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::DPRRegisterClass);
157 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
160 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
161 addRegisterClass(VT, ARM::QPRRegisterClass);
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::tGPRRegisterClass);
436 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
437 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
438 !Subtarget->isThumb1Only()) {
439 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
440 if (!Subtarget->isFPOnlySP())
441 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
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 setOperationAction(ISD::FMA, MVT::f64, Expand);
773 setOperationAction(ISD::FMA, MVT::f32, Expand);
775 // Various VFP goodness
776 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
777 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
778 if (Subtarget->hasVFP2()) {
779 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
780 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
781 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
782 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
784 // Special handling for half-precision FP.
785 if (!Subtarget->hasFP16()) {
786 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
787 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
791 // We have target-specific dag combine patterns for the following nodes:
792 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
793 setTargetDAGCombine(ISD::ADD);
794 setTargetDAGCombine(ISD::SUB);
795 setTargetDAGCombine(ISD::MUL);
797 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
798 setTargetDAGCombine(ISD::AND);
799 setTargetDAGCombine(ISD::OR);
800 setTargetDAGCombine(ISD::XOR);
803 if (Subtarget->hasV6Ops())
804 setTargetDAGCombine(ISD::SRL);
806 setStackPointerRegisterToSaveRestore(ARM::SP);
808 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
809 !Subtarget->hasVFP2())
810 setSchedulingPreference(Sched::RegPressure);
812 setSchedulingPreference(Sched::Hybrid);
814 //// temporary - rewrite interface to use type
815 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
816 maxStoresPerMemset = 16;
817 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
819 // On ARM arguments smaller than 4 bytes are extended, so all arguments
820 // are at least 4 bytes aligned.
821 setMinStackArgumentAlignment(4);
823 benefitFromCodePlacementOpt = true;
825 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
828 // FIXME: It might make sense to define the representative register class as the
829 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
830 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
831 // SPR's representative would be DPR_VFP2. This should work well if register
832 // pressure tracking were modified such that a register use would increment the
833 // pressure of the register class's representative and all of it's super
834 // classes' representatives transitively. We have not implemented this because
835 // of the difficulty prior to coalescing of modeling operand register classes
836 // due to the common occurrence of cross class copies and subregister insertions
838 std::pair<const TargetRegisterClass*, uint8_t>
839 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
840 const TargetRegisterClass *RRC = 0;
842 switch (VT.getSimpleVT().SimpleTy) {
844 return TargetLowering::findRepresentativeClass(VT);
845 // Use DPR as representative register class for all floating point
846 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
847 // the cost is 1 for both f32 and f64.
848 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
849 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
850 RRC = ARM::DPRRegisterClass;
851 // When NEON is used for SP, only half of the register file is available
852 // because operations that define both SP and DP results will be constrained
853 // to the VFP2 class (D0-D15). We currently model this constraint prior to
854 // coalescing by double-counting the SP regs. See the FIXME above.
855 if (Subtarget->useNEONForSinglePrecisionFP())
858 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
859 case MVT::v4f32: case MVT::v2f64:
860 RRC = ARM::DPRRegisterClass;
864 RRC = ARM::DPRRegisterClass;
868 RRC = ARM::DPRRegisterClass;
872 return std::make_pair(RRC, Cost);
875 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
878 case ARMISD::Wrapper: return "ARMISD::Wrapper";
879 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
880 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
881 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
882 case ARMISD::CALL: return "ARMISD::CALL";
883 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
884 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
885 case ARMISD::tCALL: return "ARMISD::tCALL";
886 case ARMISD::BRCOND: return "ARMISD::BRCOND";
887 case ARMISD::BR_JT: return "ARMISD::BR_JT";
888 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
889 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
890 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
891 case ARMISD::CMP: return "ARMISD::CMP";
892 case ARMISD::CMPZ: return "ARMISD::CMPZ";
893 case ARMISD::CMPFP: return "ARMISD::CMPFP";
894 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
895 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
896 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
898 case ARMISD::CMOV: return "ARMISD::CMOV";
899 case ARMISD::CAND: return "ARMISD::CAND";
900 case ARMISD::COR: return "ARMISD::COR";
901 case ARMISD::CXOR: return "ARMISD::CXOR";
903 case ARMISD::RBIT: return "ARMISD::RBIT";
905 case ARMISD::FTOSI: return "ARMISD::FTOSI";
906 case ARMISD::FTOUI: return "ARMISD::FTOUI";
907 case ARMISD::SITOF: return "ARMISD::SITOF";
908 case ARMISD::UITOF: return "ARMISD::UITOF";
910 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
911 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
912 case ARMISD::RRX: return "ARMISD::RRX";
914 case ARMISD::ADDC: return "ARMISD::ADDC";
915 case ARMISD::ADDE: return "ARMISD::ADDE";
916 case ARMISD::SUBC: return "ARMISD::SUBC";
917 case ARMISD::SUBE: return "ARMISD::SUBE";
919 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
920 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
922 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
923 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
925 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
927 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
929 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
931 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
932 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
934 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
936 case ARMISD::VCEQ: return "ARMISD::VCEQ";
937 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
938 case ARMISD::VCGE: return "ARMISD::VCGE";
939 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
940 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
941 case ARMISD::VCGEU: return "ARMISD::VCGEU";
942 case ARMISD::VCGT: return "ARMISD::VCGT";
943 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
944 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
945 case ARMISD::VCGTU: return "ARMISD::VCGTU";
946 case ARMISD::VTST: return "ARMISD::VTST";
948 case ARMISD::VSHL: return "ARMISD::VSHL";
949 case ARMISD::VSHRs: return "ARMISD::VSHRs";
950 case ARMISD::VSHRu: return "ARMISD::VSHRu";
951 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
952 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
953 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
954 case ARMISD::VSHRN: return "ARMISD::VSHRN";
955 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
956 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
957 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
958 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
959 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
960 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
961 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
962 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
963 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
964 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
965 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
966 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
967 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
968 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
969 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
970 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
971 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
972 case ARMISD::VDUP: return "ARMISD::VDUP";
973 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
974 case ARMISD::VEXT: return "ARMISD::VEXT";
975 case ARMISD::VREV64: return "ARMISD::VREV64";
976 case ARMISD::VREV32: return "ARMISD::VREV32";
977 case ARMISD::VREV16: return "ARMISD::VREV16";
978 case ARMISD::VZIP: return "ARMISD::VZIP";
979 case ARMISD::VUZP: return "ARMISD::VUZP";
980 case ARMISD::VTRN: return "ARMISD::VTRN";
981 case ARMISD::VTBL1: return "ARMISD::VTBL1";
982 case ARMISD::VTBL2: return "ARMISD::VTBL2";
983 case ARMISD::VMULLs: return "ARMISD::VMULLs";
984 case ARMISD::VMULLu: return "ARMISD::VMULLu";
985 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
986 case ARMISD::FMAX: return "ARMISD::FMAX";
987 case ARMISD::FMIN: return "ARMISD::FMIN";
988 case ARMISD::BFI: return "ARMISD::BFI";
989 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
990 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
991 case ARMISD::VBSL: return "ARMISD::VBSL";
992 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
993 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
994 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
995 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
996 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
997 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
998 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
999 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1000 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1001 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1002 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1003 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1004 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1005 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1006 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1007 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1008 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1009 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1010 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1011 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1015 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1016 if (!VT.isVector()) return getPointerTy();
1017 return VT.changeVectorElementTypeToInteger();
1020 /// getRegClassFor - Return the register class that should be used for the
1021 /// specified value type.
1022 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1023 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1024 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1025 // load / store 4 to 8 consecutive D registers.
1026 if (Subtarget->hasNEON()) {
1027 if (VT == MVT::v4i64)
1028 return ARM::QQPRRegisterClass;
1029 else if (VT == MVT::v8i64)
1030 return ARM::QQQQPRRegisterClass;
1032 return TargetLowering::getRegClassFor(VT);
1035 // Create a fast isel object.
1037 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1038 return ARM::createFastISel(funcInfo);
1041 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1042 /// be used for loads / stores from the global.
1043 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1044 return (Subtarget->isThumb1Only() ? 127 : 4095);
1047 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1048 unsigned NumVals = N->getNumValues();
1050 return Sched::RegPressure;
1052 for (unsigned i = 0; i != NumVals; ++i) {
1053 EVT VT = N->getValueType(i);
1054 if (VT == MVT::Glue || VT == MVT::Other)
1056 if (VT.isFloatingPoint() || VT.isVector())
1060 if (!N->isMachineOpcode())
1061 return Sched::RegPressure;
1063 // Load are scheduled for latency even if there instruction itinerary
1064 // is not available.
1065 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1066 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1068 if (MCID.getNumDefs() == 0)
1069 return Sched::RegPressure;
1070 if (!Itins->isEmpty() &&
1071 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1074 return Sched::RegPressure;
1077 //===----------------------------------------------------------------------===//
1079 //===----------------------------------------------------------------------===//
1081 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1082 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1084 default: llvm_unreachable("Unknown condition code!");
1085 case ISD::SETNE: return ARMCC::NE;
1086 case ISD::SETEQ: return ARMCC::EQ;
1087 case ISD::SETGT: return ARMCC::GT;
1088 case ISD::SETGE: return ARMCC::GE;
1089 case ISD::SETLT: return ARMCC::LT;
1090 case ISD::SETLE: return ARMCC::LE;
1091 case ISD::SETUGT: return ARMCC::HI;
1092 case ISD::SETUGE: return ARMCC::HS;
1093 case ISD::SETULT: return ARMCC::LO;
1094 case ISD::SETULE: return ARMCC::LS;
1098 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1099 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1100 ARMCC::CondCodes &CondCode2) {
1101 CondCode2 = ARMCC::AL;
1103 default: llvm_unreachable("Unknown FP condition!");
1105 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1107 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1109 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1110 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1111 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1112 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1113 case ISD::SETO: CondCode = ARMCC::VC; break;
1114 case ISD::SETUO: CondCode = ARMCC::VS; break;
1115 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1116 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1117 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1119 case ISD::SETULT: CondCode = ARMCC::LT; break;
1121 case ISD::SETULE: CondCode = ARMCC::LE; break;
1123 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1127 //===----------------------------------------------------------------------===//
1128 // Calling Convention Implementation
1129 //===----------------------------------------------------------------------===//
1131 #include "ARMGenCallingConv.inc"
1133 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1134 /// given CallingConvention value.
1135 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1137 bool isVarArg) const {
1140 llvm_unreachable("Unsupported calling convention");
1141 case CallingConv::Fast:
1142 if (Subtarget->hasVFP2() && !isVarArg) {
1143 if (!Subtarget->isAAPCS_ABI())
1144 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1145 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1146 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1149 case CallingConv::C: {
1150 // Use target triple & subtarget features to do actual dispatch.
1151 if (!Subtarget->isAAPCS_ABI())
1152 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1153 else if (Subtarget->hasVFP2() &&
1154 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1156 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1157 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1159 case CallingConv::ARM_AAPCS_VFP:
1161 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 case CallingConv::ARM_AAPCS:
1164 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_APCS:
1166 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1170 /// LowerCallResult - Lower the result values of a call into the
1171 /// appropriate copies out of appropriate physical registers.
1173 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1174 CallingConv::ID CallConv, bool isVarArg,
1175 const SmallVectorImpl<ISD::InputArg> &Ins,
1176 DebugLoc dl, SelectionDAG &DAG,
1177 SmallVectorImpl<SDValue> &InVals) const {
1179 // Assign locations to each value returned by this call.
1180 SmallVector<CCValAssign, 16> RVLocs;
1181 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1182 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1183 CCInfo.AnalyzeCallResult(Ins,
1184 CCAssignFnForNode(CallConv, /* Return*/ true,
1187 // Copy all of the result registers out of their specified physreg.
1188 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1189 CCValAssign VA = RVLocs[i];
1192 if (VA.needsCustom()) {
1193 // Handle f64 or half of a v2f64.
1194 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1196 Chain = Lo.getValue(1);
1197 InFlag = Lo.getValue(2);
1198 VA = RVLocs[++i]; // skip ahead to next loc
1199 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1201 Chain = Hi.getValue(1);
1202 InFlag = Hi.getValue(2);
1203 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1205 if (VA.getLocVT() == MVT::v2f64) {
1206 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1207 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1208 DAG.getConstant(0, MVT::i32));
1210 VA = RVLocs[++i]; // skip ahead to next loc
1211 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1212 Chain = Lo.getValue(1);
1213 InFlag = Lo.getValue(2);
1214 VA = RVLocs[++i]; // skip ahead to next loc
1215 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1216 Chain = Hi.getValue(1);
1217 InFlag = Hi.getValue(2);
1218 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1219 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1220 DAG.getConstant(1, MVT::i32));
1223 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1225 Chain = Val.getValue(1);
1226 InFlag = Val.getValue(2);
1229 switch (VA.getLocInfo()) {
1230 default: llvm_unreachable("Unknown loc info!");
1231 case CCValAssign::Full: break;
1232 case CCValAssign::BCvt:
1233 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1237 InVals.push_back(Val);
1243 /// LowerMemOpCallTo - Store the argument to the stack.
1245 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1246 SDValue StackPtr, SDValue Arg,
1247 DebugLoc dl, SelectionDAG &DAG,
1248 const CCValAssign &VA,
1249 ISD::ArgFlagsTy Flags) const {
1250 unsigned LocMemOffset = VA.getLocMemOffset();
1251 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1252 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1253 return DAG.getStore(Chain, dl, Arg, PtrOff,
1254 MachinePointerInfo::getStack(LocMemOffset),
1258 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1259 SDValue Chain, SDValue &Arg,
1260 RegsToPassVector &RegsToPass,
1261 CCValAssign &VA, CCValAssign &NextVA,
1263 SmallVector<SDValue, 8> &MemOpChains,
1264 ISD::ArgFlagsTy Flags) const {
1266 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1267 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1268 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1270 if (NextVA.isRegLoc())
1271 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1273 assert(NextVA.isMemLoc());
1274 if (StackPtr.getNode() == 0)
1275 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1277 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1283 /// LowerCall - Lowering a call into a callseq_start <-
1284 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1287 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1288 CallingConv::ID CallConv, bool isVarArg,
1289 bool doesNotRet, bool &isTailCall,
1290 const SmallVectorImpl<ISD::OutputArg> &Outs,
1291 const SmallVectorImpl<SDValue> &OutVals,
1292 const SmallVectorImpl<ISD::InputArg> &Ins,
1293 DebugLoc dl, SelectionDAG &DAG,
1294 SmallVectorImpl<SDValue> &InVals) const {
1295 MachineFunction &MF = DAG.getMachineFunction();
1296 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1297 bool IsSibCall = false;
1298 // Disable tail calls if they're not supported.
1299 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1302 // Check if it's really possible to do a tail call.
1303 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1304 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1305 Outs, OutVals, Ins, DAG);
1306 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1307 // detected sibcalls.
1314 // Analyze operands of the call, assigning locations to each operand.
1315 SmallVector<CCValAssign, 16> ArgLocs;
1316 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1317 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1318 CCInfo.AnalyzeCallOperands(Outs,
1319 CCAssignFnForNode(CallConv, /* Return*/ false,
1322 // Get a count of how many bytes are to be pushed on the stack.
1323 unsigned NumBytes = CCInfo.getNextStackOffset();
1325 // For tail calls, memory operands are available in our caller's stack.
1329 // Adjust the stack pointer for the new arguments...
1330 // These operations are automatically eliminated by the prolog/epilog pass
1332 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1334 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1336 RegsToPassVector RegsToPass;
1337 SmallVector<SDValue, 8> MemOpChains;
1339 // Walk the register/memloc assignments, inserting copies/loads. In the case
1340 // of tail call optimization, arguments are handled later.
1341 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1343 ++i, ++realArgIdx) {
1344 CCValAssign &VA = ArgLocs[i];
1345 SDValue Arg = OutVals[realArgIdx];
1346 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1347 bool isByVal = Flags.isByVal();
1349 // Promote the value if needed.
1350 switch (VA.getLocInfo()) {
1351 default: llvm_unreachable("Unknown loc info!");
1352 case CCValAssign::Full: break;
1353 case CCValAssign::SExt:
1354 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1356 case CCValAssign::ZExt:
1357 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1359 case CCValAssign::AExt:
1360 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1362 case CCValAssign::BCvt:
1363 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1367 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1368 if (VA.needsCustom()) {
1369 if (VA.getLocVT() == MVT::v2f64) {
1370 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1371 DAG.getConstant(0, MVT::i32));
1372 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1373 DAG.getConstant(1, MVT::i32));
1375 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1376 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1378 VA = ArgLocs[++i]; // skip ahead to next loc
1379 if (VA.isRegLoc()) {
1380 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1381 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1383 assert(VA.isMemLoc());
1385 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1386 dl, DAG, VA, Flags));
1389 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1390 StackPtr, MemOpChains, Flags);
1392 } else if (VA.isRegLoc()) {
1393 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1394 } else if (isByVal) {
1395 assert(VA.isMemLoc());
1396 unsigned offset = 0;
1398 // True if this byval aggregate will be split between registers
1400 if (CCInfo.isFirstByValRegValid()) {
1401 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1403 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1404 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1405 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1406 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1407 MachinePointerInfo(),
1408 false, false, false, 0);
1409 MemOpChains.push_back(Load.getValue(1));
1410 RegsToPass.push_back(std::make_pair(j, Load));
1412 offset = ARM::R4 - CCInfo.getFirstByValReg();
1413 CCInfo.clearFirstByValReg();
1416 unsigned LocMemOffset = VA.getLocMemOffset();
1417 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1420 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1421 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1422 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1424 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1425 Flags.getByValAlign(),
1426 /*isVolatile=*/false,
1427 /*AlwaysInline=*/false,
1428 MachinePointerInfo(0),
1429 MachinePointerInfo(0)));
1431 } else if (!IsSibCall) {
1432 assert(VA.isMemLoc());
1434 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1435 dl, DAG, VA, Flags));
1439 if (!MemOpChains.empty())
1440 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1441 &MemOpChains[0], MemOpChains.size());
1443 // Build a sequence of copy-to-reg nodes chained together with token chain
1444 // and flag operands which copy the outgoing args into the appropriate regs.
1446 // Tail call byval lowering might overwrite argument registers so in case of
1447 // tail call optimization the copies to registers are lowered later.
1449 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1450 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1451 RegsToPass[i].second, InFlag);
1452 InFlag = Chain.getValue(1);
1455 // For tail calls lower the arguments to the 'real' stack slot.
1457 // Force all the incoming stack arguments to be loaded from the stack
1458 // before any new outgoing arguments are stored to the stack, because the
1459 // outgoing stack slots may alias the incoming argument stack slots, and
1460 // the alias isn't otherwise explicit. This is slightly more conservative
1461 // than necessary, because it means that each store effectively depends
1462 // on every argument instead of just those arguments it would clobber.
1464 // Do not flag preceding copytoreg stuff together with the following stuff.
1466 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1467 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1468 RegsToPass[i].second, InFlag);
1469 InFlag = Chain.getValue(1);
1474 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1475 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1476 // node so that legalize doesn't hack it.
1477 bool isDirect = false;
1478 bool isARMFunc = false;
1479 bool isLocalARMFunc = false;
1480 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1482 if (EnableARMLongCalls) {
1483 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1484 && "long-calls with non-static relocation model!");
1485 // Handle a global address or an external symbol. If it's not one of
1486 // those, the target's already in a register, so we don't need to do
1488 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1489 const GlobalValue *GV = G->getGlobal();
1490 // Create a constant pool entry for the callee address
1491 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1492 ARMConstantPoolValue *CPV =
1493 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1495 // Get the address of the callee into a register
1496 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1497 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1498 Callee = DAG.getLoad(getPointerTy(), dl,
1499 DAG.getEntryNode(), CPAddr,
1500 MachinePointerInfo::getConstantPool(),
1501 false, false, false, 0);
1502 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1503 const char *Sym = S->getSymbol();
1505 // Create a constant pool entry for the callee address
1506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507 ARMConstantPoolValue *CPV =
1508 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1509 ARMPCLabelIndex, 0);
1510 // Get the address of the callee into a register
1511 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513 Callee = DAG.getLoad(getPointerTy(), dl,
1514 DAG.getEntryNode(), CPAddr,
1515 MachinePointerInfo::getConstantPool(),
1516 false, false, false, 0);
1518 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1519 const GlobalValue *GV = G->getGlobal();
1521 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1522 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1523 getTargetMachine().getRelocationModel() != Reloc::Static;
1524 isARMFunc = !Subtarget->isThumb() || isStub;
1525 // ARM call to a local ARM function is predicable.
1526 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1527 // tBX takes a register source operand.
1528 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1529 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1530 ARMConstantPoolValue *CPV =
1531 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1532 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1533 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1534 Callee = DAG.getLoad(getPointerTy(), dl,
1535 DAG.getEntryNode(), CPAddr,
1536 MachinePointerInfo::getConstantPool(),
1537 false, false, false, 0);
1538 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1539 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1540 getPointerTy(), Callee, PICLabel);
1542 // On ELF targets for PIC code, direct calls should go through the PLT
1543 unsigned OpFlags = 0;
1544 if (Subtarget->isTargetELF() &&
1545 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1546 OpFlags = ARMII::MO_PLT;
1547 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1549 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1551 bool isStub = Subtarget->isTargetDarwin() &&
1552 getTargetMachine().getRelocationModel() != Reloc::Static;
1553 isARMFunc = !Subtarget->isThumb() || isStub;
1554 // tBX takes a register source operand.
1555 const char *Sym = S->getSymbol();
1556 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1557 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1558 ARMConstantPoolValue *CPV =
1559 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1560 ARMPCLabelIndex, 4);
1561 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1562 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1563 Callee = DAG.getLoad(getPointerTy(), dl,
1564 DAG.getEntryNode(), CPAddr,
1565 MachinePointerInfo::getConstantPool(),
1566 false, false, false, 0);
1567 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1568 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1569 getPointerTy(), Callee, PICLabel);
1571 unsigned OpFlags = 0;
1572 // On ELF targets for PIC code, direct calls should go through the PLT
1573 if (Subtarget->isTargetELF() &&
1574 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1575 OpFlags = ARMII::MO_PLT;
1576 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1580 // FIXME: handle tail calls differently.
1582 if (Subtarget->isThumb()) {
1583 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1584 CallOpc = ARMISD::CALL_NOLINK;
1585 else if (doesNotRet && isDirect && !isARMFunc &&
1586 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1587 // "mov lr, pc; b _foo" to avoid confusing the RSP
1588 CallOpc = ARMISD::CALL_NOLINK;
1590 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1592 if (!isDirect && !Subtarget->hasV5TOps()) {
1593 CallOpc = ARMISD::CALL_NOLINK;
1594 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1595 // "mov lr, pc; b _foo" to avoid confusing the RSP
1596 CallOpc = ARMISD::CALL_NOLINK;
1598 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1601 std::vector<SDValue> Ops;
1602 Ops.push_back(Chain);
1603 Ops.push_back(Callee);
1605 // Add argument registers to the end of the list so that they are known live
1607 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1608 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1609 RegsToPass[i].second.getValueType()));
1611 // Add a register mask operand representing the call-preserved registers.
1612 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1613 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1614 assert(Mask && "Missing call preserved mask for calling convention");
1615 Ops.push_back(DAG.getRegisterMask(Mask));
1617 if (InFlag.getNode())
1618 Ops.push_back(InFlag);
1620 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1622 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1624 // Returns a chain and a flag for retval copy to use.
1625 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1626 InFlag = Chain.getValue(1);
1628 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1629 DAG.getIntPtrConstant(0, true), InFlag);
1631 InFlag = Chain.getValue(1);
1633 // Handle result values, copying them out of physregs into vregs that we
1635 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1639 /// HandleByVal - Every parameter *after* a byval parameter is passed
1640 /// on the stack. Remember the next parameter register to allocate,
1641 /// and then confiscate the rest of the parameter registers to insure
1644 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1645 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1646 assert((State->getCallOrPrologue() == Prologue ||
1647 State->getCallOrPrologue() == Call) &&
1648 "unhandled ParmContext");
1649 if ((!State->isFirstByValRegValid()) &&
1650 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1651 State->setFirstByValReg(reg);
1652 // At a call site, a byval parameter that is split between
1653 // registers and memory needs its size truncated here. In a
1654 // function prologue, such byval parameters are reassembled in
1655 // memory, and are not truncated.
1656 if (State->getCallOrPrologue() == Call) {
1657 unsigned excess = 4 * (ARM::R4 - reg);
1658 assert(size >= excess && "expected larger existing stack allocation");
1662 // Confiscate any remaining parameter registers to preclude their
1663 // assignment to subsequent parameters.
1664 while (State->AllocateReg(GPRArgRegs, 4))
1668 /// MatchingStackOffset - Return true if the given stack call argument is
1669 /// already available in the same position (relatively) of the caller's
1670 /// incoming argument stack.
1672 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1673 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1674 const TargetInstrInfo *TII) {
1675 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1677 if (Arg.getOpcode() == ISD::CopyFromReg) {
1678 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1679 if (!TargetRegisterInfo::isVirtualRegister(VR))
1681 MachineInstr *Def = MRI->getVRegDef(VR);
1684 if (!Flags.isByVal()) {
1685 if (!TII->isLoadFromStackSlot(Def, FI))
1690 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1691 if (Flags.isByVal())
1692 // ByVal argument is passed in as a pointer but it's now being
1693 // dereferenced. e.g.
1694 // define @foo(%struct.X* %A) {
1695 // tail call @bar(%struct.X* byval %A)
1698 SDValue Ptr = Ld->getBasePtr();
1699 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1702 FI = FINode->getIndex();
1706 assert(FI != INT_MAX);
1707 if (!MFI->isFixedObjectIndex(FI))
1709 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1712 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1713 /// for tail call optimization. Targets which want to do tail call
1714 /// optimization should implement this function.
1716 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1717 CallingConv::ID CalleeCC,
1719 bool isCalleeStructRet,
1720 bool isCallerStructRet,
1721 const SmallVectorImpl<ISD::OutputArg> &Outs,
1722 const SmallVectorImpl<SDValue> &OutVals,
1723 const SmallVectorImpl<ISD::InputArg> &Ins,
1724 SelectionDAG& DAG) const {
1725 const Function *CallerF = DAG.getMachineFunction().getFunction();
1726 CallingConv::ID CallerCC = CallerF->getCallingConv();
1727 bool CCMatch = CallerCC == CalleeCC;
1729 // Look for obvious safe cases to perform tail call optimization that do not
1730 // require ABI changes. This is what gcc calls sibcall.
1732 // Do not sibcall optimize vararg calls unless the call site is not passing
1734 if (isVarArg && !Outs.empty())
1737 // Also avoid sibcall optimization if either caller or callee uses struct
1738 // return semantics.
1739 if (isCalleeStructRet || isCallerStructRet)
1742 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1743 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1744 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1745 // support in the assembler and linker to be used. This would need to be
1746 // fixed to fully support tail calls in Thumb1.
1748 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1749 // LR. This means if we need to reload LR, it takes an extra instructions,
1750 // which outweighs the value of the tail call; but here we don't know yet
1751 // whether LR is going to be used. Probably the right approach is to
1752 // generate the tail call here and turn it back into CALL/RET in
1753 // emitEpilogue if LR is used.
1755 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1756 // but we need to make sure there are enough registers; the only valid
1757 // registers are the 4 used for parameters. We don't currently do this
1759 if (Subtarget->isThumb1Only())
1762 // If the calling conventions do not match, then we'd better make sure the
1763 // results are returned in the same way as what the caller expects.
1765 SmallVector<CCValAssign, 16> RVLocs1;
1766 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1767 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1768 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1770 SmallVector<CCValAssign, 16> RVLocs2;
1771 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1772 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1773 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1775 if (RVLocs1.size() != RVLocs2.size())
1777 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1778 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1780 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1782 if (RVLocs1[i].isRegLoc()) {
1783 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1786 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1792 // If the callee takes no arguments then go on to check the results of the
1794 if (!Outs.empty()) {
1795 // Check if stack adjustment is needed. For now, do not do this if any
1796 // argument is passed on the stack.
1797 SmallVector<CCValAssign, 16> ArgLocs;
1798 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1799 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1800 CCInfo.AnalyzeCallOperands(Outs,
1801 CCAssignFnForNode(CalleeCC, false, isVarArg));
1802 if (CCInfo.getNextStackOffset()) {
1803 MachineFunction &MF = DAG.getMachineFunction();
1805 // Check if the arguments are already laid out in the right way as
1806 // the caller's fixed stack objects.
1807 MachineFrameInfo *MFI = MF.getFrameInfo();
1808 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1809 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1810 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1812 ++i, ++realArgIdx) {
1813 CCValAssign &VA = ArgLocs[i];
1814 EVT RegVT = VA.getLocVT();
1815 SDValue Arg = OutVals[realArgIdx];
1816 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1817 if (VA.getLocInfo() == CCValAssign::Indirect)
1819 if (VA.needsCustom()) {
1820 // f64 and vector types are split into multiple registers or
1821 // register/stack-slot combinations. The types will not match
1822 // the registers; give up on memory f64 refs until we figure
1823 // out what to do about this.
1826 if (!ArgLocs[++i].isRegLoc())
1828 if (RegVT == MVT::v2f64) {
1829 if (!ArgLocs[++i].isRegLoc())
1831 if (!ArgLocs[++i].isRegLoc())
1834 } else if (!VA.isRegLoc()) {
1835 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1847 ARMTargetLowering::LowerReturn(SDValue Chain,
1848 CallingConv::ID CallConv, bool isVarArg,
1849 const SmallVectorImpl<ISD::OutputArg> &Outs,
1850 const SmallVectorImpl<SDValue> &OutVals,
1851 DebugLoc dl, SelectionDAG &DAG) const {
1853 // CCValAssign - represent the assignment of the return value to a location.
1854 SmallVector<CCValAssign, 16> RVLocs;
1856 // CCState - Info about the registers and stack slots.
1857 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1858 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1860 // Analyze outgoing return values.
1861 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1864 // If this is the first return lowered for this function, add
1865 // the regs to the liveout set for the function.
1866 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1867 for (unsigned i = 0; i != RVLocs.size(); ++i)
1868 if (RVLocs[i].isRegLoc())
1869 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1874 // Copy the result values into the output registers.
1875 for (unsigned i = 0, realRVLocIdx = 0;
1877 ++i, ++realRVLocIdx) {
1878 CCValAssign &VA = RVLocs[i];
1879 assert(VA.isRegLoc() && "Can only return in registers!");
1881 SDValue Arg = OutVals[realRVLocIdx];
1883 switch (VA.getLocInfo()) {
1884 default: llvm_unreachable("Unknown loc info!");
1885 case CCValAssign::Full: break;
1886 case CCValAssign::BCvt:
1887 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1891 if (VA.needsCustom()) {
1892 if (VA.getLocVT() == MVT::v2f64) {
1893 // Extract the first half and return it in two registers.
1894 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1895 DAG.getConstant(0, MVT::i32));
1896 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1897 DAG.getVTList(MVT::i32, MVT::i32), Half);
1899 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1900 Flag = Chain.getValue(1);
1901 VA = RVLocs[++i]; // skip ahead to next loc
1902 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1903 HalfGPRs.getValue(1), Flag);
1904 Flag = Chain.getValue(1);
1905 VA = RVLocs[++i]; // skip ahead to next loc
1907 // Extract the 2nd half and fall through to handle it as an f64 value.
1908 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1909 DAG.getConstant(1, MVT::i32));
1911 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1913 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1914 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1915 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1916 Flag = Chain.getValue(1);
1917 VA = RVLocs[++i]; // skip ahead to next loc
1918 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1921 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1923 // Guarantee that all emitted copies are
1924 // stuck together, avoiding something bad.
1925 Flag = Chain.getValue(1);
1930 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1932 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1937 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1938 if (N->getNumValues() != 1)
1940 if (!N->hasNUsesOfValue(1, 0))
1943 unsigned NumCopies = 0;
1944 SDNode* Copies[2] = { 0, 0 };
1945 SDNode *Use = *N->use_begin();
1946 if (Use->getOpcode() == ISD::CopyToReg) {
1947 Copies[NumCopies++] = Use;
1948 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1949 // f64 returned in a pair of GPRs.
1950 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1952 if (UI->getOpcode() != ISD::CopyToReg)
1954 Copies[UI.getUse().getResNo()] = *UI;
1957 } else if (Use->getOpcode() == ISD::BITCAST) {
1958 // f32 returned in a single GPR.
1959 if (!Use->hasNUsesOfValue(1, 0))
1961 Use = *Use->use_begin();
1962 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1964 Copies[NumCopies++] = Use;
1969 if (NumCopies != 1 && NumCopies != 2)
1972 bool HasRet = false;
1973 for (unsigned i = 0; i < NumCopies; ++i) {
1974 SDNode *Copy = Copies[i];
1975 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1977 if (UI->getOpcode() == ISD::CopyToReg) {
1979 if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1983 if (UI->getOpcode() != ARMISD::RET_FLAG)
1992 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1993 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1996 if (!CI->isTailCall())
1999 return !Subtarget->isThumb1Only();
2002 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2003 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2004 // one of the above mentioned nodes. It has to be wrapped because otherwise
2005 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2006 // be used to form addressing mode. These wrapped nodes will be selected
2008 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2009 EVT PtrVT = Op.getValueType();
2010 // FIXME there is no actual debug info here
2011 DebugLoc dl = Op.getDebugLoc();
2012 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2014 if (CP->isMachineConstantPoolEntry())
2015 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2016 CP->getAlignment());
2018 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2019 CP->getAlignment());
2020 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2023 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2024 return MachineJumpTableInfo::EK_Inline;
2027 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2028 SelectionDAG &DAG) const {
2029 MachineFunction &MF = DAG.getMachineFunction();
2030 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2031 unsigned ARMPCLabelIndex = 0;
2032 DebugLoc DL = Op.getDebugLoc();
2033 EVT PtrVT = getPointerTy();
2034 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2035 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2037 if (RelocM == Reloc::Static) {
2038 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2040 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2041 ARMPCLabelIndex = AFI->createPICLabelUId();
2042 ARMConstantPoolValue *CPV =
2043 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2044 ARMCP::CPBlockAddress, PCAdj);
2045 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2047 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2048 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2049 MachinePointerInfo::getConstantPool(),
2050 false, false, false, 0);
2051 if (RelocM == Reloc::Static)
2053 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2054 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2057 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2059 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2060 SelectionDAG &DAG) const {
2061 DebugLoc dl = GA->getDebugLoc();
2062 EVT PtrVT = getPointerTy();
2063 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2064 MachineFunction &MF = DAG.getMachineFunction();
2065 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2066 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2067 ARMConstantPoolValue *CPV =
2068 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2069 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2070 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2071 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2072 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2073 MachinePointerInfo::getConstantPool(),
2074 false, false, false, 0);
2075 SDValue Chain = Argument.getValue(1);
2077 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2078 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2080 // call __tls_get_addr.
2083 Entry.Node = Argument;
2084 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2085 Args.push_back(Entry);
2086 // FIXME: is there useful debug info available here?
2087 std::pair<SDValue, SDValue> CallResult =
2088 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2089 false, false, false, false,
2090 0, CallingConv::C, /*isTailCall=*/false,
2091 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2092 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2093 return CallResult.first;
2096 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2097 // "local exec" model.
2099 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2100 SelectionDAG &DAG) const {
2101 const GlobalValue *GV = GA->getGlobal();
2102 DebugLoc dl = GA->getDebugLoc();
2104 SDValue Chain = DAG.getEntryNode();
2105 EVT PtrVT = getPointerTy();
2106 // Get the Thread Pointer
2107 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2109 if (GV->isDeclaration()) {
2110 MachineFunction &MF = DAG.getMachineFunction();
2111 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2112 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2113 // Initial exec model.
2114 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2115 ARMConstantPoolValue *CPV =
2116 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2117 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2119 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2120 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2121 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2122 MachinePointerInfo::getConstantPool(),
2123 false, false, false, 0);
2124 Chain = Offset.getValue(1);
2126 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2127 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2129 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2130 MachinePointerInfo::getConstantPool(),
2131 false, false, false, 0);
2134 ARMConstantPoolValue *CPV =
2135 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2136 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2137 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2138 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2139 MachinePointerInfo::getConstantPool(),
2140 false, false, false, 0);
2143 // The address of the thread local variable is the add of the thread
2144 // pointer with the offset of the variable.
2145 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2149 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2150 // TODO: implement the "local dynamic" model
2151 assert(Subtarget->isTargetELF() &&
2152 "TLS not implemented for non-ELF targets");
2153 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2154 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2155 // otherwise use the "Local Exec" TLS Model
2156 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2157 return LowerToTLSGeneralDynamicModel(GA, DAG);
2159 return LowerToTLSExecModels(GA, DAG);
2162 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2163 SelectionDAG &DAG) const {
2164 EVT PtrVT = getPointerTy();
2165 DebugLoc dl = Op.getDebugLoc();
2166 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2167 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2168 if (RelocM == Reloc::PIC_) {
2169 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2170 ARMConstantPoolValue *CPV =
2171 ARMConstantPoolConstant::Create(GV,
2172 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2173 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2174 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2175 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2177 MachinePointerInfo::getConstantPool(),
2178 false, false, false, 0);
2179 SDValue Chain = Result.getValue(1);
2180 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2181 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2183 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2184 MachinePointerInfo::getGOT(),
2185 false, false, false, 0);
2189 // If we have T2 ops, we can materialize the address directly via movt/movw
2190 // pair. This is always cheaper.
2191 if (Subtarget->useMovt()) {
2193 // FIXME: Once remat is capable of dealing with instructions with register
2194 // operands, expand this into two nodes.
2195 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2196 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2198 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2199 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2200 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2201 MachinePointerInfo::getConstantPool(),
2202 false, false, false, 0);
2206 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2207 SelectionDAG &DAG) const {
2208 EVT PtrVT = getPointerTy();
2209 DebugLoc dl = Op.getDebugLoc();
2210 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2211 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2212 MachineFunction &MF = DAG.getMachineFunction();
2213 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2215 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2216 // update ARMFastISel::ARMMaterializeGV.
2217 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2219 // FIXME: Once remat is capable of dealing with instructions with register
2220 // operands, expand this into two nodes.
2221 if (RelocM == Reloc::Static)
2222 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2223 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2225 unsigned Wrapper = (RelocM == Reloc::PIC_)
2226 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2227 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2228 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2229 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2230 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2231 MachinePointerInfo::getGOT(),
2232 false, false, false, 0);
2236 unsigned ARMPCLabelIndex = 0;
2238 if (RelocM == Reloc::Static) {
2239 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2241 ARMPCLabelIndex = AFI->createPICLabelUId();
2242 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2243 ARMConstantPoolValue *CPV =
2244 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2246 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2248 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2250 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2251 MachinePointerInfo::getConstantPool(),
2252 false, false, false, 0);
2253 SDValue Chain = Result.getValue(1);
2255 if (RelocM == Reloc::PIC_) {
2256 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2257 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2260 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2261 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2262 false, false, false, 0);
2267 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2268 SelectionDAG &DAG) const {
2269 assert(Subtarget->isTargetELF() &&
2270 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2271 MachineFunction &MF = DAG.getMachineFunction();
2272 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2273 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2274 EVT PtrVT = getPointerTy();
2275 DebugLoc dl = Op.getDebugLoc();
2276 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2277 ARMConstantPoolValue *CPV =
2278 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2279 ARMPCLabelIndex, PCAdj);
2280 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2281 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2282 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2283 MachinePointerInfo::getConstantPool(),
2284 false, false, false, 0);
2285 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2286 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2290 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2291 DebugLoc dl = Op.getDebugLoc();
2292 SDValue Val = DAG.getConstant(0, MVT::i32);
2293 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2294 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2295 Op.getOperand(1), Val);
2299 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2300 DebugLoc dl = Op.getDebugLoc();
2301 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2302 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2306 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2307 const ARMSubtarget *Subtarget) const {
2308 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2309 DebugLoc dl = Op.getDebugLoc();
2311 default: return SDValue(); // Don't custom lower most intrinsics.
2312 case Intrinsic::arm_thread_pointer: {
2313 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2314 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2316 case Intrinsic::eh_sjlj_lsda: {
2317 MachineFunction &MF = DAG.getMachineFunction();
2318 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2319 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2320 EVT PtrVT = getPointerTy();
2321 DebugLoc dl = Op.getDebugLoc();
2322 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2324 unsigned PCAdj = (RelocM != Reloc::PIC_)
2325 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2326 ARMConstantPoolValue *CPV =
2327 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2328 ARMCP::CPLSDA, PCAdj);
2329 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2330 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2332 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2333 MachinePointerInfo::getConstantPool(),
2334 false, false, false, 0);
2336 if (RelocM == Reloc::PIC_) {
2337 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2338 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2342 case Intrinsic::arm_neon_vmulls:
2343 case Intrinsic::arm_neon_vmullu: {
2344 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2345 ? ARMISD::VMULLs : ARMISD::VMULLu;
2346 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2347 Op.getOperand(1), Op.getOperand(2));
2352 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2353 const ARMSubtarget *Subtarget) {
2354 DebugLoc dl = Op.getDebugLoc();
2355 if (!Subtarget->hasDataBarrier()) {
2356 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2357 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2359 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2360 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2361 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2362 DAG.getConstant(0, MVT::i32));
2365 SDValue Op5 = Op.getOperand(5);
2366 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2367 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2368 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2369 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2371 ARM_MB::MemBOpt DMBOpt;
2372 if (isDeviceBarrier)
2373 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2375 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2376 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2377 DAG.getConstant(DMBOpt, MVT::i32));
2381 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2382 const ARMSubtarget *Subtarget) {
2383 // FIXME: handle "fence singlethread" more efficiently.
2384 DebugLoc dl = Op.getDebugLoc();
2385 if (!Subtarget->hasDataBarrier()) {
2386 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2387 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2389 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2390 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2391 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2392 DAG.getConstant(0, MVT::i32));
2395 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2396 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2399 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2400 const ARMSubtarget *Subtarget) {
2401 // ARM pre v5TE and Thumb1 does not have preload instructions.
2402 if (!(Subtarget->isThumb2() ||
2403 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2404 // Just preserve the chain.
2405 return Op.getOperand(0);
2407 DebugLoc dl = Op.getDebugLoc();
2408 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2410 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2411 // ARMv7 with MP extension has PLDW.
2412 return Op.getOperand(0);
2414 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2415 if (Subtarget->isThumb()) {
2417 isRead = ~isRead & 1;
2418 isData = ~isData & 1;
2421 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2422 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2423 DAG.getConstant(isData, MVT::i32));
2426 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2427 MachineFunction &MF = DAG.getMachineFunction();
2428 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2430 // vastart just stores the address of the VarArgsFrameIndex slot into the
2431 // memory location argument.
2432 DebugLoc dl = Op.getDebugLoc();
2433 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2434 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2435 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2436 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2437 MachinePointerInfo(SV), false, false, 0);
2441 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2442 SDValue &Root, SelectionDAG &DAG,
2443 DebugLoc dl) const {
2444 MachineFunction &MF = DAG.getMachineFunction();
2445 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2447 const TargetRegisterClass *RC;
2448 if (AFI->isThumb1OnlyFunction())
2449 RC = ARM::tGPRRegisterClass;
2451 RC = ARM::GPRRegisterClass;
2453 // Transform the arguments stored in physical registers into virtual ones.
2454 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2455 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2458 if (NextVA.isMemLoc()) {
2459 MachineFrameInfo *MFI = MF.getFrameInfo();
2460 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2462 // Create load node to retrieve arguments from the stack.
2463 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2464 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2465 MachinePointerInfo::getFixedStack(FI),
2466 false, false, false, 0);
2468 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2469 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2472 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2476 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2477 unsigned &VARegSize, unsigned &VARegSaveSize)
2480 if (CCInfo.isFirstByValRegValid())
2481 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2483 unsigned int firstUnalloced;
2484 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2485 sizeof(GPRArgRegs) /
2486 sizeof(GPRArgRegs[0]));
2487 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2490 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2491 VARegSize = NumGPRs * 4;
2492 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2495 // The remaining GPRs hold either the beginning of variable-argument
2496 // data, or the beginning of an aggregate passed by value (usuall
2497 // byval). Either way, we allocate stack slots adjacent to the data
2498 // provided by our caller, and store the unallocated registers there.
2499 // If this is a variadic function, the va_list pointer will begin with
2500 // these values; otherwise, this reassembles a (byval) structure that
2501 // was split between registers and memory.
2503 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2504 DebugLoc dl, SDValue &Chain,
2505 unsigned ArgOffset) const {
2506 MachineFunction &MF = DAG.getMachineFunction();
2507 MachineFrameInfo *MFI = MF.getFrameInfo();
2508 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2509 unsigned firstRegToSaveIndex;
2510 if (CCInfo.isFirstByValRegValid())
2511 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2513 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2514 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2517 unsigned VARegSize, VARegSaveSize;
2518 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2519 if (VARegSaveSize) {
2520 // If this function is vararg, store any remaining integer argument regs
2521 // to their spots on the stack so that they may be loaded by deferencing
2522 // the result of va_next.
2523 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2524 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2525 ArgOffset + VARegSaveSize
2528 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2531 SmallVector<SDValue, 4> MemOps;
2532 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2533 const TargetRegisterClass *RC;
2534 if (AFI->isThumb1OnlyFunction())
2535 RC = ARM::tGPRRegisterClass;
2537 RC = ARM::GPRRegisterClass;
2539 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2540 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2542 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2543 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2545 MemOps.push_back(Store);
2546 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2547 DAG.getConstant(4, getPointerTy()));
2549 if (!MemOps.empty())
2550 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2551 &MemOps[0], MemOps.size());
2553 // This will point to the next argument passed via stack.
2554 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2558 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2559 CallingConv::ID CallConv, bool isVarArg,
2560 const SmallVectorImpl<ISD::InputArg>
2562 DebugLoc dl, SelectionDAG &DAG,
2563 SmallVectorImpl<SDValue> &InVals)
2565 MachineFunction &MF = DAG.getMachineFunction();
2566 MachineFrameInfo *MFI = MF.getFrameInfo();
2568 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2570 // Assign locations to all of the incoming arguments.
2571 SmallVector<CCValAssign, 16> ArgLocs;
2572 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2573 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2574 CCInfo.AnalyzeFormalArguments(Ins,
2575 CCAssignFnForNode(CallConv, /* Return*/ false,
2578 SmallVector<SDValue, 16> ArgValues;
2579 int lastInsIndex = -1;
2582 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2583 CCValAssign &VA = ArgLocs[i];
2585 // Arguments stored in registers.
2586 if (VA.isRegLoc()) {
2587 EVT RegVT = VA.getLocVT();
2589 if (VA.needsCustom()) {
2590 // f64 and vector types are split up into multiple registers or
2591 // combinations of registers and stack slots.
2592 if (VA.getLocVT() == MVT::v2f64) {
2593 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2595 VA = ArgLocs[++i]; // skip ahead to next loc
2597 if (VA.isMemLoc()) {
2598 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2599 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2600 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2601 MachinePointerInfo::getFixedStack(FI),
2602 false, false, false, 0);
2604 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2607 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2608 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2609 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2610 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2611 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2613 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2616 const TargetRegisterClass *RC;
2618 if (RegVT == MVT::f32)
2619 RC = ARM::SPRRegisterClass;
2620 else if (RegVT == MVT::f64)
2621 RC = ARM::DPRRegisterClass;
2622 else if (RegVT == MVT::v2f64)
2623 RC = ARM::QPRRegisterClass;
2624 else if (RegVT == MVT::i32)
2625 RC = (AFI->isThumb1OnlyFunction() ?
2626 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2628 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2630 // Transform the arguments in physical registers into virtual ones.
2631 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2632 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2635 // If this is an 8 or 16-bit value, it is really passed promoted
2636 // to 32 bits. Insert an assert[sz]ext to capture this, then
2637 // truncate to the right size.
2638 switch (VA.getLocInfo()) {
2639 default: llvm_unreachable("Unknown loc info!");
2640 case CCValAssign::Full: break;
2641 case CCValAssign::BCvt:
2642 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2644 case CCValAssign::SExt:
2645 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2646 DAG.getValueType(VA.getValVT()));
2647 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2649 case CCValAssign::ZExt:
2650 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2651 DAG.getValueType(VA.getValVT()));
2652 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2656 InVals.push_back(ArgValue);
2658 } else { // VA.isRegLoc()
2661 assert(VA.isMemLoc());
2662 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2664 int index = ArgLocs[i].getValNo();
2666 // Some Ins[] entries become multiple ArgLoc[] entries.
2667 // Process them only once.
2668 if (index != lastInsIndex)
2670 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2671 // FIXME: For now, all byval parameter objects are marked mutable.
2672 // This can be changed with more analysis.
2673 // In case of tail call optimization mark all arguments mutable.
2674 // Since they could be overwritten by lowering of arguments in case of
2676 if (Flags.isByVal()) {
2677 unsigned VARegSize, VARegSaveSize;
2678 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2679 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2680 unsigned Bytes = Flags.getByValSize() - VARegSize;
2681 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2682 int FI = MFI->CreateFixedObject(Bytes,
2683 VA.getLocMemOffset(), false);
2684 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2686 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2687 VA.getLocMemOffset(), true);
2689 // Create load nodes to retrieve arguments from the stack.
2690 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2691 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2692 MachinePointerInfo::getFixedStack(FI),
2693 false, false, false, 0));
2695 lastInsIndex = index;
2702 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2707 /// isFloatingPointZero - Return true if this is +0.0.
2708 static bool isFloatingPointZero(SDValue Op) {
2709 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2710 return CFP->getValueAPF().isPosZero();
2711 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2712 // Maybe this has already been legalized into the constant pool?
2713 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2714 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2715 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2716 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2717 return CFP->getValueAPF().isPosZero();
2723 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2724 /// the given operands.
2726 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2727 SDValue &ARMcc, SelectionDAG &DAG,
2728 DebugLoc dl) const {
2729 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2730 unsigned C = RHSC->getZExtValue();
2731 if (!isLegalICmpImmediate(C)) {
2732 // Constant does not fit, try adjusting it by one?
2737 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2738 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2739 RHS = DAG.getConstant(C-1, MVT::i32);
2744 if (C != 0 && isLegalICmpImmediate(C-1)) {
2745 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2746 RHS = DAG.getConstant(C-1, MVT::i32);
2751 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2752 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2753 RHS = DAG.getConstant(C+1, MVT::i32);
2758 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2759 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2760 RHS = DAG.getConstant(C+1, MVT::i32);
2767 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2768 ARMISD::NodeType CompareType;
2771 CompareType = ARMISD::CMP;
2776 CompareType = ARMISD::CMPZ;
2779 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2780 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2783 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2785 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2786 DebugLoc dl) const {
2788 if (!isFloatingPointZero(RHS))
2789 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2791 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2792 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2795 /// duplicateCmp - Glue values can have only one use, so this function
2796 /// duplicates a comparison node.
2798 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2799 unsigned Opc = Cmp.getOpcode();
2800 DebugLoc DL = Cmp.getDebugLoc();
2801 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2802 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2804 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2805 Cmp = Cmp.getOperand(0);
2806 Opc = Cmp.getOpcode();
2807 if (Opc == ARMISD::CMPFP)
2808 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2810 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2811 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2813 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2816 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2817 SDValue Cond = Op.getOperand(0);
2818 SDValue SelectTrue = Op.getOperand(1);
2819 SDValue SelectFalse = Op.getOperand(2);
2820 DebugLoc dl = Op.getDebugLoc();
2824 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2825 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2827 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2828 const ConstantSDNode *CMOVTrue =
2829 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2830 const ConstantSDNode *CMOVFalse =
2831 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2833 if (CMOVTrue && CMOVFalse) {
2834 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2835 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2839 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2841 False = SelectFalse;
2842 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2847 if (True.getNode() && False.getNode()) {
2848 EVT VT = Op.getValueType();
2849 SDValue ARMcc = Cond.getOperand(2);
2850 SDValue CCR = Cond.getOperand(3);
2851 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2852 assert(True.getValueType() == VT);
2853 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2858 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2859 // undefined bits before doing a full-word comparison with zero.
2860 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2861 DAG.getConstant(1, Cond.getValueType()));
2863 return DAG.getSelectCC(dl, Cond,
2864 DAG.getConstant(0, Cond.getValueType()),
2865 SelectTrue, SelectFalse, ISD::SETNE);
2868 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2869 EVT VT = Op.getValueType();
2870 SDValue LHS = Op.getOperand(0);
2871 SDValue RHS = Op.getOperand(1);
2872 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2873 SDValue TrueVal = Op.getOperand(2);
2874 SDValue FalseVal = Op.getOperand(3);
2875 DebugLoc dl = Op.getDebugLoc();
2877 if (LHS.getValueType() == MVT::i32) {
2879 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2880 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2881 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2884 ARMCC::CondCodes CondCode, CondCode2;
2885 FPCCToARMCC(CC, CondCode, CondCode2);
2887 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2888 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2889 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2890 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2892 if (CondCode2 != ARMCC::AL) {
2893 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2894 // FIXME: Needs another CMP because flag can have but one use.
2895 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2896 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2897 Result, TrueVal, ARMcc2, CCR, Cmp2);
2902 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2903 /// to morph to an integer compare sequence.
2904 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2905 const ARMSubtarget *Subtarget) {
2906 SDNode *N = Op.getNode();
2907 if (!N->hasOneUse())
2908 // Otherwise it requires moving the value from fp to integer registers.
2910 if (!N->getNumValues())
2912 EVT VT = Op.getValueType();
2913 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2914 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2915 // vmrs are very slow, e.g. cortex-a8.
2918 if (isFloatingPointZero(Op)) {
2922 return ISD::isNormalLoad(N);
2925 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2926 if (isFloatingPointZero(Op))
2927 return DAG.getConstant(0, MVT::i32);
2929 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2930 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2931 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2932 Ld->isVolatile(), Ld->isNonTemporal(),
2933 Ld->isInvariant(), Ld->getAlignment());
2935 llvm_unreachable("Unknown VFP cmp argument!");
2938 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2939 SDValue &RetVal1, SDValue &RetVal2) {
2940 if (isFloatingPointZero(Op)) {
2941 RetVal1 = DAG.getConstant(0, MVT::i32);
2942 RetVal2 = DAG.getConstant(0, MVT::i32);
2946 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2947 SDValue Ptr = Ld->getBasePtr();
2948 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2949 Ld->getChain(), Ptr,
2950 Ld->getPointerInfo(),
2951 Ld->isVolatile(), Ld->isNonTemporal(),
2952 Ld->isInvariant(), Ld->getAlignment());
2954 EVT PtrType = Ptr.getValueType();
2955 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2956 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2957 PtrType, Ptr, DAG.getConstant(4, PtrType));
2958 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2959 Ld->getChain(), NewPtr,
2960 Ld->getPointerInfo().getWithOffset(4),
2961 Ld->isVolatile(), Ld->isNonTemporal(),
2962 Ld->isInvariant(), NewAlign);
2966 llvm_unreachable("Unknown VFP cmp argument!");
2969 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2970 /// f32 and even f64 comparisons to integer ones.
2972 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2973 SDValue Chain = Op.getOperand(0);
2974 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2975 SDValue LHS = Op.getOperand(2);
2976 SDValue RHS = Op.getOperand(3);
2977 SDValue Dest = Op.getOperand(4);
2978 DebugLoc dl = Op.getDebugLoc();
2980 bool LHSSeenZero = false;
2981 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
2982 bool RHSSeenZero = false;
2983 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
2984 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
2985 // If unsafe fp math optimization is enabled and there are no other uses of
2986 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2987 // to an integer comparison.
2988 if (CC == ISD::SETOEQ)
2990 else if (CC == ISD::SETUNE)
2993 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
2995 if (LHS.getValueType() == MVT::f32) {
2996 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2997 bitcastf32Toi32(LHS, DAG), Mask);
2998 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2999 bitcastf32Toi32(RHS, DAG), Mask);
3000 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3001 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3002 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3003 Chain, Dest, ARMcc, CCR, Cmp);
3008 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3009 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3010 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3011 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3012 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3013 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3014 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3015 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3016 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3022 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3023 SDValue Chain = Op.getOperand(0);
3024 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3025 SDValue LHS = Op.getOperand(2);
3026 SDValue RHS = Op.getOperand(3);
3027 SDValue Dest = Op.getOperand(4);
3028 DebugLoc dl = Op.getDebugLoc();
3030 if (LHS.getValueType() == MVT::i32) {
3032 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3033 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3034 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3035 Chain, Dest, ARMcc, CCR, Cmp);
3038 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3040 if (getTargetMachine().Options.UnsafeFPMath &&
3041 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3042 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3043 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3044 if (Result.getNode())
3048 ARMCC::CondCodes CondCode, CondCode2;
3049 FPCCToARMCC(CC, CondCode, CondCode2);
3051 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3052 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3053 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3054 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3055 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3056 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3057 if (CondCode2 != ARMCC::AL) {
3058 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3059 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3060 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3065 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3066 SDValue Chain = Op.getOperand(0);
3067 SDValue Table = Op.getOperand(1);
3068 SDValue Index = Op.getOperand(2);
3069 DebugLoc dl = Op.getDebugLoc();
3071 EVT PTy = getPointerTy();
3072 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3073 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3074 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3075 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3076 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3077 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3078 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3079 if (Subtarget->isThumb2()) {
3080 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3081 // which does another jump to the destination. This also makes it easier
3082 // to translate it to TBB / TBH later.
3083 // FIXME: This might not work if the function is extremely large.
3084 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3085 Addr, Op.getOperand(2), JTI, UId);
3087 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3088 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3089 MachinePointerInfo::getJumpTable(),
3090 false, false, false, 0);
3091 Chain = Addr.getValue(1);
3092 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3093 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3095 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3096 MachinePointerInfo::getJumpTable(),
3097 false, false, false, 0);
3098 Chain = Addr.getValue(1);
3099 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3103 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3104 EVT VT = Op.getValueType();
3105 DebugLoc dl = Op.getDebugLoc();
3107 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3108 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3110 return DAG.UnrollVectorOp(Op.getNode());
3113 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3114 "Invalid type for custom lowering!");
3115 if (VT != MVT::v4i16)
3116 return DAG.UnrollVectorOp(Op.getNode());
3118 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3119 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3122 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3123 EVT VT = Op.getValueType();
3125 return LowerVectorFP_TO_INT(Op, DAG);
3127 DebugLoc dl = Op.getDebugLoc();
3130 switch (Op.getOpcode()) {
3131 default: llvm_unreachable("Invalid opcode!");
3132 case ISD::FP_TO_SINT:
3133 Opc = ARMISD::FTOSI;
3135 case ISD::FP_TO_UINT:
3136 Opc = ARMISD::FTOUI;
3139 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3140 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3143 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3144 EVT VT = Op.getValueType();
3145 DebugLoc dl = Op.getDebugLoc();
3147 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3148 if (VT.getVectorElementType() == MVT::f32)
3150 return DAG.UnrollVectorOp(Op.getNode());
3153 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3154 "Invalid type for custom lowering!");
3155 if (VT != MVT::v4f32)
3156 return DAG.UnrollVectorOp(Op.getNode());
3160 switch (Op.getOpcode()) {
3161 default: llvm_unreachable("Invalid opcode!");
3162 case ISD::SINT_TO_FP:
3163 CastOpc = ISD::SIGN_EXTEND;
3164 Opc = ISD::SINT_TO_FP;
3166 case ISD::UINT_TO_FP:
3167 CastOpc = ISD::ZERO_EXTEND;
3168 Opc = ISD::UINT_TO_FP;
3172 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3173 return DAG.getNode(Opc, dl, VT, Op);
3176 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3177 EVT VT = Op.getValueType();
3179 return LowerVectorINT_TO_FP(Op, DAG);
3181 DebugLoc dl = Op.getDebugLoc();
3184 switch (Op.getOpcode()) {
3185 default: llvm_unreachable("Invalid opcode!");
3186 case ISD::SINT_TO_FP:
3187 Opc = ARMISD::SITOF;
3189 case ISD::UINT_TO_FP:
3190 Opc = ARMISD::UITOF;
3194 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3195 return DAG.getNode(Opc, dl, VT, Op);
3198 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3199 // Implement fcopysign with a fabs and a conditional fneg.
3200 SDValue Tmp0 = Op.getOperand(0);
3201 SDValue Tmp1 = Op.getOperand(1);
3202 DebugLoc dl = Op.getDebugLoc();
3203 EVT VT = Op.getValueType();
3204 EVT SrcVT = Tmp1.getValueType();
3205 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3206 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3207 bool UseNEON = !InGPR && Subtarget->hasNEON();
3210 // Use VBSL to copy the sign bit.
3211 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3212 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3213 DAG.getTargetConstant(EncodedVal, MVT::i32));
3214 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3216 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3217 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3218 DAG.getConstant(32, MVT::i32));
3219 else /*if (VT == MVT::f32)*/
3220 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3221 if (SrcVT == MVT::f32) {
3222 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3224 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3225 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3226 DAG.getConstant(32, MVT::i32));
3227 } else if (VT == MVT::f32)
3228 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3229 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3230 DAG.getConstant(32, MVT::i32));
3231 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3232 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3234 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3236 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3237 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3238 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3240 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3241 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3242 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3243 if (VT == MVT::f32) {
3244 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3245 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3246 DAG.getConstant(0, MVT::i32));
3248 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3254 // Bitcast operand 1 to i32.
3255 if (SrcVT == MVT::f64)
3256 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3257 &Tmp1, 1).getValue(1);
3258 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3260 // Or in the signbit with integer operations.
3261 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3262 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3263 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3264 if (VT == MVT::f32) {
3265 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3266 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3267 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3268 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3271 // f64: Or the high part with signbit and then combine two parts.
3272 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3274 SDValue Lo = Tmp0.getValue(0);
3275 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3276 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3277 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3280 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3281 MachineFunction &MF = DAG.getMachineFunction();
3282 MachineFrameInfo *MFI = MF.getFrameInfo();
3283 MFI->setReturnAddressIsTaken(true);
3285 EVT VT = Op.getValueType();
3286 DebugLoc dl = Op.getDebugLoc();
3287 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3289 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3290 SDValue Offset = DAG.getConstant(4, MVT::i32);
3291 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3292 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3293 MachinePointerInfo(), false, false, false, 0);
3296 // Return LR, which contains the return address. Mark it an implicit live-in.
3297 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3298 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3301 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3302 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3303 MFI->setFrameAddressIsTaken(true);
3305 EVT VT = Op.getValueType();
3306 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3307 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3308 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3309 ? ARM::R7 : ARM::R11;
3310 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3312 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3313 MachinePointerInfo(),
3314 false, false, false, 0);
3318 /// ExpandBITCAST - If the target supports VFP, this function is called to
3319 /// expand a bit convert where either the source or destination type is i64 to
3320 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3321 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3322 /// vectors), since the legalizer won't know what to do with that.
3323 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3324 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3325 DebugLoc dl = N->getDebugLoc();
3326 SDValue Op = N->getOperand(0);
3328 // This function is only supposed to be called for i64 types, either as the
3329 // source or destination of the bit convert.
3330 EVT SrcVT = Op.getValueType();
3331 EVT DstVT = N->getValueType(0);
3332 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3333 "ExpandBITCAST called for non-i64 type");
3335 // Turn i64->f64 into VMOVDRR.
3336 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3337 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3338 DAG.getConstant(0, MVT::i32));
3339 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3340 DAG.getConstant(1, MVT::i32));
3341 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3342 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3345 // Turn f64->i64 into VMOVRRD.
3346 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3347 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3348 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3349 // Merge the pieces into a single i64 value.
3350 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3356 /// getZeroVector - Returns a vector of specified type with all zero elements.
3357 /// Zero vectors are used to represent vector negation and in those cases
3358 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3359 /// not support i64 elements, so sometimes the zero vectors will need to be
3360 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3362 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3363 assert(VT.isVector() && "Expected a vector type");
3364 // The canonical modified immediate encoding of a zero vector is....0!
3365 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3366 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3367 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3368 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3371 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3372 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3373 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3374 SelectionDAG &DAG) const {
3375 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3376 EVT VT = Op.getValueType();
3377 unsigned VTBits = VT.getSizeInBits();
3378 DebugLoc dl = Op.getDebugLoc();
3379 SDValue ShOpLo = Op.getOperand(0);
3380 SDValue ShOpHi = Op.getOperand(1);
3381 SDValue ShAmt = Op.getOperand(2);
3383 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3385 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3387 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3388 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3389 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3390 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3391 DAG.getConstant(VTBits, MVT::i32));
3392 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3393 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3394 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3396 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3397 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3399 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3400 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3403 SDValue Ops[2] = { Lo, Hi };
3404 return DAG.getMergeValues(Ops, 2, dl);
3407 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3408 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3409 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3410 SelectionDAG &DAG) const {
3411 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3412 EVT VT = Op.getValueType();
3413 unsigned VTBits = VT.getSizeInBits();
3414 DebugLoc dl = Op.getDebugLoc();
3415 SDValue ShOpLo = Op.getOperand(0);
3416 SDValue ShOpHi = Op.getOperand(1);
3417 SDValue ShAmt = Op.getOperand(2);
3420 assert(Op.getOpcode() == ISD::SHL_PARTS);
3421 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3422 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3423 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3424 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3425 DAG.getConstant(VTBits, MVT::i32));
3426 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3427 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3429 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3430 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3431 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3433 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3434 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3437 SDValue Ops[2] = { Lo, Hi };
3438 return DAG.getMergeValues(Ops, 2, dl);
3441 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3442 SelectionDAG &DAG) const {
3443 // The rounding mode is in bits 23:22 of the FPSCR.
3444 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3445 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3446 // so that the shift + and get folded into a bitfield extract.
3447 DebugLoc dl = Op.getDebugLoc();
3448 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3449 DAG.getConstant(Intrinsic::arm_get_fpscr,
3451 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3452 DAG.getConstant(1U << 22, MVT::i32));
3453 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3454 DAG.getConstant(22, MVT::i32));
3455 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3456 DAG.getConstant(3, MVT::i32));
3459 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3460 const ARMSubtarget *ST) {
3461 EVT VT = N->getValueType(0);
3462 DebugLoc dl = N->getDebugLoc();
3464 if (!ST->hasV6T2Ops())
3467 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3468 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3471 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3472 const ARMSubtarget *ST) {
3473 EVT VT = N->getValueType(0);
3474 DebugLoc dl = N->getDebugLoc();
3479 // Lower vector shifts on NEON to use VSHL.
3480 assert(ST->hasNEON() && "unexpected vector shift");
3482 // Left shifts translate directly to the vshiftu intrinsic.
3483 if (N->getOpcode() == ISD::SHL)
3484 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3485 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3486 N->getOperand(0), N->getOperand(1));
3488 assert((N->getOpcode() == ISD::SRA ||
3489 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3491 // NEON uses the same intrinsics for both left and right shifts. For
3492 // right shifts, the shift amounts are negative, so negate the vector of
3494 EVT ShiftVT = N->getOperand(1).getValueType();
3495 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3496 getZeroVector(ShiftVT, DAG, dl),
3498 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3499 Intrinsic::arm_neon_vshifts :
3500 Intrinsic::arm_neon_vshiftu);
3501 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3502 DAG.getConstant(vshiftInt, MVT::i32),
3503 N->getOperand(0), NegatedCount);
3506 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3507 const ARMSubtarget *ST) {
3508 EVT VT = N->getValueType(0);
3509 DebugLoc dl = N->getDebugLoc();
3511 // We can get here for a node like i32 = ISD::SHL i32, i64
3515 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3516 "Unknown shift to lower!");
3518 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3519 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3520 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3523 // If we are in thumb mode, we don't have RRX.
3524 if (ST->isThumb1Only()) return SDValue();
3526 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3527 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3528 DAG.getConstant(0, MVT::i32));
3529 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3530 DAG.getConstant(1, MVT::i32));
3532 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3533 // captures the result into a carry flag.
3534 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3535 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3537 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3538 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3540 // Merge the pieces into a single i64 value.
3541 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3544 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3545 SDValue TmpOp0, TmpOp1;
3546 bool Invert = false;
3550 SDValue Op0 = Op.getOperand(0);
3551 SDValue Op1 = Op.getOperand(1);
3552 SDValue CC = Op.getOperand(2);
3553 EVT VT = Op.getValueType();
3554 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3555 DebugLoc dl = Op.getDebugLoc();
3557 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3558 switch (SetCCOpcode) {
3559 default: llvm_unreachable("Illegal FP comparison");
3561 case ISD::SETNE: Invert = true; // Fallthrough
3563 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3565 case ISD::SETLT: Swap = true; // Fallthrough
3567 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3569 case ISD::SETLE: Swap = true; // Fallthrough
3571 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3572 case ISD::SETUGE: Swap = true; // Fallthrough
3573 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3574 case ISD::SETUGT: Swap = true; // Fallthrough
3575 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3576 case ISD::SETUEQ: Invert = true; // Fallthrough
3578 // Expand this to (OLT | OGT).
3582 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3583 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3585 case ISD::SETUO: Invert = true; // Fallthrough
3587 // Expand this to (OLT | OGE).
3591 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3592 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3596 // Integer comparisons.
3597 switch (SetCCOpcode) {
3598 default: llvm_unreachable("Illegal integer comparison");
3599 case ISD::SETNE: Invert = true;
3600 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3601 case ISD::SETLT: Swap = true;
3602 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3603 case ISD::SETLE: Swap = true;
3604 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3605 case ISD::SETULT: Swap = true;
3606 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3607 case ISD::SETULE: Swap = true;
3608 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3611 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3612 if (Opc == ARMISD::VCEQ) {
3615 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3617 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3620 // Ignore bitconvert.
3621 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3622 AndOp = AndOp.getOperand(0);
3624 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3626 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3627 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3634 std::swap(Op0, Op1);
3636 // If one of the operands is a constant vector zero, attempt to fold the
3637 // comparison to a specialized compare-against-zero form.
3639 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3641 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3642 if (Opc == ARMISD::VCGE)
3643 Opc = ARMISD::VCLEZ;
3644 else if (Opc == ARMISD::VCGT)
3645 Opc = ARMISD::VCLTZ;
3650 if (SingleOp.getNode()) {
3653 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3655 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3657 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3659 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3661 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3663 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3666 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3670 Result = DAG.getNOT(dl, Result, VT);
3675 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3676 /// valid vector constant for a NEON instruction with a "modified immediate"
3677 /// operand (e.g., VMOV). If so, return the encoded value.
3678 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3679 unsigned SplatBitSize, SelectionDAG &DAG,
3680 EVT &VT, bool is128Bits, NEONModImmType type) {
3681 unsigned OpCmode, Imm;
3683 // SplatBitSize is set to the smallest size that splats the vector, so a
3684 // zero vector will always have SplatBitSize == 8. However, NEON modified
3685 // immediate instructions others than VMOV do not support the 8-bit encoding
3686 // of a zero vector, and the default encoding of zero is supposed to be the
3691 switch (SplatBitSize) {
3693 if (type != VMOVModImm)
3695 // Any 1-byte value is OK. Op=0, Cmode=1110.
3696 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3699 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3703 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3704 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3705 if ((SplatBits & ~0xff) == 0) {
3706 // Value = 0x00nn: Op=x, Cmode=100x.
3711 if ((SplatBits & ~0xff00) == 0) {
3712 // Value = 0xnn00: Op=x, Cmode=101x.
3714 Imm = SplatBits >> 8;
3720 // NEON's 32-bit VMOV supports splat values where:
3721 // * only one byte is nonzero, or
3722 // * the least significant byte is 0xff and the second byte is nonzero, or
3723 // * the least significant 2 bytes are 0xff and the third is nonzero.
3724 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3725 if ((SplatBits & ~0xff) == 0) {
3726 // Value = 0x000000nn: Op=x, Cmode=000x.
3731 if ((SplatBits & ~0xff00) == 0) {
3732 // Value = 0x0000nn00: Op=x, Cmode=001x.
3734 Imm = SplatBits >> 8;
3737 if ((SplatBits & ~0xff0000) == 0) {
3738 // Value = 0x00nn0000: Op=x, Cmode=010x.
3740 Imm = SplatBits >> 16;
3743 if ((SplatBits & ~0xff000000) == 0) {
3744 // Value = 0xnn000000: Op=x, Cmode=011x.
3746 Imm = SplatBits >> 24;
3750 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3751 if (type == OtherModImm) return SDValue();
3753 if ((SplatBits & ~0xffff) == 0 &&
3754 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3755 // Value = 0x0000nnff: Op=x, Cmode=1100.
3757 Imm = SplatBits >> 8;
3762 if ((SplatBits & ~0xffffff) == 0 &&
3763 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3764 // Value = 0x00nnffff: Op=x, Cmode=1101.
3766 Imm = SplatBits >> 16;
3767 SplatBits |= 0xffff;
3771 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3772 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3773 // VMOV.I32. A (very) minor optimization would be to replicate the value
3774 // and fall through here to test for a valid 64-bit splat. But, then the
3775 // caller would also need to check and handle the change in size.
3779 if (type != VMOVModImm)
3781 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3782 uint64_t BitMask = 0xff;
3784 unsigned ImmMask = 1;
3786 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3787 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3790 } else if ((SplatBits & BitMask) != 0) {
3796 // Op=1, Cmode=1110.
3799 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3804 llvm_unreachable("unexpected size for isNEONModifiedImm");
3807 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3808 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3811 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3812 const ARMSubtarget *ST) const {
3813 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3816 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3817 assert(Op.getValueType() == MVT::f32 &&
3818 "ConstantFP custom lowering should only occur for f32.");
3820 // Try splatting with a VMOV.f32...
3821 APFloat FPVal = CFP->getValueAPF();
3822 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3824 DebugLoc DL = Op.getDebugLoc();
3825 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3826 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3828 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3829 DAG.getConstant(0, MVT::i32));
3832 // If that fails, try a VMOV.i32
3834 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3835 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3837 if (NewVal != SDValue()) {
3838 DebugLoc DL = Op.getDebugLoc();
3839 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3841 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3843 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3844 DAG.getConstant(0, MVT::i32));
3847 // Finally, try a VMVN.i32
3848 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3850 if (NewVal != SDValue()) {
3851 DebugLoc DL = Op.getDebugLoc();
3852 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3853 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3855 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3856 DAG.getConstant(0, MVT::i32));
3863 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3864 bool &ReverseVEXT, unsigned &Imm) {
3865 unsigned NumElts = VT.getVectorNumElements();
3866 ReverseVEXT = false;
3868 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3874 // If this is a VEXT shuffle, the immediate value is the index of the first
3875 // element. The other shuffle indices must be the successive elements after
3877 unsigned ExpectedElt = Imm;
3878 for (unsigned i = 1; i < NumElts; ++i) {
3879 // Increment the expected index. If it wraps around, it may still be
3880 // a VEXT but the source vectors must be swapped.
3882 if (ExpectedElt == NumElts * 2) {
3887 if (M[i] < 0) continue; // ignore UNDEF indices
3888 if (ExpectedElt != static_cast<unsigned>(M[i]))
3892 // Adjust the index value if the source operands will be swapped.
3899 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3900 /// instruction with the specified blocksize. (The order of the elements
3901 /// within each block of the vector is reversed.)
3902 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3903 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3904 "Only possible block sizes for VREV are: 16, 32, 64");
3906 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3910 unsigned NumElts = VT.getVectorNumElements();
3911 unsigned BlockElts = M[0] + 1;
3912 // If the first shuffle index is UNDEF, be optimistic.
3914 BlockElts = BlockSize / EltSz;
3916 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3919 for (unsigned i = 0; i < NumElts; ++i) {
3920 if (M[i] < 0) continue; // ignore UNDEF indices
3921 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3928 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3929 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3930 // range, then 0 is placed into the resulting vector. So pretty much any mask
3931 // of 8 elements can work here.
3932 return VT == MVT::v8i8 && M.size() == 8;
3935 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3936 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3940 unsigned NumElts = VT.getVectorNumElements();
3941 WhichResult = (M[0] == 0 ? 0 : 1);
3942 for (unsigned i = 0; i < NumElts; i += 2) {
3943 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3944 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3950 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3951 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3952 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3953 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3954 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3958 unsigned NumElts = VT.getVectorNumElements();
3959 WhichResult = (M[0] == 0 ? 0 : 1);
3960 for (unsigned i = 0; i < NumElts; i += 2) {
3961 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3962 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3968 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3969 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3973 unsigned NumElts = VT.getVectorNumElements();
3974 WhichResult = (M[0] == 0 ? 0 : 1);
3975 for (unsigned i = 0; i != NumElts; ++i) {
3976 if (M[i] < 0) continue; // ignore UNDEF indices
3977 if ((unsigned) M[i] != 2 * i + WhichResult)
3981 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3982 if (VT.is64BitVector() && EltSz == 32)
3988 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3989 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3990 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3991 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3992 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3996 unsigned Half = VT.getVectorNumElements() / 2;
3997 WhichResult = (M[0] == 0 ? 0 : 1);
3998 for (unsigned j = 0; j != 2; ++j) {
3999 unsigned Idx = WhichResult;
4000 for (unsigned i = 0; i != Half; ++i) {
4001 int MIdx = M[i + j * Half];
4002 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4008 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4009 if (VT.is64BitVector() && EltSz == 32)
4015 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4016 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4020 unsigned NumElts = VT.getVectorNumElements();
4021 WhichResult = (M[0] == 0 ? 0 : 1);
4022 unsigned Idx = WhichResult * NumElts / 2;
4023 for (unsigned i = 0; i != NumElts; i += 2) {
4024 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4025 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4030 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4031 if (VT.is64BitVector() && EltSz == 32)
4037 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4038 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4039 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4040 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4041 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4045 unsigned NumElts = VT.getVectorNumElements();
4046 WhichResult = (M[0] == 0 ? 0 : 1);
4047 unsigned Idx = WhichResult * NumElts / 2;
4048 for (unsigned i = 0; i != NumElts; i += 2) {
4049 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4050 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4055 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4056 if (VT.is64BitVector() && EltSz == 32)
4062 // If N is an integer constant that can be moved into a register in one
4063 // instruction, return an SDValue of such a constant (will become a MOV
4064 // instruction). Otherwise return null.
4065 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4066 const ARMSubtarget *ST, DebugLoc dl) {
4068 if (!isa<ConstantSDNode>(N))
4070 Val = cast<ConstantSDNode>(N)->getZExtValue();
4072 if (ST->isThumb1Only()) {
4073 if (Val <= 255 || ~Val <= 255)
4074 return DAG.getConstant(Val, MVT::i32);
4076 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4077 return DAG.getConstant(Val, MVT::i32);
4082 // If this is a case we can't handle, return null and let the default
4083 // expansion code take care of it.
4084 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4085 const ARMSubtarget *ST) const {
4086 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4087 DebugLoc dl = Op.getDebugLoc();
4088 EVT VT = Op.getValueType();
4090 APInt SplatBits, SplatUndef;
4091 unsigned SplatBitSize;
4093 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4094 if (SplatBitSize <= 64) {
4095 // Check if an immediate VMOV works.
4097 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4098 SplatUndef.getZExtValue(), SplatBitSize,
4099 DAG, VmovVT, VT.is128BitVector(),
4101 if (Val.getNode()) {
4102 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4103 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4106 // Try an immediate VMVN.
4107 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4108 Val = isNEONModifiedImm(NegatedImm,
4109 SplatUndef.getZExtValue(), SplatBitSize,
4110 DAG, VmovVT, VT.is128BitVector(),
4112 if (Val.getNode()) {
4113 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4114 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4117 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4118 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4119 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4121 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4122 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4128 // Scan through the operands to see if only one value is used.
4129 unsigned NumElts = VT.getVectorNumElements();
4130 bool isOnlyLowElement = true;
4131 bool usesOnlyOneValue = true;
4132 bool isConstant = true;
4134 for (unsigned i = 0; i < NumElts; ++i) {
4135 SDValue V = Op.getOperand(i);
4136 if (V.getOpcode() == ISD::UNDEF)
4139 isOnlyLowElement = false;
4140 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4143 if (!Value.getNode())
4145 else if (V != Value)
4146 usesOnlyOneValue = false;
4149 if (!Value.getNode())
4150 return DAG.getUNDEF(VT);
4152 if (isOnlyLowElement)
4153 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4155 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4157 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4158 // i32 and try again.
4159 if (usesOnlyOneValue && EltSize <= 32) {
4161 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4162 if (VT.getVectorElementType().isFloatingPoint()) {
4163 SmallVector<SDValue, 8> Ops;
4164 for (unsigned i = 0; i < NumElts; ++i)
4165 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4167 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4168 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4169 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4171 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4173 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4175 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4178 // If all elements are constants and the case above didn't get hit, fall back
4179 // to the default expansion, which will generate a load from the constant
4184 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4186 SDValue shuffle = ReconstructShuffle(Op, DAG);
4187 if (shuffle != SDValue())
4191 // Vectors with 32- or 64-bit elements can be built by directly assigning
4192 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4193 // will be legalized.
4194 if (EltSize >= 32) {
4195 // Do the expansion with floating-point types, since that is what the VFP
4196 // registers are defined to use, and since i64 is not legal.
4197 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4198 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4199 SmallVector<SDValue, 8> Ops;
4200 for (unsigned i = 0; i < NumElts; ++i)
4201 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4202 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4203 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4209 // Gather data to see if the operation can be modelled as a
4210 // shuffle in combination with VEXTs.
4211 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4212 SelectionDAG &DAG) const {
4213 DebugLoc dl = Op.getDebugLoc();
4214 EVT VT = Op.getValueType();
4215 unsigned NumElts = VT.getVectorNumElements();
4217 SmallVector<SDValue, 2> SourceVecs;
4218 SmallVector<unsigned, 2> MinElts;
4219 SmallVector<unsigned, 2> MaxElts;
4221 for (unsigned i = 0; i < NumElts; ++i) {
4222 SDValue V = Op.getOperand(i);
4223 if (V.getOpcode() == ISD::UNDEF)
4225 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4226 // A shuffle can only come from building a vector from various
4227 // elements of other vectors.
4229 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4230 VT.getVectorElementType()) {
4231 // This code doesn't know how to handle shuffles where the vector
4232 // element types do not match (this happens because type legalization
4233 // promotes the return type of EXTRACT_VECTOR_ELT).
4234 // FIXME: It might be appropriate to extend this code to handle
4235 // mismatched types.
4239 // Record this extraction against the appropriate vector if possible...
4240 SDValue SourceVec = V.getOperand(0);
4241 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4242 bool FoundSource = false;
4243 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4244 if (SourceVecs[j] == SourceVec) {
4245 if (MinElts[j] > EltNo)
4247 if (MaxElts[j] < EltNo)
4254 // Or record a new source if not...
4256 SourceVecs.push_back(SourceVec);
4257 MinElts.push_back(EltNo);
4258 MaxElts.push_back(EltNo);
4262 // Currently only do something sane when at most two source vectors
4264 if (SourceVecs.size() > 2)
4267 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4268 int VEXTOffsets[2] = {0, 0};
4270 // This loop extracts the usage patterns of the source vectors
4271 // and prepares appropriate SDValues for a shuffle if possible.
4272 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4273 if (SourceVecs[i].getValueType() == VT) {
4274 // No VEXT necessary
4275 ShuffleSrcs[i] = SourceVecs[i];
4278 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4279 // It probably isn't worth padding out a smaller vector just to
4280 // break it down again in a shuffle.
4284 // Since only 64-bit and 128-bit vectors are legal on ARM and
4285 // we've eliminated the other cases...
4286 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4287 "unexpected vector sizes in ReconstructShuffle");
4289 if (MaxElts[i] - MinElts[i] >= NumElts) {
4290 // Span too large for a VEXT to cope
4294 if (MinElts[i] >= NumElts) {
4295 // The extraction can just take the second half
4296 VEXTOffsets[i] = NumElts;
4297 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4299 DAG.getIntPtrConstant(NumElts));
4300 } else if (MaxElts[i] < NumElts) {
4301 // The extraction can just take the first half
4303 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4305 DAG.getIntPtrConstant(0));
4307 // An actual VEXT is needed
4308 VEXTOffsets[i] = MinElts[i];
4309 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4311 DAG.getIntPtrConstant(0));
4312 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4314 DAG.getIntPtrConstant(NumElts));
4315 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4316 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4320 SmallVector<int, 8> Mask;
4322 for (unsigned i = 0; i < NumElts; ++i) {
4323 SDValue Entry = Op.getOperand(i);
4324 if (Entry.getOpcode() == ISD::UNDEF) {
4329 SDValue ExtractVec = Entry.getOperand(0);
4330 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4331 .getOperand(1))->getSExtValue();
4332 if (ExtractVec == SourceVecs[0]) {
4333 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4335 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4339 // Final check before we try to produce nonsense...
4340 if (isShuffleMaskLegal(Mask, VT))
4341 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4347 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4348 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4349 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4350 /// are assumed to be legal.
4352 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4354 if (VT.getVectorNumElements() == 4 &&
4355 (VT.is128BitVector() || VT.is64BitVector())) {
4356 unsigned PFIndexes[4];
4357 for (unsigned i = 0; i != 4; ++i) {
4361 PFIndexes[i] = M[i];
4364 // Compute the index in the perfect shuffle table.
4365 unsigned PFTableIndex =
4366 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4367 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4368 unsigned Cost = (PFEntry >> 30);
4375 unsigned Imm, WhichResult;
4377 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4378 return (EltSize >= 32 ||
4379 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4380 isVREVMask(M, VT, 64) ||
4381 isVREVMask(M, VT, 32) ||
4382 isVREVMask(M, VT, 16) ||
4383 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4384 isVTBLMask(M, VT) ||
4385 isVTRNMask(M, VT, WhichResult) ||
4386 isVUZPMask(M, VT, WhichResult) ||
4387 isVZIPMask(M, VT, WhichResult) ||
4388 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4389 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4390 isVZIP_v_undef_Mask(M, VT, WhichResult));
4393 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4394 /// the specified operations to build the shuffle.
4395 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4396 SDValue RHS, SelectionDAG &DAG,
4398 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4399 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4400 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4403 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4412 OP_VUZPL, // VUZP, left result
4413 OP_VUZPR, // VUZP, right result
4414 OP_VZIPL, // VZIP, left result
4415 OP_VZIPR, // VZIP, right result
4416 OP_VTRNL, // VTRN, left result
4417 OP_VTRNR // VTRN, right result
4420 if (OpNum == OP_COPY) {
4421 if (LHSID == (1*9+2)*9+3) return LHS;
4422 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4426 SDValue OpLHS, OpRHS;
4427 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4428 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4429 EVT VT = OpLHS.getValueType();
4432 default: llvm_unreachable("Unknown shuffle opcode!");
4434 // VREV divides the vector in half and swaps within the half.
4435 if (VT.getVectorElementType() == MVT::i32 ||
4436 VT.getVectorElementType() == MVT::f32)
4437 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4438 // vrev <4 x i16> -> VREV32
4439 if (VT.getVectorElementType() == MVT::i16)
4440 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4441 // vrev <4 x i8> -> VREV16
4442 assert(VT.getVectorElementType() == MVT::i8);
4443 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4448 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4449 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4453 return DAG.getNode(ARMISD::VEXT, dl, VT,
4455 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4458 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4459 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4462 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4463 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4466 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4467 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4471 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4472 ArrayRef<int> ShuffleMask,
4473 SelectionDAG &DAG) {
4474 // Check to see if we can use the VTBL instruction.
4475 SDValue V1 = Op.getOperand(0);
4476 SDValue V2 = Op.getOperand(1);
4477 DebugLoc DL = Op.getDebugLoc();
4479 SmallVector<SDValue, 8> VTBLMask;
4480 for (ArrayRef<int>::iterator
4481 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4482 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4484 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4485 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4486 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4489 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4490 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4494 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4495 SDValue V1 = Op.getOperand(0);
4496 SDValue V2 = Op.getOperand(1);
4497 DebugLoc dl = Op.getDebugLoc();
4498 EVT VT = Op.getValueType();
4499 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4501 // Convert shuffles that are directly supported on NEON to target-specific
4502 // DAG nodes, instead of keeping them as shuffles and matching them again
4503 // during code selection. This is more efficient and avoids the possibility
4504 // of inconsistencies between legalization and selection.
4505 // FIXME: floating-point vectors should be canonicalized to integer vectors
4506 // of the same time so that they get CSEd properly.
4507 ArrayRef<int> ShuffleMask = SVN->getMask();
4509 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4510 if (EltSize <= 32) {
4511 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4512 int Lane = SVN->getSplatIndex();
4513 // If this is undef splat, generate it via "just" vdup, if possible.
4514 if (Lane == -1) Lane = 0;
4516 // Test if V1 is a SCALAR_TO_VECTOR.
4517 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4518 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4520 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4521 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4523 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4524 !isa<ConstantSDNode>(V1.getOperand(0))) {
4525 bool IsScalarToVector = true;
4526 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4527 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4528 IsScalarToVector = false;
4531 if (IsScalarToVector)
4532 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4534 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4535 DAG.getConstant(Lane, MVT::i32));
4540 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4543 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4544 DAG.getConstant(Imm, MVT::i32));
4547 if (isVREVMask(ShuffleMask, VT, 64))
4548 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4549 if (isVREVMask(ShuffleMask, VT, 32))
4550 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4551 if (isVREVMask(ShuffleMask, VT, 16))
4552 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4554 // Check for Neon shuffles that modify both input vectors in place.
4555 // If both results are used, i.e., if there are two shuffles with the same
4556 // source operands and with masks corresponding to both results of one of
4557 // these operations, DAG memoization will ensure that a single node is
4558 // used for both shuffles.
4559 unsigned WhichResult;
4560 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4561 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4562 V1, V2).getValue(WhichResult);
4563 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4564 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4565 V1, V2).getValue(WhichResult);
4566 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4567 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4568 V1, V2).getValue(WhichResult);
4570 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4571 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4572 V1, V1).getValue(WhichResult);
4573 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4574 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4575 V1, V1).getValue(WhichResult);
4576 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4577 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4578 V1, V1).getValue(WhichResult);
4581 // If the shuffle is not directly supported and it has 4 elements, use
4582 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4583 unsigned NumElts = VT.getVectorNumElements();
4585 unsigned PFIndexes[4];
4586 for (unsigned i = 0; i != 4; ++i) {
4587 if (ShuffleMask[i] < 0)
4590 PFIndexes[i] = ShuffleMask[i];
4593 // Compute the index in the perfect shuffle table.
4594 unsigned PFTableIndex =
4595 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4596 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4597 unsigned Cost = (PFEntry >> 30);
4600 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4603 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4604 if (EltSize >= 32) {
4605 // Do the expansion with floating-point types, since that is what the VFP
4606 // registers are defined to use, and since i64 is not legal.
4607 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4608 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4609 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4610 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4611 SmallVector<SDValue, 8> Ops;
4612 for (unsigned i = 0; i < NumElts; ++i) {
4613 if (ShuffleMask[i] < 0)
4614 Ops.push_back(DAG.getUNDEF(EltVT));
4616 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4617 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4618 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4621 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4622 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4625 if (VT == MVT::v8i8) {
4626 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4627 if (NewOp.getNode())
4634 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4635 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4636 SDValue Lane = Op.getOperand(2);
4637 if (!isa<ConstantSDNode>(Lane))
4643 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4644 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4645 SDValue Lane = Op.getOperand(1);
4646 if (!isa<ConstantSDNode>(Lane))
4649 SDValue Vec = Op.getOperand(0);
4650 if (Op.getValueType() == MVT::i32 &&
4651 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4652 DebugLoc dl = Op.getDebugLoc();
4653 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4659 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4660 // The only time a CONCAT_VECTORS operation can have legal types is when
4661 // two 64-bit vectors are concatenated to a 128-bit vector.
4662 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4663 "unexpected CONCAT_VECTORS");
4664 DebugLoc dl = Op.getDebugLoc();
4665 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4666 SDValue Op0 = Op.getOperand(0);
4667 SDValue Op1 = Op.getOperand(1);
4668 if (Op0.getOpcode() != ISD::UNDEF)
4669 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4670 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4671 DAG.getIntPtrConstant(0));
4672 if (Op1.getOpcode() != ISD::UNDEF)
4673 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4674 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4675 DAG.getIntPtrConstant(1));
4676 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4679 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4680 /// element has been zero/sign-extended, depending on the isSigned parameter,
4681 /// from an integer type half its size.
4682 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4684 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4685 EVT VT = N->getValueType(0);
4686 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4687 SDNode *BVN = N->getOperand(0).getNode();
4688 if (BVN->getValueType(0) != MVT::v4i32 ||
4689 BVN->getOpcode() != ISD::BUILD_VECTOR)
4691 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4692 unsigned HiElt = 1 - LoElt;
4693 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4694 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4695 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4696 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4697 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4700 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4701 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4704 if (Hi0->isNullValue() && Hi1->isNullValue())
4710 if (N->getOpcode() != ISD::BUILD_VECTOR)
4713 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4714 SDNode *Elt = N->getOperand(i).getNode();
4715 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4716 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4717 unsigned HalfSize = EltSize / 2;
4719 if (!isIntN(HalfSize, C->getSExtValue()))
4722 if (!isUIntN(HalfSize, C->getZExtValue()))
4733 /// isSignExtended - Check if a node is a vector value that is sign-extended
4734 /// or a constant BUILD_VECTOR with sign-extended elements.
4735 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4736 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4738 if (isExtendedBUILD_VECTOR(N, DAG, true))
4743 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4744 /// or a constant BUILD_VECTOR with zero-extended elements.
4745 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4746 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4748 if (isExtendedBUILD_VECTOR(N, DAG, false))
4753 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4754 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4755 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4756 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4757 return N->getOperand(0);
4758 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4759 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4760 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4761 LD->isNonTemporal(), LD->isInvariant(),
4762 LD->getAlignment());
4763 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4764 // have been legalized as a BITCAST from v4i32.
4765 if (N->getOpcode() == ISD::BITCAST) {
4766 SDNode *BVN = N->getOperand(0).getNode();
4767 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4768 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4769 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4770 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4771 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4773 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4774 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4775 EVT VT = N->getValueType(0);
4776 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4777 unsigned NumElts = VT.getVectorNumElements();
4778 MVT TruncVT = MVT::getIntegerVT(EltSize);
4779 SmallVector<SDValue, 8> Ops;
4780 for (unsigned i = 0; i != NumElts; ++i) {
4781 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4782 const APInt &CInt = C->getAPIntValue();
4783 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4785 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4786 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4789 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4790 unsigned Opcode = N->getOpcode();
4791 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4792 SDNode *N0 = N->getOperand(0).getNode();
4793 SDNode *N1 = N->getOperand(1).getNode();
4794 return N0->hasOneUse() && N1->hasOneUse() &&
4795 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4800 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4801 unsigned Opcode = N->getOpcode();
4802 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4803 SDNode *N0 = N->getOperand(0).getNode();
4804 SDNode *N1 = N->getOperand(1).getNode();
4805 return N0->hasOneUse() && N1->hasOneUse() &&
4806 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4811 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4812 // Multiplications are only custom-lowered for 128-bit vectors so that
4813 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4814 EVT VT = Op.getValueType();
4815 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4816 SDNode *N0 = Op.getOperand(0).getNode();
4817 SDNode *N1 = Op.getOperand(1).getNode();
4818 unsigned NewOpc = 0;
4820 bool isN0SExt = isSignExtended(N0, DAG);
4821 bool isN1SExt = isSignExtended(N1, DAG);
4822 if (isN0SExt && isN1SExt)
4823 NewOpc = ARMISD::VMULLs;
4825 bool isN0ZExt = isZeroExtended(N0, DAG);
4826 bool isN1ZExt = isZeroExtended(N1, DAG);
4827 if (isN0ZExt && isN1ZExt)
4828 NewOpc = ARMISD::VMULLu;
4829 else if (isN1SExt || isN1ZExt) {
4830 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4831 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4832 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4833 NewOpc = ARMISD::VMULLs;
4835 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4836 NewOpc = ARMISD::VMULLu;
4838 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4840 NewOpc = ARMISD::VMULLu;
4846 if (VT == MVT::v2i64)
4847 // Fall through to expand this. It is not legal.
4850 // Other vector multiplications are legal.
4855 // Legalize to a VMULL instruction.
4856 DebugLoc DL = Op.getDebugLoc();
4858 SDValue Op1 = SkipExtension(N1, DAG);
4860 Op0 = SkipExtension(N0, DAG);
4861 assert(Op0.getValueType().is64BitVector() &&
4862 Op1.getValueType().is64BitVector() &&
4863 "unexpected types for extended operands to VMULL");
4864 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4867 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4868 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4875 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4876 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4877 EVT Op1VT = Op1.getValueType();
4878 return DAG.getNode(N0->getOpcode(), DL, VT,
4879 DAG.getNode(NewOpc, DL, VT,
4880 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4881 DAG.getNode(NewOpc, DL, VT,
4882 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4886 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4888 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4889 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4890 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4891 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4892 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4893 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4894 // Get reciprocal estimate.
4895 // float4 recip = vrecpeq_f32(yf);
4896 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4897 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4898 // Because char has a smaller range than uchar, we can actually get away
4899 // without any newton steps. This requires that we use a weird bias
4900 // of 0xb000, however (again, this has been exhaustively tested).
4901 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4902 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4903 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4904 Y = DAG.getConstant(0xb000, MVT::i32);
4905 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4906 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4907 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4908 // Convert back to short.
4909 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4910 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4915 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4917 // Convert to float.
4918 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4919 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4920 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4921 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4922 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4923 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4925 // Use reciprocal estimate and one refinement step.
4926 // float4 recip = vrecpeq_f32(yf);
4927 // recip *= vrecpsq_f32(yf, recip);
4928 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4929 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4930 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4931 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4933 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4934 // Because short has a smaller range than ushort, we can actually get away
4935 // with only a single newton step. This requires that we use a weird bias
4936 // of 89, however (again, this has been exhaustively tested).
4937 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4938 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4939 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4940 N1 = DAG.getConstant(0x89, MVT::i32);
4941 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4942 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4943 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4944 // Convert back to integer and return.
4945 // return vmovn_s32(vcvt_s32_f32(result));
4946 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4947 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4951 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4952 EVT VT = Op.getValueType();
4953 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4954 "unexpected type for custom-lowering ISD::SDIV");
4956 DebugLoc dl = Op.getDebugLoc();
4957 SDValue N0 = Op.getOperand(0);
4958 SDValue N1 = Op.getOperand(1);
4961 if (VT == MVT::v8i8) {
4962 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4963 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4965 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4966 DAG.getIntPtrConstant(4));
4967 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4968 DAG.getIntPtrConstant(4));
4969 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4970 DAG.getIntPtrConstant(0));
4971 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4972 DAG.getIntPtrConstant(0));
4974 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4975 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4977 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4978 N0 = LowerCONCAT_VECTORS(N0, DAG);
4980 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4983 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4986 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4987 EVT VT = Op.getValueType();
4988 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4989 "unexpected type for custom-lowering ISD::UDIV");
4991 DebugLoc dl = Op.getDebugLoc();
4992 SDValue N0 = Op.getOperand(0);
4993 SDValue N1 = Op.getOperand(1);
4996 if (VT == MVT::v8i8) {
4997 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4998 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5000 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5001 DAG.getIntPtrConstant(4));
5002 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5003 DAG.getIntPtrConstant(4));
5004 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5005 DAG.getIntPtrConstant(0));
5006 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5007 DAG.getIntPtrConstant(0));
5009 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5010 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5012 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5013 N0 = LowerCONCAT_VECTORS(N0, DAG);
5015 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5016 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5021 // v4i16 sdiv ... Convert to float.
5022 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5023 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5024 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5025 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5026 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5027 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5029 // Use reciprocal estimate and two refinement steps.
5030 // float4 recip = vrecpeq_f32(yf);
5031 // recip *= vrecpsq_f32(yf, recip);
5032 // recip *= vrecpsq_f32(yf, recip);
5033 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5034 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5035 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5036 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5038 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5039 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5040 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5042 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5043 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5044 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5045 // and that it will never cause us to return an answer too large).
5046 // float4 result = as_float4(as_int4(xf*recip) + 2);
5047 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5048 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5049 N1 = DAG.getConstant(2, MVT::i32);
5050 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5051 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5052 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5053 // Convert back to integer and return.
5054 // return vmovn_u32(vcvt_s32_f32(result));
5055 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5056 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5060 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5061 EVT VT = Op.getNode()->getValueType(0);
5062 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5065 bool ExtraOp = false;
5066 switch (Op.getOpcode()) {
5067 default: llvm_unreachable("Invalid code");
5068 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5069 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5070 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5071 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5075 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5077 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5078 Op.getOperand(1), Op.getOperand(2));
5081 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5082 // Monotonic load/store is legal for all targets
5083 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5086 // Aquire/Release load/store is not legal for targets without a
5087 // dmb or equivalent available.
5093 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5094 SelectionDAG &DAG, unsigned NewOp) {
5095 DebugLoc dl = Node->getDebugLoc();
5096 assert (Node->getValueType(0) == MVT::i64 &&
5097 "Only know how to expand i64 atomics");
5099 SmallVector<SDValue, 6> Ops;
5100 Ops.push_back(Node->getOperand(0)); // Chain
5101 Ops.push_back(Node->getOperand(1)); // Ptr
5103 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5104 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5105 // High part of Val1
5106 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5107 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5108 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5109 // High part of Val1
5110 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5111 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5112 // High part of Val2
5113 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5114 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5116 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5118 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5119 cast<MemSDNode>(Node)->getMemOperand());
5120 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5121 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5122 Results.push_back(Result.getValue(2));
5125 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5126 switch (Op.getOpcode()) {
5127 default: llvm_unreachable("Don't know how to custom lower this!");
5128 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5129 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5130 case ISD::GlobalAddress:
5131 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5132 LowerGlobalAddressELF(Op, DAG);
5133 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5134 case ISD::SELECT: return LowerSELECT(Op, DAG);
5135 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5136 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5137 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5138 case ISD::VASTART: return LowerVASTART(Op, DAG);
5139 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5140 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5141 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5142 case ISD::SINT_TO_FP:
5143 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5144 case ISD::FP_TO_SINT:
5145 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5146 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5147 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5148 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5149 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5150 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5151 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5152 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5154 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5157 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5158 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5159 case ISD::SRL_PARTS:
5160 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5161 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5162 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5163 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5164 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5165 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5166 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5167 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5168 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5169 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5170 case ISD::MUL: return LowerMUL(Op, DAG);
5171 case ISD::SDIV: return LowerSDIV(Op, DAG);
5172 case ISD::UDIV: return LowerUDIV(Op, DAG);
5176 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5177 case ISD::ATOMIC_LOAD:
5178 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5182 /// ReplaceNodeResults - Replace the results of node with an illegal result
5183 /// type with new values built out of custom code.
5184 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5185 SmallVectorImpl<SDValue>&Results,
5186 SelectionDAG &DAG) const {
5188 switch (N->getOpcode()) {
5190 llvm_unreachable("Don't know how to custom expand this!");
5192 Res = ExpandBITCAST(N, DAG);
5196 Res = Expand64BitShift(N, DAG, Subtarget);
5198 case ISD::ATOMIC_LOAD_ADD:
5199 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5201 case ISD::ATOMIC_LOAD_AND:
5202 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5204 case ISD::ATOMIC_LOAD_NAND:
5205 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5207 case ISD::ATOMIC_LOAD_OR:
5208 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5210 case ISD::ATOMIC_LOAD_SUB:
5211 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5213 case ISD::ATOMIC_LOAD_XOR:
5214 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5216 case ISD::ATOMIC_SWAP:
5217 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5219 case ISD::ATOMIC_CMP_SWAP:
5220 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5224 Results.push_back(Res);
5227 //===----------------------------------------------------------------------===//
5228 // ARM Scheduler Hooks
5229 //===----------------------------------------------------------------------===//
5232 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5233 MachineBasicBlock *BB,
5234 unsigned Size) const {
5235 unsigned dest = MI->getOperand(0).getReg();
5236 unsigned ptr = MI->getOperand(1).getReg();
5237 unsigned oldval = MI->getOperand(2).getReg();
5238 unsigned newval = MI->getOperand(3).getReg();
5239 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5240 DebugLoc dl = MI->getDebugLoc();
5241 bool isThumb2 = Subtarget->isThumb2();
5243 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5245 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5246 : ARM::GPRRegisterClass);
5249 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5250 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5251 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5254 unsigned ldrOpc, strOpc;
5256 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5258 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5259 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5262 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5263 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5266 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5267 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5271 MachineFunction *MF = BB->getParent();
5272 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5273 MachineFunction::iterator It = BB;
5274 ++It; // insert the new blocks after the current block
5276 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5277 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5278 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5279 MF->insert(It, loop1MBB);
5280 MF->insert(It, loop2MBB);
5281 MF->insert(It, exitMBB);
5283 // Transfer the remainder of BB and its successor edges to exitMBB.
5284 exitMBB->splice(exitMBB->begin(), BB,
5285 llvm::next(MachineBasicBlock::iterator(MI)),
5287 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5291 // fallthrough --> loop1MBB
5292 BB->addSuccessor(loop1MBB);
5295 // ldrex dest, [ptr]
5299 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5300 if (ldrOpc == ARM::t2LDREX)
5302 AddDefaultPred(MIB);
5303 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5304 .addReg(dest).addReg(oldval));
5305 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5306 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5307 BB->addSuccessor(loop2MBB);
5308 BB->addSuccessor(exitMBB);
5311 // strex scratch, newval, [ptr]
5315 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5316 if (strOpc == ARM::t2STREX)
5318 AddDefaultPred(MIB);
5319 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5320 .addReg(scratch).addImm(0));
5321 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5322 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5323 BB->addSuccessor(loop1MBB);
5324 BB->addSuccessor(exitMBB);
5330 MI->eraseFromParent(); // The instruction is gone now.
5336 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5337 unsigned Size, unsigned BinOpcode) const {
5338 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5339 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5341 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5342 MachineFunction *MF = BB->getParent();
5343 MachineFunction::iterator It = BB;
5346 unsigned dest = MI->getOperand(0).getReg();
5347 unsigned ptr = MI->getOperand(1).getReg();
5348 unsigned incr = MI->getOperand(2).getReg();
5349 DebugLoc dl = MI->getDebugLoc();
5350 bool isThumb2 = Subtarget->isThumb2();
5352 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5354 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5355 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5358 unsigned ldrOpc, strOpc;
5360 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5362 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5363 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5366 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5367 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5370 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5371 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5375 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5376 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5377 MF->insert(It, loopMBB);
5378 MF->insert(It, exitMBB);
5380 // Transfer the remainder of BB and its successor edges to exitMBB.
5381 exitMBB->splice(exitMBB->begin(), BB,
5382 llvm::next(MachineBasicBlock::iterator(MI)),
5384 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5386 const TargetRegisterClass *TRC =
5387 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5388 unsigned scratch = MRI.createVirtualRegister(TRC);
5389 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5393 // fallthrough --> loopMBB
5394 BB->addSuccessor(loopMBB);
5398 // <binop> scratch2, dest, incr
5399 // strex scratch, scratch2, ptr
5402 // fallthrough --> exitMBB
5404 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5405 if (ldrOpc == ARM::t2LDREX)
5407 AddDefaultPred(MIB);
5409 // operand order needs to go the other way for NAND
5410 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5411 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5412 addReg(incr).addReg(dest)).addReg(0);
5414 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5415 addReg(dest).addReg(incr)).addReg(0);
5418 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5419 if (strOpc == ARM::t2STREX)
5421 AddDefaultPred(MIB);
5422 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5423 .addReg(scratch).addImm(0));
5424 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5425 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5427 BB->addSuccessor(loopMBB);
5428 BB->addSuccessor(exitMBB);
5434 MI->eraseFromParent(); // The instruction is gone now.
5440 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5441 MachineBasicBlock *BB,
5444 ARMCC::CondCodes Cond) const {
5445 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5447 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5448 MachineFunction *MF = BB->getParent();
5449 MachineFunction::iterator It = BB;
5452 unsigned dest = MI->getOperand(0).getReg();
5453 unsigned ptr = MI->getOperand(1).getReg();
5454 unsigned incr = MI->getOperand(2).getReg();
5455 unsigned oldval = dest;
5456 DebugLoc dl = MI->getDebugLoc();
5457 bool isThumb2 = Subtarget->isThumb2();
5459 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5461 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5462 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5465 unsigned ldrOpc, strOpc, extendOpc;
5467 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5469 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5470 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5471 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5474 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5475 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5476 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5479 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5480 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5485 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5486 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5487 MF->insert(It, loopMBB);
5488 MF->insert(It, exitMBB);
5490 // Transfer the remainder of BB and its successor edges to exitMBB.
5491 exitMBB->splice(exitMBB->begin(), BB,
5492 llvm::next(MachineBasicBlock::iterator(MI)),
5494 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5496 const TargetRegisterClass *TRC =
5497 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5498 unsigned scratch = MRI.createVirtualRegister(TRC);
5499 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5503 // fallthrough --> loopMBB
5504 BB->addSuccessor(loopMBB);
5508 // (sign extend dest, if required)
5510 // cmov.cond scratch2, dest, incr
5511 // strex scratch, scratch2, ptr
5514 // fallthrough --> exitMBB
5516 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5517 if (ldrOpc == ARM::t2LDREX)
5519 AddDefaultPred(MIB);
5521 // Sign extend the value, if necessary.
5522 if (signExtend && extendOpc) {
5523 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5524 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5529 // Build compare and cmov instructions.
5530 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5531 .addReg(oldval).addReg(incr));
5532 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5533 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5535 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5536 if (strOpc == ARM::t2STREX)
5538 AddDefaultPred(MIB);
5539 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5540 .addReg(scratch).addImm(0));
5541 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5542 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5544 BB->addSuccessor(loopMBB);
5545 BB->addSuccessor(exitMBB);
5551 MI->eraseFromParent(); // The instruction is gone now.
5557 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5558 unsigned Op1, unsigned Op2,
5559 bool NeedsCarry, bool IsCmpxchg) const {
5560 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5561 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5563 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5564 MachineFunction *MF = BB->getParent();
5565 MachineFunction::iterator It = BB;
5568 unsigned destlo = MI->getOperand(0).getReg();
5569 unsigned desthi = MI->getOperand(1).getReg();
5570 unsigned ptr = MI->getOperand(2).getReg();
5571 unsigned vallo = MI->getOperand(3).getReg();
5572 unsigned valhi = MI->getOperand(4).getReg();
5573 DebugLoc dl = MI->getDebugLoc();
5574 bool isThumb2 = Subtarget->isThumb2();
5576 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5578 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5579 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5580 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5583 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5584 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5586 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5587 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5589 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5590 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5592 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5593 MF->insert(It, loopMBB);
5595 MF->insert(It, contBB);
5596 MF->insert(It, cont2BB);
5598 MF->insert(It, exitMBB);
5600 // Transfer the remainder of BB and its successor edges to exitMBB.
5601 exitMBB->splice(exitMBB->begin(), BB,
5602 llvm::next(MachineBasicBlock::iterator(MI)),
5604 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5606 const TargetRegisterClass *TRC =
5607 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5608 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5612 // fallthrough --> loopMBB
5613 BB->addSuccessor(loopMBB);
5616 // ldrexd r2, r3, ptr
5617 // <binopa> r0, r2, incr
5618 // <binopb> r1, r3, incr
5619 // strexd storesuccess, r0, r1, ptr
5620 // cmp storesuccess, #0
5622 // fallthrough --> exitMBB
5624 // Note that the registers are explicitly specified because there is not any
5625 // way to force the register allocator to allocate a register pair.
5627 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5628 // need to properly enforce the restriction that the two output registers
5629 // for ldrexd must be different.
5632 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5633 .addReg(ARM::R2, RegState::Define)
5634 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5635 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5636 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5637 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5641 for (unsigned i = 0; i < 2; i++) {
5642 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5644 .addReg(i == 0 ? destlo : desthi)
5645 .addReg(i == 0 ? vallo : valhi));
5646 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5647 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5648 BB->addSuccessor(exitMBB);
5649 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5650 BB = (i == 0 ? contBB : cont2BB);
5653 // Copy to physregs for strexd
5654 unsigned setlo = MI->getOperand(5).getReg();
5655 unsigned sethi = MI->getOperand(6).getReg();
5656 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5657 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5659 // Perform binary operation
5660 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5661 .addReg(destlo).addReg(vallo))
5662 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5663 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5664 .addReg(desthi).addReg(valhi)).addReg(0);
5666 // Copy to physregs for strexd
5667 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5668 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5672 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5673 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5675 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5676 .addReg(storesuccess).addImm(0));
5677 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5678 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5680 BB->addSuccessor(loopMBB);
5681 BB->addSuccessor(exitMBB);
5687 MI->eraseFromParent(); // The instruction is gone now.
5692 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5693 /// registers the function context.
5694 void ARMTargetLowering::
5695 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5696 MachineBasicBlock *DispatchBB, int FI) const {
5697 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5698 DebugLoc dl = MI->getDebugLoc();
5699 MachineFunction *MF = MBB->getParent();
5700 MachineRegisterInfo *MRI = &MF->getRegInfo();
5701 MachineConstantPool *MCP = MF->getConstantPool();
5702 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5703 const Function *F = MF->getFunction();
5705 bool isThumb = Subtarget->isThumb();
5706 bool isThumb2 = Subtarget->isThumb2();
5708 unsigned PCLabelId = AFI->createPICLabelUId();
5709 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5710 ARMConstantPoolValue *CPV =
5711 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5712 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5714 const TargetRegisterClass *TRC =
5715 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5717 // Grab constant pool and fixed stack memory operands.
5718 MachineMemOperand *CPMMO =
5719 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5720 MachineMemOperand::MOLoad, 4, 4);
5722 MachineMemOperand *FIMMOSt =
5723 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5724 MachineMemOperand::MOStore, 4, 4);
5726 // Load the address of the dispatch MBB into the jump buffer.
5728 // Incoming value: jbuf
5729 // ldr.n r5, LCPI1_1
5732 // str r5, [$jbuf, #+4] ; &jbuf[1]
5733 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5734 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5735 .addConstantPoolIndex(CPI)
5736 .addMemOperand(CPMMO));
5737 // Set the low bit because of thumb mode.
5738 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5740 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5741 .addReg(NewVReg1, RegState::Kill)
5743 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5744 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5745 .addReg(NewVReg2, RegState::Kill)
5747 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5748 .addReg(NewVReg3, RegState::Kill)
5750 .addImm(36) // &jbuf[1] :: pc
5751 .addMemOperand(FIMMOSt));
5752 } else if (isThumb) {
5753 // Incoming value: jbuf
5754 // ldr.n r1, LCPI1_4
5758 // add r2, $jbuf, #+4 ; &jbuf[1]
5760 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5761 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5762 .addConstantPoolIndex(CPI)
5763 .addMemOperand(CPMMO));
5764 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5765 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5766 .addReg(NewVReg1, RegState::Kill)
5768 // Set the low bit because of thumb mode.
5769 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5770 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5771 .addReg(ARM::CPSR, RegState::Define)
5773 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5774 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5775 .addReg(ARM::CPSR, RegState::Define)
5776 .addReg(NewVReg2, RegState::Kill)
5777 .addReg(NewVReg3, RegState::Kill));
5778 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5779 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5781 .addImm(36)); // &jbuf[1] :: pc
5782 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5783 .addReg(NewVReg4, RegState::Kill)
5784 .addReg(NewVReg5, RegState::Kill)
5786 .addMemOperand(FIMMOSt));
5788 // Incoming value: jbuf
5791 // str r1, [$jbuf, #+4] ; &jbuf[1]
5792 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5793 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5794 .addConstantPoolIndex(CPI)
5796 .addMemOperand(CPMMO));
5797 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5798 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5799 .addReg(NewVReg1, RegState::Kill)
5800 .addImm(PCLabelId));
5801 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5802 .addReg(NewVReg2, RegState::Kill)
5804 .addImm(36) // &jbuf[1] :: pc
5805 .addMemOperand(FIMMOSt));
5809 MachineBasicBlock *ARMTargetLowering::
5810 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5811 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5812 DebugLoc dl = MI->getDebugLoc();
5813 MachineFunction *MF = MBB->getParent();
5814 MachineRegisterInfo *MRI = &MF->getRegInfo();
5815 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5816 MachineFrameInfo *MFI = MF->getFrameInfo();
5817 int FI = MFI->getFunctionContextIndex();
5819 const TargetRegisterClass *TRC =
5820 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5822 // Get a mapping of the call site numbers to all of the landing pads they're
5824 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5825 unsigned MaxCSNum = 0;
5826 MachineModuleInfo &MMI = MF->getMMI();
5827 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5829 if (!BB->isLandingPad()) continue;
5831 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5833 for (MachineBasicBlock::iterator
5834 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5835 if (!II->isEHLabel()) continue;
5837 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5838 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5840 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5841 for (SmallVectorImpl<unsigned>::iterator
5842 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5843 CSI != CSE; ++CSI) {
5844 CallSiteNumToLPad[*CSI].push_back(BB);
5845 MaxCSNum = std::max(MaxCSNum, *CSI);
5851 // Get an ordered list of the machine basic blocks for the jump table.
5852 std::vector<MachineBasicBlock*> LPadList;
5853 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5854 LPadList.reserve(CallSiteNumToLPad.size());
5855 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5856 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5857 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5858 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5859 LPadList.push_back(*II);
5860 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5864 assert(!LPadList.empty() &&
5865 "No landing pad destinations for the dispatch jump table!");
5867 // Create the jump table and associated information.
5868 MachineJumpTableInfo *JTI =
5869 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5870 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5871 unsigned UId = AFI->createJumpTableUId();
5873 // Create the MBBs for the dispatch code.
5875 // Shove the dispatch's address into the return slot in the function context.
5876 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5877 DispatchBB->setIsLandingPad();
5879 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5880 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5881 DispatchBB->addSuccessor(TrapBB);
5883 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5884 DispatchBB->addSuccessor(DispContBB);
5887 MF->insert(MF->end(), DispatchBB);
5888 MF->insert(MF->end(), DispContBB);
5889 MF->insert(MF->end(), TrapBB);
5891 // Insert code into the entry block that creates and registers the function
5893 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5895 MachineMemOperand *FIMMOLd =
5896 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5897 MachineMemOperand::MOLoad |
5898 MachineMemOperand::MOVolatile, 4, 4);
5900 if (AFI->isThumb1OnlyFunction())
5901 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5902 else if (!Subtarget->hasVFP2())
5903 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5905 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5907 unsigned NumLPads = LPadList.size();
5908 if (Subtarget->isThumb2()) {
5909 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5910 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5913 .addMemOperand(FIMMOLd));
5915 if (NumLPads < 256) {
5916 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5918 .addImm(LPadList.size()));
5920 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5921 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5922 .addImm(NumLPads & 0xFFFF));
5924 unsigned VReg2 = VReg1;
5925 if ((NumLPads & 0xFFFF0000) != 0) {
5926 VReg2 = MRI->createVirtualRegister(TRC);
5927 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5929 .addImm(NumLPads >> 16));
5932 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5937 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5942 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5943 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5944 .addJumpTableIndex(MJTI)
5947 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5950 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5951 .addReg(NewVReg3, RegState::Kill)
5953 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5955 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5956 .addReg(NewVReg4, RegState::Kill)
5958 .addJumpTableIndex(MJTI)
5960 } else if (Subtarget->isThumb()) {
5961 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5962 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5965 .addMemOperand(FIMMOLd));
5967 if (NumLPads < 256) {
5968 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5972 MachineConstantPool *ConstantPool = MF->getConstantPool();
5973 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5974 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5976 // MachineConstantPool wants an explicit alignment.
5977 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5979 Align = getTargetData()->getTypeAllocSize(C->getType());
5980 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5982 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5983 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5984 .addReg(VReg1, RegState::Define)
5985 .addConstantPoolIndex(Idx));
5986 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5991 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5996 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5997 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5998 .addReg(ARM::CPSR, RegState::Define)
6002 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6003 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6004 .addJumpTableIndex(MJTI)
6007 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6008 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6009 .addReg(ARM::CPSR, RegState::Define)
6010 .addReg(NewVReg2, RegState::Kill)
6013 MachineMemOperand *JTMMOLd =
6014 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6015 MachineMemOperand::MOLoad, 4, 4);
6017 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6018 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6019 .addReg(NewVReg4, RegState::Kill)
6021 .addMemOperand(JTMMOLd));
6023 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6024 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6025 .addReg(ARM::CPSR, RegState::Define)
6026 .addReg(NewVReg5, RegState::Kill)
6029 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6030 .addReg(NewVReg6, RegState::Kill)
6031 .addJumpTableIndex(MJTI)
6034 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6035 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6038 .addMemOperand(FIMMOLd));
6040 if (NumLPads < 256) {
6041 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6044 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6045 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6046 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6047 .addImm(NumLPads & 0xFFFF));
6049 unsigned VReg2 = VReg1;
6050 if ((NumLPads & 0xFFFF0000) != 0) {
6051 VReg2 = MRI->createVirtualRegister(TRC);
6052 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6054 .addImm(NumLPads >> 16));
6057 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6061 MachineConstantPool *ConstantPool = MF->getConstantPool();
6062 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6063 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6065 // MachineConstantPool wants an explicit alignment.
6066 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6068 Align = getTargetData()->getTypeAllocSize(C->getType());
6069 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6071 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6072 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6073 .addReg(VReg1, RegState::Define)
6074 .addConstantPoolIndex(Idx)
6076 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6078 .addReg(VReg1, RegState::Kill));
6081 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6086 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6088 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6090 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6091 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6092 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6093 .addJumpTableIndex(MJTI)
6096 MachineMemOperand *JTMMOLd =
6097 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6098 MachineMemOperand::MOLoad, 4, 4);
6099 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6101 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6102 .addReg(NewVReg3, RegState::Kill)
6105 .addMemOperand(JTMMOLd));
6107 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6108 .addReg(NewVReg5, RegState::Kill)
6110 .addJumpTableIndex(MJTI)
6114 // Add the jump table entries as successors to the MBB.
6115 MachineBasicBlock *PrevMBB = 0;
6116 for (std::vector<MachineBasicBlock*>::iterator
6117 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6118 MachineBasicBlock *CurMBB = *I;
6119 if (PrevMBB != CurMBB)
6120 DispContBB->addSuccessor(CurMBB);
6124 // N.B. the order the invoke BBs are processed in doesn't matter here.
6125 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6126 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6127 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6128 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6129 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6130 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6131 MachineBasicBlock *BB = *I;
6133 // Remove the landing pad successor from the invoke block and replace it
6134 // with the new dispatch block.
6135 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6137 while (!Successors.empty()) {
6138 MachineBasicBlock *SMBB = Successors.pop_back_val();
6139 if (SMBB->isLandingPad()) {
6140 BB->removeSuccessor(SMBB);
6141 MBBLPads.push_back(SMBB);
6145 BB->addSuccessor(DispatchBB);
6147 // Find the invoke call and mark all of the callee-saved registers as
6148 // 'implicit defined' so that they're spilled. This prevents code from
6149 // moving instructions to before the EH block, where they will never be
6151 for (MachineBasicBlock::reverse_iterator
6152 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6153 if (!II->isCall()) continue;
6155 DenseMap<unsigned, bool> DefRegs;
6156 for (MachineInstr::mop_iterator
6157 OI = II->operands_begin(), OE = II->operands_end();
6159 if (!OI->isReg()) continue;
6160 DefRegs[OI->getReg()] = true;
6163 MachineInstrBuilder MIB(&*II);
6165 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6166 unsigned Reg = SavedRegs[i];
6167 if (Subtarget->isThumb2() &&
6168 !ARM::tGPRRegisterClass->contains(Reg) &&
6169 !ARM::hGPRRegisterClass->contains(Reg))
6171 else if (Subtarget->isThumb1Only() &&
6172 !ARM::tGPRRegisterClass->contains(Reg))
6174 else if (!Subtarget->isThumb() &&
6175 !ARM::GPRRegisterClass->contains(Reg))
6178 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6185 // Mark all former landing pads as non-landing pads. The dispatch is the only
6187 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6188 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6189 (*I)->setIsLandingPad(false);
6191 // The instruction is gone now.
6192 MI->eraseFromParent();
6198 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6199 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6200 E = MBB->succ_end(); I != E; ++I)
6203 llvm_unreachable("Expecting a BB with two successors!");
6207 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6208 MachineBasicBlock *BB) const {
6209 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6210 DebugLoc dl = MI->getDebugLoc();
6211 bool isThumb2 = Subtarget->isThumb2();
6212 switch (MI->getOpcode()) {
6215 llvm_unreachable("Unexpected instr type to insert");
6217 // The Thumb2 pre-indexed stores have the same MI operands, they just
6218 // define them differently in the .td files from the isel patterns, so
6219 // they need pseudos.
6220 case ARM::t2STR_preidx:
6221 MI->setDesc(TII->get(ARM::t2STR_PRE));
6223 case ARM::t2STRB_preidx:
6224 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6226 case ARM::t2STRH_preidx:
6227 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6230 case ARM::STRi_preidx:
6231 case ARM::STRBi_preidx: {
6232 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6233 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6234 // Decode the offset.
6235 unsigned Offset = MI->getOperand(4).getImm();
6236 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6237 Offset = ARM_AM::getAM2Offset(Offset);
6241 MachineMemOperand *MMO = *MI->memoperands_begin();
6242 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6243 .addOperand(MI->getOperand(0)) // Rn_wb
6244 .addOperand(MI->getOperand(1)) // Rt
6245 .addOperand(MI->getOperand(2)) // Rn
6246 .addImm(Offset) // offset (skip GPR==zero_reg)
6247 .addOperand(MI->getOperand(5)) // pred
6248 .addOperand(MI->getOperand(6))
6249 .addMemOperand(MMO);
6250 MI->eraseFromParent();
6253 case ARM::STRr_preidx:
6254 case ARM::STRBr_preidx:
6255 case ARM::STRH_preidx: {
6257 switch (MI->getOpcode()) {
6258 default: llvm_unreachable("unexpected opcode!");
6259 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6260 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6261 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6263 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6264 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6265 MIB.addOperand(MI->getOperand(i));
6266 MI->eraseFromParent();
6269 case ARM::ATOMIC_LOAD_ADD_I8:
6270 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6271 case ARM::ATOMIC_LOAD_ADD_I16:
6272 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6273 case ARM::ATOMIC_LOAD_ADD_I32:
6274 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6276 case ARM::ATOMIC_LOAD_AND_I8:
6277 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6278 case ARM::ATOMIC_LOAD_AND_I16:
6279 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6280 case ARM::ATOMIC_LOAD_AND_I32:
6281 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6283 case ARM::ATOMIC_LOAD_OR_I8:
6284 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6285 case ARM::ATOMIC_LOAD_OR_I16:
6286 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6287 case ARM::ATOMIC_LOAD_OR_I32:
6288 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6290 case ARM::ATOMIC_LOAD_XOR_I8:
6291 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6292 case ARM::ATOMIC_LOAD_XOR_I16:
6293 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6294 case ARM::ATOMIC_LOAD_XOR_I32:
6295 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6297 case ARM::ATOMIC_LOAD_NAND_I8:
6298 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6299 case ARM::ATOMIC_LOAD_NAND_I16:
6300 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6301 case ARM::ATOMIC_LOAD_NAND_I32:
6302 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6304 case ARM::ATOMIC_LOAD_SUB_I8:
6305 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6306 case ARM::ATOMIC_LOAD_SUB_I16:
6307 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6308 case ARM::ATOMIC_LOAD_SUB_I32:
6309 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6311 case ARM::ATOMIC_LOAD_MIN_I8:
6312 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6313 case ARM::ATOMIC_LOAD_MIN_I16:
6314 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6315 case ARM::ATOMIC_LOAD_MIN_I32:
6316 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6318 case ARM::ATOMIC_LOAD_MAX_I8:
6319 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6320 case ARM::ATOMIC_LOAD_MAX_I16:
6321 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6322 case ARM::ATOMIC_LOAD_MAX_I32:
6323 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6325 case ARM::ATOMIC_LOAD_UMIN_I8:
6326 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6327 case ARM::ATOMIC_LOAD_UMIN_I16:
6328 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6329 case ARM::ATOMIC_LOAD_UMIN_I32:
6330 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6332 case ARM::ATOMIC_LOAD_UMAX_I8:
6333 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6334 case ARM::ATOMIC_LOAD_UMAX_I16:
6335 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6336 case ARM::ATOMIC_LOAD_UMAX_I32:
6337 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6339 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6340 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6341 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6343 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6344 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6345 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6348 case ARM::ATOMADD6432:
6349 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6350 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6351 /*NeedsCarry*/ true);
6352 case ARM::ATOMSUB6432:
6353 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6354 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6355 /*NeedsCarry*/ true);
6356 case ARM::ATOMOR6432:
6357 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6358 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6359 case ARM::ATOMXOR6432:
6360 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6361 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6362 case ARM::ATOMAND6432:
6363 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6364 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6365 case ARM::ATOMSWAP6432:
6366 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6367 case ARM::ATOMCMPXCHG6432:
6368 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6369 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6370 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6372 case ARM::tMOVCCr_pseudo: {
6373 // To "insert" a SELECT_CC instruction, we actually have to insert the
6374 // diamond control-flow pattern. The incoming instruction knows the
6375 // destination vreg to set, the condition code register to branch on, the
6376 // true/false values to select between, and a branch opcode to use.
6377 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6378 MachineFunction::iterator It = BB;
6384 // cmpTY ccX, r1, r2
6386 // fallthrough --> copy0MBB
6387 MachineBasicBlock *thisMBB = BB;
6388 MachineFunction *F = BB->getParent();
6389 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6390 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6391 F->insert(It, copy0MBB);
6392 F->insert(It, sinkMBB);
6394 // Transfer the remainder of BB and its successor edges to sinkMBB.
6395 sinkMBB->splice(sinkMBB->begin(), BB,
6396 llvm::next(MachineBasicBlock::iterator(MI)),
6398 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6400 BB->addSuccessor(copy0MBB);
6401 BB->addSuccessor(sinkMBB);
6403 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6404 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6407 // %FalseValue = ...
6408 // # fallthrough to sinkMBB
6411 // Update machine-CFG edges
6412 BB->addSuccessor(sinkMBB);
6415 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6418 BuildMI(*BB, BB->begin(), dl,
6419 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6420 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6421 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6423 MI->eraseFromParent(); // The pseudo instruction is gone now.
6428 case ARM::BCCZi64: {
6429 // If there is an unconditional branch to the other successor, remove it.
6430 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6432 // Compare both parts that make up the double comparison separately for
6434 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6436 unsigned LHS1 = MI->getOperand(1).getReg();
6437 unsigned LHS2 = MI->getOperand(2).getReg();
6439 AddDefaultPred(BuildMI(BB, dl,
6440 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6441 .addReg(LHS1).addImm(0));
6442 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6443 .addReg(LHS2).addImm(0)
6444 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6446 unsigned RHS1 = MI->getOperand(3).getReg();
6447 unsigned RHS2 = MI->getOperand(4).getReg();
6448 AddDefaultPred(BuildMI(BB, dl,
6449 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6450 .addReg(LHS1).addReg(RHS1));
6451 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6452 .addReg(LHS2).addReg(RHS2)
6453 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6456 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6457 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6458 if (MI->getOperand(0).getImm() == ARMCC::NE)
6459 std::swap(destMBB, exitMBB);
6461 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6462 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6464 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6466 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6468 MI->eraseFromParent(); // The pseudo instruction is gone now.
6472 case ARM::Int_eh_sjlj_setjmp:
6473 case ARM::Int_eh_sjlj_setjmp_nofp:
6474 case ARM::tInt_eh_sjlj_setjmp:
6475 case ARM::t2Int_eh_sjlj_setjmp:
6476 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6477 EmitSjLjDispatchBlock(MI, BB);
6482 // To insert an ABS instruction, we have to insert the
6483 // diamond control-flow pattern. The incoming instruction knows the
6484 // source vreg to test against 0, the destination vreg to set,
6485 // the condition code register to branch on, the
6486 // true/false values to select between, and a branch opcode to use.
6491 // BCC (branch to SinkBB if V0 >= 0)
6492 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6493 // SinkBB: V1 = PHI(V2, V3)
6494 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6495 MachineFunction::iterator BBI = BB;
6497 MachineFunction *Fn = BB->getParent();
6498 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6499 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6500 Fn->insert(BBI, RSBBB);
6501 Fn->insert(BBI, SinkBB);
6503 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6504 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6505 bool isThumb2 = Subtarget->isThumb2();
6506 MachineRegisterInfo &MRI = Fn->getRegInfo();
6507 // In Thumb mode S must not be specified if source register is the SP or
6508 // PC and if destination register is the SP, so restrict register class
6509 unsigned NewMovDstReg = MRI.createVirtualRegister(
6510 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6511 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6512 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6514 // Transfer the remainder of BB and its successor edges to sinkMBB.
6515 SinkBB->splice(SinkBB->begin(), BB,
6516 llvm::next(MachineBasicBlock::iterator(MI)),
6518 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6520 BB->addSuccessor(RSBBB);
6521 BB->addSuccessor(SinkBB);
6523 // fall through to SinkMBB
6524 RSBBB->addSuccessor(SinkBB);
6526 // insert a movs at the end of BB
6527 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6529 .addReg(ABSSrcReg, RegState::Kill)
6530 .addImm((unsigned)ARMCC::AL).addReg(0)
6531 .addReg(ARM::CPSR, RegState::Define);
6533 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6535 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6536 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6538 // insert rsbri in RSBBB
6539 // Note: BCC and rsbri will be converted into predicated rsbmi
6540 // by if-conversion pass
6541 BuildMI(*RSBBB, RSBBB->begin(), dl,
6542 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6543 .addReg(NewMovDstReg, RegState::Kill)
6544 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6546 // insert PHI in SinkBB,
6547 // reuse ABSDstReg to not change uses of ABS instruction
6548 BuildMI(*SinkBB, SinkBB->begin(), dl,
6549 TII->get(ARM::PHI), ABSDstReg)
6550 .addReg(NewRsbDstReg).addMBB(RSBBB)
6551 .addReg(NewMovDstReg).addMBB(BB);
6553 // remove ABS instruction
6554 MI->eraseFromParent();
6556 // return last added BB
6562 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6563 SDNode *Node) const {
6564 if (!MI->hasPostISelHook()) {
6565 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6566 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6570 const MCInstrDesc *MCID = &MI->getDesc();
6571 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6572 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6573 // operand is still set to noreg. If needed, set the optional operand's
6574 // register to CPSR, and remove the redundant implicit def.
6576 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6578 // Rename pseudo opcodes.
6579 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6581 const ARMBaseInstrInfo *TII =
6582 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6583 MCID = &TII->get(NewOpc);
6585 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6586 "converted opcode should be the same except for cc_out");
6590 // Add the optional cc_out operand
6591 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6593 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6595 // Any ARM instruction that sets the 's' bit should specify an optional
6596 // "cc_out" operand in the last operand position.
6597 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6598 assert(!NewOpc && "Optional cc_out operand required");
6601 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6602 // since we already have an optional CPSR def.
6603 bool definesCPSR = false;
6604 bool deadCPSR = false;
6605 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6607 const MachineOperand &MO = MI->getOperand(i);
6608 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6612 MI->RemoveOperand(i);
6617 assert(!NewOpc && "Optional cc_out operand required");
6620 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6622 assert(!MI->getOperand(ccOutIdx).getReg() &&
6623 "expect uninitialized optional cc_out operand");
6627 // If this instruction was defined with an optional CPSR def and its dag node
6628 // had a live implicit CPSR def, then activate the optional CPSR def.
6629 MachineOperand &MO = MI->getOperand(ccOutIdx);
6630 MO.setReg(ARM::CPSR);
6634 //===----------------------------------------------------------------------===//
6635 // ARM Optimization Hooks
6636 //===----------------------------------------------------------------------===//
6639 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6640 TargetLowering::DAGCombinerInfo &DCI) {
6641 SelectionDAG &DAG = DCI.DAG;
6642 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6643 EVT VT = N->getValueType(0);
6644 unsigned Opc = N->getOpcode();
6645 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6646 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6647 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6648 ISD::CondCode CC = ISD::SETCC_INVALID;
6651 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6653 SDValue CCOp = Slct.getOperand(0);
6654 if (CCOp.getOpcode() == ISD::SETCC)
6655 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6658 bool DoXform = false;
6660 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6663 if (LHS.getOpcode() == ISD::Constant &&
6664 cast<ConstantSDNode>(LHS)->isNullValue()) {
6666 } else if (CC != ISD::SETCC_INVALID &&
6667 RHS.getOpcode() == ISD::Constant &&
6668 cast<ConstantSDNode>(RHS)->isNullValue()) {
6669 std::swap(LHS, RHS);
6670 SDValue Op0 = Slct.getOperand(0);
6671 EVT OpVT = isSlctCC ? Op0.getValueType() :
6672 Op0.getOperand(0).getValueType();
6673 bool isInt = OpVT.isInteger();
6674 CC = ISD::getSetCCInverse(CC, isInt);
6676 if (!TLI.isCondCodeLegal(CC, OpVT))
6677 return SDValue(); // Inverse operator isn't legal.
6684 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6686 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6687 Slct.getOperand(0), Slct.getOperand(1), CC);
6688 SDValue CCOp = Slct.getOperand(0);
6690 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6691 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6692 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6693 CCOp, OtherOp, Result);
6698 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6699 // (only after legalization).
6700 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6701 TargetLowering::DAGCombinerInfo &DCI,
6702 const ARMSubtarget *Subtarget) {
6704 // Only perform optimization if after legalize, and if NEON is available. We
6705 // also expected both operands to be BUILD_VECTORs.
6706 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6707 || N0.getOpcode() != ISD::BUILD_VECTOR
6708 || N1.getOpcode() != ISD::BUILD_VECTOR)
6711 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6712 EVT VT = N->getValueType(0);
6713 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6716 // Check that the vector operands are of the right form.
6717 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6718 // operands, where N is the size of the formed vector.
6719 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6720 // index such that we have a pair wise add pattern.
6722 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6723 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6725 SDValue Vec = N0->getOperand(0)->getOperand(0);
6726 SDNode *V = Vec.getNode();
6727 unsigned nextIndex = 0;
6729 // For each operands to the ADD which are BUILD_VECTORs,
6730 // check to see if each of their operands are an EXTRACT_VECTOR with
6731 // the same vector and appropriate index.
6732 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6733 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6734 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6736 SDValue ExtVec0 = N0->getOperand(i);
6737 SDValue ExtVec1 = N1->getOperand(i);
6739 // First operand is the vector, verify its the same.
6740 if (V != ExtVec0->getOperand(0).getNode() ||
6741 V != ExtVec1->getOperand(0).getNode())
6744 // Second is the constant, verify its correct.
6745 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6746 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6748 // For the constant, we want to see all the even or all the odd.
6749 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6750 || C1->getZExtValue() != nextIndex+1)
6759 // Create VPADDL node.
6760 SelectionDAG &DAG = DCI.DAG;
6761 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6763 // Build operand list.
6764 SmallVector<SDValue, 8> Ops;
6765 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6766 TLI.getPointerTy()));
6768 // Input is the vector.
6771 // Get widened type and narrowed type.
6773 unsigned numElem = VT.getVectorNumElements();
6774 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6775 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6776 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6777 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6779 llvm_unreachable("Invalid vector element type for padd optimization.");
6782 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6783 widenType, &Ops[0], Ops.size());
6784 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6787 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6788 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6789 /// called with the default operands, and if that fails, with commuted
6791 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6792 TargetLowering::DAGCombinerInfo &DCI,
6793 const ARMSubtarget *Subtarget){
6795 // Attempt to create vpaddl for this add.
6796 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6797 if (Result.getNode())
6800 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6801 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6802 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6803 if (Result.getNode()) return Result;
6808 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6810 static SDValue PerformADDCombine(SDNode *N,
6811 TargetLowering::DAGCombinerInfo &DCI,
6812 const ARMSubtarget *Subtarget) {
6813 SDValue N0 = N->getOperand(0);
6814 SDValue N1 = N->getOperand(1);
6816 // First try with the default operand order.
6817 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6818 if (Result.getNode())
6821 // If that didn't work, try again with the operands commuted.
6822 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6825 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6827 static SDValue PerformSUBCombine(SDNode *N,
6828 TargetLowering::DAGCombinerInfo &DCI) {
6829 SDValue N0 = N->getOperand(0);
6830 SDValue N1 = N->getOperand(1);
6832 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6833 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6834 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6835 if (Result.getNode()) return Result;
6841 /// PerformVMULCombine
6842 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6843 /// special multiplier accumulator forwarding.
6849 static SDValue PerformVMULCombine(SDNode *N,
6850 TargetLowering::DAGCombinerInfo &DCI,
6851 const ARMSubtarget *Subtarget) {
6852 if (!Subtarget->hasVMLxForwarding())
6855 SelectionDAG &DAG = DCI.DAG;
6856 SDValue N0 = N->getOperand(0);
6857 SDValue N1 = N->getOperand(1);
6858 unsigned Opcode = N0.getOpcode();
6859 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6860 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6861 Opcode = N1.getOpcode();
6862 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6863 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6868 EVT VT = N->getValueType(0);
6869 DebugLoc DL = N->getDebugLoc();
6870 SDValue N00 = N0->getOperand(0);
6871 SDValue N01 = N0->getOperand(1);
6872 return DAG.getNode(Opcode, DL, VT,
6873 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6874 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6877 static SDValue PerformMULCombine(SDNode *N,
6878 TargetLowering::DAGCombinerInfo &DCI,
6879 const ARMSubtarget *Subtarget) {
6880 SelectionDAG &DAG = DCI.DAG;
6882 if (Subtarget->isThumb1Only())
6885 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6888 EVT VT = N->getValueType(0);
6889 if (VT.is64BitVector() || VT.is128BitVector())
6890 return PerformVMULCombine(N, DCI, Subtarget);
6894 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6898 int64_t MulAmt = C->getSExtValue();
6899 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6901 ShiftAmt = ShiftAmt & (32 - 1);
6902 SDValue V = N->getOperand(0);
6903 DebugLoc DL = N->getDebugLoc();
6906 MulAmt >>= ShiftAmt;
6909 if (isPowerOf2_32(MulAmt - 1)) {
6910 // (mul x, 2^N + 1) => (add (shl x, N), x)
6911 Res = DAG.getNode(ISD::ADD, DL, VT,
6913 DAG.getNode(ISD::SHL, DL, VT,
6915 DAG.getConstant(Log2_32(MulAmt - 1),
6917 } else if (isPowerOf2_32(MulAmt + 1)) {
6918 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6919 Res = DAG.getNode(ISD::SUB, DL, VT,
6920 DAG.getNode(ISD::SHL, DL, VT,
6922 DAG.getConstant(Log2_32(MulAmt + 1),
6928 uint64_t MulAmtAbs = -MulAmt;
6929 if (isPowerOf2_32(MulAmtAbs + 1)) {
6930 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6931 Res = DAG.getNode(ISD::SUB, DL, VT,
6933 DAG.getNode(ISD::SHL, DL, VT,
6935 DAG.getConstant(Log2_32(MulAmtAbs + 1),
6937 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
6938 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6939 Res = DAG.getNode(ISD::ADD, DL, VT,
6941 DAG.getNode(ISD::SHL, DL, VT,
6943 DAG.getConstant(Log2_32(MulAmtAbs-1),
6945 Res = DAG.getNode(ISD::SUB, DL, VT,
6946 DAG.getConstant(0, MVT::i32),Res);
6953 Res = DAG.getNode(ISD::SHL, DL, VT,
6954 Res, DAG.getConstant(ShiftAmt, MVT::i32));
6956 // Do not add new nodes to DAG combiner worklist.
6957 DCI.CombineTo(N, Res, false);
6961 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6962 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6965 SDValue FalseVal = N.getOperand(0);
6966 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6970 return C->isAllOnesValue();
6971 return C->isNullValue();
6974 /// formConditionalOp - Combine an operation with a conditional move operand
6975 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
6976 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
6977 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
6979 SDValue N0 = N->getOperand(0);
6980 SDValue N1 = N->getOperand(1);
6982 bool isAND = N->getOpcode() == ISD::AND;
6983 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
6984 if (!isCand && Commutable) {
6985 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
6993 switch (N->getOpcode()) {
6994 default: llvm_unreachable("Unexpected node");
6995 case ISD::AND: Opc = ARMISD::CAND; break;
6996 case ISD::OR: Opc = ARMISD::COR; break;
6997 case ISD::XOR: Opc = ARMISD::CXOR; break;
6999 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
7000 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
7004 static SDValue PerformANDCombine(SDNode *N,
7005 TargetLowering::DAGCombinerInfo &DCI,
7006 const ARMSubtarget *Subtarget) {
7008 // Attempt to use immediate-form VBIC
7009 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7010 DebugLoc dl = N->getDebugLoc();
7011 EVT VT = N->getValueType(0);
7012 SelectionDAG &DAG = DCI.DAG;
7014 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7017 APInt SplatBits, SplatUndef;
7018 unsigned SplatBitSize;
7021 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7022 if (SplatBitSize <= 64) {
7024 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7025 SplatUndef.getZExtValue(), SplatBitSize,
7026 DAG, VbicVT, VT.is128BitVector(),
7028 if (Val.getNode()) {
7030 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7031 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7032 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7037 if (!Subtarget->isThumb1Only()) {
7038 // (and x, (cmov -1, y, cond)) => (and.cond x, y)
7039 SDValue CAND = formConditionalOp(N, DAG, true);
7047 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7048 static SDValue PerformORCombine(SDNode *N,
7049 TargetLowering::DAGCombinerInfo &DCI,
7050 const ARMSubtarget *Subtarget) {
7051 // Attempt to use immediate-form VORR
7052 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7053 DebugLoc dl = N->getDebugLoc();
7054 EVT VT = N->getValueType(0);
7055 SelectionDAG &DAG = DCI.DAG;
7057 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7060 APInt SplatBits, SplatUndef;
7061 unsigned SplatBitSize;
7063 if (BVN && Subtarget->hasNEON() &&
7064 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7065 if (SplatBitSize <= 64) {
7067 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7068 SplatUndef.getZExtValue(), SplatBitSize,
7069 DAG, VorrVT, VT.is128BitVector(),
7071 if (Val.getNode()) {
7073 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7074 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7075 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7080 if (!Subtarget->isThumb1Only()) {
7081 // (or x, (cmov 0, y, cond)) => (or.cond x, y)
7082 SDValue COR = formConditionalOp(N, DAG, true);
7087 SDValue N0 = N->getOperand(0);
7088 if (N0.getOpcode() != ISD::AND)
7090 SDValue N1 = N->getOperand(1);
7092 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7093 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7094 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7096 unsigned SplatBitSize;
7099 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7101 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7102 HasAnyUndefs) && !HasAnyUndefs) {
7103 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7105 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7106 HasAnyUndefs) && !HasAnyUndefs &&
7107 SplatBits0 == ~SplatBits1) {
7108 // Canonicalize the vector type to make instruction selection simpler.
7109 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7110 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7111 N0->getOperand(1), N0->getOperand(0),
7113 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7118 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7121 // BFI is only available on V6T2+
7122 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7125 DebugLoc DL = N->getDebugLoc();
7126 // 1) or (and A, mask), val => ARMbfi A, val, mask
7127 // iff (val & mask) == val
7129 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7130 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7131 // && mask == ~mask2
7132 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7133 // && ~mask == mask2
7134 // (i.e., copy a bitfield value into another bitfield of the same width)
7139 SDValue N00 = N0.getOperand(0);
7141 // The value and the mask need to be constants so we can verify this is
7142 // actually a bitfield set. If the mask is 0xffff, we can do better
7143 // via a movt instruction, so don't use BFI in that case.
7144 SDValue MaskOp = N0.getOperand(1);
7145 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7148 unsigned Mask = MaskC->getZExtValue();
7152 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7153 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7155 unsigned Val = N1C->getZExtValue();
7156 if ((Val & ~Mask) != Val)
7159 if (ARM::isBitFieldInvertedMask(Mask)) {
7160 Val >>= CountTrailingZeros_32(~Mask);
7162 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7163 DAG.getConstant(Val, MVT::i32),
7164 DAG.getConstant(Mask, MVT::i32));
7166 // Do not add new nodes to DAG combiner worklist.
7167 DCI.CombineTo(N, Res, false);
7170 } else if (N1.getOpcode() == ISD::AND) {
7171 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7172 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7175 unsigned Mask2 = N11C->getZExtValue();
7177 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7179 if (ARM::isBitFieldInvertedMask(Mask) &&
7181 // The pack halfword instruction works better for masks that fit it,
7182 // so use that when it's available.
7183 if (Subtarget->hasT2ExtractPack() &&
7184 (Mask == 0xffff || Mask == 0xffff0000))
7187 unsigned amt = CountTrailingZeros_32(Mask2);
7188 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7189 DAG.getConstant(amt, MVT::i32));
7190 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7191 DAG.getConstant(Mask, MVT::i32));
7192 // Do not add new nodes to DAG combiner worklist.
7193 DCI.CombineTo(N, Res, false);
7195 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7197 // The pack halfword instruction works better for masks that fit it,
7198 // so use that when it's available.
7199 if (Subtarget->hasT2ExtractPack() &&
7200 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7203 unsigned lsb = CountTrailingZeros_32(Mask);
7204 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7205 DAG.getConstant(lsb, MVT::i32));
7206 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7207 DAG.getConstant(Mask2, MVT::i32));
7208 // Do not add new nodes to DAG combiner worklist.
7209 DCI.CombineTo(N, Res, false);
7214 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7215 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7216 ARM::isBitFieldInvertedMask(~Mask)) {
7217 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7218 // where lsb(mask) == #shamt and masked bits of B are known zero.
7219 SDValue ShAmt = N00.getOperand(1);
7220 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7221 unsigned LSB = CountTrailingZeros_32(Mask);
7225 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7226 DAG.getConstant(~Mask, MVT::i32));
7228 // Do not add new nodes to DAG combiner worklist.
7229 DCI.CombineTo(N, Res, false);
7235 static SDValue PerformXORCombine(SDNode *N,
7236 TargetLowering::DAGCombinerInfo &DCI,
7237 const ARMSubtarget *Subtarget) {
7238 EVT VT = N->getValueType(0);
7239 SelectionDAG &DAG = DCI.DAG;
7241 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7244 if (!Subtarget->isThumb1Only()) {
7245 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7246 SDValue CXOR = formConditionalOp(N, DAG, true);
7254 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7255 /// the bits being cleared by the AND are not demanded by the BFI.
7256 static SDValue PerformBFICombine(SDNode *N,
7257 TargetLowering::DAGCombinerInfo &DCI) {
7258 SDValue N1 = N->getOperand(1);
7259 if (N1.getOpcode() == ISD::AND) {
7260 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7263 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7264 unsigned LSB = CountTrailingZeros_32(~InvMask);
7265 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7266 unsigned Mask = (1 << Width)-1;
7267 unsigned Mask2 = N11C->getZExtValue();
7268 if ((Mask & (~Mask2)) == 0)
7269 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7270 N->getOperand(0), N1.getOperand(0),
7276 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7277 /// ARMISD::VMOVRRD.
7278 static SDValue PerformVMOVRRDCombine(SDNode *N,
7279 TargetLowering::DAGCombinerInfo &DCI) {
7280 // vmovrrd(vmovdrr x, y) -> x,y
7281 SDValue InDouble = N->getOperand(0);
7282 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7283 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7285 // vmovrrd(load f64) -> (load i32), (load i32)
7286 SDNode *InNode = InDouble.getNode();
7287 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7288 InNode->getValueType(0) == MVT::f64 &&
7289 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7290 !cast<LoadSDNode>(InNode)->isVolatile()) {
7291 // TODO: Should this be done for non-FrameIndex operands?
7292 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7294 SelectionDAG &DAG = DCI.DAG;
7295 DebugLoc DL = LD->getDebugLoc();
7296 SDValue BasePtr = LD->getBasePtr();
7297 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7298 LD->getPointerInfo(), LD->isVolatile(),
7299 LD->isNonTemporal(), LD->isInvariant(),
7300 LD->getAlignment());
7302 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7303 DAG.getConstant(4, MVT::i32));
7304 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7305 LD->getPointerInfo(), LD->isVolatile(),
7306 LD->isNonTemporal(), LD->isInvariant(),
7307 std::min(4U, LD->getAlignment() / 2));
7309 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7310 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7311 DCI.RemoveFromWorklist(LD);
7319 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7320 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7321 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7322 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7323 SDValue Op0 = N->getOperand(0);
7324 SDValue Op1 = N->getOperand(1);
7325 if (Op0.getOpcode() == ISD::BITCAST)
7326 Op0 = Op0.getOperand(0);
7327 if (Op1.getOpcode() == ISD::BITCAST)
7328 Op1 = Op1.getOperand(0);
7329 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7330 Op0.getNode() == Op1.getNode() &&
7331 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7332 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7333 N->getValueType(0), Op0.getOperand(0));
7337 /// PerformSTORECombine - Target-specific dag combine xforms for
7339 static SDValue PerformSTORECombine(SDNode *N,
7340 TargetLowering::DAGCombinerInfo &DCI) {
7341 // Bitcast an i64 store extracted from a vector to f64.
7342 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7343 StoreSDNode *St = cast<StoreSDNode>(N);
7344 SDValue StVal = St->getValue();
7345 if (!ISD::isNormalStore(St) || St->isVolatile())
7348 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7349 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7350 SelectionDAG &DAG = DCI.DAG;
7351 DebugLoc DL = St->getDebugLoc();
7352 SDValue BasePtr = St->getBasePtr();
7353 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7354 StVal.getNode()->getOperand(0), BasePtr,
7355 St->getPointerInfo(), St->isVolatile(),
7356 St->isNonTemporal(), St->getAlignment());
7358 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7359 DAG.getConstant(4, MVT::i32));
7360 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7361 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7362 St->isNonTemporal(),
7363 std::min(4U, St->getAlignment() / 2));
7366 if (StVal.getValueType() != MVT::i64 ||
7367 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7370 SelectionDAG &DAG = DCI.DAG;
7371 DebugLoc dl = StVal.getDebugLoc();
7372 SDValue IntVec = StVal.getOperand(0);
7373 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7374 IntVec.getValueType().getVectorNumElements());
7375 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7376 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7377 Vec, StVal.getOperand(1));
7378 dl = N->getDebugLoc();
7379 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7380 // Make the DAGCombiner fold the bitcasts.
7381 DCI.AddToWorklist(Vec.getNode());
7382 DCI.AddToWorklist(ExtElt.getNode());
7383 DCI.AddToWorklist(V.getNode());
7384 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7385 St->getPointerInfo(), St->isVolatile(),
7386 St->isNonTemporal(), St->getAlignment(),
7390 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7391 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7392 /// i64 vector to have f64 elements, since the value can then be loaded
7393 /// directly into a VFP register.
7394 static bool hasNormalLoadOperand(SDNode *N) {
7395 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7396 for (unsigned i = 0; i < NumElts; ++i) {
7397 SDNode *Elt = N->getOperand(i).getNode();
7398 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7404 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7405 /// ISD::BUILD_VECTOR.
7406 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7407 TargetLowering::DAGCombinerInfo &DCI){
7408 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7409 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7410 // into a pair of GPRs, which is fine when the value is used as a scalar,
7411 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7412 SelectionDAG &DAG = DCI.DAG;
7413 if (N->getNumOperands() == 2) {
7414 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7419 // Load i64 elements as f64 values so that type legalization does not split
7420 // them up into i32 values.
7421 EVT VT = N->getValueType(0);
7422 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7424 DebugLoc dl = N->getDebugLoc();
7425 SmallVector<SDValue, 8> Ops;
7426 unsigned NumElts = VT.getVectorNumElements();
7427 for (unsigned i = 0; i < NumElts; ++i) {
7428 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7430 // Make the DAGCombiner fold the bitcast.
7431 DCI.AddToWorklist(V.getNode());
7433 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7434 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7435 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7438 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7439 /// ISD::INSERT_VECTOR_ELT.
7440 static SDValue PerformInsertEltCombine(SDNode *N,
7441 TargetLowering::DAGCombinerInfo &DCI) {
7442 // Bitcast an i64 load inserted into a vector to f64.
7443 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7444 EVT VT = N->getValueType(0);
7445 SDNode *Elt = N->getOperand(1).getNode();
7446 if (VT.getVectorElementType() != MVT::i64 ||
7447 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7450 SelectionDAG &DAG = DCI.DAG;
7451 DebugLoc dl = N->getDebugLoc();
7452 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7453 VT.getVectorNumElements());
7454 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7455 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7456 // Make the DAGCombiner fold the bitcasts.
7457 DCI.AddToWorklist(Vec.getNode());
7458 DCI.AddToWorklist(V.getNode());
7459 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7460 Vec, V, N->getOperand(2));
7461 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7464 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7465 /// ISD::VECTOR_SHUFFLE.
7466 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7467 // The LLVM shufflevector instruction does not require the shuffle mask
7468 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7469 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7470 // operands do not match the mask length, they are extended by concatenating
7471 // them with undef vectors. That is probably the right thing for other
7472 // targets, but for NEON it is better to concatenate two double-register
7473 // size vector operands into a single quad-register size vector. Do that
7474 // transformation here:
7475 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7476 // shuffle(concat(v1, v2), undef)
7477 SDValue Op0 = N->getOperand(0);
7478 SDValue Op1 = N->getOperand(1);
7479 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7480 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7481 Op0.getNumOperands() != 2 ||
7482 Op1.getNumOperands() != 2)
7484 SDValue Concat0Op1 = Op0.getOperand(1);
7485 SDValue Concat1Op1 = Op1.getOperand(1);
7486 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7487 Concat1Op1.getOpcode() != ISD::UNDEF)
7489 // Skip the transformation if any of the types are illegal.
7490 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7491 EVT VT = N->getValueType(0);
7492 if (!TLI.isTypeLegal(VT) ||
7493 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7494 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7497 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7498 Op0.getOperand(0), Op1.getOperand(0));
7499 // Translate the shuffle mask.
7500 SmallVector<int, 16> NewMask;
7501 unsigned NumElts = VT.getVectorNumElements();
7502 unsigned HalfElts = NumElts/2;
7503 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7504 for (unsigned n = 0; n < NumElts; ++n) {
7505 int MaskElt = SVN->getMaskElt(n);
7507 if (MaskElt < (int)HalfElts)
7509 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7510 NewElt = HalfElts + MaskElt - NumElts;
7511 NewMask.push_back(NewElt);
7513 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7514 DAG.getUNDEF(VT), NewMask.data());
7517 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7518 /// NEON load/store intrinsics to merge base address updates.
7519 static SDValue CombineBaseUpdate(SDNode *N,
7520 TargetLowering::DAGCombinerInfo &DCI) {
7521 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7524 SelectionDAG &DAG = DCI.DAG;
7525 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7526 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7527 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7528 SDValue Addr = N->getOperand(AddrOpIdx);
7530 // Search for a use of the address operand that is an increment.
7531 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7532 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7534 if (User->getOpcode() != ISD::ADD ||
7535 UI.getUse().getResNo() != Addr.getResNo())
7538 // Check that the add is independent of the load/store. Otherwise, folding
7539 // it would create a cycle.
7540 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7543 // Find the new opcode for the updating load/store.
7545 bool isLaneOp = false;
7546 unsigned NewOpc = 0;
7547 unsigned NumVecs = 0;
7549 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7551 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7552 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7554 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7556 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7558 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7560 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7561 NumVecs = 2; isLaneOp = true; break;
7562 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7563 NumVecs = 3; isLaneOp = true; break;
7564 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7565 NumVecs = 4; isLaneOp = true; break;
7566 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7567 NumVecs = 1; isLoad = false; break;
7568 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7569 NumVecs = 2; isLoad = false; break;
7570 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7571 NumVecs = 3; isLoad = false; break;
7572 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7573 NumVecs = 4; isLoad = false; break;
7574 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7575 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7576 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7577 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7578 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7579 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7583 switch (N->getOpcode()) {
7584 default: llvm_unreachable("unexpected opcode for Neon base update");
7585 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7586 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7587 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7591 // Find the size of memory referenced by the load/store.
7594 VecTy = N->getValueType(0);
7596 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7597 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7599 NumBytes /= VecTy.getVectorNumElements();
7601 // If the increment is a constant, it must match the memory ref size.
7602 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7603 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7604 uint64_t IncVal = CInc->getZExtValue();
7605 if (IncVal != NumBytes)
7607 } else if (NumBytes >= 3 * 16) {
7608 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7609 // separate instructions that make it harder to use a non-constant update.
7613 // Create the new updating load/store node.
7615 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7617 for (n = 0; n < NumResultVecs; ++n)
7619 Tys[n++] = MVT::i32;
7620 Tys[n] = MVT::Other;
7621 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7622 SmallVector<SDValue, 8> Ops;
7623 Ops.push_back(N->getOperand(0)); // incoming chain
7624 Ops.push_back(N->getOperand(AddrOpIdx));
7626 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7627 Ops.push_back(N->getOperand(i));
7629 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7630 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7631 Ops.data(), Ops.size(),
7632 MemInt->getMemoryVT(),
7633 MemInt->getMemOperand());
7636 std::vector<SDValue> NewResults;
7637 for (unsigned i = 0; i < NumResultVecs; ++i) {
7638 NewResults.push_back(SDValue(UpdN.getNode(), i));
7640 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7641 DCI.CombineTo(N, NewResults);
7642 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7649 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7650 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7651 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7653 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7654 SelectionDAG &DAG = DCI.DAG;
7655 EVT VT = N->getValueType(0);
7656 // vldN-dup instructions only support 64-bit vectors for N > 1.
7657 if (!VT.is64BitVector())
7660 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7661 SDNode *VLD = N->getOperand(0).getNode();
7662 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7664 unsigned NumVecs = 0;
7665 unsigned NewOpc = 0;
7666 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7667 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7669 NewOpc = ARMISD::VLD2DUP;
7670 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7672 NewOpc = ARMISD::VLD3DUP;
7673 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7675 NewOpc = ARMISD::VLD4DUP;
7680 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7681 // numbers match the load.
7682 unsigned VLDLaneNo =
7683 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7684 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7686 // Ignore uses of the chain result.
7687 if (UI.getUse().getResNo() == NumVecs)
7690 if (User->getOpcode() != ARMISD::VDUPLANE ||
7691 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7695 // Create the vldN-dup node.
7698 for (n = 0; n < NumVecs; ++n)
7700 Tys[n] = MVT::Other;
7701 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7702 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7703 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7704 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7705 Ops, 2, VLDMemInt->getMemoryVT(),
7706 VLDMemInt->getMemOperand());
7709 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7711 unsigned ResNo = UI.getUse().getResNo();
7712 // Ignore uses of the chain result.
7713 if (ResNo == NumVecs)
7716 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7719 // Now the vldN-lane intrinsic is dead except for its chain result.
7720 // Update uses of the chain.
7721 std::vector<SDValue> VLDDupResults;
7722 for (unsigned n = 0; n < NumVecs; ++n)
7723 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7724 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7725 DCI.CombineTo(VLD, VLDDupResults);
7730 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7731 /// ARMISD::VDUPLANE.
7732 static SDValue PerformVDUPLANECombine(SDNode *N,
7733 TargetLowering::DAGCombinerInfo &DCI) {
7734 SDValue Op = N->getOperand(0);
7736 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7737 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7738 if (CombineVLDDUP(N, DCI))
7739 return SDValue(N, 0);
7741 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7742 // redundant. Ignore bit_converts for now; element sizes are checked below.
7743 while (Op.getOpcode() == ISD::BITCAST)
7744 Op = Op.getOperand(0);
7745 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7748 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7749 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7750 // The canonical VMOV for a zero vector uses a 32-bit element size.
7751 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7753 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7755 EVT VT = N->getValueType(0);
7756 if (EltSize > VT.getVectorElementType().getSizeInBits())
7759 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7762 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7763 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7764 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7768 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7770 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7775 APFloat APF = C->getValueAPF();
7776 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7777 != APFloat::opOK || !isExact)
7780 c0 = (I == 0) ? cN : c0;
7781 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7788 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7789 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7790 /// when the VMUL has a constant operand that is a power of 2.
7792 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7793 /// vmul.f32 d16, d17, d16
7794 /// vcvt.s32.f32 d16, d16
7796 /// vcvt.s32.f32 d16, d16, #3
7797 static SDValue PerformVCVTCombine(SDNode *N,
7798 TargetLowering::DAGCombinerInfo &DCI,
7799 const ARMSubtarget *Subtarget) {
7800 SelectionDAG &DAG = DCI.DAG;
7801 SDValue Op = N->getOperand(0);
7803 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7804 Op.getOpcode() != ISD::FMUL)
7808 SDValue N0 = Op->getOperand(0);
7809 SDValue ConstVec = Op->getOperand(1);
7810 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7812 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7813 !isConstVecPow2(ConstVec, isSigned, C))
7816 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7817 Intrinsic::arm_neon_vcvtfp2fxu;
7818 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7820 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7821 DAG.getConstant(Log2_64(C), MVT::i32));
7824 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7825 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7826 /// when the VDIV has a constant operand that is a power of 2.
7828 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7829 /// vcvt.f32.s32 d16, d16
7830 /// vdiv.f32 d16, d17, d16
7832 /// vcvt.f32.s32 d16, d16, #3
7833 static SDValue PerformVDIVCombine(SDNode *N,
7834 TargetLowering::DAGCombinerInfo &DCI,
7835 const ARMSubtarget *Subtarget) {
7836 SelectionDAG &DAG = DCI.DAG;
7837 SDValue Op = N->getOperand(0);
7838 unsigned OpOpcode = Op.getNode()->getOpcode();
7840 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7841 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7845 SDValue ConstVec = N->getOperand(1);
7846 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7848 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7849 !isConstVecPow2(ConstVec, isSigned, C))
7852 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7853 Intrinsic::arm_neon_vcvtfxu2fp;
7854 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7856 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7857 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7860 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7861 /// operand of a vector shift operation, where all the elements of the
7862 /// build_vector must have the same constant integer value.
7863 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7864 // Ignore bit_converts.
7865 while (Op.getOpcode() == ISD::BITCAST)
7866 Op = Op.getOperand(0);
7867 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7868 APInt SplatBits, SplatUndef;
7869 unsigned SplatBitSize;
7871 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7872 HasAnyUndefs, ElementBits) ||
7873 SplatBitSize > ElementBits)
7875 Cnt = SplatBits.getSExtValue();
7879 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7880 /// operand of a vector shift left operation. That value must be in the range:
7881 /// 0 <= Value < ElementBits for a left shift; or
7882 /// 0 <= Value <= ElementBits for a long left shift.
7883 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7884 assert(VT.isVector() && "vector shift count is not a vector type");
7885 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7886 if (! getVShiftImm(Op, ElementBits, Cnt))
7888 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7891 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7892 /// operand of a vector shift right operation. For a shift opcode, the value
7893 /// is positive, but for an intrinsic the value count must be negative. The
7894 /// absolute value must be in the range:
7895 /// 1 <= |Value| <= ElementBits for a right shift; or
7896 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7897 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7899 assert(VT.isVector() && "vector shift count is not a vector type");
7900 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7901 if (! getVShiftImm(Op, ElementBits, Cnt))
7905 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7908 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7909 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7910 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7913 // Don't do anything for most intrinsics.
7916 // Vector shifts: check for immediate versions and lower them.
7917 // Note: This is done during DAG combining instead of DAG legalizing because
7918 // the build_vectors for 64-bit vector element shift counts are generally
7919 // not legal, and it is hard to see their values after they get legalized to
7920 // loads from a constant pool.
7921 case Intrinsic::arm_neon_vshifts:
7922 case Intrinsic::arm_neon_vshiftu:
7923 case Intrinsic::arm_neon_vshiftls:
7924 case Intrinsic::arm_neon_vshiftlu:
7925 case Intrinsic::arm_neon_vshiftn:
7926 case Intrinsic::arm_neon_vrshifts:
7927 case Intrinsic::arm_neon_vrshiftu:
7928 case Intrinsic::arm_neon_vrshiftn:
7929 case Intrinsic::arm_neon_vqshifts:
7930 case Intrinsic::arm_neon_vqshiftu:
7931 case Intrinsic::arm_neon_vqshiftsu:
7932 case Intrinsic::arm_neon_vqshiftns:
7933 case Intrinsic::arm_neon_vqshiftnu:
7934 case Intrinsic::arm_neon_vqshiftnsu:
7935 case Intrinsic::arm_neon_vqrshiftns:
7936 case Intrinsic::arm_neon_vqrshiftnu:
7937 case Intrinsic::arm_neon_vqrshiftnsu: {
7938 EVT VT = N->getOperand(1).getValueType();
7940 unsigned VShiftOpc = 0;
7943 case Intrinsic::arm_neon_vshifts:
7944 case Intrinsic::arm_neon_vshiftu:
7945 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7946 VShiftOpc = ARMISD::VSHL;
7949 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7950 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7951 ARMISD::VSHRs : ARMISD::VSHRu);
7956 case Intrinsic::arm_neon_vshiftls:
7957 case Intrinsic::arm_neon_vshiftlu:
7958 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7960 llvm_unreachable("invalid shift count for vshll intrinsic");
7962 case Intrinsic::arm_neon_vrshifts:
7963 case Intrinsic::arm_neon_vrshiftu:
7964 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7968 case Intrinsic::arm_neon_vqshifts:
7969 case Intrinsic::arm_neon_vqshiftu:
7970 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7974 case Intrinsic::arm_neon_vqshiftsu:
7975 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7977 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7979 case Intrinsic::arm_neon_vshiftn:
7980 case Intrinsic::arm_neon_vrshiftn:
7981 case Intrinsic::arm_neon_vqshiftns:
7982 case Intrinsic::arm_neon_vqshiftnu:
7983 case Intrinsic::arm_neon_vqshiftnsu:
7984 case Intrinsic::arm_neon_vqrshiftns:
7985 case Intrinsic::arm_neon_vqrshiftnu:
7986 case Intrinsic::arm_neon_vqrshiftnsu:
7987 // Narrowing shifts require an immediate right shift.
7988 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7990 llvm_unreachable("invalid shift count for narrowing vector shift "
7994 llvm_unreachable("unhandled vector shift");
7998 case Intrinsic::arm_neon_vshifts:
7999 case Intrinsic::arm_neon_vshiftu:
8000 // Opcode already set above.
8002 case Intrinsic::arm_neon_vshiftls:
8003 case Intrinsic::arm_neon_vshiftlu:
8004 if (Cnt == VT.getVectorElementType().getSizeInBits())
8005 VShiftOpc = ARMISD::VSHLLi;
8007 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8008 ARMISD::VSHLLs : ARMISD::VSHLLu);
8010 case Intrinsic::arm_neon_vshiftn:
8011 VShiftOpc = ARMISD::VSHRN; break;
8012 case Intrinsic::arm_neon_vrshifts:
8013 VShiftOpc = ARMISD::VRSHRs; break;
8014 case Intrinsic::arm_neon_vrshiftu:
8015 VShiftOpc = ARMISD::VRSHRu; break;
8016 case Intrinsic::arm_neon_vrshiftn:
8017 VShiftOpc = ARMISD::VRSHRN; break;
8018 case Intrinsic::arm_neon_vqshifts:
8019 VShiftOpc = ARMISD::VQSHLs; break;
8020 case Intrinsic::arm_neon_vqshiftu:
8021 VShiftOpc = ARMISD::VQSHLu; break;
8022 case Intrinsic::arm_neon_vqshiftsu:
8023 VShiftOpc = ARMISD::VQSHLsu; break;
8024 case Intrinsic::arm_neon_vqshiftns:
8025 VShiftOpc = ARMISD::VQSHRNs; break;
8026 case Intrinsic::arm_neon_vqshiftnu:
8027 VShiftOpc = ARMISD::VQSHRNu; break;
8028 case Intrinsic::arm_neon_vqshiftnsu:
8029 VShiftOpc = ARMISD::VQSHRNsu; break;
8030 case Intrinsic::arm_neon_vqrshiftns:
8031 VShiftOpc = ARMISD::VQRSHRNs; break;
8032 case Intrinsic::arm_neon_vqrshiftnu:
8033 VShiftOpc = ARMISD::VQRSHRNu; break;
8034 case Intrinsic::arm_neon_vqrshiftnsu:
8035 VShiftOpc = ARMISD::VQRSHRNsu; break;
8038 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8039 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8042 case Intrinsic::arm_neon_vshiftins: {
8043 EVT VT = N->getOperand(1).getValueType();
8045 unsigned VShiftOpc = 0;
8047 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8048 VShiftOpc = ARMISD::VSLI;
8049 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8050 VShiftOpc = ARMISD::VSRI;
8052 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8055 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8056 N->getOperand(1), N->getOperand(2),
8057 DAG.getConstant(Cnt, MVT::i32));
8060 case Intrinsic::arm_neon_vqrshifts:
8061 case Intrinsic::arm_neon_vqrshiftu:
8062 // No immediate versions of these to check for.
8069 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8070 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8071 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8072 /// vector element shift counts are generally not legal, and it is hard to see
8073 /// their values after they get legalized to loads from a constant pool.
8074 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8075 const ARMSubtarget *ST) {
8076 EVT VT = N->getValueType(0);
8077 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8078 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8079 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8080 SDValue N1 = N->getOperand(1);
8081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8082 SDValue N0 = N->getOperand(0);
8083 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8084 DAG.MaskedValueIsZero(N0.getOperand(0),
8085 APInt::getHighBitsSet(32, 16)))
8086 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8090 // Nothing to be done for scalar shifts.
8091 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8092 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8095 assert(ST->hasNEON() && "unexpected vector shift");
8098 switch (N->getOpcode()) {
8099 default: llvm_unreachable("unexpected shift opcode");
8102 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8103 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8104 DAG.getConstant(Cnt, MVT::i32));
8109 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8110 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8111 ARMISD::VSHRs : ARMISD::VSHRu);
8112 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8113 DAG.getConstant(Cnt, MVT::i32));
8119 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8120 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8121 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8122 const ARMSubtarget *ST) {
8123 SDValue N0 = N->getOperand(0);
8125 // Check for sign- and zero-extensions of vector extract operations of 8-
8126 // and 16-bit vector elements. NEON supports these directly. They are
8127 // handled during DAG combining because type legalization will promote them
8128 // to 32-bit types and it is messy to recognize the operations after that.
8129 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8130 SDValue Vec = N0.getOperand(0);
8131 SDValue Lane = N0.getOperand(1);
8132 EVT VT = N->getValueType(0);
8133 EVT EltVT = N0.getValueType();
8134 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8136 if (VT == MVT::i32 &&
8137 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8138 TLI.isTypeLegal(Vec.getValueType()) &&
8139 isa<ConstantSDNode>(Lane)) {
8142 switch (N->getOpcode()) {
8143 default: llvm_unreachable("unexpected opcode");
8144 case ISD::SIGN_EXTEND:
8145 Opc = ARMISD::VGETLANEs;
8147 case ISD::ZERO_EXTEND:
8148 case ISD::ANY_EXTEND:
8149 Opc = ARMISD::VGETLANEu;
8152 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8159 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8160 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8161 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8162 const ARMSubtarget *ST) {
8163 // If the target supports NEON, try to use vmax/vmin instructions for f32
8164 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8165 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8166 // a NaN; only do the transformation when it matches that behavior.
8168 // For now only do this when using NEON for FP operations; if using VFP, it
8169 // is not obvious that the benefit outweighs the cost of switching to the
8171 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8172 N->getValueType(0) != MVT::f32)
8175 SDValue CondLHS = N->getOperand(0);
8176 SDValue CondRHS = N->getOperand(1);
8177 SDValue LHS = N->getOperand(2);
8178 SDValue RHS = N->getOperand(3);
8179 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8181 unsigned Opcode = 0;
8183 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8184 IsReversed = false; // x CC y ? x : y
8185 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8186 IsReversed = true ; // x CC y ? y : x
8200 // If LHS is NaN, an ordered comparison will be false and the result will
8201 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8202 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8203 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8204 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8206 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8207 // will return -0, so vmin can only be used for unsafe math or if one of
8208 // the operands is known to be nonzero.
8209 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8210 !DAG.getTarget().Options.UnsafeFPMath &&
8211 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8213 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8222 // If LHS is NaN, an ordered comparison will be false and the result will
8223 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8224 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8225 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8226 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8228 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8229 // will return +0, so vmax can only be used for unsafe math or if one of
8230 // the operands is known to be nonzero.
8231 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8232 !DAG.getTarget().Options.UnsafeFPMath &&
8233 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8235 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8241 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8244 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8246 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8247 SDValue Cmp = N->getOperand(4);
8248 if (Cmp.getOpcode() != ARMISD::CMPZ)
8249 // Only looking at EQ and NE cases.
8252 EVT VT = N->getValueType(0);
8253 DebugLoc dl = N->getDebugLoc();
8254 SDValue LHS = Cmp.getOperand(0);
8255 SDValue RHS = Cmp.getOperand(1);
8256 SDValue FalseVal = N->getOperand(0);
8257 SDValue TrueVal = N->getOperand(1);
8258 SDValue ARMcc = N->getOperand(2);
8259 ARMCC::CondCodes CC =
8260 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8278 /// FIXME: Turn this into a target neutral optimization?
8280 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8281 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8282 N->getOperand(3), Cmp);
8283 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8285 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8286 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8287 N->getOperand(3), NewCmp);
8290 if (Res.getNode()) {
8291 APInt KnownZero, KnownOne;
8292 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8293 // Capture demanded bits information that would be otherwise lost.
8294 if (KnownZero == 0xfffffffe)
8295 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8296 DAG.getValueType(MVT::i1));
8297 else if (KnownZero == 0xffffff00)
8298 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8299 DAG.getValueType(MVT::i8));
8300 else if (KnownZero == 0xffff0000)
8301 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8302 DAG.getValueType(MVT::i16));
8308 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8309 DAGCombinerInfo &DCI) const {
8310 switch (N->getOpcode()) {
8312 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8313 case ISD::SUB: return PerformSUBCombine(N, DCI);
8314 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8315 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8316 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8317 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8318 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8319 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8320 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8321 case ISD::STORE: return PerformSTORECombine(N, DCI);
8322 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8323 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8324 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8325 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8326 case ISD::FP_TO_SINT:
8327 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8328 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8329 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8332 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8333 case ISD::SIGN_EXTEND:
8334 case ISD::ZERO_EXTEND:
8335 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8336 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8337 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8338 case ARMISD::VLD2DUP:
8339 case ARMISD::VLD3DUP:
8340 case ARMISD::VLD4DUP:
8341 return CombineBaseUpdate(N, DCI);
8342 case ISD::INTRINSIC_VOID:
8343 case ISD::INTRINSIC_W_CHAIN:
8344 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8345 case Intrinsic::arm_neon_vld1:
8346 case Intrinsic::arm_neon_vld2:
8347 case Intrinsic::arm_neon_vld3:
8348 case Intrinsic::arm_neon_vld4:
8349 case Intrinsic::arm_neon_vld2lane:
8350 case Intrinsic::arm_neon_vld3lane:
8351 case Intrinsic::arm_neon_vld4lane:
8352 case Intrinsic::arm_neon_vst1:
8353 case Intrinsic::arm_neon_vst2:
8354 case Intrinsic::arm_neon_vst3:
8355 case Intrinsic::arm_neon_vst4:
8356 case Intrinsic::arm_neon_vst2lane:
8357 case Intrinsic::arm_neon_vst3lane:
8358 case Intrinsic::arm_neon_vst4lane:
8359 return CombineBaseUpdate(N, DCI);
8367 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8369 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8372 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8373 if (!Subtarget->allowsUnalignedMem())
8376 switch (VT.getSimpleVT().SimpleTy) {
8383 // FIXME: VLD1 etc with standard alignment is legal.
8387 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8388 unsigned AlignCheck) {
8389 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8390 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8393 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8394 unsigned DstAlign, unsigned SrcAlign,
8397 MachineFunction &MF) const {
8398 const Function *F = MF.getFunction();
8400 // See if we can use NEON instructions for this...
8402 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8403 Subtarget->hasNEON()) {
8404 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8406 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8411 // Lowering to i32/i16 if the size permits.
8414 } else if (Size >= 2) {
8418 // Let the target-independent logic figure it out.
8422 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8427 switch (VT.getSimpleVT().SimpleTy) {
8428 default: return false;
8443 if ((V & (Scale - 1)) != 0)
8446 return V == (V & ((1LL << 5) - 1));
8449 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8450 const ARMSubtarget *Subtarget) {
8457 switch (VT.getSimpleVT().SimpleTy) {
8458 default: return false;
8463 // + imm12 or - imm8
8465 return V == (V & ((1LL << 8) - 1));
8466 return V == (V & ((1LL << 12) - 1));
8469 // Same as ARM mode. FIXME: NEON?
8470 if (!Subtarget->hasVFP2())
8475 return V == (V & ((1LL << 8) - 1));
8479 /// isLegalAddressImmediate - Return true if the integer value can be used
8480 /// as the offset of the target addressing mode for load / store of the
8482 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8483 const ARMSubtarget *Subtarget) {
8490 if (Subtarget->isThumb1Only())
8491 return isLegalT1AddressImmediate(V, VT);
8492 else if (Subtarget->isThumb2())
8493 return isLegalT2AddressImmediate(V, VT, Subtarget);
8498 switch (VT.getSimpleVT().SimpleTy) {
8499 default: return false;
8504 return V == (V & ((1LL << 12) - 1));
8507 return V == (V & ((1LL << 8) - 1));
8510 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8515 return V == (V & ((1LL << 8) - 1));
8519 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8521 int Scale = AM.Scale;
8525 switch (VT.getSimpleVT().SimpleTy) {
8526 default: return false;
8535 return Scale == 2 || Scale == 4 || Scale == 8;
8538 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8542 // Note, we allow "void" uses (basically, uses that aren't loads or
8543 // stores), because arm allows folding a scale into many arithmetic
8544 // operations. This should be made more precise and revisited later.
8546 // Allow r << imm, but the imm has to be a multiple of two.
8547 if (Scale & 1) return false;
8548 return isPowerOf2_32(Scale);
8552 /// isLegalAddressingMode - Return true if the addressing mode represented
8553 /// by AM is legal for this target, for a load/store of the specified type.
8554 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8556 EVT VT = getValueType(Ty, true);
8557 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8560 // Can never fold addr of global into load/store.
8565 case 0: // no scale reg, must be "r+i" or "r", or "i".
8568 if (Subtarget->isThumb1Only())
8572 // ARM doesn't support any R+R*scale+imm addr modes.
8579 if (Subtarget->isThumb2())
8580 return isLegalT2ScaledAddressingMode(AM, VT);
8582 int Scale = AM.Scale;
8583 switch (VT.getSimpleVT().SimpleTy) {
8584 default: return false;
8588 if (Scale < 0) Scale = -Scale;
8592 return isPowerOf2_32(Scale & ~1);
8596 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8601 // Note, we allow "void" uses (basically, uses that aren't loads or
8602 // stores), because arm allows folding a scale into many arithmetic
8603 // operations. This should be made more precise and revisited later.
8605 // Allow r << imm, but the imm has to be a multiple of two.
8606 if (Scale & 1) return false;
8607 return isPowerOf2_32(Scale);
8613 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8614 /// icmp immediate, that is the target has icmp instructions which can compare
8615 /// a register against the immediate without having to materialize the
8616 /// immediate into a register.
8617 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8618 // Thumb2 and ARM modes can use cmn for negative immediates.
8619 if (!Subtarget->isThumb())
8620 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
8621 if (Subtarget->isThumb2())
8622 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
8623 // Thumb1 doesn't have cmn, and only 8-bit immediates.
8624 return Imm >= 0 && Imm <= 255;
8627 /// isLegalAddImmediate - Return true if the specified immediate is legal
8628 /// add immediate, that is the target has add instructions which can add
8629 /// a register with the immediate without having to materialize the
8630 /// immediate into a register.
8631 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8632 return ARM_AM::getSOImmVal(Imm) != -1;
8635 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8636 bool isSEXTLoad, SDValue &Base,
8637 SDValue &Offset, bool &isInc,
8638 SelectionDAG &DAG) {
8639 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8642 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8644 Base = Ptr->getOperand(0);
8645 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8646 int RHSC = (int)RHS->getZExtValue();
8647 if (RHSC < 0 && RHSC > -256) {
8648 assert(Ptr->getOpcode() == ISD::ADD);
8650 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8654 isInc = (Ptr->getOpcode() == ISD::ADD);
8655 Offset = Ptr->getOperand(1);
8657 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8659 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8660 int RHSC = (int)RHS->getZExtValue();
8661 if (RHSC < 0 && RHSC > -0x1000) {
8662 assert(Ptr->getOpcode() == ISD::ADD);
8664 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8665 Base = Ptr->getOperand(0);
8670 if (Ptr->getOpcode() == ISD::ADD) {
8672 ARM_AM::ShiftOpc ShOpcVal=
8673 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8674 if (ShOpcVal != ARM_AM::no_shift) {
8675 Base = Ptr->getOperand(1);
8676 Offset = Ptr->getOperand(0);
8678 Base = Ptr->getOperand(0);
8679 Offset = Ptr->getOperand(1);
8684 isInc = (Ptr->getOpcode() == ISD::ADD);
8685 Base = Ptr->getOperand(0);
8686 Offset = Ptr->getOperand(1);
8690 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8694 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8695 bool isSEXTLoad, SDValue &Base,
8696 SDValue &Offset, bool &isInc,
8697 SelectionDAG &DAG) {
8698 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8701 Base = Ptr->getOperand(0);
8702 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8703 int RHSC = (int)RHS->getZExtValue();
8704 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8705 assert(Ptr->getOpcode() == ISD::ADD);
8707 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8709 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8710 isInc = Ptr->getOpcode() == ISD::ADD;
8711 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8719 /// getPreIndexedAddressParts - returns true by value, base pointer and
8720 /// offset pointer and addressing mode by reference if the node's address
8721 /// can be legally represented as pre-indexed load / store address.
8723 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8725 ISD::MemIndexedMode &AM,
8726 SelectionDAG &DAG) const {
8727 if (Subtarget->isThumb1Only())
8732 bool isSEXTLoad = false;
8733 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8734 Ptr = LD->getBasePtr();
8735 VT = LD->getMemoryVT();
8736 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8737 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8738 Ptr = ST->getBasePtr();
8739 VT = ST->getMemoryVT();
8744 bool isLegal = false;
8745 if (Subtarget->isThumb2())
8746 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8747 Offset, isInc, DAG);
8749 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8750 Offset, isInc, DAG);
8754 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8758 /// getPostIndexedAddressParts - returns true by value, base pointer and
8759 /// offset pointer and addressing mode by reference if this node can be
8760 /// combined with a load / store to form a post-indexed load / store.
8761 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8764 ISD::MemIndexedMode &AM,
8765 SelectionDAG &DAG) const {
8766 if (Subtarget->isThumb1Only())
8771 bool isSEXTLoad = false;
8772 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8773 VT = LD->getMemoryVT();
8774 Ptr = LD->getBasePtr();
8775 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8776 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8777 VT = ST->getMemoryVT();
8778 Ptr = ST->getBasePtr();
8783 bool isLegal = false;
8784 if (Subtarget->isThumb2())
8785 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8788 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8794 // Swap base ptr and offset to catch more post-index load / store when
8795 // it's legal. In Thumb2 mode, offset must be an immediate.
8796 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8797 !Subtarget->isThumb2())
8798 std::swap(Base, Offset);
8800 // Post-indexed load / store update the base pointer.
8805 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8809 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8812 const SelectionDAG &DAG,
8813 unsigned Depth) const {
8814 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
8815 switch (Op.getOpcode()) {
8817 case ARMISD::CMOV: {
8818 // Bits are known zero/one if known on the LHS and RHS.
8819 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
8820 if (KnownZero == 0 && KnownOne == 0) return;
8822 APInt KnownZeroRHS, KnownOneRHS;
8823 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
8824 KnownZero &= KnownZeroRHS;
8825 KnownOne &= KnownOneRHS;
8831 //===----------------------------------------------------------------------===//
8832 // ARM Inline Assembly Support
8833 //===----------------------------------------------------------------------===//
8835 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8836 // Looking for "rev" which is V6+.
8837 if (!Subtarget->hasV6Ops())
8840 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8841 std::string AsmStr = IA->getAsmString();
8842 SmallVector<StringRef, 4> AsmPieces;
8843 SplitString(AsmStr, AsmPieces, ";\n");
8845 switch (AsmPieces.size()) {
8846 default: return false;
8848 AsmStr = AsmPieces[0];
8850 SplitString(AsmStr, AsmPieces, " \t,");
8853 if (AsmPieces.size() == 3 &&
8854 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8855 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8856 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8857 if (Ty && Ty->getBitWidth() == 32)
8858 return IntrinsicLowering::LowerToByteSwap(CI);
8866 /// getConstraintType - Given a constraint letter, return the type of
8867 /// constraint it is for this target.
8868 ARMTargetLowering::ConstraintType
8869 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8870 if (Constraint.size() == 1) {
8871 switch (Constraint[0]) {
8873 case 'l': return C_RegisterClass;
8874 case 'w': return C_RegisterClass;
8875 case 'h': return C_RegisterClass;
8876 case 'x': return C_RegisterClass;
8877 case 't': return C_RegisterClass;
8878 case 'j': return C_Other; // Constant for movw.
8879 // An address with a single base register. Due to the way we
8880 // currently handle addresses it is the same as an 'r' memory constraint.
8881 case 'Q': return C_Memory;
8883 } else if (Constraint.size() == 2) {
8884 switch (Constraint[0]) {
8886 // All 'U+' constraints are addresses.
8887 case 'U': return C_Memory;
8890 return TargetLowering::getConstraintType(Constraint);
8893 /// Examine constraint type and operand type and determine a weight value.
8894 /// This object must already have been set up with the operand type
8895 /// and the current alternative constraint selected.
8896 TargetLowering::ConstraintWeight
8897 ARMTargetLowering::getSingleConstraintMatchWeight(
8898 AsmOperandInfo &info, const char *constraint) const {
8899 ConstraintWeight weight = CW_Invalid;
8900 Value *CallOperandVal = info.CallOperandVal;
8901 // If we don't have a value, we can't do a match,
8902 // but allow it at the lowest weight.
8903 if (CallOperandVal == NULL)
8905 Type *type = CallOperandVal->getType();
8906 // Look at the constraint type.
8907 switch (*constraint) {
8909 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8912 if (type->isIntegerTy()) {
8913 if (Subtarget->isThumb())
8914 weight = CW_SpecificReg;
8916 weight = CW_Register;
8920 if (type->isFloatingPointTy())
8921 weight = CW_Register;
8927 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8929 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8931 if (Constraint.size() == 1) {
8932 // GCC ARM Constraint Letters
8933 switch (Constraint[0]) {
8934 case 'l': // Low regs or general regs.
8935 if (Subtarget->isThumb())
8936 return RCPair(0U, ARM::tGPRRegisterClass);
8938 return RCPair(0U, ARM::GPRRegisterClass);
8939 case 'h': // High regs or no regs.
8940 if (Subtarget->isThumb())
8941 return RCPair(0U, ARM::hGPRRegisterClass);
8944 return RCPair(0U, ARM::GPRRegisterClass);
8947 return RCPair(0U, ARM::SPRRegisterClass);
8948 if (VT.getSizeInBits() == 64)
8949 return RCPair(0U, ARM::DPRRegisterClass);
8950 if (VT.getSizeInBits() == 128)
8951 return RCPair(0U, ARM::QPRRegisterClass);
8955 return RCPair(0U, ARM::SPR_8RegisterClass);
8956 if (VT.getSizeInBits() == 64)
8957 return RCPair(0U, ARM::DPR_8RegisterClass);
8958 if (VT.getSizeInBits() == 128)
8959 return RCPair(0U, ARM::QPR_8RegisterClass);
8963 return RCPair(0U, ARM::SPRRegisterClass);
8967 if (StringRef("{cc}").equals_lower(Constraint))
8968 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8970 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8973 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8974 /// vector. If it is invalid, don't add anything to Ops.
8975 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8976 std::string &Constraint,
8977 std::vector<SDValue>&Ops,
8978 SelectionDAG &DAG) const {
8979 SDValue Result(0, 0);
8981 // Currently only support length 1 constraints.
8982 if (Constraint.length() != 1) return;
8984 char ConstraintLetter = Constraint[0];
8985 switch (ConstraintLetter) {
8988 case 'I': case 'J': case 'K': case 'L':
8989 case 'M': case 'N': case 'O':
8990 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8994 int64_t CVal64 = C->getSExtValue();
8995 int CVal = (int) CVal64;
8996 // None of these constraints allow values larger than 32 bits. Check
8997 // that the value fits in an int.
9001 switch (ConstraintLetter) {
9003 // Constant suitable for movw, must be between 0 and
9005 if (Subtarget->hasV6T2Ops())
9006 if (CVal >= 0 && CVal <= 65535)
9010 if (Subtarget->isThumb1Only()) {
9011 // This must be a constant between 0 and 255, for ADD
9013 if (CVal >= 0 && CVal <= 255)
9015 } else if (Subtarget->isThumb2()) {
9016 // A constant that can be used as an immediate value in a
9017 // data-processing instruction.
9018 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9021 // A constant that can be used as an immediate value in a
9022 // data-processing instruction.
9023 if (ARM_AM::getSOImmVal(CVal) != -1)
9029 if (Subtarget->isThumb()) { // FIXME thumb2
9030 // This must be a constant between -255 and -1, for negated ADD
9031 // immediates. This can be used in GCC with an "n" modifier that
9032 // prints the negated value, for use with SUB instructions. It is
9033 // not useful otherwise but is implemented for compatibility.
9034 if (CVal >= -255 && CVal <= -1)
9037 // This must be a constant between -4095 and 4095. It is not clear
9038 // what this constraint is intended for. Implemented for
9039 // compatibility with GCC.
9040 if (CVal >= -4095 && CVal <= 4095)
9046 if (Subtarget->isThumb1Only()) {
9047 // A 32-bit value where only one byte has a nonzero value. Exclude
9048 // zero to match GCC. This constraint is used by GCC internally for
9049 // constants that can be loaded with a move/shift combination.
9050 // It is not useful otherwise but is implemented for compatibility.
9051 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9053 } else if (Subtarget->isThumb2()) {
9054 // A constant whose bitwise inverse can be used as an immediate
9055 // value in a data-processing instruction. This can be used in GCC
9056 // with a "B" modifier that prints the inverted value, for use with
9057 // BIC and MVN instructions. It is not useful otherwise but is
9058 // implemented for compatibility.
9059 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9062 // A constant whose bitwise inverse can be used as an immediate
9063 // value in a data-processing instruction. This can be used in GCC
9064 // with a "B" modifier that prints the inverted value, for use with
9065 // BIC and MVN instructions. It is not useful otherwise but is
9066 // implemented for compatibility.
9067 if (ARM_AM::getSOImmVal(~CVal) != -1)
9073 if (Subtarget->isThumb1Only()) {
9074 // This must be a constant between -7 and 7,
9075 // for 3-operand ADD/SUB immediate instructions.
9076 if (CVal >= -7 && CVal < 7)
9078 } else if (Subtarget->isThumb2()) {
9079 // A constant whose negation can be used as an immediate value in a
9080 // data-processing instruction. This can be used in GCC with an "n"
9081 // modifier that prints the negated value, for use with SUB
9082 // instructions. It is not useful otherwise but is implemented for
9084 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9087 // A constant whose negation can be used as an immediate value in a
9088 // data-processing instruction. This can be used in GCC with an "n"
9089 // modifier that prints the negated value, for use with SUB
9090 // instructions. It is not useful otherwise but is implemented for
9092 if (ARM_AM::getSOImmVal(-CVal) != -1)
9098 if (Subtarget->isThumb()) { // FIXME thumb2
9099 // This must be a multiple of 4 between 0 and 1020, for
9100 // ADD sp + immediate.
9101 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9104 // A power of two or a constant between 0 and 32. This is used in
9105 // GCC for the shift amount on shifted register operands, but it is
9106 // useful in general for any shift amounts.
9107 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9113 if (Subtarget->isThumb()) { // FIXME thumb2
9114 // This must be a constant between 0 and 31, for shift amounts.
9115 if (CVal >= 0 && CVal <= 31)
9121 if (Subtarget->isThumb()) { // FIXME thumb2
9122 // This must be a multiple of 4 between -508 and 508, for
9123 // ADD/SUB sp = sp + immediate.
9124 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9129 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9133 if (Result.getNode()) {
9134 Ops.push_back(Result);
9137 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9141 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9142 // The ARM target isn't yet aware of offsets.
9146 bool ARM::isBitFieldInvertedMask(unsigned v) {
9147 if (v == 0xffffffff)
9149 // there can be 1's on either or both "outsides", all the "inside"
9151 unsigned int lsb = 0, msb = 31;
9152 while (v & (1 << msb)) --msb;
9153 while (v & (1 << lsb)) ++lsb;
9154 for (unsigned int i = lsb; i <= msb; ++i) {
9161 /// isFPImmLegal - Returns true if the target can instruction select the
9162 /// specified FP immediate natively. If false, the legalizer will
9163 /// materialize the FP immediate as a load from a constant pool.
9164 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9165 if (!Subtarget->hasVFP3())
9168 return ARM_AM::getFP32Imm(Imm) != -1;
9170 return ARM_AM::getFP64Imm(Imm) != -1;
9174 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9175 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9176 /// specified in the intrinsic calls.
9177 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9179 unsigned Intrinsic) const {
9180 switch (Intrinsic) {
9181 case Intrinsic::arm_neon_vld1:
9182 case Intrinsic::arm_neon_vld2:
9183 case Intrinsic::arm_neon_vld3:
9184 case Intrinsic::arm_neon_vld4:
9185 case Intrinsic::arm_neon_vld2lane:
9186 case Intrinsic::arm_neon_vld3lane:
9187 case Intrinsic::arm_neon_vld4lane: {
9188 Info.opc = ISD::INTRINSIC_W_CHAIN;
9189 // Conservatively set memVT to the entire set of vectors loaded.
9190 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9191 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9192 Info.ptrVal = I.getArgOperand(0);
9194 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9195 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9196 Info.vol = false; // volatile loads with NEON intrinsics not supported
9197 Info.readMem = true;
9198 Info.writeMem = false;
9201 case Intrinsic::arm_neon_vst1:
9202 case Intrinsic::arm_neon_vst2:
9203 case Intrinsic::arm_neon_vst3:
9204 case Intrinsic::arm_neon_vst4:
9205 case Intrinsic::arm_neon_vst2lane:
9206 case Intrinsic::arm_neon_vst3lane:
9207 case Intrinsic::arm_neon_vst4lane: {
9208 Info.opc = ISD::INTRINSIC_VOID;
9209 // Conservatively set memVT to the entire set of vectors stored.
9210 unsigned NumElts = 0;
9211 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9212 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9213 if (!ArgTy->isVectorTy())
9215 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9217 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9218 Info.ptrVal = I.getArgOperand(0);
9220 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9221 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9222 Info.vol = false; // volatile stores with NEON intrinsics not supported
9223 Info.readMem = false;
9224 Info.writeMem = true;
9227 case Intrinsic::arm_strexd: {
9228 Info.opc = ISD::INTRINSIC_W_CHAIN;
9229 Info.memVT = MVT::i64;
9230 Info.ptrVal = I.getArgOperand(2);
9234 Info.readMem = false;
9235 Info.writeMem = true;
9238 case Intrinsic::arm_ldrexd: {
9239 Info.opc = ISD::INTRINSIC_W_CHAIN;
9240 Info.memVT = MVT::i64;
9241 Info.ptrVal = I.getArgOperand(0);
9245 Info.readMem = true;
9246 Info.writeMem = false;