetDarwin()) {
181 // Uses VFP for Thumb libfuncs if available.
182 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
183 // Single-precision floating-point arithmetic.
184 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
185 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
186 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
187 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
189 // Double-precision floating-point arithmetic.
190 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
191 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
192 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
193 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
195 // Single-precision comparisons.
196 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
197 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
198 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
199 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
200 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
201 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
202 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
203 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
205 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
214 // Double-precision comparisons.
215 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
216 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
217 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
218 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
219 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
220 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
221 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
222 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
224 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
233 // Floating-point to integer conversions.
234 // i64 conversions are done via library routines even when generating VFP
235 // instructions, so use the same ones.
236 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
237 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
238 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
239 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
241 // Conversions between floating types.
242 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
243 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
245 // Integer to floating-point conversions.
246 // i64 conversions are done via library routines even when generating VFP
247 // instructions, so use the same ones.
248 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
249 // e.g., __floatunsidf vs. __floatunssidfvfp.
250 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
251 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
252 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
253 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
257 // These libcalls are not available in 32-bit.
258 setLibcallName(RTLIB::SHL_I128, 0);
259 setLibcallName(RTLIB::SRL_I128, 0);
260 setLibcallName(RTLIB::SRA_I128, 0);
262 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
263 // Double-precision floating-point arithmetic helper functions
264 // RTABI chapter 4.1.2, Table 2
265 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
266 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
267 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
268 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
269 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
270 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
274 // Double-precision floating-point comparison helper functions
275 // RTABI chapter 4.1.2, Table 3
276 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
277 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
278 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
279 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
280 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
281 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
282 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
283 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
284 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
285 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
286 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
287 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
288 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
289 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
290 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
291 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
292 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
301 // Single-precision floating-point arithmetic helper functions
302 // RTABI chapter 4.1.2, Table 4
303 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
304 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
305 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
306 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
307 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
308 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
312 // Single-precision floating-point comparison helper functions
313 // RTABI chapter 4.1.2, Table 5
314 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
315 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
316 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
317 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
318 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
319 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
320 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
321 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
322 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
323 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
324 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
325 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
326 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
327 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
328 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
329 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
330 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
339 // Floating-point to integer conversions.
340 // RTABI chapter 4.1.2, Table 6
341 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
342 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
343 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
344 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
345 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
346 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
347 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
348 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
349 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
358 // Conversions between floating types.
359 // RTABI chapter 4.1.2, Table 7
360 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
361 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
362 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
363 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
365 // Integer to floating-point conversions.
366 // RTABI chapter 4.1.2, Table 8
367 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
368 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
369 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
370 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
371 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
372 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
373 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
374 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
375 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
384 // Long long helper functions
385 // RTABI chapter 4.2, Table 9
386 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
387 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
388 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
389 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
390 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
397 // Integer division functions
398 // RTABI chapter 4.3.1
399 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
400 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
401 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
402 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
403 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
404 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
405 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
407 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
417 // RTABI chapter 4.3.4
418 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
419 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
420 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
421 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
422 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
423 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
426 // Use divmod compiler-rt calls for iOS 5.0 and later.
427 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
428 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
429 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
430 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
433 if (Subtarget->isThumb1Only())
434 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
436 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
437 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
438 !Subtarget->isThumb1Only()) {
439 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
440 if (!Subtarget->isFPOnlySP())
441 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
443 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
446 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
447 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
448 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
449 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
450 setTruncStoreAction((MVT::SimpleValueType)VT,
451 (MVT::SimpleValueType)InnerVT, Expand);
452 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
453 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
457 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
459 if (Subtarget->hasNEON()) {
460 addDRTypeForNEON(MVT::v2f32);
461 addDRTypeForNEON(MVT::v8i8);
462 addDRTypeForNEON(MVT::v4i16);
463 addDRTypeForNEON(MVT::v2i32);
464 addDRTypeForNEON(MVT::v1i64);
466 addQRTypeForNEON(MVT::v4f32);
467 addQRTypeForNEON(MVT::v2f64);
468 addQRTypeForNEON(MVT::v16i8);
469 addQRTypeForNEON(MVT::v8i16);
470 addQRTypeForNEON(MVT::v4i32);
471 addQRTypeForNEON(MVT::v2i64);
473 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
474 // neither Neon nor VFP support any arithmetic operations on it.
475 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
476 // supported for v4f32.
477 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
478 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
479 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
480 // FIXME: Code duplication: FDIV and FREM are expanded always, see
481 // ARMTargetLowering::addTypeForNEON method for details.
482 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
483 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
484 // FIXME: Create unittest.
485 // In another words, find a way when "copysign" appears in DAG with vector
487 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
488 // FIXME: Code duplication: SETCC has custom operation action, see
489 // ARMTargetLowering::addTypeForNEON method for details.
490 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
491 // FIXME: Create unittest for FNEG and for FABS.
492 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
493 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
495 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
496 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
497 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
498 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
499 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
500 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
501 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
502 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
503 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
504 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
505 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
506 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
507 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
508 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
509 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
511 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
512 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
513 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
514 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
515 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
518 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
519 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
520 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
522 // Neon does not support some operations on v1i64 and v2i64 types.
523 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
524 // Custom handling for some quad-vector types to detect VMULL.
525 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
526 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
527 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
528 // Custom handling for some vector types to avoid expensive expansions
529 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
532 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
533 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
534 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
535 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
536 // a destination type that is wider than the source, and nor does
537 // it have a FP_TO_[SU]INT instruction with a narrower destination than
539 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
540 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
541 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
542 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
544 setTargetDAGCombine(ISD::INTRINSIC_VOID);
545 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
546 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
547 setTargetDAGCombine(ISD::SHL);
548 setTargetDAGCombine(ISD::SRL);
549 setTargetDAGCombine(ISD::SRA);
550 setTargetDAGCombine(ISD::SIGN_EXTEND);
551 setTargetDAGCombine(ISD::ZERO_EXTEND);
552 setTargetDAGCombine(ISD::ANY_EXTEND);
553 setTargetDAGCombine(ISD::SELECT_CC);
554 setTargetDAGCombine(ISD::BUILD_VECTOR);
555 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
556 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
557 setTargetDAGCombine(ISD::STORE);
558 setTargetDAGCombine(ISD::FP_TO_SINT);
559 setTargetDAGCombine(ISD::FP_TO_UINT);
560 setTargetDAGCombine(ISD::FDIV);
562 // It is legal to extload from v4i8 to v4i16 or v4i32.
563 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
564 MVT::v4i16, MVT::v2i16,
566 for (unsigned i = 0; i < 6; ++i) {
567 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
568 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
569 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
573 computeRegisterProperties();
575 // ARM does not have f32 extending load.
576 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
578 // ARM does not have i1 sign extending load.
579 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
581 // ARM supports all 4 flavors of integer indexed load / store.
582 if (!Subtarget->isThumb1Only()) {
583 for (unsigned im = (unsigned)ISD::PRE_INC;
584 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
585 setIndexedLoadAction(im, MVT::i1, Legal);
586 setIndexedLoadAction(im, MVT::i8, Legal);
587 setIndexedLoadAction(im, MVT::i16, Legal);
588 setIndexedLoadAction(im, MVT::i32, Legal);
589 setIndexedStoreAction(im, MVT::i1, Legal);
590 setIndexedStoreAction(im, MVT::i8, Legal);
591 setIndexedStoreAction(im, MVT::i16, Legal);
592 setIndexedStoreAction(im, MVT::i32, Legal);
596 // i64 operation support.
597 setOperationAction(ISD::MUL, MVT::i64, Expand);
598 setOperationAction(ISD::MULHU, MVT::i32, Expand);
599 if (Subtarget->isThumb1Only()) {
600 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
601 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
603 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
604 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
605 setOperationAction(ISD::MULHS, MVT::i32, Expand);
607 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
608 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
609 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
610 setOperationAction(ISD::SRL, MVT::i64, Custom);
611 setOperationAction(ISD::SRA, MVT::i64, Custom);
613 if (!Subtarget->isThumb1Only()) {
614 // FIXME: We should do this for Thumb1 as well.
615 setOperationAction(ISD::ADDC, MVT::i32, Custom);
616 setOperationAction(ISD::ADDE, MVT::i32, Custom);
617 setOperationAction(ISD::SUBC, MVT::i32, Custom);
618 setOperationAction(ISD::SUBE, MVT::i32, Custom);
621 // ARM does not have ROTL.
622 setOperationAction(ISD::ROTL, MVT::i32, Expand);
623 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
624 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
625 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
626 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
628 // These just redirect to CTTZ and CTLZ on ARM.
629 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
630 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
632 // Only ARMv6 has BSWAP.
633 if (!Subtarget->hasV6Ops())
634 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
636 // These are expanded into libcalls.
637 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
638 // v7M has a hardware divider
639 setOperationAction(ISD::SDIV, MVT::i32, Expand);
640 setOperationAction(ISD::UDIV, MVT::i32, Expand);
642 setOperationAction(ISD::SREM, MVT::i32, Expand);
643 setOperationAction(ISD::UREM, MVT::i32, Expand);
644 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
645 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
647 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
648 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
649 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
650 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
651 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
653 setOperationAction(ISD::TRAP, MVT::Other, Legal);
655 // Use the default implementation.
656 setOperationAction(ISD::VASTART, MVT::Other, Custom);
657 setOperationAction(ISD::VAARG, MVT::Other, Expand);
658 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
659 setOperationAction(ISD::VAEND, MVT::Other, Expand);
660 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
661 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
663 if (!Subtarget->isTargetDarwin()) {
664 // Non-Darwin platforms may return values in these registers via the
665 // personality function.
666 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
667 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
668 setExceptionPointerRegister(ARM::R0);
669 setExceptionSelectorRegister(ARM::R1);
672 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
673 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
674 // the default expansion.
675 // FIXME: This should be checking for v6k, not just v6.
676 if (Subtarget->hasDataBarrier() ||
677 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
678 // membarrier needs custom lowering; the rest are legal and handled
680 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
681 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
682 // Custom lowering for 64-bit ops
683 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
684 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
685 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
686 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
690 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
691 setInsertFencesForAtomic(true);
693 // Set them all for expansion, which will force libcalls.
694 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
695 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
696 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
697 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
698 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
699 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
708 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
709 // Unordered/Monotonic case.
710 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
711 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
712 // Since the libcalls include locking, fold in the fences
713 setShouldFoldAtomicFences(true);
716 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
718 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
719 if (!Subtarget->hasV6Ops()) {
720 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
721 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
725 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
726 !Subtarget->isThumb1Only()) {
727 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
728 // iff target supports vfp2.
729 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
730 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
733 // We want to custom lower some of our intrinsics.
734 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
735 if (Subtarget->isTargetDarwin()) {
736 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
737 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
738 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
741 setOperationAction(ISD::SETCC, MVT::i32, Expand);
742 setOperationAction(ISD::SETCC, MVT::f32, Expand);
743 setOperationAction(ISD::SETCC, MVT::f64, Expand);
744 setOperationAction(ISD::SELECT, MVT::i32, Custom);
745 setOperationAction(ISD::SELECT, MVT::f32, Custom);
746 setOperationAction(ISD::SELECT, MVT::f64, Custom);
747 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
748 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
749 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
751 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
752 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
753 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
754 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
755 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
757 // We don't support sin/cos/fmod/copysign/pow
758 setOperationAction(ISD::FSIN, MVT::f64, Expand);
759 setOperationAction(ISD::FSIN, MVT::f32, Expand);
760 setOperationAction(ISD::FCOS, MVT::f32, Expand);
761 setOperationAction(ISD::FCOS, MVT::f64, Expand);
762 setOperationAction(ISD::FREM, MVT::f64, Expand);
763 setOperationAction(ISD::FREM, MVT::f32, Expand);
764 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
765 !Subtarget->isThumb1Only()) {
766 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
767 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
769 setOperationAction(ISD::FPOW, MVT::f64, Expand);
770 setOperationAction(ISD::FPOW, MVT::f32, Expand);
772 if (!Subtarget->hasVFP4()) {
773 setOperationAction(ISD::FMA, MVT::f64, Expand);
774 setOperationAction(ISD::FMA, MVT::f32, Expand);
777 // Various VFP goodness
778 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
779 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
780 if (Subtarget->hasVFP2()) {
781 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
782 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
783 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
784 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
786 // Special handling for half-precision FP.
787 if (!Subtarget->hasFP16()) {
788 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
789 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
793 // We have target-specific dag combine patterns for the following nodes:
794 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
795 setTargetDAGCombine(ISD::ADD);
796 setTargetDAGCombine(ISD::SUB);
797 setTargetDAGCombine(ISD::MUL);
799 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
800 setTargetDAGCombine(ISD::AND);
801 setTargetDAGCombine(ISD::OR);
802 setTargetDAGCombine(ISD::XOR);
805 if (Subtarget->hasV6Ops())
806 setTargetDAGCombine(ISD::SRL);
808 setStackPointerRegisterToSaveRestore(ARM::SP);
810 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
811 !Subtarget->hasVFP2())
812 setSchedulingPreference(Sched::RegPressure);
814 setSchedulingPreference(Sched::Hybrid);
816 //// temporary - rewrite interface to use type
817 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
818 maxStoresPerMemset = 16;
819 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
821 // On ARM arguments smaller than 4 bytes are extended, so all arguments
822 // are at least 4 bytes aligned.
823 setMinStackArgumentAlignment(4);
825 benefitFromCodePlacementOpt = true;
827 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
830 // FIXME: It might make sense to define the representative register class as the
831 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
832 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
833 // SPR's representative would be DPR_VFP2. This should work well if register
834 // pressure tracking were modified such that a register use would increment the
835 // pressure of the register class's representative and all of it's super
836 // classes' representatives transitively. We have not implemented this because
837 // of the difficulty prior to coalescing of modeling operand register classes
838 // due to the common occurrence of cross class copies and subregister insertions
840 std::pair<const TargetRegisterClass*, uint8_t>
841 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
842 const TargetRegisterClass *RRC = 0;
844 switch (VT.getSimpleVT().SimpleTy) {
846 return TargetLowering::findRepresentativeClass(VT);
847 // Use DPR as representative register class for all floating point
848 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
849 // the cost is 1 for both f32 and f64.
850 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
851 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
852 RRC = &ARM::DPRRegClass;
853 // When NEON is used for SP, only half of the register file is available
854 // because operations that define both SP and DP results will be constrained
855 // to the VFP2 class (D0-D15). We currently model this constraint prior to
856 // coalescing by double-counting the SP regs. See the FIXME above.
857 if (Subtarget->useNEONForSinglePrecisionFP())
860 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
861 case MVT::v4f32: case MVT::v2f64:
862 RRC = &ARM::DPRRegClass;
866 RRC = &ARM::DPRRegClass;
870 RRC = &ARM::DPRRegClass;
874 return std::make_pair(RRC, Cost);
877 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
880 case ARMISD::Wrapper: return "ARMISD::Wrapper";
881 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
882 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
883 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
884 case ARMISD::CALL: return "ARMISD::CALL";
885 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
886 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
887 case ARMISD::tCALL: return "ARMISD::tCALL";
888 case ARMISD::BRCOND: return "ARMISD::BRCOND";
889 case ARMISD::BR_JT: return "ARMISD::BR_JT";
890 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
891 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
892 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
893 case ARMISD::CMP: return "ARMISD::CMP";
894 case ARMISD::CMPZ: return "ARMISD::CMPZ";
895 case ARMISD::CMPFP: return "ARMISD::CMPFP";
896 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
897 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
898 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
900 case ARMISD::CMOV: return "ARMISD::CMOV";
901 case ARMISD::CAND: return "ARMISD::CAND";
902 case ARMISD::COR: return "ARMISD::COR";
903 case ARMISD::CXOR: return "ARMISD::CXOR";
905 case ARMISD::RBIT: return "ARMISD::RBIT";
907 case ARMISD::FTOSI: return "ARMISD::FTOSI";
908 case ARMISD::FTOUI: return "ARMISD::FTOUI";
909 case ARMISD::SITOF: return "ARMISD::SITOF";
910 case ARMISD::UITOF: return "ARMISD::UITOF";
912 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
913 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
914 case ARMISD::RRX: return "ARMISD::RRX";
916 case ARMISD::ADDC: return "ARMISD::ADDC";
917 case ARMISD::ADDE: return "ARMISD::ADDE";
918 case ARMISD::SUBC: return "ARMISD::SUBC";
919 case ARMISD::SUBE: return "ARMISD::SUBE";
921 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
922 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
924 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
925 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
927 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
929 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
931 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
933 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
934 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
936 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
938 case ARMISD::VCEQ: return "ARMISD::VCEQ";
939 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
940 case ARMISD::VCGE: return "ARMISD::VCGE";
941 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
942 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
943 case ARMISD::VCGEU: return "ARMISD::VCGEU";
944 case ARMISD::VCGT: return "ARMISD::VCGT";
945 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
946 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
947 case ARMISD::VCGTU: return "ARMISD::VCGTU";
948 case ARMISD::VTST: return "ARMISD::VTST";
950 case ARMISD::VSHL: return "ARMISD::VSHL";
951 case ARMISD::VSHRs: return "ARMISD::VSHRs";
952 case ARMISD::VSHRu: return "ARMISD::VSHRu";
953 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
954 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
955 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
956 case ARMISD::VSHRN: return "ARMISD::VSHRN";
957 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
958 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
959 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
960 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
961 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
962 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
963 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
964 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
965 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
966 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
967 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
968 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
969 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
970 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
971 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
972 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
973 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
974 case ARMISD::VDUP: return "ARMISD::VDUP";
975 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
976 case ARMISD::VEXT: return "ARMISD::VEXT";
977 case ARMISD::VREV64: return "ARMISD::VREV64";
978 case ARMISD::VREV32: return "ARMISD::VREV32";
979 case ARMISD::VREV16: return "ARMISD::VREV16";
980 case ARMISD::VZIP: return "ARMISD::VZIP";
981 case ARMISD::VUZP: return "ARMISD::VUZP";
982 case ARMISD::VTRN: return "ARMISD::VTRN";
983 case ARMISD::VTBL1: return "ARMISD::VTBL1";
984 case ARMISD::VTBL2: return "ARMISD::VTBL2";
985 case ARMISD::VMULLs: return "ARMISD::VMULLs";
986 case ARMISD::VMULLu: return "ARMISD::VMULLu";
987 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
988 case ARMISD::FMAX: return "ARMISD::FMAX";
989 case ARMISD::FMIN: return "ARMISD::FMIN";
990 case ARMISD::BFI: return "ARMISD::BFI";
991 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
992 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
993 case ARMISD::VBSL: return "ARMISD::VBSL";
994 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
995 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
996 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
997 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
998 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
999 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1000 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1001 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1002 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1003 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1004 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1005 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1006 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1007 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1008 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1009 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1010 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1011 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1012 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1013 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1017 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1018 if (!VT.isVector()) return getPointerTy();
1019 return VT.changeVectorElementTypeToInteger();
1022 /// getRegClassFor - Return the register class that should be used for the
1023 /// specified value type.
1024 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1025 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1026 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1027 // load / store 4 to 8 consecutive D registers.
1028 if (Subtarget->hasNEON()) {
1029 if (VT == MVT::v4i64)
1030 return &ARM::QQPRRegClass;
1031 if (VT == MVT::v8i64)
1032 return &ARM::QQQQPRRegClass;
1034 return TargetLowering::getRegClassFor(VT);
1037 // Create a fast isel object.
1039 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1040 return ARM::createFastISel(funcInfo);
1043 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1044 /// be used for loads / stores from the global.
1045 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1046 return (Subtarget->isThumb1Only() ? 127 : 4095);
1049 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1050 unsigned NumVals = N->getNumValues();
1052 return Sched::RegPressure;
1054 for (unsigned i = 0; i != NumVals; ++i) {
1055 EVT VT = N->getValueType(i);
1056 if (VT == MVT::Glue || VT == MVT::Other)
1058 if (VT.isFloatingPoint() || VT.isVector())
1062 if (!N->isMachineOpcode())
1063 return Sched::RegPressure;
1065 // Load are scheduled for latency even if there instruction itinerary
1066 // is not available.
1067 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1068 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1070 if (MCID.getNumDefs() == 0)
1071 return Sched::RegPressure;
1072 if (!Itins->isEmpty() &&
1073 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1076 return Sched::RegPressure;
1079 //===----------------------------------------------------------------------===//
1081 //===----------------------------------------------------------------------===//
1083 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1084 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1086 default: llvm_unreachable("Unknown condition code!");
1087 case ISD::SETNE: return ARMCC::NE;
1088 case ISD::SETEQ: return ARMCC::EQ;
1089 case ISD::SETGT: return ARMCC::GT;
1090 case ISD::SETGE: return ARMCC::GE;
1091 case ISD::SETLT: return ARMCC::LT;
1092 case ISD::SETLE: return ARMCC::LE;
1093 case ISD::SETUGT: return ARMCC::HI;
1094 case ISD::SETUGE: return ARMCC::HS;
1095 case ISD::SETULT: return ARMCC::LO;
1096 case ISD::SETULE: return ARMCC::LS;
1100 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1101 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1102 ARMCC::CondCodes &CondCode2) {
1103 CondCode2 = ARMCC::AL;
1105 default: llvm_unreachable("Unknown FP condition!");
1107 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1109 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1111 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1112 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1113 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1114 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1115 case ISD::SETO: CondCode = ARMCC::VC; break;
1116 case ISD::SETUO: CondCode = ARMCC::VS; break;
1117 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1118 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1119 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1121 case ISD::SETULT: CondCode = ARMCC::LT; break;
1123 case ISD::SETULE: CondCode = ARMCC::LE; break;
1125 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1129 //===----------------------------------------------------------------------===//
1130 // Calling Convention Implementation
1131 //===----------------------------------------------------------------------===//
1133 #include "ARMGenCallingConv.inc"
1135 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1136 /// given CallingConvention value.
1137 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1139 bool isVarArg) const {
1142 llvm_unreachable("Unsupported calling convention");
1143 case CallingConv::Fast:
1144 if (Subtarget->hasVFP2() && !isVarArg) {
1145 if (!Subtarget->isAAPCS_ABI())
1146 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1147 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1148 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1151 case CallingConv::C: {
1152 // Use target triple & subtarget features to do actual dispatch.
1153 if (!Subtarget->isAAPCS_ABI())
1154 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1155 else if (Subtarget->hasVFP2() &&
1156 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1158 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1159 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1161 case CallingConv::ARM_AAPCS_VFP:
1163 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1165 case CallingConv::ARM_AAPCS:
1166 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1167 case CallingConv::ARM_APCS:
1168 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1172 /// LowerCallResult - Lower the result values of a call into the
1173 /// appropriate copies out of appropriate physical registers.
1175 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1176 CallingConv::ID CallConv, bool isVarArg,
1177 const SmallVectorImpl<ISD::InputArg> &Ins,
1178 DebugLoc dl, SelectionDAG &DAG,
1179 SmallVectorImpl<SDValue> &InVals) const {
1181 // Assign locations to each value returned by this call.
1182 SmallVector<CCValAssign, 16> RVLocs;
1183 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1184 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1185 CCInfo.AnalyzeCallResult(Ins,
1186 CCAssignFnForNode(CallConv, /* Return*/ true,
1189 // Copy all of the result registers out of their specified physreg.
1190 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1191 CCValAssign VA = RVLocs[i];
1194 if (VA.needsCustom()) {
1195 // Handle f64 or half of a v2f64.
1196 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1198 Chain = Lo.getValue(1);
1199 InFlag = Lo.getValue(2);
1200 VA = RVLocs[++i]; // skip ahead to next loc
1201 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1203 Chain = Hi.getValue(1);
1204 InFlag = Hi.getValue(2);
1205 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1207 if (VA.getLocVT() == MVT::v2f64) {
1208 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1209 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1210 DAG.getConstant(0, MVT::i32));
1212 VA = RVLocs[++i]; // skip ahead to next loc
1213 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1214 Chain = Lo.getValue(1);
1215 InFlag = Lo.getValue(2);
1216 VA = RVLocs[++i]; // skip ahead to next loc
1217 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1218 Chain = Hi.getValue(1);
1219 InFlag = Hi.getValue(2);
1220 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1221 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1222 DAG.getConstant(1, MVT::i32));
1225 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1227 Chain = Val.getValue(1);
1228 InFlag = Val.getValue(2);
1231 switch (VA.getLocInfo()) {
1232 default: llvm_unreachable("Unknown loc info!");
1233 case CCValAssign::Full: break;
1234 case CCValAssign::BCvt:
1235 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1239 InVals.push_back(Val);
1245 /// LowerMemOpCallTo - Store the argument to the stack.
1247 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1248 SDValue StackPtr, SDValue Arg,
1249 DebugLoc dl, SelectionDAG &DAG,
1250 const CCValAssign &VA,
1251 ISD::ArgFlagsTy Flags) const {
1252 unsigned LocMemOffset = VA.getLocMemOffset();
1253 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1254 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1255 return DAG.getStore(Chain, dl, Arg, PtrOff,
1256 MachinePointerInfo::getStack(LocMemOffset),
1260 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1261 SDValue Chain, SDValue &Arg,
1262 RegsToPassVector &RegsToPass,
1263 CCValAssign &VA, CCValAssign &NextVA,
1265 SmallVector<SDValue, 8> &MemOpChains,
1266 ISD::ArgFlagsTy Flags) const {
1268 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1269 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1270 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1272 if (NextVA.isRegLoc())
1273 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1275 assert(NextVA.isMemLoc());
1276 if (StackPtr.getNode() == 0)
1277 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1279 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1285 /// LowerCall - Lowering a call into a callseq_start <-
1286 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1289 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1290 CallingConv::ID CallConv, bool isVarArg,
1291 bool doesNotRet, bool &isTailCall,
1292 const SmallVectorImpl<ISD::OutputArg> &Outs,
1293 const SmallVectorImpl<SDValue> &OutVals,
1294 const SmallVectorImpl<ISD::InputArg> &Ins,
1295 DebugLoc dl, SelectionDAG &DAG,
1296 SmallVectorImpl<SDValue> &InVals) const {
1297 MachineFunction &MF = DAG.getMachineFunction();
1298 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1299 bool IsSibCall = false;
1300 // Disable tail calls if they're not supported.
1301 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1304 // Check if it's really possible to do a tail call.
1305 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1306 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1307 Outs, OutVals, Ins, DAG);
1308 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1309 // detected sibcalls.
1316 // Analyze operands of the call, assigning locations to each operand.
1317 SmallVector<CCValAssign, 16> ArgLocs;
1318 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1319 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1320 CCInfo.AnalyzeCallOperands(Outs,
1321 CCAssignFnForNode(CallConv, /* Return*/ false,
1324 // Get a count of how many bytes are to be pushed on the stack.
1325 unsigned NumBytes = CCInfo.getNextStackOffset();
1327 // For tail calls, memory operands are available in our caller's stack.
1331 // Adjust the stack pointer for the new arguments...
1332 // These operations are automatically eliminated by the prolog/epilog pass
1334 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1336 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1338 RegsToPassVector RegsToPass;
1339 SmallVector<SDValue, 8> MemOpChains;
1341 // Walk the register/memloc assignments, inserting copies/loads. In the case
1342 // of tail call optimization, arguments are handled later.
1343 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1345 ++i, ++realArgIdx) {
1346 CCValAssign &VA = ArgLocs[i];
1347 SDValue Arg = OutVals[realArgIdx];
1348 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1349 bool isByVal = Flags.isByVal();
1351 // Promote the value if needed.
1352 switch (VA.getLocInfo()) {
1353 default: llvm_unreachable("Unknown loc info!");
1354 case CCValAssign::Full: break;
1355 case CCValAssign::SExt:
1356 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1358 case CCValAssign::ZExt:
1359 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1361 case CCValAssign::AExt:
1362 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1364 case CCValAssign::BCvt:
1365 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1369 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1370 if (VA.needsCustom()) {
1371 if (VA.getLocVT() == MVT::v2f64) {
1372 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1373 DAG.getConstant(0, MVT::i32));
1374 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1375 DAG.getConstant(1, MVT::i32));
1377 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1378 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1380 VA = ArgLocs[++i]; // skip ahead to next loc
1381 if (VA.isRegLoc()) {
1382 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1383 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1385 assert(VA.isMemLoc());
1387 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1388 dl, DAG, VA, Flags));
1391 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1392 StackPtr, MemOpChains, Flags);
1394 } else if (VA.isRegLoc()) {
1395 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1396 } else if (isByVal) {
1397 assert(VA.isMemLoc());
1398 unsigned offset = 0;
1400 // True if this byval aggregate will be split between registers
1402 if (CCInfo.isFirstByValRegValid()) {
1403 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1405 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1406 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1407 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1408 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1409 MachinePointerInfo(),
1410 false, false, false, 0);
1411 MemOpChains.push_back(Load.getValue(1));
1412 RegsToPass.push_back(std::make_pair(j, Load));
1414 offset = ARM::R4 - CCInfo.getFirstByValReg();
1415 CCInfo.clearFirstByValReg();
1418 unsigned LocMemOffset = VA.getLocMemOffset();
1419 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1420 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1422 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1423 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1424 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1426 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1427 Flags.getByValAlign(),
1428 /*isVolatile=*/false,
1429 /*AlwaysInline=*/false,
1430 MachinePointerInfo(0),
1431 MachinePointerInfo(0)));
1433 } else if (!IsSibCall) {
1434 assert(VA.isMemLoc());
1436 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1437 dl, DAG, VA, Flags));
1441 if (!MemOpChains.empty())
1442 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1443 &MemOpChains[0], MemOpChains.size());
1445 // Build a sequence of copy-to-reg nodes chained together with token chain
1446 // and flag operands which copy the outgoing args into the appropriate regs.
1448 // Tail call byval lowering might overwrite argument registers so in case of
1449 // tail call optimization the copies to registers are lowered later.
1451 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1452 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1453 RegsToPass[i].second, InFlag);
1454 InFlag = Chain.getValue(1);
1457 // For tail calls lower the arguments to the 'real' stack slot.
1459 // Force all the incoming stack arguments to be loaded from the stack
1460 // before any new outgoing arguments are stored to the stack, because the
1461 // outgoing stack slots may alias the incoming argument stack slots, and
1462 // the alias isn't otherwise explicit. This is slightly more conservative
1463 // than necessary, because it means that each store effectively depends
1464 // on every argument instead of just those arguments it would clobber.
1466 // Do not flag preceding copytoreg stuff together with the following stuff.
1468 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1469 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1470 RegsToPass[i].second, InFlag);
1471 InFlag = Chain.getValue(1);
1476 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1477 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1478 // node so that legalize doesn't hack it.
1479 bool isDirect = false;
1480 bool isARMFunc = false;
1481 bool isLocalARMFunc = false;
1482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1484 if (EnableARMLongCalls) {
1485 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1486 && "long-calls with non-static relocation model!");
1487 // Handle a global address or an external symbol. If it's not one of
1488 // those, the target's already in a register, so we don't need to do
1490 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1491 const GlobalValue *GV = G->getGlobal();
1492 // Create a constant pool entry for the callee address
1493 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1494 ARMConstantPoolValue *CPV =
1495 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1497 // Get the address of the callee into a register
1498 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1499 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1500 Callee = DAG.getLoad(getPointerTy(), dl,
1501 DAG.getEntryNode(), CPAddr,
1502 MachinePointerInfo::getConstantPool(),
1503 false, false, false, 0);
1504 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1505 const char *Sym = S->getSymbol();
1507 // Create a constant pool entry for the callee address
1508 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1509 ARMConstantPoolValue *CPV =
1510 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1511 ARMPCLabelIndex, 0);
1512 // Get the address of the callee into a register
1513 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1514 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1515 Callee = DAG.getLoad(getPointerTy(), dl,
1516 DAG.getEntryNode(), CPAddr,
1517 MachinePointerInfo::getConstantPool(),
1518 false, false, false, 0);
1520 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1521 const GlobalValue *GV = G->getGlobal();
1523 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1524 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1525 getTargetMachine().getRelocationModel() != Reloc::Static;
1526 isARMFunc = !Subtarget->isThumb() || isStub;
1527 // ARM call to a local ARM function is predicable.
1528 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1529 // tBX takes a register source operand.
1530 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1531 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1532 ARMConstantPoolValue *CPV =
1533 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1534 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1535 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1536 Callee = DAG.getLoad(getPointerTy(), dl,
1537 DAG.getEntryNode(), CPAddr,
1538 MachinePointerInfo::getConstantPool(),
1539 false, false, false, 0);
1540 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1541 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1542 getPointerTy(), Callee, PICLabel);
1544 // On ELF targets for PIC code, direct calls should go through the PLT
1545 unsigned OpFlags = 0;
1546 if (Subtarget->isTargetELF() &&
1547 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1548 OpFlags = ARMII::MO_PLT;
1549 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1551 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1553 bool isStub = Subtarget->isTargetDarwin() &&
1554 getTargetMachine().getRelocationModel() != Reloc::Static;
1555 isARMFunc = !Subtarget->isThumb() || isStub;
1556 // tBX takes a register source operand.
1557 const char *Sym = S->getSymbol();
1558 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1559 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1560 ARMConstantPoolValue *CPV =
1561 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1562 ARMPCLabelIndex, 4);
1563 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1564 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1565 Callee = DAG.getLoad(getPointerTy(), dl,
1566 DAG.getEntryNode(), CPAddr,
1567 MachinePointerInfo::getConstantPool(),
1568 false, false, false, 0);
1569 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1570 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1571 getPointerTy(), Callee, PICLabel);
1573 unsigned OpFlags = 0;
1574 // On ELF targets for PIC code, direct calls should go through the PLT
1575 if (Subtarget->isTargetELF() &&
1576 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1577 OpFlags = ARMII::MO_PLT;
1578 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1582 // FIXME: handle tail calls differently.
1584 if (Subtarget->isThumb()) {
1585 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1586 CallOpc = ARMISD::CALL_NOLINK;
1587 else if (doesNotRet && isDirect && !isARMFunc &&
1588 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1589 // "mov lr, pc; b _foo" to avoid confusing the RSP
1590 CallOpc = ARMISD::CALL_NOLINK;
1592 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1594 if (!isDirect && !Subtarget->hasV5TOps()) {
1595 CallOpc = ARMISD::CALL_NOLINK;
1596 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1597 // "mov lr, pc; b _foo" to avoid confusing the RSP
1598 CallOpc = ARMISD::CALL_NOLINK;
1600 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1603 std::vector<SDValue> Ops;
1604 Ops.push_back(Chain);
1605 Ops.push_back(Callee);
1607 // Add argument registers to the end of the list so that they are known live
1609 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1610 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1611 RegsToPass[i].second.getValueType()));
1613 // Add a register mask operand representing the call-preserved registers.
1614 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1615 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1616 assert(Mask && "Missing call preserved mask for calling convention");
1617 Ops.push_back(DAG.getRegisterMask(Mask));
1619 if (InFlag.getNode())
1620 Ops.push_back(InFlag);
1622 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1624 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1626 // Returns a chain and a flag for retval copy to use.
1627 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1628 InFlag = Chain.getValue(1);
1630 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1631 DAG.getIntPtrConstant(0, true), InFlag);
1633 InFlag = Chain.getValue(1);
1635 // Handle result values, copying them out of physregs into vregs that we
1637 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1641 /// HandleByVal - Every parameter *after* a byval parameter is passed
1642 /// on the stack. Remember the next parameter register to allocate,
1643 /// and then confiscate the rest of the parameter registers to insure
1646 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1647 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1648 assert((State->getCallOrPrologue() == Prologue ||
1649 State->getCallOrPrologue() == Call) &&
1650 "unhandled ParmContext");
1651 if ((!State->isFirstByValRegValid()) &&
1652 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1653 State->setFirstByValReg(reg);
1654 // At a call site, a byval parameter that is split between
1655 // registers and memory needs its size truncated here. In a
1656 // function prologue, such byval parameters are reassembled in
1657 // memory, and are not truncated.
1658 if (State->getCallOrPrologue() == Call) {
1659 unsigned excess = 4 * (ARM::R4 - reg);
1660 assert(size >= excess && "expected larger existing stack allocation");
1664 // Confiscate any remaining parameter registers to preclude their
1665 // assignment to subsequent parameters.
1666 while (State->AllocateReg(GPRArgRegs, 4))
1670 /// MatchingStackOffset - Return true if the given stack call argument is
1671 /// already available in the same position (relatively) of the caller's
1672 /// incoming argument stack.
1674 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1675 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1676 const TargetInstrInfo *TII) {
1677 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1679 if (Arg.getOpcode() == ISD::CopyFromReg) {
1680 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1681 if (!TargetRegisterInfo::isVirtualRegister(VR))
1683 MachineInstr *Def = MRI->getVRegDef(VR);
1686 if (!Flags.isByVal()) {
1687 if (!TII->isLoadFromStackSlot(Def, FI))
1692 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1693 if (Flags.isByVal())
1694 // ByVal argument is passed in as a pointer but it's now being
1695 // dereferenced. e.g.
1696 // define @foo(%struct.X* %A) {
1697 // tail call @bar(%struct.X* byval %A)
1700 SDValue Ptr = Ld->getBasePtr();
1701 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1704 FI = FINode->getIndex();
1708 assert(FI != INT_MAX);
1709 if (!MFI->isFixedObjectIndex(FI))
1711 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1714 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1715 /// for tail call optimization. Targets which want to do tail call
1716 /// optimization should implement this function.
1718 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1719 CallingConv::ID CalleeCC,
1721 bool isCalleeStructRet,
1722 bool isCallerStructRet,
1723 const SmallVectorImpl<ISD::OutputArg> &Outs,
1724 const SmallVectorImpl<SDValue> &OutVals,
1725 const SmallVectorImpl<ISD::InputArg> &Ins,
1726 SelectionDAG& DAG) const {
1727 const Function *CallerF = DAG.getMachineFunction().getFunction();
1728 CallingConv::ID CallerCC = CallerF->getCallingConv();
1729 bool CCMatch = CallerCC == CalleeCC;
1731 // Look for obvious safe cases to perform tail call optimization that do not
1732 // require ABI changes. This is what gcc calls sibcall.
1734 // Do not sibcall optimize vararg calls unless the call site is not passing
1736 if (isVarArg && !Outs.empty())
1739 // Also avoid sibcall optimization if either caller or callee uses struct
1740 // return semantics.
1741 if (isCalleeStructRet || isCallerStructRet)
1744 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1745 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1746 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1747 // support in the assembler and linker to be used. This would need to be
1748 // fixed to fully support tail calls in Thumb1.
1750 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1751 // LR. This means if we need to reload LR, it takes an extra instructions,
1752 // which outweighs the value of the tail call; but here we don't know yet
1753 // whether LR is going to be used. Probably the right approach is to
1754 // generate the tail call here and turn it back into CALL/RET in
1755 // emitEpilogue if LR is used.
1757 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1758 // but we need to make sure there are enough registers; the only valid
1759 // registers are the 4 used for parameters. We don't currently do this
1761 if (Subtarget->isThumb1Only())
1764 // If the calling conventions do not match, then we'd better make sure the
1765 // results are returned in the same way as what the caller expects.
1767 SmallVector<CCValAssign, 16> RVLocs1;
1768 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1769 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1770 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1772 SmallVector<CCValAssign, 16> RVLocs2;
1773 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1774 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1775 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1777 if (RVLocs1.size() != RVLocs2.size())
1779 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1780 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1782 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1784 if (RVLocs1[i].isRegLoc()) {
1785 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1788 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1794 // If the callee takes no arguments then go on to check the results of the
1796 if (!Outs.empty()) {
1797 // Check if stack adjustment is needed. For now, do not do this if any
1798 // argument is passed on the stack.
1799 SmallVector<CCValAssign, 16> ArgLocs;
1800 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1801 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1802 CCInfo.AnalyzeCallOperands(Outs,
1803 CCAssignFnForNode(CalleeCC, false, isVarArg));
1804 if (CCInfo.getNextStackOffset()) {
1805 MachineFunction &MF = DAG.getMachineFunction();
1807 // Check if the arguments are already laid out in the right way as
1808 // the caller's fixed stack objects.
1809 MachineFrameInfo *MFI = MF.getFrameInfo();
1810 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1811 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1812 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1814 ++i, ++realArgIdx) {
1815 CCValAssign &VA = ArgLocs[i];
1816 EVT RegVT = VA.getLocVT();
1817 SDValue Arg = OutVals[realArgIdx];
1818 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1819 if (VA.getLocInfo() == CCValAssign::Indirect)
1821 if (VA.needsCustom()) {
1822 // f64 and vector types are split into multiple registers or
1823 // register/stack-slot combinations. The types will not match
1824 // the registers; give up on memory f64 refs until we figure
1825 // out what to do about this.
1828 if (!ArgLocs[++i].isRegLoc())
1830 if (RegVT == MVT::v2f64) {
1831 if (!ArgLocs[++i].isRegLoc())
1833 if (!ArgLocs[++i].isRegLoc())
1836 } else if (!VA.isRegLoc()) {
1837 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1849 ARMTargetLowering::LowerReturn(SDValue Chain,
1850 CallingConv::ID CallConv, bool isVarArg,
1851 const SmallVectorImpl<ISD::OutputArg> &Outs,
1852 const SmallVectorImpl<SDValue> &OutVals,
1853 DebugLoc dl, SelectionDAG &DAG) const {
1855 // CCValAssign - represent the assignment of the return value to a location.
1856 SmallVector<CCValAssign, 16> RVLocs;
1858 // CCState - Info about the registers and stack slots.
1859 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1860 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1862 // Analyze outgoing return values.
1863 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1866 // If this is the first return lowered for this function, add
1867 // the regs to the liveout set for the function.
1868 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1869 for (unsigned i = 0; i != RVLocs.size(); ++i)
1870 if (RVLocs[i].isRegLoc())
1871 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1876 // Copy the result values into the output registers.
1877 for (unsigned i = 0, realRVLocIdx = 0;
1879 ++i, ++realRVLocIdx) {
1880 CCValAssign &VA = RVLocs[i];
1881 assert(VA.isRegLoc() && "Can only return in registers!");
1883 SDValue Arg = OutVals[realRVLocIdx];
1885 switch (VA.getLocInfo()) {
1886 default: llvm_unreachable("Unknown loc info!");
1887 case CCValAssign::Full: break;
1888 case CCValAssign::BCvt:
1889 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1893 if (VA.needsCustom()) {
1894 if (VA.getLocVT() == MVT::v2f64) {
1895 // Extract the first half and return it in two registers.
1896 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1897 DAG.getConstant(0, MVT::i32));
1898 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1899 DAG.getVTList(MVT::i32, MVT::i32), Half);
1901 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1902 Flag = Chain.getValue(1);
1903 VA = RVLocs[++i]; // skip ahead to next loc
1904 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1905 HalfGPRs.getValue(1), Flag);
1906 Flag = Chain.getValue(1);
1907 VA = RVLocs[++i]; // skip ahead to next loc
1909 // Extract the 2nd half and fall through to handle it as an f64 value.
1910 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1911 DAG.getConstant(1, MVT::i32));
1913 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1915 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1916 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1917 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1918 Flag = Chain.getValue(1);
1919 VA = RVLocs[++i]; // skip ahead to next loc
1920 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1923 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1925 // Guarantee that all emitted copies are
1926 // stuck together, avoiding something bad.
1927 Flag = Chain.getValue(1);
1932 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1934 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1939 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1940 if (N->getNumValues() != 1)
1942 if (!N->hasNUsesOfValue(1, 0))
1945 SDValue TCChain = Chain;
1946 SDNode *Copy = *N->use_begin();
1947 if (Copy->getOpcode() == ISD::CopyToReg) {
1948 // If the copy has a glue operand, we conservatively assume it isn't safe to
1949 // perform a tail call.
1950 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1952 TCChain = Copy->getOperand(0);
1953 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1954 SDNode *VMov = Copy;
1955 // f64 returned in a pair of GPRs.
1956 SmallPtrSet<SDNode*, 2> Copies;
1957 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1959 if (UI->getOpcode() != ISD::CopyToReg)
1963 if (Copies.size() > 2)
1966 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1968 SDValue UseChain = UI->getOperand(0);
1969 if (Copies.count(UseChain.getNode()))
1976 } else if (Copy->getOpcode() == ISD::BITCAST) {
1977 // f32 returned in a single GPR.
1978 if (!Copy->hasOneUse())
1980 Copy = *Copy->use_begin();
1981 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
1983 Chain = Copy->getOperand(0);
1988 bool HasRet = false;
1989 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1991 if (UI->getOpcode() != ARMISD::RET_FLAG)
2003 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2004 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2007 if (!CI->isTailCall())
2010 return !Subtarget->isThumb1Only();
2013 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2014 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2015 // one of the above mentioned nodes. It has to be wrapped because otherwise
2016 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2017 // be used to form addressing mode. These wrapped nodes will be selected
2019 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2020 EVT PtrVT = Op.getValueType();
2021 // FIXME there is no actual debug info here
2022 DebugLoc dl = Op.getDebugLoc();
2023 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2025 if (CP->isMachineConstantPoolEntry())
2026 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2027 CP->getAlignment());
2029 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2030 CP->getAlignment());
2031 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2034 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2035 return MachineJumpTableInfo::EK_Inline;
2038 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2039 SelectionDAG &DAG) const {
2040 MachineFunction &MF = DAG.getMachineFunction();
2041 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2042 unsigned ARMPCLabelIndex = 0;
2043 DebugLoc DL = Op.getDebugLoc();
2044 EVT PtrVT = getPointerTy();
2045 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2046 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2048 if (RelocM == Reloc::Static) {
2049 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2051 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2052 ARMPCLabelIndex = AFI->createPICLabelUId();
2053 ARMConstantPoolValue *CPV =
2054 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2055 ARMCP::CPBlockAddress, PCAdj);
2056 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2058 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2059 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2060 MachinePointerInfo::getConstantPool(),
2061 false, false, false, 0);
2062 if (RelocM == Reloc::Static)
2064 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2065 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2068 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2070 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2071 SelectionDAG &DAG) const {
2072 DebugLoc dl = GA->getDebugLoc();
2073 EVT PtrVT = getPointerTy();
2074 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2075 MachineFunction &MF = DAG.getMachineFunction();
2076 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2077 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2078 ARMConstantPoolValue *CPV =
2079 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2080 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2081 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2082 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2083 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2084 MachinePointerInfo::getConstantPool(),
2085 false, false, false, 0);
2086 SDValue Chain = Argument.getValue(1);
2088 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2089 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2091 // call __tls_get_addr.
2094 Entry.Node = Argument;
2095 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2096 Args.push_back(Entry);
2097 // FIXME: is there useful debug info available here?
2098 std::pair<SDValue, SDValue> CallResult =
2099 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2100 false, false, false, false,
2101 0, CallingConv::C, /*isTailCall=*/false,
2102 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2103 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2104 return CallResult.first;
2107 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2108 // "local exec" model.
2110 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2112 TLSModel::Model model) const {
2113 const GlobalValue *GV = GA->getGlobal();
2114 DebugLoc dl = GA->getDebugLoc();
2116 SDValue Chain = DAG.getEntryNode();
2117 EVT PtrVT = getPointerTy();
2118 // Get the Thread Pointer
2119 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2121 if (model == TLSModel::InitialExec) {
2122 MachineFunction &MF = DAG.getMachineFunction();
2123 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2124 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2125 // Initial exec model.
2126 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2127 ARMConstantPoolValue *CPV =
2128 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2129 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2131 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2132 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2133 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2134 MachinePointerInfo::getConstantPool(),
2135 false, false, false, 0);
2136 Chain = Offset.getValue(1);
2138 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2139 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2141 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2142 MachinePointerInfo::getConstantPool(),
2143 false, false, false, 0);
2146 assert(model == TLSModel::LocalExec);
2147 ARMConstantPoolValue *CPV =
2148 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2149 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2150 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2151 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2152 MachinePointerInfo::getConstantPool(),
2153 false, false, false, 0);
2156 // The address of the thread local variable is the add of the thread
2157 // pointer with the offset of the variable.
2158 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2162 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2163 // TODO: implement the "local dynamic" model
2164 assert(Subtarget->isTargetELF() &&
2165 "TLS not implemented for non-ELF targets");
2166 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2168 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2171 case TLSModel::GeneralDynamic:
2172 case TLSModel::LocalDynamic:
2173 return LowerToTLSGeneralDynamicModel(GA, DAG);
2174 case TLSModel::InitialExec:
2175 case TLSModel::LocalExec:
2176 return LowerToTLSExecModels(GA, DAG, model);
2178 llvm_unreachable("bogus TLS model");
2181 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2182 SelectionDAG &DAG) const {
2183 EVT PtrVT = getPointerTy();
2184 DebugLoc dl = Op.getDebugLoc();
2185 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2186 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2187 if (RelocM == Reloc::PIC_) {
2188 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2189 ARMConstantPoolValue *CPV =
2190 ARMConstantPoolConstant::Create(GV,
2191 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2192 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2193 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2194 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2196 MachinePointerInfo::getConstantPool(),
2197 false, false, false, 0);
2198 SDValue Chain = Result.getValue(1);
2199 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2200 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2202 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2203 MachinePointerInfo::getGOT(),
2204 false, false, false, 0);
2208 // If we have T2 ops, we can materialize the address directly via movt/movw
2209 // pair. This is always cheaper.
2210 if (Subtarget->useMovt()) {
2212 // FIXME: Once remat is capable of dealing with instructions with register
2213 // operands, expand this into two nodes.
2214 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2215 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2217 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2218 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2219 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2220 MachinePointerInfo::getConstantPool(),
2221 false, false, false, 0);
2225 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2226 SelectionDAG &DAG) const {
2227 EVT PtrVT = getPointerTy();
2228 DebugLoc dl = Op.getDebugLoc();
2229 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2230 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2231 MachineFunction &MF = DAG.getMachineFunction();
2232 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2234 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2235 // update ARMFastISel::ARMMaterializeGV.
2236 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2238 // FIXME: Once remat is capable of dealing with instructions with register
2239 // operands, expand this into two nodes.
2240 if (RelocM == Reloc::Static)
2241 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2242 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2244 unsigned Wrapper = (RelocM == Reloc::PIC_)
2245 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2246 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2247 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2248 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2249 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2250 MachinePointerInfo::getGOT(),
2251 false, false, false, 0);
2255 unsigned ARMPCLabelIndex = 0;
2257 if (RelocM == Reloc::Static) {
2258 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2260 ARMPCLabelIndex = AFI->createPICLabelUId();
2261 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2262 ARMConstantPoolValue *CPV =
2263 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2265 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2267 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2269 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2270 MachinePointerInfo::getConstantPool(),
2271 false, false, false, 0);
2272 SDValue Chain = Result.getValue(1);
2274 if (RelocM == Reloc::PIC_) {
2275 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2276 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2279 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2280 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2281 false, false, false, 0);
2286 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2287 SelectionDAG &DAG) const {
2288 assert(Subtarget->isTargetELF() &&
2289 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2290 MachineFunction &MF = DAG.getMachineFunction();
2291 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2292 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2293 EVT PtrVT = getPointerTy();
2294 DebugLoc dl = Op.getDebugLoc();
2295 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2296 ARMConstantPoolValue *CPV =
2297 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2298 ARMPCLabelIndex, PCAdj);
2299 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2300 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2301 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2302 MachinePointerInfo::getConstantPool(),
2303 false, false, false, 0);
2304 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2305 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2309 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2310 DebugLoc dl = Op.getDebugLoc();
2311 SDValue Val = DAG.getConstant(0, MVT::i32);
2312 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2313 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2314 Op.getOperand(1), Val);
2318 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2319 DebugLoc dl = Op.getDebugLoc();
2320 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2321 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2325 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2326 const ARMSubtarget *Subtarget) const {
2327 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2328 DebugLoc dl = Op.getDebugLoc();
2330 default: return SDValue(); // Don't custom lower most intrinsics.
2331 case Intrinsic::arm_thread_pointer: {
2332 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2333 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2335 case Intrinsic::eh_sjlj_lsda: {
2336 MachineFunction &MF = DAG.getMachineFunction();
2337 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2338 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2339 EVT PtrVT = getPointerTy();
2340 DebugLoc dl = Op.getDebugLoc();
2341 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2343 unsigned PCAdj = (RelocM != Reloc::PIC_)
2344 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2345 ARMConstantPoolValue *CPV =
2346 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2347 ARMCP::CPLSDA, PCAdj);
2348 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2349 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2351 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2352 MachinePointerInfo::getConstantPool(),
2353 false, false, false, 0);
2355 if (RelocM == Reloc::PIC_) {
2356 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2357 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2361 case Intrinsic::arm_neon_vmulls:
2362 case Intrinsic::arm_neon_vmullu: {
2363 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2364 ? ARMISD::VMULLs : ARMISD::VMULLu;
2365 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2366 Op.getOperand(1), Op.getOperand(2));
2371 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2372 const ARMSubtarget *Subtarget) {
2373 DebugLoc dl = Op.getDebugLoc();
2374 if (!Subtarget->hasDataBarrier()) {
2375 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2376 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2378 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2379 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2380 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2381 DAG.getConstant(0, MVT::i32));
2384 SDValue Op5 = Op.getOperand(5);
2385 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2386 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2387 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2388 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2390 ARM_MB::MemBOpt DMBOpt;
2391 if (isDeviceBarrier)
2392 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2394 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2395 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2396 DAG.getConstant(DMBOpt, MVT::i32));
2400 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2401 const ARMSubtarget *Subtarget) {
2402 // FIXME: handle "fence singlethread" more efficiently.
2403 DebugLoc dl = Op.getDebugLoc();
2404 if (!Subtarget->hasDataBarrier()) {
2405 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2406 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2408 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2409 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2410 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2411 DAG.getConstant(0, MVT::i32));
2414 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2415 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2418 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2419 const ARMSubtarget *Subtarget) {
2420 // ARM pre v5TE and Thumb1 does not have preload instructions.
2421 if (!(Subtarget->isThumb2() ||
2422 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2423 // Just preserve the chain.
2424 return Op.getOperand(0);
2426 DebugLoc dl = Op.getDebugLoc();
2427 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2429 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2430 // ARMv7 with MP extension has PLDW.
2431 return Op.getOperand(0);
2433 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2434 if (Subtarget->isThumb()) {
2436 isRead = ~isRead & 1;
2437 isData = ~isData & 1;
2440 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2441 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2442 DAG.getConstant(isData, MVT::i32));
2445 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2446 MachineFunction &MF = DAG.getMachineFunction();
2447 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2449 // vastart just stores the address of the VarArgsFrameIndex slot into the
2450 // memory location argument.
2451 DebugLoc dl = Op.getDebugLoc();
2452 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2453 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2454 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2455 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2456 MachinePointerInfo(SV), false, false, 0);
2460 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2461 SDValue &Root, SelectionDAG &DAG,
2462 DebugLoc dl) const {
2463 MachineFunction &MF = DAG.getMachineFunction();
2464 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2466 const TargetRegisterClass *RC;
2467 if (AFI->isThumb1OnlyFunction())
2468 RC = &ARM::tGPRRegClass;
2470 RC = &ARM::GPRRegClass;
2472 // Transform the arguments stored in physical registers into virtual ones.
2473 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2474 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2477 if (NextVA.isMemLoc()) {
2478 MachineFrameInfo *MFI = MF.getFrameInfo();
2479 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2481 // Create load node to retrieve arguments from the stack.
2482 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2483 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2484 MachinePointerInfo::getFixedStack(FI),
2485 false, false, false, 0);
2487 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2488 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2491 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2495 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2496 unsigned &VARegSize, unsigned &VARegSaveSize)
2499 if (CCInfo.isFirstByValRegValid())
2500 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2502 unsigned int firstUnalloced;
2503 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2504 sizeof(GPRArgRegs) /
2505 sizeof(GPRArgRegs[0]));
2506 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2509 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2510 VARegSize = NumGPRs * 4;
2511 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2514 // The remaining GPRs hold either the beginning of variable-argument
2515 // data, or the beginning of an aggregate passed by value (usuall
2516 // byval). Either way, we allocate stack slots adjacent to the data
2517 // provided by our caller, and store the unallocated registers there.
2518 // If this is a variadic function, the va_list pointer will begin with
2519 // these values; otherwise, this reassembles a (byval) structure that
2520 // was split between registers and memory.
2522 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2523 DebugLoc dl, SDValue &Chain,
2524 unsigned ArgOffset) const {
2525 MachineFunction &MF = DAG.getMachineFunction();
2526 MachineFrameInfo *MFI = MF.getFrameInfo();
2527 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2528 unsigned firstRegToSaveIndex;
2529 if (CCInfo.isFirstByValRegValid())
2530 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2532 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2533 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2536 unsigned VARegSize, VARegSaveSize;
2537 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2538 if (VARegSaveSize) {
2539 // If this function is vararg, store any remaining integer argument regs
2540 // to their spots on the stack so that they may be loaded by deferencing
2541 // the result of va_next.
2542 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2543 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2544 ArgOffset + VARegSaveSize
2547 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2550 SmallVector<SDValue, 4> MemOps;
2551 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2552 const TargetRegisterClass *RC;
2553 if (AFI->isThumb1OnlyFunction())
2554 RC = &ARM::tGPRRegClass;
2556 RC = &ARM::GPRRegClass;
2558 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2559 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2561 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2562 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2564 MemOps.push_back(Store);
2565 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2566 DAG.getConstant(4, getPointerTy()));
2568 if (!MemOps.empty())
2569 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2570 &MemOps[0], MemOps.size());
2572 // This will point to the next argument passed via stack.
2573 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2577 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2578 CallingConv::ID CallConv, bool isVarArg,
2579 const SmallVectorImpl<ISD::InputArg>
2581 DebugLoc dl, SelectionDAG &DAG,
2582 SmallVectorImpl<SDValue> &InVals)
2584 MachineFunction &MF = DAG.getMachineFunction();
2585 MachineFrameInfo *MFI = MF.getFrameInfo();
2587 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2589 // Assign locations to all of the incoming arguments.
2590 SmallVector<CCValAssign, 16> ArgLocs;
2591 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2592 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2593 CCInfo.AnalyzeFormalArguments(Ins,
2594 CCAssignFnForNode(CallConv, /* Return*/ false,
2597 SmallVector<SDValue, 16> ArgValues;
2598 int lastInsIndex = -1;
2601 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2602 CCValAssign &VA = ArgLocs[i];
2604 // Arguments stored in registers.
2605 if (VA.isRegLoc()) {
2606 EVT RegVT = VA.getLocVT();
2608 if (VA.needsCustom()) {
2609 // f64 and vector types are split up into multiple registers or
2610 // combinations of registers and stack slots.
2611 if (VA.getLocVT() == MVT::v2f64) {
2612 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2614 VA = ArgLocs[++i]; // skip ahead to next loc
2616 if (VA.isMemLoc()) {
2617 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2618 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2619 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2620 MachinePointerInfo::getFixedStack(FI),
2621 false, false, false, 0);
2623 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2626 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2627 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2628 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2629 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2630 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2632 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2635 const TargetRegisterClass *RC;
2637 if (RegVT == MVT::f32)
2638 RC = &ARM::SPRRegClass;
2639 else if (RegVT == MVT::f64)
2640 RC = &ARM::DPRRegClass;
2641 else if (RegVT == MVT::v2f64)
2642 RC = &ARM::QPRRegClass;
2643 else if (RegVT == MVT::i32)
2644 RC = AFI->isThumb1OnlyFunction() ?
2645 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2646 (const TargetRegisterClass*)&ARM::GPRRegClass;
2648 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2650 // Transform the arguments in physical registers into virtual ones.
2651 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2652 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2655 // If this is an 8 or 16-bit value, it is really passed promoted
2656 // to 32 bits. Insert an assert[sz]ext to capture this, then
2657 // truncate to the right size.
2658 switch (VA.getLocInfo()) {
2659 default: llvm_unreachable("Unknown loc info!");
2660 case CCValAssign::Full: break;
2661 case CCValAssign::BCvt:
2662 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2664 case CCValAssign::SExt:
2665 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2666 DAG.getValueType(VA.getValVT()));
2667 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2669 case CCValAssign::ZExt:
2670 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2671 DAG.getValueType(VA.getValVT()));
2672 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2676 InVals.push_back(ArgValue);
2678 } else { // VA.isRegLoc()
2681 assert(VA.isMemLoc());
2682 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2684 int index = ArgLocs[i].getValNo();
2686 // Some Ins[] entries become multiple ArgLoc[] entries.
2687 // Process them only once.
2688 if (index != lastInsIndex)
2690 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2691 // FIXME: For now, all byval parameter objects are marked mutable.
2692 // This can be changed with more analysis.
2693 // In case of tail call optimization mark all arguments mutable.
2694 // Since they could be overwritten by lowering of arguments in case of
2696 if (Flags.isByVal()) {
2697 unsigned VARegSize, VARegSaveSize;
2698 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2699 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2700 unsigned Bytes = Flags.getByValSize() - VARegSize;
2701 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2702 int FI = MFI->CreateFixedObject(Bytes,
2703 VA.getLocMemOffset(), false);
2704 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2706 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2707 VA.getLocMemOffset(), true);
2709 // Create load nodes to retrieve arguments from the stack.
2710 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2711 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2712 MachinePointerInfo::getFixedStack(FI),
2713 false, false, false, 0));
2715 lastInsIndex = index;
2722 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2727 /// isFloatingPointZero - Return true if this is +0.0.
2728 static bool isFloatingPointZero(SDValue Op) {
2729 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2730 return CFP->getValueAPF().isPosZero();
2731 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2732 // Maybe this has already been legalized into the constant pool?
2733 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2734 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2735 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2736 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2737 return CFP->getValueAPF().isPosZero();
2743 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2744 /// the given operands.
2746 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2747 SDValue &ARMcc, SelectionDAG &DAG,
2748 DebugLoc dl) const {
2749 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2750 unsigned C = RHSC->getZExtValue();
2751 if (!isLegalICmpImmediate(C)) {
2752 // Constant does not fit, try adjusting it by one?
2757 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2758 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2759 RHS = DAG.getConstant(C-1, MVT::i32);
2764 if (C != 0 && isLegalICmpImmediate(C-1)) {
2765 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2766 RHS = DAG.getConstant(C-1, MVT::i32);
2771 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2772 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2773 RHS = DAG.getConstant(C+1, MVT::i32);
2778 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2779 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2780 RHS = DAG.getConstant(C+1, MVT::i32);
2787 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2788 ARMISD::NodeType CompareType;
2791 CompareType = ARMISD::CMP;
2796 CompareType = ARMISD::CMPZ;
2799 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2800 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2803 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2805 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2806 DebugLoc dl) const {
2808 if (!isFloatingPointZero(RHS))
2809 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2811 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2812 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2815 /// duplicateCmp - Glue values can have only one use, so this function
2816 /// duplicates a comparison node.
2818 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2819 unsigned Opc = Cmp.getOpcode();
2820 DebugLoc DL = Cmp.getDebugLoc();
2821 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2822 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2824 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2825 Cmp = Cmp.getOperand(0);
2826 Opc = Cmp.getOpcode();
2827 if (Opc == ARMISD::CMPFP)
2828 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2830 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2831 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2833 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2836 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2837 SDValue Cond = Op.getOperand(0);
2838 SDValue SelectTrue = Op.getOperand(1);
2839 SDValue SelectFalse = Op.getOperand(2);
2840 DebugLoc dl = Op.getDebugLoc();
2844 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2845 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2847 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2848 const ConstantSDNode *CMOVTrue =
2849 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2850 const ConstantSDNode *CMOVFalse =
2851 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2853 if (CMOVTrue && CMOVFalse) {
2854 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2855 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2859 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2861 False = SelectFalse;
2862 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2867 if (True.getNode() && False.getNode()) {
2868 EVT VT = Op.getValueType();
2869 SDValue ARMcc = Cond.getOperand(2);
2870 SDValue CCR = Cond.getOperand(3);
2871 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2872 assert(True.getValueType() == VT);
2873 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2878 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2879 // undefined bits before doing a full-word comparison with zero.
2880 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2881 DAG.getConstant(1, Cond.getValueType()));
2883 return DAG.getSelectCC(dl, Cond,
2884 DAG.getConstant(0, Cond.getValueType()),
2885 SelectTrue, SelectFalse, ISD::SETNE);
2888 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2889 EVT VT = Op.getValueType();
2890 SDValue LHS = Op.getOperand(0);
2891 SDValue RHS = Op.getOperand(1);
2892 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2893 SDValue TrueVal = Op.getOperand(2);
2894 SDValue FalseVal = Op.getOperand(3);
2895 DebugLoc dl = Op.getDebugLoc();
2897 if (LHS.getValueType() == MVT::i32) {
2899 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2900 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2901 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2904 ARMCC::CondCodes CondCode, CondCode2;
2905 FPCCToARMCC(CC, CondCode, CondCode2);
2907 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2908 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2909 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2910 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2912 if (CondCode2 != ARMCC::AL) {
2913 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2914 // FIXME: Needs another CMP because flag can have but one use.
2915 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2916 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2917 Result, TrueVal, ARMcc2, CCR, Cmp2);
2922 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2923 /// to morph to an integer compare sequence.
2924 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2925 const ARMSubtarget *Subtarget) {
2926 SDNode *N = Op.getNode();
2927 if (!N->hasOneUse())
2928 // Otherwise it requires moving the value from fp to integer registers.
2930 if (!N->getNumValues())
2932 EVT VT = Op.getValueType();
2933 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2934 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2935 // vmrs are very slow, e.g. cortex-a8.
2938 if (isFloatingPointZero(Op)) {
2942 return ISD::isNormalLoad(N);
2945 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2946 if (isFloatingPointZero(Op))
2947 return DAG.getConstant(0, MVT::i32);
2949 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2950 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2951 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2952 Ld->isVolatile(), Ld->isNonTemporal(),
2953 Ld->isInvariant(), Ld->getAlignment());
2955 llvm_unreachable("Unknown VFP cmp argument!");
2958 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2959 SDValue &RetVal1, SDValue &RetVal2) {
2960 if (isFloatingPointZero(Op)) {
2961 RetVal1 = DAG.getConstant(0, MVT::i32);
2962 RetVal2 = DAG.getConstant(0, MVT::i32);
2966 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2967 SDValue Ptr = Ld->getBasePtr();
2968 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2969 Ld->getChain(), Ptr,
2970 Ld->getPointerInfo(),
2971 Ld->isVolatile(), Ld->isNonTemporal(),
2972 Ld->isInvariant(), Ld->getAlignment());
2974 EVT PtrType = Ptr.getValueType();
2975 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2976 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2977 PtrType, Ptr, DAG.getConstant(4, PtrType));
2978 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2979 Ld->getChain(), NewPtr,
2980 Ld->getPointerInfo().getWithOffset(4),
2981 Ld->isVolatile(), Ld->isNonTemporal(),
2982 Ld->isInvariant(), NewAlign);
2986 llvm_unreachable("Unknown VFP cmp argument!");
2989 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2990 /// f32 and even f64 comparisons to integer ones.
2992 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2993 SDValue Chain = Op.getOperand(0);
2994 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2995 SDValue LHS = Op.getOperand(2);
2996 SDValue RHS = Op.getOperand(3);
2997 SDValue Dest = Op.getOperand(4);
2998 DebugLoc dl = Op.getDebugLoc();
3000 bool LHSSeenZero = false;
3001 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3002 bool RHSSeenZero = false;
3003 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3004 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3005 // If unsafe fp math optimization is enabled and there are no other uses of
3006 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3007 // to an integer comparison.
3008 if (CC == ISD::SETOEQ)
3010 else if (CC == ISD::SETUNE)
3013 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3015 if (LHS.getValueType() == MVT::f32) {
3016 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3017 bitcastf32Toi32(LHS, DAG), Mask);
3018 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3019 bitcastf32Toi32(RHS, DAG), Mask);
3020 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3021 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3022 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3023 Chain, Dest, ARMcc, CCR, Cmp);
3028 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3029 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3030 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3031 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3032 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3033 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3034 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3035 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3036 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3042 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3043 SDValue Chain = Op.getOperand(0);
3044 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3045 SDValue LHS = Op.getOperand(2);
3046 SDValue RHS = Op.getOperand(3);
3047 SDValue Dest = Op.getOperand(4);
3048 DebugLoc dl = Op.getDebugLoc();
3050 if (LHS.getValueType() == MVT::i32) {
3052 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3053 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3054 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3055 Chain, Dest, ARMcc, CCR, Cmp);
3058 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3060 if (getTargetMachine().Options.UnsafeFPMath &&
3061 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3062 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3063 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3064 if (Result.getNode())
3068 ARMCC::CondCodes CondCode, CondCode2;
3069 FPCCToARMCC(CC, CondCode, CondCode2);
3071 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3072 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3073 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3074 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3075 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3076 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3077 if (CondCode2 != ARMCC::AL) {
3078 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3079 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3080 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3085 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3086 SDValue Chain = Op.getOperand(0);
3087 SDValue Table = Op.getOperand(1);
3088 SDValue Index = Op.getOperand(2);
3089 DebugLoc dl = Op.getDebugLoc();
3091 EVT PTy = getPointerTy();
3092 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3093 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3094 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3095 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3096 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3097 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3098 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3099 if (Subtarget->isThumb2()) {
3100 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3101 // which does another jump to the destination. This also makes it easier
3102 // to translate it to TBB / TBH later.
3103 // FIXME: This might not work if the function is extremely large.
3104 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3105 Addr, Op.getOperand(2), JTI, UId);
3107 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3108 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3109 MachinePointerInfo::getJumpTable(),
3110 false, false, false, 0);
3111 Chain = Addr.getValue(1);
3112 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3113 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3115 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3116 MachinePointerInfo::getJumpTable(),
3117 false, false, false, 0);
3118 Chain = Addr.getValue(1);
3119 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3123 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3124 EVT VT = Op.getValueType();
3125 DebugLoc dl = Op.getDebugLoc();
3127 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3128 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3130 return DAG.UnrollVectorOp(Op.getNode());
3133 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3134 "Invalid type for custom lowering!");
3135 if (VT != MVT::v4i16)
3136 return DAG.UnrollVectorOp(Op.getNode());
3138 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3139 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3142 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3143 EVT VT = Op.getValueType();
3145 return LowerVectorFP_TO_INT(Op, DAG);
3147 DebugLoc dl = Op.getDebugLoc();
3150 switch (Op.getOpcode()) {
3151 default: llvm_unreachable("Invalid opcode!");
3152 case ISD::FP_TO_SINT:
3153 Opc = ARMISD::FTOSI;
3155 case ISD::FP_TO_UINT:
3156 Opc = ARMISD::FTOUI;
3159 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3160 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3163 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3164 EVT VT = Op.getValueType();
3165 DebugLoc dl = Op.getDebugLoc();
3167 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3168 if (VT.getVectorElementType() == MVT::f32)
3170 return DAG.UnrollVectorOp(Op.getNode());
3173 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3174 "Invalid type for custom lowering!");
3175 if (VT != MVT::v4f32)
3176 return DAG.UnrollVectorOp(Op.getNode());
3180 switch (Op.getOpcode()) {
3181 default: llvm_unreachable("Invalid opcode!");
3182 case ISD::SINT_TO_FP:
3183 CastOpc = ISD::SIGN_EXTEND;
3184 Opc = ISD::SINT_TO_FP;
3186 case ISD::UINT_TO_FP:
3187 CastOpc = ISD::ZERO_EXTEND;
3188 Opc = ISD::UINT_TO_FP;
3192 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3193 return DAG.getNode(Opc, dl, VT, Op);
3196 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3197 EVT VT = Op.getValueType();
3199 return LowerVectorINT_TO_FP(Op, DAG);
3201 DebugLoc dl = Op.getDebugLoc();
3204 switch (Op.getOpcode()) {
3205 default: llvm_unreachable("Invalid opcode!");
3206 case ISD::SINT_TO_FP:
3207 Opc = ARMISD::SITOF;
3209 case ISD::UINT_TO_FP:
3210 Opc = ARMISD::UITOF;
3214 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3215 return DAG.getNode(Opc, dl, VT, Op);
3218 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3219 // Implement fcopysign with a fabs and a conditional fneg.
3220 SDValue Tmp0 = Op.getOperand(0);
3221 SDValue Tmp1 = Op.getOperand(1);
3222 DebugLoc dl = Op.getDebugLoc();
3223 EVT VT = Op.getValueType();
3224 EVT SrcVT = Tmp1.getValueType();
3225 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3226 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3227 bool UseNEON = !InGPR && Subtarget->hasNEON();
3230 // Use VBSL to copy the sign bit.
3231 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3232 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3233 DAG.getTargetConstant(EncodedVal, MVT::i32));
3234 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3236 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3237 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3238 DAG.getConstant(32, MVT::i32));
3239 else /*if (VT == MVT::f32)*/
3240 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3241 if (SrcVT == MVT::f32) {
3242 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3244 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3245 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3246 DAG.getConstant(32, MVT::i32));
3247 } else if (VT == MVT::f32)
3248 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3249 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3250 DAG.getConstant(32, MVT::i32));
3251 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3252 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3254 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3256 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3257 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3258 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3260 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3261 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3262 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3263 if (VT == MVT::f32) {
3264 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3265 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3266 DAG.getConstant(0, MVT::i32));
3268 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3274 // Bitcast operand 1 to i32.
3275 if (SrcVT == MVT::f64)
3276 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3277 &Tmp1, 1).getValue(1);
3278 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3280 // Or in the signbit with integer operations.
3281 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3282 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3283 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3284 if (VT == MVT::f32) {
3285 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3286 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3287 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3288 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3291 // f64: Or the high part with signbit and then combine two parts.
3292 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3294 SDValue Lo = Tmp0.getValue(0);
3295 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3296 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3297 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3300 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3301 MachineFunction &MF = DAG.getMachineFunction();
3302 MachineFrameInfo *MFI = MF.getFrameInfo();
3303 MFI->setReturnAddressIsTaken(true);
3305 EVT VT = Op.getValueType();
3306 DebugLoc dl = Op.getDebugLoc();
3307 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3309 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3310 SDValue Offset = DAG.getConstant(4, MVT::i32);
3311 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3312 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3313 MachinePointerInfo(), false, false, false, 0);
3316 // Return LR, which contains the return address. Mark it an implicit live-in.
3317 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3318 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3321 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3322 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3323 MFI->setFrameAddressIsTaken(true);
3325 EVT VT = Op.getValueType();
3326 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3327 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3328 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3329 ? ARM::R7 : ARM::R11;
3330 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3332 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3333 MachinePointerInfo(),
3334 false, false, false, 0);
3338 /// ExpandBITCAST - If the target supports VFP, this function is called to
3339 /// expand a bit convert where either the source or destination type is i64 to
3340 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3341 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3342 /// vectors), since the legalizer won't know what to do with that.
3343 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3344 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3345 DebugLoc dl = N->getDebugLoc();
3346 SDValue Op = N->getOperand(0);
3348 // This function is only supposed to be called for i64 types, either as the
3349 // source or destination of the bit convert.
3350 EVT SrcVT = Op.getValueType();
3351 EVT DstVT = N->getValueType(0);
3352 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3353 "ExpandBITCAST called for non-i64 type");
3355 // Turn i64->f64 into VMOVDRR.
3356 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3357 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3358 DAG.getConstant(0, MVT::i32));
3359 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3360 DAG.getConstant(1, MVT::i32));
3361 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3362 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3365 // Turn f64->i64 into VMOVRRD.
3366 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3367 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3368 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3369 // Merge the pieces into a single i64 value.
3370 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3376 /// getZeroVector - Returns a vector of specified type with all zero elements.
3377 /// Zero vectors are used to represent vector negation and in those cases
3378 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3379 /// not support i64 elements, so sometimes the zero vectors will need to be
3380 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3382 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3383 assert(VT.isVector() && "Expected a vector type");
3384 // The canonical modified immediate encoding of a zero vector is....0!
3385 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3386 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3387 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3388 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3391 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3392 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3393 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3394 SelectionDAG &DAG) const {
3395 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3396 EVT VT = Op.getValueType();
3397 unsigned VTBits = VT.getSizeInBits();
3398 DebugLoc dl = Op.getDebugLoc();
3399 SDValue ShOpLo = Op.getOperand(0);
3400 SDValue ShOpHi = Op.getOperand(1);
3401 SDValue ShAmt = Op.getOperand(2);
3403 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3405 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3407 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3408 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3409 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3410 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3411 DAG.getConstant(VTBits, MVT::i32));
3412 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3413 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3414 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3416 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3417 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3419 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3420 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3423 SDValue Ops[2] = { Lo, Hi };
3424 return DAG.getMergeValues(Ops, 2, dl);
3427 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3428 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3429 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3430 SelectionDAG &DAG) const {
3431 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3432 EVT VT = Op.getValueType();
3433 unsigned VTBits = VT.getSizeInBits();
3434 DebugLoc dl = Op.getDebugLoc();
3435 SDValue ShOpLo = Op.getOperand(0);
3436 SDValue ShOpHi = Op.getOperand(1);
3437 SDValue ShAmt = Op.getOperand(2);
3440 assert(Op.getOpcode() == ISD::SHL_PARTS);
3441 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3442 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3443 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3444 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3445 DAG.getConstant(VTBits, MVT::i32));
3446 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3447 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3449 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3450 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3451 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3453 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3454 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3457 SDValue Ops[2] = { Lo, Hi };
3458 return DAG.getMergeValues(Ops, 2, dl);
3461 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3462 SelectionDAG &DAG) const {
3463 // The rounding mode is in bits 23:22 of the FPSCR.
3464 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3465 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3466 // so that the shift + and get folded into a bitfield extract.
3467 DebugLoc dl = Op.getDebugLoc();
3468 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3469 DAG.getConstant(Intrinsic::arm_get_fpscr,
3471 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3472 DAG.getConstant(1U << 22, MVT::i32));
3473 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3474 DAG.getConstant(22, MVT::i32));
3475 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3476 DAG.getConstant(3, MVT::i32));
3479 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3480 const ARMSubtarget *ST) {
3481 EVT VT = N->getValueType(0);
3482 DebugLoc dl = N->getDebugLoc();
3484 if (!ST->hasV6T2Ops())
3487 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3488 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3491 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3492 const ARMSubtarget *ST) {
3493 EVT VT = N->getValueType(0);
3494 DebugLoc dl = N->getDebugLoc();
3499 // Lower vector shifts on NEON to use VSHL.
3500 assert(ST->hasNEON() && "unexpected vector shift");
3502 // Left shifts translate directly to the vshiftu intrinsic.
3503 if (N->getOpcode() == ISD::SHL)
3504 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3505 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3506 N->getOperand(0), N->getOperand(1));
3508 assert((N->getOpcode() == ISD::SRA ||
3509 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3511 // NEON uses the same intrinsics for both left and right shifts. For
3512 // right shifts, the shift amounts are negative, so negate the vector of
3514 EVT ShiftVT = N->getOperand(1).getValueType();
3515 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3516 getZeroVector(ShiftVT, DAG, dl),
3518 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3519 Intrinsic::arm_neon_vshifts :
3520 Intrinsic::arm_neon_vshiftu);
3521 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3522 DAG.getConstant(vshiftInt, MVT::i32),
3523 N->getOperand(0), NegatedCount);
3526 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3527 const ARMSubtarget *ST) {
3528 EVT VT = N->getValueType(0);
3529 DebugLoc dl = N->getDebugLoc();
3531 // We can get here for a node like i32 = ISD::SHL i32, i64
3535 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3536 "Unknown shift to lower!");
3538 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3539 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3540 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3543 // If we are in thumb mode, we don't have RRX.
3544 if (ST->isThumb1Only()) return SDValue();
3546 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3547 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3548 DAG.getConstant(0, MVT::i32));
3549 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3550 DAG.getConstant(1, MVT::i32));
3552 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3553 // captures the result into a carry flag.
3554 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3555 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3557 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3558 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3560 // Merge the pieces into a single i64 value.
3561 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3564 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3565 SDValue TmpOp0, TmpOp1;
3566 bool Invert = false;
3570 SDValue Op0 = Op.getOperand(0);
3571 SDValue Op1 = Op.getOperand(1);
3572 SDValue CC = Op.getOperand(2);
3573 EVT VT = Op.getValueType();
3574 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3575 DebugLoc dl = Op.getDebugLoc();
3577 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3578 switch (SetCCOpcode) {
3579 default: llvm_unreachable("Illegal FP comparison");
3581 case ISD::SETNE: Invert = true; // Fallthrough
3583 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3585 case ISD::SETLT: Swap = true; // Fallthrough
3587 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3589 case ISD::SETLE: Swap = true; // Fallthrough
3591 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3592 case ISD::SETUGE: Swap = true; // Fallthrough
3593 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3594 case ISD::SETUGT: Swap = true; // Fallthrough
3595 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3596 case ISD::SETUEQ: Invert = true; // Fallthrough
3598 // Expand this to (OLT | OGT).
3602 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3603 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3605 case ISD::SETUO: Invert = true; // Fallthrough
3607 // Expand this to (OLT | OGE).
3611 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3612 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3616 // Integer comparisons.
3617 switch (SetCCOpcode) {
3618 default: llvm_unreachable("Illegal integer comparison");
3619 case ISD::SETNE: Invert = true;
3620 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3621 case ISD::SETLT: Swap = true;
3622 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3623 case ISD::SETLE: Swap = true;
3624 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3625 case ISD::SETULT: Swap = true;
3626 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3627 case ISD::SETULE: Swap = true;
3628 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3631 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3632 if (Opc == ARMISD::VCEQ) {
3635 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3637 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3640 // Ignore bitconvert.
3641 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3642 AndOp = AndOp.getOperand(0);
3644 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3646 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3647 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3654 std::swap(Op0, Op1);
3656 // If one of the operands is a constant vector zero, attempt to fold the
3657 // comparison to a specialized compare-against-zero form.
3659 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3661 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3662 if (Opc == ARMISD::VCGE)
3663 Opc = ARMISD::VCLEZ;
3664 else if (Opc == ARMISD::VCGT)
3665 Opc = ARMISD::VCLTZ;
3670 if (SingleOp.getNode()) {
3673 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3675 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3677 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3679 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3681 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3683 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3686 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3690 Result = DAG.getNOT(dl, Result, VT);
3695 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3696 /// valid vector constant for a NEON instruction with a "modified immediate"
3697 /// operand (e.g., VMOV). If so, return the encoded value.
3698 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3699 unsigned SplatBitSize, SelectionDAG &DAG,
3700 EVT &VT, bool is128Bits, NEONModImmType type) {
3701 unsigned OpCmode, Imm;
3703 // SplatBitSize is set to the smallest size that splats the vector, so a
3704 // zero vector will always have SplatBitSize == 8. However, NEON modified
3705 // immediate instructions others than VMOV do not support the 8-bit encoding
3706 // of a zero vector, and the default encoding of zero is supposed to be the
3711 switch (SplatBitSize) {
3713 if (type != VMOVModImm)
3715 // Any 1-byte value is OK. Op=0, Cmode=1110.
3716 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3719 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3723 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3724 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3725 if ((SplatBits & ~0xff) == 0) {
3726 // Value = 0x00nn: Op=x, Cmode=100x.
3731 if ((SplatBits & ~0xff00) == 0) {
3732 // Value = 0xnn00: Op=x, Cmode=101x.
3734 Imm = SplatBits >> 8;
3740 // NEON's 32-bit VMOV supports splat values where:
3741 // * only one byte is nonzero, or
3742 // * the least significant byte is 0xff and the second byte is nonzero, or
3743 // * the least significant 2 bytes are 0xff and the third is nonzero.
3744 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3745 if ((SplatBits & ~0xff) == 0) {
3746 // Value = 0x000000nn: Op=x, Cmode=000x.
3751 if ((SplatBits & ~0xff00) == 0) {
3752 // Value = 0x0000nn00: Op=x, Cmode=001x.
3754 Imm = SplatBits >> 8;
3757 if ((SplatBits & ~0xff0000) == 0) {
3758 // Value = 0x00nn0000: Op=x, Cmode=010x.
3760 Imm = SplatBits >> 16;
3763 if ((SplatBits & ~0xff000000) == 0) {
3764 // Value = 0xnn000000: Op=x, Cmode=011x.
3766 Imm = SplatBits >> 24;
3770 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3771 if (type == OtherModImm) return SDValue();
3773 if ((SplatBits & ~0xffff) == 0 &&
3774 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3775 // Value = 0x0000nnff: Op=x, Cmode=1100.
3777 Imm = SplatBits >> 8;
3782 if ((SplatBits & ~0xffffff) == 0 &&
3783 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3784 // Value = 0x00nnffff: Op=x, Cmode=1101.
3786 Imm = SplatBits >> 16;
3787 SplatBits |= 0xffff;
3791 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3792 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3793 // VMOV.I32. A (very) minor optimization would be to replicate the value
3794 // and fall through here to test for a valid 64-bit splat. But, then the
3795 // caller would also need to check and handle the change in size.
3799 if (type != VMOVModImm)
3801 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3802 uint64_t BitMask = 0xff;
3804 unsigned ImmMask = 1;
3806 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3807 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3810 } else if ((SplatBits & BitMask) != 0) {
3816 // Op=1, Cmode=1110.
3819 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3824 llvm_unreachable("unexpected size for isNEONModifiedImm");
3827 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3828 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3831 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3832 const ARMSubtarget *ST) const {
3833 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3836 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3837 assert(Op.getValueType() == MVT::f32 &&
3838 "ConstantFP custom lowering should only occur for f32.");
3840 // Try splatting with a VMOV.f32...
3841 APFloat FPVal = CFP->getValueAPF();
3842 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3844 DebugLoc DL = Op.getDebugLoc();
3845 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3846 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3848 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3849 DAG.getConstant(0, MVT::i32));
3852 // If that fails, try a VMOV.i32
3854 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3855 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3857 if (NewVal != SDValue()) {
3858 DebugLoc DL = Op.getDebugLoc();
3859 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3861 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3863 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3864 DAG.getConstant(0, MVT::i32));
3867 // Finally, try a VMVN.i32
3868 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3870 if (NewVal != SDValue()) {
3871 DebugLoc DL = Op.getDebugLoc();
3872 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3873 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3875 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3876 DAG.getConstant(0, MVT::i32));
3883 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3884 bool &ReverseVEXT, unsigned &Imm) {
3885 unsigned NumElts = VT.getVectorNumElements();
3886 ReverseVEXT = false;
3888 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3894 // If this is a VEXT shuffle, the immediate value is the index of the first
3895 // element. The other shuffle indices must be the successive elements after
3897 unsigned ExpectedElt = Imm;
3898 for (unsigned i = 1; i < NumElts; ++i) {
3899 // Increment the expected index. If it wraps around, it may still be
3900 // a VEXT but the source vectors must be swapped.
3902 if (ExpectedElt == NumElts * 2) {
3907 if (M[i] < 0) continue; // ignore UNDEF indices
3908 if (ExpectedElt != static_cast<unsigned>(M[i]))
3912 // Adjust the index value if the source operands will be swapped.
3919 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3920 /// instruction with the specified blocksize. (The order of the elements
3921 /// within each block of the vector is reversed.)
3922 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3923 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3924 "Only possible block sizes for VREV are: 16, 32, 64");
3926 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3930 unsigned NumElts = VT.getVectorNumElements();
3931 unsigned BlockElts = M[0] + 1;
3932 // If the first shuffle index is UNDEF, be optimistic.
3934 BlockElts = BlockSize / EltSz;
3936 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3939 for (unsigned i = 0; i < NumElts; ++i) {
3940 if (M[i] < 0) continue; // ignore UNDEF indices
3941 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3948 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3949 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3950 // range, then 0 is placed into the resulting vector. So pretty much any mask
3951 // of 8 elements can work here.
3952 return VT == MVT::v8i8 && M.size() == 8;
3955 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3956 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3960 unsigned NumElts = VT.getVectorNumElements();
3961 WhichResult = (M[0] == 0 ? 0 : 1);
3962 for (unsigned i = 0; i < NumElts; i += 2) {
3963 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3964 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3970 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3971 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3972 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3973 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3974 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3978 unsigned NumElts = VT.getVectorNumElements();
3979 WhichResult = (M[0] == 0 ? 0 : 1);
3980 for (unsigned i = 0; i < NumElts; i += 2) {
3981 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3982 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3988 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3989 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3993 unsigned NumElts = VT.getVectorNumElements();
3994 WhichResult = (M[0] == 0 ? 0 : 1);
3995 for (unsigned i = 0; i != NumElts; ++i) {
3996 if (M[i] < 0) continue; // ignore UNDEF indices
3997 if ((unsigned) M[i] != 2 * i + WhichResult)
4001 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4002 if (VT.is64BitVector() && EltSz == 32)
4008 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4009 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4010 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4011 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4012 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4016 unsigned Half = VT.getVectorNumElements() / 2;
4017 WhichResult = (M[0] == 0 ? 0 : 1);
4018 for (unsigned j = 0; j != 2; ++j) {
4019 unsigned Idx = WhichResult;
4020 for (unsigned i = 0; i != Half; ++i) {
4021 int MIdx = M[i + j * Half];
4022 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4028 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4029 if (VT.is64BitVector() && EltSz == 32)
4035 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4036 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4040 unsigned NumElts = VT.getVectorNumElements();
4041 WhichResult = (M[0] == 0 ? 0 : 1);
4042 unsigned Idx = WhichResult * NumElts / 2;
4043 for (unsigned i = 0; i != NumElts; i += 2) {
4044 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4045 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4050 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4051 if (VT.is64BitVector() && EltSz == 32)
4057 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4058 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4059 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4060 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4061 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4065 unsigned NumElts = VT.getVectorNumElements();
4066 WhichResult = (M[0] == 0 ? 0 : 1);
4067 unsigned Idx = WhichResult * NumElts / 2;
4068 for (unsigned i = 0; i != NumElts; i += 2) {
4069 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4070 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4075 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4076 if (VT.is64BitVector() && EltSz == 32)
4082 // If N is an integer constant that can be moved into a register in one
4083 // instruction, return an SDValue of such a constant (will become a MOV
4084 // instruction). Otherwise return null.
4085 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4086 const ARMSubtarget *ST, DebugLoc dl) {
4088 if (!isa<ConstantSDNode>(N))
4090 Val = cast<ConstantSDNode>(N)->getZExtValue();
4092 if (ST->isThumb1Only()) {
4093 if (Val <= 255 || ~Val <= 255)
4094 return DAG.getConstant(Val, MVT::i32);
4096 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4097 return DAG.getConstant(Val, MVT::i32);
4102 // If this is a case we can't handle, return null and let the default
4103 // expansion code take care of it.
4104 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4105 const ARMSubtarget *ST) const {
4106 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4107 DebugLoc dl = Op.getDebugLoc();
4108 EVT VT = Op.getValueType();
4110 APInt SplatBits, SplatUndef;
4111 unsigned SplatBitSize;
4113 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4114 if (SplatBitSize <= 64) {
4115 // Check if an immediate VMOV works.
4117 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4118 SplatUndef.getZExtValue(), SplatBitSize,
4119 DAG, VmovVT, VT.is128BitVector(),
4121 if (Val.getNode()) {
4122 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4123 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4126 // Try an immediate VMVN.
4127 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4128 Val = isNEONModifiedImm(NegatedImm,
4129 SplatUndef.getZExtValue(), SplatBitSize,
4130 DAG, VmovVT, VT.is128BitVector(),
4132 if (Val.getNode()) {
4133 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4134 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4137 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4138 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4139 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4141 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4142 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4148 // Scan through the operands to see if only one value is used.
4149 unsigned NumElts = VT.getVectorNumElements();
4150 bool isOnlyLowElement = true;
4151 bool usesOnlyOneValue = true;
4152 bool isConstant = true;
4154 for (unsigned i = 0; i < NumElts; ++i) {
4155 SDValue V = Op.getOperand(i);
4156 if (V.getOpcode() == ISD::UNDEF)
4159 isOnlyLowElement = false;
4160 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4163 if (!Value.getNode())
4165 else if (V != Value)
4166 usesOnlyOneValue = false;
4169 if (!Value.getNode())
4170 return DAG.getUNDEF(VT);
4172 if (isOnlyLowElement)
4173 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4175 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4177 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4178 // i32 and try again.
4179 if (usesOnlyOneValue && EltSize <= 32) {
4181 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4182 if (VT.getVectorElementType().isFloatingPoint()) {
4183 SmallVector<SDValue, 8> Ops;
4184 for (unsigned i = 0; i < NumElts; ++i)
4185 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4187 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4188 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4189 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4191 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4193 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4195 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4198 // If all elements are constants and the case above didn't get hit, fall back
4199 // to the default expansion, which will generate a load from the constant
4204 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4206 SDValue shuffle = ReconstructShuffle(Op, DAG);
4207 if (shuffle != SDValue())
4211 // Vectors with 32- or 64-bit elements can be built by directly assigning
4212 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4213 // will be legalized.
4214 if (EltSize >= 32) {
4215 // Do the expansion with floating-point types, since that is what the VFP
4216 // registers are defined to use, and since i64 is not legal.
4217 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4218 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4219 SmallVector<SDValue, 8> Ops;
4220 for (unsigned i = 0; i < NumElts; ++i)
4221 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4222 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4223 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4229 // Gather data to see if the operation can be modelled as a
4230 // shuffle in combination with VEXTs.
4231 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4232 SelectionDAG &DAG) const {
4233 DebugLoc dl = Op.getDebugLoc();
4234 EVT VT = Op.getValueType();
4235 unsigned NumElts = VT.getVectorNumElements();
4237 SmallVector<SDValue, 2> SourceVecs;
4238 SmallVector<unsigned, 2> MinElts;
4239 SmallVector<unsigned, 2> MaxElts;
4241 for (unsigned i = 0; i < NumElts; ++i) {
4242 SDValue V = Op.getOperand(i);
4243 if (V.getOpcode() == ISD::UNDEF)
4245 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4246 // A shuffle can only come from building a vector from various
4247 // elements of other vectors.
4249 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4250 VT.getVectorElementType()) {
4251 // This code doesn't know how to handle shuffles where the vector
4252 // element types do not match (this happens because type legalization
4253 // promotes the return type of EXTRACT_VECTOR_ELT).
4254 // FIXME: It might be appropriate to extend this code to handle
4255 // mismatched types.
4259 // Record this extraction against the appropriate vector if possible...
4260 SDValue SourceVec = V.getOperand(0);
4261 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4262 bool FoundSource = false;
4263 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4264 if (SourceVecs[j] == SourceVec) {
4265 if (MinElts[j] > EltNo)
4267 if (MaxElts[j] < EltNo)
4274 // Or record a new source if not...
4276 SourceVecs.push_back(SourceVec);
4277 MinElts.push_back(EltNo);
4278 MaxElts.push_back(EltNo);
4282 // Currently only do something sane when at most two source vectors
4284 if (SourceVecs.size() > 2)
4287 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4288 int VEXTOffsets[2] = {0, 0};
4290 // This loop extracts the usage patterns of the source vectors
4291 // and prepares appropriate SDValues for a shuffle if possible.
4292 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4293 if (SourceVecs[i].getValueType() == VT) {
4294 // No VEXT necessary
4295 ShuffleSrcs[i] = SourceVecs[i];
4298 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4299 // It probably isn't worth padding out a smaller vector just to
4300 // break it down again in a shuffle.
4304 // Since only 64-bit and 128-bit vectors are legal on ARM and
4305 // we've eliminated the other cases...
4306 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4307 "unexpected vector sizes in ReconstructShuffle");
4309 if (MaxElts[i] - MinElts[i] >= NumElts) {
4310 // Span too large for a VEXT to cope
4314 if (MinElts[i] >= NumElts) {
4315 // The extraction can just take the second half
4316 VEXTOffsets[i] = NumElts;
4317 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4319 DAG.getIntPtrConstant(NumElts));
4320 } else if (MaxElts[i] < NumElts) {
4321 // The extraction can just take the first half
4323 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4325 DAG.getIntPtrConstant(0));
4327 // An actual VEXT is needed
4328 VEXTOffsets[i] = MinElts[i];
4329 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4331 DAG.getIntPtrConstant(0));
4332 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4334 DAG.getIntPtrConstant(NumElts));
4335 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4336 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4340 SmallVector<int, 8> Mask;
4342 for (unsigned i = 0; i < NumElts; ++i) {
4343 SDValue Entry = Op.getOperand(i);
4344 if (Entry.getOpcode() == ISD::UNDEF) {
4349 SDValue ExtractVec = Entry.getOperand(0);
4350 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4351 .getOperand(1))->getSExtValue();
4352 if (ExtractVec == SourceVecs[0]) {
4353 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4355 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4359 // Final check before we try to produce nonsense...
4360 if (isShuffleMaskLegal(Mask, VT))
4361 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4367 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4368 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4369 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4370 /// are assumed to be legal.
4372 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4374 if (VT.getVectorNumElements() == 4 &&
4375 (VT.is128BitVector() || VT.is64BitVector())) {
4376 unsigned PFIndexes[4];
4377 for (unsigned i = 0; i != 4; ++i) {
4381 PFIndexes[i] = M[i];
4384 // Compute the index in the perfect shuffle table.
4385 unsigned PFTableIndex =
4386 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4387 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4388 unsigned Cost = (PFEntry >> 30);
4395 unsigned Imm, WhichResult;
4397 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4398 return (EltSize >= 32 ||
4399 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4400 isVREVMask(M, VT, 64) ||
4401 isVREVMask(M, VT, 32) ||
4402 isVREVMask(M, VT, 16) ||
4403 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4404 isVTBLMask(M, VT) ||
4405 isVTRNMask(M, VT, WhichResult) ||
4406 isVUZPMask(M, VT, WhichResult) ||
4407 isVZIPMask(M, VT, WhichResult) ||
4408 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4409 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4410 isVZIP_v_undef_Mask(M, VT, WhichResult));
4413 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4414 /// the specified operations to build the shuffle.
4415 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4416 SDValue RHS, SelectionDAG &DAG,
4418 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4419 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4420 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4423 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4432 OP_VUZPL, // VUZP, left result
4433 OP_VUZPR, // VUZP, right result
4434 OP_VZIPL, // VZIP, left result
4435 OP_VZIPR, // VZIP, right result
4436 OP_VTRNL, // VTRN, left result
4437 OP_VTRNR // VTRN, right result
4440 if (OpNum == OP_COPY) {
4441 if (LHSID == (1*9+2)*9+3) return LHS;
4442 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4446 SDValue OpLHS, OpRHS;
4447 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4448 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4449 EVT VT = OpLHS.getValueType();
4452 default: llvm_unreachable("Unknown shuffle opcode!");
4454 // VREV divides the vector in half and swaps within the half.
4455 if (VT.getVectorElementType() == MVT::i32 ||
4456 VT.getVectorElementType() == MVT::f32)
4457 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4458 // vrev <4 x i16> -> VREV32
4459 if (VT.getVectorElementType() == MVT::i16)
4460 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4461 // vrev <4 x i8> -> VREV16
4462 assert(VT.getVectorElementType() == MVT::i8);
4463 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4468 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4469 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4473 return DAG.getNode(ARMISD::VEXT, dl, VT,
4475 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4478 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4479 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4482 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4483 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4486 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4487 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4491 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4492 ArrayRef<int> ShuffleMask,
4493 SelectionDAG &DAG) {
4494 // Check to see if we can use the VTBL instruction.
4495 SDValue V1 = Op.getOperand(0);
4496 SDValue V2 = Op.getOperand(1);
4497 DebugLoc DL = Op.getDebugLoc();
4499 SmallVector<SDValue, 8> VTBLMask;
4500 for (ArrayRef<int>::iterator
4501 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4502 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4504 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4505 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4506 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4509 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4510 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4514 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4515 SDValue V1 = Op.getOperand(0);
4516 SDValue V2 = Op.getOperand(1);
4517 DebugLoc dl = Op.getDebugLoc();
4518 EVT VT = Op.getValueType();
4519 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4521 // Convert shuffles that are directly supported on NEON to target-specific
4522 // DAG nodes, instead of keeping them as shuffles and matching them again
4523 // during code selection. This is more efficient and avoids the possibility
4524 // of inconsistencies between legalization and selection.
4525 // FIXME: floating-point vectors should be canonicalized to integer vectors
4526 // of the same time so that they get CSEd properly.
4527 ArrayRef<int> ShuffleMask = SVN->getMask();
4529 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4530 if (EltSize <= 32) {
4531 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4532 int Lane = SVN->getSplatIndex();
4533 // If this is undef splat, generate it via "just" vdup, if possible.
4534 if (Lane == -1) Lane = 0;
4536 // Test if V1 is a SCALAR_TO_VECTOR.
4537 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4538 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4540 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4541 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4543 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4544 !isa<ConstantSDNode>(V1.getOperand(0))) {
4545 bool IsScalarToVector = true;
4546 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4547 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4548 IsScalarToVector = false;
4551 if (IsScalarToVector)
4552 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4554 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4555 DAG.getConstant(Lane, MVT::i32));
4560 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4563 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4564 DAG.getConstant(Imm, MVT::i32));
4567 if (isVREVMask(ShuffleMask, VT, 64))
4568 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4569 if (isVREVMask(ShuffleMask, VT, 32))
4570 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4571 if (isVREVMask(ShuffleMask, VT, 16))
4572 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4574 // Check for Neon shuffles that modify both input vectors in place.
4575 // If both results are used, i.e., if there are two shuffles with the same
4576 // source operands and with masks corresponding to both results of one of
4577 // these operations, DAG memoization will ensure that a single node is
4578 // used for both shuffles.
4579 unsigned WhichResult;
4580 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4581 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4582 V1, V2).getValue(WhichResult);
4583 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4584 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4585 V1, V2).getValue(WhichResult);
4586 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4587 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4588 V1, V2).getValue(WhichResult);
4590 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4591 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4592 V1, V1).getValue(WhichResult);
4593 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4594 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4595 V1, V1).getValue(WhichResult);
4596 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4597 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4598 V1, V1).getValue(WhichResult);
4601 // If the shuffle is not directly supported and it has 4 elements, use
4602 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4603 unsigned NumElts = VT.getVectorNumElements();
4605 unsigned PFIndexes[4];
4606 for (unsigned i = 0; i != 4; ++i) {
4607 if (ShuffleMask[i] < 0)
4610 PFIndexes[i] = ShuffleMask[i];
4613 // Compute the index in the perfect shuffle table.
4614 unsigned PFTableIndex =
4615 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4616 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4617 unsigned Cost = (PFEntry >> 30);
4620 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4623 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4624 if (EltSize >= 32) {
4625 // Do the expansion with floating-point types, since that is what the VFP
4626 // registers are defined to use, and since i64 is not legal.
4627 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4628 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4629 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4630 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4631 SmallVector<SDValue, 8> Ops;
4632 for (unsigned i = 0; i < NumElts; ++i) {
4633 if (ShuffleMask[i] < 0)
4634 Ops.push_back(DAG.getUNDEF(EltVT));
4636 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4637 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4638 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4641 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4642 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4645 if (VT == MVT::v8i8) {
4646 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4647 if (NewOp.getNode())
4654 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4655 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4656 SDValue Lane = Op.getOperand(2);
4657 if (!isa<ConstantSDNode>(Lane))
4663 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4664 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4665 SDValue Lane = Op.getOperand(1);
4666 if (!isa<ConstantSDNode>(Lane))
4669 SDValue Vec = Op.getOperand(0);
4670 if (Op.getValueType() == MVT::i32 &&
4671 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4672 DebugLoc dl = Op.getDebugLoc();
4673 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4679 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4680 // The only time a CONCAT_VECTORS operation can have legal types is when
4681 // two 64-bit vectors are concatenated to a 128-bit vector.
4682 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4683 "unexpected CONCAT_VECTORS");
4684 DebugLoc dl = Op.getDebugLoc();
4685 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4686 SDValue Op0 = Op.getOperand(0);
4687 SDValue Op1 = Op.getOperand(1);
4688 if (Op0.getOpcode() != ISD::UNDEF)
4689 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4690 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4691 DAG.getIntPtrConstant(0));
4692 if (Op1.getOpcode() != ISD::UNDEF)
4693 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4694 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4695 DAG.getIntPtrConstant(1));
4696 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4699 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4700 /// element has been zero/sign-extended, depending on the isSigned parameter,
4701 /// from an integer type half its size.
4702 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4704 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4705 EVT VT = N->getValueType(0);
4706 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4707 SDNode *BVN = N->getOperand(0).getNode();
4708 if (BVN->getValueType(0) != MVT::v4i32 ||
4709 BVN->getOpcode() != ISD::BUILD_VECTOR)
4711 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4712 unsigned HiElt = 1 - LoElt;
4713 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4714 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4715 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4716 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4717 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4720 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4721 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4724 if (Hi0->isNullValue() && Hi1->isNullValue())
4730 if (N->getOpcode() != ISD::BUILD_VECTOR)
4733 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4734 SDNode *Elt = N->getOperand(i).getNode();
4735 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4736 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4737 unsigned HalfSize = EltSize / 2;
4739 if (!isIntN(HalfSize, C->getSExtValue()))
4742 if (!isUIntN(HalfSize, C->getZExtValue()))
4753 /// isSignExtended - Check if a node is a vector value that is sign-extended
4754 /// or a constant BUILD_VECTOR with sign-extended elements.
4755 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4756 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4758 if (isExtendedBUILD_VECTOR(N, DAG, true))
4763 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4764 /// or a constant BUILD_VECTOR with zero-extended elements.
4765 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4766 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4768 if (isExtendedBUILD_VECTOR(N, DAG, false))
4773 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4774 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4775 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4776 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4777 return N->getOperand(0);
4778 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4779 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4780 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4781 LD->isNonTemporal(), LD->isInvariant(),
4782 LD->getAlignment());
4783 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4784 // have been legalized as a BITCAST from v4i32.
4785 if (N->getOpcode() == ISD::BITCAST) {
4786 SDNode *BVN = N->getOperand(0).getNode();
4787 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4788 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4789 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4790 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4791 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4793 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4794 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4795 EVT VT = N->getValueType(0);
4796 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4797 unsigned NumElts = VT.getVectorNumElements();
4798 MVT TruncVT = MVT::getIntegerVT(EltSize);
4799 SmallVector<SDValue, 8> Ops;
4800 for (unsigned i = 0; i != NumElts; ++i) {
4801 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4802 const APInt &CInt = C->getAPIntValue();
4803 // Element types smaller than 32 bits are not legal, so use i32 elements.
4804 // The values are implicitly truncated so sext vs. zext doesn't matter.
4805 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4807 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4808 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4811 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4812 unsigned Opcode = N->getOpcode();
4813 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4814 SDNode *N0 = N->getOperand(0).getNode();
4815 SDNode *N1 = N->getOperand(1).getNode();
4816 return N0->hasOneUse() && N1->hasOneUse() &&
4817 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4822 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4823 unsigned Opcode = N->getOpcode();
4824 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4825 SDNode *N0 = N->getOperand(0).getNode();
4826 SDNode *N1 = N->getOperand(1).getNode();
4827 return N0->hasOneUse() && N1->hasOneUse() &&
4828 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4833 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4834 // Multiplications are only custom-lowered for 128-bit vectors so that
4835 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4836 EVT VT = Op.getValueType();
4837 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4838 SDNode *N0 = Op.getOperand(0).getNode();
4839 SDNode *N1 = Op.getOperand(1).getNode();
4840 unsigned NewOpc = 0;
4842 bool isN0SExt = isSignExtended(N0, DAG);
4843 bool isN1SExt = isSignExtended(N1, DAG);
4844 if (isN0SExt && isN1SExt)
4845 NewOpc = ARMISD::VMULLs;
4847 bool isN0ZExt = isZeroExtended(N0, DAG);
4848 bool isN1ZExt = isZeroExtended(N1, DAG);
4849 if (isN0ZExt && isN1ZExt)
4850 NewOpc = ARMISD::VMULLu;
4851 else if (isN1SExt || isN1ZExt) {
4852 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4853 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4854 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4855 NewOpc = ARMISD::VMULLs;
4857 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4858 NewOpc = ARMISD::VMULLu;
4860 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4862 NewOpc = ARMISD::VMULLu;
4868 if (VT == MVT::v2i64)
4869 // Fall through to expand this. It is not legal.
4872 // Other vector multiplications are legal.
4877 // Legalize to a VMULL instruction.
4878 DebugLoc DL = Op.getDebugLoc();
4880 SDValue Op1 = SkipExtension(N1, DAG);
4882 Op0 = SkipExtension(N0, DAG);
4883 assert(Op0.getValueType().is64BitVector() &&
4884 Op1.getValueType().is64BitVector() &&
4885 "unexpected types for extended operands to VMULL");
4886 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4889 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4890 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4897 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4898 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4899 EVT Op1VT = Op1.getValueType();
4900 return DAG.getNode(N0->getOpcode(), DL, VT,
4901 DAG.getNode(NewOpc, DL, VT,
4902 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4903 DAG.getNode(NewOpc, DL, VT,
4904 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4908 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4910 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4911 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4912 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4913 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4914 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4915 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4916 // Get reciprocal estimate.
4917 // float4 recip = vrecpeq_f32(yf);
4918 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4919 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4920 // Because char has a smaller range than uchar, we can actually get away
4921 // without any newton steps. This requires that we use a weird bias
4922 // of 0xb000, however (again, this has been exhaustively tested).
4923 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4924 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4925 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4926 Y = DAG.getConstant(0xb000, MVT::i32);
4927 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4928 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4929 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4930 // Convert back to short.
4931 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4932 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4937 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4939 // Convert to float.
4940 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4941 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4942 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4943 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4944 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4945 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4947 // Use reciprocal estimate and one refinement step.
4948 // float4 recip = vrecpeq_f32(yf);
4949 // recip *= vrecpsq_f32(yf, recip);
4950 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4951 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4952 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4953 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4955 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4956 // Because short has a smaller range than ushort, we can actually get away
4957 // with only a single newton step. This requires that we use a weird bias
4958 // of 89, however (again, this has been exhaustively tested).
4959 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4960 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4961 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4962 N1 = DAG.getConstant(0x89, MVT::i32);
4963 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4964 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4965 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4966 // Convert back to integer and return.
4967 // return vmovn_s32(vcvt_s32_f32(result));
4968 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4969 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4973 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4974 EVT VT = Op.getValueType();
4975 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4976 "unexpected type for custom-lowering ISD::SDIV");
4978 DebugLoc dl = Op.getDebugLoc();
4979 SDValue N0 = Op.getOperand(0);
4980 SDValue N1 = Op.getOperand(1);
4983 if (VT == MVT::v8i8) {
4984 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4985 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4987 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4988 DAG.getIntPtrConstant(4));
4989 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4990 DAG.getIntPtrConstant(4));
4991 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4992 DAG.getIntPtrConstant(0));
4993 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4994 DAG.getIntPtrConstant(0));
4996 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4997 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4999 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5000 N0 = LowerCONCAT_VECTORS(N0, DAG);
5002 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5005 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5008 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5009 EVT VT = Op.getValueType();
5010 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5011 "unexpected type for custom-lowering ISD::UDIV");
5013 DebugLoc dl = Op.getDebugLoc();
5014 SDValue N0 = Op.getOperand(0);
5015 SDValue N1 = Op.getOperand(1);
5018 if (VT == MVT::v8i8) {
5019 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5020 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5022 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5023 DAG.getIntPtrConstant(4));
5024 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5025 DAG.getIntPtrConstant(4));
5026 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5027 DAG.getIntPtrConstant(0));
5028 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5029 DAG.getIntPtrConstant(0));
5031 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5032 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5034 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5035 N0 = LowerCONCAT_VECTORS(N0, DAG);
5037 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5038 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5043 // v4i16 sdiv ... Convert to float.
5044 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5045 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5046 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5047 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5048 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5049 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5051 // Use reciprocal estimate and two refinement steps.
5052 // float4 recip = vrecpeq_f32(yf);
5053 // recip *= vrecpsq_f32(yf, recip);
5054 // recip *= vrecpsq_f32(yf, recip);
5055 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5056 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5057 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5058 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5060 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5061 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5062 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5064 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5065 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5066 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5067 // and that it will never cause us to return an answer too large).
5068 // float4 result = as_float4(as_int4(xf*recip) + 2);
5069 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5070 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5071 N1 = DAG.getConstant(2, MVT::i32);
5072 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5073 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5074 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5075 // Convert back to integer and return.
5076 // return vmovn_u32(vcvt_s32_f32(result));
5077 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5078 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5082 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5083 EVT VT = Op.getNode()->getValueType(0);
5084 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5087 bool ExtraOp = false;
5088 switch (Op.getOpcode()) {
5089 default: llvm_unreachable("Invalid code");
5090 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5091 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5092 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5093 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5097 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5099 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5100 Op.getOperand(1), Op.getOperand(2));
5103 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5104 // Monotonic load/store is legal for all targets
5105 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5108 // Aquire/Release load/store is not legal for targets without a
5109 // dmb or equivalent available.
5115 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5116 SelectionDAG &DAG, unsigned NewOp) {
5117 DebugLoc dl = Node->getDebugLoc();
5118 assert (Node->getValueType(0) == MVT::i64 &&
5119 "Only know how to expand i64 atomics");
5121 SmallVector<SDValue, 6> Ops;
5122 Ops.push_back(Node->getOperand(0)); // Chain
5123 Ops.push_back(Node->getOperand(1)); // Ptr
5125 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5126 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5127 // High part of Val1
5128 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5129 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5130 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5131 // High part of Val1
5132 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5133 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5134 // High part of Val2
5135 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5136 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5138 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5140 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5141 cast<MemSDNode>(Node)->getMemOperand());
5142 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5143 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5144 Results.push_back(Result.getValue(2));
5147 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5148 switch (Op.getOpcode()) {
5149 default: llvm_unreachable("Don't know how to custom lower this!");
5150 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5151 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5152 case ISD::GlobalAddress:
5153 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5154 LowerGlobalAddressELF(Op, DAG);
5155 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5156 case ISD::SELECT: return LowerSELECT(Op, DAG);
5157 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5158 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5159 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5160 case ISD::VASTART: return LowerVASTART(Op, DAG);
5161 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5162 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5163 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5164 case ISD::SINT_TO_FP:
5165 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5166 case ISD::FP_TO_SINT:
5167 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5168 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5169 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5170 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5171 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5172 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5173 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5174 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5176 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5179 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5180 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5181 case ISD::SRL_PARTS:
5182 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5183 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5184 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5185 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5186 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5187 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5188 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5189 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5190 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5191 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5192 case ISD::MUL: return LowerMUL(Op, DAG);
5193 case ISD::SDIV: return LowerSDIV(Op, DAG);
5194 case ISD::UDIV: return LowerUDIV(Op, DAG);
5198 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5199 case ISD::ATOMIC_LOAD:
5200 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5204 /// ReplaceNodeResults - Replace the results of node with an illegal result
5205 /// type with new values built out of custom code.
5206 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5207 SmallVectorImpl<SDValue>&Results,
5208 SelectionDAG &DAG) const {
5210 switch (N->getOpcode()) {
5212 llvm_unreachable("Don't know how to custom expand this!");
5214 Res = ExpandBITCAST(N, DAG);
5218 Res = Expand64BitShift(N, DAG, Subtarget);
5220 case ISD::ATOMIC_LOAD_ADD:
5221 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5223 case ISD::ATOMIC_LOAD_AND:
5224 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5226 case ISD::ATOMIC_LOAD_NAND:
5227 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5229 case ISD::ATOMIC_LOAD_OR:
5230 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5232 case ISD::ATOMIC_LOAD_SUB:
5233 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5235 case ISD::ATOMIC_LOAD_XOR:
5236 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5238 case ISD::ATOMIC_SWAP:
5239 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5241 case ISD::ATOMIC_CMP_SWAP:
5242 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5246 Results.push_back(Res);
5249 //===----------------------------------------------------------------------===//
5250 // ARM Scheduler Hooks
5251 //===----------------------------------------------------------------------===//
5254 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5255 MachineBasicBlock *BB,
5256 unsigned Size) const {
5257 unsigned dest = MI->getOperand(0).getReg();
5258 unsigned ptr = MI->getOperand(1).getReg();
5259 unsigned oldval = MI->getOperand(2).getReg();
5260 unsigned newval = MI->getOperand(3).getReg();
5261 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5262 DebugLoc dl = MI->getDebugLoc();
5263 bool isThumb2 = Subtarget->isThumb2();
5265 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5266 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5267 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5268 (const TargetRegisterClass*)&ARM::GPRRegClass);
5271 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5272 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5273 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5276 unsigned ldrOpc, strOpc;
5278 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5280 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5281 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5284 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5285 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5288 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5289 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5293 MachineFunction *MF = BB->getParent();
5294 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5295 MachineFunction::iterator It = BB;
5296 ++It; // insert the new blocks after the current block
5298 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5299 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5300 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5301 MF->insert(It, loop1MBB);
5302 MF->insert(It, loop2MBB);
5303 MF->insert(It, exitMBB);
5305 // Transfer the remainder of BB and its successor edges to exitMBB.
5306 exitMBB->splice(exitMBB->begin(), BB,
5307 llvm::next(MachineBasicBlock::iterator(MI)),
5309 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5313 // fallthrough --> loop1MBB
5314 BB->addSuccessor(loop1MBB);
5317 // ldrex dest, [ptr]
5321 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5322 if (ldrOpc == ARM::t2LDREX)
5324 AddDefaultPred(MIB);
5325 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5326 .addReg(dest).addReg(oldval));
5327 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5328 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5329 BB->addSuccessor(loop2MBB);
5330 BB->addSuccessor(exitMBB);
5333 // strex scratch, newval, [ptr]
5337 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5338 if (strOpc == ARM::t2STREX)
5340 AddDefaultPred(MIB);
5341 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5342 .addReg(scratch).addImm(0));
5343 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5344 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5345 BB->addSuccessor(loop1MBB);
5346 BB->addSuccessor(exitMBB);
5352 MI->eraseFromParent(); // The instruction is gone now.
5358 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5359 unsigned Size, unsigned BinOpcode) const {
5360 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5361 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5363 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5364 MachineFunction *MF = BB->getParent();
5365 MachineFunction::iterator It = BB;
5368 unsigned dest = MI->getOperand(0).getReg();
5369 unsigned ptr = MI->getOperand(1).getReg();
5370 unsigned incr = MI->getOperand(2).getReg();
5371 DebugLoc dl = MI->getDebugLoc();
5372 bool isThumb2 = Subtarget->isThumb2();
5374 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5376 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5377 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5380 unsigned ldrOpc, strOpc;
5382 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5384 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5385 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5388 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5389 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5392 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5393 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5397 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5398 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5399 MF->insert(It, loopMBB);
5400 MF->insert(It, exitMBB);
5402 // Transfer the remainder of BB and its successor edges to exitMBB.
5403 exitMBB->splice(exitMBB->begin(), BB,
5404 llvm::next(MachineBasicBlock::iterator(MI)),
5406 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5408 const TargetRegisterClass *TRC = isThumb2 ?
5409 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5410 (const TargetRegisterClass*)&ARM::GPRRegClass;
5411 unsigned scratch = MRI.createVirtualRegister(TRC);
5412 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5416 // fallthrough --> loopMBB
5417 BB->addSuccessor(loopMBB);
5421 // <binop> scratch2, dest, incr
5422 // strex scratch, scratch2, ptr
5425 // fallthrough --> exitMBB
5427 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5428 if (ldrOpc == ARM::t2LDREX)
5430 AddDefaultPred(MIB);
5432 // operand order needs to go the other way for NAND
5433 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5434 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5435 addReg(incr).addReg(dest)).addReg(0);
5437 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5438 addReg(dest).addReg(incr)).addReg(0);
5441 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5442 if (strOpc == ARM::t2STREX)
5444 AddDefaultPred(MIB);
5445 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5446 .addReg(scratch).addImm(0));
5447 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5448 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5450 BB->addSuccessor(loopMBB);
5451 BB->addSuccessor(exitMBB);
5457 MI->eraseFromParent(); // The instruction is gone now.
5463 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5464 MachineBasicBlock *BB,
5467 ARMCC::CondCodes Cond) const {
5468 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5470 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5471 MachineFunction *MF = BB->getParent();
5472 MachineFunction::iterator It = BB;
5475 unsigned dest = MI->getOperand(0).getReg();
5476 unsigned ptr = MI->getOperand(1).getReg();
5477 unsigned incr = MI->getOperand(2).getReg();
5478 unsigned oldval = dest;
5479 DebugLoc dl = MI->getDebugLoc();
5480 bool isThumb2 = Subtarget->isThumb2();
5482 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5484 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5485 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5488 unsigned ldrOpc, strOpc, extendOpc;
5490 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5492 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5493 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5494 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5497 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5498 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5499 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5502 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5503 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5508 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5509 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5510 MF->insert(It, loopMBB);
5511 MF->insert(It, exitMBB);
5513 // Transfer the remainder of BB and its successor edges to exitMBB.
5514 exitMBB->splice(exitMBB->begin(), BB,
5515 llvm::next(MachineBasicBlock::iterator(MI)),
5517 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5519 const TargetRegisterClass *TRC = isThumb2 ?
5520 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5521 (const TargetRegisterClass*)&ARM::GPRRegClass;
5522 unsigned scratch = MRI.createVirtualRegister(TRC);
5523 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5527 // fallthrough --> loopMBB
5528 BB->addSuccessor(loopMBB);
5532 // (sign extend dest, if required)
5534 // cmov.cond scratch2, dest, incr
5535 // strex scratch, scratch2, ptr
5538 // fallthrough --> exitMBB
5540 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5541 if (ldrOpc == ARM::t2LDREX)
5543 AddDefaultPred(MIB);
5545 // Sign extend the value, if necessary.
5546 if (signExtend && extendOpc) {
5547 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5548 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5553 // Build compare and cmov instructions.
5554 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5555 .addReg(oldval).addReg(incr));
5556 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5557 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5559 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5560 if (strOpc == ARM::t2STREX)
5562 AddDefaultPred(MIB);
5563 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5564 .addReg(scratch).addImm(0));
5565 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5566 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5568 BB->addSuccessor(loopMBB);
5569 BB->addSuccessor(exitMBB);
5575 MI->eraseFromParent(); // The instruction is gone now.
5581 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5582 unsigned Op1, unsigned Op2,
5583 bool NeedsCarry, bool IsCmpxchg) const {
5584 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5585 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5587 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5588 MachineFunction *MF = BB->getParent();
5589 MachineFunction::iterator It = BB;
5592 unsigned destlo = MI->getOperand(0).getReg();
5593 unsigned desthi = MI->getOperand(1).getReg();
5594 unsigned ptr = MI->getOperand(2).getReg();
5595 unsigned vallo = MI->getOperand(3).getReg();
5596 unsigned valhi = MI->getOperand(4).getReg();
5597 DebugLoc dl = MI->getDebugLoc();
5598 bool isThumb2 = Subtarget->isThumb2();
5600 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5602 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5603 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5604 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5607 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5608 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5610 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5611 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5613 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5614 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5616 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5617 MF->insert(It, loopMBB);
5619 MF->insert(It, contBB);
5620 MF->insert(It, cont2BB);
5622 MF->insert(It, exitMBB);
5624 // Transfer the remainder of BB and its successor edges to exitMBB.
5625 exitMBB->splice(exitMBB->begin(), BB,
5626 llvm::next(MachineBasicBlock::iterator(MI)),
5628 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5630 const TargetRegisterClass *TRC = isThumb2 ?
5631 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5632 (const TargetRegisterClass*)&ARM::GPRRegClass;
5633 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5637 // fallthrough --> loopMBB
5638 BB->addSuccessor(loopMBB);
5641 // ldrexd r2, r3, ptr
5642 // <binopa> r0, r2, incr
5643 // <binopb> r1, r3, incr
5644 // strexd storesuccess, r0, r1, ptr
5645 // cmp storesuccess, #0
5647 // fallthrough --> exitMBB
5649 // Note that the registers are explicitly specified because there is not any
5650 // way to force the register allocator to allocate a register pair.
5652 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5653 // need to properly enforce the restriction that the two output registers
5654 // for ldrexd must be different.
5657 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5658 .addReg(ARM::R2, RegState::Define)
5659 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5660 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5661 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5662 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5666 for (unsigned i = 0; i < 2; i++) {
5667 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5669 .addReg(i == 0 ? destlo : desthi)
5670 .addReg(i == 0 ? vallo : valhi));
5671 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5672 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5673 BB->addSuccessor(exitMBB);
5674 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5675 BB = (i == 0 ? contBB : cont2BB);
5678 // Copy to physregs for strexd
5679 unsigned setlo = MI->getOperand(5).getReg();
5680 unsigned sethi = MI->getOperand(6).getReg();
5681 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5682 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5684 // Perform binary operation
5685 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5686 .addReg(destlo).addReg(vallo))
5687 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5688 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5689 .addReg(desthi).addReg(valhi)).addReg(0);
5691 // Copy to physregs for strexd
5692 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5693 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5697 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5698 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5700 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5701 .addReg(storesuccess).addImm(0));
5702 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5703 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5705 BB->addSuccessor(loopMBB);
5706 BB->addSuccessor(exitMBB);
5712 MI->eraseFromParent(); // The instruction is gone now.
5717 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5718 /// registers the function context.
5719 void ARMTargetLowering::
5720 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5721 MachineBasicBlock *DispatchBB, int FI) const {
5722 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5723 DebugLoc dl = MI->getDebugLoc();
5724 MachineFunction *MF = MBB->getParent();
5725 MachineRegisterInfo *MRI = &MF->getRegInfo();
5726 MachineConstantPool *MCP = MF->getConstantPool();
5727 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5728 const Function *F = MF->getFunction();
5730 bool isThumb = Subtarget->isThumb();
5731 bool isThumb2 = Subtarget->isThumb2();
5733 unsigned PCLabelId = AFI->createPICLabelUId();
5734 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5735 ARMConstantPoolValue *CPV =
5736 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5737 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5739 const TargetRegisterClass *TRC = isThumb ?
5740 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5741 (const TargetRegisterClass*)&ARM::GPRRegClass;
5743 // Grab constant pool and fixed stack memory operands.
5744 MachineMemOperand *CPMMO =
5745 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5746 MachineMemOperand::MOLoad, 4, 4);
5748 MachineMemOperand *FIMMOSt =
5749 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5750 MachineMemOperand::MOStore, 4, 4);
5752 // Load the address of the dispatch MBB into the jump buffer.
5754 // Incoming value: jbuf
5755 // ldr.n r5, LCPI1_1
5758 // str r5, [$jbuf, #+4] ; &jbuf[1]
5759 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5760 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5761 .addConstantPoolIndex(CPI)
5762 .addMemOperand(CPMMO));
5763 // Set the low bit because of thumb mode.
5764 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5766 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5767 .addReg(NewVReg1, RegState::Kill)
5769 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5770 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5771 .addReg(NewVReg2, RegState::Kill)
5773 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5774 .addReg(NewVReg3, RegState::Kill)
5776 .addImm(36) // &jbuf[1] :: pc
5777 .addMemOperand(FIMMOSt));
5778 } else if (isThumb) {
5779 // Incoming value: jbuf
5780 // ldr.n r1, LCPI1_4
5784 // add r2, $jbuf, #+4 ; &jbuf[1]
5786 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5787 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5788 .addConstantPoolIndex(CPI)
5789 .addMemOperand(CPMMO));
5790 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5791 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5792 .addReg(NewVReg1, RegState::Kill)
5794 // Set the low bit because of thumb mode.
5795 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5796 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5797 .addReg(ARM::CPSR, RegState::Define)
5799 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5800 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5801 .addReg(ARM::CPSR, RegState::Define)
5802 .addReg(NewVReg2, RegState::Kill)
5803 .addReg(NewVReg3, RegState::Kill));
5804 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5805 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5807 .addImm(36)); // &jbuf[1] :: pc
5808 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5809 .addReg(NewVReg4, RegState::Kill)
5810 .addReg(NewVReg5, RegState::Kill)
5812 .addMemOperand(FIMMOSt));
5814 // Incoming value: jbuf
5817 // str r1, [$jbuf, #+4] ; &jbuf[1]
5818 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5819 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5820 .addConstantPoolIndex(CPI)
5822 .addMemOperand(CPMMO));
5823 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5824 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5825 .addReg(NewVReg1, RegState::Kill)
5826 .addImm(PCLabelId));
5827 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5828 .addReg(NewVReg2, RegState::Kill)
5830 .addImm(36) // &jbuf[1] :: pc
5831 .addMemOperand(FIMMOSt));
5835 MachineBasicBlock *ARMTargetLowering::
5836 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5837 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5838 DebugLoc dl = MI->getDebugLoc();
5839 MachineFunction *MF = MBB->getParent();
5840 MachineRegisterInfo *MRI = &MF->getRegInfo();
5841 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5842 MachineFrameInfo *MFI = MF->getFrameInfo();
5843 int FI = MFI->getFunctionContextIndex();
5845 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5846 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5847 (const TargetRegisterClass*)&ARM::GPRRegClass;
5849 // Get a mapping of the call site numbers to all of the landing pads they're
5851 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5852 unsigned MaxCSNum = 0;
5853 MachineModuleInfo &MMI = MF->getMMI();
5854 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5856 if (!BB->isLandingPad()) continue;
5858 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5860 for (MachineBasicBlock::iterator
5861 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5862 if (!II->isEHLabel()) continue;
5864 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5865 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5867 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5868 for (SmallVectorImpl<unsigned>::iterator
5869 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5870 CSI != CSE; ++CSI) {
5871 CallSiteNumToLPad[*CSI].push_back(BB);
5872 MaxCSNum = std::max(MaxCSNum, *CSI);
5878 // Get an ordered list of the machine basic blocks for the jump table.
5879 std::vector<MachineBasicBlock*> LPadList;
5880 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5881 LPadList.reserve(CallSiteNumToLPad.size());
5882 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5883 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5884 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5885 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5886 LPadList.push_back(*II);
5887 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5891 assert(!LPadList.empty() &&
5892 "No landing pad destinations for the dispatch jump table!");
5894 // Create the jump table and associated information.
5895 MachineJumpTableInfo *JTI =
5896 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5897 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5898 unsigned UId = AFI->createJumpTableUId();
5900 // Create the MBBs for the dispatch code.
5902 // Shove the dispatch's address into the return slot in the function context.
5903 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5904 DispatchBB->setIsLandingPad();
5906 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5907 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5908 DispatchBB->addSuccessor(TrapBB);
5910 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5911 DispatchBB->addSuccessor(DispContBB);
5914 MF->insert(MF->end(), DispatchBB);
5915 MF->insert(MF->end(), DispContBB);
5916 MF->insert(MF->end(), TrapBB);
5918 // Insert code into the entry block that creates and registers the function
5920 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5922 MachineMemOperand *FIMMOLd =
5923 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5924 MachineMemOperand::MOLoad |
5925 MachineMemOperand::MOVolatile, 4, 4);
5927 if (AFI->isThumb1OnlyFunction())
5928 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5929 else if (!Subtarget->hasVFP2())
5930 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5932 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5934 unsigned NumLPads = LPadList.size();
5935 if (Subtarget->isThumb2()) {
5936 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5937 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5940 .addMemOperand(FIMMOLd));
5942 if (NumLPads < 256) {
5943 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5945 .addImm(LPadList.size()));
5947 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5948 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5949 .addImm(NumLPads & 0xFFFF));
5951 unsigned VReg2 = VReg1;
5952 if ((NumLPads & 0xFFFF0000) != 0) {
5953 VReg2 = MRI->createVirtualRegister(TRC);
5954 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5956 .addImm(NumLPads >> 16));
5959 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5964 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5969 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5970 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5971 .addJumpTableIndex(MJTI)
5974 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5977 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5978 .addReg(NewVReg3, RegState::Kill)
5980 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5982 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5983 .addReg(NewVReg4, RegState::Kill)
5985 .addJumpTableIndex(MJTI)
5987 } else if (Subtarget->isThumb()) {
5988 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5989 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5992 .addMemOperand(FIMMOLd));
5994 if (NumLPads < 256) {
5995 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5999 MachineConstantPool *ConstantPool = MF->getConstantPool();
6000 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6001 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6003 // MachineConstantPool wants an explicit alignment.
6004 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6006 Align = getTargetData()->getTypeAllocSize(C->getType());
6007 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6009 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6010 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6011 .addReg(VReg1, RegState::Define)
6012 .addConstantPoolIndex(Idx));
6013 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6018 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6023 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6024 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6025 .addReg(ARM::CPSR, RegState::Define)
6029 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6030 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6031 .addJumpTableIndex(MJTI)
6034 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6035 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6036 .addReg(ARM::CPSR, RegState::Define)
6037 .addReg(NewVReg2, RegState::Kill)
6040 MachineMemOperand *JTMMOLd =
6041 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6042 MachineMemOperand::MOLoad, 4, 4);
6044 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6045 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6046 .addReg(NewVReg4, RegState::Kill)
6048 .addMemOperand(JTMMOLd));
6050 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6051 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6052 .addReg(ARM::CPSR, RegState::Define)
6053 .addReg(NewVReg5, RegState::Kill)
6056 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6057 .addReg(NewVReg6, RegState::Kill)
6058 .addJumpTableIndex(MJTI)
6061 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6062 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6065 .addMemOperand(FIMMOLd));
6067 if (NumLPads < 256) {
6068 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6071 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6072 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6073 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6074 .addImm(NumLPads & 0xFFFF));
6076 unsigned VReg2 = VReg1;
6077 if ((NumLPads & 0xFFFF0000) != 0) {
6078 VReg2 = MRI->createVirtualRegister(TRC);
6079 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6081 .addImm(NumLPads >> 16));
6084 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6088 MachineConstantPool *ConstantPool = MF->getConstantPool();
6089 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6090 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6092 // MachineConstantPool wants an explicit alignment.
6093 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6095 Align = getTargetData()->getTypeAllocSize(C->getType());
6096 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6098 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6099 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6100 .addReg(VReg1, RegState::Define)
6101 .addConstantPoolIndex(Idx)
6103 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6105 .addReg(VReg1, RegState::Kill));
6108 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6113 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6115 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6117 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6118 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6119 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6120 .addJumpTableIndex(MJTI)
6123 MachineMemOperand *JTMMOLd =
6124 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6125 MachineMemOperand::MOLoad, 4, 4);
6126 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6128 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6129 .addReg(NewVReg3, RegState::Kill)
6132 .addMemOperand(JTMMOLd));
6134 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6135 .addReg(NewVReg5, RegState::Kill)
6137 .addJumpTableIndex(MJTI)
6141 // Add the jump table entries as successors to the MBB.
6142 MachineBasicBlock *PrevMBB = 0;
6143 for (std::vector<MachineBasicBlock*>::iterator
6144 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6145 MachineBasicBlock *CurMBB = *I;
6146 if (PrevMBB != CurMBB)
6147 DispContBB->addSuccessor(CurMBB);
6151 // N.B. the order the invoke BBs are processed in doesn't matter here.
6152 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6153 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6154 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6155 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6156 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6157 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6158 MachineBasicBlock *BB = *I;
6160 // Remove the landing pad successor from the invoke block and replace it
6161 // with the new dispatch block.
6162 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6164 while (!Successors.empty()) {
6165 MachineBasicBlock *SMBB = Successors.pop_back_val();
6166 if (SMBB->isLandingPad()) {
6167 BB->removeSuccessor(SMBB);
6168 MBBLPads.push_back(SMBB);
6172 BB->addSuccessor(DispatchBB);
6174 // Find the invoke call and mark all of the callee-saved registers as
6175 // 'implicit defined' so that they're spilled. This prevents code from
6176 // moving instructions to before the EH block, where they will never be
6178 for (MachineBasicBlock::reverse_iterator
6179 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6180 if (!II->isCall()) continue;
6182 DenseMap<unsigned, bool> DefRegs;
6183 for (MachineInstr::mop_iterator
6184 OI = II->operands_begin(), OE = II->operands_end();
6186 if (!OI->isReg()) continue;
6187 DefRegs[OI->getReg()] = true;
6190 MachineInstrBuilder MIB(&*II);
6192 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6193 unsigned Reg = SavedRegs[i];
6194 if (Subtarget->isThumb2() &&
6195 !ARM::tGPRRegClass.contains(Reg) &&
6196 !ARM::hGPRRegClass.contains(Reg))
6198 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6200 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6203 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6210 // Mark all former landing pads as non-landing pads. The dispatch is the only
6212 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6213 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6214 (*I)->setIsLandingPad(false);
6216 // The instruction is gone now.
6217 MI->eraseFromParent();
6223 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6224 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6225 E = MBB->succ_end(); I != E; ++I)
6228 llvm_unreachable("Expecting a BB with two successors!");
6232 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6233 MachineBasicBlock *BB) const {
6234 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6235 DebugLoc dl = MI->getDebugLoc();
6236 bool isThumb2 = Subtarget->isThumb2();
6237 switch (MI->getOpcode()) {
6240 llvm_unreachable("Unexpected instr type to insert");
6242 // The Thumb2 pre-indexed stores have the same MI operands, they just
6243 // define them differently in the .td files from the isel patterns, so
6244 // they need pseudos.
6245 case ARM::t2STR_preidx:
6246 MI->setDesc(TII->get(ARM::t2STR_PRE));
6248 case ARM::t2STRB_preidx:
6249 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6251 case ARM::t2STRH_preidx:
6252 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6255 case ARM::STRi_preidx:
6256 case ARM::STRBi_preidx: {
6257 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6258 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6259 // Decode the offset.
6260 unsigned Offset = MI->getOperand(4).getImm();
6261 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6262 Offset = ARM_AM::getAM2Offset(Offset);
6266 MachineMemOperand *MMO = *MI->memoperands_begin();
6267 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6268 .addOperand(MI->getOperand(0)) // Rn_wb
6269 .addOperand(MI->getOperand(1)) // Rt
6270 .addOperand(MI->getOperand(2)) // Rn
6271 .addImm(Offset) // offset (skip GPR==zero_reg)
6272 .addOperand(MI->getOperand(5)) // pred
6273 .addOperand(MI->getOperand(6))
6274 .addMemOperand(MMO);
6275 MI->eraseFromParent();
6278 case ARM::STRr_preidx:
6279 case ARM::STRBr_preidx:
6280 case ARM::STRH_preidx: {
6282 switch (MI->getOpcode()) {
6283 default: llvm_unreachable("unexpected opcode!");
6284 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6285 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6286 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6288 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6289 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6290 MIB.addOperand(MI->getOperand(i));
6291 MI->eraseFromParent();
6294 case ARM::ATOMIC_LOAD_ADD_I8:
6295 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6296 case ARM::ATOMIC_LOAD_ADD_I16:
6297 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6298 case ARM::ATOMIC_LOAD_ADD_I32:
6299 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6301 case ARM::ATOMIC_LOAD_AND_I8:
6302 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6303 case ARM::ATOMIC_LOAD_AND_I16:
6304 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6305 case ARM::ATOMIC_LOAD_AND_I32:
6306 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6308 case ARM::ATOMIC_LOAD_OR_I8:
6309 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6310 case ARM::ATOMIC_LOAD_OR_I16:
6311 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6312 case ARM::ATOMIC_LOAD_OR_I32:
6313 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6315 case ARM::ATOMIC_LOAD_XOR_I8:
6316 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6317 case ARM::ATOMIC_LOAD_XOR_I16:
6318 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6319 case ARM::ATOMIC_LOAD_XOR_I32:
6320 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6322 case ARM::ATOMIC_LOAD_NAND_I8:
6323 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6324 case ARM::ATOMIC_LOAD_NAND_I16:
6325 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6326 case ARM::ATOMIC_LOAD_NAND_I32:
6327 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6329 case ARM::ATOMIC_LOAD_SUB_I8:
6330 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6331 case ARM::ATOMIC_LOAD_SUB_I16:
6332 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6333 case ARM::ATOMIC_LOAD_SUB_I32:
6334 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6336 case ARM::ATOMIC_LOAD_MIN_I8:
6337 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6338 case ARM::ATOMIC_LOAD_MIN_I16:
6339 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6340 case ARM::ATOMIC_LOAD_MIN_I32:
6341 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6343 case ARM::ATOMIC_LOAD_MAX_I8:
6344 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6345 case ARM::ATOMIC_LOAD_MAX_I16:
6346 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6347 case ARM::ATOMIC_LOAD_MAX_I32:
6348 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6350 case ARM::ATOMIC_LOAD_UMIN_I8:
6351 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6352 case ARM::ATOMIC_LOAD_UMIN_I16:
6353 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6354 case ARM::ATOMIC_LOAD_UMIN_I32:
6355 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6357 case ARM::ATOMIC_LOAD_UMAX_I8:
6358 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6359 case ARM::ATOMIC_LOAD_UMAX_I16:
6360 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6361 case ARM::ATOMIC_LOAD_UMAX_I32:
6362 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6364 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6365 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6366 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6368 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6369 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6370 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6373 case ARM::ATOMADD6432:
6374 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6375 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6376 /*NeedsCarry*/ true);
6377 case ARM::ATOMSUB6432:
6378 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6379 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6380 /*NeedsCarry*/ true);
6381 case ARM::ATOMOR6432:
6382 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6383 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6384 case ARM::ATOMXOR6432:
6385 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6386 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6387 case ARM::ATOMAND6432:
6388 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6389 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6390 case ARM::ATOMSWAP6432:
6391 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6392 case ARM::ATOMCMPXCHG6432:
6393 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6394 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6395 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6397 case ARM::tMOVCCr_pseudo: {
6398 // To "insert" a SELECT_CC instruction, we actually have to insert the
6399 // diamond control-flow pattern. The incoming instruction knows the
6400 // destination vreg to set, the condition code register to branch on, the
6401 // true/false values to select between, and a branch opcode to use.
6402 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6403 MachineFunction::iterator It = BB;
6409 // cmpTY ccX, r1, r2
6411 // fallthrough --> copy0MBB
6412 MachineBasicBlock *thisMBB = BB;
6413 MachineFunction *F = BB->getParent();
6414 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6415 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6416 F->insert(It, copy0MBB);
6417 F->insert(It, sinkMBB);
6419 // Transfer the remainder of BB and its successor edges to sinkMBB.
6420 sinkMBB->splice(sinkMBB->begin(), BB,
6421 llvm::next(MachineBasicBlock::iterator(MI)),
6423 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6425 BB->addSuccessor(copy0MBB);
6426 BB->addSuccessor(sinkMBB);
6428 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6429 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6432 // %FalseValue = ...
6433 // # fallthrough to sinkMBB
6436 // Update machine-CFG edges
6437 BB->addSuccessor(sinkMBB);
6440 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6443 BuildMI(*BB, BB->begin(), dl,
6444 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6445 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6446 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6448 MI->eraseFromParent(); // The pseudo instruction is gone now.
6453 case ARM::BCCZi64: {
6454 // If there is an unconditional branch to the other successor, remove it.
6455 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6457 // Compare both parts that make up the double comparison separately for
6459 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6461 unsigned LHS1 = MI->getOperand(1).getReg();
6462 unsigned LHS2 = MI->getOperand(2).getReg();
6464 AddDefaultPred(BuildMI(BB, dl,
6465 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6466 .addReg(LHS1).addImm(0));
6467 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6468 .addReg(LHS2).addImm(0)
6469 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6471 unsigned RHS1 = MI->getOperand(3).getReg();
6472 unsigned RHS2 = MI->getOperand(4).getReg();
6473 AddDefaultPred(BuildMI(BB, dl,
6474 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6475 .addReg(LHS1).addReg(RHS1));
6476 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6477 .addReg(LHS2).addReg(RHS2)
6478 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6481 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6482 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6483 if (MI->getOperand(0).getImm() == ARMCC::NE)
6484 std::swap(destMBB, exitMBB);
6486 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6487 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6489 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6491 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6493 MI->eraseFromParent(); // The pseudo instruction is gone now.
6497 case ARM::Int_eh_sjlj_setjmp:
6498 case ARM::Int_eh_sjlj_setjmp_nofp:
6499 case ARM::tInt_eh_sjlj_setjmp:
6500 case ARM::t2Int_eh_sjlj_setjmp:
6501 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6502 EmitSjLjDispatchBlock(MI, BB);
6507 // To insert an ABS instruction, we have to insert the
6508 // diamond control-flow pattern. The incoming instruction knows the
6509 // source vreg to test against 0, the destination vreg to set,
6510 // the condition code register to branch on, the
6511 // true/false values to select between, and a branch opcode to use.
6516 // BCC (branch to SinkBB if V0 >= 0)
6517 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6518 // SinkBB: V1 = PHI(V2, V3)
6519 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6520 MachineFunction::iterator BBI = BB;
6522 MachineFunction *Fn = BB->getParent();
6523 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6524 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6525 Fn->insert(BBI, RSBBB);
6526 Fn->insert(BBI, SinkBB);
6528 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6529 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6530 bool isThumb2 = Subtarget->isThumb2();
6531 MachineRegisterInfo &MRI = Fn->getRegInfo();
6532 // In Thumb mode S must not be specified if source register is the SP or
6533 // PC and if destination register is the SP, so restrict register class
6534 unsigned NewMovDstReg = MRI.createVirtualRegister(isThumb2 ?
6535 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6536 (const TargetRegisterClass*)&ARM::GPRRegClass);
6537 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6538 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6539 (const TargetRegisterClass*)&ARM::GPRRegClass);
6541 // Transfer the remainder of BB and its successor edges to sinkMBB.
6542 SinkBB->splice(SinkBB->begin(), BB,
6543 llvm::next(MachineBasicBlock::iterator(MI)),
6545 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6547 BB->addSuccessor(RSBBB);
6548 BB->addSuccessor(SinkBB);
6550 // fall through to SinkMBB
6551 RSBBB->addSuccessor(SinkBB);
6553 // insert a movs at the end of BB
6554 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6556 .addReg(ABSSrcReg, RegState::Kill)
6557 .addImm((unsigned)ARMCC::AL).addReg(0)
6558 .addReg(ARM::CPSR, RegState::Define);
6560 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6562 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6563 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6565 // insert rsbri in RSBBB
6566 // Note: BCC and rsbri will be converted into predicated rsbmi
6567 // by if-conversion pass
6568 BuildMI(*RSBBB, RSBBB->begin(), dl,
6569 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6570 .addReg(NewMovDstReg, RegState::Kill)
6571 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6573 // insert PHI in SinkBB,
6574 // reuse ABSDstReg to not change uses of ABS instruction
6575 BuildMI(*SinkBB, SinkBB->begin(), dl,
6576 TII->get(ARM::PHI), ABSDstReg)
6577 .addReg(NewRsbDstReg).addMBB(RSBBB)
6578 .addReg(NewMovDstReg).addMBB(BB);
6580 // remove ABS instruction
6581 MI->eraseFromParent();
6583 // return last added BB
6589 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6590 SDNode *Node) const {
6591 if (!MI->hasPostISelHook()) {
6592 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6593 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6597 const MCInstrDesc *MCID = &MI->getDesc();
6598 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6599 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6600 // operand is still set to noreg. If needed, set the optional operand's
6601 // register to CPSR, and remove the redundant implicit def.
6603 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6605 // Rename pseudo opcodes.
6606 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6608 const ARMBaseInstrInfo *TII =
6609 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6610 MCID = &TII->get(NewOpc);
6612 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6613 "converted opcode should be the same except for cc_out");
6617 // Add the optional cc_out operand
6618 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6620 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6622 // Any ARM instruction that sets the 's' bit should specify an optional
6623 // "cc_out" operand in the last operand position.
6624 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6625 assert(!NewOpc && "Optional cc_out operand required");
6628 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6629 // since we already have an optional CPSR def.
6630 bool definesCPSR = false;
6631 bool deadCPSR = false;
6632 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6634 const MachineOperand &MO = MI->getOperand(i);
6635 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6639 MI->RemoveOperand(i);
6644 assert(!NewOpc && "Optional cc_out operand required");
6647 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6649 assert(!MI->getOperand(ccOutIdx).getReg() &&
6650 "expect uninitialized optional cc_out operand");
6654 // If this instruction was defined with an optional CPSR def and its dag node
6655 // had a live implicit CPSR def, then activate the optional CPSR def.
6656 MachineOperand &MO = MI->getOperand(ccOutIdx);
6657 MO.setReg(ARM::CPSR);
6661 //===----------------------------------------------------------------------===//
6662 // ARM Optimization Hooks
6663 //===----------------------------------------------------------------------===//
6666 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6667 TargetLowering::DAGCombinerInfo &DCI) {
6668 SelectionDAG &DAG = DCI.DAG;
6669 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6670 EVT VT = N->getValueType(0);
6671 unsigned Opc = N->getOpcode();
6672 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6673 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6674 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6675 ISD::CondCode CC = ISD::SETCC_INVALID;
6678 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6680 SDValue CCOp = Slct.getOperand(0);
6681 if (CCOp.getOpcode() == ISD::SETCC)
6682 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6685 bool DoXform = false;
6687 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6690 if (LHS.getOpcode() == ISD::Constant &&
6691 cast<ConstantSDNode>(LHS)->isNullValue()) {
6693 } else if (CC != ISD::SETCC_INVALID &&
6694 RHS.getOpcode() == ISD::Constant &&
6695 cast<ConstantSDNode>(RHS)->isNullValue()) {
6696 std::swap(LHS, RHS);
6697 SDValue Op0 = Slct.getOperand(0);
6698 EVT OpVT = isSlctCC ? Op0.getValueType() :
6699 Op0.getOperand(0).getValueType();
6700 bool isInt = OpVT.isInteger();
6701 CC = ISD::getSetCCInverse(CC, isInt);
6703 if (!TLI.isCondCodeLegal(CC, OpVT))
6704 return SDValue(); // Inverse operator isn't legal.
6711 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6713 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6714 Slct.getOperand(0), Slct.getOperand(1), CC);
6715 SDValue CCOp = Slct.getOperand(0);
6717 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6718 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6719 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6720 CCOp, OtherOp, Result);
6725 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6726 // (only after legalization).
6727 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6728 TargetLowering::DAGCombinerInfo &DCI,
6729 const ARMSubtarget *Subtarget) {
6731 // Only perform optimization if after legalize, and if NEON is available. We
6732 // also expected both operands to be BUILD_VECTORs.
6733 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6734 || N0.getOpcode() != ISD::BUILD_VECTOR
6735 || N1.getOpcode() != ISD::BUILD_VECTOR)
6738 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6739 EVT VT = N->getValueType(0);
6740 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6743 // Check that the vector operands are of the right form.
6744 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6745 // operands, where N is the size of the formed vector.
6746 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6747 // index such that we have a pair wise add pattern.
6749 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6750 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6752 SDValue Vec = N0->getOperand(0)->getOperand(0);
6753 SDNode *V = Vec.getNode();
6754 unsigned nextIndex = 0;
6756 // For each operands to the ADD which are BUILD_VECTORs,
6757 // check to see if each of their operands are an EXTRACT_VECTOR with
6758 // the same vector and appropriate index.
6759 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6760 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6761 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6763 SDValue ExtVec0 = N0->getOperand(i);
6764 SDValue ExtVec1 = N1->getOperand(i);
6766 // First operand is the vector, verify its the same.
6767 if (V != ExtVec0->getOperand(0).getNode() ||
6768 V != ExtVec1->getOperand(0).getNode())
6771 // Second is the constant, verify its correct.
6772 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6773 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6775 // For the constant, we want to see all the even or all the odd.
6776 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6777 || C1->getZExtValue() != nextIndex+1)
6786 // Create VPADDL node.
6787 SelectionDAG &DAG = DCI.DAG;
6788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6790 // Build operand list.
6791 SmallVector<SDValue, 8> Ops;
6792 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6793 TLI.getPointerTy()));
6795 // Input is the vector.
6798 // Get widened type and narrowed type.
6800 unsigned numElem = VT.getVectorNumElements();
6801 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6802 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6803 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6804 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6806 llvm_unreachable("Invalid vector element type for padd optimization.");
6809 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6810 widenType, &Ops[0], Ops.size());
6811 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6814 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6815 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6816 /// called with the default operands, and if that fails, with commuted
6818 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6819 TargetLowering::DAGCombinerInfo &DCI,
6820 const ARMSubtarget *Subtarget){
6822 // Attempt to create vpaddl for this add.
6823 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6824 if (Result.getNode())
6827 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6828 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6829 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6830 if (Result.getNode()) return Result;
6835 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6837 static SDValue PerformADDCombine(SDNode *N,
6838 TargetLowering::DAGCombinerInfo &DCI,
6839 const ARMSubtarget *Subtarget) {
6840 SDValue N0 = N->getOperand(0);
6841 SDValue N1 = N->getOperand(1);
6843 // First try with the default operand order.
6844 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6845 if (Result.getNode())
6848 // If that didn't work, try again with the operands commuted.
6849 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6852 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6854 static SDValue PerformSUBCombine(SDNode *N,
6855 TargetLowering::DAGCombinerInfo &DCI) {
6856 SDValue N0 = N->getOperand(0);
6857 SDValue N1 = N->getOperand(1);
6859 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6860 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6861 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6862 if (Result.getNode()) return Result;
6868 /// PerformVMULCombine
6869 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6870 /// special multiplier accumulator forwarding.
6876 static SDValue PerformVMULCombine(SDNode *N,
6877 TargetLowering::DAGCombinerInfo &DCI,
6878 const ARMSubtarget *Subtarget) {
6879 if (!Subtarget->hasVMLxForwarding())
6882 SelectionDAG &DAG = DCI.DAG;
6883 SDValue N0 = N->getOperand(0);
6884 SDValue N1 = N->getOperand(1);
6885 unsigned Opcode = N0.getOpcode();
6886 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6887 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6888 Opcode = N1.getOpcode();
6889 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6890 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6895 EVT VT = N->getValueType(0);
6896 DebugLoc DL = N->getDebugLoc();
6897 SDValue N00 = N0->getOperand(0);
6898 SDValue N01 = N0->getOperand(1);
6899 return DAG.getNode(Opcode, DL, VT,
6900 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6901 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6904 static SDValue PerformMULCombine(SDNode *N,
6905 TargetLowering::DAGCombinerInfo &DCI,
6906 const ARMSubtarget *Subtarget) {
6907 SelectionDAG &DAG = DCI.DAG;
6909 if (Subtarget->isThumb1Only())
6912 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6915 EVT VT = N->getValueType(0);
6916 if (VT.is64BitVector() || VT.is128BitVector())
6917 return PerformVMULCombine(N, DCI, Subtarget);
6921 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6925 int64_t MulAmt = C->getSExtValue();
6926 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6928 ShiftAmt = ShiftAmt & (32 - 1);
6929 SDValue V = N->getOperand(0);
6930 DebugLoc DL = N->getDebugLoc();
6933 MulAmt >>= ShiftAmt;
6936 if (isPowerOf2_32(MulAmt - 1)) {
6937 // (mul x, 2^N + 1) => (add (shl x, N), x)
6938 Res = DAG.getNode(ISD::ADD, DL, VT,
6940 DAG.getNode(ISD::SHL, DL, VT,
6942 DAG.getConstant(Log2_32(MulAmt - 1),
6944 } else if (isPowerOf2_32(MulAmt + 1)) {
6945 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6946 Res = DAG.getNode(ISD::SUB, DL, VT,
6947 DAG.getNode(ISD::SHL, DL, VT,
6949 DAG.getConstant(Log2_32(MulAmt + 1),
6955 uint64_t MulAmtAbs = -MulAmt;
6956 if (isPowerOf2_32(MulAmtAbs + 1)) {
6957 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6958 Res = DAG.getNode(ISD::SUB, DL, VT,
6960 DAG.getNode(ISD::SHL, DL, VT,
6962 DAG.getConstant(Log2_32(MulAmtAbs + 1),
6964 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
6965 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6966 Res = DAG.getNode(ISD::ADD, DL, VT,
6968 DAG.getNode(ISD::SHL, DL, VT,
6970 DAG.getConstant(Log2_32(MulAmtAbs-1),
6972 Res = DAG.getNode(ISD::SUB, DL, VT,
6973 DAG.getConstant(0, MVT::i32),Res);
6980 Res = DAG.getNode(ISD::SHL, DL, VT,
6981 Res, DAG.getConstant(ShiftAmt, MVT::i32));
6983 // Do not add new nodes to DAG combiner worklist.
6984 DCI.CombineTo(N, Res, false);
6988 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6989 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6992 SDValue FalseVal = N.getOperand(0);
6993 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6997 return C->isAllOnesValue();
6998 return C->isNullValue();
7001 /// formConditionalOp - Combine an operation with a conditional move operand
7002 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
7003 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
7004 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
7006 SDValue N0 = N->getOperand(0);
7007 SDValue N1 = N->getOperand(1);
7009 bool isAND = N->getOpcode() == ISD::AND;
7010 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
7011 if (!isCand && Commutable) {
7012 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
7020 switch (N->getOpcode()) {
7021 default: llvm_unreachable("Unexpected node");
7022 case ISD::AND: Opc = ARMISD::CAND; break;
7023 case ISD::OR: Opc = ARMISD::COR; break;
7024 case ISD::XOR: Opc = ARMISD::CXOR; break;
7026 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
7027 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
7031 static SDValue PerformANDCombine(SDNode *N,
7032 TargetLowering::DAGCombinerInfo &DCI,
7033 const ARMSubtarget *Subtarget) {
7035 // Attempt to use immediate-form VBIC
7036 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7037 DebugLoc dl = N->getDebugLoc();
7038 EVT VT = N->getValueType(0);
7039 SelectionDAG &DAG = DCI.DAG;
7041 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7044 APInt SplatBits, SplatUndef;
7045 unsigned SplatBitSize;
7048 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7049 if (SplatBitSize <= 64) {
7051 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7052 SplatUndef.getZExtValue(), SplatBitSize,
7053 DAG, VbicVT, VT.is128BitVector(),
7055 if (Val.getNode()) {
7057 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7058 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7059 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7064 if (!Subtarget->isThumb1Only()) {
7065 // (and x, (cmov -1, y, cond)) => (and.cond x, y)
7066 SDValue CAND = formConditionalOp(N, DAG, true);
7074 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7075 static SDValue PerformORCombine(SDNode *N,
7076 TargetLowering::DAGCombinerInfo &DCI,
7077 const ARMSubtarget *Subtarget) {
7078 // Attempt to use immediate-form VORR
7079 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7080 DebugLoc dl = N->getDebugLoc();
7081 EVT VT = N->getValueType(0);
7082 SelectionDAG &DAG = DCI.DAG;
7084 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7087 APInt SplatBits, SplatUndef;
7088 unsigned SplatBitSize;
7090 if (BVN && Subtarget->hasNEON() &&
7091 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7092 if (SplatBitSize <= 64) {
7094 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7095 SplatUndef.getZExtValue(), SplatBitSize,
7096 DAG, VorrVT, VT.is128BitVector(),
7098 if (Val.getNode()) {
7100 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7101 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7102 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7107 if (!Subtarget->isThumb1Only()) {
7108 // (or x, (cmov 0, y, cond)) => (or.cond x, y)
7109 SDValue COR = formConditionalOp(N, DAG, true);
7114 SDValue N0 = N->getOperand(0);
7115 if (N0.getOpcode() != ISD::AND)
7117 SDValue N1 = N->getOperand(1);
7119 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7120 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7121 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7123 unsigned SplatBitSize;
7126 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7128 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7129 HasAnyUndefs) && !HasAnyUndefs) {
7130 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7132 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7133 HasAnyUndefs) && !HasAnyUndefs &&
7134 SplatBits0 == ~SplatBits1) {
7135 // Canonicalize the vector type to make instruction selection simpler.
7136 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7137 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7138 N0->getOperand(1), N0->getOperand(0),
7140 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7145 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7148 // BFI is only available on V6T2+
7149 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7152 DebugLoc DL = N->getDebugLoc();
7153 // 1) or (and A, mask), val => ARMbfi A, val, mask
7154 // iff (val & mask) == val
7156 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7157 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7158 // && mask == ~mask2
7159 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7160 // && ~mask == mask2
7161 // (i.e., copy a bitfield value into another bitfield of the same width)
7166 SDValue N00 = N0.getOperand(0);
7168 // The value and the mask need to be constants so we can verify this is
7169 // actually a bitfield set. If the mask is 0xffff, we can do better
7170 // via a movt instruction, so don't use BFI in that case.
7171 SDValue MaskOp = N0.getOperand(1);
7172 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7175 unsigned Mask = MaskC->getZExtValue();
7179 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7180 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7182 unsigned Val = N1C->getZExtValue();
7183 if ((Val & ~Mask) != Val)
7186 if (ARM::isBitFieldInvertedMask(Mask)) {
7187 Val >>= CountTrailingZeros_32(~Mask);
7189 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7190 DAG.getConstant(Val, MVT::i32),
7191 DAG.getConstant(Mask, MVT::i32));
7193 // Do not add new nodes to DAG combiner worklist.
7194 DCI.CombineTo(N, Res, false);
7197 } else if (N1.getOpcode() == ISD::AND) {
7198 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7199 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7202 unsigned Mask2 = N11C->getZExtValue();
7204 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7206 if (ARM::isBitFieldInvertedMask(Mask) &&
7208 // The pack halfword instruction works better for masks that fit it,
7209 // so use that when it's available.
7210 if (Subtarget->hasT2ExtractPack() &&
7211 (Mask == 0xffff || Mask == 0xffff0000))
7214 unsigned amt = CountTrailingZeros_32(Mask2);
7215 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7216 DAG.getConstant(amt, MVT::i32));
7217 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7218 DAG.getConstant(Mask, MVT::i32));
7219 // Do not add new nodes to DAG combiner worklist.
7220 DCI.CombineTo(N, Res, false);
7222 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7224 // The pack halfword instruction works better for masks that fit it,
7225 // so use that when it's available.
7226 if (Subtarget->hasT2ExtractPack() &&
7227 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7230 unsigned lsb = CountTrailingZeros_32(Mask);
7231 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7232 DAG.getConstant(lsb, MVT::i32));
7233 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7234 DAG.getConstant(Mask2, MVT::i32));
7235 // Do not add new nodes to DAG combiner worklist.
7236 DCI.CombineTo(N, Res, false);
7241 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7242 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7243 ARM::isBitFieldInvertedMask(~Mask)) {
7244 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7245 // where lsb(mask) == #shamt and masked bits of B are known zero.
7246 SDValue ShAmt = N00.getOperand(1);
7247 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7248 unsigned LSB = CountTrailingZeros_32(Mask);
7252 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7253 DAG.getConstant(~Mask, MVT::i32));
7255 // Do not add new nodes to DAG combiner worklist.
7256 DCI.CombineTo(N, Res, false);
7262 static SDValue PerformXORCombine(SDNode *N,
7263 TargetLowering::DAGCombinerInfo &DCI,
7264 const ARMSubtarget *Subtarget) {
7265 EVT VT = N->getValueType(0);
7266 SelectionDAG &DAG = DCI.DAG;
7268 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7271 if (!Subtarget->isThumb1Only()) {
7272 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7273 SDValue CXOR = formConditionalOp(N, DAG, true);
7281 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7282 /// the bits being cleared by the AND are not demanded by the BFI.
7283 static SDValue PerformBFICombine(SDNode *N,
7284 TargetLowering::DAGCombinerInfo &DCI) {
7285 SDValue N1 = N->getOperand(1);
7286 if (N1.getOpcode() == ISD::AND) {
7287 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7290 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7291 unsigned LSB = CountTrailingZeros_32(~InvMask);
7292 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7293 unsigned Mask = (1 << Width)-1;
7294 unsigned Mask2 = N11C->getZExtValue();
7295 if ((Mask & (~Mask2)) == 0)
7296 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7297 N->getOperand(0), N1.getOperand(0),
7303 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7304 /// ARMISD::VMOVRRD.
7305 static SDValue PerformVMOVRRDCombine(SDNode *N,
7306 TargetLowering::DAGCombinerInfo &DCI) {
7307 // vmovrrd(vmovdrr x, y) -> x,y
7308 SDValue InDouble = N->getOperand(0);
7309 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7310 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7312 // vmovrrd(load f64) -> (load i32), (load i32)
7313 SDNode *InNode = InDouble.getNode();
7314 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7315 InNode->getValueType(0) == MVT::f64 &&
7316 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7317 !cast<LoadSDNode>(InNode)->isVolatile()) {
7318 // TODO: Should this be done for non-FrameIndex operands?
7319 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7321 SelectionDAG &DAG = DCI.DAG;
7322 DebugLoc DL = LD->getDebugLoc();
7323 SDValue BasePtr = LD->getBasePtr();
7324 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7325 LD->getPointerInfo(), LD->isVolatile(),
7326 LD->isNonTemporal(), LD->isInvariant(),
7327 LD->getAlignment());
7329 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7330 DAG.getConstant(4, MVT::i32));
7331 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7332 LD->getPointerInfo(), LD->isVolatile(),
7333 LD->isNonTemporal(), LD->isInvariant(),
7334 std::min(4U, LD->getAlignment() / 2));
7336 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7337 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7338 DCI.RemoveFromWorklist(LD);
7346 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7347 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7348 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7349 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7350 SDValue Op0 = N->getOperand(0);
7351 SDValue Op1 = N->getOperand(1);
7352 if (Op0.getOpcode() == ISD::BITCAST)
7353 Op0 = Op0.getOperand(0);
7354 if (Op1.getOpcode() == ISD::BITCAST)
7355 Op1 = Op1.getOperand(0);
7356 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7357 Op0.getNode() == Op1.getNode() &&
7358 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7359 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7360 N->getValueType(0), Op0.getOperand(0));
7364 /// PerformSTORECombine - Target-specific dag combine xforms for
7366 static SDValue PerformSTORECombine(SDNode *N,
7367 TargetLowering::DAGCombinerInfo &DCI) {
7368 StoreSDNode *St = cast<StoreSDNode>(N);
7369 if (St->isVolatile())
7372 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7373 // pack all of the elements in one place. Next, store to memory in fewer
7375 SDValue StVal = St->getValue();
7376 EVT VT = StVal.getValueType();
7377 if (St->isTruncatingStore() && VT.isVector()) {
7378 SelectionDAG &DAG = DCI.DAG;
7379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7380 EVT StVT = St->getMemoryVT();
7381 unsigned NumElems = VT.getVectorNumElements();
7382 assert(StVT != VT && "Cannot truncate to the same type");
7383 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7384 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7386 // From, To sizes and ElemCount must be pow of two
7387 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7389 // We are going to use the original vector elt for storing.
7390 // Accumulated smaller vector elements must be a multiple of the store size.
7391 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7393 unsigned SizeRatio = FromEltSz / ToEltSz;
7394 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7396 // Create a type on which we perform the shuffle.
7397 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7398 NumElems*SizeRatio);
7399 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7401 DebugLoc DL = St->getDebugLoc();
7402 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
7403 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
7404 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
7406 // Can't shuffle using an illegal type.
7407 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
7409 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
7410 DAG.getUNDEF(WideVec.getValueType()),
7412 // At this point all of the data is stored at the bottom of the
7413 // register. We now need to save it to mem.
7415 // Find the largest store unit
7416 MVT StoreType = MVT::i8;
7417 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
7418 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
7419 MVT Tp = (MVT::SimpleValueType)tp;
7420 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
7423 // Didn't find a legal store type.
7424 if (!TLI.isTypeLegal(StoreType))
7427 // Bitcast the original vector into a vector of store-size units
7428 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
7429 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
7430 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
7431 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
7432 SmallVector<SDValue, 8> Chains;
7433 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
7434 TLI.getPointerTy());
7435 SDValue BasePtr = St->getBasePtr();
7437 // Perform one or more big stores into memory.
7438 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
7439 for (unsigned I = 0; I < E; I++) {
7440 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
7441 StoreType, ShuffWide,
7442 DAG.getIntPtrConstant(I));
7443 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
7444 St->getPointerInfo(), St->isVolatile(),
7445 St->isNonTemporal(), St->getAlignment());
7446 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
7448 Chains.push_back(Ch);
7450 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
7454 if (!ISD::isNormalStore(St))
7457 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
7458 // ARM stores of arguments in the same cache line.
7459 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7460 StVal.getNode()->hasOneUse()) {
7461 SelectionDAG &DAG = DCI.DAG;
7462 DebugLoc DL = St->getDebugLoc();
7463 SDValue BasePtr = St->getBasePtr();
7464 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7465 StVal.getNode()->getOperand(0), BasePtr,
7466 St->getPointerInfo(), St->isVolatile(),
7467 St->isNonTemporal(), St->getAlignment());
7469 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7470 DAG.getConstant(4, MVT::i32));
7471 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7472 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7473 St->isNonTemporal(),
7474 std::min(4U, St->getAlignment() / 2));
7477 if (StVal.getValueType() != MVT::i64 ||
7478 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7481 // Bitcast an i64 store extracted from a vector to f64.
7482 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7483 SelectionDAG &DAG = DCI.DAG;
7484 DebugLoc dl = StVal.getDebugLoc();
7485 SDValue IntVec = StVal.getOperand(0);
7486 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7487 IntVec.getValueType().getVectorNumElements());
7488 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7489 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7490 Vec, StVal.getOperand(1));
7491 dl = N->getDebugLoc();
7492 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7493 // Make the DAGCombiner fold the bitcasts.
7494 DCI.AddToWorklist(Vec.getNode());
7495 DCI.AddToWorklist(ExtElt.getNode());
7496 DCI.AddToWorklist(V.getNode());
7497 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7498 St->getPointerInfo(), St->isVolatile(),
7499 St->isNonTemporal(), St->getAlignment(),
7503 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7504 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7505 /// i64 vector to have f64 elements, since the value can then be loaded
7506 /// directly into a VFP register.
7507 static bool hasNormalLoadOperand(SDNode *N) {
7508 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7509 for (unsigned i = 0; i < NumElts; ++i) {
7510 SDNode *Elt = N->getOperand(i).getNode();
7511 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7517 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7518 /// ISD::BUILD_VECTOR.
7519 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7520 TargetLowering::DAGCombinerInfo &DCI){
7521 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7522 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7523 // into a pair of GPRs, which is fine when the value is used as a scalar,
7524 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7525 SelectionDAG &DAG = DCI.DAG;
7526 if (N->getNumOperands() == 2) {
7527 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7532 // Load i64 elements as f64 values so that type legalization does not split
7533 // them up into i32 values.
7534 EVT VT = N->getValueType(0);
7535 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7537 DebugLoc dl = N->getDebugLoc();
7538 SmallVector<SDValue, 8> Ops;
7539 unsigned NumElts = VT.getVectorNumElements();
7540 for (unsigned i = 0; i < NumElts; ++i) {
7541 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7543 // Make the DAGCombiner fold the bitcast.
7544 DCI.AddToWorklist(V.getNode());
7546 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7547 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7548 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7551 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7552 /// ISD::INSERT_VECTOR_ELT.
7553 static SDValue PerformInsertEltCombine(SDNode *N,
7554 TargetLowering::DAGCombinerInfo &DCI) {
7555 // Bitcast an i64 load inserted into a vector to f64.
7556 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7557 EVT VT = N->getValueType(0);
7558 SDNode *Elt = N->getOperand(1).getNode();
7559 if (VT.getVectorElementType() != MVT::i64 ||
7560 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7563 SelectionDAG &DAG = DCI.DAG;
7564 DebugLoc dl = N->getDebugLoc();
7565 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7566 VT.getVectorNumElements());
7567 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7568 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7569 // Make the DAGCombiner fold the bitcasts.
7570 DCI.AddToWorklist(Vec.getNode());
7571 DCI.AddToWorklist(V.getNode());
7572 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7573 Vec, V, N->getOperand(2));
7574 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7577 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7578 /// ISD::VECTOR_SHUFFLE.
7579 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7580 // The LLVM shufflevector instruction does not require the shuffle mask
7581 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7582 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7583 // operands do not match the mask length, they are extended by concatenating
7584 // them with undef vectors. That is probably the right thing for other
7585 // targets, but for NEON it is better to concatenate two double-register
7586 // size vector operands into a single quad-register size vector. Do that
7587 // transformation here:
7588 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7589 // shuffle(concat(v1, v2), undef)
7590 SDValue Op0 = N->getOperand(0);
7591 SDValue Op1 = N->getOperand(1);
7592 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7593 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7594 Op0.getNumOperands() != 2 ||
7595 Op1.getNumOperands() != 2)
7597 SDValue Concat0Op1 = Op0.getOperand(1);
7598 SDValue Concat1Op1 = Op1.getOperand(1);
7599 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7600 Concat1Op1.getOpcode() != ISD::UNDEF)
7602 // Skip the transformation if any of the types are illegal.
7603 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7604 EVT VT = N->getValueType(0);
7605 if (!TLI.isTypeLegal(VT) ||
7606 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7607 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7610 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7611 Op0.getOperand(0), Op1.getOperand(0));
7612 // Translate the shuffle mask.
7613 SmallVector<int, 16> NewMask;
7614 unsigned NumElts = VT.getVectorNumElements();
7615 unsigned HalfElts = NumElts/2;
7616 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7617 for (unsigned n = 0; n < NumElts; ++n) {
7618 int MaskElt = SVN->getMaskElt(n);
7620 if (MaskElt < (int)HalfElts)
7622 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7623 NewElt = HalfElts + MaskElt - NumElts;
7624 NewMask.push_back(NewElt);
7626 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7627 DAG.getUNDEF(VT), NewMask.data());
7630 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7631 /// NEON load/store intrinsics to merge base address updates.
7632 static SDValue CombineBaseUpdate(SDNode *N,
7633 TargetLowering::DAGCombinerInfo &DCI) {
7634 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7637 SelectionDAG &DAG = DCI.DAG;
7638 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7639 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7640 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7641 SDValue Addr = N->getOperand(AddrOpIdx);
7643 // Search for a use of the address operand that is an increment.
7644 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7645 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7647 if (User->getOpcode() != ISD::ADD ||
7648 UI.getUse().getResNo() != Addr.getResNo())
7651 // Check that the add is independent of the load/store. Otherwise, folding
7652 // it would create a cycle.
7653 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7656 // Find the new opcode for the updating load/store.
7658 bool isLaneOp = false;
7659 unsigned NewOpc = 0;
7660 unsigned NumVecs = 0;
7662 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7664 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7665 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7667 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7669 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7671 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7673 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7674 NumVecs = 2; isLaneOp = true; break;
7675 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7676 NumVecs = 3; isLaneOp = true; break;
7677 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7678 NumVecs = 4; isLaneOp = true; break;
7679 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7680 NumVecs = 1; isLoad = false; break;
7681 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7682 NumVecs = 2; isLoad = false; break;
7683 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7684 NumVecs = 3; isLoad = false; break;
7685 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7686 NumVecs = 4; isLoad = false; break;
7687 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7688 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7689 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7690 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7691 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7692 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7696 switch (N->getOpcode()) {
7697 default: llvm_unreachable("unexpected opcode for Neon base update");
7698 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7699 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7700 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7704 // Find the size of memory referenced by the load/store.
7707 VecTy = N->getValueType(0);
7709 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7710 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7712 NumBytes /= VecTy.getVectorNumElements();
7714 // If the increment is a constant, it must match the memory ref size.
7715 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7716 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7717 uint64_t IncVal = CInc->getZExtValue();
7718 if (IncVal != NumBytes)
7720 } else if (NumBytes >= 3 * 16) {
7721 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7722 // separate instructions that make it harder to use a non-constant update.
7726 // Create the new updating load/store node.
7728 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7730 for (n = 0; n < NumResultVecs; ++n)
7732 Tys[n++] = MVT::i32;
7733 Tys[n] = MVT::Other;
7734 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7735 SmallVector<SDValue, 8> Ops;
7736 Ops.push_back(N->getOperand(0)); // incoming chain
7737 Ops.push_back(N->getOperand(AddrOpIdx));
7739 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7740 Ops.push_back(N->getOperand(i));
7742 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7743 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7744 Ops.data(), Ops.size(),
7745 MemInt->getMemoryVT(),
7746 MemInt->getMemOperand());
7749 std::vector<SDValue> NewResults;
7750 for (unsigned i = 0; i < NumResultVecs; ++i) {
7751 NewResults.push_back(SDValue(UpdN.getNode(), i));
7753 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7754 DCI.CombineTo(N, NewResults);
7755 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7762 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7763 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7764 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7766 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7767 SelectionDAG &DAG = DCI.DAG;
7768 EVT VT = N->getValueType(0);
7769 // vldN-dup instructions only support 64-bit vectors for N > 1.
7770 if (!VT.is64BitVector())
7773 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7774 SDNode *VLD = N->getOperand(0).getNode();
7775 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7777 unsigned NumVecs = 0;
7778 unsigned NewOpc = 0;
7779 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7780 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7782 NewOpc = ARMISD::VLD2DUP;
7783 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7785 NewOpc = ARMISD::VLD3DUP;
7786 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7788 NewOpc = ARMISD::VLD4DUP;
7793 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7794 // numbers match the load.
7795 unsigned VLDLaneNo =
7796 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7797 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7799 // Ignore uses of the chain result.
7800 if (UI.getUse().getResNo() == NumVecs)
7803 if (User->getOpcode() != ARMISD::VDUPLANE ||
7804 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7808 // Create the vldN-dup node.
7811 for (n = 0; n < NumVecs; ++n)
7813 Tys[n] = MVT::Other;
7814 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7815 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7816 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7817 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7818 Ops, 2, VLDMemInt->getMemoryVT(),
7819 VLDMemInt->getMemOperand());
7822 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7824 unsigned ResNo = UI.getUse().getResNo();
7825 // Ignore uses of the chain result.
7826 if (ResNo == NumVecs)
7829 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7832 // Now the vldN-lane intrinsic is dead except for its chain result.
7833 // Update uses of the chain.
7834 std::vector<SDValue> VLDDupResults;
7835 for (unsigned n = 0; n < NumVecs; ++n)
7836 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7837 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7838 DCI.CombineTo(VLD, VLDDupResults);
7843 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7844 /// ARMISD::VDUPLANE.
7845 static SDValue PerformVDUPLANECombine(SDNode *N,
7846 TargetLowering::DAGCombinerInfo &DCI) {
7847 SDValue Op = N->getOperand(0);
7849 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7850 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7851 if (CombineVLDDUP(N, DCI))
7852 return SDValue(N, 0);
7854 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7855 // redundant. Ignore bit_converts for now; element sizes are checked below.
7856 while (Op.getOpcode() == ISD::BITCAST)
7857 Op = Op.getOperand(0);
7858 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7861 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7862 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7863 // The canonical VMOV for a zero vector uses a 32-bit element size.
7864 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7866 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7868 EVT VT = N->getValueType(0);
7869 if (EltSize > VT.getVectorElementType().getSizeInBits())
7872 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7875 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7876 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7877 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7881 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7883 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7888 APFloat APF = C->getValueAPF();
7889 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7890 != APFloat::opOK || !isExact)
7893 c0 = (I == 0) ? cN : c0;
7894 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7901 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7902 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7903 /// when the VMUL has a constant operand that is a power of 2.
7905 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7906 /// vmul.f32 d16, d17, d16
7907 /// vcvt.s32.f32 d16, d16
7909 /// vcvt.s32.f32 d16, d16, #3
7910 static SDValue PerformVCVTCombine(SDNode *N,
7911 TargetLowering::DAGCombinerInfo &DCI,
7912 const ARMSubtarget *Subtarget) {
7913 SelectionDAG &DAG = DCI.DAG;
7914 SDValue Op = N->getOperand(0);
7916 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7917 Op.getOpcode() != ISD::FMUL)
7921 SDValue N0 = Op->getOperand(0);
7922 SDValue ConstVec = Op->getOperand(1);
7923 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7925 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7926 !isConstVecPow2(ConstVec, isSigned, C))
7929 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7930 Intrinsic::arm_neon_vcvtfp2fxu;
7931 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7933 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7934 DAG.getConstant(Log2_64(C), MVT::i32));
7937 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7938 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7939 /// when the VDIV has a constant operand that is a power of 2.
7941 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7942 /// vcvt.f32.s32 d16, d16
7943 /// vdiv.f32 d16, d17, d16
7945 /// vcvt.f32.s32 d16, d16, #3
7946 static SDValue PerformVDIVCombine(SDNode *N,
7947 TargetLowering::DAGCombinerInfo &DCI,
7948 const ARMSubtarget *Subtarget) {
7949 SelectionDAG &DAG = DCI.DAG;
7950 SDValue Op = N->getOperand(0);
7951 unsigned OpOpcode = Op.getNode()->getOpcode();
7953 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7954 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7958 SDValue ConstVec = N->getOperand(1);
7959 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7961 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7962 !isConstVecPow2(ConstVec, isSigned, C))
7965 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7966 Intrinsic::arm_neon_vcvtfxu2fp;
7967 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7969 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7970 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7973 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7974 /// operand of a vector shift operation, where all the elements of the
7975 /// build_vector must have the same constant integer value.
7976 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7977 // Ignore bit_converts.
7978 while (Op.getOpcode() == ISD::BITCAST)
7979 Op = Op.getOperand(0);
7980 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7981 APInt SplatBits, SplatUndef;
7982 unsigned SplatBitSize;
7984 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7985 HasAnyUndefs, ElementBits) ||
7986 SplatBitSize > ElementBits)
7988 Cnt = SplatBits.getSExtValue();
7992 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7993 /// operand of a vector shift left operation. That value must be in the range:
7994 /// 0 <= Value < ElementBits for a left shift; or
7995 /// 0 <= Value <= ElementBits for a long left shift.
7996 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7997 assert(VT.isVector() && "vector shift count is not a vector type");
7998 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7999 if (! getVShiftImm(Op, ElementBits, Cnt))
8001 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8004 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8005 /// operand of a vector shift right operation. For a shift opcode, the value
8006 /// is positive, but for an intrinsic the value count must be negative. The
8007 /// absolute value must be in the range:
8008 /// 1 <= |Value| <= ElementBits for a right shift; or
8009 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8010 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8012 assert(VT.isVector() && "vector shift count is not a vector type");
8013 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8014 if (! getVShiftImm(Op, ElementBits, Cnt))
8018 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8021 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8022 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8023 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8026 // Don't do anything for most intrinsics.
8029 // Vector shifts: check for immediate versions and lower them.
8030 // Note: This is done during DAG combining instead of DAG legalizing because
8031 // the build_vectors for 64-bit vector element shift counts are generally
8032 // not legal, and it is hard to see their values after they get legalized to
8033 // loads from a constant pool.
8034 case Intrinsic::arm_neon_vshifts:
8035 case Intrinsic::arm_neon_vshiftu:
8036 case Intrinsic::arm_neon_vshiftls:
8037 case Intrinsic::arm_neon_vshiftlu:
8038 case Intrinsic::arm_neon_vshiftn:
8039 case Intrinsic::arm_neon_vrshifts:
8040 case Intrinsic::arm_neon_vrshiftu:
8041 case Intrinsic::arm_neon_vrshiftn:
8042 case Intrinsic::arm_neon_vqshifts:
8043 case Intrinsic::arm_neon_vqshiftu:
8044 case Intrinsic::arm_neon_vqshiftsu:
8045 case Intrinsic::arm_neon_vqshiftns:
8046 case Intrinsic::arm_neon_vqshiftnu:
8047 case Intrinsic::arm_neon_vqshiftnsu:
8048 case Intrinsic::arm_neon_vqrshiftns:
8049 case Intrinsic::arm_neon_vqrshiftnu:
8050 case Intrinsic::arm_neon_vqrshiftnsu: {
8051 EVT VT = N->getOperand(1).getValueType();
8053 unsigned VShiftOpc = 0;
8056 case Intrinsic::arm_neon_vshifts:
8057 case Intrinsic::arm_neon_vshiftu:
8058 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8059 VShiftOpc = ARMISD::VSHL;
8062 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8063 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8064 ARMISD::VSHRs : ARMISD::VSHRu);
8069 case Intrinsic::arm_neon_vshiftls:
8070 case Intrinsic::arm_neon_vshiftlu:
8071 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8073 llvm_unreachable("invalid shift count for vshll intrinsic");
8075 case Intrinsic::arm_neon_vrshifts:
8076 case Intrinsic::arm_neon_vrshiftu:
8077 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8081 case Intrinsic::arm_neon_vqshifts:
8082 case Intrinsic::arm_neon_vqshiftu:
8083 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8087 case Intrinsic::arm_neon_vqshiftsu:
8088 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8090 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8092 case Intrinsic::arm_neon_vshiftn:
8093 case Intrinsic::arm_neon_vrshiftn:
8094 case Intrinsic::arm_neon_vqshiftns:
8095 case Intrinsic::arm_neon_vqshiftnu:
8096 case Intrinsic::arm_neon_vqshiftnsu:
8097 case Intrinsic::arm_neon_vqrshiftns:
8098 case Intrinsic::arm_neon_vqrshiftnu:
8099 case Intrinsic::arm_neon_vqrshiftnsu:
8100 // Narrowing shifts require an immediate right shift.
8101 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8103 llvm_unreachable("invalid shift count for narrowing vector shift "
8107 llvm_unreachable("unhandled vector shift");
8111 case Intrinsic::arm_neon_vshifts:
8112 case Intrinsic::arm_neon_vshiftu:
8113 // Opcode already set above.
8115 case Intrinsic::arm_neon_vshiftls:
8116 case Intrinsic::arm_neon_vshiftlu:
8117 if (Cnt == VT.getVectorElementType().getSizeInBits())
8118 VShiftOpc = ARMISD::VSHLLi;
8120 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8121 ARMISD::VSHLLs : ARMISD::VSHLLu);
8123 case Intrinsic::arm_neon_vshiftn:
8124 VShiftOpc = ARMISD::VSHRN; break;
8125 case Intrinsic::arm_neon_vrshifts:
8126 VShiftOpc = ARMISD::VRSHRs; break;
8127 case Intrinsic::arm_neon_vrshiftu:
8128 VShiftOpc = ARMISD::VRSHRu; break;
8129 case Intrinsic::arm_neon_vrshiftn:
8130 VShiftOpc = ARMISD::VRSHRN; break;
8131 case Intrinsic::arm_neon_vqshifts:
8132 VShiftOpc = ARMISD::VQSHLs; break;
8133 case Intrinsic::arm_neon_vqshiftu:
8134 VShiftOpc = ARMISD::VQSHLu; break;
8135 case Intrinsic::arm_neon_vqshiftsu:
8136 VShiftOpc = ARMISD::VQSHLsu; break;
8137 case Intrinsic::arm_neon_vqshiftns:
8138 VShiftOpc = ARMISD::VQSHRNs; break;
8139 case Intrinsic::arm_neon_vqshiftnu:
8140 VShiftOpc = ARMISD::VQSHRNu; break;
8141 case Intrinsic::arm_neon_vqshiftnsu:
8142 VShiftOpc = ARMISD::VQSHRNsu; break;
8143 case Intrinsic::arm_neon_vqrshiftns:
8144 VShiftOpc = ARMISD::VQRSHRNs; break;
8145 case Intrinsic::arm_neon_vqrshiftnu:
8146 VShiftOpc = ARMISD::VQRSHRNu; break;
8147 case Intrinsic::arm_neon_vqrshiftnsu:
8148 VShiftOpc = ARMISD::VQRSHRNsu; break;
8151 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8152 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8155 case Intrinsic::arm_neon_vshiftins: {
8156 EVT VT = N->getOperand(1).getValueType();
8158 unsigned VShiftOpc = 0;
8160 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8161 VShiftOpc = ARMISD::VSLI;
8162 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8163 VShiftOpc = ARMISD::VSRI;
8165 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8168 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8169 N->getOperand(1), N->getOperand(2),
8170 DAG.getConstant(Cnt, MVT::i32));
8173 case Intrinsic::arm_neon_vqrshifts:
8174 case Intrinsic::arm_neon_vqrshiftu:
8175 // No immediate versions of these to check for.
8182 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8183 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8184 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8185 /// vector element shift counts are generally not legal, and it is hard to see
8186 /// their values after they get legalized to loads from a constant pool.
8187 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8188 const ARMSubtarget *ST) {
8189 EVT VT = N->getValueType(0);
8190 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8191 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8192 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8193 SDValue N1 = N->getOperand(1);
8194 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8195 SDValue N0 = N->getOperand(0);
8196 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8197 DAG.MaskedValueIsZero(N0.getOperand(0),
8198 APInt::getHighBitsSet(32, 16)))
8199 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8203 // Nothing to be done for scalar shifts.
8204 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8205 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8208 assert(ST->hasNEON() && "unexpected vector shift");
8211 switch (N->getOpcode()) {
8212 default: llvm_unreachable("unexpected shift opcode");
8215 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8216 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8217 DAG.getConstant(Cnt, MVT::i32));
8222 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8223 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8224 ARMISD::VSHRs : ARMISD::VSHRu);
8225 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8226 DAG.getConstant(Cnt, MVT::i32));
8232 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8233 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8234 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8235 const ARMSubtarget *ST) {
8236 SDValue N0 = N->getOperand(0);
8238 // Check for sign- and zero-extensions of vector extract operations of 8-
8239 // and 16-bit vector elements. NEON supports these directly. They are
8240 // handled during DAG combining because type legalization will promote them
8241 // to 32-bit types and it is messy to recognize the operations after that.
8242 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8243 SDValue Vec = N0.getOperand(0);
8244 SDValue Lane = N0.getOperand(1);
8245 EVT VT = N->getValueType(0);
8246 EVT EltVT = N0.getValueType();
8247 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8249 if (VT == MVT::i32 &&
8250 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8251 TLI.isTypeLegal(Vec.getValueType()) &&
8252 isa<ConstantSDNode>(Lane)) {
8255 switch (N->getOpcode()) {
8256 default: llvm_unreachable("unexpected opcode");
8257 case ISD::SIGN_EXTEND:
8258 Opc = ARMISD::VGETLANEs;
8260 case ISD::ZERO_EXTEND:
8261 case ISD::ANY_EXTEND:
8262 Opc = ARMISD::VGETLANEu;
8265 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8272 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8273 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8274 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8275 const ARMSubtarget *ST) {
8276 // If the target supports NEON, try to use vmax/vmin instructions for f32
8277 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8278 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8279 // a NaN; only do the transformation when it matches that behavior.
8281 // For now only do this when using NEON for FP operations; if using VFP, it
8282 // is not obvious that the benefit outweighs the cost of switching to the
8284 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8285 N->getValueType(0) != MVT::f32)
8288 SDValue CondLHS = N->getOperand(0);
8289 SDValue CondRHS = N->getOperand(1);
8290 SDValue LHS = N->getOperand(2);
8291 SDValue RHS = N->getOperand(3);
8292 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8294 unsigned Opcode = 0;
8296 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8297 IsReversed = false; // x CC y ? x : y
8298 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8299 IsReversed = true ; // x CC y ? y : x
8313 // If LHS is NaN, an ordered comparison will be false and the result will
8314 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8315 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8316 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8317 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8319 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8320 // will return -0, so vmin can only be used for unsafe math or if one of
8321 // the operands is known to be nonzero.
8322 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8323 !DAG.getTarget().Options.UnsafeFPMath &&
8324 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8326 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8335 // If LHS is NaN, an ordered comparison will be false and the result will
8336 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8337 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8338 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8339 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8341 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8342 // will return +0, so vmax can only be used for unsafe math or if one of
8343 // the operands is known to be nonzero.
8344 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8345 !DAG.getTarget().Options.UnsafeFPMath &&
8346 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8348 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8354 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8357 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8359 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8360 SDValue Cmp = N->getOperand(4);
8361 if (Cmp.getOpcode() != ARMISD::CMPZ)
8362 // Only looking at EQ and NE cases.
8365 EVT VT = N->getValueType(0);
8366 DebugLoc dl = N->getDebugLoc();
8367 SDValue LHS = Cmp.getOperand(0);
8368 SDValue RHS = Cmp.getOperand(1);
8369 SDValue FalseVal = N->getOperand(0);
8370 SDValue TrueVal = N->getOperand(1);
8371 SDValue ARMcc = N->getOperand(2);
8372 ARMCC::CondCodes CC =
8373 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8391 /// FIXME: Turn this into a target neutral optimization?
8393 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8394 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8395 N->getOperand(3), Cmp);
8396 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8398 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8399 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8400 N->getOperand(3), NewCmp);
8403 if (Res.getNode()) {
8404 APInt KnownZero, KnownOne;
8405 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8406 // Capture demanded bits information that would be otherwise lost.
8407 if (KnownZero == 0xfffffffe)
8408 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8409 DAG.getValueType(MVT::i1));
8410 else if (KnownZero == 0xffffff00)
8411 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8412 DAG.getValueType(MVT::i8));
8413 else if (KnownZero == 0xffff0000)
8414 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8415 DAG.getValueType(MVT::i16));
8421 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8422 DAGCombinerInfo &DCI) const {
8423 switch (N->getOpcode()) {
8425 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8426 case ISD::SUB: return PerformSUBCombine(N, DCI);
8427 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8428 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8429 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8430 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8431 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8432 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8433 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8434 case ISD::STORE: return PerformSTORECombine(N, DCI);
8435 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8436 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8437 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8438 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8439 case ISD::FP_TO_SINT:
8440 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8441 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8442 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8445 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8446 case ISD::SIGN_EXTEND:
8447 case ISD::ZERO_EXTEND:
8448 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8449 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8450 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8451 case ARMISD::VLD2DUP:
8452 case ARMISD::VLD3DUP:
8453 case ARMISD::VLD4DUP:
8454 return CombineBaseUpdate(N, DCI);
8455 case ISD::INTRINSIC_VOID:
8456 case ISD::INTRINSIC_W_CHAIN:
8457 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8458 case Intrinsic::arm_neon_vld1:
8459 case Intrinsic::arm_neon_vld2:
8460 case Intrinsic::arm_neon_vld3:
8461 case Intrinsic::arm_neon_vld4:
8462 case Intrinsic::arm_neon_vld2lane:
8463 case Intrinsic::arm_neon_vld3lane:
8464 case Intrinsic::arm_neon_vld4lane:
8465 case Intrinsic::arm_neon_vst1:
8466 case Intrinsic::arm_neon_vst2:
8467 case Intrinsic::arm_neon_vst3:
8468 case Intrinsic::arm_neon_vst4:
8469 case Intrinsic::arm_neon_vst2lane:
8470 case Intrinsic::arm_neon_vst3lane:
8471 case Intrinsic::arm_neon_vst4lane:
8472 return CombineBaseUpdate(N, DCI);
8480 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8482 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8485 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8486 if (!Subtarget->allowsUnalignedMem())
8489 switch (VT.getSimpleVT().SimpleTy) {
8496 // FIXME: VLD1 etc with standard alignment is legal.
8500 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8501 unsigned AlignCheck) {
8502 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8503 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8506 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8507 unsigned DstAlign, unsigned SrcAlign,
8510 MachineFunction &MF) const {
8511 const Function *F = MF.getFunction();
8513 // See if we can use NEON instructions for this...
8515 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8516 Subtarget->hasNEON()) {
8517 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8519 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8524 // Lowering to i32/i16 if the size permits.
8527 } else if (Size >= 2) {
8531 // Let the target-independent logic figure it out.
8535 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8540 switch (VT.getSimpleVT().SimpleTy) {
8541 default: return false;
8556 if ((V & (Scale - 1)) != 0)
8559 return V == (V & ((1LL << 5) - 1));
8562 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8563 const ARMSubtarget *Subtarget) {
8570 switch (VT.getSimpleVT().SimpleTy) {
8571 default: return false;
8576 // + imm12 or - imm8
8578 return V == (V & ((1LL << 8) - 1));
8579 return V == (V & ((1LL << 12) - 1));
8582 // Same as ARM mode. FIXME: NEON?
8583 if (!Subtarget->hasVFP2())
8588 return V == (V & ((1LL << 8) - 1));
8592 /// isLegalAddressImmediate - Return true if the integer value can be used
8593 /// as the offset of the target addressing mode for load / store of the
8595 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8596 const ARMSubtarget *Subtarget) {
8603 if (Subtarget->isThumb1Only())
8604 return isLegalT1AddressImmediate(V, VT);
8605 else if (Subtarget->isThumb2())
8606 return isLegalT2AddressImmediate(V, VT, Subtarget);
8611 switch (VT.getSimpleVT().SimpleTy) {
8612 default: return false;
8617 return V == (V & ((1LL << 12) - 1));
8620 return V == (V & ((1LL << 8) - 1));
8623 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8628 return V == (V & ((1LL << 8) - 1));
8632 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8634 int Scale = AM.Scale;
8638 switch (VT.getSimpleVT().SimpleTy) {
8639 default: return false;
8648 return Scale == 2 || Scale == 4 || Scale == 8;
8651 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8655 // Note, we allow "void" uses (basically, uses that aren't loads or
8656 // stores), because arm allows folding a scale into many arithmetic
8657 // operations. This should be made more precise and revisited later.
8659 // Allow r << imm, but the imm has to be a multiple of two.
8660 if (Scale & 1) return false;
8661 return isPowerOf2_32(Scale);
8665 /// isLegalAddressingMode - Return true if the addressing mode represented
8666 /// by AM is legal for this target, for a load/store of the specified type.
8667 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8669 EVT VT = getValueType(Ty, true);
8670 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8673 // Can never fold addr of global into load/store.
8678 case 0: // no scale reg, must be "r+i" or "r", or "i".
8681 if (Subtarget->isThumb1Only())
8685 // ARM doesn't support any R+R*scale+imm addr modes.
8692 if (Subtarget->isThumb2())
8693 return isLegalT2ScaledAddressingMode(AM, VT);
8695 int Scale = AM.Scale;
8696 switch (VT.getSimpleVT().SimpleTy) {
8697 default: return false;
8701 if (Scale < 0) Scale = -Scale;
8705 return isPowerOf2_32(Scale & ~1);
8709 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8714 // Note, we allow "void" uses (basically, uses that aren't loads or
8715 // stores), because arm allows folding a scale into many arithmetic
8716 // operations. This should be made more precise and revisited later.
8718 // Allow r << imm, but the imm has to be a multiple of two.
8719 if (Scale & 1) return false;
8720 return isPowerOf2_32(Scale);
8726 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8727 /// icmp immediate, that is the target has icmp instructions which can compare
8728 /// a register against the immediate without having to materialize the
8729 /// immediate into a register.
8730 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8731 // Thumb2 and ARM modes can use cmn for negative immediates.
8732 if (!Subtarget->isThumb())
8733 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
8734 if (Subtarget->isThumb2())
8735 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
8736 // Thumb1 doesn't have cmn, and only 8-bit immediates.
8737 return Imm >= 0 && Imm <= 255;
8740 /// isLegalAddImmediate - Return true if the specified immediate is legal
8741 /// add immediate, that is the target has add instructions which can add
8742 /// a register with the immediate without having to materialize the
8743 /// immediate into a register.
8744 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8745 return ARM_AM::getSOImmVal(Imm) != -1;
8748 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8749 bool isSEXTLoad, SDValue &Base,
8750 SDValue &Offset, bool &isInc,
8751 SelectionDAG &DAG) {
8752 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8755 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8757 Base = Ptr->getOperand(0);
8758 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8759 int RHSC = (int)RHS->getZExtValue();
8760 if (RHSC < 0 && RHSC > -256) {
8761 assert(Ptr->getOpcode() == ISD::ADD);
8763 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8767 isInc = (Ptr->getOpcode() == ISD::ADD);
8768 Offset = Ptr->getOperand(1);
8770 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8772 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8773 int RHSC = (int)RHS->getZExtValue();
8774 if (RHSC < 0 && RHSC > -0x1000) {
8775 assert(Ptr->getOpcode() == ISD::ADD);
8777 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8778 Base = Ptr->getOperand(0);
8783 if (Ptr->getOpcode() == ISD::ADD) {
8785 ARM_AM::ShiftOpc ShOpcVal=
8786 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8787 if (ShOpcVal != ARM_AM::no_shift) {
8788 Base = Ptr->getOperand(1);
8789 Offset = Ptr->getOperand(0);
8791 Base = Ptr->getOperand(0);
8792 Offset = Ptr->getOperand(1);
8797 isInc = (Ptr->getOpcode() == ISD::ADD);
8798 Base = Ptr->getOperand(0);
8799 Offset = Ptr->getOperand(1);
8803 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8807 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8808 bool isSEXTLoad, SDValue &Base,
8809 SDValue &Offset, bool &isInc,
8810 SelectionDAG &DAG) {
8811 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8814 Base = Ptr->getOperand(0);
8815 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8816 int RHSC = (int)RHS->getZExtValue();
8817 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8818 assert(Ptr->getOpcode() == ISD::ADD);
8820 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8822 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8823 isInc = Ptr->getOpcode() == ISD::ADD;
8824 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8832 /// getPreIndexedAddressParts - returns true by value, base pointer and
8833 /// offset pointer and addressing mode by reference if the node's address
8834 /// can be legally represented as pre-indexed load / store address.
8836 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8838 ISD::MemIndexedMode &AM,
8839 SelectionDAG &DAG) const {
8840 if (Subtarget->isThumb1Only())
8845 bool isSEXTLoad = false;
8846 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8847 Ptr = LD->getBasePtr();
8848 VT = LD->getMemoryVT();
8849 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8850 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8851 Ptr = ST->getBasePtr();
8852 VT = ST->getMemoryVT();
8857 bool isLegal = false;
8858 if (Subtarget->isThumb2())
8859 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8860 Offset, isInc, DAG);
8862 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8863 Offset, isInc, DAG);
8867 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8871 /// getPostIndexedAddressParts - returns true by value, base pointer and
8872 /// offset pointer and addressing mode by reference if this node can be
8873 /// combined with a load / store to form a post-indexed load / store.
8874 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8877 ISD::MemIndexedMode &AM,
8878 SelectionDAG &DAG) const {
8879 if (Subtarget->isThumb1Only())
8884 bool isSEXTLoad = false;
8885 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8886 VT = LD->getMemoryVT();
8887 Ptr = LD->getBasePtr();
8888 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8889 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8890 VT = ST->getMemoryVT();
8891 Ptr = ST->getBasePtr();
8896 bool isLegal = false;
8897 if (Subtarget->isThumb2())
8898 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8901 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8907 // Swap base ptr and offset to catch more post-index load / store when
8908 // it's legal. In Thumb2 mode, offset must be an immediate.
8909 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8910 !Subtarget->isThumb2())
8911 std::swap(Base, Offset);
8913 // Post-indexed load / store update the base pointer.
8918 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8922 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8925 const SelectionDAG &DAG,
8926 unsigned Depth) const {
8927 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
8928 switch (Op.getOpcode()) {
8930 case ARMISD::CMOV: {
8931 // Bits are known zero/one if known on the LHS and RHS.
8932 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
8933 if (KnownZero == 0 && KnownOne == 0) return;
8935 APInt KnownZeroRHS, KnownOneRHS;
8936 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
8937 KnownZero &= KnownZeroRHS;
8938 KnownOne &= KnownOneRHS;
8944 //===----------------------------------------------------------------------===//
8945 // ARM Inline Assembly Support
8946 //===----------------------------------------------------------------------===//
8948 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8949 // Looking for "rev" which is V6+.
8950 if (!Subtarget->hasV6Ops())
8953 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8954 std::string AsmStr = IA->getAsmString();
8955 SmallVector<StringRef, 4> AsmPieces;
8956 SplitString(AsmStr, AsmPieces, ";\n");
8958 switch (AsmPieces.size()) {
8959 default: return false;
8961 AsmStr = AsmPieces[0];
8963 SplitString(AsmStr, AsmPieces, " \t,");
8966 if (AsmPieces.size() == 3 &&
8967 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8968 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8969 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8970 if (Ty && Ty->getBitWidth() == 32)
8971 return IntrinsicLowering::LowerToByteSwap(CI);
8979 /// getConstraintType - Given a constraint letter, return the type of
8980 /// constraint it is for this target.
8981 ARMTargetLowering::ConstraintType
8982 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8983 if (Constraint.size() == 1) {
8984 switch (Constraint[0]) {
8986 case 'l': return C_RegisterClass;
8987 case 'w': return C_RegisterClass;
8988 case 'h': return C_RegisterClass;
8989 case 'x': return C_RegisterClass;
8990 case 't': return C_RegisterClass;
8991 case 'j': return C_Other; // Constant for movw.
8992 // An address with a single base register. Due to the way we
8993 // currently handle addresses it is the same as an 'r' memory constraint.
8994 case 'Q': return C_Memory;
8996 } else if (Constraint.size() == 2) {
8997 switch (Constraint[0]) {
8999 // All 'U+' constraints are addresses.
9000 case 'U': return C_Memory;
9003 return TargetLowering::getConstraintType(Constraint);
9006 /// Examine constraint type and operand type and determine a weight value.
9007 /// This object must already have been set up with the operand type
9008 /// and the current alternative constraint selected.
9009 TargetLowering::ConstraintWeight
9010 ARMTargetLowering::getSingleConstraintMatchWeight(
9011 AsmOperandInfo &info, const char *constraint) const {
9012 ConstraintWeight weight = CW_Invalid;
9013 Value *CallOperandVal = info.CallOperandVal;
9014 // If we don't have a value, we can't do a match,
9015 // but allow it at the lowest weight.
9016 if (CallOperandVal == NULL)
9018 Type *type = CallOperandVal->getType();
9019 // Look at the constraint type.
9020 switch (*constraint) {
9022 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9025 if (type->isIntegerTy()) {
9026 if (Subtarget->isThumb())
9027 weight = CW_SpecificReg;
9029 weight = CW_Register;
9033 if (type->isFloatingPointTy())
9034 weight = CW_Register;
9040 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9042 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9044 if (Constraint.size() == 1) {
9045 // GCC ARM Constraint Letters
9046 switch (Constraint[0]) {
9047 case 'l': // Low regs or general regs.
9048 if (Subtarget->isThumb())
9049 return RCPair(0U, &ARM::tGPRRegClass);
9050 return RCPair(0U, &ARM::GPRRegClass);
9051 case 'h': // High regs or no regs.
9052 if (Subtarget->isThumb())
9053 return RCPair(0U, &ARM::hGPRRegClass);
9056 return RCPair(0U, &ARM::GPRRegClass);
9059 return RCPair(0U, &ARM::SPRRegClass);
9060 if (VT.getSizeInBits() == 64)
9061 return RCPair(0U, &ARM::DPRRegClass);
9062 if (VT.getSizeInBits() == 128)
9063 return RCPair(0U, &ARM::QPRRegClass);
9067 return RCPair(0U, &ARM::SPR_8RegClass);
9068 if (VT.getSizeInBits() == 64)
9069 return RCPair(0U, &ARM::DPR_8RegClass);
9070 if (VT.getSizeInBits() == 128)
9071 return RCPair(0U, &ARM::QPR_8RegClass);
9075 return RCPair(0U, &ARM::SPRRegClass);
9079 if (StringRef("{cc}").equals_lower(Constraint))
9080 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9082 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9085 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9086 /// vector. If it is invalid, don't add anything to Ops.
9087 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9088 std::string &Constraint,
9089 std::vector<SDValue>&Ops,
9090 SelectionDAG &DAG) const {
9091 SDValue Result(0, 0);
9093 // Currently only support length 1 constraints.
9094 if (Constraint.length() != 1) return;
9096 char ConstraintLetter = Constraint[0];
9097 switch (ConstraintLetter) {
9100 case 'I': case 'J': case 'K': case 'L':
9101 case 'M': case 'N': case 'O':
9102 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9106 int64_t CVal64 = C->getSExtValue();
9107 int CVal = (int) CVal64;
9108 // None of these constraints allow values larger than 32 bits. Check
9109 // that the value fits in an int.
9113 switch (ConstraintLetter) {
9115 // Constant suitable for movw, must be between 0 and
9117 if (Subtarget->hasV6T2Ops())
9118 if (CVal >= 0 && CVal <= 65535)
9122 if (Subtarget->isThumb1Only()) {
9123 // This must be a constant between 0 and 255, for ADD
9125 if (CVal >= 0 && CVal <= 255)
9127 } else if (Subtarget->isThumb2()) {
9128 // A constant that can be used as an immediate value in a
9129 // data-processing instruction.
9130 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9133 // A constant that can be used as an immediate value in a
9134 // data-processing instruction.
9135 if (ARM_AM::getSOImmVal(CVal) != -1)
9141 if (Subtarget->isThumb()) { // FIXME thumb2
9142 // This must be a constant between -255 and -1, for negated ADD
9143 // immediates. This can be used in GCC with an "n" modifier that
9144 // prints the negated value, for use with SUB instructions. It is
9145 // not useful otherwise but is implemented for compatibility.
9146 if (CVal >= -255 && CVal <= -1)
9149 // This must be a constant between -4095 and 4095. It is not clear
9150 // what this constraint is intended for. Implemented for
9151 // compatibility with GCC.
9152 if (CVal >= -4095 && CVal <= 4095)
9158 if (Subtarget->isThumb1Only()) {
9159 // A 32-bit value where only one byte has a nonzero value. Exclude
9160 // zero to match GCC. This constraint is used by GCC internally for
9161 // constants that can be loaded with a move/shift combination.
9162 // It is not useful otherwise but is implemented for compatibility.
9163 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9165 } else if (Subtarget->isThumb2()) {
9166 // A constant whose bitwise inverse can be used as an immediate
9167 // value in a data-processing instruction. This can be used in GCC
9168 // with a "B" modifier that prints the inverted value, for use with
9169 // BIC and MVN instructions. It is not useful otherwise but is
9170 // implemented for compatibility.
9171 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9174 // A constant whose bitwise inverse can be used as an immediate
9175 // value in a data-processing instruction. This can be used in GCC
9176 // with a "B" modifier that prints the inverted value, for use with
9177 // BIC and MVN instructions. It is not useful otherwise but is
9178 // implemented for compatibility.
9179 if (ARM_AM::getSOImmVal(~CVal) != -1)
9185 if (Subtarget->isThumb1Only()) {
9186 // This must be a constant between -7 and 7,
9187 // for 3-operand ADD/SUB immediate instructions.
9188 if (CVal >= -7 && CVal < 7)
9190 } else if (Subtarget->isThumb2()) {
9191 // A constant whose negation can be used as an immediate value in a
9192 // data-processing instruction. This can be used in GCC with an "n"
9193 // modifier that prints the negated value, for use with SUB
9194 // instructions. It is not useful otherwise but is implemented for
9196 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9199 // A constant whose negation can be used as an immediate value in a
9200 // data-processing instruction. This can be used in GCC with an "n"
9201 // modifier that prints the negated value, for use with SUB
9202 // instructions. It is not useful otherwise but is implemented for
9204 if (ARM_AM::getSOImmVal(-CVal) != -1)
9210 if (Subtarget->isThumb()) { // FIXME thumb2
9211 // This must be a multiple of 4 between 0 and 1020, for
9212 // ADD sp + immediate.
9213 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9216 // A power of two or a constant between 0 and 32. This is used in
9217 // GCC for the shift amount on shifted register operands, but it is
9218 // useful in general for any shift amounts.
9219 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9225 if (Subtarget->isThumb()) { // FIXME thumb2
9226 // This must be a constant between 0 and 31, for shift amounts.
9227 if (CVal >= 0 && CVal <= 31)
9233 if (Subtarget->isThumb()) { // FIXME thumb2
9234 // This must be a multiple of 4 between -508 and 508, for
9235 // ADD/SUB sp = sp + immediate.
9236 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9241 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9245 if (Result.getNode()) {
9246 Ops.push_back(Result);
9249 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9253 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9254 // The ARM target isn't yet aware of offsets.
9258 bool ARM::isBitFieldInvertedMask(unsigned v) {
9259 if (v == 0xffffffff)
9261 // there can be 1's on either or both "outsides", all the "inside"
9263 unsigned int lsb = 0, msb = 31;
9264 while (v & (1 << msb)) --msb;
9265 while (v & (1 << lsb)) ++lsb;
9266 for (unsigned int i = lsb; i <= msb; ++i) {
9273 /// isFPImmLegal - Returns true if the target can instruction select the
9274 /// specified FP immediate natively. If false, the legalizer will
9275 /// materialize the FP immediate as a load from a constant pool.
9276 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9277 if (!Subtarget->hasVFP3())
9280 return ARM_AM::getFP32Imm(Imm) != -1;
9282 return ARM_AM::getFP64Imm(Imm) != -1;
9286 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9287 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9288 /// specified in the intrinsic calls.
9289 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9291 unsigned Intrinsic) const {
9292 switch (Intrinsic) {
9293 case Intrinsic::arm_neon_vld1:
9294 case Intrinsic::arm_neon_vld2:
9295 case Intrinsic::arm_neon_vld3:
9296 case Intrinsic::arm_neon_vld4:
9297 case Intrinsic::arm_neon_vld2lane:
9298 case Intrinsic::arm_neon_vld3lane:
9299 case Intrinsic::arm_neon_vld4lane: {
9300 Info.opc = ISD::INTRINSIC_W_CHAIN;
9301 // Conservatively set memVT to the entire set of vectors loaded.
9302 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9303 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9304 Info.ptrVal = I.getArgOperand(0);
9306 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9307 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9308 Info.vol = false; // volatile loads with NEON intrinsics not supported
9309 Info.readMem = true;
9310 Info.writeMem = false;
9313 case Intrinsic::arm_neon_vst1:
9314 case Intrinsic::arm_neon_vst2:
9315 case Intrinsic::arm_neon_vst3:
9316 case Intrinsic::arm_neon_vst4:
9317 case Intrinsic::arm_neon_vst2lane:
9318 case Intrinsic::arm_neon_vst3lane:
9319 case Intrinsic::arm_neon_vst4lane: {
9320 Info.opc = ISD::INTRINSIC_VOID;
9321 // Conservatively set memVT to the entire set of vectors stored.
9322 unsigned NumElts = 0;
9323 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9324 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9325 if (!ArgTy->isVectorTy())
9327 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9329 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9330 Info.ptrVal = I.getArgOperand(0);
9332 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9333 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9334 Info.vol = false; // volatile stores with NEON intrinsics not supported
9335 Info.readMem = false;
9336 Info.writeMem = true;
9339 case Intrinsic::arm_strexd: {
9340 Info.opc = ISD::INTRINSIC_W_CHAIN;
9341 Info.memVT = MVT::i64;
9342 Info.ptrVal = I.getArgOperand(2);
9346 Info.readMem = false;
9347 Info.writeMem = true;
9350 case Intrinsic::arm_ldrexd: {
9351 Info.opc = ISD::INTRINSIC_W_CHAIN;
9352 Info.memVT = MVT::i64;
9353 Info.ptrVal = I.getArgOperand(0);
9357 Info.readMem = true;
9358 Info.writeMem = false;