1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "MCTargetDesc/ARMAddressingModes.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/IntrinsicLowering.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/ADT/StringExtras.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
59 // This option should go away when tail calls fully work.
61 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
62 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
66 EnableARMLongCalls("arm-long-calls", cl::Hidden,
67 cl::desc("Generate calls via indirect call instructions"),
71 ARMInterworking("arm-interworking", cl::Hidden,
72 cl::desc("Enable / disable ARM interworking (for debugging only)"),
76 class ARMCCState : public CCState {
78 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
79 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
80 LLVMContext &C, ParmContext PC)
81 : CCState(CC, isVarArg, MF, TM, locs, C) {
82 assert(((PC == Call) || (PC == Prologue)) &&
83 "ARMCCState users must specify whether their context is call"
84 "or prologue generation.");
90 // The APCS parameter registers.
91 static const unsigned GPRArgRegs[] = {
92 ARM::R0, ARM::R1, ARM::R2, ARM::R3
95 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
96 EVT PromotedBitwiseVT) {
97 if (VT != PromotedLdStVT) {
98 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
99 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
100 PromotedLdStVT.getSimpleVT());
102 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
104 PromotedLdStVT.getSimpleVT());
107 EVT ElemTy = VT.getVectorElementType();
108 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
109 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
112 if (ElemTy == MVT::i32) {
113 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
114 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
115 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
120 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
121 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
123 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
124 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
125 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
126 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
127 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
128 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
129 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
130 if (VT.isInteger()) {
131 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
132 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
133 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
136 // Promote all bit-wise operations.
137 if (VT.isInteger() && VT != PromotedBitwiseVT) {
138 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
139 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
140 PromotedBitwiseVT.getSimpleVT());
141 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
142 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
143 PromotedBitwiseVT.getSimpleVT());
144 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
145 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
146 PromotedBitwiseVT.getSimpleVT());
149 // Neon does not support vector divide/remainder operations.
150 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
154 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
155 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
158 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
159 addRegisterClass(VT, ARM::DPRRegisterClass);
160 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
163 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
164 addRegisterClass(VT, ARM::QPRRegisterClass);
165 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
168 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
169 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
170 return new TargetLoweringObjectFileMachO();
172 return new ARMElfTargetObjectFile();
175 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
176 : TargetLowering(TM, createTLOF(TM)) {
177 Subtarget = &TM.getSubtarget<ARMSubtarget>();
178 RegInfo = TM.getRegisterInfo();
179 Itins = TM.getInstrItineraryData();
181 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
183 if (Subtarget->isTargetDarwin()) {
184 // Uses VFP for Thumb libfuncs if available.
185 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
186 // Single-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
188 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
189 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
190 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
192 // Double-precision floating-point arithmetic.
193 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
194 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
195 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
196 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
198 // Single-precision comparisons.
199 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
200 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
201 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
202 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
203 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
204 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
205 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
206 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
208 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
217 // Double-precision comparisons.
218 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
219 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
220 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
221 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
222 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
223 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
224 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
225 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
227 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
233 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
234 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
236 // Floating-point to integer conversions.
237 // i64 conversions are done via library routines even when generating VFP
238 // instructions, so use the same ones.
239 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
240 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
241 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
242 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
244 // Conversions between floating types.
245 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
246 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
248 // Integer to floating-point conversions.
249 // i64 conversions are done via library routines even when generating VFP
250 // instructions, so use the same ones.
251 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
252 // e.g., __floatunsidf vs. __floatunssidfvfp.
253 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
254 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
255 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
256 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
260 // These libcalls are not available in 32-bit.
261 setLibcallName(RTLIB::SHL_I128, 0);
262 setLibcallName(RTLIB::SRL_I128, 0);
263 setLibcallName(RTLIB::SRA_I128, 0);
265 if (Subtarget->isAAPCS_ABI()) {
266 // Double-precision floating-point arithmetic helper functions
267 // RTABI chapter 4.1.2, Table 2
268 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
269 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
270 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
271 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
272 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
277 // Double-precision floating-point comparison helper functions
278 // RTABI chapter 4.1.2, Table 3
279 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
280 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
281 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
282 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
283 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
284 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
286 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
287 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
288 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
289 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
290 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
291 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
292 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
293 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
294 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
295 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
301 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
302 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
304 // Single-precision floating-point arithmetic helper functions
305 // RTABI chapter 4.1.2, Table 4
306 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
307 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
308 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
309 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
310 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
315 // Single-precision floating-point comparison helper functions
316 // RTABI chapter 4.1.2, Table 5
317 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
318 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
319 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
320 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
321 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
322 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
324 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
325 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
326 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
327 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
328 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
329 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
330 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
331 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
332 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
333 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
339 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
340 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
342 // Floating-point to integer conversions.
343 // RTABI chapter 4.1.2, Table 6
344 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
345 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
346 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
347 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
348 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
349 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
350 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
351 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
352 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
361 // Conversions between floating types.
362 // RTABI chapter 4.1.2, Table 7
363 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
364 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
365 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
366 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
368 // Integer to floating-point conversions.
369 // RTABI chapter 4.1.2, Table 8
370 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
371 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
372 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
373 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
374 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
375 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
376 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
377 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
378 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
387 // Long long helper functions
388 // RTABI chapter 4.2, Table 9
389 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
390 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
391 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
392 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
393 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
394 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
395 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
399 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
400 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
402 // Integer division functions
403 // RTABI chapter 4.3.1
404 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
405 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
406 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
407 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
408 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
409 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
410 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
415 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
418 // RTABI chapter 4.3.4
419 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
420 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
421 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
424 // Use divmod compiler-rt calls for iOS 5.0 and later.
425 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
426 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
427 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
428 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
431 if (Subtarget->isThumb1Only())
432 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
434 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
435 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
436 !Subtarget->isThumb1Only()) {
437 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
438 if (!Subtarget->isFPOnlySP())
439 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
441 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
444 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
445 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
446 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
447 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
448 setTruncStoreAction((MVT::SimpleValueType)VT,
449 (MVT::SimpleValueType)InnerVT, Expand);
450 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
452 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
455 if (Subtarget->hasNEON()) {
456 addDRTypeForNEON(MVT::v2f32);
457 addDRTypeForNEON(MVT::v8i8);
458 addDRTypeForNEON(MVT::v4i16);
459 addDRTypeForNEON(MVT::v2i32);
460 addDRTypeForNEON(MVT::v1i64);
462 addQRTypeForNEON(MVT::v4f32);
463 addQRTypeForNEON(MVT::v2f64);
464 addQRTypeForNEON(MVT::v16i8);
465 addQRTypeForNEON(MVT::v8i16);
466 addQRTypeForNEON(MVT::v4i32);
467 addQRTypeForNEON(MVT::v2i64);
469 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
470 // neither Neon nor VFP support any arithmetic operations on it.
471 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
472 // supported for v4f32.
473 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
474 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
475 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
476 // FIXME: Code duplication: FDIV and FREM are expanded always, see
477 // ARMTargetLowering::addTypeForNEON method for details.
478 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
479 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
480 // FIXME: Create unittest.
481 // In another words, find a way when "copysign" appears in DAG with vector
483 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
484 // FIXME: Code duplication: SETCC has custom operation action, see
485 // ARMTargetLowering::addTypeForNEON method for details.
486 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
487 // FIXME: Create unittest for FNEG and for FABS.
488 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
489 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
490 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
492 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
493 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
495 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
498 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
500 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
501 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
502 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
503 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
504 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
507 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
508 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
509 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
510 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
512 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
515 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
518 // Neon does not support some operations on v1i64 and v2i64 types.
519 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
520 // Custom handling for some quad-vector types to detect VMULL.
521 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
522 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
523 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
524 // Custom handling for some vector types to avoid expensive expansions
525 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
526 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
527 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
528 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
529 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
530 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
531 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
532 // a destination type that is wider than the source.
533 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
534 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
536 setTargetDAGCombine(ISD::INTRINSIC_VOID);
537 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
538 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
539 setTargetDAGCombine(ISD::SHL);
540 setTargetDAGCombine(ISD::SRL);
541 setTargetDAGCombine(ISD::SRA);
542 setTargetDAGCombine(ISD::SIGN_EXTEND);
543 setTargetDAGCombine(ISD::ZERO_EXTEND);
544 setTargetDAGCombine(ISD::ANY_EXTEND);
545 setTargetDAGCombine(ISD::SELECT_CC);
546 setTargetDAGCombine(ISD::BUILD_VECTOR);
547 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
548 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
549 setTargetDAGCombine(ISD::STORE);
550 setTargetDAGCombine(ISD::FP_TO_SINT);
551 setTargetDAGCombine(ISD::FP_TO_UINT);
552 setTargetDAGCombine(ISD::FDIV);
554 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
557 computeRegisterProperties();
559 // ARM does not have f32 extending load.
560 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
562 // ARM does not have i1 sign extending load.
563 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
565 // ARM supports all 4 flavors of integer indexed load / store.
566 if (!Subtarget->isThumb1Only()) {
567 for (unsigned im = (unsigned)ISD::PRE_INC;
568 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
569 setIndexedLoadAction(im, MVT::i1, Legal);
570 setIndexedLoadAction(im, MVT::i8, Legal);
571 setIndexedLoadAction(im, MVT::i16, Legal);
572 setIndexedLoadAction(im, MVT::i32, Legal);
573 setIndexedStoreAction(im, MVT::i1, Legal);
574 setIndexedStoreAction(im, MVT::i8, Legal);
575 setIndexedStoreAction(im, MVT::i16, Legal);
576 setIndexedStoreAction(im, MVT::i32, Legal);
580 // i64 operation support.
581 setOperationAction(ISD::MUL, MVT::i64, Expand);
582 setOperationAction(ISD::MULHU, MVT::i32, Expand);
583 if (Subtarget->isThumb1Only()) {
584 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
585 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
587 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
588 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
589 setOperationAction(ISD::MULHS, MVT::i32, Expand);
591 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
592 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
593 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
594 setOperationAction(ISD::SRL, MVT::i64, Custom);
595 setOperationAction(ISD::SRA, MVT::i64, Custom);
597 if (!Subtarget->isThumb1Only()) {
598 // FIXME: We should do this for Thumb1 as well.
599 setOperationAction(ISD::ADDC, MVT::i32, Custom);
600 setOperationAction(ISD::ADDE, MVT::i32, Custom);
601 setOperationAction(ISD::SUBC, MVT::i32, Custom);
602 setOperationAction(ISD::SUBE, MVT::i32, Custom);
605 // ARM does not have ROTL.
606 setOperationAction(ISD::ROTL, MVT::i32, Expand);
607 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
608 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
609 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
610 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
612 // These just redirect to CTTZ and CTLZ on ARM.
613 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
614 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
616 // Only ARMv6 has BSWAP.
617 if (!Subtarget->hasV6Ops())
618 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
620 // These are expanded into libcalls.
621 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
622 // v7M has a hardware divider
623 setOperationAction(ISD::SDIV, MVT::i32, Expand);
624 setOperationAction(ISD::UDIV, MVT::i32, Expand);
626 setOperationAction(ISD::SREM, MVT::i32, Expand);
627 setOperationAction(ISD::UREM, MVT::i32, Expand);
628 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
629 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
631 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
632 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
633 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
634 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
635 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
637 setOperationAction(ISD::TRAP, MVT::Other, Legal);
639 // Use the default implementation.
640 setOperationAction(ISD::VASTART, MVT::Other, Custom);
641 setOperationAction(ISD::VAARG, MVT::Other, Expand);
642 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
643 setOperationAction(ISD::VAEND, MVT::Other, Expand);
644 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
645 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
646 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
647 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
648 setExceptionPointerRegister(ARM::R0);
649 setExceptionSelectorRegister(ARM::R1);
651 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
652 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
653 // the default expansion.
654 // FIXME: This should be checking for v6k, not just v6.
655 if (Subtarget->hasDataBarrier() ||
656 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
657 // membarrier needs custom lowering; the rest are legal and handled
659 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
660 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
661 // Custom lowering for 64-bit ops
662 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
663 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
664 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
665 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
666 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
667 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
668 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
669 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
670 setInsertFencesForAtomic(true);
672 // Set them all for expansion, which will force libcalls.
673 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
674 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
675 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
676 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
677 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
678 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
679 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
680 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
681 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
682 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
683 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
684 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
685 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
686 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
687 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
688 // Unordered/Monotonic case.
689 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
690 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
691 // Since the libcalls include locking, fold in the fences
692 setShouldFoldAtomicFences(true);
695 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
697 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
698 if (!Subtarget->hasV6Ops()) {
699 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
700 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
702 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
704 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
705 !Subtarget->isThumb1Only()) {
706 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
707 // iff target supports vfp2.
708 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
709 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
712 // We want to custom lower some of our intrinsics.
713 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
714 if (Subtarget->isTargetDarwin()) {
715 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
716 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
717 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
720 setOperationAction(ISD::SETCC, MVT::i32, Expand);
721 setOperationAction(ISD::SETCC, MVT::f32, Expand);
722 setOperationAction(ISD::SETCC, MVT::f64, Expand);
723 setOperationAction(ISD::SELECT, MVT::i32, Custom);
724 setOperationAction(ISD::SELECT, MVT::f32, Custom);
725 setOperationAction(ISD::SELECT, MVT::f64, Custom);
726 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
727 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
728 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
730 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
731 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
732 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
733 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
734 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
736 // We don't support sin/cos/fmod/copysign/pow
737 setOperationAction(ISD::FSIN, MVT::f64, Expand);
738 setOperationAction(ISD::FSIN, MVT::f32, Expand);
739 setOperationAction(ISD::FCOS, MVT::f32, Expand);
740 setOperationAction(ISD::FCOS, MVT::f64, Expand);
741 setOperationAction(ISD::FREM, MVT::f64, Expand);
742 setOperationAction(ISD::FREM, MVT::f32, Expand);
743 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
744 !Subtarget->isThumb1Only()) {
745 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
746 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
748 setOperationAction(ISD::FPOW, MVT::f64, Expand);
749 setOperationAction(ISD::FPOW, MVT::f32, Expand);
751 setOperationAction(ISD::FMA, MVT::f64, Expand);
752 setOperationAction(ISD::FMA, MVT::f32, Expand);
754 // Various VFP goodness
755 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
756 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
757 if (Subtarget->hasVFP2()) {
758 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
759 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
760 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
761 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
763 // Special handling for half-precision FP.
764 if (!Subtarget->hasFP16()) {
765 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
766 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
770 // We have target-specific dag combine patterns for the following nodes:
771 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
772 setTargetDAGCombine(ISD::ADD);
773 setTargetDAGCombine(ISD::SUB);
774 setTargetDAGCombine(ISD::MUL);
776 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
777 setTargetDAGCombine(ISD::OR);
778 if (Subtarget->hasNEON())
779 setTargetDAGCombine(ISD::AND);
781 setStackPointerRegisterToSaveRestore(ARM::SP);
783 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
784 !Subtarget->hasVFP2())
785 setSchedulingPreference(Sched::RegPressure);
787 setSchedulingPreference(Sched::Hybrid);
789 //// temporary - rewrite interface to use type
790 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
791 maxStoresPerMemset = 16;
792 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
794 // On ARM arguments smaller than 4 bytes are extended, so all arguments
795 // are at least 4 bytes aligned.
796 setMinStackArgumentAlignment(4);
798 benefitFromCodePlacementOpt = true;
800 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
803 // FIXME: It might make sense to define the representative register class as the
804 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
805 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
806 // SPR's representative would be DPR_VFP2. This should work well if register
807 // pressure tracking were modified such that a register use would increment the
808 // pressure of the register class's representative and all of it's super
809 // classes' representatives transitively. We have not implemented this because
810 // of the difficulty prior to coalescing of modeling operand register classes
811 // due to the common occurrence of cross class copies and subregister insertions
813 std::pair<const TargetRegisterClass*, uint8_t>
814 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
815 const TargetRegisterClass *RRC = 0;
817 switch (VT.getSimpleVT().SimpleTy) {
819 return TargetLowering::findRepresentativeClass(VT);
820 // Use DPR as representative register class for all floating point
821 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
822 // the cost is 1 for both f32 and f64.
823 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
824 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
825 RRC = ARM::DPRRegisterClass;
826 // When NEON is used for SP, only half of the register file is available
827 // because operations that define both SP and DP results will be constrained
828 // to the VFP2 class (D0-D15). We currently model this constraint prior to
829 // coalescing by double-counting the SP regs. See the FIXME above.
830 if (Subtarget->useNEONForSinglePrecisionFP())
833 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
834 case MVT::v4f32: case MVT::v2f64:
835 RRC = ARM::DPRRegisterClass;
839 RRC = ARM::DPRRegisterClass;
843 RRC = ARM::DPRRegisterClass;
847 return std::make_pair(RRC, Cost);
850 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
853 case ARMISD::Wrapper: return "ARMISD::Wrapper";
854 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
855 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
856 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
857 case ARMISD::CALL: return "ARMISD::CALL";
858 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
859 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
860 case ARMISD::tCALL: return "ARMISD::tCALL";
861 case ARMISD::BRCOND: return "ARMISD::BRCOND";
862 case ARMISD::BR_JT: return "ARMISD::BR_JT";
863 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
864 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
865 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
866 case ARMISD::CMP: return "ARMISD::CMP";
867 case ARMISD::CMPZ: return "ARMISD::CMPZ";
868 case ARMISD::CMPFP: return "ARMISD::CMPFP";
869 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
870 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
871 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
872 case ARMISD::CMOV: return "ARMISD::CMOV";
874 case ARMISD::RBIT: return "ARMISD::RBIT";
876 case ARMISD::FTOSI: return "ARMISD::FTOSI";
877 case ARMISD::FTOUI: return "ARMISD::FTOUI";
878 case ARMISD::SITOF: return "ARMISD::SITOF";
879 case ARMISD::UITOF: return "ARMISD::UITOF";
881 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
882 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
883 case ARMISD::RRX: return "ARMISD::RRX";
885 case ARMISD::ADDC: return "ARMISD::ADDC";
886 case ARMISD::ADDE: return "ARMISD::ADDE";
887 case ARMISD::SUBC: return "ARMISD::SUBC";
888 case ARMISD::SUBE: return "ARMISD::SUBE";
890 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
891 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
893 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
894 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
896 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
898 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
900 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
902 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
903 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
905 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
907 case ARMISD::VCEQ: return "ARMISD::VCEQ";
908 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
909 case ARMISD::VCGE: return "ARMISD::VCGE";
910 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
911 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
912 case ARMISD::VCGEU: return "ARMISD::VCGEU";
913 case ARMISD::VCGT: return "ARMISD::VCGT";
914 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
915 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
916 case ARMISD::VCGTU: return "ARMISD::VCGTU";
917 case ARMISD::VTST: return "ARMISD::VTST";
919 case ARMISD::VSHL: return "ARMISD::VSHL";
920 case ARMISD::VSHRs: return "ARMISD::VSHRs";
921 case ARMISD::VSHRu: return "ARMISD::VSHRu";
922 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
923 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
924 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
925 case ARMISD::VSHRN: return "ARMISD::VSHRN";
926 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
927 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
928 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
929 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
930 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
931 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
932 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
933 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
934 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
935 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
936 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
937 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
938 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
939 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
940 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
941 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
942 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
943 case ARMISD::VDUP: return "ARMISD::VDUP";
944 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
945 case ARMISD::VEXT: return "ARMISD::VEXT";
946 case ARMISD::VREV64: return "ARMISD::VREV64";
947 case ARMISD::VREV32: return "ARMISD::VREV32";
948 case ARMISD::VREV16: return "ARMISD::VREV16";
949 case ARMISD::VZIP: return "ARMISD::VZIP";
950 case ARMISD::VUZP: return "ARMISD::VUZP";
951 case ARMISD::VTRN: return "ARMISD::VTRN";
952 case ARMISD::VTBL1: return "ARMISD::VTBL1";
953 case ARMISD::VTBL2: return "ARMISD::VTBL2";
954 case ARMISD::VMULLs: return "ARMISD::VMULLs";
955 case ARMISD::VMULLu: return "ARMISD::VMULLu";
956 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
957 case ARMISD::FMAX: return "ARMISD::FMAX";
958 case ARMISD::FMIN: return "ARMISD::FMIN";
959 case ARMISD::BFI: return "ARMISD::BFI";
960 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
961 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
962 case ARMISD::VBSL: return "ARMISD::VBSL";
963 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
964 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
965 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
966 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
967 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
968 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
969 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
970 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
971 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
972 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
973 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
974 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
975 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
976 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
977 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
978 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
979 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
980 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
981 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
982 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
986 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
987 if (!VT.isVector()) return getPointerTy();
988 return VT.changeVectorElementTypeToInteger();
991 /// getRegClassFor - Return the register class that should be used for the
992 /// specified value type.
993 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
994 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
995 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
996 // load / store 4 to 8 consecutive D registers.
997 if (Subtarget->hasNEON()) {
998 if (VT == MVT::v4i64)
999 return ARM::QQPRRegisterClass;
1000 else if (VT == MVT::v8i64)
1001 return ARM::QQQQPRRegisterClass;
1003 return TargetLowering::getRegClassFor(VT);
1006 // Create a fast isel object.
1008 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1009 return ARM::createFastISel(funcInfo);
1012 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1013 /// be used for loads / stores from the global.
1014 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1015 return (Subtarget->isThumb1Only() ? 127 : 4095);
1018 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1019 unsigned NumVals = N->getNumValues();
1021 return Sched::RegPressure;
1023 for (unsigned i = 0; i != NumVals; ++i) {
1024 EVT VT = N->getValueType(i);
1025 if (VT == MVT::Glue || VT == MVT::Other)
1027 if (VT.isFloatingPoint() || VT.isVector())
1031 if (!N->isMachineOpcode())
1032 return Sched::RegPressure;
1034 // Load are scheduled for latency even if there instruction itinerary
1035 // is not available.
1036 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1037 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1039 if (MCID.getNumDefs() == 0)
1040 return Sched::RegPressure;
1041 if (!Itins->isEmpty() &&
1042 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1045 return Sched::RegPressure;
1048 //===----------------------------------------------------------------------===//
1050 //===----------------------------------------------------------------------===//
1052 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1053 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1055 default: llvm_unreachable("Unknown condition code!");
1056 case ISD::SETNE: return ARMCC::NE;
1057 case ISD::SETEQ: return ARMCC::EQ;
1058 case ISD::SETGT: return ARMCC::GT;
1059 case ISD::SETGE: return ARMCC::GE;
1060 case ISD::SETLT: return ARMCC::LT;
1061 case ISD::SETLE: return ARMCC::LE;
1062 case ISD::SETUGT: return ARMCC::HI;
1063 case ISD::SETUGE: return ARMCC::HS;
1064 case ISD::SETULT: return ARMCC::LO;
1065 case ISD::SETULE: return ARMCC::LS;
1069 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1070 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1071 ARMCC::CondCodes &CondCode2) {
1072 CondCode2 = ARMCC::AL;
1074 default: llvm_unreachable("Unknown FP condition!");
1076 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1078 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1080 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1081 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1082 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1083 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1084 case ISD::SETO: CondCode = ARMCC::VC; break;
1085 case ISD::SETUO: CondCode = ARMCC::VS; break;
1086 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1087 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1088 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1090 case ISD::SETULT: CondCode = ARMCC::LT; break;
1092 case ISD::SETULE: CondCode = ARMCC::LE; break;
1094 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1098 //===----------------------------------------------------------------------===//
1099 // Calling Convention Implementation
1100 //===----------------------------------------------------------------------===//
1102 #include "ARMGenCallingConv.inc"
1104 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1105 /// given CallingConvention value.
1106 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1108 bool isVarArg) const {
1111 llvm_unreachable("Unsupported calling convention");
1112 case CallingConv::Fast:
1113 if (Subtarget->hasVFP2() && !isVarArg) {
1114 if (!Subtarget->isAAPCS_ABI())
1115 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1116 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1117 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1120 case CallingConv::C: {
1121 // Use target triple & subtarget features to do actual dispatch.
1122 if (!Subtarget->isAAPCS_ABI())
1123 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1124 else if (Subtarget->hasVFP2() &&
1125 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1127 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1128 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1130 case CallingConv::ARM_AAPCS_VFP:
1131 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1132 case CallingConv::ARM_AAPCS:
1133 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1134 case CallingConv::ARM_APCS:
1135 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1139 /// LowerCallResult - Lower the result values of a call into the
1140 /// appropriate copies out of appropriate physical registers.
1142 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1143 CallingConv::ID CallConv, bool isVarArg,
1144 const SmallVectorImpl<ISD::InputArg> &Ins,
1145 DebugLoc dl, SelectionDAG &DAG,
1146 SmallVectorImpl<SDValue> &InVals) const {
1148 // Assign locations to each value returned by this call.
1149 SmallVector<CCValAssign, 16> RVLocs;
1150 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1151 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1152 CCInfo.AnalyzeCallResult(Ins,
1153 CCAssignFnForNode(CallConv, /* Return*/ true,
1156 // Copy all of the result registers out of their specified physreg.
1157 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1158 CCValAssign VA = RVLocs[i];
1161 if (VA.needsCustom()) {
1162 // Handle f64 or half of a v2f64.
1163 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1165 Chain = Lo.getValue(1);
1166 InFlag = Lo.getValue(2);
1167 VA = RVLocs[++i]; // skip ahead to next loc
1168 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1170 Chain = Hi.getValue(1);
1171 InFlag = Hi.getValue(2);
1172 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1174 if (VA.getLocVT() == MVT::v2f64) {
1175 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1176 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1177 DAG.getConstant(0, MVT::i32));
1179 VA = RVLocs[++i]; // skip ahead to next loc
1180 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1181 Chain = Lo.getValue(1);
1182 InFlag = Lo.getValue(2);
1183 VA = RVLocs[++i]; // skip ahead to next loc
1184 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1185 Chain = Hi.getValue(1);
1186 InFlag = Hi.getValue(2);
1187 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1188 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1189 DAG.getConstant(1, MVT::i32));
1192 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1194 Chain = Val.getValue(1);
1195 InFlag = Val.getValue(2);
1198 switch (VA.getLocInfo()) {
1199 default: llvm_unreachable("Unknown loc info!");
1200 case CCValAssign::Full: break;
1201 case CCValAssign::BCvt:
1202 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1206 InVals.push_back(Val);
1212 /// LowerMemOpCallTo - Store the argument to the stack.
1214 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1215 SDValue StackPtr, SDValue Arg,
1216 DebugLoc dl, SelectionDAG &DAG,
1217 const CCValAssign &VA,
1218 ISD::ArgFlagsTy Flags) const {
1219 unsigned LocMemOffset = VA.getLocMemOffset();
1220 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1221 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1222 return DAG.getStore(Chain, dl, Arg, PtrOff,
1223 MachinePointerInfo::getStack(LocMemOffset),
1227 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1228 SDValue Chain, SDValue &Arg,
1229 RegsToPassVector &RegsToPass,
1230 CCValAssign &VA, CCValAssign &NextVA,
1232 SmallVector<SDValue, 8> &MemOpChains,
1233 ISD::ArgFlagsTy Flags) const {
1235 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1236 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1237 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1239 if (NextVA.isRegLoc())
1240 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1242 assert(NextVA.isMemLoc());
1243 if (StackPtr.getNode() == 0)
1244 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1246 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1252 /// LowerCall - Lowering a call into a callseq_start <-
1253 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1256 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1257 CallingConv::ID CallConv, bool isVarArg,
1259 const SmallVectorImpl<ISD::OutputArg> &Outs,
1260 const SmallVectorImpl<SDValue> &OutVals,
1261 const SmallVectorImpl<ISD::InputArg> &Ins,
1262 DebugLoc dl, SelectionDAG &DAG,
1263 SmallVectorImpl<SDValue> &InVals) const {
1264 MachineFunction &MF = DAG.getMachineFunction();
1265 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1266 bool IsSibCall = false;
1267 // Disable tail calls if they're not supported.
1268 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1271 // Check if it's really possible to do a tail call.
1272 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1273 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1274 Outs, OutVals, Ins, DAG);
1275 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1276 // detected sibcalls.
1283 // Analyze operands of the call, assigning locations to each operand.
1284 SmallVector<CCValAssign, 16> ArgLocs;
1285 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1286 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1287 CCInfo.AnalyzeCallOperands(Outs,
1288 CCAssignFnForNode(CallConv, /* Return*/ false,
1291 // Get a count of how many bytes are to be pushed on the stack.
1292 unsigned NumBytes = CCInfo.getNextStackOffset();
1294 // For tail calls, memory operands are available in our caller's stack.
1298 // Adjust the stack pointer for the new arguments...
1299 // These operations are automatically eliminated by the prolog/epilog pass
1301 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1303 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1305 RegsToPassVector RegsToPass;
1306 SmallVector<SDValue, 8> MemOpChains;
1308 // Walk the register/memloc assignments, inserting copies/loads. In the case
1309 // of tail call optimization, arguments are handled later.
1310 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1312 ++i, ++realArgIdx) {
1313 CCValAssign &VA = ArgLocs[i];
1314 SDValue Arg = OutVals[realArgIdx];
1315 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1316 bool isByVal = Flags.isByVal();
1318 // Promote the value if needed.
1319 switch (VA.getLocInfo()) {
1320 default: llvm_unreachable("Unknown loc info!");
1321 case CCValAssign::Full: break;
1322 case CCValAssign::SExt:
1323 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1325 case CCValAssign::ZExt:
1326 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1328 case CCValAssign::AExt:
1329 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1331 case CCValAssign::BCvt:
1332 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1336 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1337 if (VA.needsCustom()) {
1338 if (VA.getLocVT() == MVT::v2f64) {
1339 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1340 DAG.getConstant(0, MVT::i32));
1341 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1342 DAG.getConstant(1, MVT::i32));
1344 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1345 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1347 VA = ArgLocs[++i]; // skip ahead to next loc
1348 if (VA.isRegLoc()) {
1349 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1350 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1352 assert(VA.isMemLoc());
1354 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1355 dl, DAG, VA, Flags));
1358 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1359 StackPtr, MemOpChains, Flags);
1361 } else if (VA.isRegLoc()) {
1362 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1363 } else if (isByVal) {
1364 assert(VA.isMemLoc());
1365 unsigned offset = 0;
1367 // True if this byval aggregate will be split between registers
1369 if (CCInfo.isFirstByValRegValid()) {
1370 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1372 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1373 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1374 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1375 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1376 MachinePointerInfo(),
1377 false, false, false, 0);
1378 MemOpChains.push_back(Load.getValue(1));
1379 RegsToPass.push_back(std::make_pair(j, Load));
1381 offset = ARM::R4 - CCInfo.getFirstByValReg();
1382 CCInfo.clearFirstByValReg();
1385 unsigned LocMemOffset = VA.getLocMemOffset();
1386 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1387 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1389 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1390 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1391 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1393 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1394 Flags.getByValAlign(),
1395 /*isVolatile=*/false,
1396 /*AlwaysInline=*/false,
1397 MachinePointerInfo(0),
1398 MachinePointerInfo(0)));
1400 } else if (!IsSibCall) {
1401 assert(VA.isMemLoc());
1403 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1404 dl, DAG, VA, Flags));
1408 if (!MemOpChains.empty())
1409 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1410 &MemOpChains[0], MemOpChains.size());
1412 // Build a sequence of copy-to-reg nodes chained together with token chain
1413 // and flag operands which copy the outgoing args into the appropriate regs.
1415 // Tail call byval lowering might overwrite argument registers so in case of
1416 // tail call optimization the copies to registers are lowered later.
1418 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1419 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1420 RegsToPass[i].second, InFlag);
1421 InFlag = Chain.getValue(1);
1424 // For tail calls lower the arguments to the 'real' stack slot.
1426 // Force all the incoming stack arguments to be loaded from the stack
1427 // before any new outgoing arguments are stored to the stack, because the
1428 // outgoing stack slots may alias the incoming argument stack slots, and
1429 // the alias isn't otherwise explicit. This is slightly more conservative
1430 // than necessary, because it means that each store effectively depends
1431 // on every argument instead of just those arguments it would clobber.
1433 // Do not flag preceding copytoreg stuff together with the following stuff.
1435 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1436 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1437 RegsToPass[i].second, InFlag);
1438 InFlag = Chain.getValue(1);
1443 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1444 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1445 // node so that legalize doesn't hack it.
1446 bool isDirect = false;
1447 bool isARMFunc = false;
1448 bool isLocalARMFunc = false;
1449 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1451 if (EnableARMLongCalls) {
1452 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1453 && "long-calls with non-static relocation model!");
1454 // Handle a global address or an external symbol. If it's not one of
1455 // those, the target's already in a register, so we don't need to do
1457 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1458 const GlobalValue *GV = G->getGlobal();
1459 // Create a constant pool entry for the callee address
1460 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1461 ARMConstantPoolValue *CPV =
1462 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1464 // Get the address of the callee into a register
1465 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1466 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1467 Callee = DAG.getLoad(getPointerTy(), dl,
1468 DAG.getEntryNode(), CPAddr,
1469 MachinePointerInfo::getConstantPool(),
1470 false, false, false, 0);
1471 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1472 const char *Sym = S->getSymbol();
1474 // Create a constant pool entry for the callee address
1475 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1476 ARMConstantPoolValue *CPV =
1477 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1478 ARMPCLabelIndex, 0);
1479 // Get the address of the callee into a register
1480 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1481 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1482 Callee = DAG.getLoad(getPointerTy(), dl,
1483 DAG.getEntryNode(), CPAddr,
1484 MachinePointerInfo::getConstantPool(),
1485 false, false, false, 0);
1487 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1488 const GlobalValue *GV = G->getGlobal();
1490 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1491 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1492 getTargetMachine().getRelocationModel() != Reloc::Static;
1493 isARMFunc = !Subtarget->isThumb() || isStub;
1494 // ARM call to a local ARM function is predicable.
1495 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1496 // tBX takes a register source operand.
1497 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1498 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1499 ARMConstantPoolValue *CPV =
1500 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1501 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1502 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1503 Callee = DAG.getLoad(getPointerTy(), dl,
1504 DAG.getEntryNode(), CPAddr,
1505 MachinePointerInfo::getConstantPool(),
1506 false, false, false, 0);
1507 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1508 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1509 getPointerTy(), Callee, PICLabel);
1511 // On ELF targets for PIC code, direct calls should go through the PLT
1512 unsigned OpFlags = 0;
1513 if (Subtarget->isTargetELF() &&
1514 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1515 OpFlags = ARMII::MO_PLT;
1516 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1518 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1520 bool isStub = Subtarget->isTargetDarwin() &&
1521 getTargetMachine().getRelocationModel() != Reloc::Static;
1522 isARMFunc = !Subtarget->isThumb() || isStub;
1523 // tBX takes a register source operand.
1524 const char *Sym = S->getSymbol();
1525 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1526 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1527 ARMConstantPoolValue *CPV =
1528 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1529 ARMPCLabelIndex, 4);
1530 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1531 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1532 Callee = DAG.getLoad(getPointerTy(), dl,
1533 DAG.getEntryNode(), CPAddr,
1534 MachinePointerInfo::getConstantPool(),
1535 false, false, false, 0);
1536 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1537 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1538 getPointerTy(), Callee, PICLabel);
1540 unsigned OpFlags = 0;
1541 // On ELF targets for PIC code, direct calls should go through the PLT
1542 if (Subtarget->isTargetELF() &&
1543 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1544 OpFlags = ARMII::MO_PLT;
1545 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1549 // FIXME: handle tail calls differently.
1551 if (Subtarget->isThumb()) {
1552 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1553 CallOpc = ARMISD::CALL_NOLINK;
1555 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1557 CallOpc = (isDirect || Subtarget->hasV5TOps())
1558 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1559 : ARMISD::CALL_NOLINK;
1562 std::vector<SDValue> Ops;
1563 Ops.push_back(Chain);
1564 Ops.push_back(Callee);
1566 // Add argument registers to the end of the list so that they are known live
1568 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1569 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1570 RegsToPass[i].second.getValueType()));
1572 if (InFlag.getNode())
1573 Ops.push_back(InFlag);
1575 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1577 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1579 // Returns a chain and a flag for retval copy to use.
1580 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1581 InFlag = Chain.getValue(1);
1583 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1584 DAG.getIntPtrConstant(0, true), InFlag);
1586 InFlag = Chain.getValue(1);
1588 // Handle result values, copying them out of physregs into vregs that we
1590 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1594 /// HandleByVal - Every parameter *after* a byval parameter is passed
1595 /// on the stack. Remember the next parameter register to allocate,
1596 /// and then confiscate the rest of the parameter registers to insure
1599 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1600 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1601 assert((State->getCallOrPrologue() == Prologue ||
1602 State->getCallOrPrologue() == Call) &&
1603 "unhandled ParmContext");
1604 if ((!State->isFirstByValRegValid()) &&
1605 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1606 State->setFirstByValReg(reg);
1607 // At a call site, a byval parameter that is split between
1608 // registers and memory needs its size truncated here. In a
1609 // function prologue, such byval parameters are reassembled in
1610 // memory, and are not truncated.
1611 if (State->getCallOrPrologue() == Call) {
1612 unsigned excess = 4 * (ARM::R4 - reg);
1613 assert(size >= excess && "expected larger existing stack allocation");
1617 // Confiscate any remaining parameter registers to preclude their
1618 // assignment to subsequent parameters.
1619 while (State->AllocateReg(GPRArgRegs, 4))
1623 /// MatchingStackOffset - Return true if the given stack call argument is
1624 /// already available in the same position (relatively) of the caller's
1625 /// incoming argument stack.
1627 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1628 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1629 const ARMInstrInfo *TII) {
1630 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1632 if (Arg.getOpcode() == ISD::CopyFromReg) {
1633 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1634 if (!TargetRegisterInfo::isVirtualRegister(VR))
1636 MachineInstr *Def = MRI->getVRegDef(VR);
1639 if (!Flags.isByVal()) {
1640 if (!TII->isLoadFromStackSlot(Def, FI))
1645 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1646 if (Flags.isByVal())
1647 // ByVal argument is passed in as a pointer but it's now being
1648 // dereferenced. e.g.
1649 // define @foo(%struct.X* %A) {
1650 // tail call @bar(%struct.X* byval %A)
1653 SDValue Ptr = Ld->getBasePtr();
1654 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1657 FI = FINode->getIndex();
1661 assert(FI != INT_MAX);
1662 if (!MFI->isFixedObjectIndex(FI))
1664 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1667 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1668 /// for tail call optimization. Targets which want to do tail call
1669 /// optimization should implement this function.
1671 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1672 CallingConv::ID CalleeCC,
1674 bool isCalleeStructRet,
1675 bool isCallerStructRet,
1676 const SmallVectorImpl<ISD::OutputArg> &Outs,
1677 const SmallVectorImpl<SDValue> &OutVals,
1678 const SmallVectorImpl<ISD::InputArg> &Ins,
1679 SelectionDAG& DAG) const {
1680 const Function *CallerF = DAG.getMachineFunction().getFunction();
1681 CallingConv::ID CallerCC = CallerF->getCallingConv();
1682 bool CCMatch = CallerCC == CalleeCC;
1684 // Look for obvious safe cases to perform tail call optimization that do not
1685 // require ABI changes. This is what gcc calls sibcall.
1687 // Do not sibcall optimize vararg calls unless the call site is not passing
1689 if (isVarArg && !Outs.empty())
1692 // Also avoid sibcall optimization if either caller or callee uses struct
1693 // return semantics.
1694 if (isCalleeStructRet || isCallerStructRet)
1697 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1698 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1699 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1700 // support in the assembler and linker to be used. This would need to be
1701 // fixed to fully support tail calls in Thumb1.
1703 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1704 // LR. This means if we need to reload LR, it takes an extra instructions,
1705 // which outweighs the value of the tail call; but here we don't know yet
1706 // whether LR is going to be used. Probably the right approach is to
1707 // generate the tail call here and turn it back into CALL/RET in
1708 // emitEpilogue if LR is used.
1710 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1711 // but we need to make sure there are enough registers; the only valid
1712 // registers are the 4 used for parameters. We don't currently do this
1714 if (Subtarget->isThumb1Only())
1717 // If the calling conventions do not match, then we'd better make sure the
1718 // results are returned in the same way as what the caller expects.
1720 SmallVector<CCValAssign, 16> RVLocs1;
1721 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1722 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1723 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1725 SmallVector<CCValAssign, 16> RVLocs2;
1726 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1727 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1728 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1730 if (RVLocs1.size() != RVLocs2.size())
1732 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1733 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1735 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1737 if (RVLocs1[i].isRegLoc()) {
1738 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1741 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1747 // If the callee takes no arguments then go on to check the results of the
1749 if (!Outs.empty()) {
1750 // Check if stack adjustment is needed. For now, do not do this if any
1751 // argument is passed on the stack.
1752 SmallVector<CCValAssign, 16> ArgLocs;
1753 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1754 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1755 CCInfo.AnalyzeCallOperands(Outs,
1756 CCAssignFnForNode(CalleeCC, false, isVarArg));
1757 if (CCInfo.getNextStackOffset()) {
1758 MachineFunction &MF = DAG.getMachineFunction();
1760 // Check if the arguments are already laid out in the right way as
1761 // the caller's fixed stack objects.
1762 MachineFrameInfo *MFI = MF.getFrameInfo();
1763 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1764 const ARMInstrInfo *TII =
1765 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1766 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1768 ++i, ++realArgIdx) {
1769 CCValAssign &VA = ArgLocs[i];
1770 EVT RegVT = VA.getLocVT();
1771 SDValue Arg = OutVals[realArgIdx];
1772 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1773 if (VA.getLocInfo() == CCValAssign::Indirect)
1775 if (VA.needsCustom()) {
1776 // f64 and vector types are split into multiple registers or
1777 // register/stack-slot combinations. The types will not match
1778 // the registers; give up on memory f64 refs until we figure
1779 // out what to do about this.
1782 if (!ArgLocs[++i].isRegLoc())
1784 if (RegVT == MVT::v2f64) {
1785 if (!ArgLocs[++i].isRegLoc())
1787 if (!ArgLocs[++i].isRegLoc())
1790 } else if (!VA.isRegLoc()) {
1791 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1803 ARMTargetLowering::LowerReturn(SDValue Chain,
1804 CallingConv::ID CallConv, bool isVarArg,
1805 const SmallVectorImpl<ISD::OutputArg> &Outs,
1806 const SmallVectorImpl<SDValue> &OutVals,
1807 DebugLoc dl, SelectionDAG &DAG) const {
1809 // CCValAssign - represent the assignment of the return value to a location.
1810 SmallVector<CCValAssign, 16> RVLocs;
1812 // CCState - Info about the registers and stack slots.
1813 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1814 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1816 // Analyze outgoing return values.
1817 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1820 // If this is the first return lowered for this function, add
1821 // the regs to the liveout set for the function.
1822 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1823 for (unsigned i = 0; i != RVLocs.size(); ++i)
1824 if (RVLocs[i].isRegLoc())
1825 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1830 // Copy the result values into the output registers.
1831 for (unsigned i = 0, realRVLocIdx = 0;
1833 ++i, ++realRVLocIdx) {
1834 CCValAssign &VA = RVLocs[i];
1835 assert(VA.isRegLoc() && "Can only return in registers!");
1837 SDValue Arg = OutVals[realRVLocIdx];
1839 switch (VA.getLocInfo()) {
1840 default: llvm_unreachable("Unknown loc info!");
1841 case CCValAssign::Full: break;
1842 case CCValAssign::BCvt:
1843 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1847 if (VA.needsCustom()) {
1848 if (VA.getLocVT() == MVT::v2f64) {
1849 // Extract the first half and return it in two registers.
1850 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1851 DAG.getConstant(0, MVT::i32));
1852 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1853 DAG.getVTList(MVT::i32, MVT::i32), Half);
1855 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1856 Flag = Chain.getValue(1);
1857 VA = RVLocs[++i]; // skip ahead to next loc
1858 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1859 HalfGPRs.getValue(1), Flag);
1860 Flag = Chain.getValue(1);
1861 VA = RVLocs[++i]; // skip ahead to next loc
1863 // Extract the 2nd half and fall through to handle it as an f64 value.
1864 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1865 DAG.getConstant(1, MVT::i32));
1867 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1869 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1870 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1871 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1872 Flag = Chain.getValue(1);
1873 VA = RVLocs[++i]; // skip ahead to next loc
1874 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1877 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1879 // Guarantee that all emitted copies are
1880 // stuck together, avoiding something bad.
1881 Flag = Chain.getValue(1);
1886 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1888 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1893 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1894 if (N->getNumValues() != 1)
1896 if (!N->hasNUsesOfValue(1, 0))
1899 unsigned NumCopies = 0;
1901 SDNode *Use = *N->use_begin();
1902 if (Use->getOpcode() == ISD::CopyToReg) {
1903 Copies[NumCopies++] = Use;
1904 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1905 // f64 returned in a pair of GPRs.
1906 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1908 if (UI->getOpcode() != ISD::CopyToReg)
1910 Copies[UI.getUse().getResNo()] = *UI;
1913 } else if (Use->getOpcode() == ISD::BITCAST) {
1914 // f32 returned in a single GPR.
1915 if (!Use->hasNUsesOfValue(1, 0))
1917 Use = *Use->use_begin();
1918 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1920 Copies[NumCopies++] = Use;
1925 if (NumCopies != 1 && NumCopies != 2)
1928 bool HasRet = false;
1929 for (unsigned i = 0; i < NumCopies; ++i) {
1930 SDNode *Copy = Copies[i];
1931 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1933 if (UI->getOpcode() == ISD::CopyToReg) {
1935 if (Use == Copies[0] || Use == Copies[1])
1939 if (UI->getOpcode() != ARMISD::RET_FLAG)
1948 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1949 if (!EnableARMTailCalls)
1952 if (!CI->isTailCall())
1955 return !Subtarget->isThumb1Only();
1958 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1959 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1960 // one of the above mentioned nodes. It has to be wrapped because otherwise
1961 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1962 // be used to form addressing mode. These wrapped nodes will be selected
1964 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1965 EVT PtrVT = Op.getValueType();
1966 // FIXME there is no actual debug info here
1967 DebugLoc dl = Op.getDebugLoc();
1968 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1970 if (CP->isMachineConstantPoolEntry())
1971 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1972 CP->getAlignment());
1974 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1975 CP->getAlignment());
1976 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1979 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1980 return MachineJumpTableInfo::EK_Inline;
1983 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1984 SelectionDAG &DAG) const {
1985 MachineFunction &MF = DAG.getMachineFunction();
1986 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1987 unsigned ARMPCLabelIndex = 0;
1988 DebugLoc DL = Op.getDebugLoc();
1989 EVT PtrVT = getPointerTy();
1990 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1991 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1993 if (RelocM == Reloc::Static) {
1994 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1996 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1997 ARMPCLabelIndex = AFI->createPICLabelUId();
1998 ARMConstantPoolValue *CPV =
1999 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2000 ARMCP::CPBlockAddress, PCAdj);
2001 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2003 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2004 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2005 MachinePointerInfo::getConstantPool(),
2006 false, false, false, 0);
2007 if (RelocM == Reloc::Static)
2009 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2010 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2013 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2015 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2016 SelectionDAG &DAG) const {
2017 DebugLoc dl = GA->getDebugLoc();
2018 EVT PtrVT = getPointerTy();
2019 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2020 MachineFunction &MF = DAG.getMachineFunction();
2021 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2022 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2023 ARMConstantPoolValue *CPV =
2024 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2025 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2026 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2027 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2028 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2029 MachinePointerInfo::getConstantPool(),
2030 false, false, false, 0);
2031 SDValue Chain = Argument.getValue(1);
2033 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2034 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2036 // call __tls_get_addr.
2039 Entry.Node = Argument;
2040 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2041 Args.push_back(Entry);
2042 // FIXME: is there useful debug info available here?
2043 std::pair<SDValue, SDValue> CallResult =
2044 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2045 false, false, false, false,
2046 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2047 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2048 return CallResult.first;
2051 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2052 // "local exec" model.
2054 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2055 SelectionDAG &DAG) const {
2056 const GlobalValue *GV = GA->getGlobal();
2057 DebugLoc dl = GA->getDebugLoc();
2059 SDValue Chain = DAG.getEntryNode();
2060 EVT PtrVT = getPointerTy();
2061 // Get the Thread Pointer
2062 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2064 if (GV->isDeclaration()) {
2065 MachineFunction &MF = DAG.getMachineFunction();
2066 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2067 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2068 // Initial exec model.
2069 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2070 ARMConstantPoolValue *CPV =
2071 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2072 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2074 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2075 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2076 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2077 MachinePointerInfo::getConstantPool(),
2078 false, false, false, 0);
2079 Chain = Offset.getValue(1);
2081 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2082 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2084 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2085 MachinePointerInfo::getConstantPool(),
2086 false, false, false, 0);
2089 ARMConstantPoolValue *CPV =
2090 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2091 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2092 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2093 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2094 MachinePointerInfo::getConstantPool(),
2095 false, false, false, 0);
2098 // The address of the thread local variable is the add of the thread
2099 // pointer with the offset of the variable.
2100 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2104 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2105 // TODO: implement the "local dynamic" model
2106 assert(Subtarget->isTargetELF() &&
2107 "TLS not implemented for non-ELF targets");
2108 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2109 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2110 // otherwise use the "Local Exec" TLS Model
2111 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2112 return LowerToTLSGeneralDynamicModel(GA, DAG);
2114 return LowerToTLSExecModels(GA, DAG);
2117 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2118 SelectionDAG &DAG) const {
2119 EVT PtrVT = getPointerTy();
2120 DebugLoc dl = Op.getDebugLoc();
2121 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2122 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2123 if (RelocM == Reloc::PIC_) {
2124 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2125 ARMConstantPoolValue *CPV =
2126 ARMConstantPoolConstant::Create(GV,
2127 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2128 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2129 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2130 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2132 MachinePointerInfo::getConstantPool(),
2133 false, false, false, 0);
2134 SDValue Chain = Result.getValue(1);
2135 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2136 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2138 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2139 MachinePointerInfo::getGOT(),
2140 false, false, false, 0);
2144 // If we have T2 ops, we can materialize the address directly via movt/movw
2145 // pair. This is always cheaper.
2146 if (Subtarget->useMovt()) {
2148 // FIXME: Once remat is capable of dealing with instructions with register
2149 // operands, expand this into two nodes.
2150 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2151 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2153 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2154 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2155 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2156 MachinePointerInfo::getConstantPool(),
2157 false, false, false, 0);
2161 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2162 SelectionDAG &DAG) const {
2163 EVT PtrVT = getPointerTy();
2164 DebugLoc dl = Op.getDebugLoc();
2165 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2166 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2167 MachineFunction &MF = DAG.getMachineFunction();
2168 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2170 // FIXME: Enable this for static codegen when tool issues are fixed.
2171 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2173 // FIXME: Once remat is capable of dealing with instructions with register
2174 // operands, expand this into two nodes.
2175 if (RelocM == Reloc::Static)
2176 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2177 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2179 unsigned Wrapper = (RelocM == Reloc::PIC_)
2180 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2181 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2182 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2183 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2184 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2185 MachinePointerInfo::getGOT(),
2186 false, false, false, 0);
2190 unsigned ARMPCLabelIndex = 0;
2192 if (RelocM == Reloc::Static) {
2193 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2195 ARMPCLabelIndex = AFI->createPICLabelUId();
2196 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2197 ARMConstantPoolValue *CPV =
2198 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2200 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2202 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2204 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2205 MachinePointerInfo::getConstantPool(),
2206 false, false, false, 0);
2207 SDValue Chain = Result.getValue(1);
2209 if (RelocM == Reloc::PIC_) {
2210 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2211 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2214 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2215 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2216 false, false, false, 0);
2221 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2222 SelectionDAG &DAG) const {
2223 assert(Subtarget->isTargetELF() &&
2224 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2225 MachineFunction &MF = DAG.getMachineFunction();
2226 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2227 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2228 EVT PtrVT = getPointerTy();
2229 DebugLoc dl = Op.getDebugLoc();
2230 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2231 ARMConstantPoolValue *CPV =
2232 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2233 ARMPCLabelIndex, PCAdj);
2234 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2235 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2236 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2237 MachinePointerInfo::getConstantPool(),
2238 false, false, false, 0);
2239 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2240 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2244 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2245 DebugLoc dl = Op.getDebugLoc();
2246 SDValue Val = DAG.getConstant(0, MVT::i32);
2247 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2248 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2249 Op.getOperand(1), Val);
2253 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2254 DebugLoc dl = Op.getDebugLoc();
2255 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2256 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2260 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2261 const ARMSubtarget *Subtarget) const {
2262 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2263 DebugLoc dl = Op.getDebugLoc();
2265 default: return SDValue(); // Don't custom lower most intrinsics.
2266 case Intrinsic::arm_thread_pointer: {
2267 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2268 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2270 case Intrinsic::eh_sjlj_lsda: {
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 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2278 unsigned PCAdj = (RelocM != Reloc::PIC_)
2279 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2280 ARMConstantPoolValue *CPV =
2281 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2282 ARMCP::CPLSDA, PCAdj);
2283 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2284 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2286 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2287 MachinePointerInfo::getConstantPool(),
2288 false, false, false, 0);
2290 if (RelocM == Reloc::PIC_) {
2291 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2292 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2296 case Intrinsic::arm_neon_vmulls:
2297 case Intrinsic::arm_neon_vmullu: {
2298 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2299 ? ARMISD::VMULLs : ARMISD::VMULLu;
2300 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2301 Op.getOperand(1), Op.getOperand(2));
2306 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2307 const ARMSubtarget *Subtarget) {
2308 DebugLoc dl = Op.getDebugLoc();
2309 if (!Subtarget->hasDataBarrier()) {
2310 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2311 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2313 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2314 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2315 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2316 DAG.getConstant(0, MVT::i32));
2319 SDValue Op5 = Op.getOperand(5);
2320 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2321 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2322 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2323 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2325 ARM_MB::MemBOpt DMBOpt;
2326 if (isDeviceBarrier)
2327 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2329 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2330 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2331 DAG.getConstant(DMBOpt, MVT::i32));
2335 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2336 const ARMSubtarget *Subtarget) {
2337 // FIXME: handle "fence singlethread" more efficiently.
2338 DebugLoc dl = Op.getDebugLoc();
2339 if (!Subtarget->hasDataBarrier()) {
2340 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2341 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2343 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2344 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2345 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2346 DAG.getConstant(0, MVT::i32));
2349 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2350 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2353 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2354 const ARMSubtarget *Subtarget) {
2355 // ARM pre v5TE and Thumb1 does not have preload instructions.
2356 if (!(Subtarget->isThumb2() ||
2357 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2358 // Just preserve the chain.
2359 return Op.getOperand(0);
2361 DebugLoc dl = Op.getDebugLoc();
2362 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2364 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2365 // ARMv7 with MP extension has PLDW.
2366 return Op.getOperand(0);
2368 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2369 if (Subtarget->isThumb()) {
2371 isRead = ~isRead & 1;
2372 isData = ~isData & 1;
2375 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2376 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2377 DAG.getConstant(isData, MVT::i32));
2380 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2381 MachineFunction &MF = DAG.getMachineFunction();
2382 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2384 // vastart just stores the address of the VarArgsFrameIndex slot into the
2385 // memory location argument.
2386 DebugLoc dl = Op.getDebugLoc();
2387 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2388 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2389 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2390 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2391 MachinePointerInfo(SV), false, false, 0);
2395 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2396 SDValue &Root, SelectionDAG &DAG,
2397 DebugLoc dl) const {
2398 MachineFunction &MF = DAG.getMachineFunction();
2399 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2401 TargetRegisterClass *RC;
2402 if (AFI->isThumb1OnlyFunction())
2403 RC = ARM::tGPRRegisterClass;
2405 RC = ARM::GPRRegisterClass;
2407 // Transform the arguments stored in physical registers into virtual ones.
2408 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2409 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2412 if (NextVA.isMemLoc()) {
2413 MachineFrameInfo *MFI = MF.getFrameInfo();
2414 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2416 // Create load node to retrieve arguments from the stack.
2417 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2418 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2419 MachinePointerInfo::getFixedStack(FI),
2420 false, false, false, 0);
2422 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2423 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2426 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2430 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2431 unsigned &VARegSize, unsigned &VARegSaveSize)
2434 if (CCInfo.isFirstByValRegValid())
2435 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2437 unsigned int firstUnalloced;
2438 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2439 sizeof(GPRArgRegs) /
2440 sizeof(GPRArgRegs[0]));
2441 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2444 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2445 VARegSize = NumGPRs * 4;
2446 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2449 // The remaining GPRs hold either the beginning of variable-argument
2450 // data, or the beginning of an aggregate passed by value (usuall
2451 // byval). Either way, we allocate stack slots adjacent to the data
2452 // provided by our caller, and store the unallocated registers there.
2453 // If this is a variadic function, the va_list pointer will begin with
2454 // these values; otherwise, this reassembles a (byval) structure that
2455 // was split between registers and memory.
2457 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2458 DebugLoc dl, SDValue &Chain,
2459 unsigned ArgOffset) const {
2460 MachineFunction &MF = DAG.getMachineFunction();
2461 MachineFrameInfo *MFI = MF.getFrameInfo();
2462 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2463 unsigned firstRegToSaveIndex;
2464 if (CCInfo.isFirstByValRegValid())
2465 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2467 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2468 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2471 unsigned VARegSize, VARegSaveSize;
2472 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2473 if (VARegSaveSize) {
2474 // If this function is vararg, store any remaining integer argument regs
2475 // to their spots on the stack so that they may be loaded by deferencing
2476 // the result of va_next.
2477 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2478 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2479 ArgOffset + VARegSaveSize
2482 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2485 SmallVector<SDValue, 4> MemOps;
2486 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2487 TargetRegisterClass *RC;
2488 if (AFI->isThumb1OnlyFunction())
2489 RC = ARM::tGPRRegisterClass;
2491 RC = ARM::GPRRegisterClass;
2493 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2494 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2496 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2497 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2499 MemOps.push_back(Store);
2500 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2501 DAG.getConstant(4, getPointerTy()));
2503 if (!MemOps.empty())
2504 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2505 &MemOps[0], MemOps.size());
2507 // This will point to the next argument passed via stack.
2508 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2512 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2513 CallingConv::ID CallConv, bool isVarArg,
2514 const SmallVectorImpl<ISD::InputArg>
2516 DebugLoc dl, SelectionDAG &DAG,
2517 SmallVectorImpl<SDValue> &InVals)
2519 MachineFunction &MF = DAG.getMachineFunction();
2520 MachineFrameInfo *MFI = MF.getFrameInfo();
2522 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2524 // Assign locations to all of the incoming arguments.
2525 SmallVector<CCValAssign, 16> ArgLocs;
2526 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2527 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2528 CCInfo.AnalyzeFormalArguments(Ins,
2529 CCAssignFnForNode(CallConv, /* Return*/ false,
2532 SmallVector<SDValue, 16> ArgValues;
2533 int lastInsIndex = -1;
2536 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2537 CCValAssign &VA = ArgLocs[i];
2539 // Arguments stored in registers.
2540 if (VA.isRegLoc()) {
2541 EVT RegVT = VA.getLocVT();
2543 if (VA.needsCustom()) {
2544 // f64 and vector types are split up into multiple registers or
2545 // combinations of registers and stack slots.
2546 if (VA.getLocVT() == MVT::v2f64) {
2547 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2549 VA = ArgLocs[++i]; // skip ahead to next loc
2551 if (VA.isMemLoc()) {
2552 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2553 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2554 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2555 MachinePointerInfo::getFixedStack(FI),
2556 false, false, false, 0);
2558 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2561 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2562 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2563 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2564 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2565 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2567 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2570 TargetRegisterClass *RC;
2572 if (RegVT == MVT::f32)
2573 RC = ARM::SPRRegisterClass;
2574 else if (RegVT == MVT::f64)
2575 RC = ARM::DPRRegisterClass;
2576 else if (RegVT == MVT::v2f64)
2577 RC = ARM::QPRRegisterClass;
2578 else if (RegVT == MVT::i32)
2579 RC = (AFI->isThumb1OnlyFunction() ?
2580 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2582 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2584 // Transform the arguments in physical registers into virtual ones.
2585 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2586 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2589 // If this is an 8 or 16-bit value, it is really passed promoted
2590 // to 32 bits. Insert an assert[sz]ext to capture this, then
2591 // truncate to the right size.
2592 switch (VA.getLocInfo()) {
2593 default: llvm_unreachable("Unknown loc info!");
2594 case CCValAssign::Full: break;
2595 case CCValAssign::BCvt:
2596 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2598 case CCValAssign::SExt:
2599 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2600 DAG.getValueType(VA.getValVT()));
2601 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2603 case CCValAssign::ZExt:
2604 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2605 DAG.getValueType(VA.getValVT()));
2606 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2610 InVals.push_back(ArgValue);
2612 } else { // VA.isRegLoc()
2615 assert(VA.isMemLoc());
2616 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2618 int index = ArgLocs[i].getValNo();
2620 // Some Ins[] entries become multiple ArgLoc[] entries.
2621 // Process them only once.
2622 if (index != lastInsIndex)
2624 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2625 // FIXME: For now, all byval parameter objects are marked mutable.
2626 // This can be changed with more analysis.
2627 // In case of tail call optimization mark all arguments mutable.
2628 // Since they could be overwritten by lowering of arguments in case of
2630 if (Flags.isByVal()) {
2631 unsigned VARegSize, VARegSaveSize;
2632 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2633 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2634 unsigned Bytes = Flags.getByValSize() - VARegSize;
2635 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2636 int FI = MFI->CreateFixedObject(Bytes,
2637 VA.getLocMemOffset(), false);
2638 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2640 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2641 VA.getLocMemOffset(), true);
2643 // Create load nodes to retrieve arguments from the stack.
2644 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2645 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2646 MachinePointerInfo::getFixedStack(FI),
2647 false, false, false, 0));
2649 lastInsIndex = index;
2656 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2661 /// isFloatingPointZero - Return true if this is +0.0.
2662 static bool isFloatingPointZero(SDValue Op) {
2663 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2664 return CFP->getValueAPF().isPosZero();
2665 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2666 // Maybe this has already been legalized into the constant pool?
2667 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2668 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2669 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2670 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2671 return CFP->getValueAPF().isPosZero();
2677 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2678 /// the given operands.
2680 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2681 SDValue &ARMcc, SelectionDAG &DAG,
2682 DebugLoc dl) const {
2683 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2684 unsigned C = RHSC->getZExtValue();
2685 if (!isLegalICmpImmediate(C)) {
2686 // Constant does not fit, try adjusting it by one?
2691 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2692 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2693 RHS = DAG.getConstant(C-1, MVT::i32);
2698 if (C != 0 && isLegalICmpImmediate(C-1)) {
2699 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2700 RHS = DAG.getConstant(C-1, MVT::i32);
2705 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2706 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2707 RHS = DAG.getConstant(C+1, MVT::i32);
2712 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2713 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2714 RHS = DAG.getConstant(C+1, MVT::i32);
2721 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2722 ARMISD::NodeType CompareType;
2725 CompareType = ARMISD::CMP;
2730 CompareType = ARMISD::CMPZ;
2733 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2734 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2737 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2739 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2740 DebugLoc dl) const {
2742 if (!isFloatingPointZero(RHS))
2743 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2745 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2746 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2749 /// duplicateCmp - Glue values can have only one use, so this function
2750 /// duplicates a comparison node.
2752 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2753 unsigned Opc = Cmp.getOpcode();
2754 DebugLoc DL = Cmp.getDebugLoc();
2755 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2756 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2758 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2759 Cmp = Cmp.getOperand(0);
2760 Opc = Cmp.getOpcode();
2761 if (Opc == ARMISD::CMPFP)
2762 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2764 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2765 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2767 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2770 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2771 SDValue Cond = Op.getOperand(0);
2772 SDValue SelectTrue = Op.getOperand(1);
2773 SDValue SelectFalse = Op.getOperand(2);
2774 DebugLoc dl = Op.getDebugLoc();
2778 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2779 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2781 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2782 const ConstantSDNode *CMOVTrue =
2783 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2784 const ConstantSDNode *CMOVFalse =
2785 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2787 if (CMOVTrue && CMOVFalse) {
2788 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2789 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2793 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2795 False = SelectFalse;
2796 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2801 if (True.getNode() && False.getNode()) {
2802 EVT VT = Op.getValueType();
2803 SDValue ARMcc = Cond.getOperand(2);
2804 SDValue CCR = Cond.getOperand(3);
2805 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2806 assert(True.getValueType() == VT);
2807 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2812 return DAG.getSelectCC(dl, Cond,
2813 DAG.getConstant(0, Cond.getValueType()),
2814 SelectTrue, SelectFalse, ISD::SETNE);
2817 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2818 EVT VT = Op.getValueType();
2819 SDValue LHS = Op.getOperand(0);
2820 SDValue RHS = Op.getOperand(1);
2821 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2822 SDValue TrueVal = Op.getOperand(2);
2823 SDValue FalseVal = Op.getOperand(3);
2824 DebugLoc dl = Op.getDebugLoc();
2826 if (LHS.getValueType() == MVT::i32) {
2828 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2829 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2830 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2833 ARMCC::CondCodes CondCode, CondCode2;
2834 FPCCToARMCC(CC, CondCode, CondCode2);
2836 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2837 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2838 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2839 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2841 if (CondCode2 != ARMCC::AL) {
2842 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2843 // FIXME: Needs another CMP because flag can have but one use.
2844 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2845 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2846 Result, TrueVal, ARMcc2, CCR, Cmp2);
2851 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2852 /// to morph to an integer compare sequence.
2853 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2854 const ARMSubtarget *Subtarget) {
2855 SDNode *N = Op.getNode();
2856 if (!N->hasOneUse())
2857 // Otherwise it requires moving the value from fp to integer registers.
2859 if (!N->getNumValues())
2861 EVT VT = Op.getValueType();
2862 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2863 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2864 // vmrs are very slow, e.g. cortex-a8.
2867 if (isFloatingPointZero(Op)) {
2871 return ISD::isNormalLoad(N);
2874 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2875 if (isFloatingPointZero(Op))
2876 return DAG.getConstant(0, MVT::i32);
2878 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2879 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2880 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2881 Ld->isVolatile(), Ld->isNonTemporal(),
2882 Ld->isInvariant(), Ld->getAlignment());
2884 llvm_unreachable("Unknown VFP cmp argument!");
2887 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2888 SDValue &RetVal1, SDValue &RetVal2) {
2889 if (isFloatingPointZero(Op)) {
2890 RetVal1 = DAG.getConstant(0, MVT::i32);
2891 RetVal2 = DAG.getConstant(0, MVT::i32);
2895 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2896 SDValue Ptr = Ld->getBasePtr();
2897 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2898 Ld->getChain(), Ptr,
2899 Ld->getPointerInfo(),
2900 Ld->isVolatile(), Ld->isNonTemporal(),
2901 Ld->isInvariant(), Ld->getAlignment());
2903 EVT PtrType = Ptr.getValueType();
2904 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2905 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2906 PtrType, Ptr, DAG.getConstant(4, PtrType));
2907 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2908 Ld->getChain(), NewPtr,
2909 Ld->getPointerInfo().getWithOffset(4),
2910 Ld->isVolatile(), Ld->isNonTemporal(),
2911 Ld->isInvariant(), NewAlign);
2915 llvm_unreachable("Unknown VFP cmp argument!");
2918 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2919 /// f32 and even f64 comparisons to integer ones.
2921 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2922 SDValue Chain = Op.getOperand(0);
2923 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2924 SDValue LHS = Op.getOperand(2);
2925 SDValue RHS = Op.getOperand(3);
2926 SDValue Dest = Op.getOperand(4);
2927 DebugLoc dl = Op.getDebugLoc();
2929 bool SeenZero = false;
2930 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2931 canChangeToInt(RHS, SeenZero, Subtarget) &&
2932 // If one of the operand is zero, it's safe to ignore the NaN case since
2933 // we only care about equality comparisons.
2934 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2935 // If unsafe fp math optimization is enabled and there are no other uses of
2936 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2937 // to an integer comparison.
2938 if (CC == ISD::SETOEQ)
2940 else if (CC == ISD::SETUNE)
2944 if (LHS.getValueType() == MVT::f32) {
2945 LHS = bitcastf32Toi32(LHS, DAG);
2946 RHS = bitcastf32Toi32(RHS, DAG);
2947 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2948 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2949 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2950 Chain, Dest, ARMcc, CCR, Cmp);
2955 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2956 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2957 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2958 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2959 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2960 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2961 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2967 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2968 SDValue Chain = Op.getOperand(0);
2969 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2970 SDValue LHS = Op.getOperand(2);
2971 SDValue RHS = Op.getOperand(3);
2972 SDValue Dest = Op.getOperand(4);
2973 DebugLoc dl = Op.getDebugLoc();
2975 if (LHS.getValueType() == MVT::i32) {
2977 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2978 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2979 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2980 Chain, Dest, ARMcc, CCR, Cmp);
2983 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2985 if (getTargetMachine().Options.UnsafeFPMath &&
2986 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2987 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2988 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2989 if (Result.getNode())
2993 ARMCC::CondCodes CondCode, CondCode2;
2994 FPCCToARMCC(CC, CondCode, CondCode2);
2996 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2997 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2998 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2999 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3000 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3001 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3002 if (CondCode2 != ARMCC::AL) {
3003 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3004 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3005 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3010 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3011 SDValue Chain = Op.getOperand(0);
3012 SDValue Table = Op.getOperand(1);
3013 SDValue Index = Op.getOperand(2);
3014 DebugLoc dl = Op.getDebugLoc();
3016 EVT PTy = getPointerTy();
3017 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3018 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3019 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3020 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3021 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3022 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3023 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3024 if (Subtarget->isThumb2()) {
3025 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3026 // which does another jump to the destination. This also makes it easier
3027 // to translate it to TBB / TBH later.
3028 // FIXME: This might not work if the function is extremely large.
3029 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3030 Addr, Op.getOperand(2), JTI, UId);
3032 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3033 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3034 MachinePointerInfo::getJumpTable(),
3035 false, false, false, 0);
3036 Chain = Addr.getValue(1);
3037 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3038 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3040 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3041 MachinePointerInfo::getJumpTable(),
3042 false, false, false, 0);
3043 Chain = Addr.getValue(1);
3044 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3048 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3049 EVT VT = Op.getValueType();
3050 assert(VT.getVectorElementType() == MVT::i32 && "Unexpected custom lowering");
3052 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3054 return DAG.UnrollVectorOp(Op.getNode());
3057 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3058 EVT VT = Op.getValueType();
3060 return LowerVectorFP_TO_INT(Op, DAG);
3062 DebugLoc dl = Op.getDebugLoc();
3065 switch (Op.getOpcode()) {
3067 assert(0 && "Invalid opcode!");
3068 case ISD::FP_TO_SINT:
3069 Opc = ARMISD::FTOSI;
3071 case ISD::FP_TO_UINT:
3072 Opc = ARMISD::FTOUI;
3075 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3076 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3079 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3080 EVT VT = Op.getValueType();
3081 DebugLoc dl = Op.getDebugLoc();
3083 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3084 if (VT.getVectorElementType() == MVT::f32)
3086 return DAG.UnrollVectorOp(Op.getNode());
3089 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3090 "Invalid type for custom lowering!");
3091 if (VT != MVT::v4f32)
3092 return DAG.UnrollVectorOp(Op.getNode());
3096 switch (Op.getOpcode()) {
3098 assert(0 && "Invalid opcode!");
3099 case ISD::SINT_TO_FP:
3100 CastOpc = ISD::SIGN_EXTEND;
3101 Opc = ISD::SINT_TO_FP;
3103 case ISD::UINT_TO_FP:
3104 CastOpc = ISD::ZERO_EXTEND;
3105 Opc = ISD::UINT_TO_FP;
3109 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3110 return DAG.getNode(Opc, dl, VT, Op);
3113 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3114 EVT VT = Op.getValueType();
3116 return LowerVectorINT_TO_FP(Op, DAG);
3118 DebugLoc dl = Op.getDebugLoc();
3121 switch (Op.getOpcode()) {
3123 assert(0 && "Invalid opcode!");
3124 case ISD::SINT_TO_FP:
3125 Opc = ARMISD::SITOF;
3127 case ISD::UINT_TO_FP:
3128 Opc = ARMISD::UITOF;
3132 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3133 return DAG.getNode(Opc, dl, VT, Op);
3136 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3137 // Implement fcopysign with a fabs and a conditional fneg.
3138 SDValue Tmp0 = Op.getOperand(0);
3139 SDValue Tmp1 = Op.getOperand(1);
3140 DebugLoc dl = Op.getDebugLoc();
3141 EVT VT = Op.getValueType();
3142 EVT SrcVT = Tmp1.getValueType();
3143 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3144 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3145 bool UseNEON = !InGPR && Subtarget->hasNEON();
3148 // Use VBSL to copy the sign bit.
3149 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3150 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3151 DAG.getTargetConstant(EncodedVal, MVT::i32));
3152 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3154 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3155 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3156 DAG.getConstant(32, MVT::i32));
3157 else /*if (VT == MVT::f32)*/
3158 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3159 if (SrcVT == MVT::f32) {
3160 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3162 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3163 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3164 DAG.getConstant(32, MVT::i32));
3165 } else if (VT == MVT::f32)
3166 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3167 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3168 DAG.getConstant(32, MVT::i32));
3169 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3170 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3172 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3174 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3175 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3176 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3178 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3179 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3180 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3181 if (VT == MVT::f32) {
3182 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3183 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3184 DAG.getConstant(0, MVT::i32));
3186 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3192 // Bitcast operand 1 to i32.
3193 if (SrcVT == MVT::f64)
3194 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3195 &Tmp1, 1).getValue(1);
3196 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3198 // Or in the signbit with integer operations.
3199 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3200 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3201 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3202 if (VT == MVT::f32) {
3203 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3204 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3205 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3206 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3209 // f64: Or the high part with signbit and then combine two parts.
3210 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3212 SDValue Lo = Tmp0.getValue(0);
3213 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3214 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3215 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3218 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3219 MachineFunction &MF = DAG.getMachineFunction();
3220 MachineFrameInfo *MFI = MF.getFrameInfo();
3221 MFI->setReturnAddressIsTaken(true);
3223 EVT VT = Op.getValueType();
3224 DebugLoc dl = Op.getDebugLoc();
3225 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3227 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3228 SDValue Offset = DAG.getConstant(4, MVT::i32);
3229 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3230 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3231 MachinePointerInfo(), false, false, false, 0);
3234 // Return LR, which contains the return address. Mark it an implicit live-in.
3235 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3236 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3239 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3240 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3241 MFI->setFrameAddressIsTaken(true);
3243 EVT VT = Op.getValueType();
3244 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3245 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3246 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3247 ? ARM::R7 : ARM::R11;
3248 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3250 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3251 MachinePointerInfo(),
3252 false, false, false, 0);
3256 /// ExpandBITCAST - If the target supports VFP, this function is called to
3257 /// expand a bit convert where either the source or destination type is i64 to
3258 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3259 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3260 /// vectors), since the legalizer won't know what to do with that.
3261 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3262 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3263 DebugLoc dl = N->getDebugLoc();
3264 SDValue Op = N->getOperand(0);
3266 // This function is only supposed to be called for i64 types, either as the
3267 // source or destination of the bit convert.
3268 EVT SrcVT = Op.getValueType();
3269 EVT DstVT = N->getValueType(0);
3270 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3271 "ExpandBITCAST called for non-i64 type");
3273 // Turn i64->f64 into VMOVDRR.
3274 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3275 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3276 DAG.getConstant(0, MVT::i32));
3277 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3278 DAG.getConstant(1, MVT::i32));
3279 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3280 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3283 // Turn f64->i64 into VMOVRRD.
3284 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3285 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3286 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3287 // Merge the pieces into a single i64 value.
3288 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3294 /// getZeroVector - Returns a vector of specified type with all zero elements.
3295 /// Zero vectors are used to represent vector negation and in those cases
3296 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3297 /// not support i64 elements, so sometimes the zero vectors will need to be
3298 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3300 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3301 assert(VT.isVector() && "Expected a vector type");
3302 // The canonical modified immediate encoding of a zero vector is....0!
3303 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3304 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3305 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3306 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3309 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3310 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3311 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3312 SelectionDAG &DAG) const {
3313 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3314 EVT VT = Op.getValueType();
3315 unsigned VTBits = VT.getSizeInBits();
3316 DebugLoc dl = Op.getDebugLoc();
3317 SDValue ShOpLo = Op.getOperand(0);
3318 SDValue ShOpHi = Op.getOperand(1);
3319 SDValue ShAmt = Op.getOperand(2);
3321 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3323 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3325 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3326 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3327 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3328 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3329 DAG.getConstant(VTBits, MVT::i32));
3330 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3331 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3332 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3334 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3335 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3337 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3338 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3341 SDValue Ops[2] = { Lo, Hi };
3342 return DAG.getMergeValues(Ops, 2, dl);
3345 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3346 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3347 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3348 SelectionDAG &DAG) const {
3349 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3350 EVT VT = Op.getValueType();
3351 unsigned VTBits = VT.getSizeInBits();
3352 DebugLoc dl = Op.getDebugLoc();
3353 SDValue ShOpLo = Op.getOperand(0);
3354 SDValue ShOpHi = Op.getOperand(1);
3355 SDValue ShAmt = Op.getOperand(2);
3358 assert(Op.getOpcode() == ISD::SHL_PARTS);
3359 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3360 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3361 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3362 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3363 DAG.getConstant(VTBits, MVT::i32));
3364 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3365 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3367 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3368 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3369 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3371 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3372 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3375 SDValue Ops[2] = { Lo, Hi };
3376 return DAG.getMergeValues(Ops, 2, dl);
3379 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3380 SelectionDAG &DAG) const {
3381 // The rounding mode is in bits 23:22 of the FPSCR.
3382 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3383 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3384 // so that the shift + and get folded into a bitfield extract.
3385 DebugLoc dl = Op.getDebugLoc();
3386 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3387 DAG.getConstant(Intrinsic::arm_get_fpscr,
3389 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3390 DAG.getConstant(1U << 22, MVT::i32));
3391 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3392 DAG.getConstant(22, MVT::i32));
3393 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3394 DAG.getConstant(3, MVT::i32));
3397 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3398 const ARMSubtarget *ST) {
3399 EVT VT = N->getValueType(0);
3400 DebugLoc dl = N->getDebugLoc();
3402 if (!ST->hasV6T2Ops())
3405 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3406 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3409 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3410 const ARMSubtarget *ST) {
3411 EVT VT = N->getValueType(0);
3412 DebugLoc dl = N->getDebugLoc();
3417 // Lower vector shifts on NEON to use VSHL.
3418 assert(ST->hasNEON() && "unexpected vector shift");
3420 // Left shifts translate directly to the vshiftu intrinsic.
3421 if (N->getOpcode() == ISD::SHL)
3422 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3423 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3424 N->getOperand(0), N->getOperand(1));
3426 assert((N->getOpcode() == ISD::SRA ||
3427 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3429 // NEON uses the same intrinsics for both left and right shifts. For
3430 // right shifts, the shift amounts are negative, so negate the vector of
3432 EVT ShiftVT = N->getOperand(1).getValueType();
3433 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3434 getZeroVector(ShiftVT, DAG, dl),
3436 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3437 Intrinsic::arm_neon_vshifts :
3438 Intrinsic::arm_neon_vshiftu);
3439 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3440 DAG.getConstant(vshiftInt, MVT::i32),
3441 N->getOperand(0), NegatedCount);
3444 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3445 const ARMSubtarget *ST) {
3446 EVT VT = N->getValueType(0);
3447 DebugLoc dl = N->getDebugLoc();
3449 // We can get here for a node like i32 = ISD::SHL i32, i64
3453 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3454 "Unknown shift to lower!");
3456 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3457 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3458 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3461 // If we are in thumb mode, we don't have RRX.
3462 if (ST->isThumb1Only()) return SDValue();
3464 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3465 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3466 DAG.getConstant(0, MVT::i32));
3467 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3468 DAG.getConstant(1, MVT::i32));
3470 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3471 // captures the result into a carry flag.
3472 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3473 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3475 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3476 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3478 // Merge the pieces into a single i64 value.
3479 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3482 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3483 SDValue TmpOp0, TmpOp1;
3484 bool Invert = false;
3488 SDValue Op0 = Op.getOperand(0);
3489 SDValue Op1 = Op.getOperand(1);
3490 SDValue CC = Op.getOperand(2);
3491 EVT VT = Op.getValueType();
3492 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3493 DebugLoc dl = Op.getDebugLoc();
3495 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3496 switch (SetCCOpcode) {
3497 default: llvm_unreachable("Illegal FP comparison"); break;
3499 case ISD::SETNE: Invert = true; // Fallthrough
3501 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3503 case ISD::SETLT: Swap = true; // Fallthrough
3505 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3507 case ISD::SETLE: Swap = true; // Fallthrough
3509 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3510 case ISD::SETUGE: Swap = true; // Fallthrough
3511 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3512 case ISD::SETUGT: Swap = true; // Fallthrough
3513 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3514 case ISD::SETUEQ: Invert = true; // Fallthrough
3516 // Expand this to (OLT | OGT).
3520 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3521 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3523 case ISD::SETUO: Invert = true; // Fallthrough
3525 // Expand this to (OLT | OGE).
3529 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3530 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3534 // Integer comparisons.
3535 switch (SetCCOpcode) {
3536 default: llvm_unreachable("Illegal integer comparison"); break;
3537 case ISD::SETNE: Invert = true;
3538 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3539 case ISD::SETLT: Swap = true;
3540 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3541 case ISD::SETLE: Swap = true;
3542 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3543 case ISD::SETULT: Swap = true;
3544 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3545 case ISD::SETULE: Swap = true;
3546 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3549 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3550 if (Opc == ARMISD::VCEQ) {
3553 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3555 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3558 // Ignore bitconvert.
3559 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3560 AndOp = AndOp.getOperand(0);
3562 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3564 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3565 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3572 std::swap(Op0, Op1);
3574 // If one of the operands is a constant vector zero, attempt to fold the
3575 // comparison to a specialized compare-against-zero form.
3577 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3579 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3580 if (Opc == ARMISD::VCGE)
3581 Opc = ARMISD::VCLEZ;
3582 else if (Opc == ARMISD::VCGT)
3583 Opc = ARMISD::VCLTZ;
3588 if (SingleOp.getNode()) {
3591 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3593 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3595 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3597 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3599 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3601 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3604 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3608 Result = DAG.getNOT(dl, Result, VT);
3613 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3614 /// valid vector constant for a NEON instruction with a "modified immediate"
3615 /// operand (e.g., VMOV). If so, return the encoded value.
3616 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3617 unsigned SplatBitSize, SelectionDAG &DAG,
3618 EVT &VT, bool is128Bits, NEONModImmType type) {
3619 unsigned OpCmode, Imm;
3621 // SplatBitSize is set to the smallest size that splats the vector, so a
3622 // zero vector will always have SplatBitSize == 8. However, NEON modified
3623 // immediate instructions others than VMOV do not support the 8-bit encoding
3624 // of a zero vector, and the default encoding of zero is supposed to be the
3629 switch (SplatBitSize) {
3631 if (type != VMOVModImm)
3633 // Any 1-byte value is OK. Op=0, Cmode=1110.
3634 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3637 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3641 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3642 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3643 if ((SplatBits & ~0xff) == 0) {
3644 // Value = 0x00nn: Op=x, Cmode=100x.
3649 if ((SplatBits & ~0xff00) == 0) {
3650 // Value = 0xnn00: Op=x, Cmode=101x.
3652 Imm = SplatBits >> 8;
3658 // NEON's 32-bit VMOV supports splat values where:
3659 // * only one byte is nonzero, or
3660 // * the least significant byte is 0xff and the second byte is nonzero, or
3661 // * the least significant 2 bytes are 0xff and the third is nonzero.
3662 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3663 if ((SplatBits & ~0xff) == 0) {
3664 // Value = 0x000000nn: Op=x, Cmode=000x.
3669 if ((SplatBits & ~0xff00) == 0) {
3670 // Value = 0x0000nn00: Op=x, Cmode=001x.
3672 Imm = SplatBits >> 8;
3675 if ((SplatBits & ~0xff0000) == 0) {
3676 // Value = 0x00nn0000: Op=x, Cmode=010x.
3678 Imm = SplatBits >> 16;
3681 if ((SplatBits & ~0xff000000) == 0) {
3682 // Value = 0xnn000000: Op=x, Cmode=011x.
3684 Imm = SplatBits >> 24;
3688 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3689 if (type == OtherModImm) return SDValue();
3691 if ((SplatBits & ~0xffff) == 0 &&
3692 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3693 // Value = 0x0000nnff: Op=x, Cmode=1100.
3695 Imm = SplatBits >> 8;
3700 if ((SplatBits & ~0xffffff) == 0 &&
3701 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3702 // Value = 0x00nnffff: Op=x, Cmode=1101.
3704 Imm = SplatBits >> 16;
3705 SplatBits |= 0xffff;
3709 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3710 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3711 // VMOV.I32. A (very) minor optimization would be to replicate the value
3712 // and fall through here to test for a valid 64-bit splat. But, then the
3713 // caller would also need to check and handle the change in size.
3717 if (type != VMOVModImm)
3719 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3720 uint64_t BitMask = 0xff;
3722 unsigned ImmMask = 1;
3724 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3725 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3728 } else if ((SplatBits & BitMask) != 0) {
3734 // Op=1, Cmode=1110.
3737 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3742 llvm_unreachable("unexpected size for isNEONModifiedImm");
3746 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3747 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3750 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3751 bool &ReverseVEXT, unsigned &Imm) {
3752 unsigned NumElts = VT.getVectorNumElements();
3753 ReverseVEXT = false;
3755 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3761 // If this is a VEXT shuffle, the immediate value is the index of the first
3762 // element. The other shuffle indices must be the successive elements after
3764 unsigned ExpectedElt = Imm;
3765 for (unsigned i = 1; i < NumElts; ++i) {
3766 // Increment the expected index. If it wraps around, it may still be
3767 // a VEXT but the source vectors must be swapped.
3769 if (ExpectedElt == NumElts * 2) {
3774 if (M[i] < 0) continue; // ignore UNDEF indices
3775 if (ExpectedElt != static_cast<unsigned>(M[i]))
3779 // Adjust the index value if the source operands will be swapped.
3786 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3787 /// instruction with the specified blocksize. (The order of the elements
3788 /// within each block of the vector is reversed.)
3789 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3790 unsigned BlockSize) {
3791 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3792 "Only possible block sizes for VREV are: 16, 32, 64");
3794 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3798 unsigned NumElts = VT.getVectorNumElements();
3799 unsigned BlockElts = M[0] + 1;
3800 // If the first shuffle index is UNDEF, be optimistic.
3802 BlockElts = BlockSize / EltSz;
3804 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3807 for (unsigned i = 0; i < NumElts; ++i) {
3808 if (M[i] < 0) continue; // ignore UNDEF indices
3809 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3816 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3817 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3818 // range, then 0 is placed into the resulting vector. So pretty much any mask
3819 // of 8 elements can work here.
3820 return VT == MVT::v8i8 && M.size() == 8;
3823 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3824 unsigned &WhichResult) {
3825 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3829 unsigned NumElts = VT.getVectorNumElements();
3830 WhichResult = (M[0] == 0 ? 0 : 1);
3831 for (unsigned i = 0; i < NumElts; i += 2) {
3832 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3833 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3839 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3840 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3841 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3842 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3843 unsigned &WhichResult) {
3844 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3848 unsigned NumElts = VT.getVectorNumElements();
3849 WhichResult = (M[0] == 0 ? 0 : 1);
3850 for (unsigned i = 0; i < NumElts; i += 2) {
3851 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3852 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3858 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3859 unsigned &WhichResult) {
3860 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3864 unsigned NumElts = VT.getVectorNumElements();
3865 WhichResult = (M[0] == 0 ? 0 : 1);
3866 for (unsigned i = 0; i != NumElts; ++i) {
3867 if (M[i] < 0) continue; // ignore UNDEF indices
3868 if ((unsigned) M[i] != 2 * i + WhichResult)
3872 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3873 if (VT.is64BitVector() && EltSz == 32)
3879 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3880 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3881 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3882 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3883 unsigned &WhichResult) {
3884 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3888 unsigned Half = VT.getVectorNumElements() / 2;
3889 WhichResult = (M[0] == 0 ? 0 : 1);
3890 for (unsigned j = 0; j != 2; ++j) {
3891 unsigned Idx = WhichResult;
3892 for (unsigned i = 0; i != Half; ++i) {
3893 int MIdx = M[i + j * Half];
3894 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3900 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3901 if (VT.is64BitVector() && EltSz == 32)
3907 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3908 unsigned &WhichResult) {
3909 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3913 unsigned NumElts = VT.getVectorNumElements();
3914 WhichResult = (M[0] == 0 ? 0 : 1);
3915 unsigned Idx = WhichResult * NumElts / 2;
3916 for (unsigned i = 0; i != NumElts; i += 2) {
3917 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3918 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3923 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3924 if (VT.is64BitVector() && EltSz == 32)
3930 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3931 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3932 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3933 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3934 unsigned &WhichResult) {
3935 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3939 unsigned NumElts = VT.getVectorNumElements();
3940 WhichResult = (M[0] == 0 ? 0 : 1);
3941 unsigned Idx = WhichResult * NumElts / 2;
3942 for (unsigned i = 0; i != NumElts; i += 2) {
3943 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3944 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3949 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3950 if (VT.is64BitVector() && EltSz == 32)
3956 // If N is an integer constant that can be moved into a register in one
3957 // instruction, return an SDValue of such a constant (will become a MOV
3958 // instruction). Otherwise return null.
3959 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3960 const ARMSubtarget *ST, DebugLoc dl) {
3962 if (!isa<ConstantSDNode>(N))
3964 Val = cast<ConstantSDNode>(N)->getZExtValue();
3966 if (ST->isThumb1Only()) {
3967 if (Val <= 255 || ~Val <= 255)
3968 return DAG.getConstant(Val, MVT::i32);
3970 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3971 return DAG.getConstant(Val, MVT::i32);
3976 // If this is a case we can't handle, return null and let the default
3977 // expansion code take care of it.
3978 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3979 const ARMSubtarget *ST) const {
3980 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3981 DebugLoc dl = Op.getDebugLoc();
3982 EVT VT = Op.getValueType();
3984 APInt SplatBits, SplatUndef;
3985 unsigned SplatBitSize;
3987 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3988 if (SplatBitSize <= 64) {
3989 // Check if an immediate VMOV works.
3991 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3992 SplatUndef.getZExtValue(), SplatBitSize,
3993 DAG, VmovVT, VT.is128BitVector(),
3995 if (Val.getNode()) {
3996 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3997 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4000 // Try an immediate VMVN.
4001 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4002 Val = isNEONModifiedImm(NegatedImm,
4003 SplatUndef.getZExtValue(), SplatBitSize,
4004 DAG, VmovVT, VT.is128BitVector(),
4006 if (Val.getNode()) {
4007 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4008 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4011 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4012 if (VT == MVT::v2f32 || VT == MVT::v4f32) {
4013 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4015 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4016 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4022 // Scan through the operands to see if only one value is used.
4023 unsigned NumElts = VT.getVectorNumElements();
4024 bool isOnlyLowElement = true;
4025 bool usesOnlyOneValue = true;
4026 bool isConstant = true;
4028 for (unsigned i = 0; i < NumElts; ++i) {
4029 SDValue V = Op.getOperand(i);
4030 if (V.getOpcode() == ISD::UNDEF)
4033 isOnlyLowElement = false;
4034 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4037 if (!Value.getNode())
4039 else if (V != Value)
4040 usesOnlyOneValue = false;
4043 if (!Value.getNode())
4044 return DAG.getUNDEF(VT);
4046 if (isOnlyLowElement)
4047 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4049 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4051 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4052 // i32 and try again.
4053 if (usesOnlyOneValue && EltSize <= 32) {
4055 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4056 if (VT.getVectorElementType().isFloatingPoint()) {
4057 SmallVector<SDValue, 8> Ops;
4058 for (unsigned i = 0; i < NumElts; ++i)
4059 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4061 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4062 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4063 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4065 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4067 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4069 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4072 // If all elements are constants and the case above didn't get hit, fall back
4073 // to the default expansion, which will generate a load from the constant
4078 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4080 SDValue shuffle = ReconstructShuffle(Op, DAG);
4081 if (shuffle != SDValue())
4085 // Vectors with 32- or 64-bit elements can be built by directly assigning
4086 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4087 // will be legalized.
4088 if (EltSize >= 32) {
4089 // Do the expansion with floating-point types, since that is what the VFP
4090 // registers are defined to use, and since i64 is not legal.
4091 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4092 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4093 SmallVector<SDValue, 8> Ops;
4094 for (unsigned i = 0; i < NumElts; ++i)
4095 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4096 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4097 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4103 // Gather data to see if the operation can be modelled as a
4104 // shuffle in combination with VEXTs.
4105 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4106 SelectionDAG &DAG) const {
4107 DebugLoc dl = Op.getDebugLoc();
4108 EVT VT = Op.getValueType();
4109 unsigned NumElts = VT.getVectorNumElements();
4111 SmallVector<SDValue, 2> SourceVecs;
4112 SmallVector<unsigned, 2> MinElts;
4113 SmallVector<unsigned, 2> MaxElts;
4115 for (unsigned i = 0; i < NumElts; ++i) {
4116 SDValue V = Op.getOperand(i);
4117 if (V.getOpcode() == ISD::UNDEF)
4119 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4120 // A shuffle can only come from building a vector from various
4121 // elements of other vectors.
4123 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4124 VT.getVectorElementType()) {
4125 // This code doesn't know how to handle shuffles where the vector
4126 // element types do not match (this happens because type legalization
4127 // promotes the return type of EXTRACT_VECTOR_ELT).
4128 // FIXME: It might be appropriate to extend this code to handle
4129 // mismatched types.
4133 // Record this extraction against the appropriate vector if possible...
4134 SDValue SourceVec = V.getOperand(0);
4135 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4136 bool FoundSource = false;
4137 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4138 if (SourceVecs[j] == SourceVec) {
4139 if (MinElts[j] > EltNo)
4141 if (MaxElts[j] < EltNo)
4148 // Or record a new source if not...
4150 SourceVecs.push_back(SourceVec);
4151 MinElts.push_back(EltNo);
4152 MaxElts.push_back(EltNo);
4156 // Currently only do something sane when at most two source vectors
4158 if (SourceVecs.size() > 2)
4161 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4162 int VEXTOffsets[2] = {0, 0};
4164 // This loop extracts the usage patterns of the source vectors
4165 // and prepares appropriate SDValues for a shuffle if possible.
4166 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4167 if (SourceVecs[i].getValueType() == VT) {
4168 // No VEXT necessary
4169 ShuffleSrcs[i] = SourceVecs[i];
4172 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4173 // It probably isn't worth padding out a smaller vector just to
4174 // break it down again in a shuffle.
4178 // Since only 64-bit and 128-bit vectors are legal on ARM and
4179 // we've eliminated the other cases...
4180 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4181 "unexpected vector sizes in ReconstructShuffle");
4183 if (MaxElts[i] - MinElts[i] >= NumElts) {
4184 // Span too large for a VEXT to cope
4188 if (MinElts[i] >= NumElts) {
4189 // The extraction can just take the second half
4190 VEXTOffsets[i] = NumElts;
4191 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4193 DAG.getIntPtrConstant(NumElts));
4194 } else if (MaxElts[i] < NumElts) {
4195 // The extraction can just take the first half
4197 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4199 DAG.getIntPtrConstant(0));
4201 // An actual VEXT is needed
4202 VEXTOffsets[i] = MinElts[i];
4203 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4205 DAG.getIntPtrConstant(0));
4206 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4208 DAG.getIntPtrConstant(NumElts));
4209 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4210 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4214 SmallVector<int, 8> Mask;
4216 for (unsigned i = 0; i < NumElts; ++i) {
4217 SDValue Entry = Op.getOperand(i);
4218 if (Entry.getOpcode() == ISD::UNDEF) {
4223 SDValue ExtractVec = Entry.getOperand(0);
4224 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4225 .getOperand(1))->getSExtValue();
4226 if (ExtractVec == SourceVecs[0]) {
4227 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4229 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4233 // Final check before we try to produce nonsense...
4234 if (isShuffleMaskLegal(Mask, VT))
4235 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4241 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4242 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4243 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4244 /// are assumed to be legal.
4246 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4248 if (VT.getVectorNumElements() == 4 &&
4249 (VT.is128BitVector() || VT.is64BitVector())) {
4250 unsigned PFIndexes[4];
4251 for (unsigned i = 0; i != 4; ++i) {
4255 PFIndexes[i] = M[i];
4258 // Compute the index in the perfect shuffle table.
4259 unsigned PFTableIndex =
4260 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4261 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4262 unsigned Cost = (PFEntry >> 30);
4269 unsigned Imm, WhichResult;
4271 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4272 return (EltSize >= 32 ||
4273 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4274 isVREVMask(M, VT, 64) ||
4275 isVREVMask(M, VT, 32) ||
4276 isVREVMask(M, VT, 16) ||
4277 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4278 isVTBLMask(M, VT) ||
4279 isVTRNMask(M, VT, WhichResult) ||
4280 isVUZPMask(M, VT, WhichResult) ||
4281 isVZIPMask(M, VT, WhichResult) ||
4282 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4283 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4284 isVZIP_v_undef_Mask(M, VT, WhichResult));
4287 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4288 /// the specified operations to build the shuffle.
4289 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4290 SDValue RHS, SelectionDAG &DAG,
4292 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4293 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4294 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4297 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4306 OP_VUZPL, // VUZP, left result
4307 OP_VUZPR, // VUZP, right result
4308 OP_VZIPL, // VZIP, left result
4309 OP_VZIPR, // VZIP, right result
4310 OP_VTRNL, // VTRN, left result
4311 OP_VTRNR // VTRN, right result
4314 if (OpNum == OP_COPY) {
4315 if (LHSID == (1*9+2)*9+3) return LHS;
4316 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4320 SDValue OpLHS, OpRHS;
4321 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4322 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4323 EVT VT = OpLHS.getValueType();
4326 default: llvm_unreachable("Unknown shuffle opcode!");
4328 // VREV divides the vector in half and swaps within the half.
4329 if (VT.getVectorElementType() == MVT::i32 ||
4330 VT.getVectorElementType() == MVT::f32)
4331 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4332 // vrev <4 x i16> -> VREV32
4333 if (VT.getVectorElementType() == MVT::i16)
4334 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4335 // vrev <4 x i8> -> VREV16
4336 assert(VT.getVectorElementType() == MVT::i8);
4337 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4342 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4343 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4347 return DAG.getNode(ARMISD::VEXT, dl, VT,
4349 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4352 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4353 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4356 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4357 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4360 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4361 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4365 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4366 SmallVectorImpl<int> &ShuffleMask,
4367 SelectionDAG &DAG) {
4368 // Check to see if we can use the VTBL instruction.
4369 SDValue V1 = Op.getOperand(0);
4370 SDValue V2 = Op.getOperand(1);
4371 DebugLoc DL = Op.getDebugLoc();
4373 SmallVector<SDValue, 8> VTBLMask;
4374 for (SmallVectorImpl<int>::iterator
4375 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4376 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4378 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4379 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4380 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4383 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4384 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4388 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4389 SDValue V1 = Op.getOperand(0);
4390 SDValue V2 = Op.getOperand(1);
4391 DebugLoc dl = Op.getDebugLoc();
4392 EVT VT = Op.getValueType();
4393 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4394 SmallVector<int, 8> ShuffleMask;
4396 // Convert shuffles that are directly supported on NEON to target-specific
4397 // DAG nodes, instead of keeping them as shuffles and matching them again
4398 // during code selection. This is more efficient and avoids the possibility
4399 // of inconsistencies between legalization and selection.
4400 // FIXME: floating-point vectors should be canonicalized to integer vectors
4401 // of the same time so that they get CSEd properly.
4402 SVN->getMask(ShuffleMask);
4404 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4405 if (EltSize <= 32) {
4406 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4407 int Lane = SVN->getSplatIndex();
4408 // If this is undef splat, generate it via "just" vdup, if possible.
4409 if (Lane == -1) Lane = 0;
4411 // Test if V1 is a SCALAR_TO_VECTOR.
4412 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4413 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4415 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4416 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4418 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4419 !isa<ConstantSDNode>(V1.getOperand(0))) {
4420 bool IsScalarToVector = true;
4421 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4422 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4423 IsScalarToVector = false;
4426 if (IsScalarToVector)
4427 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4429 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4430 DAG.getConstant(Lane, MVT::i32));
4435 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4438 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4439 DAG.getConstant(Imm, MVT::i32));
4442 if (isVREVMask(ShuffleMask, VT, 64))
4443 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4444 if (isVREVMask(ShuffleMask, VT, 32))
4445 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4446 if (isVREVMask(ShuffleMask, VT, 16))
4447 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4449 // Check for Neon shuffles that modify both input vectors in place.
4450 // If both results are used, i.e., if there are two shuffles with the same
4451 // source operands and with masks corresponding to both results of one of
4452 // these operations, DAG memoization will ensure that a single node is
4453 // used for both shuffles.
4454 unsigned WhichResult;
4455 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4456 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4457 V1, V2).getValue(WhichResult);
4458 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4459 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4460 V1, V2).getValue(WhichResult);
4461 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4462 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4463 V1, V2).getValue(WhichResult);
4465 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4466 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4467 V1, V1).getValue(WhichResult);
4468 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4469 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4470 V1, V1).getValue(WhichResult);
4471 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4472 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4473 V1, V1).getValue(WhichResult);
4476 // If the shuffle is not directly supported and it has 4 elements, use
4477 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4478 unsigned NumElts = VT.getVectorNumElements();
4480 unsigned PFIndexes[4];
4481 for (unsigned i = 0; i != 4; ++i) {
4482 if (ShuffleMask[i] < 0)
4485 PFIndexes[i] = ShuffleMask[i];
4488 // Compute the index in the perfect shuffle table.
4489 unsigned PFTableIndex =
4490 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4491 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4492 unsigned Cost = (PFEntry >> 30);
4495 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4498 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4499 if (EltSize >= 32) {
4500 // Do the expansion with floating-point types, since that is what the VFP
4501 // registers are defined to use, and since i64 is not legal.
4502 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4503 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4504 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4505 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4506 SmallVector<SDValue, 8> Ops;
4507 for (unsigned i = 0; i < NumElts; ++i) {
4508 if (ShuffleMask[i] < 0)
4509 Ops.push_back(DAG.getUNDEF(EltVT));
4511 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4512 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4513 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4516 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4517 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4520 if (VT == MVT::v8i8) {
4521 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4522 if (NewOp.getNode())
4529 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4530 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4531 SDValue Lane = Op.getOperand(2);
4532 if (!isa<ConstantSDNode>(Lane))
4538 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4539 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4540 SDValue Lane = Op.getOperand(1);
4541 if (!isa<ConstantSDNode>(Lane))
4544 SDValue Vec = Op.getOperand(0);
4545 if (Op.getValueType() == MVT::i32 &&
4546 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4547 DebugLoc dl = Op.getDebugLoc();
4548 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4554 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4555 // The only time a CONCAT_VECTORS operation can have legal types is when
4556 // two 64-bit vectors are concatenated to a 128-bit vector.
4557 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4558 "unexpected CONCAT_VECTORS");
4559 DebugLoc dl = Op.getDebugLoc();
4560 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4561 SDValue Op0 = Op.getOperand(0);
4562 SDValue Op1 = Op.getOperand(1);
4563 if (Op0.getOpcode() != ISD::UNDEF)
4564 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4565 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4566 DAG.getIntPtrConstant(0));
4567 if (Op1.getOpcode() != ISD::UNDEF)
4568 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4569 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4570 DAG.getIntPtrConstant(1));
4571 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4574 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4575 /// element has been zero/sign-extended, depending on the isSigned parameter,
4576 /// from an integer type half its size.
4577 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4579 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4580 EVT VT = N->getValueType(0);
4581 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4582 SDNode *BVN = N->getOperand(0).getNode();
4583 if (BVN->getValueType(0) != MVT::v4i32 ||
4584 BVN->getOpcode() != ISD::BUILD_VECTOR)
4586 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4587 unsigned HiElt = 1 - LoElt;
4588 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4589 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4590 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4591 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4592 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4595 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4596 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4599 if (Hi0->isNullValue() && Hi1->isNullValue())
4605 if (N->getOpcode() != ISD::BUILD_VECTOR)
4608 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4609 SDNode *Elt = N->getOperand(i).getNode();
4610 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4611 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4612 unsigned HalfSize = EltSize / 2;
4614 if (!isIntN(HalfSize, C->getSExtValue()))
4617 if (!isUIntN(HalfSize, C->getZExtValue()))
4628 /// isSignExtended - Check if a node is a vector value that is sign-extended
4629 /// or a constant BUILD_VECTOR with sign-extended elements.
4630 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4631 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4633 if (isExtendedBUILD_VECTOR(N, DAG, true))
4638 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4639 /// or a constant BUILD_VECTOR with zero-extended elements.
4640 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4641 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4643 if (isExtendedBUILD_VECTOR(N, DAG, false))
4648 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4649 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4650 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4651 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4652 return N->getOperand(0);
4653 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4654 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4655 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4656 LD->isNonTemporal(), LD->isInvariant(),
4657 LD->getAlignment());
4658 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4659 // have been legalized as a BITCAST from v4i32.
4660 if (N->getOpcode() == ISD::BITCAST) {
4661 SDNode *BVN = N->getOperand(0).getNode();
4662 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4663 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4664 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4665 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4666 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4668 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4669 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4670 EVT VT = N->getValueType(0);
4671 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4672 unsigned NumElts = VT.getVectorNumElements();
4673 MVT TruncVT = MVT::getIntegerVT(EltSize);
4674 SmallVector<SDValue, 8> Ops;
4675 for (unsigned i = 0; i != NumElts; ++i) {
4676 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4677 const APInt &CInt = C->getAPIntValue();
4678 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4680 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4681 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4684 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4685 unsigned Opcode = N->getOpcode();
4686 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4687 SDNode *N0 = N->getOperand(0).getNode();
4688 SDNode *N1 = N->getOperand(1).getNode();
4689 return N0->hasOneUse() && N1->hasOneUse() &&
4690 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4695 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4696 unsigned Opcode = N->getOpcode();
4697 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4698 SDNode *N0 = N->getOperand(0).getNode();
4699 SDNode *N1 = N->getOperand(1).getNode();
4700 return N0->hasOneUse() && N1->hasOneUse() &&
4701 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4706 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4707 // Multiplications are only custom-lowered for 128-bit vectors so that
4708 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4709 EVT VT = Op.getValueType();
4710 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4711 SDNode *N0 = Op.getOperand(0).getNode();
4712 SDNode *N1 = Op.getOperand(1).getNode();
4713 unsigned NewOpc = 0;
4715 bool isN0SExt = isSignExtended(N0, DAG);
4716 bool isN1SExt = isSignExtended(N1, DAG);
4717 if (isN0SExt && isN1SExt)
4718 NewOpc = ARMISD::VMULLs;
4720 bool isN0ZExt = isZeroExtended(N0, DAG);
4721 bool isN1ZExt = isZeroExtended(N1, DAG);
4722 if (isN0ZExt && isN1ZExt)
4723 NewOpc = ARMISD::VMULLu;
4724 else if (isN1SExt || isN1ZExt) {
4725 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4726 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4727 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4728 NewOpc = ARMISD::VMULLs;
4730 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4731 NewOpc = ARMISD::VMULLu;
4733 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4735 NewOpc = ARMISD::VMULLu;
4741 if (VT == MVT::v2i64)
4742 // Fall through to expand this. It is not legal.
4745 // Other vector multiplications are legal.
4750 // Legalize to a VMULL instruction.
4751 DebugLoc DL = Op.getDebugLoc();
4753 SDValue Op1 = SkipExtension(N1, DAG);
4755 Op0 = SkipExtension(N0, DAG);
4756 assert(Op0.getValueType().is64BitVector() &&
4757 Op1.getValueType().is64BitVector() &&
4758 "unexpected types for extended operands to VMULL");
4759 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4762 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4763 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4770 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4771 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4772 EVT Op1VT = Op1.getValueType();
4773 return DAG.getNode(N0->getOpcode(), DL, VT,
4774 DAG.getNode(NewOpc, DL, VT,
4775 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4776 DAG.getNode(NewOpc, DL, VT,
4777 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4781 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4783 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4784 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4785 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4786 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4787 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4788 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4789 // Get reciprocal estimate.
4790 // float4 recip = vrecpeq_f32(yf);
4791 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4792 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4793 // Because char has a smaller range than uchar, we can actually get away
4794 // without any newton steps. This requires that we use a weird bias
4795 // of 0xb000, however (again, this has been exhaustively tested).
4796 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4797 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4798 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4799 Y = DAG.getConstant(0xb000, MVT::i32);
4800 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4801 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4802 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4803 // Convert back to short.
4804 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4805 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4810 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4812 // Convert to float.
4813 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4814 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4815 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4816 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4817 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4818 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4820 // Use reciprocal estimate and one refinement step.
4821 // float4 recip = vrecpeq_f32(yf);
4822 // recip *= vrecpsq_f32(yf, recip);
4823 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4824 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4825 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4826 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4828 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4829 // Because short has a smaller range than ushort, we can actually get away
4830 // with only a single newton step. This requires that we use a weird bias
4831 // of 89, however (again, this has been exhaustively tested).
4832 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4833 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4834 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4835 N1 = DAG.getConstant(0x89, MVT::i32);
4836 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4837 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4838 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4839 // Convert back to integer and return.
4840 // return vmovn_s32(vcvt_s32_f32(result));
4841 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4842 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4846 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4847 EVT VT = Op.getValueType();
4848 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4849 "unexpected type for custom-lowering ISD::SDIV");
4851 DebugLoc dl = Op.getDebugLoc();
4852 SDValue N0 = Op.getOperand(0);
4853 SDValue N1 = Op.getOperand(1);
4856 if (VT == MVT::v8i8) {
4857 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4858 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4860 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4861 DAG.getIntPtrConstant(4));
4862 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4863 DAG.getIntPtrConstant(4));
4864 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4865 DAG.getIntPtrConstant(0));
4866 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4867 DAG.getIntPtrConstant(0));
4869 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4870 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4872 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4873 N0 = LowerCONCAT_VECTORS(N0, DAG);
4875 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4878 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4881 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4882 EVT VT = Op.getValueType();
4883 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4884 "unexpected type for custom-lowering ISD::UDIV");
4886 DebugLoc dl = Op.getDebugLoc();
4887 SDValue N0 = Op.getOperand(0);
4888 SDValue N1 = Op.getOperand(1);
4891 if (VT == MVT::v8i8) {
4892 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4893 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4895 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4896 DAG.getIntPtrConstant(4));
4897 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4898 DAG.getIntPtrConstant(4));
4899 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4900 DAG.getIntPtrConstant(0));
4901 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4902 DAG.getIntPtrConstant(0));
4904 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4905 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4907 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4908 N0 = LowerCONCAT_VECTORS(N0, DAG);
4910 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4911 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4916 // v4i16 sdiv ... Convert to float.
4917 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4918 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4919 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4920 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4921 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4922 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4924 // Use reciprocal estimate and two refinement steps.
4925 // float4 recip = vrecpeq_f32(yf);
4926 // recip *= vrecpsq_f32(yf, recip);
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), BN1);
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 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4935 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4937 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4938 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4939 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4940 // and that it will never cause us to return an answer too large).
4941 // float4 result = as_float4(as_int4(xf*recip) + 2);
4942 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4943 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4944 N1 = DAG.getConstant(2, MVT::i32);
4945 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4946 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4947 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4948 // Convert back to integer and return.
4949 // return vmovn_u32(vcvt_s32_f32(result));
4950 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4951 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4955 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4956 EVT VT = Op.getNode()->getValueType(0);
4957 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4960 bool ExtraOp = false;
4961 switch (Op.getOpcode()) {
4962 default: assert(0 && "Invalid code");
4963 case ISD::ADDC: Opc = ARMISD::ADDC; break;
4964 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4965 case ISD::SUBC: Opc = ARMISD::SUBC; break;
4966 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4970 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4972 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4973 Op.getOperand(1), Op.getOperand(2));
4976 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
4977 // Monotonic load/store is legal for all targets
4978 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
4981 // Aquire/Release load/store is not legal for targets without a
4982 // dmb or equivalent available.
4988 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
4989 SelectionDAG &DAG, unsigned NewOp) {
4990 DebugLoc dl = Node->getDebugLoc();
4991 assert (Node->getValueType(0) == MVT::i64 &&
4992 "Only know how to expand i64 atomics");
4994 SmallVector<SDValue, 6> Ops;
4995 Ops.push_back(Node->getOperand(0)); // Chain
4996 Ops.push_back(Node->getOperand(1)); // Ptr
4998 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4999 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5000 // High part of Val1
5001 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5002 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5003 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5004 // High part of Val1
5005 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5006 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5007 // High part of Val2
5008 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5009 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5011 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5013 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5014 cast<MemSDNode>(Node)->getMemOperand());
5015 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5016 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5017 Results.push_back(Result.getValue(2));
5020 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5021 switch (Op.getOpcode()) {
5022 default: llvm_unreachable("Don't know how to custom lower this!");
5023 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5024 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5025 case ISD::GlobalAddress:
5026 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5027 LowerGlobalAddressELF(Op, DAG);
5028 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5029 case ISD::SELECT: return LowerSELECT(Op, DAG);
5030 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5031 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5032 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5033 case ISD::VASTART: return LowerVASTART(Op, DAG);
5034 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5035 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5036 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5037 case ISD::SINT_TO_FP:
5038 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5039 case ISD::FP_TO_SINT:
5040 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5041 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5042 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5043 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5044 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5045 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5046 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5047 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5049 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5052 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5053 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5054 case ISD::SRL_PARTS:
5055 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5056 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5057 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5058 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5059 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5060 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5061 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5062 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5063 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5064 case ISD::MUL: return LowerMUL(Op, DAG);
5065 case ISD::SDIV: return LowerSDIV(Op, DAG);
5066 case ISD::UDIV: return LowerUDIV(Op, DAG);
5070 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5071 case ISD::ATOMIC_LOAD:
5072 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5077 /// ReplaceNodeResults - Replace the results of node with an illegal result
5078 /// type with new values built out of custom code.
5079 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5080 SmallVectorImpl<SDValue>&Results,
5081 SelectionDAG &DAG) const {
5083 switch (N->getOpcode()) {
5085 llvm_unreachable("Don't know how to custom expand this!");
5088 Res = ExpandBITCAST(N, DAG);
5092 Res = Expand64BitShift(N, DAG, Subtarget);
5094 case ISD::ATOMIC_LOAD_ADD:
5095 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5097 case ISD::ATOMIC_LOAD_AND:
5098 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5100 case ISD::ATOMIC_LOAD_NAND:
5101 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5103 case ISD::ATOMIC_LOAD_OR:
5104 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5106 case ISD::ATOMIC_LOAD_SUB:
5107 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5109 case ISD::ATOMIC_LOAD_XOR:
5110 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5112 case ISD::ATOMIC_SWAP:
5113 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5115 case ISD::ATOMIC_CMP_SWAP:
5116 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5120 Results.push_back(Res);
5123 //===----------------------------------------------------------------------===//
5124 // ARM Scheduler Hooks
5125 //===----------------------------------------------------------------------===//
5128 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5129 MachineBasicBlock *BB,
5130 unsigned Size) const {
5131 unsigned dest = MI->getOperand(0).getReg();
5132 unsigned ptr = MI->getOperand(1).getReg();
5133 unsigned oldval = MI->getOperand(2).getReg();
5134 unsigned newval = MI->getOperand(3).getReg();
5135 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5136 DebugLoc dl = MI->getDebugLoc();
5137 bool isThumb2 = Subtarget->isThumb2();
5139 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5141 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5142 : ARM::GPRRegisterClass);
5145 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5146 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5147 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5150 unsigned ldrOpc, strOpc;
5152 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5154 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5155 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5158 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5159 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5162 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5163 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5167 MachineFunction *MF = BB->getParent();
5168 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5169 MachineFunction::iterator It = BB;
5170 ++It; // insert the new blocks after the current block
5172 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5173 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5174 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5175 MF->insert(It, loop1MBB);
5176 MF->insert(It, loop2MBB);
5177 MF->insert(It, exitMBB);
5179 // Transfer the remainder of BB and its successor edges to exitMBB.
5180 exitMBB->splice(exitMBB->begin(), BB,
5181 llvm::next(MachineBasicBlock::iterator(MI)),
5183 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5187 // fallthrough --> loop1MBB
5188 BB->addSuccessor(loop1MBB);
5191 // ldrex dest, [ptr]
5195 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5196 if (ldrOpc == ARM::t2LDREX)
5198 AddDefaultPred(MIB);
5199 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5200 .addReg(dest).addReg(oldval));
5201 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5202 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5203 BB->addSuccessor(loop2MBB);
5204 BB->addSuccessor(exitMBB);
5207 // strex scratch, newval, [ptr]
5211 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5212 if (strOpc == ARM::t2STREX)
5214 AddDefaultPred(MIB);
5215 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5216 .addReg(scratch).addImm(0));
5217 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5218 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5219 BB->addSuccessor(loop1MBB);
5220 BB->addSuccessor(exitMBB);
5226 MI->eraseFromParent(); // The instruction is gone now.
5232 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5233 unsigned Size, unsigned BinOpcode) const {
5234 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5235 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5237 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5238 MachineFunction *MF = BB->getParent();
5239 MachineFunction::iterator It = BB;
5242 unsigned dest = MI->getOperand(0).getReg();
5243 unsigned ptr = MI->getOperand(1).getReg();
5244 unsigned incr = MI->getOperand(2).getReg();
5245 DebugLoc dl = MI->getDebugLoc();
5246 bool isThumb2 = Subtarget->isThumb2();
5248 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5250 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5251 MRI.constrainRegClass(ptr, 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 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5272 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5273 MF->insert(It, loopMBB);
5274 MF->insert(It, exitMBB);
5276 // Transfer the remainder of BB and its successor edges to exitMBB.
5277 exitMBB->splice(exitMBB->begin(), BB,
5278 llvm::next(MachineBasicBlock::iterator(MI)),
5280 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5282 TargetRegisterClass *TRC =
5283 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5284 unsigned scratch = MRI.createVirtualRegister(TRC);
5285 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5289 // fallthrough --> loopMBB
5290 BB->addSuccessor(loopMBB);
5294 // <binop> scratch2, dest, incr
5295 // strex scratch, scratch2, ptr
5298 // fallthrough --> exitMBB
5300 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5301 if (ldrOpc == ARM::t2LDREX)
5303 AddDefaultPred(MIB);
5305 // operand order needs to go the other way for NAND
5306 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5307 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5308 addReg(incr).addReg(dest)).addReg(0);
5310 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5311 addReg(dest).addReg(incr)).addReg(0);
5314 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5315 if (strOpc == ARM::t2STREX)
5317 AddDefaultPred(MIB);
5318 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5319 .addReg(scratch).addImm(0));
5320 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5321 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5323 BB->addSuccessor(loopMBB);
5324 BB->addSuccessor(exitMBB);
5330 MI->eraseFromParent(); // The instruction is gone now.
5336 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5337 MachineBasicBlock *BB,
5340 ARMCC::CondCodes Cond) const {
5341 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5343 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5344 MachineFunction *MF = BB->getParent();
5345 MachineFunction::iterator It = BB;
5348 unsigned dest = MI->getOperand(0).getReg();
5349 unsigned ptr = MI->getOperand(1).getReg();
5350 unsigned incr = MI->getOperand(2).getReg();
5351 unsigned oldval = dest;
5352 DebugLoc dl = MI->getDebugLoc();
5353 bool isThumb2 = Subtarget->isThumb2();
5355 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5357 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5358 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5361 unsigned ldrOpc, strOpc, extendOpc;
5363 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5365 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5366 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5367 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5370 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5371 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5372 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5375 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5376 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5381 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5382 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5383 MF->insert(It, loopMBB);
5384 MF->insert(It, exitMBB);
5386 // Transfer the remainder of BB and its successor edges to exitMBB.
5387 exitMBB->splice(exitMBB->begin(), BB,
5388 llvm::next(MachineBasicBlock::iterator(MI)),
5390 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5392 TargetRegisterClass *TRC =
5393 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5394 unsigned scratch = MRI.createVirtualRegister(TRC);
5395 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5399 // fallthrough --> loopMBB
5400 BB->addSuccessor(loopMBB);
5404 // (sign extend dest, if required)
5406 // cmov.cond scratch2, dest, incr
5407 // strex scratch, scratch2, ptr
5410 // fallthrough --> exitMBB
5412 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5413 if (ldrOpc == ARM::t2LDREX)
5415 AddDefaultPred(MIB);
5417 // Sign extend the value, if necessary.
5418 if (signExtend && extendOpc) {
5419 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5420 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5425 // Build compare and cmov instructions.
5426 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5427 .addReg(oldval).addReg(incr));
5428 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5429 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5431 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5432 if (strOpc == ARM::t2STREX)
5434 AddDefaultPred(MIB);
5435 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5436 .addReg(scratch).addImm(0));
5437 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5438 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5440 BB->addSuccessor(loopMBB);
5441 BB->addSuccessor(exitMBB);
5447 MI->eraseFromParent(); // The instruction is gone now.
5453 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5454 unsigned Op1, unsigned Op2,
5455 bool NeedsCarry, bool IsCmpxchg) const {
5456 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5457 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5459 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5460 MachineFunction *MF = BB->getParent();
5461 MachineFunction::iterator It = BB;
5464 unsigned destlo = MI->getOperand(0).getReg();
5465 unsigned desthi = MI->getOperand(1).getReg();
5466 unsigned ptr = MI->getOperand(2).getReg();
5467 unsigned vallo = MI->getOperand(3).getReg();
5468 unsigned valhi = MI->getOperand(4).getReg();
5469 DebugLoc dl = MI->getDebugLoc();
5470 bool isThumb2 = Subtarget->isThumb2();
5472 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5474 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5475 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5476 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5479 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5480 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5482 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5483 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5485 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5486 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5488 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5489 MF->insert(It, loopMBB);
5491 MF->insert(It, contBB);
5492 MF->insert(It, cont2BB);
5494 MF->insert(It, exitMBB);
5496 // Transfer the remainder of BB and its successor edges to exitMBB.
5497 exitMBB->splice(exitMBB->begin(), BB,
5498 llvm::next(MachineBasicBlock::iterator(MI)),
5500 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5502 TargetRegisterClass *TRC =
5503 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5504 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5508 // fallthrough --> loopMBB
5509 BB->addSuccessor(loopMBB);
5512 // ldrexd r2, r3, ptr
5513 // <binopa> r0, r2, incr
5514 // <binopb> r1, r3, incr
5515 // strexd storesuccess, r0, r1, ptr
5516 // cmp storesuccess, #0
5518 // fallthrough --> exitMBB
5520 // Note that the registers are explicitly specified because there is not any
5521 // way to force the register allocator to allocate a register pair.
5523 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5524 // need to properly enforce the restriction that the two output registers
5525 // for ldrexd must be different.
5528 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5529 .addReg(ARM::R2, RegState::Define)
5530 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5531 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5532 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5533 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5537 for (unsigned i = 0; i < 2; i++) {
5538 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5540 .addReg(i == 0 ? destlo : desthi)
5541 .addReg(i == 0 ? vallo : valhi));
5542 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5543 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5544 BB->addSuccessor(exitMBB);
5545 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5546 BB = (i == 0 ? contBB : cont2BB);
5549 // Copy to physregs for strexd
5550 unsigned setlo = MI->getOperand(5).getReg();
5551 unsigned sethi = MI->getOperand(6).getReg();
5552 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5553 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5555 // Perform binary operation
5556 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5557 .addReg(destlo).addReg(vallo))
5558 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5559 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5560 .addReg(desthi).addReg(valhi)).addReg(0);
5562 // Copy to physregs for strexd
5563 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5564 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5568 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5569 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5571 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5572 .addReg(storesuccess).addImm(0));
5573 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5574 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5576 BB->addSuccessor(loopMBB);
5577 BB->addSuccessor(exitMBB);
5583 MI->eraseFromParent(); // The instruction is gone now.
5588 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5589 /// registers the function context.
5590 void ARMTargetLowering::
5591 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5592 MachineBasicBlock *DispatchBB, int FI) const {
5593 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5594 DebugLoc dl = MI->getDebugLoc();
5595 MachineFunction *MF = MBB->getParent();
5596 MachineRegisterInfo *MRI = &MF->getRegInfo();
5597 MachineConstantPool *MCP = MF->getConstantPool();
5598 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5599 const Function *F = MF->getFunction();
5601 bool isThumb = Subtarget->isThumb();
5602 bool isThumb2 = Subtarget->isThumb2();
5604 unsigned PCLabelId = AFI->createPICLabelUId();
5605 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5606 ARMConstantPoolValue *CPV =
5607 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5608 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5610 const TargetRegisterClass *TRC =
5611 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5613 // Grab constant pool and fixed stack memory operands.
5614 MachineMemOperand *CPMMO =
5615 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5616 MachineMemOperand::MOLoad, 4, 4);
5618 MachineMemOperand *FIMMOSt =
5619 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5620 MachineMemOperand::MOStore, 4, 4);
5622 // Load the address of the dispatch MBB into the jump buffer.
5624 // Incoming value: jbuf
5625 // ldr.n r5, LCPI1_1
5628 // str r5, [$jbuf, #+4] ; &jbuf[1]
5629 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5630 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5631 .addConstantPoolIndex(CPI)
5632 .addMemOperand(CPMMO));
5633 // Set the low bit because of thumb mode.
5634 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5636 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5637 .addReg(NewVReg1, RegState::Kill)
5639 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5640 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5641 .addReg(NewVReg2, RegState::Kill)
5643 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5644 .addReg(NewVReg3, RegState::Kill)
5646 .addImm(36) // &jbuf[1] :: pc
5647 .addMemOperand(FIMMOSt));
5648 } else if (isThumb) {
5649 // Incoming value: jbuf
5650 // ldr.n r1, LCPI1_4
5654 // add r2, $jbuf, #+4 ; &jbuf[1]
5656 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5657 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5658 .addConstantPoolIndex(CPI)
5659 .addMemOperand(CPMMO));
5660 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5661 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5662 .addReg(NewVReg1, RegState::Kill)
5664 // Set the low bit because of thumb mode.
5665 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5666 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5667 .addReg(ARM::CPSR, RegState::Define)
5669 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5670 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5671 .addReg(ARM::CPSR, RegState::Define)
5672 .addReg(NewVReg2, RegState::Kill)
5673 .addReg(NewVReg3, RegState::Kill));
5674 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5675 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5677 .addImm(36)); // &jbuf[1] :: pc
5678 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5679 .addReg(NewVReg4, RegState::Kill)
5680 .addReg(NewVReg5, RegState::Kill)
5682 .addMemOperand(FIMMOSt));
5684 // Incoming value: jbuf
5687 // str r1, [$jbuf, #+4] ; &jbuf[1]
5688 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5689 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5690 .addConstantPoolIndex(CPI)
5692 .addMemOperand(CPMMO));
5693 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5694 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5695 .addReg(NewVReg1, RegState::Kill)
5696 .addImm(PCLabelId));
5697 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5698 .addReg(NewVReg2, RegState::Kill)
5700 .addImm(36) // &jbuf[1] :: pc
5701 .addMemOperand(FIMMOSt));
5705 MachineBasicBlock *ARMTargetLowering::
5706 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5707 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5708 DebugLoc dl = MI->getDebugLoc();
5709 MachineFunction *MF = MBB->getParent();
5710 MachineRegisterInfo *MRI = &MF->getRegInfo();
5711 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5712 MachineFrameInfo *MFI = MF->getFrameInfo();
5713 int FI = MFI->getFunctionContextIndex();
5715 const TargetRegisterClass *TRC =
5716 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5718 // Get a mapping of the call site numbers to all of the landing pads they're
5720 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5721 unsigned MaxCSNum = 0;
5722 MachineModuleInfo &MMI = MF->getMMI();
5723 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5724 if (!BB->isLandingPad()) continue;
5726 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5728 for (MachineBasicBlock::iterator
5729 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5730 if (!II->isEHLabel()) continue;
5732 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5733 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5735 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5736 for (SmallVectorImpl<unsigned>::iterator
5737 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5738 CSI != CSE; ++CSI) {
5739 CallSiteNumToLPad[*CSI].push_back(BB);
5740 MaxCSNum = std::max(MaxCSNum, *CSI);
5746 // Get an ordered list of the machine basic blocks for the jump table.
5747 std::vector<MachineBasicBlock*> LPadList;
5748 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5749 LPadList.reserve(CallSiteNumToLPad.size());
5750 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5751 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5752 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5753 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5754 LPadList.push_back(*II);
5755 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5759 assert(!LPadList.empty() &&
5760 "No landing pad destinations for the dispatch jump table!");
5762 // Create the jump table and associated information.
5763 MachineJumpTableInfo *JTI =
5764 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5765 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5766 unsigned UId = AFI->createJumpTableUId();
5768 // Create the MBBs for the dispatch code.
5770 // Shove the dispatch's address into the return slot in the function context.
5771 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5772 DispatchBB->setIsLandingPad();
5774 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5775 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5776 DispatchBB->addSuccessor(TrapBB);
5778 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5779 DispatchBB->addSuccessor(DispContBB);
5782 MF->insert(MF->end(), DispatchBB);
5783 MF->insert(MF->end(), DispContBB);
5784 MF->insert(MF->end(), TrapBB);
5786 // Insert code into the entry block that creates and registers the function
5788 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5790 MachineMemOperand *FIMMOLd =
5791 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5792 MachineMemOperand::MOLoad |
5793 MachineMemOperand::MOVolatile, 4, 4);
5795 BuildMI(DispatchBB, dl, TII->get(ARM::eh_sjlj_dispatchsetup));
5797 unsigned NumLPads = LPadList.size();
5798 if (Subtarget->isThumb2()) {
5799 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5800 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5803 .addMemOperand(FIMMOLd));
5805 if (NumLPads < 256) {
5806 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5808 .addImm(LPadList.size()));
5810 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5811 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5812 .addImm(NumLPads & 0xFFFF));
5814 unsigned VReg2 = VReg1;
5815 if ((NumLPads & 0xFFFF0000) != 0) {
5816 VReg2 = MRI->createVirtualRegister(TRC);
5817 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5819 .addImm(NumLPads >> 16));
5822 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5827 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5832 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5833 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5834 .addJumpTableIndex(MJTI)
5837 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5840 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5841 .addReg(NewVReg3, RegState::Kill)
5843 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5845 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5846 .addReg(NewVReg4, RegState::Kill)
5848 .addJumpTableIndex(MJTI)
5850 } else if (Subtarget->isThumb()) {
5851 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5852 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5855 .addMemOperand(FIMMOLd));
5857 if (NumLPads < 256) {
5858 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5862 MachineConstantPool *ConstantPool = MF->getConstantPool();
5863 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5864 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5866 // MachineConstantPool wants an explicit alignment.
5867 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5869 Align = getTargetData()->getTypeAllocSize(C->getType());
5870 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5872 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5873 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5874 .addReg(VReg1, RegState::Define)
5875 .addConstantPoolIndex(Idx));
5876 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5881 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5886 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5887 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5888 .addReg(ARM::CPSR, RegState::Define)
5892 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5893 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5894 .addJumpTableIndex(MJTI)
5897 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5898 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5899 .addReg(ARM::CPSR, RegState::Define)
5900 .addReg(NewVReg2, RegState::Kill)
5903 MachineMemOperand *JTMMOLd =
5904 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5905 MachineMemOperand::MOLoad, 4, 4);
5907 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5908 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5909 .addReg(NewVReg4, RegState::Kill)
5911 .addMemOperand(JTMMOLd));
5913 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5914 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5915 .addReg(ARM::CPSR, RegState::Define)
5916 .addReg(NewVReg5, RegState::Kill)
5919 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5920 .addReg(NewVReg6, RegState::Kill)
5921 .addJumpTableIndex(MJTI)
5924 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5925 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5928 .addMemOperand(FIMMOLd));
5930 if (NumLPads < 256) {
5931 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5934 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5935 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5936 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5937 .addImm(NumLPads & 0xFFFF));
5939 unsigned VReg2 = VReg1;
5940 if ((NumLPads & 0xFFFF0000) != 0) {
5941 VReg2 = MRI->createVirtualRegister(TRC);
5942 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5944 .addImm(NumLPads >> 16));
5947 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5951 MachineConstantPool *ConstantPool = MF->getConstantPool();
5952 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5953 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5955 // MachineConstantPool wants an explicit alignment.
5956 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5958 Align = getTargetData()->getTypeAllocSize(C->getType());
5959 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5961 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5962 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
5963 .addReg(VReg1, RegState::Define)
5964 .addConstantPoolIndex(Idx)
5966 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5968 .addReg(VReg1, RegState::Kill));
5971 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
5976 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5978 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
5980 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5981 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5982 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
5983 .addJumpTableIndex(MJTI)
5986 MachineMemOperand *JTMMOLd =
5987 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5988 MachineMemOperand::MOLoad, 4, 4);
5989 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5991 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
5992 .addReg(NewVReg3, RegState::Kill)
5995 .addMemOperand(JTMMOLd));
5997 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
5998 .addReg(NewVReg5, RegState::Kill)
6000 .addJumpTableIndex(MJTI)
6004 // Add the jump table entries as successors to the MBB.
6005 MachineBasicBlock *PrevMBB = 0;
6006 for (std::vector<MachineBasicBlock*>::iterator
6007 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6008 MachineBasicBlock *CurMBB = *I;
6009 if (PrevMBB != CurMBB)
6010 DispContBB->addSuccessor(CurMBB);
6014 // N.B. the order the invoke BBs are processed in doesn't matter here.
6015 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6016 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6017 const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
6018 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6019 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6020 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6021 MachineBasicBlock *BB = *I;
6023 // Remove the landing pad successor from the invoke block and replace it
6024 // with the new dispatch block.
6025 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6027 while (!Successors.empty()) {
6028 MachineBasicBlock *SMBB = Successors.pop_back_val();
6029 if (SMBB->isLandingPad()) {
6030 BB->removeSuccessor(SMBB);
6031 MBBLPads.push_back(SMBB);
6035 BB->addSuccessor(DispatchBB);
6037 // Find the invoke call and mark all of the callee-saved registers as
6038 // 'implicit defined' so that they're spilled. This prevents code from
6039 // moving instructions to before the EH block, where they will never be
6041 for (MachineBasicBlock::reverse_iterator
6042 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6043 if (!II->isCall()) continue;
6045 DenseMap<unsigned, bool> DefRegs;
6046 for (MachineInstr::mop_iterator
6047 OI = II->operands_begin(), OE = II->operands_end();
6049 if (!OI->isReg()) continue;
6050 DefRegs[OI->getReg()] = true;
6053 MachineInstrBuilder MIB(&*II);
6055 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6056 unsigned Reg = SavedRegs[i];
6057 if (Subtarget->isThumb2() &&
6058 !ARM::tGPRRegisterClass->contains(Reg) &&
6059 !ARM::hGPRRegisterClass->contains(Reg))
6061 else if (Subtarget->isThumb1Only() &&
6062 !ARM::tGPRRegisterClass->contains(Reg))
6064 else if (!Subtarget->isThumb() &&
6065 !ARM::GPRRegisterClass->contains(Reg))
6068 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6075 // Mark all former landing pads as non-landing pads. The dispatch is the only
6077 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6078 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6079 (*I)->setIsLandingPad(false);
6081 // The instruction is gone now.
6082 MI->eraseFromParent();
6088 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6089 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6090 E = MBB->succ_end(); I != E; ++I)
6093 llvm_unreachable("Expecting a BB with two successors!");
6097 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6098 MachineBasicBlock *BB) const {
6099 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6100 DebugLoc dl = MI->getDebugLoc();
6101 bool isThumb2 = Subtarget->isThumb2();
6102 switch (MI->getOpcode()) {
6105 llvm_unreachable("Unexpected instr type to insert");
6107 // The Thumb2 pre-indexed stores have the same MI operands, they just
6108 // define them differently in the .td files from the isel patterns, so
6109 // they need pseudos.
6110 case ARM::t2STR_preidx:
6111 MI->setDesc(TII->get(ARM::t2STR_PRE));
6113 case ARM::t2STRB_preidx:
6114 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6116 case ARM::t2STRH_preidx:
6117 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6120 case ARM::STRi_preidx:
6121 case ARM::STRBi_preidx: {
6122 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6123 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6124 // Decode the offset.
6125 unsigned Offset = MI->getOperand(4).getImm();
6126 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6127 Offset = ARM_AM::getAM2Offset(Offset);
6131 MachineMemOperand *MMO = *MI->memoperands_begin();
6132 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6133 .addOperand(MI->getOperand(0)) // Rn_wb
6134 .addOperand(MI->getOperand(1)) // Rt
6135 .addOperand(MI->getOperand(2)) // Rn
6136 .addImm(Offset) // offset (skip GPR==zero_reg)
6137 .addOperand(MI->getOperand(5)) // pred
6138 .addOperand(MI->getOperand(6))
6139 .addMemOperand(MMO);
6140 MI->eraseFromParent();
6143 case ARM::STRr_preidx:
6144 case ARM::STRBr_preidx:
6145 case ARM::STRH_preidx: {
6147 switch (MI->getOpcode()) {
6148 default: llvm_unreachable("unexpected opcode!");
6149 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6150 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6151 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6153 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6154 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6155 MIB.addOperand(MI->getOperand(i));
6156 MI->eraseFromParent();
6159 case ARM::ATOMIC_LOAD_ADD_I8:
6160 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6161 case ARM::ATOMIC_LOAD_ADD_I16:
6162 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6163 case ARM::ATOMIC_LOAD_ADD_I32:
6164 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6166 case ARM::ATOMIC_LOAD_AND_I8:
6167 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6168 case ARM::ATOMIC_LOAD_AND_I16:
6169 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6170 case ARM::ATOMIC_LOAD_AND_I32:
6171 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6173 case ARM::ATOMIC_LOAD_OR_I8:
6174 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6175 case ARM::ATOMIC_LOAD_OR_I16:
6176 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6177 case ARM::ATOMIC_LOAD_OR_I32:
6178 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6180 case ARM::ATOMIC_LOAD_XOR_I8:
6181 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6182 case ARM::ATOMIC_LOAD_XOR_I16:
6183 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6184 case ARM::ATOMIC_LOAD_XOR_I32:
6185 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6187 case ARM::ATOMIC_LOAD_NAND_I8:
6188 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6189 case ARM::ATOMIC_LOAD_NAND_I16:
6190 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6191 case ARM::ATOMIC_LOAD_NAND_I32:
6192 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6194 case ARM::ATOMIC_LOAD_SUB_I8:
6195 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6196 case ARM::ATOMIC_LOAD_SUB_I16:
6197 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6198 case ARM::ATOMIC_LOAD_SUB_I32:
6199 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6201 case ARM::ATOMIC_LOAD_MIN_I8:
6202 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6203 case ARM::ATOMIC_LOAD_MIN_I16:
6204 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6205 case ARM::ATOMIC_LOAD_MIN_I32:
6206 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6208 case ARM::ATOMIC_LOAD_MAX_I8:
6209 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6210 case ARM::ATOMIC_LOAD_MAX_I16:
6211 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6212 case ARM::ATOMIC_LOAD_MAX_I32:
6213 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6215 case ARM::ATOMIC_LOAD_UMIN_I8:
6216 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6217 case ARM::ATOMIC_LOAD_UMIN_I16:
6218 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6219 case ARM::ATOMIC_LOAD_UMIN_I32:
6220 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6222 case ARM::ATOMIC_LOAD_UMAX_I8:
6223 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6224 case ARM::ATOMIC_LOAD_UMAX_I16:
6225 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6226 case ARM::ATOMIC_LOAD_UMAX_I32:
6227 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6229 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6230 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6231 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6233 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6234 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6235 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6238 case ARM::ATOMADD6432:
6239 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6240 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6241 /*NeedsCarry*/ true);
6242 case ARM::ATOMSUB6432:
6243 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6244 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6245 /*NeedsCarry*/ true);
6246 case ARM::ATOMOR6432:
6247 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6248 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6249 case ARM::ATOMXOR6432:
6250 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6251 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6252 case ARM::ATOMAND6432:
6253 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6254 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6255 case ARM::ATOMSWAP6432:
6256 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6257 case ARM::ATOMCMPXCHG6432:
6258 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6259 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6260 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6262 case ARM::tMOVCCr_pseudo: {
6263 // To "insert" a SELECT_CC instruction, we actually have to insert the
6264 // diamond control-flow pattern. The incoming instruction knows the
6265 // destination vreg to set, the condition code register to branch on, the
6266 // true/false values to select between, and a branch opcode to use.
6267 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6268 MachineFunction::iterator It = BB;
6274 // cmpTY ccX, r1, r2
6276 // fallthrough --> copy0MBB
6277 MachineBasicBlock *thisMBB = BB;
6278 MachineFunction *F = BB->getParent();
6279 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6280 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6281 F->insert(It, copy0MBB);
6282 F->insert(It, sinkMBB);
6284 // Transfer the remainder of BB and its successor edges to sinkMBB.
6285 sinkMBB->splice(sinkMBB->begin(), BB,
6286 llvm::next(MachineBasicBlock::iterator(MI)),
6288 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6290 BB->addSuccessor(copy0MBB);
6291 BB->addSuccessor(sinkMBB);
6293 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6294 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6297 // %FalseValue = ...
6298 // # fallthrough to sinkMBB
6301 // Update machine-CFG edges
6302 BB->addSuccessor(sinkMBB);
6305 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6308 BuildMI(*BB, BB->begin(), dl,
6309 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6310 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6311 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6313 MI->eraseFromParent(); // The pseudo instruction is gone now.
6318 case ARM::BCCZi64: {
6319 // If there is an unconditional branch to the other successor, remove it.
6320 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6322 // Compare both parts that make up the double comparison separately for
6324 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6326 unsigned LHS1 = MI->getOperand(1).getReg();
6327 unsigned LHS2 = MI->getOperand(2).getReg();
6329 AddDefaultPred(BuildMI(BB, dl,
6330 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6331 .addReg(LHS1).addImm(0));
6332 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6333 .addReg(LHS2).addImm(0)
6334 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6336 unsigned RHS1 = MI->getOperand(3).getReg();
6337 unsigned RHS2 = MI->getOperand(4).getReg();
6338 AddDefaultPred(BuildMI(BB, dl,
6339 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6340 .addReg(LHS1).addReg(RHS1));
6341 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6342 .addReg(LHS2).addReg(RHS2)
6343 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6346 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6347 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6348 if (MI->getOperand(0).getImm() == ARMCC::NE)
6349 std::swap(destMBB, exitMBB);
6351 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6352 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6354 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6356 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6358 MI->eraseFromParent(); // The pseudo instruction is gone now.
6362 case ARM::Int_eh_sjlj_setjmp:
6363 case ARM::Int_eh_sjlj_setjmp_nofp:
6364 case ARM::tInt_eh_sjlj_setjmp:
6365 case ARM::t2Int_eh_sjlj_setjmp:
6366 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6367 EmitSjLjDispatchBlock(MI, BB);
6372 // To insert an ABS instruction, we have to insert the
6373 // diamond control-flow pattern. The incoming instruction knows the
6374 // source vreg to test against 0, the destination vreg to set,
6375 // the condition code register to branch on, the
6376 // true/false values to select between, and a branch opcode to use.
6381 // BCC (branch to SinkBB if V0 >= 0)
6382 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6383 // SinkBB: V1 = PHI(V2, V3)
6384 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6385 MachineFunction::iterator BBI = BB;
6387 MachineFunction *Fn = BB->getParent();
6388 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6389 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6390 Fn->insert(BBI, RSBBB);
6391 Fn->insert(BBI, SinkBB);
6393 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6394 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6395 bool isThumb2 = Subtarget->isThumb2();
6396 MachineRegisterInfo &MRI = Fn->getRegInfo();
6397 // In Thumb mode S must not be specified if source register is the SP or
6398 // PC and if destination register is the SP, so restrict register class
6399 unsigned NewMovDstReg = MRI.createVirtualRegister(
6400 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6401 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6402 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6404 // Transfer the remainder of BB and its successor edges to sinkMBB.
6405 SinkBB->splice(SinkBB->begin(), BB,
6406 llvm::next(MachineBasicBlock::iterator(MI)),
6408 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6410 BB->addSuccessor(RSBBB);
6411 BB->addSuccessor(SinkBB);
6413 // fall through to SinkMBB
6414 RSBBB->addSuccessor(SinkBB);
6416 // insert a movs at the end of BB
6417 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6419 .addReg(ABSSrcReg, RegState::Kill)
6420 .addImm((unsigned)ARMCC::AL).addReg(0)
6421 .addReg(ARM::CPSR, RegState::Define);
6423 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6425 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6426 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6428 // insert rsbri in RSBBB
6429 // Note: BCC and rsbri will be converted into predicated rsbmi
6430 // by if-conversion pass
6431 BuildMI(*RSBBB, RSBBB->begin(), dl,
6432 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6433 .addReg(NewMovDstReg, RegState::Kill)
6434 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6436 // insert PHI in SinkBB,
6437 // reuse ABSDstReg to not change uses of ABS instruction
6438 BuildMI(*SinkBB, SinkBB->begin(), dl,
6439 TII->get(ARM::PHI), ABSDstReg)
6440 .addReg(NewRsbDstReg).addMBB(RSBBB)
6441 .addReg(NewMovDstReg).addMBB(BB);
6443 // remove ABS instruction
6444 MI->eraseFromParent();
6446 // return last added BB
6452 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6453 SDNode *Node) const {
6454 if (!MI->hasPostISelHook()) {
6455 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6456 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6460 const MCInstrDesc *MCID = &MI->getDesc();
6461 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6462 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6463 // operand is still set to noreg. If needed, set the optional operand's
6464 // register to CPSR, and remove the redundant implicit def.
6466 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6468 // Rename pseudo opcodes.
6469 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6471 const ARMBaseInstrInfo *TII =
6472 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6473 MCID = &TII->get(NewOpc);
6475 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6476 "converted opcode should be the same except for cc_out");
6480 // Add the optional cc_out operand
6481 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6483 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6485 // Any ARM instruction that sets the 's' bit should specify an optional
6486 // "cc_out" operand in the last operand position.
6487 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6488 assert(!NewOpc && "Optional cc_out operand required");
6491 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6492 // since we already have an optional CPSR def.
6493 bool definesCPSR = false;
6494 bool deadCPSR = false;
6495 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6497 const MachineOperand &MO = MI->getOperand(i);
6498 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6502 MI->RemoveOperand(i);
6507 assert(!NewOpc && "Optional cc_out operand required");
6510 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6512 assert(!MI->getOperand(ccOutIdx).getReg() &&
6513 "expect uninitialized optional cc_out operand");
6517 // If this instruction was defined with an optional CPSR def and its dag node
6518 // had a live implicit CPSR def, then activate the optional CPSR def.
6519 MachineOperand &MO = MI->getOperand(ccOutIdx);
6520 MO.setReg(ARM::CPSR);
6524 //===----------------------------------------------------------------------===//
6525 // ARM Optimization Hooks
6526 //===----------------------------------------------------------------------===//
6529 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6530 TargetLowering::DAGCombinerInfo &DCI) {
6531 SelectionDAG &DAG = DCI.DAG;
6532 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6533 EVT VT = N->getValueType(0);
6534 unsigned Opc = N->getOpcode();
6535 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6536 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6537 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6538 ISD::CondCode CC = ISD::SETCC_INVALID;
6541 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6543 SDValue CCOp = Slct.getOperand(0);
6544 if (CCOp.getOpcode() == ISD::SETCC)
6545 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6548 bool DoXform = false;
6550 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6553 if (LHS.getOpcode() == ISD::Constant &&
6554 cast<ConstantSDNode>(LHS)->isNullValue()) {
6556 } else if (CC != ISD::SETCC_INVALID &&
6557 RHS.getOpcode() == ISD::Constant &&
6558 cast<ConstantSDNode>(RHS)->isNullValue()) {
6559 std::swap(LHS, RHS);
6560 SDValue Op0 = Slct.getOperand(0);
6561 EVT OpVT = isSlctCC ? Op0.getValueType() :
6562 Op0.getOperand(0).getValueType();
6563 bool isInt = OpVT.isInteger();
6564 CC = ISD::getSetCCInverse(CC, isInt);
6566 if (!TLI.isCondCodeLegal(CC, OpVT))
6567 return SDValue(); // Inverse operator isn't legal.
6574 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6576 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6577 Slct.getOperand(0), Slct.getOperand(1), CC);
6578 SDValue CCOp = Slct.getOperand(0);
6580 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6581 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6582 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6583 CCOp, OtherOp, Result);
6588 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6589 // (only after legalization).
6590 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6591 TargetLowering::DAGCombinerInfo &DCI,
6592 const ARMSubtarget *Subtarget) {
6594 // Only perform optimization if after legalize, and if NEON is available. We
6595 // also expected both operands to be BUILD_VECTORs.
6596 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6597 || N0.getOpcode() != ISD::BUILD_VECTOR
6598 || N1.getOpcode() != ISD::BUILD_VECTOR)
6601 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6602 EVT VT = N->getValueType(0);
6603 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6606 // Check that the vector operands are of the right form.
6607 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6608 // operands, where N is the size of the formed vector.
6609 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6610 // index such that we have a pair wise add pattern.
6612 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6613 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6615 SDValue Vec = N0->getOperand(0)->getOperand(0);
6616 SDNode *V = Vec.getNode();
6617 unsigned nextIndex = 0;
6619 // For each operands to the ADD which are BUILD_VECTORs,
6620 // check to see if each of their operands are an EXTRACT_VECTOR with
6621 // the same vector and appropriate index.
6622 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6623 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6624 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6626 SDValue ExtVec0 = N0->getOperand(i);
6627 SDValue ExtVec1 = N1->getOperand(i);
6629 // First operand is the vector, verify its the same.
6630 if (V != ExtVec0->getOperand(0).getNode() ||
6631 V != ExtVec1->getOperand(0).getNode())
6634 // Second is the constant, verify its correct.
6635 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6636 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6638 // For the constant, we want to see all the even or all the odd.
6639 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6640 || C1->getZExtValue() != nextIndex+1)
6649 // Create VPADDL node.
6650 SelectionDAG &DAG = DCI.DAG;
6651 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6653 // Build operand list.
6654 SmallVector<SDValue, 8> Ops;
6655 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6656 TLI.getPointerTy()));
6658 // Input is the vector.
6661 // Get widened type and narrowed type.
6663 unsigned numElem = VT.getVectorNumElements();
6664 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6665 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6666 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6667 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6669 assert(0 && "Invalid vector element type for padd optimization.");
6672 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6673 widenType, &Ops[0], Ops.size());
6674 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6677 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6678 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6679 /// called with the default operands, and if that fails, with commuted
6681 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6682 TargetLowering::DAGCombinerInfo &DCI,
6683 const ARMSubtarget *Subtarget){
6685 // Attempt to create vpaddl for this add.
6686 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6687 if (Result.getNode())
6690 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6691 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6692 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6693 if (Result.getNode()) return Result;
6698 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6700 static SDValue PerformADDCombine(SDNode *N,
6701 TargetLowering::DAGCombinerInfo &DCI,
6702 const ARMSubtarget *Subtarget) {
6703 SDValue N0 = N->getOperand(0);
6704 SDValue N1 = N->getOperand(1);
6706 // First try with the default operand order.
6707 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6708 if (Result.getNode())
6711 // If that didn't work, try again with the operands commuted.
6712 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6715 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6717 static SDValue PerformSUBCombine(SDNode *N,
6718 TargetLowering::DAGCombinerInfo &DCI) {
6719 SDValue N0 = N->getOperand(0);
6720 SDValue N1 = N->getOperand(1);
6722 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6723 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6724 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6725 if (Result.getNode()) return Result;
6731 /// PerformVMULCombine
6732 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6733 /// special multiplier accumulator forwarding.
6739 static SDValue PerformVMULCombine(SDNode *N,
6740 TargetLowering::DAGCombinerInfo &DCI,
6741 const ARMSubtarget *Subtarget) {
6742 if (!Subtarget->hasVMLxForwarding())
6745 SelectionDAG &DAG = DCI.DAG;
6746 SDValue N0 = N->getOperand(0);
6747 SDValue N1 = N->getOperand(1);
6748 unsigned Opcode = N0.getOpcode();
6749 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6750 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6751 Opcode = N1.getOpcode();
6752 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6753 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6758 EVT VT = N->getValueType(0);
6759 DebugLoc DL = N->getDebugLoc();
6760 SDValue N00 = N0->getOperand(0);
6761 SDValue N01 = N0->getOperand(1);
6762 return DAG.getNode(Opcode, DL, VT,
6763 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6764 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6767 static SDValue PerformMULCombine(SDNode *N,
6768 TargetLowering::DAGCombinerInfo &DCI,
6769 const ARMSubtarget *Subtarget) {
6770 SelectionDAG &DAG = DCI.DAG;
6772 if (Subtarget->isThumb1Only())
6775 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6778 EVT VT = N->getValueType(0);
6779 if (VT.is64BitVector() || VT.is128BitVector())
6780 return PerformVMULCombine(N, DCI, Subtarget);
6784 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6788 uint64_t MulAmt = C->getZExtValue();
6789 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6790 ShiftAmt = ShiftAmt & (32 - 1);
6791 SDValue V = N->getOperand(0);
6792 DebugLoc DL = N->getDebugLoc();
6795 MulAmt >>= ShiftAmt;
6796 if (isPowerOf2_32(MulAmt - 1)) {
6797 // (mul x, 2^N + 1) => (add (shl x, N), x)
6798 Res = DAG.getNode(ISD::ADD, DL, VT,
6799 V, DAG.getNode(ISD::SHL, DL, VT,
6800 V, DAG.getConstant(Log2_32(MulAmt-1),
6802 } else if (isPowerOf2_32(MulAmt + 1)) {
6803 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6804 Res = DAG.getNode(ISD::SUB, DL, VT,
6805 DAG.getNode(ISD::SHL, DL, VT,
6806 V, DAG.getConstant(Log2_32(MulAmt+1),
6813 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6814 DAG.getConstant(ShiftAmt, MVT::i32));
6816 // Do not add new nodes to DAG combiner worklist.
6817 DCI.CombineTo(N, Res, false);
6821 static SDValue PerformANDCombine(SDNode *N,
6822 TargetLowering::DAGCombinerInfo &DCI) {
6824 // Attempt to use immediate-form VBIC
6825 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6826 DebugLoc dl = N->getDebugLoc();
6827 EVT VT = N->getValueType(0);
6828 SelectionDAG &DAG = DCI.DAG;
6830 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6833 APInt SplatBits, SplatUndef;
6834 unsigned SplatBitSize;
6837 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6838 if (SplatBitSize <= 64) {
6840 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6841 SplatUndef.getZExtValue(), SplatBitSize,
6842 DAG, VbicVT, VT.is128BitVector(),
6844 if (Val.getNode()) {
6846 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6847 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6848 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6856 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6857 static SDValue PerformORCombine(SDNode *N,
6858 TargetLowering::DAGCombinerInfo &DCI,
6859 const ARMSubtarget *Subtarget) {
6860 // Attempt to use immediate-form VORR
6861 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6862 DebugLoc dl = N->getDebugLoc();
6863 EVT VT = N->getValueType(0);
6864 SelectionDAG &DAG = DCI.DAG;
6866 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6869 APInt SplatBits, SplatUndef;
6870 unsigned SplatBitSize;
6872 if (BVN && Subtarget->hasNEON() &&
6873 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6874 if (SplatBitSize <= 64) {
6876 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6877 SplatUndef.getZExtValue(), SplatBitSize,
6878 DAG, VorrVT, VT.is128BitVector(),
6880 if (Val.getNode()) {
6882 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6883 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6884 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6889 SDValue N0 = N->getOperand(0);
6890 if (N0.getOpcode() != ISD::AND)
6892 SDValue N1 = N->getOperand(1);
6894 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6895 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6896 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6898 unsigned SplatBitSize;
6901 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6903 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6904 HasAnyUndefs) && !HasAnyUndefs) {
6905 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6907 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6908 HasAnyUndefs) && !HasAnyUndefs &&
6909 SplatBits0 == ~SplatBits1) {
6910 // Canonicalize the vector type to make instruction selection simpler.
6911 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6912 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6913 N0->getOperand(1), N0->getOperand(0),
6915 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6920 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6923 // BFI is only available on V6T2+
6924 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6927 DebugLoc DL = N->getDebugLoc();
6928 // 1) or (and A, mask), val => ARMbfi A, val, mask
6929 // iff (val & mask) == val
6931 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6932 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6933 // && mask == ~mask2
6934 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6935 // && ~mask == mask2
6936 // (i.e., copy a bitfield value into another bitfield of the same width)
6941 SDValue N00 = N0.getOperand(0);
6943 // The value and the mask need to be constants so we can verify this is
6944 // actually a bitfield set. If the mask is 0xffff, we can do better
6945 // via a movt instruction, so don't use BFI in that case.
6946 SDValue MaskOp = N0.getOperand(1);
6947 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6950 unsigned Mask = MaskC->getZExtValue();
6954 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6955 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6957 unsigned Val = N1C->getZExtValue();
6958 if ((Val & ~Mask) != Val)
6961 if (ARM::isBitFieldInvertedMask(Mask)) {
6962 Val >>= CountTrailingZeros_32(~Mask);
6964 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6965 DAG.getConstant(Val, MVT::i32),
6966 DAG.getConstant(Mask, MVT::i32));
6968 // Do not add new nodes to DAG combiner worklist.
6969 DCI.CombineTo(N, Res, false);
6972 } else if (N1.getOpcode() == ISD::AND) {
6973 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6974 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6977 unsigned Mask2 = N11C->getZExtValue();
6979 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
6981 if (ARM::isBitFieldInvertedMask(Mask) &&
6983 // The pack halfword instruction works better for masks that fit it,
6984 // so use that when it's available.
6985 if (Subtarget->hasT2ExtractPack() &&
6986 (Mask == 0xffff || Mask == 0xffff0000))
6989 unsigned amt = CountTrailingZeros_32(Mask2);
6990 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6991 DAG.getConstant(amt, MVT::i32));
6992 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6993 DAG.getConstant(Mask, MVT::i32));
6994 // Do not add new nodes to DAG combiner worklist.
6995 DCI.CombineTo(N, Res, false);
6997 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
6999 // The pack halfword instruction works better for masks that fit it,
7000 // so use that when it's available.
7001 if (Subtarget->hasT2ExtractPack() &&
7002 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7005 unsigned lsb = CountTrailingZeros_32(Mask);
7006 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7007 DAG.getConstant(lsb, MVT::i32));
7008 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7009 DAG.getConstant(Mask2, MVT::i32));
7010 // Do not add new nodes to DAG combiner worklist.
7011 DCI.CombineTo(N, Res, false);
7016 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7017 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7018 ARM::isBitFieldInvertedMask(~Mask)) {
7019 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7020 // where lsb(mask) == #shamt and masked bits of B are known zero.
7021 SDValue ShAmt = N00.getOperand(1);
7022 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7023 unsigned LSB = CountTrailingZeros_32(Mask);
7027 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7028 DAG.getConstant(~Mask, MVT::i32));
7030 // Do not add new nodes to DAG combiner worklist.
7031 DCI.CombineTo(N, Res, false);
7037 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7038 /// the bits being cleared by the AND are not demanded by the BFI.
7039 static SDValue PerformBFICombine(SDNode *N,
7040 TargetLowering::DAGCombinerInfo &DCI) {
7041 SDValue N1 = N->getOperand(1);
7042 if (N1.getOpcode() == ISD::AND) {
7043 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7046 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7047 unsigned LSB = CountTrailingZeros_32(~InvMask);
7048 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7049 unsigned Mask = (1 << Width)-1;
7050 unsigned Mask2 = N11C->getZExtValue();
7051 if ((Mask & (~Mask2)) == 0)
7052 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7053 N->getOperand(0), N1.getOperand(0),
7059 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7060 /// ARMISD::VMOVRRD.
7061 static SDValue PerformVMOVRRDCombine(SDNode *N,
7062 TargetLowering::DAGCombinerInfo &DCI) {
7063 // vmovrrd(vmovdrr x, y) -> x,y
7064 SDValue InDouble = N->getOperand(0);
7065 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7066 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7068 // vmovrrd(load f64) -> (load i32), (load i32)
7069 SDNode *InNode = InDouble.getNode();
7070 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7071 InNode->getValueType(0) == MVT::f64 &&
7072 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7073 !cast<LoadSDNode>(InNode)->isVolatile()) {
7074 // TODO: Should this be done for non-FrameIndex operands?
7075 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7077 SelectionDAG &DAG = DCI.DAG;
7078 DebugLoc DL = LD->getDebugLoc();
7079 SDValue BasePtr = LD->getBasePtr();
7080 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7081 LD->getPointerInfo(), LD->isVolatile(),
7082 LD->isNonTemporal(), LD->isInvariant(),
7083 LD->getAlignment());
7085 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7086 DAG.getConstant(4, MVT::i32));
7087 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7088 LD->getPointerInfo(), LD->isVolatile(),
7089 LD->isNonTemporal(), LD->isInvariant(),
7090 std::min(4U, LD->getAlignment() / 2));
7092 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7093 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7094 DCI.RemoveFromWorklist(LD);
7102 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7103 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7104 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7105 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7106 SDValue Op0 = N->getOperand(0);
7107 SDValue Op1 = N->getOperand(1);
7108 if (Op0.getOpcode() == ISD::BITCAST)
7109 Op0 = Op0.getOperand(0);
7110 if (Op1.getOpcode() == ISD::BITCAST)
7111 Op1 = Op1.getOperand(0);
7112 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7113 Op0.getNode() == Op1.getNode() &&
7114 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7115 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7116 N->getValueType(0), Op0.getOperand(0));
7120 /// PerformSTORECombine - Target-specific dag combine xforms for
7122 static SDValue PerformSTORECombine(SDNode *N,
7123 TargetLowering::DAGCombinerInfo &DCI) {
7124 // Bitcast an i64 store extracted from a vector to f64.
7125 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7126 StoreSDNode *St = cast<StoreSDNode>(N);
7127 SDValue StVal = St->getValue();
7128 if (!ISD::isNormalStore(St) || St->isVolatile())
7131 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7132 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7133 SelectionDAG &DAG = DCI.DAG;
7134 DebugLoc DL = St->getDebugLoc();
7135 SDValue BasePtr = St->getBasePtr();
7136 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7137 StVal.getNode()->getOperand(0), BasePtr,
7138 St->getPointerInfo(), St->isVolatile(),
7139 St->isNonTemporal(), St->getAlignment());
7141 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7142 DAG.getConstant(4, MVT::i32));
7143 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7144 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7145 St->isNonTemporal(),
7146 std::min(4U, St->getAlignment() / 2));
7149 if (StVal.getValueType() != MVT::i64 ||
7150 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7153 SelectionDAG &DAG = DCI.DAG;
7154 DebugLoc dl = StVal.getDebugLoc();
7155 SDValue IntVec = StVal.getOperand(0);
7156 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7157 IntVec.getValueType().getVectorNumElements());
7158 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7159 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7160 Vec, StVal.getOperand(1));
7161 dl = N->getDebugLoc();
7162 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7163 // Make the DAGCombiner fold the bitcasts.
7164 DCI.AddToWorklist(Vec.getNode());
7165 DCI.AddToWorklist(ExtElt.getNode());
7166 DCI.AddToWorklist(V.getNode());
7167 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7168 St->getPointerInfo(), St->isVolatile(),
7169 St->isNonTemporal(), St->getAlignment(),
7173 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7174 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7175 /// i64 vector to have f64 elements, since the value can then be loaded
7176 /// directly into a VFP register.
7177 static bool hasNormalLoadOperand(SDNode *N) {
7178 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7179 for (unsigned i = 0; i < NumElts; ++i) {
7180 SDNode *Elt = N->getOperand(i).getNode();
7181 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7187 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7188 /// ISD::BUILD_VECTOR.
7189 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7190 TargetLowering::DAGCombinerInfo &DCI){
7191 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7192 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7193 // into a pair of GPRs, which is fine when the value is used as a scalar,
7194 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7195 SelectionDAG &DAG = DCI.DAG;
7196 if (N->getNumOperands() == 2) {
7197 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7202 // Load i64 elements as f64 values so that type legalization does not split
7203 // them up into i32 values.
7204 EVT VT = N->getValueType(0);
7205 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7207 DebugLoc dl = N->getDebugLoc();
7208 SmallVector<SDValue, 8> Ops;
7209 unsigned NumElts = VT.getVectorNumElements();
7210 for (unsigned i = 0; i < NumElts; ++i) {
7211 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7213 // Make the DAGCombiner fold the bitcast.
7214 DCI.AddToWorklist(V.getNode());
7216 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7217 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7218 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7221 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7222 /// ISD::INSERT_VECTOR_ELT.
7223 static SDValue PerformInsertEltCombine(SDNode *N,
7224 TargetLowering::DAGCombinerInfo &DCI) {
7225 // Bitcast an i64 load inserted into a vector to f64.
7226 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7227 EVT VT = N->getValueType(0);
7228 SDNode *Elt = N->getOperand(1).getNode();
7229 if (VT.getVectorElementType() != MVT::i64 ||
7230 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7233 SelectionDAG &DAG = DCI.DAG;
7234 DebugLoc dl = N->getDebugLoc();
7235 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7236 VT.getVectorNumElements());
7237 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7238 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7239 // Make the DAGCombiner fold the bitcasts.
7240 DCI.AddToWorklist(Vec.getNode());
7241 DCI.AddToWorklist(V.getNode());
7242 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7243 Vec, V, N->getOperand(2));
7244 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7247 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7248 /// ISD::VECTOR_SHUFFLE.
7249 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7250 // The LLVM shufflevector instruction does not require the shuffle mask
7251 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7252 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7253 // operands do not match the mask length, they are extended by concatenating
7254 // them with undef vectors. That is probably the right thing for other
7255 // targets, but for NEON it is better to concatenate two double-register
7256 // size vector operands into a single quad-register size vector. Do that
7257 // transformation here:
7258 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7259 // shuffle(concat(v1, v2), undef)
7260 SDValue Op0 = N->getOperand(0);
7261 SDValue Op1 = N->getOperand(1);
7262 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7263 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7264 Op0.getNumOperands() != 2 ||
7265 Op1.getNumOperands() != 2)
7267 SDValue Concat0Op1 = Op0.getOperand(1);
7268 SDValue Concat1Op1 = Op1.getOperand(1);
7269 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7270 Concat1Op1.getOpcode() != ISD::UNDEF)
7272 // Skip the transformation if any of the types are illegal.
7273 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7274 EVT VT = N->getValueType(0);
7275 if (!TLI.isTypeLegal(VT) ||
7276 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7277 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7280 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7281 Op0.getOperand(0), Op1.getOperand(0));
7282 // Translate the shuffle mask.
7283 SmallVector<int, 16> NewMask;
7284 unsigned NumElts = VT.getVectorNumElements();
7285 unsigned HalfElts = NumElts/2;
7286 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7287 for (unsigned n = 0; n < NumElts; ++n) {
7288 int MaskElt = SVN->getMaskElt(n);
7290 if (MaskElt < (int)HalfElts)
7292 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7293 NewElt = HalfElts + MaskElt - NumElts;
7294 NewMask.push_back(NewElt);
7296 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7297 DAG.getUNDEF(VT), NewMask.data());
7300 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7301 /// NEON load/store intrinsics to merge base address updates.
7302 static SDValue CombineBaseUpdate(SDNode *N,
7303 TargetLowering::DAGCombinerInfo &DCI) {
7304 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7307 SelectionDAG &DAG = DCI.DAG;
7308 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7309 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7310 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7311 SDValue Addr = N->getOperand(AddrOpIdx);
7313 // Search for a use of the address operand that is an increment.
7314 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7315 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7317 if (User->getOpcode() != ISD::ADD ||
7318 UI.getUse().getResNo() != Addr.getResNo())
7321 // Check that the add is independent of the load/store. Otherwise, folding
7322 // it would create a cycle.
7323 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7326 // Find the new opcode for the updating load/store.
7328 bool isLaneOp = false;
7329 unsigned NewOpc = 0;
7330 unsigned NumVecs = 0;
7332 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7334 default: assert(0 && "unexpected intrinsic for Neon base update");
7335 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7337 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7339 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7341 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7343 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7344 NumVecs = 2; isLaneOp = true; break;
7345 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7346 NumVecs = 3; isLaneOp = true; break;
7347 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7348 NumVecs = 4; isLaneOp = true; break;
7349 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7350 NumVecs = 1; isLoad = false; break;
7351 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7352 NumVecs = 2; isLoad = false; break;
7353 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7354 NumVecs = 3; isLoad = false; break;
7355 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7356 NumVecs = 4; isLoad = false; break;
7357 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7358 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7359 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7360 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7361 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7362 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7366 switch (N->getOpcode()) {
7367 default: assert(0 && "unexpected opcode for Neon base update");
7368 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7369 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7370 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7374 // Find the size of memory referenced by the load/store.
7377 VecTy = N->getValueType(0);
7379 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7380 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7382 NumBytes /= VecTy.getVectorNumElements();
7384 // If the increment is a constant, it must match the memory ref size.
7385 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7386 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7387 uint64_t IncVal = CInc->getZExtValue();
7388 if (IncVal != NumBytes)
7390 } else if (NumBytes >= 3 * 16) {
7391 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7392 // separate instructions that make it harder to use a non-constant update.
7396 // Create the new updating load/store node.
7398 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7400 for (n = 0; n < NumResultVecs; ++n)
7402 Tys[n++] = MVT::i32;
7403 Tys[n] = MVT::Other;
7404 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7405 SmallVector<SDValue, 8> Ops;
7406 Ops.push_back(N->getOperand(0)); // incoming chain
7407 Ops.push_back(N->getOperand(AddrOpIdx));
7409 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7410 Ops.push_back(N->getOperand(i));
7412 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7413 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7414 Ops.data(), Ops.size(),
7415 MemInt->getMemoryVT(),
7416 MemInt->getMemOperand());
7419 std::vector<SDValue> NewResults;
7420 for (unsigned i = 0; i < NumResultVecs; ++i) {
7421 NewResults.push_back(SDValue(UpdN.getNode(), i));
7423 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7424 DCI.CombineTo(N, NewResults);
7425 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7432 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7433 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7434 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7436 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7437 SelectionDAG &DAG = DCI.DAG;
7438 EVT VT = N->getValueType(0);
7439 // vldN-dup instructions only support 64-bit vectors for N > 1.
7440 if (!VT.is64BitVector())
7443 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7444 SDNode *VLD = N->getOperand(0).getNode();
7445 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7447 unsigned NumVecs = 0;
7448 unsigned NewOpc = 0;
7449 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7450 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7452 NewOpc = ARMISD::VLD2DUP;
7453 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7455 NewOpc = ARMISD::VLD3DUP;
7456 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7458 NewOpc = ARMISD::VLD4DUP;
7463 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7464 // numbers match the load.
7465 unsigned VLDLaneNo =
7466 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7467 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7469 // Ignore uses of the chain result.
7470 if (UI.getUse().getResNo() == NumVecs)
7473 if (User->getOpcode() != ARMISD::VDUPLANE ||
7474 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7478 // Create the vldN-dup node.
7481 for (n = 0; n < NumVecs; ++n)
7483 Tys[n] = MVT::Other;
7484 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7485 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7486 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7487 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7488 Ops, 2, VLDMemInt->getMemoryVT(),
7489 VLDMemInt->getMemOperand());
7492 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7494 unsigned ResNo = UI.getUse().getResNo();
7495 // Ignore uses of the chain result.
7496 if (ResNo == NumVecs)
7499 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7502 // Now the vldN-lane intrinsic is dead except for its chain result.
7503 // Update uses of the chain.
7504 std::vector<SDValue> VLDDupResults;
7505 for (unsigned n = 0; n < NumVecs; ++n)
7506 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7507 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7508 DCI.CombineTo(VLD, VLDDupResults);
7513 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7514 /// ARMISD::VDUPLANE.
7515 static SDValue PerformVDUPLANECombine(SDNode *N,
7516 TargetLowering::DAGCombinerInfo &DCI) {
7517 SDValue Op = N->getOperand(0);
7519 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7520 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7521 if (CombineVLDDUP(N, DCI))
7522 return SDValue(N, 0);
7524 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7525 // redundant. Ignore bit_converts for now; element sizes are checked below.
7526 while (Op.getOpcode() == ISD::BITCAST)
7527 Op = Op.getOperand(0);
7528 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7531 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7532 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7533 // The canonical VMOV for a zero vector uses a 32-bit element size.
7534 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7536 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7538 EVT VT = N->getValueType(0);
7539 if (EltSize > VT.getVectorElementType().getSizeInBits())
7542 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7545 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7546 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7547 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7551 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7553 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7558 APFloat APF = C->getValueAPF();
7559 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7560 != APFloat::opOK || !isExact)
7563 c0 = (I == 0) ? cN : c0;
7564 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7571 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7572 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7573 /// when the VMUL has a constant operand that is a power of 2.
7575 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7576 /// vmul.f32 d16, d17, d16
7577 /// vcvt.s32.f32 d16, d16
7579 /// vcvt.s32.f32 d16, d16, #3
7580 static SDValue PerformVCVTCombine(SDNode *N,
7581 TargetLowering::DAGCombinerInfo &DCI,
7582 const ARMSubtarget *Subtarget) {
7583 SelectionDAG &DAG = DCI.DAG;
7584 SDValue Op = N->getOperand(0);
7586 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7587 Op.getOpcode() != ISD::FMUL)
7591 SDValue N0 = Op->getOperand(0);
7592 SDValue ConstVec = Op->getOperand(1);
7593 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7595 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7596 !isConstVecPow2(ConstVec, isSigned, C))
7599 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7600 Intrinsic::arm_neon_vcvtfp2fxu;
7601 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7603 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7604 DAG.getConstant(Log2_64(C), MVT::i32));
7607 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7608 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7609 /// when the VDIV has a constant operand that is a power of 2.
7611 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7612 /// vcvt.f32.s32 d16, d16
7613 /// vdiv.f32 d16, d17, d16
7615 /// vcvt.f32.s32 d16, d16, #3
7616 static SDValue PerformVDIVCombine(SDNode *N,
7617 TargetLowering::DAGCombinerInfo &DCI,
7618 const ARMSubtarget *Subtarget) {
7619 SelectionDAG &DAG = DCI.DAG;
7620 SDValue Op = N->getOperand(0);
7621 unsigned OpOpcode = Op.getNode()->getOpcode();
7623 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7624 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7628 SDValue ConstVec = N->getOperand(1);
7629 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7631 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7632 !isConstVecPow2(ConstVec, isSigned, C))
7635 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7636 Intrinsic::arm_neon_vcvtfxu2fp;
7637 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7639 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7640 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7643 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7644 /// operand of a vector shift operation, where all the elements of the
7645 /// build_vector must have the same constant integer value.
7646 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7647 // Ignore bit_converts.
7648 while (Op.getOpcode() == ISD::BITCAST)
7649 Op = Op.getOperand(0);
7650 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7651 APInt SplatBits, SplatUndef;
7652 unsigned SplatBitSize;
7654 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7655 HasAnyUndefs, ElementBits) ||
7656 SplatBitSize > ElementBits)
7658 Cnt = SplatBits.getSExtValue();
7662 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7663 /// operand of a vector shift left operation. That value must be in the range:
7664 /// 0 <= Value < ElementBits for a left shift; or
7665 /// 0 <= Value <= ElementBits for a long left shift.
7666 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7667 assert(VT.isVector() && "vector shift count is not a vector type");
7668 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7669 if (! getVShiftImm(Op, ElementBits, Cnt))
7671 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7674 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7675 /// operand of a vector shift right operation. For a shift opcode, the value
7676 /// is positive, but for an intrinsic the value count must be negative. The
7677 /// absolute value must be in the range:
7678 /// 1 <= |Value| <= ElementBits for a right shift; or
7679 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7680 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7682 assert(VT.isVector() && "vector shift count is not a vector type");
7683 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7684 if (! getVShiftImm(Op, ElementBits, Cnt))
7688 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7691 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7692 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7693 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7696 // Don't do anything for most intrinsics.
7699 // Vector shifts: check for immediate versions and lower them.
7700 // Note: This is done during DAG combining instead of DAG legalizing because
7701 // the build_vectors for 64-bit vector element shift counts are generally
7702 // not legal, and it is hard to see their values after they get legalized to
7703 // loads from a constant pool.
7704 case Intrinsic::arm_neon_vshifts:
7705 case Intrinsic::arm_neon_vshiftu:
7706 case Intrinsic::arm_neon_vshiftls:
7707 case Intrinsic::arm_neon_vshiftlu:
7708 case Intrinsic::arm_neon_vshiftn:
7709 case Intrinsic::arm_neon_vrshifts:
7710 case Intrinsic::arm_neon_vrshiftu:
7711 case Intrinsic::arm_neon_vrshiftn:
7712 case Intrinsic::arm_neon_vqshifts:
7713 case Intrinsic::arm_neon_vqshiftu:
7714 case Intrinsic::arm_neon_vqshiftsu:
7715 case Intrinsic::arm_neon_vqshiftns:
7716 case Intrinsic::arm_neon_vqshiftnu:
7717 case Intrinsic::arm_neon_vqshiftnsu:
7718 case Intrinsic::arm_neon_vqrshiftns:
7719 case Intrinsic::arm_neon_vqrshiftnu:
7720 case Intrinsic::arm_neon_vqrshiftnsu: {
7721 EVT VT = N->getOperand(1).getValueType();
7723 unsigned VShiftOpc = 0;
7726 case Intrinsic::arm_neon_vshifts:
7727 case Intrinsic::arm_neon_vshiftu:
7728 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7729 VShiftOpc = ARMISD::VSHL;
7732 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7733 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7734 ARMISD::VSHRs : ARMISD::VSHRu);
7739 case Intrinsic::arm_neon_vshiftls:
7740 case Intrinsic::arm_neon_vshiftlu:
7741 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7743 llvm_unreachable("invalid shift count for vshll intrinsic");
7745 case Intrinsic::arm_neon_vrshifts:
7746 case Intrinsic::arm_neon_vrshiftu:
7747 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7751 case Intrinsic::arm_neon_vqshifts:
7752 case Intrinsic::arm_neon_vqshiftu:
7753 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7757 case Intrinsic::arm_neon_vqshiftsu:
7758 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7760 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7762 case Intrinsic::arm_neon_vshiftn:
7763 case Intrinsic::arm_neon_vrshiftn:
7764 case Intrinsic::arm_neon_vqshiftns:
7765 case Intrinsic::arm_neon_vqshiftnu:
7766 case Intrinsic::arm_neon_vqshiftnsu:
7767 case Intrinsic::arm_neon_vqrshiftns:
7768 case Intrinsic::arm_neon_vqrshiftnu:
7769 case Intrinsic::arm_neon_vqrshiftnsu:
7770 // Narrowing shifts require an immediate right shift.
7771 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7773 llvm_unreachable("invalid shift count for narrowing vector shift "
7777 llvm_unreachable("unhandled vector shift");
7781 case Intrinsic::arm_neon_vshifts:
7782 case Intrinsic::arm_neon_vshiftu:
7783 // Opcode already set above.
7785 case Intrinsic::arm_neon_vshiftls:
7786 case Intrinsic::arm_neon_vshiftlu:
7787 if (Cnt == VT.getVectorElementType().getSizeInBits())
7788 VShiftOpc = ARMISD::VSHLLi;
7790 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7791 ARMISD::VSHLLs : ARMISD::VSHLLu);
7793 case Intrinsic::arm_neon_vshiftn:
7794 VShiftOpc = ARMISD::VSHRN; break;
7795 case Intrinsic::arm_neon_vrshifts:
7796 VShiftOpc = ARMISD::VRSHRs; break;
7797 case Intrinsic::arm_neon_vrshiftu:
7798 VShiftOpc = ARMISD::VRSHRu; break;
7799 case Intrinsic::arm_neon_vrshiftn:
7800 VShiftOpc = ARMISD::VRSHRN; break;
7801 case Intrinsic::arm_neon_vqshifts:
7802 VShiftOpc = ARMISD::VQSHLs; break;
7803 case Intrinsic::arm_neon_vqshiftu:
7804 VShiftOpc = ARMISD::VQSHLu; break;
7805 case Intrinsic::arm_neon_vqshiftsu:
7806 VShiftOpc = ARMISD::VQSHLsu; break;
7807 case Intrinsic::arm_neon_vqshiftns:
7808 VShiftOpc = ARMISD::VQSHRNs; break;
7809 case Intrinsic::arm_neon_vqshiftnu:
7810 VShiftOpc = ARMISD::VQSHRNu; break;
7811 case Intrinsic::arm_neon_vqshiftnsu:
7812 VShiftOpc = ARMISD::VQSHRNsu; break;
7813 case Intrinsic::arm_neon_vqrshiftns:
7814 VShiftOpc = ARMISD::VQRSHRNs; break;
7815 case Intrinsic::arm_neon_vqrshiftnu:
7816 VShiftOpc = ARMISD::VQRSHRNu; break;
7817 case Intrinsic::arm_neon_vqrshiftnsu:
7818 VShiftOpc = ARMISD::VQRSHRNsu; break;
7821 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7822 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7825 case Intrinsic::arm_neon_vshiftins: {
7826 EVT VT = N->getOperand(1).getValueType();
7828 unsigned VShiftOpc = 0;
7830 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7831 VShiftOpc = ARMISD::VSLI;
7832 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7833 VShiftOpc = ARMISD::VSRI;
7835 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7838 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7839 N->getOperand(1), N->getOperand(2),
7840 DAG.getConstant(Cnt, MVT::i32));
7843 case Intrinsic::arm_neon_vqrshifts:
7844 case Intrinsic::arm_neon_vqrshiftu:
7845 // No immediate versions of these to check for.
7852 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
7853 /// lowers them. As with the vector shift intrinsics, this is done during DAG
7854 /// combining instead of DAG legalizing because the build_vectors for 64-bit
7855 /// vector element shift counts are generally not legal, and it is hard to see
7856 /// their values after they get legalized to loads from a constant pool.
7857 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7858 const ARMSubtarget *ST) {
7859 EVT VT = N->getValueType(0);
7861 // Nothing to be done for scalar shifts.
7862 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7863 if (!VT.isVector() || !TLI.isTypeLegal(VT))
7866 assert(ST->hasNEON() && "unexpected vector shift");
7869 switch (N->getOpcode()) {
7870 default: llvm_unreachable("unexpected shift opcode");
7873 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7874 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7875 DAG.getConstant(Cnt, MVT::i32));
7880 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7881 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7882 ARMISD::VSHRs : ARMISD::VSHRu);
7883 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7884 DAG.getConstant(Cnt, MVT::i32));
7890 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7891 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7892 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7893 const ARMSubtarget *ST) {
7894 SDValue N0 = N->getOperand(0);
7896 // Check for sign- and zero-extensions of vector extract operations of 8-
7897 // and 16-bit vector elements. NEON supports these directly. They are
7898 // handled during DAG combining because type legalization will promote them
7899 // to 32-bit types and it is messy to recognize the operations after that.
7900 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7901 SDValue Vec = N0.getOperand(0);
7902 SDValue Lane = N0.getOperand(1);
7903 EVT VT = N->getValueType(0);
7904 EVT EltVT = N0.getValueType();
7905 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7907 if (VT == MVT::i32 &&
7908 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7909 TLI.isTypeLegal(Vec.getValueType()) &&
7910 isa<ConstantSDNode>(Lane)) {
7913 switch (N->getOpcode()) {
7914 default: llvm_unreachable("unexpected opcode");
7915 case ISD::SIGN_EXTEND:
7916 Opc = ARMISD::VGETLANEs;
7918 case ISD::ZERO_EXTEND:
7919 case ISD::ANY_EXTEND:
7920 Opc = ARMISD::VGETLANEu;
7923 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7930 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7931 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7932 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7933 const ARMSubtarget *ST) {
7934 // If the target supports NEON, try to use vmax/vmin instructions for f32
7935 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
7936 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
7937 // a NaN; only do the transformation when it matches that behavior.
7939 // For now only do this when using NEON for FP operations; if using VFP, it
7940 // is not obvious that the benefit outweighs the cost of switching to the
7942 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7943 N->getValueType(0) != MVT::f32)
7946 SDValue CondLHS = N->getOperand(0);
7947 SDValue CondRHS = N->getOperand(1);
7948 SDValue LHS = N->getOperand(2);
7949 SDValue RHS = N->getOperand(3);
7950 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7952 unsigned Opcode = 0;
7954 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7955 IsReversed = false; // x CC y ? x : y
7956 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7957 IsReversed = true ; // x CC y ? y : x
7971 // If LHS is NaN, an ordered comparison will be false and the result will
7972 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
7973 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7974 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
7975 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7977 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
7978 // will return -0, so vmin can only be used for unsafe math or if one of
7979 // the operands is known to be nonzero.
7980 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
7981 !DAG.getTarget().Options.UnsafeFPMath &&
7982 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7984 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7993 // If LHS is NaN, an ordered comparison will be false and the result will
7994 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
7995 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7996 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7997 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7999 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8000 // will return +0, so vmax can only be used for unsafe math or if one of
8001 // the operands is known to be nonzero.
8002 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8003 !DAG.getTarget().Options.UnsafeFPMath &&
8004 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8006 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8012 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8015 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8017 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8018 SDValue Cmp = N->getOperand(4);
8019 if (Cmp.getOpcode() != ARMISD::CMPZ)
8020 // Only looking at EQ and NE cases.
8023 EVT VT = N->getValueType(0);
8024 DebugLoc dl = N->getDebugLoc();
8025 SDValue LHS = Cmp.getOperand(0);
8026 SDValue RHS = Cmp.getOperand(1);
8027 SDValue FalseVal = N->getOperand(0);
8028 SDValue TrueVal = N->getOperand(1);
8029 SDValue ARMcc = N->getOperand(2);
8030 ARMCC::CondCodes CC =
8031 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8049 /// FIXME: Turn this into a target neutral optimization?
8051 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8052 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8053 N->getOperand(3), Cmp);
8054 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8056 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8057 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8058 N->getOperand(3), NewCmp);
8061 if (Res.getNode()) {
8062 APInt KnownZero, KnownOne;
8063 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8064 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8065 // Capture demanded bits information that would be otherwise lost.
8066 if (KnownZero == 0xfffffffe)
8067 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8068 DAG.getValueType(MVT::i1));
8069 else if (KnownZero == 0xffffff00)
8070 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8071 DAG.getValueType(MVT::i8));
8072 else if (KnownZero == 0xffff0000)
8073 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8074 DAG.getValueType(MVT::i16));
8080 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8081 DAGCombinerInfo &DCI) const {
8082 switch (N->getOpcode()) {
8084 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8085 case ISD::SUB: return PerformSUBCombine(N, DCI);
8086 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8087 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8088 case ISD::AND: return PerformANDCombine(N, DCI);
8089 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8090 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8091 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8092 case ISD::STORE: return PerformSTORECombine(N, DCI);
8093 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8094 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8095 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8096 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8097 case ISD::FP_TO_SINT:
8098 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8099 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8100 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8103 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8104 case ISD::SIGN_EXTEND:
8105 case ISD::ZERO_EXTEND:
8106 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8107 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8108 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8109 case ARMISD::VLD2DUP:
8110 case ARMISD::VLD3DUP:
8111 case ARMISD::VLD4DUP:
8112 return CombineBaseUpdate(N, DCI);
8113 case ISD::INTRINSIC_VOID:
8114 case ISD::INTRINSIC_W_CHAIN:
8115 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8116 case Intrinsic::arm_neon_vld1:
8117 case Intrinsic::arm_neon_vld2:
8118 case Intrinsic::arm_neon_vld3:
8119 case Intrinsic::arm_neon_vld4:
8120 case Intrinsic::arm_neon_vld2lane:
8121 case Intrinsic::arm_neon_vld3lane:
8122 case Intrinsic::arm_neon_vld4lane:
8123 case Intrinsic::arm_neon_vst1:
8124 case Intrinsic::arm_neon_vst2:
8125 case Intrinsic::arm_neon_vst3:
8126 case Intrinsic::arm_neon_vst4:
8127 case Intrinsic::arm_neon_vst2lane:
8128 case Intrinsic::arm_neon_vst3lane:
8129 case Intrinsic::arm_neon_vst4lane:
8130 return CombineBaseUpdate(N, DCI);
8138 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8140 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8143 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8144 if (!Subtarget->allowsUnalignedMem())
8147 switch (VT.getSimpleVT().SimpleTy) {
8154 // FIXME: VLD1 etc with standard alignment is legal.
8158 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8159 unsigned AlignCheck) {
8160 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8161 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8164 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8165 unsigned DstAlign, unsigned SrcAlign,
8168 MachineFunction &MF) const {
8169 const Function *F = MF.getFunction();
8171 // See if we can use NEON instructions for this...
8173 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8174 Subtarget->hasNEON()) {
8175 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8177 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8182 // Lowering to i32/i16 if the size permits.
8185 } else if (Size >= 2) {
8189 // Let the target-independent logic figure it out.
8193 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8198 switch (VT.getSimpleVT().SimpleTy) {
8199 default: return false;
8214 if ((V & (Scale - 1)) != 0)
8217 return V == (V & ((1LL << 5) - 1));
8220 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8221 const ARMSubtarget *Subtarget) {
8228 switch (VT.getSimpleVT().SimpleTy) {
8229 default: return false;
8234 // + imm12 or - imm8
8236 return V == (V & ((1LL << 8) - 1));
8237 return V == (V & ((1LL << 12) - 1));
8240 // Same as ARM mode. FIXME: NEON?
8241 if (!Subtarget->hasVFP2())
8246 return V == (V & ((1LL << 8) - 1));
8250 /// isLegalAddressImmediate - Return true if the integer value can be used
8251 /// as the offset of the target addressing mode for load / store of the
8253 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8254 const ARMSubtarget *Subtarget) {
8261 if (Subtarget->isThumb1Only())
8262 return isLegalT1AddressImmediate(V, VT);
8263 else if (Subtarget->isThumb2())
8264 return isLegalT2AddressImmediate(V, VT, Subtarget);
8269 switch (VT.getSimpleVT().SimpleTy) {
8270 default: return false;
8275 return V == (V & ((1LL << 12) - 1));
8278 return V == (V & ((1LL << 8) - 1));
8281 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8286 return V == (V & ((1LL << 8) - 1));
8290 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8292 int Scale = AM.Scale;
8296 switch (VT.getSimpleVT().SimpleTy) {
8297 default: return false;
8306 return Scale == 2 || Scale == 4 || Scale == 8;
8309 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8313 // Note, we allow "void" uses (basically, uses that aren't loads or
8314 // stores), because arm allows folding a scale into many arithmetic
8315 // operations. This should be made more precise and revisited later.
8317 // Allow r << imm, but the imm has to be a multiple of two.
8318 if (Scale & 1) return false;
8319 return isPowerOf2_32(Scale);
8323 /// isLegalAddressingMode - Return true if the addressing mode represented
8324 /// by AM is legal for this target, for a load/store of the specified type.
8325 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8327 EVT VT = getValueType(Ty, true);
8328 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8331 // Can never fold addr of global into load/store.
8336 case 0: // no scale reg, must be "r+i" or "r", or "i".
8339 if (Subtarget->isThumb1Only())
8343 // ARM doesn't support any R+R*scale+imm addr modes.
8350 if (Subtarget->isThumb2())
8351 return isLegalT2ScaledAddressingMode(AM, VT);
8353 int Scale = AM.Scale;
8354 switch (VT.getSimpleVT().SimpleTy) {
8355 default: return false;
8359 if (Scale < 0) Scale = -Scale;
8363 return isPowerOf2_32(Scale & ~1);
8367 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8372 // Note, we allow "void" uses (basically, uses that aren't loads or
8373 // stores), because arm allows folding a scale into many arithmetic
8374 // operations. This should be made more precise and revisited later.
8376 // Allow r << imm, but the imm has to be a multiple of two.
8377 if (Scale & 1) return false;
8378 return isPowerOf2_32(Scale);
8385 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8386 /// icmp immediate, that is the target has icmp instructions which can compare
8387 /// a register against the immediate without having to materialize the
8388 /// immediate into a register.
8389 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8390 if (!Subtarget->isThumb())
8391 return ARM_AM::getSOImmVal(Imm) != -1;
8392 if (Subtarget->isThumb2())
8393 return ARM_AM::getT2SOImmVal(Imm) != -1;
8394 return Imm >= 0 && Imm <= 255;
8397 /// isLegalAddImmediate - Return true if the specified immediate is legal
8398 /// add immediate, that is the target has add instructions which can add
8399 /// a register with the immediate without having to materialize the
8400 /// immediate into a register.
8401 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8402 return ARM_AM::getSOImmVal(Imm) != -1;
8405 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8406 bool isSEXTLoad, SDValue &Base,
8407 SDValue &Offset, bool &isInc,
8408 SelectionDAG &DAG) {
8409 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8412 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8414 Base = Ptr->getOperand(0);
8415 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8416 int RHSC = (int)RHS->getZExtValue();
8417 if (RHSC < 0 && RHSC > -256) {
8418 assert(Ptr->getOpcode() == ISD::ADD);
8420 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8424 isInc = (Ptr->getOpcode() == ISD::ADD);
8425 Offset = Ptr->getOperand(1);
8427 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8429 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8430 int RHSC = (int)RHS->getZExtValue();
8431 if (RHSC < 0 && RHSC > -0x1000) {
8432 assert(Ptr->getOpcode() == ISD::ADD);
8434 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8435 Base = Ptr->getOperand(0);
8440 if (Ptr->getOpcode() == ISD::ADD) {
8442 ARM_AM::ShiftOpc ShOpcVal=
8443 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8444 if (ShOpcVal != ARM_AM::no_shift) {
8445 Base = Ptr->getOperand(1);
8446 Offset = Ptr->getOperand(0);
8448 Base = Ptr->getOperand(0);
8449 Offset = Ptr->getOperand(1);
8454 isInc = (Ptr->getOpcode() == ISD::ADD);
8455 Base = Ptr->getOperand(0);
8456 Offset = Ptr->getOperand(1);
8460 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8464 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8465 bool isSEXTLoad, SDValue &Base,
8466 SDValue &Offset, bool &isInc,
8467 SelectionDAG &DAG) {
8468 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8471 Base = Ptr->getOperand(0);
8472 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8473 int RHSC = (int)RHS->getZExtValue();
8474 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8475 assert(Ptr->getOpcode() == ISD::ADD);
8477 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8479 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8480 isInc = Ptr->getOpcode() == ISD::ADD;
8481 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8489 /// getPreIndexedAddressParts - returns true by value, base pointer and
8490 /// offset pointer and addressing mode by reference if the node's address
8491 /// can be legally represented as pre-indexed load / store address.
8493 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8495 ISD::MemIndexedMode &AM,
8496 SelectionDAG &DAG) const {
8497 if (Subtarget->isThumb1Only())
8502 bool isSEXTLoad = false;
8503 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8504 Ptr = LD->getBasePtr();
8505 VT = LD->getMemoryVT();
8506 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8507 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8508 Ptr = ST->getBasePtr();
8509 VT = ST->getMemoryVT();
8514 bool isLegal = false;
8515 if (Subtarget->isThumb2())
8516 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8517 Offset, isInc, DAG);
8519 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8520 Offset, isInc, DAG);
8524 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8528 /// getPostIndexedAddressParts - returns true by value, base pointer and
8529 /// offset pointer and addressing mode by reference if this node can be
8530 /// combined with a load / store to form a post-indexed load / store.
8531 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8534 ISD::MemIndexedMode &AM,
8535 SelectionDAG &DAG) const {
8536 if (Subtarget->isThumb1Only())
8541 bool isSEXTLoad = false;
8542 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8543 VT = LD->getMemoryVT();
8544 Ptr = LD->getBasePtr();
8545 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8546 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8547 VT = ST->getMemoryVT();
8548 Ptr = ST->getBasePtr();
8553 bool isLegal = false;
8554 if (Subtarget->isThumb2())
8555 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8558 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8564 // Swap base ptr and offset to catch more post-index load / store when
8565 // it's legal. In Thumb2 mode, offset must be an immediate.
8566 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8567 !Subtarget->isThumb2())
8568 std::swap(Base, Offset);
8570 // Post-indexed load / store update the base pointer.
8575 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8579 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8583 const SelectionDAG &DAG,
8584 unsigned Depth) const {
8585 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8586 switch (Op.getOpcode()) {
8588 case ARMISD::CMOV: {
8589 // Bits are known zero/one if known on the LHS and RHS.
8590 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8591 if (KnownZero == 0 && KnownOne == 0) return;
8593 APInt KnownZeroRHS, KnownOneRHS;
8594 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8595 KnownZeroRHS, KnownOneRHS, Depth+1);
8596 KnownZero &= KnownZeroRHS;
8597 KnownOne &= KnownOneRHS;
8603 //===----------------------------------------------------------------------===//
8604 // ARM Inline Assembly Support
8605 //===----------------------------------------------------------------------===//
8607 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8608 // Looking for "rev" which is V6+.
8609 if (!Subtarget->hasV6Ops())
8612 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8613 std::string AsmStr = IA->getAsmString();
8614 SmallVector<StringRef, 4> AsmPieces;
8615 SplitString(AsmStr, AsmPieces, ";\n");
8617 switch (AsmPieces.size()) {
8618 default: return false;
8620 AsmStr = AsmPieces[0];
8622 SplitString(AsmStr, AsmPieces, " \t,");
8625 if (AsmPieces.size() == 3 &&
8626 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8627 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8628 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8629 if (Ty && Ty->getBitWidth() == 32)
8630 return IntrinsicLowering::LowerToByteSwap(CI);
8638 /// getConstraintType - Given a constraint letter, return the type of
8639 /// constraint it is for this target.
8640 ARMTargetLowering::ConstraintType
8641 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8642 if (Constraint.size() == 1) {
8643 switch (Constraint[0]) {
8645 case 'l': return C_RegisterClass;
8646 case 'w': return C_RegisterClass;
8647 case 'h': return C_RegisterClass;
8648 case 'x': return C_RegisterClass;
8649 case 't': return C_RegisterClass;
8650 case 'j': return C_Other; // Constant for movw.
8651 // An address with a single base register. Due to the way we
8652 // currently handle addresses it is the same as an 'r' memory constraint.
8653 case 'Q': return C_Memory;
8655 } else if (Constraint.size() == 2) {
8656 switch (Constraint[0]) {
8658 // All 'U+' constraints are addresses.
8659 case 'U': return C_Memory;
8662 return TargetLowering::getConstraintType(Constraint);
8665 /// Examine constraint type and operand type and determine a weight value.
8666 /// This object must already have been set up with the operand type
8667 /// and the current alternative constraint selected.
8668 TargetLowering::ConstraintWeight
8669 ARMTargetLowering::getSingleConstraintMatchWeight(
8670 AsmOperandInfo &info, const char *constraint) const {
8671 ConstraintWeight weight = CW_Invalid;
8672 Value *CallOperandVal = info.CallOperandVal;
8673 // If we don't have a value, we can't do a match,
8674 // but allow it at the lowest weight.
8675 if (CallOperandVal == NULL)
8677 Type *type = CallOperandVal->getType();
8678 // Look at the constraint type.
8679 switch (*constraint) {
8681 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8684 if (type->isIntegerTy()) {
8685 if (Subtarget->isThumb())
8686 weight = CW_SpecificReg;
8688 weight = CW_Register;
8692 if (type->isFloatingPointTy())
8693 weight = CW_Register;
8699 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8701 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8703 if (Constraint.size() == 1) {
8704 // GCC ARM Constraint Letters
8705 switch (Constraint[0]) {
8706 case 'l': // Low regs or general regs.
8707 if (Subtarget->isThumb())
8708 return RCPair(0U, ARM::tGPRRegisterClass);
8710 return RCPair(0U, ARM::GPRRegisterClass);
8711 case 'h': // High regs or no regs.
8712 if (Subtarget->isThumb())
8713 return RCPair(0U, ARM::hGPRRegisterClass);
8716 return RCPair(0U, ARM::GPRRegisterClass);
8719 return RCPair(0U, ARM::SPRRegisterClass);
8720 if (VT.getSizeInBits() == 64)
8721 return RCPair(0U, ARM::DPRRegisterClass);
8722 if (VT.getSizeInBits() == 128)
8723 return RCPair(0U, ARM::QPRRegisterClass);
8727 return RCPair(0U, ARM::SPR_8RegisterClass);
8728 if (VT.getSizeInBits() == 64)
8729 return RCPair(0U, ARM::DPR_8RegisterClass);
8730 if (VT.getSizeInBits() == 128)
8731 return RCPair(0U, ARM::QPR_8RegisterClass);
8735 return RCPair(0U, ARM::SPRRegisterClass);
8739 if (StringRef("{cc}").equals_lower(Constraint))
8740 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8742 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8745 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8746 /// vector. If it is invalid, don't add anything to Ops.
8747 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8748 std::string &Constraint,
8749 std::vector<SDValue>&Ops,
8750 SelectionDAG &DAG) const {
8751 SDValue Result(0, 0);
8753 // Currently only support length 1 constraints.
8754 if (Constraint.length() != 1) return;
8756 char ConstraintLetter = Constraint[0];
8757 switch (ConstraintLetter) {
8760 case 'I': case 'J': case 'K': case 'L':
8761 case 'M': case 'N': case 'O':
8762 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8766 int64_t CVal64 = C->getSExtValue();
8767 int CVal = (int) CVal64;
8768 // None of these constraints allow values larger than 32 bits. Check
8769 // that the value fits in an int.
8773 switch (ConstraintLetter) {
8775 // Constant suitable for movw, must be between 0 and
8777 if (Subtarget->hasV6T2Ops())
8778 if (CVal >= 0 && CVal <= 65535)
8782 if (Subtarget->isThumb1Only()) {
8783 // This must be a constant between 0 and 255, for ADD
8785 if (CVal >= 0 && CVal <= 255)
8787 } else if (Subtarget->isThumb2()) {
8788 // A constant that can be used as an immediate value in a
8789 // data-processing instruction.
8790 if (ARM_AM::getT2SOImmVal(CVal) != -1)
8793 // A constant that can be used as an immediate value in a
8794 // data-processing instruction.
8795 if (ARM_AM::getSOImmVal(CVal) != -1)
8801 if (Subtarget->isThumb()) { // FIXME thumb2
8802 // This must be a constant between -255 and -1, for negated ADD
8803 // immediates. This can be used in GCC with an "n" modifier that
8804 // prints the negated value, for use with SUB instructions. It is
8805 // not useful otherwise but is implemented for compatibility.
8806 if (CVal >= -255 && CVal <= -1)
8809 // This must be a constant between -4095 and 4095. It is not clear
8810 // what this constraint is intended for. Implemented for
8811 // compatibility with GCC.
8812 if (CVal >= -4095 && CVal <= 4095)
8818 if (Subtarget->isThumb1Only()) {
8819 // A 32-bit value where only one byte has a nonzero value. Exclude
8820 // zero to match GCC. This constraint is used by GCC internally for
8821 // constants that can be loaded with a move/shift combination.
8822 // It is not useful otherwise but is implemented for compatibility.
8823 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8825 } else if (Subtarget->isThumb2()) {
8826 // A constant whose bitwise inverse can be used as an immediate
8827 // value in a data-processing instruction. This can be used in GCC
8828 // with a "B" modifier that prints the inverted value, for use with
8829 // BIC and MVN instructions. It is not useful otherwise but is
8830 // implemented for compatibility.
8831 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8834 // A constant whose bitwise inverse can be used as an immediate
8835 // value in a data-processing instruction. This can be used in GCC
8836 // with a "B" modifier that prints the inverted value, for use with
8837 // BIC and MVN instructions. It is not useful otherwise but is
8838 // implemented for compatibility.
8839 if (ARM_AM::getSOImmVal(~CVal) != -1)
8845 if (Subtarget->isThumb1Only()) {
8846 // This must be a constant between -7 and 7,
8847 // for 3-operand ADD/SUB immediate instructions.
8848 if (CVal >= -7 && CVal < 7)
8850 } else if (Subtarget->isThumb2()) {
8851 // A constant whose negation can be used as an immediate value in a
8852 // data-processing instruction. This can be used in GCC with an "n"
8853 // modifier that prints the negated value, for use with SUB
8854 // instructions. It is not useful otherwise but is implemented for
8856 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8859 // A constant whose negation can be used as an immediate value in a
8860 // data-processing instruction. This can be used in GCC with an "n"
8861 // modifier that prints the negated value, for use with SUB
8862 // instructions. It is not useful otherwise but is implemented for
8864 if (ARM_AM::getSOImmVal(-CVal) != -1)
8870 if (Subtarget->isThumb()) { // FIXME thumb2
8871 // This must be a multiple of 4 between 0 and 1020, for
8872 // ADD sp + immediate.
8873 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
8876 // A power of two or a constant between 0 and 32. This is used in
8877 // GCC for the shift amount on shifted register operands, but it is
8878 // useful in general for any shift amounts.
8879 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8885 if (Subtarget->isThumb()) { // FIXME thumb2
8886 // This must be a constant between 0 and 31, for shift amounts.
8887 if (CVal >= 0 && CVal <= 31)
8893 if (Subtarget->isThumb()) { // FIXME thumb2
8894 // This must be a multiple of 4 between -508 and 508, for
8895 // ADD/SUB sp = sp + immediate.
8896 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8901 Result = DAG.getTargetConstant(CVal, Op.getValueType());
8905 if (Result.getNode()) {
8906 Ops.push_back(Result);
8909 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8913 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8914 // The ARM target isn't yet aware of offsets.
8918 bool ARM::isBitFieldInvertedMask(unsigned v) {
8919 if (v == 0xffffffff)
8921 // there can be 1's on either or both "outsides", all the "inside"
8923 unsigned int lsb = 0, msb = 31;
8924 while (v & (1 << msb)) --msb;
8925 while (v & (1 << lsb)) ++lsb;
8926 for (unsigned int i = lsb; i <= msb; ++i) {
8933 /// isFPImmLegal - Returns true if the target can instruction select the
8934 /// specified FP immediate natively. If false, the legalizer will
8935 /// materialize the FP immediate as a load from a constant pool.
8936 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8937 if (!Subtarget->hasVFP3())
8940 return ARM_AM::getFP32Imm(Imm) != -1;
8942 return ARM_AM::getFP64Imm(Imm) != -1;
8946 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8947 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
8948 /// specified in the intrinsic calls.
8949 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8951 unsigned Intrinsic) const {
8952 switch (Intrinsic) {
8953 case Intrinsic::arm_neon_vld1:
8954 case Intrinsic::arm_neon_vld2:
8955 case Intrinsic::arm_neon_vld3:
8956 case Intrinsic::arm_neon_vld4:
8957 case Intrinsic::arm_neon_vld2lane:
8958 case Intrinsic::arm_neon_vld3lane:
8959 case Intrinsic::arm_neon_vld4lane: {
8960 Info.opc = ISD::INTRINSIC_W_CHAIN;
8961 // Conservatively set memVT to the entire set of vectors loaded.
8962 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8963 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8964 Info.ptrVal = I.getArgOperand(0);
8966 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8967 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8968 Info.vol = false; // volatile loads with NEON intrinsics not supported
8969 Info.readMem = true;
8970 Info.writeMem = false;
8973 case Intrinsic::arm_neon_vst1:
8974 case Intrinsic::arm_neon_vst2:
8975 case Intrinsic::arm_neon_vst3:
8976 case Intrinsic::arm_neon_vst4:
8977 case Intrinsic::arm_neon_vst2lane:
8978 case Intrinsic::arm_neon_vst3lane:
8979 case Intrinsic::arm_neon_vst4lane: {
8980 Info.opc = ISD::INTRINSIC_VOID;
8981 // Conservatively set memVT to the entire set of vectors stored.
8982 unsigned NumElts = 0;
8983 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8984 Type *ArgTy = I.getArgOperand(ArgI)->getType();
8985 if (!ArgTy->isVectorTy())
8987 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8989 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8990 Info.ptrVal = I.getArgOperand(0);
8992 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8993 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8994 Info.vol = false; // volatile stores with NEON intrinsics not supported
8995 Info.readMem = false;
8996 Info.writeMem = true;
8999 case Intrinsic::arm_strexd: {
9000 Info.opc = ISD::INTRINSIC_W_CHAIN;
9001 Info.memVT = MVT::i64;
9002 Info.ptrVal = I.getArgOperand(2);
9006 Info.readMem = false;
9007 Info.writeMem = true;
9010 case Intrinsic::arm_ldrexd: {
9011 Info.opc = ISD::INTRINSIC_W_CHAIN;
9012 Info.memVT = MVT::i64;
9013 Info.ptrVal = I.getArgOperand(0);
9017 Info.readMem = true;
9018 Info.writeMem = false;