1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "MCTargetDesc/ARMAddressingModes.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/IntrinsicLowering.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
94 EVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
97 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
98 PromotedLdStVT.getSimpleVT());
100 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
101 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
102 PromotedLdStVT.getSimpleVT());
105 EVT ElemTy = VT.getVectorElementType();
106 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
107 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
108 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
109 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
110 if (ElemTy == MVT::i32) {
111 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
113 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
114 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
116 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
119 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
121 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
122 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
123 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
124 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
125 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
126 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
127 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
128 if (VT.isInteger()) {
129 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
130 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
131 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
134 // Promote all bit-wise operations.
135 if (VT.isInteger() && VT != PromotedBitwiseVT) {
136 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
137 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
138 PromotedBitwiseVT.getSimpleVT());
139 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
140 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
141 PromotedBitwiseVT.getSimpleVT());
142 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
143 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
144 PromotedBitwiseVT.getSimpleVT());
147 // Neon does not support vector divide/remainder operations.
148 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
149 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
150 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
156 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
157 addRegisterClass(VT, &ARM::DPRRegClass);
158 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
161 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
162 addRegisterClass(VT, &ARM::QPRRegClass);
163 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
166 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
167 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
168 return new TargetLoweringObjectFileMachO();
170 return new ARMElfTargetObjectFile();
173 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
174 : TargetLowering(TM, createTLOF(TM)) {
175 Subtarget = &TM.getSubtarget<ARMSubtarget>();
176 RegInfo = TM.getRegisterInfo();
177 Itins = TM.getInstrItineraryData();
179 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
181 if (Subtarget->isTargetDarwin()) {
182 // Uses VFP for Thumb libfuncs if available.
183 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
184 // Single-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
186 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
187 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
188 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
190 // Double-precision floating-point arithmetic.
191 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
192 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
193 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
194 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
196 // Single-precision comparisons.
197 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
198 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
199 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
200 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
201 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
202 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
203 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
204 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
206 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
215 // Double-precision comparisons.
216 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
217 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
218 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
219 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
220 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
221 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
222 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
223 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
225 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
234 // Floating-point to integer conversions.
235 // i64 conversions are done via library routines even when generating VFP
236 // instructions, so use the same ones.
237 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
238 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
239 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
240 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
242 // Conversions between floating types.
243 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
244 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
246 // Integer to floating-point conversions.
247 // i64 conversions are done via library routines even when generating VFP
248 // instructions, so use the same ones.
249 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
250 // e.g., __floatunsidf vs. __floatunssidfvfp.
251 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
252 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
253 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
254 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
258 // These libcalls are not available in 32-bit.
259 setLibcallName(RTLIB::SHL_I128, 0);
260 setLibcallName(RTLIB::SRL_I128, 0);
261 setLibcallName(RTLIB::SRA_I128, 0);
263 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
264 // Double-precision floating-point arithmetic helper functions
265 // RTABI chapter 4.1.2, Table 2
266 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
267 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
268 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
269 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
270 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
275 // Double-precision floating-point comparison helper functions
276 // RTABI chapter 4.1.2, Table 3
277 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
278 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
279 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
280 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
281 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
282 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
284 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
286 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
287 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
288 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
289 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
290 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
291 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
292 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
293 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
302 // Single-precision floating-point arithmetic helper functions
303 // RTABI chapter 4.1.2, Table 4
304 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
305 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
306 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
307 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
308 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
313 // Single-precision floating-point comparison helper functions
314 // RTABI chapter 4.1.2, Table 5
315 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
316 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
317 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
318 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
319 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
320 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
322 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
324 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
325 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
326 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
327 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
328 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
329 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
330 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
331 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
340 // Floating-point to integer conversions.
341 // RTABI chapter 4.1.2, Table 6
342 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
343 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
345 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
346 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
347 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
348 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
349 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
350 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
359 // Conversions between floating types.
360 // RTABI chapter 4.1.2, Table 7
361 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
362 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
363 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
366 // Integer to floating-point conversions.
367 // RTABI chapter 4.1.2, Table 8
368 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
369 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
370 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
371 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
372 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
373 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
374 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
375 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385 // Long long helper functions
386 // RTABI chapter 4.2, Table 9
387 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
388 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
389 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
390 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
391 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
398 // Integer division functions
399 // RTABI chapter 4.3.1
400 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
401 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
402 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
403 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
404 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
405 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
407 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
408 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
415 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
418 // RTABI chapter 4.3.4
419 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
420 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
421 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
422 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
423 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
424 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
427 // Use divmod compiler-rt calls for iOS 5.0 and later.
428 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
429 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
430 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
431 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
434 if (Subtarget->isThumb1Only())
435 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
437 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
438 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
439 !Subtarget->isThumb1Only()) {
440 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
441 if (!Subtarget->isFPOnlySP())
442 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
444 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
447 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
448 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
449 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
450 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
451 setTruncStoreAction((MVT::SimpleValueType)VT,
452 (MVT::SimpleValueType)InnerVT, Expand);
453 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
455 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
458 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
460 if (Subtarget->hasNEON()) {
461 addDRTypeForNEON(MVT::v2f32);
462 addDRTypeForNEON(MVT::v8i8);
463 addDRTypeForNEON(MVT::v4i16);
464 addDRTypeForNEON(MVT::v2i32);
465 addDRTypeForNEON(MVT::v1i64);
467 addQRTypeForNEON(MVT::v4f32);
468 addQRTypeForNEON(MVT::v2f64);
469 addQRTypeForNEON(MVT::v16i8);
470 addQRTypeForNEON(MVT::v8i16);
471 addQRTypeForNEON(MVT::v4i32);
472 addQRTypeForNEON(MVT::v2i64);
474 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
475 // neither Neon nor VFP support any arithmetic operations on it.
476 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
477 // supported for v4f32.
478 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
479 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
480 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
481 // FIXME: Code duplication: FDIV and FREM are expanded always, see
482 // ARMTargetLowering::addTypeForNEON method for details.
483 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
484 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
485 // FIXME: Create unittest.
486 // In another words, find a way when "copysign" appears in DAG with vector
488 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
489 // FIXME: Code duplication: SETCC has custom operation action, see
490 // ARMTargetLowering::addTypeForNEON method for details.
491 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
492 // FIXME: Create unittest for FNEG and for FABS.
493 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
494 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
495 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
496 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
497 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
498 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
499 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
500 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
501 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
502 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
503 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
504 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
505 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
506 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
508 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
509 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
510 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
512 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
513 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
514 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
515 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
516 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
517 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
518 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
520 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
521 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
523 // Neon does not support some operations on v1i64 and v2i64 types.
524 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
525 // Custom handling for some quad-vector types to detect VMULL.
526 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
527 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
528 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
529 // Custom handling for some vector types to avoid expensive expansions
530 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
531 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
532 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
533 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
534 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
535 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
536 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
537 // a destination type that is wider than the source, and nor does
538 // it have a FP_TO_[SU]INT instruction with a narrower destination than
540 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
541 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
542 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
543 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
545 setTargetDAGCombine(ISD::INTRINSIC_VOID);
546 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
547 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
548 setTargetDAGCombine(ISD::SHL);
549 setTargetDAGCombine(ISD::SRL);
550 setTargetDAGCombine(ISD::SRA);
551 setTargetDAGCombine(ISD::SIGN_EXTEND);
552 setTargetDAGCombine(ISD::ZERO_EXTEND);
553 setTargetDAGCombine(ISD::ANY_EXTEND);
554 setTargetDAGCombine(ISD::SELECT_CC);
555 setTargetDAGCombine(ISD::BUILD_VECTOR);
556 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
557 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
558 setTargetDAGCombine(ISD::STORE);
559 setTargetDAGCombine(ISD::FP_TO_SINT);
560 setTargetDAGCombine(ISD::FP_TO_UINT);
561 setTargetDAGCombine(ISD::FDIV);
563 // It is legal to extload from v4i8 to v4i16 or v4i32.
564 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
565 MVT::v4i16, MVT::v2i16,
567 for (unsigned i = 0; i < 6; ++i) {
568 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
569 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
570 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
574 computeRegisterProperties();
576 // ARM does not have f32 extending load.
577 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
579 // ARM does not have i1 sign extending load.
580 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
582 // ARM supports all 4 flavors of integer indexed load / store.
583 if (!Subtarget->isThumb1Only()) {
584 for (unsigned im = (unsigned)ISD::PRE_INC;
585 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
586 setIndexedLoadAction(im, MVT::i1, Legal);
587 setIndexedLoadAction(im, MVT::i8, Legal);
588 setIndexedLoadAction(im, MVT::i16, Legal);
589 setIndexedLoadAction(im, MVT::i32, Legal);
590 setIndexedStoreAction(im, MVT::i1, Legal);
591 setIndexedStoreAction(im, MVT::i8, Legal);
592 setIndexedStoreAction(im, MVT::i16, Legal);
593 setIndexedStoreAction(im, MVT::i32, Legal);
597 // i64 operation support.
598 setOperationAction(ISD::MUL, MVT::i64, Expand);
599 setOperationAction(ISD::MULHU, MVT::i32, Expand);
600 if (Subtarget->isThumb1Only()) {
601 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
602 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
604 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
605 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
606 setOperationAction(ISD::MULHS, MVT::i32, Expand);
608 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
609 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
610 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRL, MVT::i64, Custom);
612 setOperationAction(ISD::SRA, MVT::i64, Custom);
614 if (!Subtarget->isThumb1Only()) {
615 // FIXME: We should do this for Thumb1 as well.
616 setOperationAction(ISD::ADDC, MVT::i32, Custom);
617 setOperationAction(ISD::ADDE, MVT::i32, Custom);
618 setOperationAction(ISD::SUBC, MVT::i32, Custom);
619 setOperationAction(ISD::SUBE, MVT::i32, Custom);
622 // ARM does not have ROTL.
623 setOperationAction(ISD::ROTL, MVT::i32, Expand);
624 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
625 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
626 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
627 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
629 // These just redirect to CTTZ and CTLZ on ARM.
630 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
631 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
633 // Only ARMv6 has BSWAP.
634 if (!Subtarget->hasV6Ops())
635 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
637 // These are expanded into libcalls.
638 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
639 // v7M has a hardware divider
640 setOperationAction(ISD::SDIV, MVT::i32, Expand);
641 setOperationAction(ISD::UDIV, MVT::i32, Expand);
643 setOperationAction(ISD::SREM, MVT::i32, Expand);
644 setOperationAction(ISD::UREM, MVT::i32, Expand);
645 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
646 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
648 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
649 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
650 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
651 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
652 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
654 setOperationAction(ISD::TRAP, MVT::Other, Legal);
656 // Use the default implementation.
657 setOperationAction(ISD::VASTART, MVT::Other, Custom);
658 setOperationAction(ISD::VAARG, MVT::Other, Expand);
659 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
660 setOperationAction(ISD::VAEND, MVT::Other, Expand);
661 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
662 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
664 if (!Subtarget->isTargetDarwin()) {
665 // Non-Darwin platforms may return values in these registers via the
666 // personality function.
667 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
668 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
669 setExceptionPointerRegister(ARM::R0);
670 setExceptionSelectorRegister(ARM::R1);
673 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
674 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
675 // the default expansion.
676 // FIXME: This should be checking for v6k, not just v6.
677 if (Subtarget->hasDataBarrier() ||
678 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
679 // membarrier needs custom lowering; the rest are legal and handled
681 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
682 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
683 // Custom lowering for 64-bit ops
684 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
685 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
686 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
691 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
692 setInsertFencesForAtomic(true);
694 // Set them all for expansion, which will force libcalls.
695 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
696 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
697 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
698 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
699 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
709 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
710 // Unordered/Monotonic case.
711 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
712 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
713 // Since the libcalls include locking, fold in the fences
714 setShouldFoldAtomicFences(true);
717 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
719 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
720 if (!Subtarget->hasV6Ops()) {
721 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
722 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
726 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
727 !Subtarget->isThumb1Only()) {
728 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
729 // iff target supports vfp2.
730 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
731 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
734 // We want to custom lower some of our intrinsics.
735 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
736 if (Subtarget->isTargetDarwin()) {
737 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
738 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
739 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
742 setOperationAction(ISD::SETCC, MVT::i32, Expand);
743 setOperationAction(ISD::SETCC, MVT::f32, Expand);
744 setOperationAction(ISD::SETCC, MVT::f64, Expand);
745 setOperationAction(ISD::SELECT, MVT::i32, Custom);
746 setOperationAction(ISD::SELECT, MVT::f32, Custom);
747 setOperationAction(ISD::SELECT, MVT::f64, Custom);
748 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
749 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
752 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
753 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
754 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
755 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
756 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
758 // We don't support sin/cos/fmod/copysign/pow
759 setOperationAction(ISD::FSIN, MVT::f64, Expand);
760 setOperationAction(ISD::FSIN, MVT::f32, Expand);
761 setOperationAction(ISD::FCOS, MVT::f32, Expand);
762 setOperationAction(ISD::FCOS, MVT::f64, Expand);
763 setOperationAction(ISD::FREM, MVT::f64, Expand);
764 setOperationAction(ISD::FREM, MVT::f32, Expand);
765 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
766 !Subtarget->isThumb1Only()) {
767 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
768 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
770 setOperationAction(ISD::FPOW, MVT::f64, Expand);
771 setOperationAction(ISD::FPOW, MVT::f32, Expand);
773 if (!Subtarget->hasVFP4()) {
774 setOperationAction(ISD::FMA, MVT::f64, Expand);
775 setOperationAction(ISD::FMA, MVT::f32, Expand);
778 // Various VFP goodness
779 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
780 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
781 if (Subtarget->hasVFP2()) {
782 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
783 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
784 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
785 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
787 // Special handling for half-precision FP.
788 if (!Subtarget->hasFP16()) {
789 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
790 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
794 // We have target-specific dag combine patterns for the following nodes:
795 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
796 setTargetDAGCombine(ISD::ADD);
797 setTargetDAGCombine(ISD::SUB);
798 setTargetDAGCombine(ISD::MUL);
800 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
801 setTargetDAGCombine(ISD::AND);
802 setTargetDAGCombine(ISD::OR);
803 setTargetDAGCombine(ISD::XOR);
806 if (Subtarget->hasV6Ops())
807 setTargetDAGCombine(ISD::SRL);
809 setStackPointerRegisterToSaveRestore(ARM::SP);
811 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
812 !Subtarget->hasVFP2())
813 setSchedulingPreference(Sched::RegPressure);
815 setSchedulingPreference(Sched::Hybrid);
817 //// temporary - rewrite interface to use type
818 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
819 maxStoresPerMemset = 16;
820 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
822 // On ARM arguments smaller than 4 bytes are extended, so all arguments
823 // are at least 4 bytes aligned.
824 setMinStackArgumentAlignment(4);
826 benefitFromCodePlacementOpt = true;
828 // Prefer likely predicted branches to selects on out-of-order cores.
829 predictableSelectIsExpensive = Subtarget->isCortexA9();
831 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
834 // FIXME: It might make sense to define the representative register class as the
835 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
836 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
837 // SPR's representative would be DPR_VFP2. This should work well if register
838 // pressure tracking were modified such that a register use would increment the
839 // pressure of the register class's representative and all of it's super
840 // classes' representatives transitively. We have not implemented this because
841 // of the difficulty prior to coalescing of modeling operand register classes
842 // due to the common occurrence of cross class copies and subregister insertions
844 std::pair<const TargetRegisterClass*, uint8_t>
845 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
846 const TargetRegisterClass *RRC = 0;
848 switch (VT.getSimpleVT().SimpleTy) {
850 return TargetLowering::findRepresentativeClass(VT);
851 // Use DPR as representative register class for all floating point
852 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
853 // the cost is 1 for both f32 and f64.
854 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
855 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
856 RRC = &ARM::DPRRegClass;
857 // When NEON is used for SP, only half of the register file is available
858 // because operations that define both SP and DP results will be constrained
859 // to the VFP2 class (D0-D15). We currently model this constraint prior to
860 // coalescing by double-counting the SP regs. See the FIXME above.
861 if (Subtarget->useNEONForSinglePrecisionFP())
864 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
865 case MVT::v4f32: case MVT::v2f64:
866 RRC = &ARM::DPRRegClass;
870 RRC = &ARM::DPRRegClass;
874 RRC = &ARM::DPRRegClass;
878 return std::make_pair(RRC, Cost);
881 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
884 case ARMISD::Wrapper: return "ARMISD::Wrapper";
885 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
886 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
887 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
888 case ARMISD::CALL: return "ARMISD::CALL";
889 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
890 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
891 case ARMISD::tCALL: return "ARMISD::tCALL";
892 case ARMISD::BRCOND: return "ARMISD::BRCOND";
893 case ARMISD::BR_JT: return "ARMISD::BR_JT";
894 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
895 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
896 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
897 case ARMISD::CMP: return "ARMISD::CMP";
898 case ARMISD::CMPZ: return "ARMISD::CMPZ";
899 case ARMISD::CMPFP: return "ARMISD::CMPFP";
900 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
901 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
902 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
904 case ARMISD::CMOV: return "ARMISD::CMOV";
905 case ARMISD::CAND: return "ARMISD::CAND";
906 case ARMISD::COR: return "ARMISD::COR";
907 case ARMISD::CXOR: return "ARMISD::CXOR";
909 case ARMISD::RBIT: return "ARMISD::RBIT";
911 case ARMISD::FTOSI: return "ARMISD::FTOSI";
912 case ARMISD::FTOUI: return "ARMISD::FTOUI";
913 case ARMISD::SITOF: return "ARMISD::SITOF";
914 case ARMISD::UITOF: return "ARMISD::UITOF";
916 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
917 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
918 case ARMISD::RRX: return "ARMISD::RRX";
920 case ARMISD::ADDC: return "ARMISD::ADDC";
921 case ARMISD::ADDE: return "ARMISD::ADDE";
922 case ARMISD::SUBC: return "ARMISD::SUBC";
923 case ARMISD::SUBE: return "ARMISD::SUBE";
925 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
926 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
928 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
929 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
931 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
933 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
935 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
937 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
938 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
940 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
942 case ARMISD::VCEQ: return "ARMISD::VCEQ";
943 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
944 case ARMISD::VCGE: return "ARMISD::VCGE";
945 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
946 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
947 case ARMISD::VCGEU: return "ARMISD::VCGEU";
948 case ARMISD::VCGT: return "ARMISD::VCGT";
949 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
950 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
951 case ARMISD::VCGTU: return "ARMISD::VCGTU";
952 case ARMISD::VTST: return "ARMISD::VTST";
954 case ARMISD::VSHL: return "ARMISD::VSHL";
955 case ARMISD::VSHRs: return "ARMISD::VSHRs";
956 case ARMISD::VSHRu: return "ARMISD::VSHRu";
957 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
958 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
959 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
960 case ARMISD::VSHRN: return "ARMISD::VSHRN";
961 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
962 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
963 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
964 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
965 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
966 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
967 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
968 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
969 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
970 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
971 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
972 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
973 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
974 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
975 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
976 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
977 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
978 case ARMISD::VDUP: return "ARMISD::VDUP";
979 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
980 case ARMISD::VEXT: return "ARMISD::VEXT";
981 case ARMISD::VREV64: return "ARMISD::VREV64";
982 case ARMISD::VREV32: return "ARMISD::VREV32";
983 case ARMISD::VREV16: return "ARMISD::VREV16";
984 case ARMISD::VZIP: return "ARMISD::VZIP";
985 case ARMISD::VUZP: return "ARMISD::VUZP";
986 case ARMISD::VTRN: return "ARMISD::VTRN";
987 case ARMISD::VTBL1: return "ARMISD::VTBL1";
988 case ARMISD::VTBL2: return "ARMISD::VTBL2";
989 case ARMISD::VMULLs: return "ARMISD::VMULLs";
990 case ARMISD::VMULLu: return "ARMISD::VMULLu";
991 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
992 case ARMISD::FMAX: return "ARMISD::FMAX";
993 case ARMISD::FMIN: return "ARMISD::FMIN";
994 case ARMISD::BFI: return "ARMISD::BFI";
995 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
996 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
997 case ARMISD::VBSL: return "ARMISD::VBSL";
998 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
999 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1000 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1001 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1002 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1003 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1004 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1005 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1006 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1007 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1008 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1009 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1010 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1011 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1012 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1013 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1014 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1015 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1016 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1017 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1021 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1022 if (!VT.isVector()) return getPointerTy();
1023 return VT.changeVectorElementTypeToInteger();
1026 /// getRegClassFor - Return the register class that should be used for the
1027 /// specified value type.
1028 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1029 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1030 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1031 // load / store 4 to 8 consecutive D registers.
1032 if (Subtarget->hasNEON()) {
1033 if (VT == MVT::v4i64)
1034 return &ARM::QQPRRegClass;
1035 if (VT == MVT::v8i64)
1036 return &ARM::QQQQPRRegClass;
1038 return TargetLowering::getRegClassFor(VT);
1041 // Create a fast isel object.
1043 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1044 return ARM::createFastISel(funcInfo);
1047 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1048 /// be used for loads / stores from the global.
1049 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1050 return (Subtarget->isThumb1Only() ? 127 : 4095);
1053 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1054 unsigned NumVals = N->getNumValues();
1056 return Sched::RegPressure;
1058 for (unsigned i = 0; i != NumVals; ++i) {
1059 EVT VT = N->getValueType(i);
1060 if (VT == MVT::Glue || VT == MVT::Other)
1062 if (VT.isFloatingPoint() || VT.isVector())
1066 if (!N->isMachineOpcode())
1067 return Sched::RegPressure;
1069 // Load are scheduled for latency even if there instruction itinerary
1070 // is not available.
1071 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1072 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1074 if (MCID.getNumDefs() == 0)
1075 return Sched::RegPressure;
1076 if (!Itins->isEmpty() &&
1077 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1080 return Sched::RegPressure;
1083 //===----------------------------------------------------------------------===//
1085 //===----------------------------------------------------------------------===//
1087 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1088 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1090 default: llvm_unreachable("Unknown condition code!");
1091 case ISD::SETNE: return ARMCC::NE;
1092 case ISD::SETEQ: return ARMCC::EQ;
1093 case ISD::SETGT: return ARMCC::GT;
1094 case ISD::SETGE: return ARMCC::GE;
1095 case ISD::SETLT: return ARMCC::LT;
1096 case ISD::SETLE: return ARMCC::LE;
1097 case ISD::SETUGT: return ARMCC::HI;
1098 case ISD::SETUGE: return ARMCC::HS;
1099 case ISD::SETULT: return ARMCC::LO;
1100 case ISD::SETULE: return ARMCC::LS;
1104 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1105 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1106 ARMCC::CondCodes &CondCode2) {
1107 CondCode2 = ARMCC::AL;
1109 default: llvm_unreachable("Unknown FP condition!");
1111 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1113 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1115 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1116 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1117 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1118 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1119 case ISD::SETO: CondCode = ARMCC::VC; break;
1120 case ISD::SETUO: CondCode = ARMCC::VS; break;
1121 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1122 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1123 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1125 case ISD::SETULT: CondCode = ARMCC::LT; break;
1127 case ISD::SETULE: CondCode = ARMCC::LE; break;
1129 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1133 //===----------------------------------------------------------------------===//
1134 // Calling Convention Implementation
1135 //===----------------------------------------------------------------------===//
1137 #include "ARMGenCallingConv.inc"
1139 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1140 /// given CallingConvention value.
1141 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1143 bool isVarArg) const {
1146 llvm_unreachable("Unsupported calling convention");
1147 case CallingConv::Fast:
1148 if (Subtarget->hasVFP2() && !isVarArg) {
1149 if (!Subtarget->isAAPCS_ABI())
1150 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1151 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1152 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1155 case CallingConv::C: {
1156 // Use target triple & subtarget features to do actual dispatch.
1157 if (!Subtarget->isAAPCS_ABI())
1158 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1159 else if (Subtarget->hasVFP2() &&
1160 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1162 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_AAPCS_VFP:
1167 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1169 case CallingConv::ARM_AAPCS:
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1171 case CallingConv::ARM_APCS:
1172 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1176 /// LowerCallResult - Lower the result values of a call into the
1177 /// appropriate copies out of appropriate physical registers.
1179 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1180 CallingConv::ID CallConv, bool isVarArg,
1181 const SmallVectorImpl<ISD::InputArg> &Ins,
1182 DebugLoc dl, SelectionDAG &DAG,
1183 SmallVectorImpl<SDValue> &InVals) const {
1185 // Assign locations to each value returned by this call.
1186 SmallVector<CCValAssign, 16> RVLocs;
1187 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1188 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1189 CCInfo.AnalyzeCallResult(Ins,
1190 CCAssignFnForNode(CallConv, /* Return*/ true,
1193 // Copy all of the result registers out of their specified physreg.
1194 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1195 CCValAssign VA = RVLocs[i];
1198 if (VA.needsCustom()) {
1199 // Handle f64 or half of a v2f64.
1200 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1202 Chain = Lo.getValue(1);
1203 InFlag = Lo.getValue(2);
1204 VA = RVLocs[++i]; // skip ahead to next loc
1205 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1207 Chain = Hi.getValue(1);
1208 InFlag = Hi.getValue(2);
1209 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1211 if (VA.getLocVT() == MVT::v2f64) {
1212 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1213 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1214 DAG.getConstant(0, MVT::i32));
1216 VA = RVLocs[++i]; // skip ahead to next loc
1217 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1218 Chain = Lo.getValue(1);
1219 InFlag = Lo.getValue(2);
1220 VA = RVLocs[++i]; // skip ahead to next loc
1221 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1222 Chain = Hi.getValue(1);
1223 InFlag = Hi.getValue(2);
1224 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1225 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1226 DAG.getConstant(1, MVT::i32));
1229 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1231 Chain = Val.getValue(1);
1232 InFlag = Val.getValue(2);
1235 switch (VA.getLocInfo()) {
1236 default: llvm_unreachable("Unknown loc info!");
1237 case CCValAssign::Full: break;
1238 case CCValAssign::BCvt:
1239 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1243 InVals.push_back(Val);
1249 /// LowerMemOpCallTo - Store the argument to the stack.
1251 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1252 SDValue StackPtr, SDValue Arg,
1253 DebugLoc dl, SelectionDAG &DAG,
1254 const CCValAssign &VA,
1255 ISD::ArgFlagsTy Flags) const {
1256 unsigned LocMemOffset = VA.getLocMemOffset();
1257 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1258 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1259 return DAG.getStore(Chain, dl, Arg, PtrOff,
1260 MachinePointerInfo::getStack(LocMemOffset),
1264 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1265 SDValue Chain, SDValue &Arg,
1266 RegsToPassVector &RegsToPass,
1267 CCValAssign &VA, CCValAssign &NextVA,
1269 SmallVector<SDValue, 8> &MemOpChains,
1270 ISD::ArgFlagsTy Flags) const {
1272 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1273 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1274 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1276 if (NextVA.isRegLoc())
1277 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1279 assert(NextVA.isMemLoc());
1280 if (StackPtr.getNode() == 0)
1281 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1283 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1289 /// LowerCall - Lowering a call into a callseq_start <-
1290 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1293 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1294 SmallVectorImpl<SDValue> &InVals) const {
1295 SelectionDAG &DAG = CLI.DAG;
1296 DebugLoc &dl = CLI.DL;
1297 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1298 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1299 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1300 SDValue Chain = CLI.Chain;
1301 SDValue Callee = CLI.Callee;
1302 bool &isTailCall = CLI.IsTailCall;
1303 CallingConv::ID CallConv = CLI.CallConv;
1304 bool doesNotRet = CLI.DoesNotReturn;
1305 bool isVarArg = CLI.IsVarArg;
1307 MachineFunction &MF = DAG.getMachineFunction();
1308 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1309 bool IsSibCall = false;
1310 // Disable tail calls if they're not supported.
1311 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1314 // Check if it's really possible to do a tail call.
1315 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1316 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1317 Outs, OutVals, Ins, DAG);
1318 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1319 // detected sibcalls.
1326 // Analyze operands of the call, assigning locations to each operand.
1327 SmallVector<CCValAssign, 16> ArgLocs;
1328 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1329 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1330 CCInfo.AnalyzeCallOperands(Outs,
1331 CCAssignFnForNode(CallConv, /* Return*/ false,
1334 // Get a count of how many bytes are to be pushed on the stack.
1335 unsigned NumBytes = CCInfo.getNextStackOffset();
1337 // For tail calls, memory operands are available in our caller's stack.
1341 // Adjust the stack pointer for the new arguments...
1342 // These operations are automatically eliminated by the prolog/epilog pass
1344 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1346 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1348 RegsToPassVector RegsToPass;
1349 SmallVector<SDValue, 8> MemOpChains;
1351 // Walk the register/memloc assignments, inserting copies/loads. In the case
1352 // of tail call optimization, arguments are handled later.
1353 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1355 ++i, ++realArgIdx) {
1356 CCValAssign &VA = ArgLocs[i];
1357 SDValue Arg = OutVals[realArgIdx];
1358 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1359 bool isByVal = Flags.isByVal();
1361 // Promote the value if needed.
1362 switch (VA.getLocInfo()) {
1363 default: llvm_unreachable("Unknown loc info!");
1364 case CCValAssign::Full: break;
1365 case CCValAssign::SExt:
1366 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1368 case CCValAssign::ZExt:
1369 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1371 case CCValAssign::AExt:
1372 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1374 case CCValAssign::BCvt:
1375 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1379 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1380 if (VA.needsCustom()) {
1381 if (VA.getLocVT() == MVT::v2f64) {
1382 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1383 DAG.getConstant(0, MVT::i32));
1384 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1385 DAG.getConstant(1, MVT::i32));
1387 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1388 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1390 VA = ArgLocs[++i]; // skip ahead to next loc
1391 if (VA.isRegLoc()) {
1392 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1393 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1395 assert(VA.isMemLoc());
1397 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1398 dl, DAG, VA, Flags));
1401 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1402 StackPtr, MemOpChains, Flags);
1404 } else if (VA.isRegLoc()) {
1405 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1406 } else if (isByVal) {
1407 assert(VA.isMemLoc());
1408 unsigned offset = 0;
1410 // True if this byval aggregate will be split between registers
1412 if (CCInfo.isFirstByValRegValid()) {
1413 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1415 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1416 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1417 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1418 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1419 MachinePointerInfo(),
1420 false, false, false, 0);
1421 MemOpChains.push_back(Load.getValue(1));
1422 RegsToPass.push_back(std::make_pair(j, Load));
1424 offset = ARM::R4 - CCInfo.getFirstByValReg();
1425 CCInfo.clearFirstByValReg();
1428 if (Flags.getByValSize() - 4*offset > 0) {
1429 unsigned LocMemOffset = VA.getLocMemOffset();
1430 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1431 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1433 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1434 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1435 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1437 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1439 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1440 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1441 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1442 Ops, array_lengthof(Ops)));
1444 } else if (!IsSibCall) {
1445 assert(VA.isMemLoc());
1447 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1448 dl, DAG, VA, Flags));
1452 if (!MemOpChains.empty())
1453 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1454 &MemOpChains[0], MemOpChains.size());
1456 // Build a sequence of copy-to-reg nodes chained together with token chain
1457 // and flag operands which copy the outgoing args into the appropriate regs.
1459 // Tail call byval lowering might overwrite argument registers so in case of
1460 // tail call optimization the copies to registers are lowered later.
1462 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1463 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1464 RegsToPass[i].second, InFlag);
1465 InFlag = Chain.getValue(1);
1468 // For tail calls lower the arguments to the 'real' stack slot.
1470 // Force all the incoming stack arguments to be loaded from the stack
1471 // before any new outgoing arguments are stored to the stack, because the
1472 // outgoing stack slots may alias the incoming argument stack slots, and
1473 // the alias isn't otherwise explicit. This is slightly more conservative
1474 // than necessary, because it means that each store effectively depends
1475 // on every argument instead of just those arguments it would clobber.
1477 // Do not flag preceding copytoreg stuff together with the following stuff.
1479 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1480 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1481 RegsToPass[i].second, InFlag);
1482 InFlag = Chain.getValue(1);
1487 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1488 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1489 // node so that legalize doesn't hack it.
1490 bool isDirect = false;
1491 bool isARMFunc = false;
1492 bool isLocalARMFunc = false;
1493 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1495 if (EnableARMLongCalls) {
1496 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1497 && "long-calls with non-static relocation model!");
1498 // Handle a global address or an external symbol. If it's not one of
1499 // those, the target's already in a register, so we don't need to do
1501 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1502 const GlobalValue *GV = G->getGlobal();
1503 // Create a constant pool entry for the callee address
1504 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1505 ARMConstantPoolValue *CPV =
1506 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1508 // Get the address of the callee into a register
1509 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1510 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1511 Callee = DAG.getLoad(getPointerTy(), dl,
1512 DAG.getEntryNode(), CPAddr,
1513 MachinePointerInfo::getConstantPool(),
1514 false, false, false, 0);
1515 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1516 const char *Sym = S->getSymbol();
1518 // Create a constant pool entry for the callee address
1519 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1520 ARMConstantPoolValue *CPV =
1521 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1522 ARMPCLabelIndex, 0);
1523 // Get the address of the callee into a register
1524 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1525 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1526 Callee = DAG.getLoad(getPointerTy(), dl,
1527 DAG.getEntryNode(), CPAddr,
1528 MachinePointerInfo::getConstantPool(),
1529 false, false, false, 0);
1531 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1532 const GlobalValue *GV = G->getGlobal();
1534 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1535 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1536 getTargetMachine().getRelocationModel() != Reloc::Static;
1537 isARMFunc = !Subtarget->isThumb() || isStub;
1538 // ARM call to a local ARM function is predicable.
1539 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1540 // tBX takes a register source operand.
1541 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1542 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1543 ARMConstantPoolValue *CPV =
1544 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1545 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1546 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1547 Callee = DAG.getLoad(getPointerTy(), dl,
1548 DAG.getEntryNode(), CPAddr,
1549 MachinePointerInfo::getConstantPool(),
1550 false, false, false, 0);
1551 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1552 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1553 getPointerTy(), Callee, PICLabel);
1555 // On ELF targets for PIC code, direct calls should go through the PLT
1556 unsigned OpFlags = 0;
1557 if (Subtarget->isTargetELF() &&
1558 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1559 OpFlags = ARMII::MO_PLT;
1560 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1562 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1564 bool isStub = Subtarget->isTargetDarwin() &&
1565 getTargetMachine().getRelocationModel() != Reloc::Static;
1566 isARMFunc = !Subtarget->isThumb() || isStub;
1567 // tBX takes a register source operand.
1568 const char *Sym = S->getSymbol();
1569 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1570 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1571 ARMConstantPoolValue *CPV =
1572 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1573 ARMPCLabelIndex, 4);
1574 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1575 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1576 Callee = DAG.getLoad(getPointerTy(), dl,
1577 DAG.getEntryNode(), CPAddr,
1578 MachinePointerInfo::getConstantPool(),
1579 false, false, false, 0);
1580 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1581 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1582 getPointerTy(), Callee, PICLabel);
1584 unsigned OpFlags = 0;
1585 // On ELF targets for PIC code, direct calls should go through the PLT
1586 if (Subtarget->isTargetELF() &&
1587 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1588 OpFlags = ARMII::MO_PLT;
1589 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1593 // FIXME: handle tail calls differently.
1595 if (Subtarget->isThumb()) {
1596 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1597 CallOpc = ARMISD::CALL_NOLINK;
1598 else if (doesNotRet && isDirect && !isARMFunc &&
1599 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1600 // "mov lr, pc; b _foo" to avoid confusing the RSP
1601 CallOpc = ARMISD::CALL_NOLINK;
1603 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1605 if (!isDirect && !Subtarget->hasV5TOps()) {
1606 CallOpc = ARMISD::CALL_NOLINK;
1607 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1608 // "mov lr, pc; b _foo" to avoid confusing the RSP
1609 CallOpc = ARMISD::CALL_NOLINK;
1611 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1614 std::vector<SDValue> Ops;
1615 Ops.push_back(Chain);
1616 Ops.push_back(Callee);
1618 // Add argument registers to the end of the list so that they are known live
1620 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1621 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1622 RegsToPass[i].second.getValueType()));
1624 // Add a register mask operand representing the call-preserved registers.
1625 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1626 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1627 assert(Mask && "Missing call preserved mask for calling convention");
1628 Ops.push_back(DAG.getRegisterMask(Mask));
1630 if (InFlag.getNode())
1631 Ops.push_back(InFlag);
1633 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1635 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1637 // Returns a chain and a flag for retval copy to use.
1638 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1639 InFlag = Chain.getValue(1);
1641 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1642 DAG.getIntPtrConstant(0, true), InFlag);
1644 InFlag = Chain.getValue(1);
1646 // Handle result values, copying them out of physregs into vregs that we
1648 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1652 /// HandleByVal - Every parameter *after* a byval parameter is passed
1653 /// on the stack. Remember the next parameter register to allocate,
1654 /// and then confiscate the rest of the parameter registers to insure
1657 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1658 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1659 assert((State->getCallOrPrologue() == Prologue ||
1660 State->getCallOrPrologue() == Call) &&
1661 "unhandled ParmContext");
1662 if ((!State->isFirstByValRegValid()) &&
1663 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1664 State->setFirstByValReg(reg);
1665 // At a call site, a byval parameter that is split between
1666 // registers and memory needs its size truncated here. In a
1667 // function prologue, such byval parameters are reassembled in
1668 // memory, and are not truncated.
1669 if (State->getCallOrPrologue() == Call) {
1670 unsigned excess = 4 * (ARM::R4 - reg);
1671 assert(size >= excess && "expected larger existing stack allocation");
1675 // Confiscate any remaining parameter registers to preclude their
1676 // assignment to subsequent parameters.
1677 while (State->AllocateReg(GPRArgRegs, 4))
1681 /// MatchingStackOffset - Return true if the given stack call argument is
1682 /// already available in the same position (relatively) of the caller's
1683 /// incoming argument stack.
1685 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1686 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1687 const TargetInstrInfo *TII) {
1688 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1690 if (Arg.getOpcode() == ISD::CopyFromReg) {
1691 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1692 if (!TargetRegisterInfo::isVirtualRegister(VR))
1694 MachineInstr *Def = MRI->getVRegDef(VR);
1697 if (!Flags.isByVal()) {
1698 if (!TII->isLoadFromStackSlot(Def, FI))
1703 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1704 if (Flags.isByVal())
1705 // ByVal argument is passed in as a pointer but it's now being
1706 // dereferenced. e.g.
1707 // define @foo(%struct.X* %A) {
1708 // tail call @bar(%struct.X* byval %A)
1711 SDValue Ptr = Ld->getBasePtr();
1712 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1715 FI = FINode->getIndex();
1719 assert(FI != INT_MAX);
1720 if (!MFI->isFixedObjectIndex(FI))
1722 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1725 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1726 /// for tail call optimization. Targets which want to do tail call
1727 /// optimization should implement this function.
1729 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1730 CallingConv::ID CalleeCC,
1732 bool isCalleeStructRet,
1733 bool isCallerStructRet,
1734 const SmallVectorImpl<ISD::OutputArg> &Outs,
1735 const SmallVectorImpl<SDValue> &OutVals,
1736 const SmallVectorImpl<ISD::InputArg> &Ins,
1737 SelectionDAG& DAG) const {
1738 const Function *CallerF = DAG.getMachineFunction().getFunction();
1739 CallingConv::ID CallerCC = CallerF->getCallingConv();
1740 bool CCMatch = CallerCC == CalleeCC;
1742 // Look for obvious safe cases to perform tail call optimization that do not
1743 // require ABI changes. This is what gcc calls sibcall.
1745 // Do not sibcall optimize vararg calls unless the call site is not passing
1747 if (isVarArg && !Outs.empty())
1750 // Also avoid sibcall optimization if either caller or callee uses struct
1751 // return semantics.
1752 if (isCalleeStructRet || isCallerStructRet)
1755 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1756 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1757 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1758 // support in the assembler and linker to be used. This would need to be
1759 // fixed to fully support tail calls in Thumb1.
1761 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1762 // LR. This means if we need to reload LR, it takes an extra instructions,
1763 // which outweighs the value of the tail call; but here we don't know yet
1764 // whether LR is going to be used. Probably the right approach is to
1765 // generate the tail call here and turn it back into CALL/RET in
1766 // emitEpilogue if LR is used.
1768 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1769 // but we need to make sure there are enough registers; the only valid
1770 // registers are the 4 used for parameters. We don't currently do this
1772 if (Subtarget->isThumb1Only())
1775 // If the calling conventions do not match, then we'd better make sure the
1776 // results are returned in the same way as what the caller expects.
1778 SmallVector<CCValAssign, 16> RVLocs1;
1779 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1780 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1781 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1783 SmallVector<CCValAssign, 16> RVLocs2;
1784 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1785 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1786 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1788 if (RVLocs1.size() != RVLocs2.size())
1790 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1791 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1793 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1795 if (RVLocs1[i].isRegLoc()) {
1796 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1799 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1805 // If the callee takes no arguments then go on to check the results of the
1807 if (!Outs.empty()) {
1808 // Check if stack adjustment is needed. For now, do not do this if any
1809 // argument is passed on the stack.
1810 SmallVector<CCValAssign, 16> ArgLocs;
1811 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1812 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1813 CCInfo.AnalyzeCallOperands(Outs,
1814 CCAssignFnForNode(CalleeCC, false, isVarArg));
1815 if (CCInfo.getNextStackOffset()) {
1816 MachineFunction &MF = DAG.getMachineFunction();
1818 // Check if the arguments are already laid out in the right way as
1819 // the caller's fixed stack objects.
1820 MachineFrameInfo *MFI = MF.getFrameInfo();
1821 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1822 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1823 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1825 ++i, ++realArgIdx) {
1826 CCValAssign &VA = ArgLocs[i];
1827 EVT RegVT = VA.getLocVT();
1828 SDValue Arg = OutVals[realArgIdx];
1829 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1830 if (VA.getLocInfo() == CCValAssign::Indirect)
1832 if (VA.needsCustom()) {
1833 // f64 and vector types are split into multiple registers or
1834 // register/stack-slot combinations. The types will not match
1835 // the registers; give up on memory f64 refs until we figure
1836 // out what to do about this.
1839 if (!ArgLocs[++i].isRegLoc())
1841 if (RegVT == MVT::v2f64) {
1842 if (!ArgLocs[++i].isRegLoc())
1844 if (!ArgLocs[++i].isRegLoc())
1847 } else if (!VA.isRegLoc()) {
1848 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1860 ARMTargetLowering::LowerReturn(SDValue Chain,
1861 CallingConv::ID CallConv, bool isVarArg,
1862 const SmallVectorImpl<ISD::OutputArg> &Outs,
1863 const SmallVectorImpl<SDValue> &OutVals,
1864 DebugLoc dl, SelectionDAG &DAG) const {
1866 // CCValAssign - represent the assignment of the return value to a location.
1867 SmallVector<CCValAssign, 16> RVLocs;
1869 // CCState - Info about the registers and stack slots.
1870 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1871 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1873 // Analyze outgoing return values.
1874 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1877 // If this is the first return lowered for this function, add
1878 // the regs to the liveout set for the function.
1879 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1880 for (unsigned i = 0; i != RVLocs.size(); ++i)
1881 if (RVLocs[i].isRegLoc())
1882 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1887 // Copy the result values into the output registers.
1888 for (unsigned i = 0, realRVLocIdx = 0;
1890 ++i, ++realRVLocIdx) {
1891 CCValAssign &VA = RVLocs[i];
1892 assert(VA.isRegLoc() && "Can only return in registers!");
1894 SDValue Arg = OutVals[realRVLocIdx];
1896 switch (VA.getLocInfo()) {
1897 default: llvm_unreachable("Unknown loc info!");
1898 case CCValAssign::Full: break;
1899 case CCValAssign::BCvt:
1900 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1904 if (VA.needsCustom()) {
1905 if (VA.getLocVT() == MVT::v2f64) {
1906 // Extract the first half and return it in two registers.
1907 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1908 DAG.getConstant(0, MVT::i32));
1909 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1910 DAG.getVTList(MVT::i32, MVT::i32), Half);
1912 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1913 Flag = Chain.getValue(1);
1914 VA = RVLocs[++i]; // skip ahead to next loc
1915 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1916 HalfGPRs.getValue(1), Flag);
1917 Flag = Chain.getValue(1);
1918 VA = RVLocs[++i]; // skip ahead to next loc
1920 // Extract the 2nd half and fall through to handle it as an f64 value.
1921 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1922 DAG.getConstant(1, MVT::i32));
1924 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1926 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1927 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1928 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1929 Flag = Chain.getValue(1);
1930 VA = RVLocs[++i]; // skip ahead to next loc
1931 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1934 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1936 // Guarantee that all emitted copies are
1937 // stuck together, avoiding something bad.
1938 Flag = Chain.getValue(1);
1943 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1945 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1950 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1951 if (N->getNumValues() != 1)
1953 if (!N->hasNUsesOfValue(1, 0))
1956 SDValue TCChain = Chain;
1957 SDNode *Copy = *N->use_begin();
1958 if (Copy->getOpcode() == ISD::CopyToReg) {
1959 // If the copy has a glue operand, we conservatively assume it isn't safe to
1960 // perform a tail call.
1961 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1963 TCChain = Copy->getOperand(0);
1964 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1965 SDNode *VMov = Copy;
1966 // f64 returned in a pair of GPRs.
1967 SmallPtrSet<SDNode*, 2> Copies;
1968 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1970 if (UI->getOpcode() != ISD::CopyToReg)
1974 if (Copies.size() > 2)
1977 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1979 SDValue UseChain = UI->getOperand(0);
1980 if (Copies.count(UseChain.getNode()))
1987 } else if (Copy->getOpcode() == ISD::BITCAST) {
1988 // f32 returned in a single GPR.
1989 if (!Copy->hasOneUse())
1991 Copy = *Copy->use_begin();
1992 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
1994 Chain = Copy->getOperand(0);
1999 bool HasRet = false;
2000 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2002 if (UI->getOpcode() != ARMISD::RET_FLAG)
2014 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2015 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2018 if (!CI->isTailCall())
2021 return !Subtarget->isThumb1Only();
2024 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2025 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2026 // one of the above mentioned nodes. It has to be wrapped because otherwise
2027 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2028 // be used to form addressing mode. These wrapped nodes will be selected
2030 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2031 EVT PtrVT = Op.getValueType();
2032 // FIXME there is no actual debug info here
2033 DebugLoc dl = Op.getDebugLoc();
2034 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2036 if (CP->isMachineConstantPoolEntry())
2037 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2038 CP->getAlignment());
2040 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2041 CP->getAlignment());
2042 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2045 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2046 return MachineJumpTableInfo::EK_Inline;
2049 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2050 SelectionDAG &DAG) const {
2051 MachineFunction &MF = DAG.getMachineFunction();
2052 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2053 unsigned ARMPCLabelIndex = 0;
2054 DebugLoc DL = Op.getDebugLoc();
2055 EVT PtrVT = getPointerTy();
2056 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2057 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2059 if (RelocM == Reloc::Static) {
2060 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2062 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2063 ARMPCLabelIndex = AFI->createPICLabelUId();
2064 ARMConstantPoolValue *CPV =
2065 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2066 ARMCP::CPBlockAddress, PCAdj);
2067 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2069 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2070 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2071 MachinePointerInfo::getConstantPool(),
2072 false, false, false, 0);
2073 if (RelocM == Reloc::Static)
2075 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2076 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2079 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2081 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2082 SelectionDAG &DAG) const {
2083 DebugLoc dl = GA->getDebugLoc();
2084 EVT PtrVT = getPointerTy();
2085 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2086 MachineFunction &MF = DAG.getMachineFunction();
2087 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2088 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2089 ARMConstantPoolValue *CPV =
2090 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2091 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2092 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2093 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2094 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2095 MachinePointerInfo::getConstantPool(),
2096 false, false, false, 0);
2097 SDValue Chain = Argument.getValue(1);
2099 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2100 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2102 // call __tls_get_addr.
2105 Entry.Node = Argument;
2106 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2107 Args.push_back(Entry);
2108 // FIXME: is there useful debug info available here?
2109 TargetLowering::CallLoweringInfo CLI(Chain,
2110 (Type *) Type::getInt32Ty(*DAG.getContext()),
2111 false, false, false, false,
2112 0, CallingConv::C, /*isTailCall=*/false,
2113 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2114 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2115 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2116 return CallResult.first;
2119 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2120 // "local exec" model.
2122 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2124 TLSModel::Model model) const {
2125 const GlobalValue *GV = GA->getGlobal();
2126 DebugLoc dl = GA->getDebugLoc();
2128 SDValue Chain = DAG.getEntryNode();
2129 EVT PtrVT = getPointerTy();
2130 // Get the Thread Pointer
2131 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2133 if (model == TLSModel::InitialExec) {
2134 MachineFunction &MF = DAG.getMachineFunction();
2135 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2136 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2137 // Initial exec model.
2138 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2139 ARMConstantPoolValue *CPV =
2140 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2141 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2143 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2144 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2145 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2146 MachinePointerInfo::getConstantPool(),
2147 false, false, false, 0);
2148 Chain = Offset.getValue(1);
2150 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2151 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2153 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2154 MachinePointerInfo::getConstantPool(),
2155 false, false, false, 0);
2158 assert(model == TLSModel::LocalExec);
2159 ARMConstantPoolValue *CPV =
2160 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2161 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2162 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2163 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2164 MachinePointerInfo::getConstantPool(),
2165 false, false, false, 0);
2168 // The address of the thread local variable is the add of the thread
2169 // pointer with the offset of the variable.
2170 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2174 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2175 // TODO: implement the "local dynamic" model
2176 assert(Subtarget->isTargetELF() &&
2177 "TLS not implemented for non-ELF targets");
2178 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2180 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2183 case TLSModel::GeneralDynamic:
2184 case TLSModel::LocalDynamic:
2185 return LowerToTLSGeneralDynamicModel(GA, DAG);
2186 case TLSModel::InitialExec:
2187 case TLSModel::LocalExec:
2188 return LowerToTLSExecModels(GA, DAG, model);
2190 llvm_unreachable("bogus TLS model");
2193 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2194 SelectionDAG &DAG) const {
2195 EVT PtrVT = getPointerTy();
2196 DebugLoc dl = Op.getDebugLoc();
2197 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2198 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2199 if (RelocM == Reloc::PIC_) {
2200 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2201 ARMConstantPoolValue *CPV =
2202 ARMConstantPoolConstant::Create(GV,
2203 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2204 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2205 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2206 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2208 MachinePointerInfo::getConstantPool(),
2209 false, false, false, 0);
2210 SDValue Chain = Result.getValue(1);
2211 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2212 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2214 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2215 MachinePointerInfo::getGOT(),
2216 false, false, false, 0);
2220 // If we have T2 ops, we can materialize the address directly via movt/movw
2221 // pair. This is always cheaper.
2222 if (Subtarget->useMovt()) {
2224 // FIXME: Once remat is capable of dealing with instructions with register
2225 // operands, expand this into two nodes.
2226 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2227 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2229 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2230 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2231 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2232 MachinePointerInfo::getConstantPool(),
2233 false, false, false, 0);
2237 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2238 SelectionDAG &DAG) const {
2239 EVT PtrVT = getPointerTy();
2240 DebugLoc dl = Op.getDebugLoc();
2241 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2242 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2243 MachineFunction &MF = DAG.getMachineFunction();
2244 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2246 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2247 // update ARMFastISel::ARMMaterializeGV.
2248 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2250 // FIXME: Once remat is capable of dealing with instructions with register
2251 // operands, expand this into two nodes.
2252 if (RelocM == Reloc::Static)
2253 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2254 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2256 unsigned Wrapper = (RelocM == Reloc::PIC_)
2257 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2258 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2259 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2260 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2261 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2262 MachinePointerInfo::getGOT(),
2263 false, false, false, 0);
2267 unsigned ARMPCLabelIndex = 0;
2269 if (RelocM == Reloc::Static) {
2270 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2272 ARMPCLabelIndex = AFI->createPICLabelUId();
2273 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2274 ARMConstantPoolValue *CPV =
2275 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2277 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2279 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2281 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2282 MachinePointerInfo::getConstantPool(),
2283 false, false, false, 0);
2284 SDValue Chain = Result.getValue(1);
2286 if (RelocM == Reloc::PIC_) {
2287 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2288 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2291 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2292 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2293 false, false, false, 0);
2298 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2299 SelectionDAG &DAG) const {
2300 assert(Subtarget->isTargetELF() &&
2301 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2302 MachineFunction &MF = DAG.getMachineFunction();
2303 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2304 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2305 EVT PtrVT = getPointerTy();
2306 DebugLoc dl = Op.getDebugLoc();
2307 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2308 ARMConstantPoolValue *CPV =
2309 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2310 ARMPCLabelIndex, PCAdj);
2311 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2312 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2313 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2314 MachinePointerInfo::getConstantPool(),
2315 false, false, false, 0);
2316 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2317 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2321 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2322 DebugLoc dl = Op.getDebugLoc();
2323 SDValue Val = DAG.getConstant(0, MVT::i32);
2324 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2325 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2326 Op.getOperand(1), Val);
2330 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2331 DebugLoc dl = Op.getDebugLoc();
2332 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2333 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2337 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2338 const ARMSubtarget *Subtarget) const {
2339 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2340 DebugLoc dl = Op.getDebugLoc();
2342 default: return SDValue(); // Don't custom lower most intrinsics.
2343 case Intrinsic::arm_thread_pointer: {
2344 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2345 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2347 case Intrinsic::eh_sjlj_lsda: {
2348 MachineFunction &MF = DAG.getMachineFunction();
2349 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2350 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2351 EVT PtrVT = getPointerTy();
2352 DebugLoc dl = Op.getDebugLoc();
2353 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2355 unsigned PCAdj = (RelocM != Reloc::PIC_)
2356 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2357 ARMConstantPoolValue *CPV =
2358 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2359 ARMCP::CPLSDA, PCAdj);
2360 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2361 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2363 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2364 MachinePointerInfo::getConstantPool(),
2365 false, false, false, 0);
2367 if (RelocM == Reloc::PIC_) {
2368 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2369 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2373 case Intrinsic::arm_neon_vmulls:
2374 case Intrinsic::arm_neon_vmullu: {
2375 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2376 ? ARMISD::VMULLs : ARMISD::VMULLu;
2377 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2378 Op.getOperand(1), Op.getOperand(2));
2383 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2384 const ARMSubtarget *Subtarget) {
2385 DebugLoc dl = Op.getDebugLoc();
2386 if (!Subtarget->hasDataBarrier()) {
2387 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2388 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2390 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2391 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2392 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2393 DAG.getConstant(0, MVT::i32));
2396 SDValue Op5 = Op.getOperand(5);
2397 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2398 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2399 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2400 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2402 ARM_MB::MemBOpt DMBOpt;
2403 if (isDeviceBarrier)
2404 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2406 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2407 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2408 DAG.getConstant(DMBOpt, MVT::i32));
2412 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2413 const ARMSubtarget *Subtarget) {
2414 // FIXME: handle "fence singlethread" more efficiently.
2415 DebugLoc dl = Op.getDebugLoc();
2416 if (!Subtarget->hasDataBarrier()) {
2417 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2418 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2420 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2421 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2422 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2423 DAG.getConstant(0, MVT::i32));
2426 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2427 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2430 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2431 const ARMSubtarget *Subtarget) {
2432 // ARM pre v5TE and Thumb1 does not have preload instructions.
2433 if (!(Subtarget->isThumb2() ||
2434 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2435 // Just preserve the chain.
2436 return Op.getOperand(0);
2438 DebugLoc dl = Op.getDebugLoc();
2439 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2441 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2442 // ARMv7 with MP extension has PLDW.
2443 return Op.getOperand(0);
2445 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2446 if (Subtarget->isThumb()) {
2448 isRead = ~isRead & 1;
2449 isData = ~isData & 1;
2452 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2453 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2454 DAG.getConstant(isData, MVT::i32));
2457 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2458 MachineFunction &MF = DAG.getMachineFunction();
2459 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2461 // vastart just stores the address of the VarArgsFrameIndex slot into the
2462 // memory location argument.
2463 DebugLoc dl = Op.getDebugLoc();
2464 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2465 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2466 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2467 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2468 MachinePointerInfo(SV), false, false, 0);
2472 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2473 SDValue &Root, SelectionDAG &DAG,
2474 DebugLoc dl) const {
2475 MachineFunction &MF = DAG.getMachineFunction();
2476 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2478 const TargetRegisterClass *RC;
2479 if (AFI->isThumb1OnlyFunction())
2480 RC = &ARM::tGPRRegClass;
2482 RC = &ARM::GPRRegClass;
2484 // Transform the arguments stored in physical registers into virtual ones.
2485 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2486 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2489 if (NextVA.isMemLoc()) {
2490 MachineFrameInfo *MFI = MF.getFrameInfo();
2491 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2493 // Create load node to retrieve arguments from the stack.
2494 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2495 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2496 MachinePointerInfo::getFixedStack(FI),
2497 false, false, false, 0);
2499 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2500 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2503 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2507 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2508 unsigned &VARegSize, unsigned &VARegSaveSize)
2511 if (CCInfo.isFirstByValRegValid())
2512 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2514 unsigned int firstUnalloced;
2515 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2516 sizeof(GPRArgRegs) /
2517 sizeof(GPRArgRegs[0]));
2518 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2521 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2522 VARegSize = NumGPRs * 4;
2523 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2526 // The remaining GPRs hold either the beginning of variable-argument
2527 // data, or the beginning of an aggregate passed by value (usuall
2528 // byval). Either way, we allocate stack slots adjacent to the data
2529 // provided by our caller, and store the unallocated registers there.
2530 // If this is a variadic function, the va_list pointer will begin with
2531 // these values; otherwise, this reassembles a (byval) structure that
2532 // was split between registers and memory.
2534 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2535 DebugLoc dl, SDValue &Chain,
2536 unsigned ArgOffset) const {
2537 MachineFunction &MF = DAG.getMachineFunction();
2538 MachineFrameInfo *MFI = MF.getFrameInfo();
2539 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2540 unsigned firstRegToSaveIndex;
2541 if (CCInfo.isFirstByValRegValid())
2542 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2544 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2545 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2548 unsigned VARegSize, VARegSaveSize;
2549 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2550 if (VARegSaveSize) {
2551 // If this function is vararg, store any remaining integer argument regs
2552 // to their spots on the stack so that they may be loaded by deferencing
2553 // the result of va_next.
2554 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2555 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2556 ArgOffset + VARegSaveSize
2559 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2562 SmallVector<SDValue, 4> MemOps;
2563 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2564 const TargetRegisterClass *RC;
2565 if (AFI->isThumb1OnlyFunction())
2566 RC = &ARM::tGPRRegClass;
2568 RC = &ARM::GPRRegClass;
2570 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2571 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2573 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2574 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2576 MemOps.push_back(Store);
2577 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2578 DAG.getConstant(4, getPointerTy()));
2580 if (!MemOps.empty())
2581 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2582 &MemOps[0], MemOps.size());
2584 // This will point to the next argument passed via stack.
2585 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2589 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2590 CallingConv::ID CallConv, bool isVarArg,
2591 const SmallVectorImpl<ISD::InputArg>
2593 DebugLoc dl, SelectionDAG &DAG,
2594 SmallVectorImpl<SDValue> &InVals)
2596 MachineFunction &MF = DAG.getMachineFunction();
2597 MachineFrameInfo *MFI = MF.getFrameInfo();
2599 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2601 // Assign locations to all of the incoming arguments.
2602 SmallVector<CCValAssign, 16> ArgLocs;
2603 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2604 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2605 CCInfo.AnalyzeFormalArguments(Ins,
2606 CCAssignFnForNode(CallConv, /* Return*/ false,
2609 SmallVector<SDValue, 16> ArgValues;
2610 int lastInsIndex = -1;
2613 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2614 CCValAssign &VA = ArgLocs[i];
2616 // Arguments stored in registers.
2617 if (VA.isRegLoc()) {
2618 EVT RegVT = VA.getLocVT();
2620 if (VA.needsCustom()) {
2621 // f64 and vector types are split up into multiple registers or
2622 // combinations of registers and stack slots.
2623 if (VA.getLocVT() == MVT::v2f64) {
2624 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2626 VA = ArgLocs[++i]; // skip ahead to next loc
2628 if (VA.isMemLoc()) {
2629 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2630 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2631 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2632 MachinePointerInfo::getFixedStack(FI),
2633 false, false, false, 0);
2635 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2638 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2639 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2640 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2641 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2642 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2644 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2647 const TargetRegisterClass *RC;
2649 if (RegVT == MVT::f32)
2650 RC = &ARM::SPRRegClass;
2651 else if (RegVT == MVT::f64)
2652 RC = &ARM::DPRRegClass;
2653 else if (RegVT == MVT::v2f64)
2654 RC = &ARM::QPRRegClass;
2655 else if (RegVT == MVT::i32)
2656 RC = AFI->isThumb1OnlyFunction() ?
2657 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2658 (const TargetRegisterClass*)&ARM::GPRRegClass;
2660 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2662 // Transform the arguments in physical registers into virtual ones.
2663 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2664 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2667 // If this is an 8 or 16-bit value, it is really passed promoted
2668 // to 32 bits. Insert an assert[sz]ext to capture this, then
2669 // truncate to the right size.
2670 switch (VA.getLocInfo()) {
2671 default: llvm_unreachable("Unknown loc info!");
2672 case CCValAssign::Full: break;
2673 case CCValAssign::BCvt:
2674 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2676 case CCValAssign::SExt:
2677 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2678 DAG.getValueType(VA.getValVT()));
2679 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2681 case CCValAssign::ZExt:
2682 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2683 DAG.getValueType(VA.getValVT()));
2684 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2688 InVals.push_back(ArgValue);
2690 } else { // VA.isRegLoc()
2693 assert(VA.isMemLoc());
2694 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2696 int index = ArgLocs[i].getValNo();
2698 // Some Ins[] entries become multiple ArgLoc[] entries.
2699 // Process them only once.
2700 if (index != lastInsIndex)
2702 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2703 // FIXME: For now, all byval parameter objects are marked mutable.
2704 // This can be changed with more analysis.
2705 // In case of tail call optimization mark all arguments mutable.
2706 // Since they could be overwritten by lowering of arguments in case of
2708 if (Flags.isByVal()) {
2709 unsigned VARegSize, VARegSaveSize;
2710 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2711 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2712 unsigned Bytes = Flags.getByValSize() - VARegSize;
2713 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2714 int FI = MFI->CreateFixedObject(Bytes,
2715 VA.getLocMemOffset(), false);
2716 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2718 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2719 VA.getLocMemOffset(), true);
2721 // Create load nodes to retrieve arguments from the stack.
2722 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2723 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2724 MachinePointerInfo::getFixedStack(FI),
2725 false, false, false, 0));
2727 lastInsIndex = index;
2734 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2739 /// isFloatingPointZero - Return true if this is +0.0.
2740 static bool isFloatingPointZero(SDValue Op) {
2741 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2742 return CFP->getValueAPF().isPosZero();
2743 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2744 // Maybe this has already been legalized into the constant pool?
2745 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2746 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2747 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2748 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2749 return CFP->getValueAPF().isPosZero();
2755 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2756 /// the given operands.
2758 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2759 SDValue &ARMcc, SelectionDAG &DAG,
2760 DebugLoc dl) const {
2761 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2762 unsigned C = RHSC->getZExtValue();
2763 if (!isLegalICmpImmediate(C)) {
2764 // Constant does not fit, try adjusting it by one?
2769 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2770 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2771 RHS = DAG.getConstant(C-1, MVT::i32);
2776 if (C != 0 && isLegalICmpImmediate(C-1)) {
2777 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2778 RHS = DAG.getConstant(C-1, MVT::i32);
2783 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2784 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2785 RHS = DAG.getConstant(C+1, MVT::i32);
2790 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2791 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2792 RHS = DAG.getConstant(C+1, MVT::i32);
2799 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2800 ARMISD::NodeType CompareType;
2803 CompareType = ARMISD::CMP;
2808 CompareType = ARMISD::CMPZ;
2811 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2812 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2815 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2817 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2818 DebugLoc dl) const {
2820 if (!isFloatingPointZero(RHS))
2821 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2823 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2824 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2827 /// duplicateCmp - Glue values can have only one use, so this function
2828 /// duplicates a comparison node.
2830 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2831 unsigned Opc = Cmp.getOpcode();
2832 DebugLoc DL = Cmp.getDebugLoc();
2833 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2834 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2836 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2837 Cmp = Cmp.getOperand(0);
2838 Opc = Cmp.getOpcode();
2839 if (Opc == ARMISD::CMPFP)
2840 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2842 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2843 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2845 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2848 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2849 SDValue Cond = Op.getOperand(0);
2850 SDValue SelectTrue = Op.getOperand(1);
2851 SDValue SelectFalse = Op.getOperand(2);
2852 DebugLoc dl = Op.getDebugLoc();
2856 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2857 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2859 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2860 const ConstantSDNode *CMOVTrue =
2861 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2862 const ConstantSDNode *CMOVFalse =
2863 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2865 if (CMOVTrue && CMOVFalse) {
2866 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2867 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2871 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2873 False = SelectFalse;
2874 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2879 if (True.getNode() && False.getNode()) {
2880 EVT VT = Op.getValueType();
2881 SDValue ARMcc = Cond.getOperand(2);
2882 SDValue CCR = Cond.getOperand(3);
2883 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2884 assert(True.getValueType() == VT);
2885 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2890 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2891 // undefined bits before doing a full-word comparison with zero.
2892 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2893 DAG.getConstant(1, Cond.getValueType()));
2895 return DAG.getSelectCC(dl, Cond,
2896 DAG.getConstant(0, Cond.getValueType()),
2897 SelectTrue, SelectFalse, ISD::SETNE);
2900 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2901 EVT VT = Op.getValueType();
2902 SDValue LHS = Op.getOperand(0);
2903 SDValue RHS = Op.getOperand(1);
2904 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2905 SDValue TrueVal = Op.getOperand(2);
2906 SDValue FalseVal = Op.getOperand(3);
2907 DebugLoc dl = Op.getDebugLoc();
2909 if (LHS.getValueType() == MVT::i32) {
2911 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2912 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2913 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2916 ARMCC::CondCodes CondCode, CondCode2;
2917 FPCCToARMCC(CC, CondCode, CondCode2);
2919 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2920 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2921 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2922 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2924 if (CondCode2 != ARMCC::AL) {
2925 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2926 // FIXME: Needs another CMP because flag can have but one use.
2927 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2928 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2929 Result, TrueVal, ARMcc2, CCR, Cmp2);
2934 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2935 /// to morph to an integer compare sequence.
2936 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2937 const ARMSubtarget *Subtarget) {
2938 SDNode *N = Op.getNode();
2939 if (!N->hasOneUse())
2940 // Otherwise it requires moving the value from fp to integer registers.
2942 if (!N->getNumValues())
2944 EVT VT = Op.getValueType();
2945 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2946 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2947 // vmrs are very slow, e.g. cortex-a8.
2950 if (isFloatingPointZero(Op)) {
2954 return ISD::isNormalLoad(N);
2957 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2958 if (isFloatingPointZero(Op))
2959 return DAG.getConstant(0, MVT::i32);
2961 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2962 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2963 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2964 Ld->isVolatile(), Ld->isNonTemporal(),
2965 Ld->isInvariant(), Ld->getAlignment());
2967 llvm_unreachable("Unknown VFP cmp argument!");
2970 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2971 SDValue &RetVal1, SDValue &RetVal2) {
2972 if (isFloatingPointZero(Op)) {
2973 RetVal1 = DAG.getConstant(0, MVT::i32);
2974 RetVal2 = DAG.getConstant(0, MVT::i32);
2978 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2979 SDValue Ptr = Ld->getBasePtr();
2980 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2981 Ld->getChain(), Ptr,
2982 Ld->getPointerInfo(),
2983 Ld->isVolatile(), Ld->isNonTemporal(),
2984 Ld->isInvariant(), Ld->getAlignment());
2986 EVT PtrType = Ptr.getValueType();
2987 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2988 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2989 PtrType, Ptr, DAG.getConstant(4, PtrType));
2990 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2991 Ld->getChain(), NewPtr,
2992 Ld->getPointerInfo().getWithOffset(4),
2993 Ld->isVolatile(), Ld->isNonTemporal(),
2994 Ld->isInvariant(), NewAlign);
2998 llvm_unreachable("Unknown VFP cmp argument!");
3001 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3002 /// f32 and even f64 comparisons to integer ones.
3004 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3005 SDValue Chain = Op.getOperand(0);
3006 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3007 SDValue LHS = Op.getOperand(2);
3008 SDValue RHS = Op.getOperand(3);
3009 SDValue Dest = Op.getOperand(4);
3010 DebugLoc dl = Op.getDebugLoc();
3012 bool LHSSeenZero = false;
3013 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3014 bool RHSSeenZero = false;
3015 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3016 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3017 // If unsafe fp math optimization is enabled and there are no other uses of
3018 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3019 // to an integer comparison.
3020 if (CC == ISD::SETOEQ)
3022 else if (CC == ISD::SETUNE)
3025 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3027 if (LHS.getValueType() == MVT::f32) {
3028 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3029 bitcastf32Toi32(LHS, DAG), Mask);
3030 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3031 bitcastf32Toi32(RHS, DAG), Mask);
3032 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3033 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3034 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3035 Chain, Dest, ARMcc, CCR, Cmp);
3040 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3041 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3042 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3043 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3044 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3045 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3046 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3047 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3048 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3054 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3055 SDValue Chain = Op.getOperand(0);
3056 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3057 SDValue LHS = Op.getOperand(2);
3058 SDValue RHS = Op.getOperand(3);
3059 SDValue Dest = Op.getOperand(4);
3060 DebugLoc dl = Op.getDebugLoc();
3062 if (LHS.getValueType() == MVT::i32) {
3064 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3065 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3066 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3067 Chain, Dest, ARMcc, CCR, Cmp);
3070 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3072 if (getTargetMachine().Options.UnsafeFPMath &&
3073 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3074 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3075 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3076 if (Result.getNode())
3080 ARMCC::CondCodes CondCode, CondCode2;
3081 FPCCToARMCC(CC, CondCode, CondCode2);
3083 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3084 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3085 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3086 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3087 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3088 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3089 if (CondCode2 != ARMCC::AL) {
3090 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3091 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3092 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3097 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3098 SDValue Chain = Op.getOperand(0);
3099 SDValue Table = Op.getOperand(1);
3100 SDValue Index = Op.getOperand(2);
3101 DebugLoc dl = Op.getDebugLoc();
3103 EVT PTy = getPointerTy();
3104 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3105 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3106 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3107 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3108 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3109 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3110 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3111 if (Subtarget->isThumb2()) {
3112 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3113 // which does another jump to the destination. This also makes it easier
3114 // to translate it to TBB / TBH later.
3115 // FIXME: This might not work if the function is extremely large.
3116 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3117 Addr, Op.getOperand(2), JTI, UId);
3119 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3120 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3121 MachinePointerInfo::getJumpTable(),
3122 false, false, false, 0);
3123 Chain = Addr.getValue(1);
3124 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3125 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3127 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3128 MachinePointerInfo::getJumpTable(),
3129 false, false, false, 0);
3130 Chain = Addr.getValue(1);
3131 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3135 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3136 EVT VT = Op.getValueType();
3137 DebugLoc dl = Op.getDebugLoc();
3139 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3140 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3142 return DAG.UnrollVectorOp(Op.getNode());
3145 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3146 "Invalid type for custom lowering!");
3147 if (VT != MVT::v4i16)
3148 return DAG.UnrollVectorOp(Op.getNode());
3150 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3151 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3154 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3155 EVT VT = Op.getValueType();
3157 return LowerVectorFP_TO_INT(Op, DAG);
3159 DebugLoc dl = Op.getDebugLoc();
3162 switch (Op.getOpcode()) {
3163 default: llvm_unreachable("Invalid opcode!");
3164 case ISD::FP_TO_SINT:
3165 Opc = ARMISD::FTOSI;
3167 case ISD::FP_TO_UINT:
3168 Opc = ARMISD::FTOUI;
3171 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3172 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3175 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3176 EVT VT = Op.getValueType();
3177 DebugLoc dl = Op.getDebugLoc();
3179 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3180 if (VT.getVectorElementType() == MVT::f32)
3182 return DAG.UnrollVectorOp(Op.getNode());
3185 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3186 "Invalid type for custom lowering!");
3187 if (VT != MVT::v4f32)
3188 return DAG.UnrollVectorOp(Op.getNode());
3192 switch (Op.getOpcode()) {
3193 default: llvm_unreachable("Invalid opcode!");
3194 case ISD::SINT_TO_FP:
3195 CastOpc = ISD::SIGN_EXTEND;
3196 Opc = ISD::SINT_TO_FP;
3198 case ISD::UINT_TO_FP:
3199 CastOpc = ISD::ZERO_EXTEND;
3200 Opc = ISD::UINT_TO_FP;
3204 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3205 return DAG.getNode(Opc, dl, VT, Op);
3208 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3209 EVT VT = Op.getValueType();
3211 return LowerVectorINT_TO_FP(Op, DAG);
3213 DebugLoc dl = Op.getDebugLoc();
3216 switch (Op.getOpcode()) {
3217 default: llvm_unreachable("Invalid opcode!");
3218 case ISD::SINT_TO_FP:
3219 Opc = ARMISD::SITOF;
3221 case ISD::UINT_TO_FP:
3222 Opc = ARMISD::UITOF;
3226 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3227 return DAG.getNode(Opc, dl, VT, Op);
3230 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3231 // Implement fcopysign with a fabs and a conditional fneg.
3232 SDValue Tmp0 = Op.getOperand(0);
3233 SDValue Tmp1 = Op.getOperand(1);
3234 DebugLoc dl = Op.getDebugLoc();
3235 EVT VT = Op.getValueType();
3236 EVT SrcVT = Tmp1.getValueType();
3237 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3238 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3239 bool UseNEON = !InGPR && Subtarget->hasNEON();
3242 // Use VBSL to copy the sign bit.
3243 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3244 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3245 DAG.getTargetConstant(EncodedVal, MVT::i32));
3246 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3248 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3249 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3250 DAG.getConstant(32, MVT::i32));
3251 else /*if (VT == MVT::f32)*/
3252 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3253 if (SrcVT == MVT::f32) {
3254 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3256 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3257 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3258 DAG.getConstant(32, MVT::i32));
3259 } else if (VT == MVT::f32)
3260 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3261 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3262 DAG.getConstant(32, MVT::i32));
3263 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3264 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3266 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3268 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3269 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3270 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3272 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3273 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3274 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3275 if (VT == MVT::f32) {
3276 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3277 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3278 DAG.getConstant(0, MVT::i32));
3280 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3286 // Bitcast operand 1 to i32.
3287 if (SrcVT == MVT::f64)
3288 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3289 &Tmp1, 1).getValue(1);
3290 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3292 // Or in the signbit with integer operations.
3293 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3294 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3295 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3296 if (VT == MVT::f32) {
3297 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3298 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3299 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3300 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3303 // f64: Or the high part with signbit and then combine two parts.
3304 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3306 SDValue Lo = Tmp0.getValue(0);
3307 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3308 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3309 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3312 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3313 MachineFunction &MF = DAG.getMachineFunction();
3314 MachineFrameInfo *MFI = MF.getFrameInfo();
3315 MFI->setReturnAddressIsTaken(true);
3317 EVT VT = Op.getValueType();
3318 DebugLoc dl = Op.getDebugLoc();
3319 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3321 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3322 SDValue Offset = DAG.getConstant(4, MVT::i32);
3323 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3324 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3325 MachinePointerInfo(), false, false, false, 0);
3328 // Return LR, which contains the return address. Mark it an implicit live-in.
3329 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3330 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3333 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3334 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3335 MFI->setFrameAddressIsTaken(true);
3337 EVT VT = Op.getValueType();
3338 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3339 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3340 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3341 ? ARM::R7 : ARM::R11;
3342 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3344 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3345 MachinePointerInfo(),
3346 false, false, false, 0);
3350 /// ExpandBITCAST - If the target supports VFP, this function is called to
3351 /// expand a bit convert where either the source or destination type is i64 to
3352 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3353 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3354 /// vectors), since the legalizer won't know what to do with that.
3355 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3356 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3357 DebugLoc dl = N->getDebugLoc();
3358 SDValue Op = N->getOperand(0);
3360 // This function is only supposed to be called for i64 types, either as the
3361 // source or destination of the bit convert.
3362 EVT SrcVT = Op.getValueType();
3363 EVT DstVT = N->getValueType(0);
3364 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3365 "ExpandBITCAST called for non-i64 type");
3367 // Turn i64->f64 into VMOVDRR.
3368 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3369 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3370 DAG.getConstant(0, MVT::i32));
3371 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3372 DAG.getConstant(1, MVT::i32));
3373 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3374 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3377 // Turn f64->i64 into VMOVRRD.
3378 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3379 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3380 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3381 // Merge the pieces into a single i64 value.
3382 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3388 /// getZeroVector - Returns a vector of specified type with all zero elements.
3389 /// Zero vectors are used to represent vector negation and in those cases
3390 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3391 /// not support i64 elements, so sometimes the zero vectors will need to be
3392 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3394 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3395 assert(VT.isVector() && "Expected a vector type");
3396 // The canonical modified immediate encoding of a zero vector is....0!
3397 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3398 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3399 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3400 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3403 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3404 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3405 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3406 SelectionDAG &DAG) const {
3407 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3408 EVT VT = Op.getValueType();
3409 unsigned VTBits = VT.getSizeInBits();
3410 DebugLoc dl = Op.getDebugLoc();
3411 SDValue ShOpLo = Op.getOperand(0);
3412 SDValue ShOpHi = Op.getOperand(1);
3413 SDValue ShAmt = Op.getOperand(2);
3415 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3417 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3419 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3420 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3421 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3422 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3423 DAG.getConstant(VTBits, MVT::i32));
3424 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3425 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3426 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3428 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3429 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3431 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3432 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3435 SDValue Ops[2] = { Lo, Hi };
3436 return DAG.getMergeValues(Ops, 2, dl);
3439 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3440 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3441 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3442 SelectionDAG &DAG) const {
3443 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3444 EVT VT = Op.getValueType();
3445 unsigned VTBits = VT.getSizeInBits();
3446 DebugLoc dl = Op.getDebugLoc();
3447 SDValue ShOpLo = Op.getOperand(0);
3448 SDValue ShOpHi = Op.getOperand(1);
3449 SDValue ShAmt = Op.getOperand(2);
3452 assert(Op.getOpcode() == ISD::SHL_PARTS);
3453 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3454 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3455 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3456 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3457 DAG.getConstant(VTBits, MVT::i32));
3458 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3459 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3461 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3462 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3463 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3465 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3466 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3469 SDValue Ops[2] = { Lo, Hi };
3470 return DAG.getMergeValues(Ops, 2, dl);
3473 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3474 SelectionDAG &DAG) const {
3475 // The rounding mode is in bits 23:22 of the FPSCR.
3476 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3477 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3478 // so that the shift + and get folded into a bitfield extract.
3479 DebugLoc dl = Op.getDebugLoc();
3480 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3481 DAG.getConstant(Intrinsic::arm_get_fpscr,
3483 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3484 DAG.getConstant(1U << 22, MVT::i32));
3485 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3486 DAG.getConstant(22, MVT::i32));
3487 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3488 DAG.getConstant(3, MVT::i32));
3491 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3492 const ARMSubtarget *ST) {
3493 EVT VT = N->getValueType(0);
3494 DebugLoc dl = N->getDebugLoc();
3496 if (!ST->hasV6T2Ops())
3499 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3500 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3503 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3504 const ARMSubtarget *ST) {
3505 EVT VT = N->getValueType(0);
3506 DebugLoc dl = N->getDebugLoc();
3511 // Lower vector shifts on NEON to use VSHL.
3512 assert(ST->hasNEON() && "unexpected vector shift");
3514 // Left shifts translate directly to the vshiftu intrinsic.
3515 if (N->getOpcode() == ISD::SHL)
3516 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3517 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3518 N->getOperand(0), N->getOperand(1));
3520 assert((N->getOpcode() == ISD::SRA ||
3521 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3523 // NEON uses the same intrinsics for both left and right shifts. For
3524 // right shifts, the shift amounts are negative, so negate the vector of
3526 EVT ShiftVT = N->getOperand(1).getValueType();
3527 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3528 getZeroVector(ShiftVT, DAG, dl),
3530 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3531 Intrinsic::arm_neon_vshifts :
3532 Intrinsic::arm_neon_vshiftu);
3533 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3534 DAG.getConstant(vshiftInt, MVT::i32),
3535 N->getOperand(0), NegatedCount);
3538 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3539 const ARMSubtarget *ST) {
3540 EVT VT = N->getValueType(0);
3541 DebugLoc dl = N->getDebugLoc();
3543 // We can get here for a node like i32 = ISD::SHL i32, i64
3547 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3548 "Unknown shift to lower!");
3550 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3551 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3552 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3555 // If we are in thumb mode, we don't have RRX.
3556 if (ST->isThumb1Only()) return SDValue();
3558 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3559 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3560 DAG.getConstant(0, MVT::i32));
3561 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3562 DAG.getConstant(1, MVT::i32));
3564 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3565 // captures the result into a carry flag.
3566 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3567 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3569 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3570 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3572 // Merge the pieces into a single i64 value.
3573 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3576 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3577 SDValue TmpOp0, TmpOp1;
3578 bool Invert = false;
3582 SDValue Op0 = Op.getOperand(0);
3583 SDValue Op1 = Op.getOperand(1);
3584 SDValue CC = Op.getOperand(2);
3585 EVT VT = Op.getValueType();
3586 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3587 DebugLoc dl = Op.getDebugLoc();
3589 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3590 switch (SetCCOpcode) {
3591 default: llvm_unreachable("Illegal FP comparison");
3593 case ISD::SETNE: Invert = true; // Fallthrough
3595 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3597 case ISD::SETLT: Swap = true; // Fallthrough
3599 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3601 case ISD::SETLE: Swap = true; // Fallthrough
3603 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3604 case ISD::SETUGE: Swap = true; // Fallthrough
3605 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3606 case ISD::SETUGT: Swap = true; // Fallthrough
3607 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3608 case ISD::SETUEQ: Invert = true; // Fallthrough
3610 // Expand this to (OLT | OGT).
3614 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3615 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3617 case ISD::SETUO: Invert = true; // Fallthrough
3619 // Expand this to (OLT | OGE).
3623 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3624 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3628 // Integer comparisons.
3629 switch (SetCCOpcode) {
3630 default: llvm_unreachable("Illegal integer comparison");
3631 case ISD::SETNE: Invert = true;
3632 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3633 case ISD::SETLT: Swap = true;
3634 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3635 case ISD::SETLE: Swap = true;
3636 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3637 case ISD::SETULT: Swap = true;
3638 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3639 case ISD::SETULE: Swap = true;
3640 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3643 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3644 if (Opc == ARMISD::VCEQ) {
3647 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3649 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3652 // Ignore bitconvert.
3653 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3654 AndOp = AndOp.getOperand(0);
3656 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3658 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3659 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3666 std::swap(Op0, Op1);
3668 // If one of the operands is a constant vector zero, attempt to fold the
3669 // comparison to a specialized compare-against-zero form.
3671 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3673 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3674 if (Opc == ARMISD::VCGE)
3675 Opc = ARMISD::VCLEZ;
3676 else if (Opc == ARMISD::VCGT)
3677 Opc = ARMISD::VCLTZ;
3682 if (SingleOp.getNode()) {
3685 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3687 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3689 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3691 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3693 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3695 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3698 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3702 Result = DAG.getNOT(dl, Result, VT);
3707 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3708 /// valid vector constant for a NEON instruction with a "modified immediate"
3709 /// operand (e.g., VMOV). If so, return the encoded value.
3710 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3711 unsigned SplatBitSize, SelectionDAG &DAG,
3712 EVT &VT, bool is128Bits, NEONModImmType type) {
3713 unsigned OpCmode, Imm;
3715 // SplatBitSize is set to the smallest size that splats the vector, so a
3716 // zero vector will always have SplatBitSize == 8. However, NEON modified
3717 // immediate instructions others than VMOV do not support the 8-bit encoding
3718 // of a zero vector, and the default encoding of zero is supposed to be the
3723 switch (SplatBitSize) {
3725 if (type != VMOVModImm)
3727 // Any 1-byte value is OK. Op=0, Cmode=1110.
3728 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3731 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3735 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3736 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3737 if ((SplatBits & ~0xff) == 0) {
3738 // Value = 0x00nn: Op=x, Cmode=100x.
3743 if ((SplatBits & ~0xff00) == 0) {
3744 // Value = 0xnn00: Op=x, Cmode=101x.
3746 Imm = SplatBits >> 8;
3752 // NEON's 32-bit VMOV supports splat values where:
3753 // * only one byte is nonzero, or
3754 // * the least significant byte is 0xff and the second byte is nonzero, or
3755 // * the least significant 2 bytes are 0xff and the third is nonzero.
3756 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3757 if ((SplatBits & ~0xff) == 0) {
3758 // Value = 0x000000nn: Op=x, Cmode=000x.
3763 if ((SplatBits & ~0xff00) == 0) {
3764 // Value = 0x0000nn00: Op=x, Cmode=001x.
3766 Imm = SplatBits >> 8;
3769 if ((SplatBits & ~0xff0000) == 0) {
3770 // Value = 0x00nn0000: Op=x, Cmode=010x.
3772 Imm = SplatBits >> 16;
3775 if ((SplatBits & ~0xff000000) == 0) {
3776 // Value = 0xnn000000: Op=x, Cmode=011x.
3778 Imm = SplatBits >> 24;
3782 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3783 if (type == OtherModImm) return SDValue();
3785 if ((SplatBits & ~0xffff) == 0 &&
3786 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3787 // Value = 0x0000nnff: Op=x, Cmode=1100.
3789 Imm = SplatBits >> 8;
3794 if ((SplatBits & ~0xffffff) == 0 &&
3795 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3796 // Value = 0x00nnffff: Op=x, Cmode=1101.
3798 Imm = SplatBits >> 16;
3799 SplatBits |= 0xffff;
3803 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3804 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3805 // VMOV.I32. A (very) minor optimization would be to replicate the value
3806 // and fall through here to test for a valid 64-bit splat. But, then the
3807 // caller would also need to check and handle the change in size.
3811 if (type != VMOVModImm)
3813 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3814 uint64_t BitMask = 0xff;
3816 unsigned ImmMask = 1;
3818 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3819 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3822 } else if ((SplatBits & BitMask) != 0) {
3828 // Op=1, Cmode=1110.
3831 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3836 llvm_unreachable("unexpected size for isNEONModifiedImm");
3839 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3840 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3843 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3844 const ARMSubtarget *ST) const {
3845 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3848 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3849 assert(Op.getValueType() == MVT::f32 &&
3850 "ConstantFP custom lowering should only occur for f32.");
3852 // Try splatting with a VMOV.f32...
3853 APFloat FPVal = CFP->getValueAPF();
3854 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3856 DebugLoc DL = Op.getDebugLoc();
3857 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3858 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3860 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3861 DAG.getConstant(0, MVT::i32));
3864 // If that fails, try a VMOV.i32
3866 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3867 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3869 if (NewVal != SDValue()) {
3870 DebugLoc DL = Op.getDebugLoc();
3871 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3873 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3875 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3876 DAG.getConstant(0, MVT::i32));
3879 // Finally, try a VMVN.i32
3880 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3882 if (NewVal != SDValue()) {
3883 DebugLoc DL = Op.getDebugLoc();
3884 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3885 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3887 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3888 DAG.getConstant(0, MVT::i32));
3895 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3896 bool &ReverseVEXT, unsigned &Imm) {
3897 unsigned NumElts = VT.getVectorNumElements();
3898 ReverseVEXT = false;
3900 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3906 // If this is a VEXT shuffle, the immediate value is the index of the first
3907 // element. The other shuffle indices must be the successive elements after
3909 unsigned ExpectedElt = Imm;
3910 for (unsigned i = 1; i < NumElts; ++i) {
3911 // Increment the expected index. If it wraps around, it may still be
3912 // a VEXT but the source vectors must be swapped.
3914 if (ExpectedElt == NumElts * 2) {
3919 if (M[i] < 0) continue; // ignore UNDEF indices
3920 if (ExpectedElt != static_cast<unsigned>(M[i]))
3924 // Adjust the index value if the source operands will be swapped.
3931 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3932 /// instruction with the specified blocksize. (The order of the elements
3933 /// within each block of the vector is reversed.)
3934 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3935 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3936 "Only possible block sizes for VREV are: 16, 32, 64");
3938 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3942 unsigned NumElts = VT.getVectorNumElements();
3943 unsigned BlockElts = M[0] + 1;
3944 // If the first shuffle index is UNDEF, be optimistic.
3946 BlockElts = BlockSize / EltSz;
3948 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3951 for (unsigned i = 0; i < NumElts; ++i) {
3952 if (M[i] < 0) continue; // ignore UNDEF indices
3953 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3960 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3961 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3962 // range, then 0 is placed into the resulting vector. So pretty much any mask
3963 // of 8 elements can work here.
3964 return VT == MVT::v8i8 && M.size() == 8;
3967 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3968 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3972 unsigned NumElts = VT.getVectorNumElements();
3973 WhichResult = (M[0] == 0 ? 0 : 1);
3974 for (unsigned i = 0; i < NumElts; i += 2) {
3975 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3976 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3982 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3983 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3984 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3985 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3986 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3990 unsigned NumElts = VT.getVectorNumElements();
3991 WhichResult = (M[0] == 0 ? 0 : 1);
3992 for (unsigned i = 0; i < NumElts; i += 2) {
3993 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3994 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4000 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4001 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4005 unsigned NumElts = VT.getVectorNumElements();
4006 WhichResult = (M[0] == 0 ? 0 : 1);
4007 for (unsigned i = 0; i != NumElts; ++i) {
4008 if (M[i] < 0) continue; // ignore UNDEF indices
4009 if ((unsigned) M[i] != 2 * i + WhichResult)
4013 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4014 if (VT.is64BitVector() && EltSz == 32)
4020 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4021 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4022 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4023 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4024 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4028 unsigned Half = VT.getVectorNumElements() / 2;
4029 WhichResult = (M[0] == 0 ? 0 : 1);
4030 for (unsigned j = 0; j != 2; ++j) {
4031 unsigned Idx = WhichResult;
4032 for (unsigned i = 0; i != Half; ++i) {
4033 int MIdx = M[i + j * Half];
4034 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4040 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4041 if (VT.is64BitVector() && EltSz == 32)
4047 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4048 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4052 unsigned NumElts = VT.getVectorNumElements();
4053 WhichResult = (M[0] == 0 ? 0 : 1);
4054 unsigned Idx = WhichResult * NumElts / 2;
4055 for (unsigned i = 0; i != NumElts; i += 2) {
4056 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4057 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4062 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4063 if (VT.is64BitVector() && EltSz == 32)
4069 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4070 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4071 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4072 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4073 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4077 unsigned NumElts = VT.getVectorNumElements();
4078 WhichResult = (M[0] == 0 ? 0 : 1);
4079 unsigned Idx = WhichResult * NumElts / 2;
4080 for (unsigned i = 0; i != NumElts; i += 2) {
4081 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4082 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4087 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4088 if (VT.is64BitVector() && EltSz == 32)
4094 // If N is an integer constant that can be moved into a register in one
4095 // instruction, return an SDValue of such a constant (will become a MOV
4096 // instruction). Otherwise return null.
4097 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4098 const ARMSubtarget *ST, DebugLoc dl) {
4100 if (!isa<ConstantSDNode>(N))
4102 Val = cast<ConstantSDNode>(N)->getZExtValue();
4104 if (ST->isThumb1Only()) {
4105 if (Val <= 255 || ~Val <= 255)
4106 return DAG.getConstant(Val, MVT::i32);
4108 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4109 return DAG.getConstant(Val, MVT::i32);
4114 // If this is a case we can't handle, return null and let the default
4115 // expansion code take care of it.
4116 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4117 const ARMSubtarget *ST) const {
4118 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4119 DebugLoc dl = Op.getDebugLoc();
4120 EVT VT = Op.getValueType();
4122 APInt SplatBits, SplatUndef;
4123 unsigned SplatBitSize;
4125 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4126 if (SplatBitSize <= 64) {
4127 // Check if an immediate VMOV works.
4129 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4130 SplatUndef.getZExtValue(), SplatBitSize,
4131 DAG, VmovVT, VT.is128BitVector(),
4133 if (Val.getNode()) {
4134 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4135 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4138 // Try an immediate VMVN.
4139 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4140 Val = isNEONModifiedImm(NegatedImm,
4141 SplatUndef.getZExtValue(), SplatBitSize,
4142 DAG, VmovVT, VT.is128BitVector(),
4144 if (Val.getNode()) {
4145 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4146 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4149 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4150 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4151 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4153 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4154 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4160 // Scan through the operands to see if only one value is used.
4161 unsigned NumElts = VT.getVectorNumElements();
4162 bool isOnlyLowElement = true;
4163 bool usesOnlyOneValue = true;
4164 bool isConstant = true;
4166 for (unsigned i = 0; i < NumElts; ++i) {
4167 SDValue V = Op.getOperand(i);
4168 if (V.getOpcode() == ISD::UNDEF)
4171 isOnlyLowElement = false;
4172 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4175 if (!Value.getNode())
4177 else if (V != Value)
4178 usesOnlyOneValue = false;
4181 if (!Value.getNode())
4182 return DAG.getUNDEF(VT);
4184 if (isOnlyLowElement)
4185 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4187 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4189 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4190 // i32 and try again.
4191 if (usesOnlyOneValue && EltSize <= 32) {
4193 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4194 if (VT.getVectorElementType().isFloatingPoint()) {
4195 SmallVector<SDValue, 8> Ops;
4196 for (unsigned i = 0; i < NumElts; ++i)
4197 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4199 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4200 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4201 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4203 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4205 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4207 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4210 // If all elements are constants and the case above didn't get hit, fall back
4211 // to the default expansion, which will generate a load from the constant
4216 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4218 SDValue shuffle = ReconstructShuffle(Op, DAG);
4219 if (shuffle != SDValue())
4223 // Vectors with 32- or 64-bit elements can be built by directly assigning
4224 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4225 // will be legalized.
4226 if (EltSize >= 32) {
4227 // Do the expansion with floating-point types, since that is what the VFP
4228 // registers are defined to use, and since i64 is not legal.
4229 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4230 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4231 SmallVector<SDValue, 8> Ops;
4232 for (unsigned i = 0; i < NumElts; ++i)
4233 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4234 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4235 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4241 // Gather data to see if the operation can be modelled as a
4242 // shuffle in combination with VEXTs.
4243 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4244 SelectionDAG &DAG) const {
4245 DebugLoc dl = Op.getDebugLoc();
4246 EVT VT = Op.getValueType();
4247 unsigned NumElts = VT.getVectorNumElements();
4249 SmallVector<SDValue, 2> SourceVecs;
4250 SmallVector<unsigned, 2> MinElts;
4251 SmallVector<unsigned, 2> MaxElts;
4253 for (unsigned i = 0; i < NumElts; ++i) {
4254 SDValue V = Op.getOperand(i);
4255 if (V.getOpcode() == ISD::UNDEF)
4257 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4258 // A shuffle can only come from building a vector from various
4259 // elements of other vectors.
4261 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4262 VT.getVectorElementType()) {
4263 // This code doesn't know how to handle shuffles where the vector
4264 // element types do not match (this happens because type legalization
4265 // promotes the return type of EXTRACT_VECTOR_ELT).
4266 // FIXME: It might be appropriate to extend this code to handle
4267 // mismatched types.
4271 // Record this extraction against the appropriate vector if possible...
4272 SDValue SourceVec = V.getOperand(0);
4273 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4274 bool FoundSource = false;
4275 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4276 if (SourceVecs[j] == SourceVec) {
4277 if (MinElts[j] > EltNo)
4279 if (MaxElts[j] < EltNo)
4286 // Or record a new source if not...
4288 SourceVecs.push_back(SourceVec);
4289 MinElts.push_back(EltNo);
4290 MaxElts.push_back(EltNo);
4294 // Currently only do something sane when at most two source vectors
4296 if (SourceVecs.size() > 2)
4299 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4300 int VEXTOffsets[2] = {0, 0};
4302 // This loop extracts the usage patterns of the source vectors
4303 // and prepares appropriate SDValues for a shuffle if possible.
4304 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4305 if (SourceVecs[i].getValueType() == VT) {
4306 // No VEXT necessary
4307 ShuffleSrcs[i] = SourceVecs[i];
4310 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4311 // It probably isn't worth padding out a smaller vector just to
4312 // break it down again in a shuffle.
4316 // Since only 64-bit and 128-bit vectors are legal on ARM and
4317 // we've eliminated the other cases...
4318 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4319 "unexpected vector sizes in ReconstructShuffle");
4321 if (MaxElts[i] - MinElts[i] >= NumElts) {
4322 // Span too large for a VEXT to cope
4326 if (MinElts[i] >= NumElts) {
4327 // The extraction can just take the second half
4328 VEXTOffsets[i] = NumElts;
4329 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4331 DAG.getIntPtrConstant(NumElts));
4332 } else if (MaxElts[i] < NumElts) {
4333 // The extraction can just take the first half
4335 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4337 DAG.getIntPtrConstant(0));
4339 // An actual VEXT is needed
4340 VEXTOffsets[i] = MinElts[i];
4341 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4343 DAG.getIntPtrConstant(0));
4344 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4346 DAG.getIntPtrConstant(NumElts));
4347 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4348 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4352 SmallVector<int, 8> Mask;
4354 for (unsigned i = 0; i < NumElts; ++i) {
4355 SDValue Entry = Op.getOperand(i);
4356 if (Entry.getOpcode() == ISD::UNDEF) {
4361 SDValue ExtractVec = Entry.getOperand(0);
4362 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4363 .getOperand(1))->getSExtValue();
4364 if (ExtractVec == SourceVecs[0]) {
4365 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4367 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4371 // Final check before we try to produce nonsense...
4372 if (isShuffleMaskLegal(Mask, VT))
4373 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4379 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4380 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4381 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4382 /// are assumed to be legal.
4384 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4386 if (VT.getVectorNumElements() == 4 &&
4387 (VT.is128BitVector() || VT.is64BitVector())) {
4388 unsigned PFIndexes[4];
4389 for (unsigned i = 0; i != 4; ++i) {
4393 PFIndexes[i] = M[i];
4396 // Compute the index in the perfect shuffle table.
4397 unsigned PFTableIndex =
4398 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4399 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4400 unsigned Cost = (PFEntry >> 30);
4407 unsigned Imm, WhichResult;
4409 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4410 return (EltSize >= 32 ||
4411 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4412 isVREVMask(M, VT, 64) ||
4413 isVREVMask(M, VT, 32) ||
4414 isVREVMask(M, VT, 16) ||
4415 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4416 isVTBLMask(M, VT) ||
4417 isVTRNMask(M, VT, WhichResult) ||
4418 isVUZPMask(M, VT, WhichResult) ||
4419 isVZIPMask(M, VT, WhichResult) ||
4420 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4421 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4422 isVZIP_v_undef_Mask(M, VT, WhichResult));
4425 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4426 /// the specified operations to build the shuffle.
4427 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4428 SDValue RHS, SelectionDAG &DAG,
4430 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4431 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4432 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4435 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4444 OP_VUZPL, // VUZP, left result
4445 OP_VUZPR, // VUZP, right result
4446 OP_VZIPL, // VZIP, left result
4447 OP_VZIPR, // VZIP, right result
4448 OP_VTRNL, // VTRN, left result
4449 OP_VTRNR // VTRN, right result
4452 if (OpNum == OP_COPY) {
4453 if (LHSID == (1*9+2)*9+3) return LHS;
4454 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4458 SDValue OpLHS, OpRHS;
4459 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4460 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4461 EVT VT = OpLHS.getValueType();
4464 default: llvm_unreachable("Unknown shuffle opcode!");
4466 // VREV divides the vector in half and swaps within the half.
4467 if (VT.getVectorElementType() == MVT::i32 ||
4468 VT.getVectorElementType() == MVT::f32)
4469 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4470 // vrev <4 x i16> -> VREV32
4471 if (VT.getVectorElementType() == MVT::i16)
4472 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4473 // vrev <4 x i8> -> VREV16
4474 assert(VT.getVectorElementType() == MVT::i8);
4475 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4480 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4481 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4485 return DAG.getNode(ARMISD::VEXT, dl, VT,
4487 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4490 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4491 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4494 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4495 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4498 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4499 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4503 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4504 ArrayRef<int> ShuffleMask,
4505 SelectionDAG &DAG) {
4506 // Check to see if we can use the VTBL instruction.
4507 SDValue V1 = Op.getOperand(0);
4508 SDValue V2 = Op.getOperand(1);
4509 DebugLoc DL = Op.getDebugLoc();
4511 SmallVector<SDValue, 8> VTBLMask;
4512 for (ArrayRef<int>::iterator
4513 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4514 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4516 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4517 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4518 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4521 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4522 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4526 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4527 SDValue V1 = Op.getOperand(0);
4528 SDValue V2 = Op.getOperand(1);
4529 DebugLoc dl = Op.getDebugLoc();
4530 EVT VT = Op.getValueType();
4531 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4533 // Convert shuffles that are directly supported on NEON to target-specific
4534 // DAG nodes, instead of keeping them as shuffles and matching them again
4535 // during code selection. This is more efficient and avoids the possibility
4536 // of inconsistencies between legalization and selection.
4537 // FIXME: floating-point vectors should be canonicalized to integer vectors
4538 // of the same time so that they get CSEd properly.
4539 ArrayRef<int> ShuffleMask = SVN->getMask();
4541 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4542 if (EltSize <= 32) {
4543 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4544 int Lane = SVN->getSplatIndex();
4545 // If this is undef splat, generate it via "just" vdup, if possible.
4546 if (Lane == -1) Lane = 0;
4548 // Test if V1 is a SCALAR_TO_VECTOR.
4549 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4550 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4552 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4553 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4555 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4556 !isa<ConstantSDNode>(V1.getOperand(0))) {
4557 bool IsScalarToVector = true;
4558 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4559 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4560 IsScalarToVector = false;
4563 if (IsScalarToVector)
4564 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4566 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4567 DAG.getConstant(Lane, MVT::i32));
4572 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4575 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4576 DAG.getConstant(Imm, MVT::i32));
4579 if (isVREVMask(ShuffleMask, VT, 64))
4580 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4581 if (isVREVMask(ShuffleMask, VT, 32))
4582 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4583 if (isVREVMask(ShuffleMask, VT, 16))
4584 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4586 // Check for Neon shuffles that modify both input vectors in place.
4587 // If both results are used, i.e., if there are two shuffles with the same
4588 // source operands and with masks corresponding to both results of one of
4589 // these operations, DAG memoization will ensure that a single node is
4590 // used for both shuffles.
4591 unsigned WhichResult;
4592 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4593 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4594 V1, V2).getValue(WhichResult);
4595 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4596 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4597 V1, V2).getValue(WhichResult);
4598 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4599 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4600 V1, V2).getValue(WhichResult);
4602 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4603 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4604 V1, V1).getValue(WhichResult);
4605 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4606 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4607 V1, V1).getValue(WhichResult);
4608 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4609 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4610 V1, V1).getValue(WhichResult);
4613 // If the shuffle is not directly supported and it has 4 elements, use
4614 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4615 unsigned NumElts = VT.getVectorNumElements();
4617 unsigned PFIndexes[4];
4618 for (unsigned i = 0; i != 4; ++i) {
4619 if (ShuffleMask[i] < 0)
4622 PFIndexes[i] = ShuffleMask[i];
4625 // Compute the index in the perfect shuffle table.
4626 unsigned PFTableIndex =
4627 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4628 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4629 unsigned Cost = (PFEntry >> 30);
4632 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4635 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4636 if (EltSize >= 32) {
4637 // Do the expansion with floating-point types, since that is what the VFP
4638 // registers are defined to use, and since i64 is not legal.
4639 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4640 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4641 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4642 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4643 SmallVector<SDValue, 8> Ops;
4644 for (unsigned i = 0; i < NumElts; ++i) {
4645 if (ShuffleMask[i] < 0)
4646 Ops.push_back(DAG.getUNDEF(EltVT));
4648 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4649 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4650 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4653 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4654 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4657 if (VT == MVT::v8i8) {
4658 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4659 if (NewOp.getNode())
4666 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4667 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4668 SDValue Lane = Op.getOperand(2);
4669 if (!isa<ConstantSDNode>(Lane))
4675 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4676 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4677 SDValue Lane = Op.getOperand(1);
4678 if (!isa<ConstantSDNode>(Lane))
4681 SDValue Vec = Op.getOperand(0);
4682 if (Op.getValueType() == MVT::i32 &&
4683 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4684 DebugLoc dl = Op.getDebugLoc();
4685 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4691 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4692 // The only time a CONCAT_VECTORS operation can have legal types is when
4693 // two 64-bit vectors are concatenated to a 128-bit vector.
4694 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4695 "unexpected CONCAT_VECTORS");
4696 DebugLoc dl = Op.getDebugLoc();
4697 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4698 SDValue Op0 = Op.getOperand(0);
4699 SDValue Op1 = Op.getOperand(1);
4700 if (Op0.getOpcode() != ISD::UNDEF)
4701 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4702 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4703 DAG.getIntPtrConstant(0));
4704 if (Op1.getOpcode() != ISD::UNDEF)
4705 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4706 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4707 DAG.getIntPtrConstant(1));
4708 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4711 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4712 /// element has been zero/sign-extended, depending on the isSigned parameter,
4713 /// from an integer type half its size.
4714 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4716 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4717 EVT VT = N->getValueType(0);
4718 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4719 SDNode *BVN = N->getOperand(0).getNode();
4720 if (BVN->getValueType(0) != MVT::v4i32 ||
4721 BVN->getOpcode() != ISD::BUILD_VECTOR)
4723 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4724 unsigned HiElt = 1 - LoElt;
4725 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4726 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4727 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4728 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4729 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4732 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4733 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4736 if (Hi0->isNullValue() && Hi1->isNullValue())
4742 if (N->getOpcode() != ISD::BUILD_VECTOR)
4745 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4746 SDNode *Elt = N->getOperand(i).getNode();
4747 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4748 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4749 unsigned HalfSize = EltSize / 2;
4751 if (!isIntN(HalfSize, C->getSExtValue()))
4754 if (!isUIntN(HalfSize, C->getZExtValue()))
4765 /// isSignExtended - Check if a node is a vector value that is sign-extended
4766 /// or a constant BUILD_VECTOR with sign-extended elements.
4767 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4768 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4770 if (isExtendedBUILD_VECTOR(N, DAG, true))
4775 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4776 /// or a constant BUILD_VECTOR with zero-extended elements.
4777 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4778 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4780 if (isExtendedBUILD_VECTOR(N, DAG, false))
4785 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4786 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4787 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4788 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4789 return N->getOperand(0);
4790 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4791 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4792 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4793 LD->isNonTemporal(), LD->isInvariant(),
4794 LD->getAlignment());
4795 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4796 // have been legalized as a BITCAST from v4i32.
4797 if (N->getOpcode() == ISD::BITCAST) {
4798 SDNode *BVN = N->getOperand(0).getNode();
4799 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4800 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4801 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4802 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4803 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4805 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4806 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4807 EVT VT = N->getValueType(0);
4808 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4809 unsigned NumElts = VT.getVectorNumElements();
4810 MVT TruncVT = MVT::getIntegerVT(EltSize);
4811 SmallVector<SDValue, 8> Ops;
4812 for (unsigned i = 0; i != NumElts; ++i) {
4813 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4814 const APInt &CInt = C->getAPIntValue();
4815 // Element types smaller than 32 bits are not legal, so use i32 elements.
4816 // The values are implicitly truncated so sext vs. zext doesn't matter.
4817 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4819 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4820 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4823 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4824 unsigned Opcode = N->getOpcode();
4825 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4826 SDNode *N0 = N->getOperand(0).getNode();
4827 SDNode *N1 = N->getOperand(1).getNode();
4828 return N0->hasOneUse() && N1->hasOneUse() &&
4829 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4834 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4835 unsigned Opcode = N->getOpcode();
4836 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4837 SDNode *N0 = N->getOperand(0).getNode();
4838 SDNode *N1 = N->getOperand(1).getNode();
4839 return N0->hasOneUse() && N1->hasOneUse() &&
4840 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4845 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4846 // Multiplications are only custom-lowered for 128-bit vectors so that
4847 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4848 EVT VT = Op.getValueType();
4849 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4850 SDNode *N0 = Op.getOperand(0).getNode();
4851 SDNode *N1 = Op.getOperand(1).getNode();
4852 unsigned NewOpc = 0;
4854 bool isN0SExt = isSignExtended(N0, DAG);
4855 bool isN1SExt = isSignExtended(N1, DAG);
4856 if (isN0SExt && isN1SExt)
4857 NewOpc = ARMISD::VMULLs;
4859 bool isN0ZExt = isZeroExtended(N0, DAG);
4860 bool isN1ZExt = isZeroExtended(N1, DAG);
4861 if (isN0ZExt && isN1ZExt)
4862 NewOpc = ARMISD::VMULLu;
4863 else if (isN1SExt || isN1ZExt) {
4864 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4865 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4866 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4867 NewOpc = ARMISD::VMULLs;
4869 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4870 NewOpc = ARMISD::VMULLu;
4872 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4874 NewOpc = ARMISD::VMULLu;
4880 if (VT == MVT::v2i64)
4881 // Fall through to expand this. It is not legal.
4884 // Other vector multiplications are legal.
4889 // Legalize to a VMULL instruction.
4890 DebugLoc DL = Op.getDebugLoc();
4892 SDValue Op1 = SkipExtension(N1, DAG);
4894 Op0 = SkipExtension(N0, DAG);
4895 assert(Op0.getValueType().is64BitVector() &&
4896 Op1.getValueType().is64BitVector() &&
4897 "unexpected types for extended operands to VMULL");
4898 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4901 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4902 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4909 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4910 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4911 EVT Op1VT = Op1.getValueType();
4912 return DAG.getNode(N0->getOpcode(), DL, VT,
4913 DAG.getNode(NewOpc, DL, VT,
4914 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4915 DAG.getNode(NewOpc, DL, VT,
4916 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4920 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4922 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4923 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4924 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4925 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4926 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4927 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4928 // Get reciprocal estimate.
4929 // float4 recip = vrecpeq_f32(yf);
4930 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4931 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4932 // Because char has a smaller range than uchar, we can actually get away
4933 // without any newton steps. This requires that we use a weird bias
4934 // of 0xb000, however (again, this has been exhaustively tested).
4935 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4936 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4937 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4938 Y = DAG.getConstant(0xb000, MVT::i32);
4939 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4940 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4941 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4942 // Convert back to short.
4943 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4944 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4949 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4951 // Convert to float.
4952 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4953 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4954 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4955 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4956 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4957 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4959 // Use reciprocal estimate and one refinement step.
4960 // float4 recip = vrecpeq_f32(yf);
4961 // recip *= vrecpsq_f32(yf, recip);
4962 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4963 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4964 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4965 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4967 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4968 // Because short has a smaller range than ushort, we can actually get away
4969 // with only a single newton step. This requires that we use a weird bias
4970 // of 89, however (again, this has been exhaustively tested).
4971 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4972 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4973 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4974 N1 = DAG.getConstant(0x89, MVT::i32);
4975 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4976 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4977 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4978 // Convert back to integer and return.
4979 // return vmovn_s32(vcvt_s32_f32(result));
4980 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4981 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4985 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4986 EVT VT = Op.getValueType();
4987 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4988 "unexpected type for custom-lowering ISD::SDIV");
4990 DebugLoc dl = Op.getDebugLoc();
4991 SDValue N0 = Op.getOperand(0);
4992 SDValue N1 = Op.getOperand(1);
4995 if (VT == MVT::v8i8) {
4996 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4997 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4999 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5000 DAG.getIntPtrConstant(4));
5001 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5002 DAG.getIntPtrConstant(4));
5003 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5004 DAG.getIntPtrConstant(0));
5005 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5006 DAG.getIntPtrConstant(0));
5008 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5009 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5011 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5012 N0 = LowerCONCAT_VECTORS(N0, DAG);
5014 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5017 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5020 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5021 EVT VT = Op.getValueType();
5022 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5023 "unexpected type for custom-lowering ISD::UDIV");
5025 DebugLoc dl = Op.getDebugLoc();
5026 SDValue N0 = Op.getOperand(0);
5027 SDValue N1 = Op.getOperand(1);
5030 if (VT == MVT::v8i8) {
5031 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5032 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5034 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5035 DAG.getIntPtrConstant(4));
5036 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5037 DAG.getIntPtrConstant(4));
5038 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5039 DAG.getIntPtrConstant(0));
5040 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5041 DAG.getIntPtrConstant(0));
5043 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5044 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5046 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5047 N0 = LowerCONCAT_VECTORS(N0, DAG);
5049 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5050 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5055 // v4i16 sdiv ... Convert to float.
5056 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5057 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5058 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5059 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5060 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5061 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5063 // Use reciprocal estimate and two refinement steps.
5064 // float4 recip = vrecpeq_f32(yf);
5065 // recip *= vrecpsq_f32(yf, recip);
5066 // recip *= vrecpsq_f32(yf, recip);
5067 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5068 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5069 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5070 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5072 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5073 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5074 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5076 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5077 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5078 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5079 // and that it will never cause us to return an answer too large).
5080 // float4 result = as_float4(as_int4(xf*recip) + 2);
5081 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5082 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5083 N1 = DAG.getConstant(2, MVT::i32);
5084 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5085 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5086 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5087 // Convert back to integer and return.
5088 // return vmovn_u32(vcvt_s32_f32(result));
5089 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5090 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5094 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5095 EVT VT = Op.getNode()->getValueType(0);
5096 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5099 bool ExtraOp = false;
5100 switch (Op.getOpcode()) {
5101 default: llvm_unreachable("Invalid code");
5102 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5103 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5104 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5105 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5109 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5111 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5112 Op.getOperand(1), Op.getOperand(2));
5115 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5116 // Monotonic load/store is legal for all targets
5117 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5120 // Aquire/Release load/store is not legal for targets without a
5121 // dmb or equivalent available.
5127 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5128 SelectionDAG &DAG, unsigned NewOp) {
5129 DebugLoc dl = Node->getDebugLoc();
5130 assert (Node->getValueType(0) == MVT::i64 &&
5131 "Only know how to expand i64 atomics");
5133 SmallVector<SDValue, 6> Ops;
5134 Ops.push_back(Node->getOperand(0)); // Chain
5135 Ops.push_back(Node->getOperand(1)); // Ptr
5137 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5138 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5139 // High part of Val1
5140 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5141 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5142 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5143 // High part of Val1
5144 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5145 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5146 // High part of Val2
5147 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5148 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5150 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5152 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5153 cast<MemSDNode>(Node)->getMemOperand());
5154 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5155 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5156 Results.push_back(Result.getValue(2));
5159 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5160 switch (Op.getOpcode()) {
5161 default: llvm_unreachable("Don't know how to custom lower this!");
5162 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5163 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5164 case ISD::GlobalAddress:
5165 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5166 LowerGlobalAddressELF(Op, DAG);
5167 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5168 case ISD::SELECT: return LowerSELECT(Op, DAG);
5169 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5170 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5171 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5172 case ISD::VASTART: return LowerVASTART(Op, DAG);
5173 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5174 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5175 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5176 case ISD::SINT_TO_FP:
5177 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5178 case ISD::FP_TO_SINT:
5179 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5180 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5181 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5182 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5183 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5184 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5185 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5186 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5188 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5191 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5192 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5193 case ISD::SRL_PARTS:
5194 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5195 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5196 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5197 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5198 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5199 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5200 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5201 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5202 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5203 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5204 case ISD::MUL: return LowerMUL(Op, DAG);
5205 case ISD::SDIV: return LowerSDIV(Op, DAG);
5206 case ISD::UDIV: return LowerUDIV(Op, DAG);
5210 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5211 case ISD::ATOMIC_LOAD:
5212 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5216 /// ReplaceNodeResults - Replace the results of node with an illegal result
5217 /// type with new values built out of custom code.
5218 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5219 SmallVectorImpl<SDValue>&Results,
5220 SelectionDAG &DAG) const {
5222 switch (N->getOpcode()) {
5224 llvm_unreachable("Don't know how to custom expand this!");
5226 Res = ExpandBITCAST(N, DAG);
5230 Res = Expand64BitShift(N, DAG, Subtarget);
5232 case ISD::ATOMIC_LOAD_ADD:
5233 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5235 case ISD::ATOMIC_LOAD_AND:
5236 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5238 case ISD::ATOMIC_LOAD_NAND:
5239 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5241 case ISD::ATOMIC_LOAD_OR:
5242 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5244 case ISD::ATOMIC_LOAD_SUB:
5245 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5247 case ISD::ATOMIC_LOAD_XOR:
5248 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5250 case ISD::ATOMIC_SWAP:
5251 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5253 case ISD::ATOMIC_CMP_SWAP:
5254 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5258 Results.push_back(Res);
5261 //===----------------------------------------------------------------------===//
5262 // ARM Scheduler Hooks
5263 //===----------------------------------------------------------------------===//
5266 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5267 MachineBasicBlock *BB,
5268 unsigned Size) const {
5269 unsigned dest = MI->getOperand(0).getReg();
5270 unsigned ptr = MI->getOperand(1).getReg();
5271 unsigned oldval = MI->getOperand(2).getReg();
5272 unsigned newval = MI->getOperand(3).getReg();
5273 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5274 DebugLoc dl = MI->getDebugLoc();
5275 bool isThumb2 = Subtarget->isThumb2();
5277 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5278 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5279 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5280 (const TargetRegisterClass*)&ARM::GPRRegClass);
5283 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5284 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5285 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5288 unsigned ldrOpc, strOpc;
5290 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5292 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5293 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5296 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5297 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5300 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5301 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5305 MachineFunction *MF = BB->getParent();
5306 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5307 MachineFunction::iterator It = BB;
5308 ++It; // insert the new blocks after the current block
5310 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5311 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5312 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5313 MF->insert(It, loop1MBB);
5314 MF->insert(It, loop2MBB);
5315 MF->insert(It, exitMBB);
5317 // Transfer the remainder of BB and its successor edges to exitMBB.
5318 exitMBB->splice(exitMBB->begin(), BB,
5319 llvm::next(MachineBasicBlock::iterator(MI)),
5321 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5325 // fallthrough --> loop1MBB
5326 BB->addSuccessor(loop1MBB);
5329 // ldrex dest, [ptr]
5333 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5334 if (ldrOpc == ARM::t2LDREX)
5336 AddDefaultPred(MIB);
5337 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5338 .addReg(dest).addReg(oldval));
5339 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5340 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5341 BB->addSuccessor(loop2MBB);
5342 BB->addSuccessor(exitMBB);
5345 // strex scratch, newval, [ptr]
5349 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5350 if (strOpc == ARM::t2STREX)
5352 AddDefaultPred(MIB);
5353 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5354 .addReg(scratch).addImm(0));
5355 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5356 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5357 BB->addSuccessor(loop1MBB);
5358 BB->addSuccessor(exitMBB);
5364 MI->eraseFromParent(); // The instruction is gone now.
5370 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5371 unsigned Size, unsigned BinOpcode) const {
5372 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5373 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5375 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5376 MachineFunction *MF = BB->getParent();
5377 MachineFunction::iterator It = BB;
5380 unsigned dest = MI->getOperand(0).getReg();
5381 unsigned ptr = MI->getOperand(1).getReg();
5382 unsigned incr = MI->getOperand(2).getReg();
5383 DebugLoc dl = MI->getDebugLoc();
5384 bool isThumb2 = Subtarget->isThumb2();
5386 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5388 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5389 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5392 unsigned ldrOpc, strOpc;
5394 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5396 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5397 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5400 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5401 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5404 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5405 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5409 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5410 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5411 MF->insert(It, loopMBB);
5412 MF->insert(It, exitMBB);
5414 // Transfer the remainder of BB and its successor edges to exitMBB.
5415 exitMBB->splice(exitMBB->begin(), BB,
5416 llvm::next(MachineBasicBlock::iterator(MI)),
5418 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5420 const TargetRegisterClass *TRC = isThumb2 ?
5421 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5422 (const TargetRegisterClass*)&ARM::GPRRegClass;
5423 unsigned scratch = MRI.createVirtualRegister(TRC);
5424 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5428 // fallthrough --> loopMBB
5429 BB->addSuccessor(loopMBB);
5433 // <binop> scratch2, dest, incr
5434 // strex scratch, scratch2, ptr
5437 // fallthrough --> exitMBB
5439 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5440 if (ldrOpc == ARM::t2LDREX)
5442 AddDefaultPred(MIB);
5444 // operand order needs to go the other way for NAND
5445 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5446 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5447 addReg(incr).addReg(dest)).addReg(0);
5449 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5450 addReg(dest).addReg(incr)).addReg(0);
5453 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5454 if (strOpc == ARM::t2STREX)
5456 AddDefaultPred(MIB);
5457 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5458 .addReg(scratch).addImm(0));
5459 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5460 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5462 BB->addSuccessor(loopMBB);
5463 BB->addSuccessor(exitMBB);
5469 MI->eraseFromParent(); // The instruction is gone now.
5475 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5476 MachineBasicBlock *BB,
5479 ARMCC::CondCodes Cond) const {
5480 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5482 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5483 MachineFunction *MF = BB->getParent();
5484 MachineFunction::iterator It = BB;
5487 unsigned dest = MI->getOperand(0).getReg();
5488 unsigned ptr = MI->getOperand(1).getReg();
5489 unsigned incr = MI->getOperand(2).getReg();
5490 unsigned oldval = dest;
5491 DebugLoc dl = MI->getDebugLoc();
5492 bool isThumb2 = Subtarget->isThumb2();
5494 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5496 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5497 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5500 unsigned ldrOpc, strOpc, extendOpc;
5502 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5504 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5505 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5506 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5509 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5510 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5511 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5514 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5515 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5520 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5521 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5522 MF->insert(It, loopMBB);
5523 MF->insert(It, exitMBB);
5525 // Transfer the remainder of BB and its successor edges to exitMBB.
5526 exitMBB->splice(exitMBB->begin(), BB,
5527 llvm::next(MachineBasicBlock::iterator(MI)),
5529 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5531 const TargetRegisterClass *TRC = isThumb2 ?
5532 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5533 (const TargetRegisterClass*)&ARM::GPRRegClass;
5534 unsigned scratch = MRI.createVirtualRegister(TRC);
5535 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5539 // fallthrough --> loopMBB
5540 BB->addSuccessor(loopMBB);
5544 // (sign extend dest, if required)
5546 // cmov.cond scratch2, dest, incr
5547 // strex scratch, scratch2, ptr
5550 // fallthrough --> exitMBB
5552 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5553 if (ldrOpc == ARM::t2LDREX)
5555 AddDefaultPred(MIB);
5557 // Sign extend the value, if necessary.
5558 if (signExtend && extendOpc) {
5559 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5560 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5565 // Build compare and cmov instructions.
5566 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5567 .addReg(oldval).addReg(incr));
5568 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5569 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5571 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5572 if (strOpc == ARM::t2STREX)
5574 AddDefaultPred(MIB);
5575 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5576 .addReg(scratch).addImm(0));
5577 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5578 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5580 BB->addSuccessor(loopMBB);
5581 BB->addSuccessor(exitMBB);
5587 MI->eraseFromParent(); // The instruction is gone now.
5593 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5594 unsigned Op1, unsigned Op2,
5595 bool NeedsCarry, bool IsCmpxchg) const {
5596 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5597 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5599 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5600 MachineFunction *MF = BB->getParent();
5601 MachineFunction::iterator It = BB;
5604 unsigned destlo = MI->getOperand(0).getReg();
5605 unsigned desthi = MI->getOperand(1).getReg();
5606 unsigned ptr = MI->getOperand(2).getReg();
5607 unsigned vallo = MI->getOperand(3).getReg();
5608 unsigned valhi = MI->getOperand(4).getReg();
5609 DebugLoc dl = MI->getDebugLoc();
5610 bool isThumb2 = Subtarget->isThumb2();
5612 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5614 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5615 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5616 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5619 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5620 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5622 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5623 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5625 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5626 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5628 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5629 MF->insert(It, loopMBB);
5631 MF->insert(It, contBB);
5632 MF->insert(It, cont2BB);
5634 MF->insert(It, exitMBB);
5636 // Transfer the remainder of BB and its successor edges to exitMBB.
5637 exitMBB->splice(exitMBB->begin(), BB,
5638 llvm::next(MachineBasicBlock::iterator(MI)),
5640 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5642 const TargetRegisterClass *TRC = isThumb2 ?
5643 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5644 (const TargetRegisterClass*)&ARM::GPRRegClass;
5645 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5649 // fallthrough --> loopMBB
5650 BB->addSuccessor(loopMBB);
5653 // ldrexd r2, r3, ptr
5654 // <binopa> r0, r2, incr
5655 // <binopb> r1, r3, incr
5656 // strexd storesuccess, r0, r1, ptr
5657 // cmp storesuccess, #0
5659 // fallthrough --> exitMBB
5661 // Note that the registers are explicitly specified because there is not any
5662 // way to force the register allocator to allocate a register pair.
5664 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5665 // need to properly enforce the restriction that the two output registers
5666 // for ldrexd must be different.
5669 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5670 .addReg(ARM::R2, RegState::Define)
5671 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5672 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5673 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5674 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5678 for (unsigned i = 0; i < 2; i++) {
5679 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5681 .addReg(i == 0 ? destlo : desthi)
5682 .addReg(i == 0 ? vallo : valhi));
5683 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5684 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5685 BB->addSuccessor(exitMBB);
5686 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5687 BB = (i == 0 ? contBB : cont2BB);
5690 // Copy to physregs for strexd
5691 unsigned setlo = MI->getOperand(5).getReg();
5692 unsigned sethi = MI->getOperand(6).getReg();
5693 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5694 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5696 // Perform binary operation
5697 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5698 .addReg(destlo).addReg(vallo))
5699 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5700 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5701 .addReg(desthi).addReg(valhi)).addReg(0);
5703 // Copy to physregs for strexd
5704 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5705 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5709 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5710 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5712 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5713 .addReg(storesuccess).addImm(0));
5714 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5715 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5717 BB->addSuccessor(loopMBB);
5718 BB->addSuccessor(exitMBB);
5724 MI->eraseFromParent(); // The instruction is gone now.
5729 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5730 /// registers the function context.
5731 void ARMTargetLowering::
5732 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5733 MachineBasicBlock *DispatchBB, int FI) const {
5734 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5735 DebugLoc dl = MI->getDebugLoc();
5736 MachineFunction *MF = MBB->getParent();
5737 MachineRegisterInfo *MRI = &MF->getRegInfo();
5738 MachineConstantPool *MCP = MF->getConstantPool();
5739 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5740 const Function *F = MF->getFunction();
5742 bool isThumb = Subtarget->isThumb();
5743 bool isThumb2 = Subtarget->isThumb2();
5745 unsigned PCLabelId = AFI->createPICLabelUId();
5746 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5747 ARMConstantPoolValue *CPV =
5748 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5749 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5751 const TargetRegisterClass *TRC = isThumb ?
5752 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5753 (const TargetRegisterClass*)&ARM::GPRRegClass;
5755 // Grab constant pool and fixed stack memory operands.
5756 MachineMemOperand *CPMMO =
5757 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5758 MachineMemOperand::MOLoad, 4, 4);
5760 MachineMemOperand *FIMMOSt =
5761 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5762 MachineMemOperand::MOStore, 4, 4);
5764 // Load the address of the dispatch MBB into the jump buffer.
5766 // Incoming value: jbuf
5767 // ldr.n r5, LCPI1_1
5770 // str r5, [$jbuf, #+4] ; &jbuf[1]
5771 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5772 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5773 .addConstantPoolIndex(CPI)
5774 .addMemOperand(CPMMO));
5775 // Set the low bit because of thumb mode.
5776 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5778 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5779 .addReg(NewVReg1, RegState::Kill)
5781 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5782 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5783 .addReg(NewVReg2, RegState::Kill)
5785 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5786 .addReg(NewVReg3, RegState::Kill)
5788 .addImm(36) // &jbuf[1] :: pc
5789 .addMemOperand(FIMMOSt));
5790 } else if (isThumb) {
5791 // Incoming value: jbuf
5792 // ldr.n r1, LCPI1_4
5796 // add r2, $jbuf, #+4 ; &jbuf[1]
5798 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5799 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5800 .addConstantPoolIndex(CPI)
5801 .addMemOperand(CPMMO));
5802 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5803 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5804 .addReg(NewVReg1, RegState::Kill)
5806 // Set the low bit because of thumb mode.
5807 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5808 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5809 .addReg(ARM::CPSR, RegState::Define)
5811 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5812 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5813 .addReg(ARM::CPSR, RegState::Define)
5814 .addReg(NewVReg2, RegState::Kill)
5815 .addReg(NewVReg3, RegState::Kill));
5816 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5817 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5819 .addImm(36)); // &jbuf[1] :: pc
5820 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5821 .addReg(NewVReg4, RegState::Kill)
5822 .addReg(NewVReg5, RegState::Kill)
5824 .addMemOperand(FIMMOSt));
5826 // Incoming value: jbuf
5829 // str r1, [$jbuf, #+4] ; &jbuf[1]
5830 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5831 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5832 .addConstantPoolIndex(CPI)
5834 .addMemOperand(CPMMO));
5835 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5836 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5837 .addReg(NewVReg1, RegState::Kill)
5838 .addImm(PCLabelId));
5839 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5840 .addReg(NewVReg2, RegState::Kill)
5842 .addImm(36) // &jbuf[1] :: pc
5843 .addMemOperand(FIMMOSt));
5847 MachineBasicBlock *ARMTargetLowering::
5848 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5849 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5850 DebugLoc dl = MI->getDebugLoc();
5851 MachineFunction *MF = MBB->getParent();
5852 MachineRegisterInfo *MRI = &MF->getRegInfo();
5853 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5854 MachineFrameInfo *MFI = MF->getFrameInfo();
5855 int FI = MFI->getFunctionContextIndex();
5857 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5858 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5859 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
5861 // Get a mapping of the call site numbers to all of the landing pads they're
5863 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5864 unsigned MaxCSNum = 0;
5865 MachineModuleInfo &MMI = MF->getMMI();
5866 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5868 if (!BB->isLandingPad()) continue;
5870 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5872 for (MachineBasicBlock::iterator
5873 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5874 if (!II->isEHLabel()) continue;
5876 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5877 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5879 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5880 for (SmallVectorImpl<unsigned>::iterator
5881 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5882 CSI != CSE; ++CSI) {
5883 CallSiteNumToLPad[*CSI].push_back(BB);
5884 MaxCSNum = std::max(MaxCSNum, *CSI);
5890 // Get an ordered list of the machine basic blocks for the jump table.
5891 std::vector<MachineBasicBlock*> LPadList;
5892 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5893 LPadList.reserve(CallSiteNumToLPad.size());
5894 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5895 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5896 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5897 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5898 LPadList.push_back(*II);
5899 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5903 assert(!LPadList.empty() &&
5904 "No landing pad destinations for the dispatch jump table!");
5906 // Create the jump table and associated information.
5907 MachineJumpTableInfo *JTI =
5908 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5909 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5910 unsigned UId = AFI->createJumpTableUId();
5912 // Create the MBBs for the dispatch code.
5914 // Shove the dispatch's address into the return slot in the function context.
5915 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5916 DispatchBB->setIsLandingPad();
5918 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5919 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5920 DispatchBB->addSuccessor(TrapBB);
5922 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5923 DispatchBB->addSuccessor(DispContBB);
5926 MF->insert(MF->end(), DispatchBB);
5927 MF->insert(MF->end(), DispContBB);
5928 MF->insert(MF->end(), TrapBB);
5930 // Insert code into the entry block that creates and registers the function
5932 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5934 MachineMemOperand *FIMMOLd =
5935 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5936 MachineMemOperand::MOLoad |
5937 MachineMemOperand::MOVolatile, 4, 4);
5939 if (AFI->isThumb1OnlyFunction())
5940 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5941 else if (!Subtarget->hasVFP2())
5942 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5944 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5946 unsigned NumLPads = LPadList.size();
5947 if (Subtarget->isThumb2()) {
5948 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5949 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5952 .addMemOperand(FIMMOLd));
5954 if (NumLPads < 256) {
5955 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5957 .addImm(LPadList.size()));
5959 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5960 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5961 .addImm(NumLPads & 0xFFFF));
5963 unsigned VReg2 = VReg1;
5964 if ((NumLPads & 0xFFFF0000) != 0) {
5965 VReg2 = MRI->createVirtualRegister(TRC);
5966 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5968 .addImm(NumLPads >> 16));
5971 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5976 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5981 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5982 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5983 .addJumpTableIndex(MJTI)
5986 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5989 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5990 .addReg(NewVReg3, RegState::Kill)
5992 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5994 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5995 .addReg(NewVReg4, RegState::Kill)
5997 .addJumpTableIndex(MJTI)
5999 } else if (Subtarget->isThumb()) {
6000 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6001 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6004 .addMemOperand(FIMMOLd));
6006 if (NumLPads < 256) {
6007 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6011 MachineConstantPool *ConstantPool = MF->getConstantPool();
6012 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6013 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6015 // MachineConstantPool wants an explicit alignment.
6016 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6018 Align = getTargetData()->getTypeAllocSize(C->getType());
6019 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6021 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6022 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6023 .addReg(VReg1, RegState::Define)
6024 .addConstantPoolIndex(Idx));
6025 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6030 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6035 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6036 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6037 .addReg(ARM::CPSR, RegState::Define)
6041 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6042 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6043 .addJumpTableIndex(MJTI)
6046 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6047 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6048 .addReg(ARM::CPSR, RegState::Define)
6049 .addReg(NewVReg2, RegState::Kill)
6052 MachineMemOperand *JTMMOLd =
6053 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6054 MachineMemOperand::MOLoad, 4, 4);
6056 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6057 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6058 .addReg(NewVReg4, RegState::Kill)
6060 .addMemOperand(JTMMOLd));
6062 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6063 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6064 .addReg(ARM::CPSR, RegState::Define)
6065 .addReg(NewVReg5, RegState::Kill)
6068 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6069 .addReg(NewVReg6, RegState::Kill)
6070 .addJumpTableIndex(MJTI)
6073 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6074 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6077 .addMemOperand(FIMMOLd));
6079 if (NumLPads < 256) {
6080 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6083 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6084 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6085 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6086 .addImm(NumLPads & 0xFFFF));
6088 unsigned VReg2 = VReg1;
6089 if ((NumLPads & 0xFFFF0000) != 0) {
6090 VReg2 = MRI->createVirtualRegister(TRC);
6091 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6093 .addImm(NumLPads >> 16));
6096 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6100 MachineConstantPool *ConstantPool = MF->getConstantPool();
6101 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6102 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6104 // MachineConstantPool wants an explicit alignment.
6105 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6107 Align = getTargetData()->getTypeAllocSize(C->getType());
6108 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6110 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6111 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6112 .addReg(VReg1, RegState::Define)
6113 .addConstantPoolIndex(Idx)
6115 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6117 .addReg(VReg1, RegState::Kill));
6120 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6125 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6127 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6129 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6130 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6131 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6132 .addJumpTableIndex(MJTI)
6135 MachineMemOperand *JTMMOLd =
6136 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6137 MachineMemOperand::MOLoad, 4, 4);
6138 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6140 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6141 .addReg(NewVReg3, RegState::Kill)
6144 .addMemOperand(JTMMOLd));
6146 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6147 .addReg(NewVReg5, RegState::Kill)
6149 .addJumpTableIndex(MJTI)
6153 // Add the jump table entries as successors to the MBB.
6154 MachineBasicBlock *PrevMBB = 0;
6155 for (std::vector<MachineBasicBlock*>::iterator
6156 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6157 MachineBasicBlock *CurMBB = *I;
6158 if (PrevMBB != CurMBB)
6159 DispContBB->addSuccessor(CurMBB);
6163 // N.B. the order the invoke BBs are processed in doesn't matter here.
6164 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6165 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6166 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6167 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6168 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6169 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6170 MachineBasicBlock *BB = *I;
6172 // Remove the landing pad successor from the invoke block and replace it
6173 // with the new dispatch block.
6174 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6176 while (!Successors.empty()) {
6177 MachineBasicBlock *SMBB = Successors.pop_back_val();
6178 if (SMBB->isLandingPad()) {
6179 BB->removeSuccessor(SMBB);
6180 MBBLPads.push_back(SMBB);
6184 BB->addSuccessor(DispatchBB);
6186 // Find the invoke call and mark all of the callee-saved registers as
6187 // 'implicit defined' so that they're spilled. This prevents code from
6188 // moving instructions to before the EH block, where they will never be
6190 for (MachineBasicBlock::reverse_iterator
6191 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6192 if (!II->isCall()) continue;
6194 DenseMap<unsigned, bool> DefRegs;
6195 for (MachineInstr::mop_iterator
6196 OI = II->operands_begin(), OE = II->operands_end();
6198 if (!OI->isReg()) continue;
6199 DefRegs[OI->getReg()] = true;
6202 MachineInstrBuilder MIB(&*II);
6204 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6205 unsigned Reg = SavedRegs[i];
6206 if (Subtarget->isThumb2() &&
6207 !ARM::tGPRRegClass.contains(Reg) &&
6208 !ARM::hGPRRegClass.contains(Reg))
6210 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6212 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6215 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6222 // Mark all former landing pads as non-landing pads. The dispatch is the only
6224 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6225 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6226 (*I)->setIsLandingPad(false);
6228 // The instruction is gone now.
6229 MI->eraseFromParent();
6235 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6236 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6237 E = MBB->succ_end(); I != E; ++I)
6240 llvm_unreachable("Expecting a BB with two successors!");
6243 MachineBasicBlock *ARMTargetLowering::
6244 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6245 // This pseudo instruction has 3 operands: dst, src, size
6246 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6247 // Otherwise, we will generate unrolled scalar copies.
6248 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6249 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6250 MachineFunction::iterator It = BB;
6253 unsigned dest = MI->getOperand(0).getReg();
6254 unsigned src = MI->getOperand(1).getReg();
6255 unsigned SizeVal = MI->getOperand(2).getImm();
6256 unsigned Align = MI->getOperand(3).getImm();
6257 DebugLoc dl = MI->getDebugLoc();
6259 bool isThumb2 = Subtarget->isThumb2();
6260 MachineFunction *MF = BB->getParent();
6261 MachineRegisterInfo &MRI = MF->getRegInfo();
6262 unsigned ldrOpc, strOpc, UnitSize;
6264 const TargetRegisterClass *TRC = isThumb2 ?
6265 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6266 (const TargetRegisterClass*)&ARM::GPRRegClass;
6269 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6270 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6272 } else if (Align & 2) {
6273 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6274 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6277 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6278 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6281 unsigned BytesLeft = SizeVal % UnitSize;
6282 unsigned LoopSize = SizeVal - BytesLeft;
6284 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6285 // Use LDR and STR to copy.
6286 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6287 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6288 unsigned srcIn = src;
6289 unsigned destIn = dest;
6290 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6291 unsigned scratch = MRI.createVirtualRegister(TRC);
6292 unsigned srcOut = MRI.createVirtualRegister(TRC);
6293 unsigned destOut = MRI.createVirtualRegister(TRC);
6295 AddDefaultPred(BuildMI(*BB, MI, dl,
6296 TII->get(ldrOpc), scratch)
6297 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6299 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6300 .addReg(scratch).addReg(destIn)
6303 AddDefaultPred(BuildMI(*BB, MI, dl,
6304 TII->get(ldrOpc), scratch)
6305 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6308 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6309 .addReg(scratch).addReg(destIn)
6310 .addReg(0).addImm(UnitSize));
6316 // Handle the leftover bytes with LDRB and STRB.
6317 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6318 // [destOut] = STRB_POST(scratch, destIn, 1)
6319 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6320 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6321 for (unsigned i = 0; i < BytesLeft; i++) {
6322 unsigned scratch = MRI.createVirtualRegister(TRC);
6323 unsigned srcOut = MRI.createVirtualRegister(TRC);
6324 unsigned destOut = MRI.createVirtualRegister(TRC);
6326 AddDefaultPred(BuildMI(*BB, MI, dl,
6327 TII->get(ldrOpc),scratch)
6328 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6330 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6331 .addReg(scratch).addReg(destIn)
6332 .addReg(0).addImm(1));
6334 AddDefaultPred(BuildMI(*BB, MI, dl,
6335 TII->get(ldrOpc),scratch)
6336 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6338 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6339 .addReg(scratch).addReg(destIn)
6340 .addReg(0).addImm(1));
6345 MI->eraseFromParent(); // The instruction is gone now.
6349 // Expand the pseudo op to a loop.
6352 // movw varEnd, # --> with thumb2
6354 // ldrcp varEnd, idx --> without thumb2
6355 // fallthrough --> loopMBB
6357 // PHI varPhi, varEnd, varLoop
6358 // PHI srcPhi, src, srcLoop
6359 // PHI destPhi, dst, destLoop
6360 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6361 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6362 // subs varLoop, varPhi, #UnitSize
6364 // fallthrough --> exitMBB
6366 // epilogue to handle left-over bytes
6367 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6368 // [destOut] = STRB_POST(scratch, destLoop, 1)
6369 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6370 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6371 MF->insert(It, loopMBB);
6372 MF->insert(It, exitMBB);
6374 // Transfer the remainder of BB and its successor edges to exitMBB.
6375 exitMBB->splice(exitMBB->begin(), BB,
6376 llvm::next(MachineBasicBlock::iterator(MI)),
6378 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6380 // Load an immediate to varEnd.
6381 unsigned varEnd = MRI.createVirtualRegister(TRC);
6383 unsigned VReg1 = varEnd;
6384 if ((LoopSize & 0xFFFF0000) != 0)
6385 VReg1 = MRI.createVirtualRegister(TRC);
6386 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6387 .addImm(LoopSize & 0xFFFF));
6389 if ((LoopSize & 0xFFFF0000) != 0)
6390 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6392 .addImm(LoopSize >> 16));
6394 MachineConstantPool *ConstantPool = MF->getConstantPool();
6395 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6396 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6398 // MachineConstantPool wants an explicit alignment.
6399 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6401 Align = getTargetData()->getTypeAllocSize(C->getType());
6402 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6404 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6405 .addReg(varEnd, RegState::Define)
6406 .addConstantPoolIndex(Idx)
6409 BB->addSuccessor(loopMBB);
6411 // Generate the loop body:
6412 // varPhi = PHI(varLoop, varEnd)
6413 // srcPhi = PHI(srcLoop, src)
6414 // destPhi = PHI(destLoop, dst)
6415 MachineBasicBlock *entryBB = BB;
6417 unsigned varLoop = MRI.createVirtualRegister(TRC);
6418 unsigned varPhi = MRI.createVirtualRegister(TRC);
6419 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6420 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6421 unsigned destLoop = MRI.createVirtualRegister(TRC);
6422 unsigned destPhi = MRI.createVirtualRegister(TRC);
6424 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6425 .addReg(varLoop).addMBB(loopMBB)
6426 .addReg(varEnd).addMBB(entryBB);
6427 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6428 .addReg(srcLoop).addMBB(loopMBB)
6429 .addReg(src).addMBB(entryBB);
6430 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6431 .addReg(destLoop).addMBB(loopMBB)
6432 .addReg(dest).addMBB(entryBB);
6434 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6435 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6436 unsigned scratch = MRI.createVirtualRegister(TRC);
6438 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6439 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6441 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6442 .addReg(scratch).addReg(destPhi)
6445 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6446 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6449 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6450 .addReg(scratch).addReg(destPhi)
6451 .addReg(0).addImm(UnitSize));
6454 // Decrement loop variable by UnitSize.
6455 MachineInstrBuilder MIB = BuildMI(BB, dl,
6456 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6457 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6458 MIB->getOperand(5).setReg(ARM::CPSR);
6459 MIB->getOperand(5).setIsDef(true);
6461 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6462 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6464 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6465 BB->addSuccessor(loopMBB);
6466 BB->addSuccessor(exitMBB);
6468 // Add epilogue to handle BytesLeft.
6470 MachineInstr *StartOfExit = exitMBB->begin();
6471 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6472 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6474 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6475 // [destOut] = STRB_POST(scratch, destLoop, 1)
6476 unsigned srcIn = srcLoop;
6477 unsigned destIn = destLoop;
6478 for (unsigned i = 0; i < BytesLeft; i++) {
6479 unsigned scratch = MRI.createVirtualRegister(TRC);
6480 unsigned srcOut = MRI.createVirtualRegister(TRC);
6481 unsigned destOut = MRI.createVirtualRegister(TRC);
6483 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6484 TII->get(ldrOpc),scratch)
6485 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6487 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6488 .addReg(scratch).addReg(destIn)
6491 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6492 TII->get(ldrOpc),scratch)
6493 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6495 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6496 .addReg(scratch).addReg(destIn)
6497 .addReg(0).addImm(1));
6503 MI->eraseFromParent(); // The instruction is gone now.
6508 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6509 MachineBasicBlock *BB) const {
6510 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6511 DebugLoc dl = MI->getDebugLoc();
6512 bool isThumb2 = Subtarget->isThumb2();
6513 switch (MI->getOpcode()) {
6516 llvm_unreachable("Unexpected instr type to insert");
6518 // The Thumb2 pre-indexed stores have the same MI operands, they just
6519 // define them differently in the .td files from the isel patterns, so
6520 // they need pseudos.
6521 case ARM::t2STR_preidx:
6522 MI->setDesc(TII->get(ARM::t2STR_PRE));
6524 case ARM::t2STRB_preidx:
6525 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6527 case ARM::t2STRH_preidx:
6528 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6531 case ARM::STRi_preidx:
6532 case ARM::STRBi_preidx: {
6533 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6534 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6535 // Decode the offset.
6536 unsigned Offset = MI->getOperand(4).getImm();
6537 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6538 Offset = ARM_AM::getAM2Offset(Offset);
6542 MachineMemOperand *MMO = *MI->memoperands_begin();
6543 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6544 .addOperand(MI->getOperand(0)) // Rn_wb
6545 .addOperand(MI->getOperand(1)) // Rt
6546 .addOperand(MI->getOperand(2)) // Rn
6547 .addImm(Offset) // offset (skip GPR==zero_reg)
6548 .addOperand(MI->getOperand(5)) // pred
6549 .addOperand(MI->getOperand(6))
6550 .addMemOperand(MMO);
6551 MI->eraseFromParent();
6554 case ARM::STRr_preidx:
6555 case ARM::STRBr_preidx:
6556 case ARM::STRH_preidx: {
6558 switch (MI->getOpcode()) {
6559 default: llvm_unreachable("unexpected opcode!");
6560 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6561 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6562 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6564 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6565 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6566 MIB.addOperand(MI->getOperand(i));
6567 MI->eraseFromParent();
6570 case ARM::ATOMIC_LOAD_ADD_I8:
6571 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6572 case ARM::ATOMIC_LOAD_ADD_I16:
6573 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6574 case ARM::ATOMIC_LOAD_ADD_I32:
6575 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6577 case ARM::ATOMIC_LOAD_AND_I8:
6578 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6579 case ARM::ATOMIC_LOAD_AND_I16:
6580 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6581 case ARM::ATOMIC_LOAD_AND_I32:
6582 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6584 case ARM::ATOMIC_LOAD_OR_I8:
6585 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6586 case ARM::ATOMIC_LOAD_OR_I16:
6587 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6588 case ARM::ATOMIC_LOAD_OR_I32:
6589 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6591 case ARM::ATOMIC_LOAD_XOR_I8:
6592 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6593 case ARM::ATOMIC_LOAD_XOR_I16:
6594 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6595 case ARM::ATOMIC_LOAD_XOR_I32:
6596 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6598 case ARM::ATOMIC_LOAD_NAND_I8:
6599 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6600 case ARM::ATOMIC_LOAD_NAND_I16:
6601 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6602 case ARM::ATOMIC_LOAD_NAND_I32:
6603 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6605 case ARM::ATOMIC_LOAD_SUB_I8:
6606 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6607 case ARM::ATOMIC_LOAD_SUB_I16:
6608 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6609 case ARM::ATOMIC_LOAD_SUB_I32:
6610 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6612 case ARM::ATOMIC_LOAD_MIN_I8:
6613 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6614 case ARM::ATOMIC_LOAD_MIN_I16:
6615 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6616 case ARM::ATOMIC_LOAD_MIN_I32:
6617 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6619 case ARM::ATOMIC_LOAD_MAX_I8:
6620 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6621 case ARM::ATOMIC_LOAD_MAX_I16:
6622 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6623 case ARM::ATOMIC_LOAD_MAX_I32:
6624 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6626 case ARM::ATOMIC_LOAD_UMIN_I8:
6627 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6628 case ARM::ATOMIC_LOAD_UMIN_I16:
6629 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6630 case ARM::ATOMIC_LOAD_UMIN_I32:
6631 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6633 case ARM::ATOMIC_LOAD_UMAX_I8:
6634 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6635 case ARM::ATOMIC_LOAD_UMAX_I16:
6636 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6637 case ARM::ATOMIC_LOAD_UMAX_I32:
6638 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6640 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6641 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6642 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6644 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6645 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6646 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6649 case ARM::ATOMADD6432:
6650 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6651 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6652 /*NeedsCarry*/ true);
6653 case ARM::ATOMSUB6432:
6654 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6655 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6656 /*NeedsCarry*/ true);
6657 case ARM::ATOMOR6432:
6658 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6659 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6660 case ARM::ATOMXOR6432:
6661 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6662 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6663 case ARM::ATOMAND6432:
6664 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6665 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6666 case ARM::ATOMSWAP6432:
6667 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6668 case ARM::ATOMCMPXCHG6432:
6669 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6670 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6671 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6673 case ARM::tMOVCCr_pseudo: {
6674 // To "insert" a SELECT_CC instruction, we actually have to insert the
6675 // diamond control-flow pattern. The incoming instruction knows the
6676 // destination vreg to set, the condition code register to branch on, the
6677 // true/false values to select between, and a branch opcode to use.
6678 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6679 MachineFunction::iterator It = BB;
6685 // cmpTY ccX, r1, r2
6687 // fallthrough --> copy0MBB
6688 MachineBasicBlock *thisMBB = BB;
6689 MachineFunction *F = BB->getParent();
6690 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6691 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6692 F->insert(It, copy0MBB);
6693 F->insert(It, sinkMBB);
6695 // Transfer the remainder of BB and its successor edges to sinkMBB.
6696 sinkMBB->splice(sinkMBB->begin(), BB,
6697 llvm::next(MachineBasicBlock::iterator(MI)),
6699 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6701 BB->addSuccessor(copy0MBB);
6702 BB->addSuccessor(sinkMBB);
6704 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6705 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6708 // %FalseValue = ...
6709 // # fallthrough to sinkMBB
6712 // Update machine-CFG edges
6713 BB->addSuccessor(sinkMBB);
6716 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6719 BuildMI(*BB, BB->begin(), dl,
6720 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6721 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6722 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6724 MI->eraseFromParent(); // The pseudo instruction is gone now.
6729 case ARM::BCCZi64: {
6730 // If there is an unconditional branch to the other successor, remove it.
6731 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6733 // Compare both parts that make up the double comparison separately for
6735 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6737 unsigned LHS1 = MI->getOperand(1).getReg();
6738 unsigned LHS2 = MI->getOperand(2).getReg();
6740 AddDefaultPred(BuildMI(BB, dl,
6741 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6742 .addReg(LHS1).addImm(0));
6743 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6744 .addReg(LHS2).addImm(0)
6745 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6747 unsigned RHS1 = MI->getOperand(3).getReg();
6748 unsigned RHS2 = MI->getOperand(4).getReg();
6749 AddDefaultPred(BuildMI(BB, dl,
6750 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6751 .addReg(LHS1).addReg(RHS1));
6752 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6753 .addReg(LHS2).addReg(RHS2)
6754 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6757 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6758 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6759 if (MI->getOperand(0).getImm() == ARMCC::NE)
6760 std::swap(destMBB, exitMBB);
6762 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6763 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6765 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6767 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6769 MI->eraseFromParent(); // The pseudo instruction is gone now.
6773 case ARM::Int_eh_sjlj_setjmp:
6774 case ARM::Int_eh_sjlj_setjmp_nofp:
6775 case ARM::tInt_eh_sjlj_setjmp:
6776 case ARM::t2Int_eh_sjlj_setjmp:
6777 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6778 EmitSjLjDispatchBlock(MI, BB);
6783 // To insert an ABS instruction, we have to insert the
6784 // diamond control-flow pattern. The incoming instruction knows the
6785 // source vreg to test against 0, the destination vreg to set,
6786 // the condition code register to branch on, the
6787 // true/false values to select between, and a branch opcode to use.
6792 // BCC (branch to SinkBB if V0 >= 0)
6793 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6794 // SinkBB: V1 = PHI(V2, V3)
6795 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6796 MachineFunction::iterator BBI = BB;
6798 MachineFunction *Fn = BB->getParent();
6799 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6800 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6801 Fn->insert(BBI, RSBBB);
6802 Fn->insert(BBI, SinkBB);
6804 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6805 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6806 bool isThumb2 = Subtarget->isThumb2();
6807 MachineRegisterInfo &MRI = Fn->getRegInfo();
6808 // In Thumb mode S must not be specified if source register is the SP or
6809 // PC and if destination register is the SP, so restrict register class
6810 unsigned NewMovDstReg = MRI.createVirtualRegister(isThumb2 ?
6811 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6812 (const TargetRegisterClass*)&ARM::GPRRegClass);
6813 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6814 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6815 (const TargetRegisterClass*)&ARM::GPRRegClass);
6817 // Transfer the remainder of BB and its successor edges to sinkMBB.
6818 SinkBB->splice(SinkBB->begin(), BB,
6819 llvm::next(MachineBasicBlock::iterator(MI)),
6821 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6823 BB->addSuccessor(RSBBB);
6824 BB->addSuccessor(SinkBB);
6826 // fall through to SinkMBB
6827 RSBBB->addSuccessor(SinkBB);
6829 // insert a movs at the end of BB
6830 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6832 .addReg(ABSSrcReg, RegState::Kill)
6833 .addImm((unsigned)ARMCC::AL).addReg(0)
6834 .addReg(ARM::CPSR, RegState::Define);
6836 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6838 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6839 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6841 // insert rsbri in RSBBB
6842 // Note: BCC and rsbri will be converted into predicated rsbmi
6843 // by if-conversion pass
6844 BuildMI(*RSBBB, RSBBB->begin(), dl,
6845 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6846 .addReg(NewMovDstReg, RegState::Kill)
6847 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6849 // insert PHI in SinkBB,
6850 // reuse ABSDstReg to not change uses of ABS instruction
6851 BuildMI(*SinkBB, SinkBB->begin(), dl,
6852 TII->get(ARM::PHI), ABSDstReg)
6853 .addReg(NewRsbDstReg).addMBB(RSBBB)
6854 .addReg(NewMovDstReg).addMBB(BB);
6856 // remove ABS instruction
6857 MI->eraseFromParent();
6859 // return last added BB
6862 case ARM::COPY_STRUCT_BYVAL_I32:
6864 return EmitStructByval(MI, BB);
6868 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6869 SDNode *Node) const {
6870 if (!MI->hasPostISelHook()) {
6871 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6872 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6876 const MCInstrDesc *MCID = &MI->getDesc();
6877 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6878 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6879 // operand is still set to noreg. If needed, set the optional operand's
6880 // register to CPSR, and remove the redundant implicit def.
6882 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6884 // Rename pseudo opcodes.
6885 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6887 const ARMBaseInstrInfo *TII =
6888 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6889 MCID = &TII->get(NewOpc);
6891 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6892 "converted opcode should be the same except for cc_out");
6896 // Add the optional cc_out operand
6897 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6899 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6901 // Any ARM instruction that sets the 's' bit should specify an optional
6902 // "cc_out" operand in the last operand position.
6903 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6904 assert(!NewOpc && "Optional cc_out operand required");
6907 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6908 // since we already have an optional CPSR def.
6909 bool definesCPSR = false;
6910 bool deadCPSR = false;
6911 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6913 const MachineOperand &MO = MI->getOperand(i);
6914 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6918 MI->RemoveOperand(i);
6923 assert(!NewOpc && "Optional cc_out operand required");
6926 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6928 assert(!MI->getOperand(ccOutIdx).getReg() &&
6929 "expect uninitialized optional cc_out operand");
6933 // If this instruction was defined with an optional CPSR def and its dag node
6934 // had a live implicit CPSR def, then activate the optional CPSR def.
6935 MachineOperand &MO = MI->getOperand(ccOutIdx);
6936 MO.setReg(ARM::CPSR);
6940 //===----------------------------------------------------------------------===//
6941 // ARM Optimization Hooks
6942 //===----------------------------------------------------------------------===//
6945 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6946 TargetLowering::DAGCombinerInfo &DCI) {
6947 SelectionDAG &DAG = DCI.DAG;
6948 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6949 EVT VT = N->getValueType(0);
6950 unsigned Opc = N->getOpcode();
6951 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6952 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6953 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6954 ISD::CondCode CC = ISD::SETCC_INVALID;
6957 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6959 SDValue CCOp = Slct.getOperand(0);
6960 if (CCOp.getOpcode() == ISD::SETCC)
6961 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6964 bool DoXform = false;
6966 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6969 if (LHS.getOpcode() == ISD::Constant &&
6970 cast<ConstantSDNode>(LHS)->isNullValue()) {
6972 } else if (CC != ISD::SETCC_INVALID &&
6973 RHS.getOpcode() == ISD::Constant &&
6974 cast<ConstantSDNode>(RHS)->isNullValue()) {
6975 std::swap(LHS, RHS);
6976 SDValue Op0 = Slct.getOperand(0);
6977 EVT OpVT = isSlctCC ? Op0.getValueType() :
6978 Op0.getOperand(0).getValueType();
6979 bool isInt = OpVT.isInteger();
6980 CC = ISD::getSetCCInverse(CC, isInt);
6982 if (!TLI.isCondCodeLegal(CC, OpVT))
6983 return SDValue(); // Inverse operator isn't legal.
6990 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6992 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6993 Slct.getOperand(0), Slct.getOperand(1), CC);
6994 SDValue CCOp = Slct.getOperand(0);
6996 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6997 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6998 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6999 CCOp, OtherOp, Result);
7004 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7005 // (only after legalization).
7006 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7007 TargetLowering::DAGCombinerInfo &DCI,
7008 const ARMSubtarget *Subtarget) {
7010 // Only perform optimization if after legalize, and if NEON is available. We
7011 // also expected both operands to be BUILD_VECTORs.
7012 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7013 || N0.getOpcode() != ISD::BUILD_VECTOR
7014 || N1.getOpcode() != ISD::BUILD_VECTOR)
7017 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7018 EVT VT = N->getValueType(0);
7019 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7022 // Check that the vector operands are of the right form.
7023 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7024 // operands, where N is the size of the formed vector.
7025 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7026 // index such that we have a pair wise add pattern.
7028 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7029 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7031 SDValue Vec = N0->getOperand(0)->getOperand(0);
7032 SDNode *V = Vec.getNode();
7033 unsigned nextIndex = 0;
7035 // For each operands to the ADD which are BUILD_VECTORs,
7036 // check to see if each of their operands are an EXTRACT_VECTOR with
7037 // the same vector and appropriate index.
7038 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7039 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7040 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7042 SDValue ExtVec0 = N0->getOperand(i);
7043 SDValue ExtVec1 = N1->getOperand(i);
7045 // First operand is the vector, verify its the same.
7046 if (V != ExtVec0->getOperand(0).getNode() ||
7047 V != ExtVec1->getOperand(0).getNode())
7050 // Second is the constant, verify its correct.
7051 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7052 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7054 // For the constant, we want to see all the even or all the odd.
7055 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7056 || C1->getZExtValue() != nextIndex+1)
7065 // Create VPADDL node.
7066 SelectionDAG &DAG = DCI.DAG;
7067 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7069 // Build operand list.
7070 SmallVector<SDValue, 8> Ops;
7071 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7072 TLI.getPointerTy()));
7074 // Input is the vector.
7077 // Get widened type and narrowed type.
7079 unsigned numElem = VT.getVectorNumElements();
7080 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7081 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7082 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7083 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7085 llvm_unreachable("Invalid vector element type for padd optimization.");
7088 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7089 widenType, &Ops[0], Ops.size());
7090 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7093 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7094 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7095 /// called with the default operands, and if that fails, with commuted
7097 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7098 TargetLowering::DAGCombinerInfo &DCI,
7099 const ARMSubtarget *Subtarget){
7101 // Attempt to create vpaddl for this add.
7102 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7103 if (Result.getNode())
7106 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7107 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
7108 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7109 if (Result.getNode()) return Result;
7114 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7116 static SDValue PerformADDCombine(SDNode *N,
7117 TargetLowering::DAGCombinerInfo &DCI,
7118 const ARMSubtarget *Subtarget) {
7119 SDValue N0 = N->getOperand(0);
7120 SDValue N1 = N->getOperand(1);
7122 // First try with the default operand order.
7123 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7124 if (Result.getNode())
7127 // If that didn't work, try again with the operands commuted.
7128 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7131 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7133 static SDValue PerformSUBCombine(SDNode *N,
7134 TargetLowering::DAGCombinerInfo &DCI) {
7135 SDValue N0 = N->getOperand(0);
7136 SDValue N1 = N->getOperand(1);
7138 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7139 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
7140 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7141 if (Result.getNode()) return Result;
7147 /// PerformVMULCombine
7148 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7149 /// special multiplier accumulator forwarding.
7155 static SDValue PerformVMULCombine(SDNode *N,
7156 TargetLowering::DAGCombinerInfo &DCI,
7157 const ARMSubtarget *Subtarget) {
7158 if (!Subtarget->hasVMLxForwarding())
7161 SelectionDAG &DAG = DCI.DAG;
7162 SDValue N0 = N->getOperand(0);
7163 SDValue N1 = N->getOperand(1);
7164 unsigned Opcode = N0.getOpcode();
7165 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7166 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7167 Opcode = N1.getOpcode();
7168 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7169 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7174 EVT VT = N->getValueType(0);
7175 DebugLoc DL = N->getDebugLoc();
7176 SDValue N00 = N0->getOperand(0);
7177 SDValue N01 = N0->getOperand(1);
7178 return DAG.getNode(Opcode, DL, VT,
7179 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7180 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7183 static SDValue PerformMULCombine(SDNode *N,
7184 TargetLowering::DAGCombinerInfo &DCI,
7185 const ARMSubtarget *Subtarget) {
7186 SelectionDAG &DAG = DCI.DAG;
7188 if (Subtarget->isThumb1Only())
7191 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7194 EVT VT = N->getValueType(0);
7195 if (VT.is64BitVector() || VT.is128BitVector())
7196 return PerformVMULCombine(N, DCI, Subtarget);
7200 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7204 int64_t MulAmt = C->getSExtValue();
7205 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7207 ShiftAmt = ShiftAmt & (32 - 1);
7208 SDValue V = N->getOperand(0);
7209 DebugLoc DL = N->getDebugLoc();
7212 MulAmt >>= ShiftAmt;
7215 if (isPowerOf2_32(MulAmt - 1)) {
7216 // (mul x, 2^N + 1) => (add (shl x, N), x)
7217 Res = DAG.getNode(ISD::ADD, DL, VT,
7219 DAG.getNode(ISD::SHL, DL, VT,
7221 DAG.getConstant(Log2_32(MulAmt - 1),
7223 } else if (isPowerOf2_32(MulAmt + 1)) {
7224 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7225 Res = DAG.getNode(ISD::SUB, DL, VT,
7226 DAG.getNode(ISD::SHL, DL, VT,
7228 DAG.getConstant(Log2_32(MulAmt + 1),
7234 uint64_t MulAmtAbs = -MulAmt;
7235 if (isPowerOf2_32(MulAmtAbs + 1)) {
7236 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7237 Res = DAG.getNode(ISD::SUB, DL, VT,
7239 DAG.getNode(ISD::SHL, DL, VT,
7241 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7243 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7244 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7245 Res = DAG.getNode(ISD::ADD, DL, VT,
7247 DAG.getNode(ISD::SHL, DL, VT,
7249 DAG.getConstant(Log2_32(MulAmtAbs-1),
7251 Res = DAG.getNode(ISD::SUB, DL, VT,
7252 DAG.getConstant(0, MVT::i32),Res);
7259 Res = DAG.getNode(ISD::SHL, DL, VT,
7260 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7262 // Do not add new nodes to DAG combiner worklist.
7263 DCI.CombineTo(N, Res, false);
7267 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
7268 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
7271 SDValue FalseVal = N.getOperand(0);
7272 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
7276 return C->isAllOnesValue();
7277 return C->isNullValue();
7280 /// formConditionalOp - Combine an operation with a conditional move operand
7281 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
7282 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
7283 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
7285 SDValue N0 = N->getOperand(0);
7286 SDValue N1 = N->getOperand(1);
7288 bool isAND = N->getOpcode() == ISD::AND;
7289 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
7290 if (!isCand && Commutable) {
7291 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
7299 switch (N->getOpcode()) {
7300 default: llvm_unreachable("Unexpected node");
7301 case ISD::AND: Opc = ARMISD::CAND; break;
7302 case ISD::OR: Opc = ARMISD::COR; break;
7303 case ISD::XOR: Opc = ARMISD::CXOR; break;
7305 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
7306 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
7310 static SDValue PerformANDCombine(SDNode *N,
7311 TargetLowering::DAGCombinerInfo &DCI,
7312 const ARMSubtarget *Subtarget) {
7314 // Attempt to use immediate-form VBIC
7315 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7316 DebugLoc dl = N->getDebugLoc();
7317 EVT VT = N->getValueType(0);
7318 SelectionDAG &DAG = DCI.DAG;
7320 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7323 APInt SplatBits, SplatUndef;
7324 unsigned SplatBitSize;
7327 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7328 if (SplatBitSize <= 64) {
7330 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7331 SplatUndef.getZExtValue(), SplatBitSize,
7332 DAG, VbicVT, VT.is128BitVector(),
7334 if (Val.getNode()) {
7336 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7337 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7338 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7343 if (!Subtarget->isThumb1Only()) {
7344 // (and x, (cmov -1, y, cond)) => (and.cond x, y)
7345 SDValue CAND = formConditionalOp(N, DAG, true);
7353 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7354 static SDValue PerformORCombine(SDNode *N,
7355 TargetLowering::DAGCombinerInfo &DCI,
7356 const ARMSubtarget *Subtarget) {
7357 // Attempt to use immediate-form VORR
7358 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7359 DebugLoc dl = N->getDebugLoc();
7360 EVT VT = N->getValueType(0);
7361 SelectionDAG &DAG = DCI.DAG;
7363 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7366 APInt SplatBits, SplatUndef;
7367 unsigned SplatBitSize;
7369 if (BVN && Subtarget->hasNEON() &&
7370 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7371 if (SplatBitSize <= 64) {
7373 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7374 SplatUndef.getZExtValue(), SplatBitSize,
7375 DAG, VorrVT, VT.is128BitVector(),
7377 if (Val.getNode()) {
7379 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7380 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7381 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7386 if (!Subtarget->isThumb1Only()) {
7387 // (or x, (cmov 0, y, cond)) => (or.cond x, y)
7388 SDValue COR = formConditionalOp(N, DAG, true);
7393 SDValue N0 = N->getOperand(0);
7394 if (N0.getOpcode() != ISD::AND)
7396 SDValue N1 = N->getOperand(1);
7398 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7399 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7400 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7402 unsigned SplatBitSize;
7405 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7407 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7408 HasAnyUndefs) && !HasAnyUndefs) {
7409 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7411 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7412 HasAnyUndefs) && !HasAnyUndefs &&
7413 SplatBits0 == ~SplatBits1) {
7414 // Canonicalize the vector type to make instruction selection simpler.
7415 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7416 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7417 N0->getOperand(1), N0->getOperand(0),
7419 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7424 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7427 // BFI is only available on V6T2+
7428 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7431 DebugLoc DL = N->getDebugLoc();
7432 // 1) or (and A, mask), val => ARMbfi A, val, mask
7433 // iff (val & mask) == val
7435 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7436 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7437 // && mask == ~mask2
7438 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7439 // && ~mask == mask2
7440 // (i.e., copy a bitfield value into another bitfield of the same width)
7445 SDValue N00 = N0.getOperand(0);
7447 // The value and the mask need to be constants so we can verify this is
7448 // actually a bitfield set. If the mask is 0xffff, we can do better
7449 // via a movt instruction, so don't use BFI in that case.
7450 SDValue MaskOp = N0.getOperand(1);
7451 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7454 unsigned Mask = MaskC->getZExtValue();
7458 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7459 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7461 unsigned Val = N1C->getZExtValue();
7462 if ((Val & ~Mask) != Val)
7465 if (ARM::isBitFieldInvertedMask(Mask)) {
7466 Val >>= CountTrailingZeros_32(~Mask);
7468 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7469 DAG.getConstant(Val, MVT::i32),
7470 DAG.getConstant(Mask, MVT::i32));
7472 // Do not add new nodes to DAG combiner worklist.
7473 DCI.CombineTo(N, Res, false);
7476 } else if (N1.getOpcode() == ISD::AND) {
7477 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7478 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7481 unsigned Mask2 = N11C->getZExtValue();
7483 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7485 if (ARM::isBitFieldInvertedMask(Mask) &&
7487 // The pack halfword instruction works better for masks that fit it,
7488 // so use that when it's available.
7489 if (Subtarget->hasT2ExtractPack() &&
7490 (Mask == 0xffff || Mask == 0xffff0000))
7493 unsigned amt = CountTrailingZeros_32(Mask2);
7494 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7495 DAG.getConstant(amt, MVT::i32));
7496 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7497 DAG.getConstant(Mask, MVT::i32));
7498 // Do not add new nodes to DAG combiner worklist.
7499 DCI.CombineTo(N, Res, false);
7501 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7503 // The pack halfword instruction works better for masks that fit it,
7504 // so use that when it's available.
7505 if (Subtarget->hasT2ExtractPack() &&
7506 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7509 unsigned lsb = CountTrailingZeros_32(Mask);
7510 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7511 DAG.getConstant(lsb, MVT::i32));
7512 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7513 DAG.getConstant(Mask2, MVT::i32));
7514 // Do not add new nodes to DAG combiner worklist.
7515 DCI.CombineTo(N, Res, false);
7520 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7521 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7522 ARM::isBitFieldInvertedMask(~Mask)) {
7523 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7524 // where lsb(mask) == #shamt and masked bits of B are known zero.
7525 SDValue ShAmt = N00.getOperand(1);
7526 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7527 unsigned LSB = CountTrailingZeros_32(Mask);
7531 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7532 DAG.getConstant(~Mask, MVT::i32));
7534 // Do not add new nodes to DAG combiner worklist.
7535 DCI.CombineTo(N, Res, false);
7541 static SDValue PerformXORCombine(SDNode *N,
7542 TargetLowering::DAGCombinerInfo &DCI,
7543 const ARMSubtarget *Subtarget) {
7544 EVT VT = N->getValueType(0);
7545 SelectionDAG &DAG = DCI.DAG;
7547 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7550 if (!Subtarget->isThumb1Only()) {
7551 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7552 SDValue CXOR = formConditionalOp(N, DAG, true);
7560 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7561 /// the bits being cleared by the AND are not demanded by the BFI.
7562 static SDValue PerformBFICombine(SDNode *N,
7563 TargetLowering::DAGCombinerInfo &DCI) {
7564 SDValue N1 = N->getOperand(1);
7565 if (N1.getOpcode() == ISD::AND) {
7566 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7569 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7570 unsigned LSB = CountTrailingZeros_32(~InvMask);
7571 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7572 unsigned Mask = (1 << Width)-1;
7573 unsigned Mask2 = N11C->getZExtValue();
7574 if ((Mask & (~Mask2)) == 0)
7575 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7576 N->getOperand(0), N1.getOperand(0),
7582 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7583 /// ARMISD::VMOVRRD.
7584 static SDValue PerformVMOVRRDCombine(SDNode *N,
7585 TargetLowering::DAGCombinerInfo &DCI) {
7586 // vmovrrd(vmovdrr x, y) -> x,y
7587 SDValue InDouble = N->getOperand(0);
7588 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7589 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7591 // vmovrrd(load f64) -> (load i32), (load i32)
7592 SDNode *InNode = InDouble.getNode();
7593 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7594 InNode->getValueType(0) == MVT::f64 &&
7595 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7596 !cast<LoadSDNode>(InNode)->isVolatile()) {
7597 // TODO: Should this be done for non-FrameIndex operands?
7598 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7600 SelectionDAG &DAG = DCI.DAG;
7601 DebugLoc DL = LD->getDebugLoc();
7602 SDValue BasePtr = LD->getBasePtr();
7603 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7604 LD->getPointerInfo(), LD->isVolatile(),
7605 LD->isNonTemporal(), LD->isInvariant(),
7606 LD->getAlignment());
7608 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7609 DAG.getConstant(4, MVT::i32));
7610 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7611 LD->getPointerInfo(), LD->isVolatile(),
7612 LD->isNonTemporal(), LD->isInvariant(),
7613 std::min(4U, LD->getAlignment() / 2));
7615 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7616 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7617 DCI.RemoveFromWorklist(LD);
7625 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7626 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7627 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7628 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7629 SDValue Op0 = N->getOperand(0);
7630 SDValue Op1 = N->getOperand(1);
7631 if (Op0.getOpcode() == ISD::BITCAST)
7632 Op0 = Op0.getOperand(0);
7633 if (Op1.getOpcode() == ISD::BITCAST)
7634 Op1 = Op1.getOperand(0);
7635 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7636 Op0.getNode() == Op1.getNode() &&
7637 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7638 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7639 N->getValueType(0), Op0.getOperand(0));
7643 /// PerformSTORECombine - Target-specific dag combine xforms for
7645 static SDValue PerformSTORECombine(SDNode *N,
7646 TargetLowering::DAGCombinerInfo &DCI) {
7647 StoreSDNode *St = cast<StoreSDNode>(N);
7648 if (St->isVolatile())
7651 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7652 // pack all of the elements in one place. Next, store to memory in fewer
7654 SDValue StVal = St->getValue();
7655 EVT VT = StVal.getValueType();
7656 if (St->isTruncatingStore() && VT.isVector()) {
7657 SelectionDAG &DAG = DCI.DAG;
7658 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7659 EVT StVT = St->getMemoryVT();
7660 unsigned NumElems = VT.getVectorNumElements();
7661 assert(StVT != VT && "Cannot truncate to the same type");
7662 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7663 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7665 // From, To sizes and ElemCount must be pow of two
7666 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7668 // We are going to use the original vector elt for storing.
7669 // Accumulated smaller vector elements must be a multiple of the store size.
7670 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7672 unsigned SizeRatio = FromEltSz / ToEltSz;
7673 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7675 // Create a type on which we perform the shuffle.
7676 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7677 NumElems*SizeRatio);
7678 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7680 DebugLoc DL = St->getDebugLoc();
7681 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
7682 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
7683 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
7685 // Can't shuffle using an illegal type.
7686 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
7688 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
7689 DAG.getUNDEF(WideVec.getValueType()),
7691 // At this point all of the data is stored at the bottom of the
7692 // register. We now need to save it to mem.
7694 // Find the largest store unit
7695 MVT StoreType = MVT::i8;
7696 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
7697 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
7698 MVT Tp = (MVT::SimpleValueType)tp;
7699 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
7702 // Didn't find a legal store type.
7703 if (!TLI.isTypeLegal(StoreType))
7706 // Bitcast the original vector into a vector of store-size units
7707 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
7708 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
7709 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
7710 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
7711 SmallVector<SDValue, 8> Chains;
7712 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
7713 TLI.getPointerTy());
7714 SDValue BasePtr = St->getBasePtr();
7716 // Perform one or more big stores into memory.
7717 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
7718 for (unsigned I = 0; I < E; I++) {
7719 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
7720 StoreType, ShuffWide,
7721 DAG.getIntPtrConstant(I));
7722 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
7723 St->getPointerInfo(), St->isVolatile(),
7724 St->isNonTemporal(), St->getAlignment());
7725 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
7727 Chains.push_back(Ch);
7729 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
7733 if (!ISD::isNormalStore(St))
7736 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
7737 // ARM stores of arguments in the same cache line.
7738 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7739 StVal.getNode()->hasOneUse()) {
7740 SelectionDAG &DAG = DCI.DAG;
7741 DebugLoc DL = St->getDebugLoc();
7742 SDValue BasePtr = St->getBasePtr();
7743 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7744 StVal.getNode()->getOperand(0), BasePtr,
7745 St->getPointerInfo(), St->isVolatile(),
7746 St->isNonTemporal(), St->getAlignment());
7748 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7749 DAG.getConstant(4, MVT::i32));
7750 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7751 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7752 St->isNonTemporal(),
7753 std::min(4U, St->getAlignment() / 2));
7756 if (StVal.getValueType() != MVT::i64 ||
7757 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7760 // Bitcast an i64 store extracted from a vector to f64.
7761 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7762 SelectionDAG &DAG = DCI.DAG;
7763 DebugLoc dl = StVal.getDebugLoc();
7764 SDValue IntVec = StVal.getOperand(0);
7765 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7766 IntVec.getValueType().getVectorNumElements());
7767 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7768 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7769 Vec, StVal.getOperand(1));
7770 dl = N->getDebugLoc();
7771 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7772 // Make the DAGCombiner fold the bitcasts.
7773 DCI.AddToWorklist(Vec.getNode());
7774 DCI.AddToWorklist(ExtElt.getNode());
7775 DCI.AddToWorklist(V.getNode());
7776 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7777 St->getPointerInfo(), St->isVolatile(),
7778 St->isNonTemporal(), St->getAlignment(),
7782 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7783 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7784 /// i64 vector to have f64 elements, since the value can then be loaded
7785 /// directly into a VFP register.
7786 static bool hasNormalLoadOperand(SDNode *N) {
7787 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7788 for (unsigned i = 0; i < NumElts; ++i) {
7789 SDNode *Elt = N->getOperand(i).getNode();
7790 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7796 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7797 /// ISD::BUILD_VECTOR.
7798 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7799 TargetLowering::DAGCombinerInfo &DCI){
7800 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7801 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7802 // into a pair of GPRs, which is fine when the value is used as a scalar,
7803 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7804 SelectionDAG &DAG = DCI.DAG;
7805 if (N->getNumOperands() == 2) {
7806 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7811 // Load i64 elements as f64 values so that type legalization does not split
7812 // them up into i32 values.
7813 EVT VT = N->getValueType(0);
7814 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7816 DebugLoc dl = N->getDebugLoc();
7817 SmallVector<SDValue, 8> Ops;
7818 unsigned NumElts = VT.getVectorNumElements();
7819 for (unsigned i = 0; i < NumElts; ++i) {
7820 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7822 // Make the DAGCombiner fold the bitcast.
7823 DCI.AddToWorklist(V.getNode());
7825 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7826 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7827 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7830 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7831 /// ISD::INSERT_VECTOR_ELT.
7832 static SDValue PerformInsertEltCombine(SDNode *N,
7833 TargetLowering::DAGCombinerInfo &DCI) {
7834 // Bitcast an i64 load inserted into a vector to f64.
7835 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7836 EVT VT = N->getValueType(0);
7837 SDNode *Elt = N->getOperand(1).getNode();
7838 if (VT.getVectorElementType() != MVT::i64 ||
7839 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7842 SelectionDAG &DAG = DCI.DAG;
7843 DebugLoc dl = N->getDebugLoc();
7844 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7845 VT.getVectorNumElements());
7846 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7847 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7848 // Make the DAGCombiner fold the bitcasts.
7849 DCI.AddToWorklist(Vec.getNode());
7850 DCI.AddToWorklist(V.getNode());
7851 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7852 Vec, V, N->getOperand(2));
7853 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7856 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7857 /// ISD::VECTOR_SHUFFLE.
7858 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7859 // The LLVM shufflevector instruction does not require the shuffle mask
7860 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7861 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7862 // operands do not match the mask length, they are extended by concatenating
7863 // them with undef vectors. That is probably the right thing for other
7864 // targets, but for NEON it is better to concatenate two double-register
7865 // size vector operands into a single quad-register size vector. Do that
7866 // transformation here:
7867 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7868 // shuffle(concat(v1, v2), undef)
7869 SDValue Op0 = N->getOperand(0);
7870 SDValue Op1 = N->getOperand(1);
7871 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7872 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7873 Op0.getNumOperands() != 2 ||
7874 Op1.getNumOperands() != 2)
7876 SDValue Concat0Op1 = Op0.getOperand(1);
7877 SDValue Concat1Op1 = Op1.getOperand(1);
7878 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7879 Concat1Op1.getOpcode() != ISD::UNDEF)
7881 // Skip the transformation if any of the types are illegal.
7882 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7883 EVT VT = N->getValueType(0);
7884 if (!TLI.isTypeLegal(VT) ||
7885 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7886 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7889 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7890 Op0.getOperand(0), Op1.getOperand(0));
7891 // Translate the shuffle mask.
7892 SmallVector<int, 16> NewMask;
7893 unsigned NumElts = VT.getVectorNumElements();
7894 unsigned HalfElts = NumElts/2;
7895 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7896 for (unsigned n = 0; n < NumElts; ++n) {
7897 int MaskElt = SVN->getMaskElt(n);
7899 if (MaskElt < (int)HalfElts)
7901 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7902 NewElt = HalfElts + MaskElt - NumElts;
7903 NewMask.push_back(NewElt);
7905 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7906 DAG.getUNDEF(VT), NewMask.data());
7909 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7910 /// NEON load/store intrinsics to merge base address updates.
7911 static SDValue CombineBaseUpdate(SDNode *N,
7912 TargetLowering::DAGCombinerInfo &DCI) {
7913 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7916 SelectionDAG &DAG = DCI.DAG;
7917 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7918 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7919 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7920 SDValue Addr = N->getOperand(AddrOpIdx);
7922 // Search for a use of the address operand that is an increment.
7923 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7924 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7926 if (User->getOpcode() != ISD::ADD ||
7927 UI.getUse().getResNo() != Addr.getResNo())
7930 // Check that the add is independent of the load/store. Otherwise, folding
7931 // it would create a cycle.
7932 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7935 // Find the new opcode for the updating load/store.
7937 bool isLaneOp = false;
7938 unsigned NewOpc = 0;
7939 unsigned NumVecs = 0;
7941 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7943 default: llvm_unreachable("unexpected intrinsic for Neon base update");
7944 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7946 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7948 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7950 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7952 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7953 NumVecs = 2; isLaneOp = true; break;
7954 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7955 NumVecs = 3; isLaneOp = true; break;
7956 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7957 NumVecs = 4; isLaneOp = true; break;
7958 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7959 NumVecs = 1; isLoad = false; break;
7960 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7961 NumVecs = 2; isLoad = false; break;
7962 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7963 NumVecs = 3; isLoad = false; break;
7964 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7965 NumVecs = 4; isLoad = false; break;
7966 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7967 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7968 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7969 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7970 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7971 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7975 switch (N->getOpcode()) {
7976 default: llvm_unreachable("unexpected opcode for Neon base update");
7977 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7978 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7979 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7983 // Find the size of memory referenced by the load/store.
7986 VecTy = N->getValueType(0);
7988 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7989 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7991 NumBytes /= VecTy.getVectorNumElements();
7993 // If the increment is a constant, it must match the memory ref size.
7994 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7995 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7996 uint64_t IncVal = CInc->getZExtValue();
7997 if (IncVal != NumBytes)
7999 } else if (NumBytes >= 3 * 16) {
8000 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8001 // separate instructions that make it harder to use a non-constant update.
8005 // Create the new updating load/store node.
8007 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8009 for (n = 0; n < NumResultVecs; ++n)
8011 Tys[n++] = MVT::i32;
8012 Tys[n] = MVT::Other;
8013 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8014 SmallVector<SDValue, 8> Ops;
8015 Ops.push_back(N->getOperand(0)); // incoming chain
8016 Ops.push_back(N->getOperand(AddrOpIdx));
8018 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8019 Ops.push_back(N->getOperand(i));
8021 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8022 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8023 Ops.data(), Ops.size(),
8024 MemInt->getMemoryVT(),
8025 MemInt->getMemOperand());
8028 std::vector<SDValue> NewResults;
8029 for (unsigned i = 0; i < NumResultVecs; ++i) {
8030 NewResults.push_back(SDValue(UpdN.getNode(), i));
8032 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8033 DCI.CombineTo(N, NewResults);
8034 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8041 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8042 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8043 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8045 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8046 SelectionDAG &DAG = DCI.DAG;
8047 EVT VT = N->getValueType(0);
8048 // vldN-dup instructions only support 64-bit vectors for N > 1.
8049 if (!VT.is64BitVector())
8052 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8053 SDNode *VLD = N->getOperand(0).getNode();
8054 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8056 unsigned NumVecs = 0;
8057 unsigned NewOpc = 0;
8058 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8059 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8061 NewOpc = ARMISD::VLD2DUP;
8062 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8064 NewOpc = ARMISD::VLD3DUP;
8065 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8067 NewOpc = ARMISD::VLD4DUP;
8072 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8073 // numbers match the load.
8074 unsigned VLDLaneNo =
8075 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8076 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8078 // Ignore uses of the chain result.
8079 if (UI.getUse().getResNo() == NumVecs)
8082 if (User->getOpcode() != ARMISD::VDUPLANE ||
8083 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8087 // Create the vldN-dup node.
8090 for (n = 0; n < NumVecs; ++n)
8092 Tys[n] = MVT::Other;
8093 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8094 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8095 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8096 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8097 Ops, 2, VLDMemInt->getMemoryVT(),
8098 VLDMemInt->getMemOperand());
8101 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8103 unsigned ResNo = UI.getUse().getResNo();
8104 // Ignore uses of the chain result.
8105 if (ResNo == NumVecs)
8108 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8111 // Now the vldN-lane intrinsic is dead except for its chain result.
8112 // Update uses of the chain.
8113 std::vector<SDValue> VLDDupResults;
8114 for (unsigned n = 0; n < NumVecs; ++n)
8115 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8116 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8117 DCI.CombineTo(VLD, VLDDupResults);
8122 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8123 /// ARMISD::VDUPLANE.
8124 static SDValue PerformVDUPLANECombine(SDNode *N,
8125 TargetLowering::DAGCombinerInfo &DCI) {
8126 SDValue Op = N->getOperand(0);
8128 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8129 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8130 if (CombineVLDDUP(N, DCI))
8131 return SDValue(N, 0);
8133 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8134 // redundant. Ignore bit_converts for now; element sizes are checked below.
8135 while (Op.getOpcode() == ISD::BITCAST)
8136 Op = Op.getOperand(0);
8137 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8140 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8141 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8142 // The canonical VMOV for a zero vector uses a 32-bit element size.
8143 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8145 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8147 EVT VT = N->getValueType(0);
8148 if (EltSize > VT.getVectorElementType().getSizeInBits())
8151 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8154 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8155 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8156 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8160 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8162 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8167 APFloat APF = C->getValueAPF();
8168 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8169 != APFloat::opOK || !isExact)
8172 c0 = (I == 0) ? cN : c0;
8173 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8180 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8181 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8182 /// when the VMUL has a constant operand that is a power of 2.
8184 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8185 /// vmul.f32 d16, d17, d16
8186 /// vcvt.s32.f32 d16, d16
8188 /// vcvt.s32.f32 d16, d16, #3
8189 static SDValue PerformVCVTCombine(SDNode *N,
8190 TargetLowering::DAGCombinerInfo &DCI,
8191 const ARMSubtarget *Subtarget) {
8192 SelectionDAG &DAG = DCI.DAG;
8193 SDValue Op = N->getOperand(0);
8195 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8196 Op.getOpcode() != ISD::FMUL)
8200 SDValue N0 = Op->getOperand(0);
8201 SDValue ConstVec = Op->getOperand(1);
8202 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8204 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8205 !isConstVecPow2(ConstVec, isSigned, C))
8208 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8209 Intrinsic::arm_neon_vcvtfp2fxu;
8210 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8212 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8213 DAG.getConstant(Log2_64(C), MVT::i32));
8216 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8217 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8218 /// when the VDIV has a constant operand that is a power of 2.
8220 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8221 /// vcvt.f32.s32 d16, d16
8222 /// vdiv.f32 d16, d17, d16
8224 /// vcvt.f32.s32 d16, d16, #3
8225 static SDValue PerformVDIVCombine(SDNode *N,
8226 TargetLowering::DAGCombinerInfo &DCI,
8227 const ARMSubtarget *Subtarget) {
8228 SelectionDAG &DAG = DCI.DAG;
8229 SDValue Op = N->getOperand(0);
8230 unsigned OpOpcode = Op.getNode()->getOpcode();
8232 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8233 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8237 SDValue ConstVec = N->getOperand(1);
8238 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8240 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8241 !isConstVecPow2(ConstVec, isSigned, C))
8244 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8245 Intrinsic::arm_neon_vcvtfxu2fp;
8246 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8248 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8249 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8252 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8253 /// operand of a vector shift operation, where all the elements of the
8254 /// build_vector must have the same constant integer value.
8255 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8256 // Ignore bit_converts.
8257 while (Op.getOpcode() == ISD::BITCAST)
8258 Op = Op.getOperand(0);
8259 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8260 APInt SplatBits, SplatUndef;
8261 unsigned SplatBitSize;
8263 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8264 HasAnyUndefs, ElementBits) ||
8265 SplatBitSize > ElementBits)
8267 Cnt = SplatBits.getSExtValue();
8271 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8272 /// operand of a vector shift left operation. That value must be in the range:
8273 /// 0 <= Value < ElementBits for a left shift; or
8274 /// 0 <= Value <= ElementBits for a long left shift.
8275 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8276 assert(VT.isVector() && "vector shift count is not a vector type");
8277 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8278 if (! getVShiftImm(Op, ElementBits, Cnt))
8280 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8283 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8284 /// operand of a vector shift right operation. For a shift opcode, the value
8285 /// is positive, but for an intrinsic the value count must be negative. The
8286 /// absolute value must be in the range:
8287 /// 1 <= |Value| <= ElementBits for a right shift; or
8288 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8289 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8291 assert(VT.isVector() && "vector shift count is not a vector type");
8292 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8293 if (! getVShiftImm(Op, ElementBits, Cnt))
8297 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8300 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8301 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8302 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8305 // Don't do anything for most intrinsics.
8308 // Vector shifts: check for immediate versions and lower them.
8309 // Note: This is done during DAG combining instead of DAG legalizing because
8310 // the build_vectors for 64-bit vector element shift counts are generally
8311 // not legal, and it is hard to see their values after they get legalized to
8312 // loads from a constant pool.
8313 case Intrinsic::arm_neon_vshifts:
8314 case Intrinsic::arm_neon_vshiftu:
8315 case Intrinsic::arm_neon_vshiftls:
8316 case Intrinsic::arm_neon_vshiftlu:
8317 case Intrinsic::arm_neon_vshiftn:
8318 case Intrinsic::arm_neon_vrshifts:
8319 case Intrinsic::arm_neon_vrshiftu:
8320 case Intrinsic::arm_neon_vrshiftn:
8321 case Intrinsic::arm_neon_vqshifts:
8322 case Intrinsic::arm_neon_vqshiftu:
8323 case Intrinsic::arm_neon_vqshiftsu:
8324 case Intrinsic::arm_neon_vqshiftns:
8325 case Intrinsic::arm_neon_vqshiftnu:
8326 case Intrinsic::arm_neon_vqshiftnsu:
8327 case Intrinsic::arm_neon_vqrshiftns:
8328 case Intrinsic::arm_neon_vqrshiftnu:
8329 case Intrinsic::arm_neon_vqrshiftnsu: {
8330 EVT VT = N->getOperand(1).getValueType();
8332 unsigned VShiftOpc = 0;
8335 case Intrinsic::arm_neon_vshifts:
8336 case Intrinsic::arm_neon_vshiftu:
8337 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8338 VShiftOpc = ARMISD::VSHL;
8341 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8342 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8343 ARMISD::VSHRs : ARMISD::VSHRu);
8348 case Intrinsic::arm_neon_vshiftls:
8349 case Intrinsic::arm_neon_vshiftlu:
8350 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8352 llvm_unreachable("invalid shift count for vshll intrinsic");
8354 case Intrinsic::arm_neon_vrshifts:
8355 case Intrinsic::arm_neon_vrshiftu:
8356 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8360 case Intrinsic::arm_neon_vqshifts:
8361 case Intrinsic::arm_neon_vqshiftu:
8362 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8366 case Intrinsic::arm_neon_vqshiftsu:
8367 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8369 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8371 case Intrinsic::arm_neon_vshiftn:
8372 case Intrinsic::arm_neon_vrshiftn:
8373 case Intrinsic::arm_neon_vqshiftns:
8374 case Intrinsic::arm_neon_vqshiftnu:
8375 case Intrinsic::arm_neon_vqshiftnsu:
8376 case Intrinsic::arm_neon_vqrshiftns:
8377 case Intrinsic::arm_neon_vqrshiftnu:
8378 case Intrinsic::arm_neon_vqrshiftnsu:
8379 // Narrowing shifts require an immediate right shift.
8380 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8382 llvm_unreachable("invalid shift count for narrowing vector shift "
8386 llvm_unreachable("unhandled vector shift");
8390 case Intrinsic::arm_neon_vshifts:
8391 case Intrinsic::arm_neon_vshiftu:
8392 // Opcode already set above.
8394 case Intrinsic::arm_neon_vshiftls:
8395 case Intrinsic::arm_neon_vshiftlu:
8396 if (Cnt == VT.getVectorElementType().getSizeInBits())
8397 VShiftOpc = ARMISD::VSHLLi;
8399 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8400 ARMISD::VSHLLs : ARMISD::VSHLLu);
8402 case Intrinsic::arm_neon_vshiftn:
8403 VShiftOpc = ARMISD::VSHRN; break;
8404 case Intrinsic::arm_neon_vrshifts:
8405 VShiftOpc = ARMISD::VRSHRs; break;
8406 case Intrinsic::arm_neon_vrshiftu:
8407 VShiftOpc = ARMISD::VRSHRu; break;
8408 case Intrinsic::arm_neon_vrshiftn:
8409 VShiftOpc = ARMISD::VRSHRN; break;
8410 case Intrinsic::arm_neon_vqshifts:
8411 VShiftOpc = ARMISD::VQSHLs; break;
8412 case Intrinsic::arm_neon_vqshiftu:
8413 VShiftOpc = ARMISD::VQSHLu; break;
8414 case Intrinsic::arm_neon_vqshiftsu:
8415 VShiftOpc = ARMISD::VQSHLsu; break;
8416 case Intrinsic::arm_neon_vqshiftns:
8417 VShiftOpc = ARMISD::VQSHRNs; break;
8418 case Intrinsic::arm_neon_vqshiftnu:
8419 VShiftOpc = ARMISD::VQSHRNu; break;
8420 case Intrinsic::arm_neon_vqshiftnsu:
8421 VShiftOpc = ARMISD::VQSHRNsu; break;
8422 case Intrinsic::arm_neon_vqrshiftns:
8423 VShiftOpc = ARMISD::VQRSHRNs; break;
8424 case Intrinsic::arm_neon_vqrshiftnu:
8425 VShiftOpc = ARMISD::VQRSHRNu; break;
8426 case Intrinsic::arm_neon_vqrshiftnsu:
8427 VShiftOpc = ARMISD::VQRSHRNsu; break;
8430 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8431 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8434 case Intrinsic::arm_neon_vshiftins: {
8435 EVT VT = N->getOperand(1).getValueType();
8437 unsigned VShiftOpc = 0;
8439 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8440 VShiftOpc = ARMISD::VSLI;
8441 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8442 VShiftOpc = ARMISD::VSRI;
8444 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8447 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8448 N->getOperand(1), N->getOperand(2),
8449 DAG.getConstant(Cnt, MVT::i32));
8452 case Intrinsic::arm_neon_vqrshifts:
8453 case Intrinsic::arm_neon_vqrshiftu:
8454 // No immediate versions of these to check for.
8461 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8462 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8463 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8464 /// vector element shift counts are generally not legal, and it is hard to see
8465 /// their values after they get legalized to loads from a constant pool.
8466 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8467 const ARMSubtarget *ST) {
8468 EVT VT = N->getValueType(0);
8469 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8470 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8471 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8472 SDValue N1 = N->getOperand(1);
8473 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8474 SDValue N0 = N->getOperand(0);
8475 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8476 DAG.MaskedValueIsZero(N0.getOperand(0),
8477 APInt::getHighBitsSet(32, 16)))
8478 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8482 // Nothing to be done for scalar shifts.
8483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8484 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8487 assert(ST->hasNEON() && "unexpected vector shift");
8490 switch (N->getOpcode()) {
8491 default: llvm_unreachable("unexpected shift opcode");
8494 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8495 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8496 DAG.getConstant(Cnt, MVT::i32));
8501 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8502 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8503 ARMISD::VSHRs : ARMISD::VSHRu);
8504 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8505 DAG.getConstant(Cnt, MVT::i32));
8511 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8512 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8513 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8514 const ARMSubtarget *ST) {
8515 SDValue N0 = N->getOperand(0);
8517 // Check for sign- and zero-extensions of vector extract operations of 8-
8518 // and 16-bit vector elements. NEON supports these directly. They are
8519 // handled during DAG combining because type legalization will promote them
8520 // to 32-bit types and it is messy to recognize the operations after that.
8521 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8522 SDValue Vec = N0.getOperand(0);
8523 SDValue Lane = N0.getOperand(1);
8524 EVT VT = N->getValueType(0);
8525 EVT EltVT = N0.getValueType();
8526 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8528 if (VT == MVT::i32 &&
8529 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8530 TLI.isTypeLegal(Vec.getValueType()) &&
8531 isa<ConstantSDNode>(Lane)) {
8534 switch (N->getOpcode()) {
8535 default: llvm_unreachable("unexpected opcode");
8536 case ISD::SIGN_EXTEND:
8537 Opc = ARMISD::VGETLANEs;
8539 case ISD::ZERO_EXTEND:
8540 case ISD::ANY_EXTEND:
8541 Opc = ARMISD::VGETLANEu;
8544 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8551 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8552 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8553 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8554 const ARMSubtarget *ST) {
8555 // If the target supports NEON, try to use vmax/vmin instructions for f32
8556 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8557 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8558 // a NaN; only do the transformation when it matches that behavior.
8560 // For now only do this when using NEON for FP operations; if using VFP, it
8561 // is not obvious that the benefit outweighs the cost of switching to the
8563 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8564 N->getValueType(0) != MVT::f32)
8567 SDValue CondLHS = N->getOperand(0);
8568 SDValue CondRHS = N->getOperand(1);
8569 SDValue LHS = N->getOperand(2);
8570 SDValue RHS = N->getOperand(3);
8571 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8573 unsigned Opcode = 0;
8575 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8576 IsReversed = false; // x CC y ? x : y
8577 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8578 IsReversed = true ; // x CC y ? y : x
8592 // If LHS is NaN, an ordered comparison will be false and the result will
8593 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8594 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8595 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8596 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8598 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8599 // will return -0, so vmin can only be used for unsafe math or if one of
8600 // the operands is known to be nonzero.
8601 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8602 !DAG.getTarget().Options.UnsafeFPMath &&
8603 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8605 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8614 // If LHS is NaN, an ordered comparison will be false and the result will
8615 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8616 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8617 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8618 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8620 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8621 // will return +0, so vmax can only be used for unsafe math or if one of
8622 // the operands is known to be nonzero.
8623 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8624 !DAG.getTarget().Options.UnsafeFPMath &&
8625 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8627 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8633 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8636 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8638 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8639 SDValue Cmp = N->getOperand(4);
8640 if (Cmp.getOpcode() != ARMISD::CMPZ)
8641 // Only looking at EQ and NE cases.
8644 EVT VT = N->getValueType(0);
8645 DebugLoc dl = N->getDebugLoc();
8646 SDValue LHS = Cmp.getOperand(0);
8647 SDValue RHS = Cmp.getOperand(1);
8648 SDValue FalseVal = N->getOperand(0);
8649 SDValue TrueVal = N->getOperand(1);
8650 SDValue ARMcc = N->getOperand(2);
8651 ARMCC::CondCodes CC =
8652 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8670 /// FIXME: Turn this into a target neutral optimization?
8672 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8673 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8674 N->getOperand(3), Cmp);
8675 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8677 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8678 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8679 N->getOperand(3), NewCmp);
8682 if (Res.getNode()) {
8683 APInt KnownZero, KnownOne;
8684 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8685 // Capture demanded bits information that would be otherwise lost.
8686 if (KnownZero == 0xfffffffe)
8687 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8688 DAG.getValueType(MVT::i1));
8689 else if (KnownZero == 0xffffff00)
8690 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8691 DAG.getValueType(MVT::i8));
8692 else if (KnownZero == 0xffff0000)
8693 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8694 DAG.getValueType(MVT::i16));
8700 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8701 DAGCombinerInfo &DCI) const {
8702 switch (N->getOpcode()) {
8704 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8705 case ISD::SUB: return PerformSUBCombine(N, DCI);
8706 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8707 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8708 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8709 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8710 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8711 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8712 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8713 case ISD::STORE: return PerformSTORECombine(N, DCI);
8714 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8715 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8716 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8717 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8718 case ISD::FP_TO_SINT:
8719 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8720 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8721 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8724 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8725 case ISD::SIGN_EXTEND:
8726 case ISD::ZERO_EXTEND:
8727 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8728 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8729 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8730 case ARMISD::VLD2DUP:
8731 case ARMISD::VLD3DUP:
8732 case ARMISD::VLD4DUP:
8733 return CombineBaseUpdate(N, DCI);
8734 case ISD::INTRINSIC_VOID:
8735 case ISD::INTRINSIC_W_CHAIN:
8736 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8737 case Intrinsic::arm_neon_vld1:
8738 case Intrinsic::arm_neon_vld2:
8739 case Intrinsic::arm_neon_vld3:
8740 case Intrinsic::arm_neon_vld4:
8741 case Intrinsic::arm_neon_vld2lane:
8742 case Intrinsic::arm_neon_vld3lane:
8743 case Intrinsic::arm_neon_vld4lane:
8744 case Intrinsic::arm_neon_vst1:
8745 case Intrinsic::arm_neon_vst2:
8746 case Intrinsic::arm_neon_vst3:
8747 case Intrinsic::arm_neon_vst4:
8748 case Intrinsic::arm_neon_vst2lane:
8749 case Intrinsic::arm_neon_vst3lane:
8750 case Intrinsic::arm_neon_vst4lane:
8751 return CombineBaseUpdate(N, DCI);
8759 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8761 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8764 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8765 if (!Subtarget->allowsUnalignedMem())
8768 switch (VT.getSimpleVT().SimpleTy) {
8775 // FIXME: VLD1 etc with standard alignment is legal.
8779 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8780 unsigned AlignCheck) {
8781 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8782 (DstAlign == 0 || DstAlign % AlignCheck == 0));
8785 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8786 unsigned DstAlign, unsigned SrcAlign,
8789 MachineFunction &MF) const {
8790 const Function *F = MF.getFunction();
8792 // See if we can use NEON instructions for this...
8794 !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8795 Subtarget->hasNEON()) {
8796 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8798 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8803 // Lowering to i32/i16 if the size permits.
8806 } else if (Size >= 2) {
8810 // Let the target-independent logic figure it out.
8814 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8819 switch (VT.getSimpleVT().SimpleTy) {
8820 default: return false;
8835 if ((V & (Scale - 1)) != 0)
8838 return V == (V & ((1LL << 5) - 1));
8841 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8842 const ARMSubtarget *Subtarget) {
8849 switch (VT.getSimpleVT().SimpleTy) {
8850 default: return false;
8855 // + imm12 or - imm8
8857 return V == (V & ((1LL << 8) - 1));
8858 return V == (V & ((1LL << 12) - 1));
8861 // Same as ARM mode. FIXME: NEON?
8862 if (!Subtarget->hasVFP2())
8867 return V == (V & ((1LL << 8) - 1));
8871 /// isLegalAddressImmediate - Return true if the integer value can be used
8872 /// as the offset of the target addressing mode for load / store of the
8874 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8875 const ARMSubtarget *Subtarget) {
8882 if (Subtarget->isThumb1Only())
8883 return isLegalT1AddressImmediate(V, VT);
8884 else if (Subtarget->isThumb2())
8885 return isLegalT2AddressImmediate(V, VT, Subtarget);
8890 switch (VT.getSimpleVT().SimpleTy) {
8891 default: return false;
8896 return V == (V & ((1LL << 12) - 1));
8899 return V == (V & ((1LL << 8) - 1));
8902 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8907 return V == (V & ((1LL << 8) - 1));
8911 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8913 int Scale = AM.Scale;
8917 switch (VT.getSimpleVT().SimpleTy) {
8918 default: return false;
8927 return Scale == 2 || Scale == 4 || Scale == 8;
8930 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8934 // Note, we allow "void" uses (basically, uses that aren't loads or
8935 // stores), because arm allows folding a scale into many arithmetic
8936 // operations. This should be made more precise and revisited later.
8938 // Allow r << imm, but the imm has to be a multiple of two.
8939 if (Scale & 1) return false;
8940 return isPowerOf2_32(Scale);
8944 /// isLegalAddressingMode - Return true if the addressing mode represented
8945 /// by AM is legal for this target, for a load/store of the specified type.
8946 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8948 EVT VT = getValueType(Ty, true);
8949 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8952 // Can never fold addr of global into load/store.
8957 case 0: // no scale reg, must be "r+i" or "r", or "i".
8960 if (Subtarget->isThumb1Only())
8964 // ARM doesn't support any R+R*scale+imm addr modes.
8971 if (Subtarget->isThumb2())
8972 return isLegalT2ScaledAddressingMode(AM, VT);
8974 int Scale = AM.Scale;
8975 switch (VT.getSimpleVT().SimpleTy) {
8976 default: return false;
8980 if (Scale < 0) Scale = -Scale;
8984 return isPowerOf2_32(Scale & ~1);
8988 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8993 // Note, we allow "void" uses (basically, uses that aren't loads or
8994 // stores), because arm allows folding a scale into many arithmetic
8995 // operations. This should be made more precise and revisited later.
8997 // Allow r << imm, but the imm has to be a multiple of two.
8998 if (Scale & 1) return false;
8999 return isPowerOf2_32(Scale);
9005 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9006 /// icmp immediate, that is the target has icmp instructions which can compare
9007 /// a register against the immediate without having to materialize the
9008 /// immediate into a register.
9009 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9010 // Thumb2 and ARM modes can use cmn for negative immediates.
9011 if (!Subtarget->isThumb())
9012 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9013 if (Subtarget->isThumb2())
9014 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9015 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9016 return Imm >= 0 && Imm <= 255;
9019 /// isLegalAddImmediate - Return true if the specified immediate is legal
9020 /// add immediate, that is the target has add instructions which can add
9021 /// a register with the immediate without having to materialize the
9022 /// immediate into a register.
9023 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9024 return ARM_AM::getSOImmVal(Imm) != -1;
9027 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9028 bool isSEXTLoad, SDValue &Base,
9029 SDValue &Offset, bool &isInc,
9030 SelectionDAG &DAG) {
9031 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9034 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9036 Base = Ptr->getOperand(0);
9037 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9038 int RHSC = (int)RHS->getZExtValue();
9039 if (RHSC < 0 && RHSC > -256) {
9040 assert(Ptr->getOpcode() == ISD::ADD);
9042 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9046 isInc = (Ptr->getOpcode() == ISD::ADD);
9047 Offset = Ptr->getOperand(1);
9049 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9051 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9052 int RHSC = (int)RHS->getZExtValue();
9053 if (RHSC < 0 && RHSC > -0x1000) {
9054 assert(Ptr->getOpcode() == ISD::ADD);
9056 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9057 Base = Ptr->getOperand(0);
9062 if (Ptr->getOpcode() == ISD::ADD) {
9064 ARM_AM::ShiftOpc ShOpcVal=
9065 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9066 if (ShOpcVal != ARM_AM::no_shift) {
9067 Base = Ptr->getOperand(1);
9068 Offset = Ptr->getOperand(0);
9070 Base = Ptr->getOperand(0);
9071 Offset = Ptr->getOperand(1);
9076 isInc = (Ptr->getOpcode() == ISD::ADD);
9077 Base = Ptr->getOperand(0);
9078 Offset = Ptr->getOperand(1);
9082 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9086 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9087 bool isSEXTLoad, SDValue &Base,
9088 SDValue &Offset, bool &isInc,
9089 SelectionDAG &DAG) {
9090 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9093 Base = Ptr->getOperand(0);
9094 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9095 int RHSC = (int)RHS->getZExtValue();
9096 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9097 assert(Ptr->getOpcode() == ISD::ADD);
9099 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9101 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9102 isInc = Ptr->getOpcode() == ISD::ADD;
9103 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9111 /// getPreIndexedAddressParts - returns true by value, base pointer and
9112 /// offset pointer and addressing mode by reference if the node's address
9113 /// can be legally represented as pre-indexed load / store address.
9115 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9117 ISD::MemIndexedMode &AM,
9118 SelectionDAG &DAG) const {
9119 if (Subtarget->isThumb1Only())
9124 bool isSEXTLoad = false;
9125 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9126 Ptr = LD->getBasePtr();
9127 VT = LD->getMemoryVT();
9128 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9129 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9130 Ptr = ST->getBasePtr();
9131 VT = ST->getMemoryVT();
9136 bool isLegal = false;
9137 if (Subtarget->isThumb2())
9138 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9139 Offset, isInc, DAG);
9141 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9142 Offset, isInc, DAG);
9146 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9150 /// getPostIndexedAddressParts - returns true by value, base pointer and
9151 /// offset pointer and addressing mode by reference if this node can be
9152 /// combined with a load / store to form a post-indexed load / store.
9153 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9156 ISD::MemIndexedMode &AM,
9157 SelectionDAG &DAG) const {
9158 if (Subtarget->isThumb1Only())
9163 bool isSEXTLoad = false;
9164 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9165 VT = LD->getMemoryVT();
9166 Ptr = LD->getBasePtr();
9167 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9168 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9169 VT = ST->getMemoryVT();
9170 Ptr = ST->getBasePtr();
9175 bool isLegal = false;
9176 if (Subtarget->isThumb2())
9177 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9180 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9186 // Swap base ptr and offset to catch more post-index load / store when
9187 // it's legal. In Thumb2 mode, offset must be an immediate.
9188 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9189 !Subtarget->isThumb2())
9190 std::swap(Base, Offset);
9192 // Post-indexed load / store update the base pointer.
9197 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9201 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9204 const SelectionDAG &DAG,
9205 unsigned Depth) const {
9206 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9207 switch (Op.getOpcode()) {
9209 case ARMISD::CMOV: {
9210 // Bits are known zero/one if known on the LHS and RHS.
9211 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9212 if (KnownZero == 0 && KnownOne == 0) return;
9214 APInt KnownZeroRHS, KnownOneRHS;
9215 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9216 KnownZero &= KnownZeroRHS;
9217 KnownOne &= KnownOneRHS;
9223 //===----------------------------------------------------------------------===//
9224 // ARM Inline Assembly Support
9225 //===----------------------------------------------------------------------===//
9227 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9228 // Looking for "rev" which is V6+.
9229 if (!Subtarget->hasV6Ops())
9232 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9233 std::string AsmStr = IA->getAsmString();
9234 SmallVector<StringRef, 4> AsmPieces;
9235 SplitString(AsmStr, AsmPieces, ";\n");
9237 switch (AsmPieces.size()) {
9238 default: return false;
9240 AsmStr = AsmPieces[0];
9242 SplitString(AsmStr, AsmPieces, " \t,");
9245 if (AsmPieces.size() == 3 &&
9246 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9247 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9248 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9249 if (Ty && Ty->getBitWidth() == 32)
9250 return IntrinsicLowering::LowerToByteSwap(CI);
9258 /// getConstraintType - Given a constraint letter, return the type of
9259 /// constraint it is for this target.
9260 ARMTargetLowering::ConstraintType
9261 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9262 if (Constraint.size() == 1) {
9263 switch (Constraint[0]) {
9265 case 'l': return C_RegisterClass;
9266 case 'w': return C_RegisterClass;
9267 case 'h': return C_RegisterClass;
9268 case 'x': return C_RegisterClass;
9269 case 't': return C_RegisterClass;
9270 case 'j': return C_Other; // Constant for movw.
9271 // An address with a single base register. Due to the way we
9272 // currently handle addresses it is the same as an 'r' memory constraint.
9273 case 'Q': return C_Memory;
9275 } else if (Constraint.size() == 2) {
9276 switch (Constraint[0]) {
9278 // All 'U+' constraints are addresses.
9279 case 'U': return C_Memory;
9282 return TargetLowering::getConstraintType(Constraint);
9285 /// Examine constraint type and operand type and determine a weight value.
9286 /// This object must already have been set up with the operand type
9287 /// and the current alternative constraint selected.
9288 TargetLowering::ConstraintWeight
9289 ARMTargetLowering::getSingleConstraintMatchWeight(
9290 AsmOperandInfo &info, const char *constraint) const {
9291 ConstraintWeight weight = CW_Invalid;
9292 Value *CallOperandVal = info.CallOperandVal;
9293 // If we don't have a value, we can't do a match,
9294 // but allow it at the lowest weight.
9295 if (CallOperandVal == NULL)
9297 Type *type = CallOperandVal->getType();
9298 // Look at the constraint type.
9299 switch (*constraint) {
9301 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9304 if (type->isIntegerTy()) {
9305 if (Subtarget->isThumb())
9306 weight = CW_SpecificReg;
9308 weight = CW_Register;
9312 if (type->isFloatingPointTy())
9313 weight = CW_Register;
9319 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9321 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9323 if (Constraint.size() == 1) {
9324 // GCC ARM Constraint Letters
9325 switch (Constraint[0]) {
9326 case 'l': // Low regs or general regs.
9327 if (Subtarget->isThumb())
9328 return RCPair(0U, &ARM::tGPRRegClass);
9329 return RCPair(0U, &ARM::GPRRegClass);
9330 case 'h': // High regs or no regs.
9331 if (Subtarget->isThumb())
9332 return RCPair(0U, &ARM::hGPRRegClass);
9335 return RCPair(0U, &ARM::GPRRegClass);
9338 return RCPair(0U, &ARM::SPRRegClass);
9339 if (VT.getSizeInBits() == 64)
9340 return RCPair(0U, &ARM::DPRRegClass);
9341 if (VT.getSizeInBits() == 128)
9342 return RCPair(0U, &ARM::QPRRegClass);
9346 return RCPair(0U, &ARM::SPR_8RegClass);
9347 if (VT.getSizeInBits() == 64)
9348 return RCPair(0U, &ARM::DPR_8RegClass);
9349 if (VT.getSizeInBits() == 128)
9350 return RCPair(0U, &ARM::QPR_8RegClass);
9354 return RCPair(0U, &ARM::SPRRegClass);
9358 if (StringRef("{cc}").equals_lower(Constraint))
9359 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9361 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9364 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9365 /// vector. If it is invalid, don't add anything to Ops.
9366 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9367 std::string &Constraint,
9368 std::vector<SDValue>&Ops,
9369 SelectionDAG &DAG) const {
9370 SDValue Result(0, 0);
9372 // Currently only support length 1 constraints.
9373 if (Constraint.length() != 1) return;
9375 char ConstraintLetter = Constraint[0];
9376 switch (ConstraintLetter) {
9379 case 'I': case 'J': case 'K': case 'L':
9380 case 'M': case 'N': case 'O':
9381 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9385 int64_t CVal64 = C->getSExtValue();
9386 int CVal = (int) CVal64;
9387 // None of these constraints allow values larger than 32 bits. Check
9388 // that the value fits in an int.
9392 switch (ConstraintLetter) {
9394 // Constant suitable for movw, must be between 0 and
9396 if (Subtarget->hasV6T2Ops())
9397 if (CVal >= 0 && CVal <= 65535)
9401 if (Subtarget->isThumb1Only()) {
9402 // This must be a constant between 0 and 255, for ADD
9404 if (CVal >= 0 && CVal <= 255)
9406 } else if (Subtarget->isThumb2()) {
9407 // A constant that can be used as an immediate value in a
9408 // data-processing instruction.
9409 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9412 // A constant that can be used as an immediate value in a
9413 // data-processing instruction.
9414 if (ARM_AM::getSOImmVal(CVal) != -1)
9420 if (Subtarget->isThumb()) { // FIXME thumb2
9421 // This must be a constant between -255 and -1, for negated ADD
9422 // immediates. This can be used in GCC with an "n" modifier that
9423 // prints the negated value, for use with SUB instructions. It is
9424 // not useful otherwise but is implemented for compatibility.
9425 if (CVal >= -255 && CVal <= -1)
9428 // This must be a constant between -4095 and 4095. It is not clear
9429 // what this constraint is intended for. Implemented for
9430 // compatibility with GCC.
9431 if (CVal >= -4095 && CVal <= 4095)
9437 if (Subtarget->isThumb1Only()) {
9438 // A 32-bit value where only one byte has a nonzero value. Exclude
9439 // zero to match GCC. This constraint is used by GCC internally for
9440 // constants that can be loaded with a move/shift combination.
9441 // It is not useful otherwise but is implemented for compatibility.
9442 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9444 } else if (Subtarget->isThumb2()) {
9445 // A constant whose bitwise inverse can be used as an immediate
9446 // value in a data-processing instruction. This can be used in GCC
9447 // with a "B" modifier that prints the inverted value, for use with
9448 // BIC and MVN instructions. It is not useful otherwise but is
9449 // implemented for compatibility.
9450 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9453 // A constant whose bitwise inverse can be used as an immediate
9454 // value in a data-processing instruction. This can be used in GCC
9455 // with a "B" modifier that prints the inverted value, for use with
9456 // BIC and MVN instructions. It is not useful otherwise but is
9457 // implemented for compatibility.
9458 if (ARM_AM::getSOImmVal(~CVal) != -1)
9464 if (Subtarget->isThumb1Only()) {
9465 // This must be a constant between -7 and 7,
9466 // for 3-operand ADD/SUB immediate instructions.
9467 if (CVal >= -7 && CVal < 7)
9469 } else if (Subtarget->isThumb2()) {
9470 // A constant whose negation can be used as an immediate value in a
9471 // data-processing instruction. This can be used in GCC with an "n"
9472 // modifier that prints the negated value, for use with SUB
9473 // instructions. It is not useful otherwise but is implemented for
9475 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9478 // A constant whose negation can be used as an immediate value in a
9479 // data-processing instruction. This can be used in GCC with an "n"
9480 // modifier that prints the negated value, for use with SUB
9481 // instructions. It is not useful otherwise but is implemented for
9483 if (ARM_AM::getSOImmVal(-CVal) != -1)
9489 if (Subtarget->isThumb()) { // FIXME thumb2
9490 // This must be a multiple of 4 between 0 and 1020, for
9491 // ADD sp + immediate.
9492 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9495 // A power of two or a constant between 0 and 32. This is used in
9496 // GCC for the shift amount on shifted register operands, but it is
9497 // useful in general for any shift amounts.
9498 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9504 if (Subtarget->isThumb()) { // FIXME thumb2
9505 // This must be a constant between 0 and 31, for shift amounts.
9506 if (CVal >= 0 && CVal <= 31)
9512 if (Subtarget->isThumb()) { // FIXME thumb2
9513 // This must be a multiple of 4 between -508 and 508, for
9514 // ADD/SUB sp = sp + immediate.
9515 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9520 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9524 if (Result.getNode()) {
9525 Ops.push_back(Result);
9528 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9532 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9533 // The ARM target isn't yet aware of offsets.
9537 bool ARM::isBitFieldInvertedMask(unsigned v) {
9538 if (v == 0xffffffff)
9540 // there can be 1's on either or both "outsides", all the "inside"
9542 unsigned int lsb = 0, msb = 31;
9543 while (v & (1 << msb)) --msb;
9544 while (v & (1 << lsb)) ++lsb;
9545 for (unsigned int i = lsb; i <= msb; ++i) {
9552 /// isFPImmLegal - Returns true if the target can instruction select the
9553 /// specified FP immediate natively. If false, the legalizer will
9554 /// materialize the FP immediate as a load from a constant pool.
9555 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9556 if (!Subtarget->hasVFP3())
9559 return ARM_AM::getFP32Imm(Imm) != -1;
9561 return ARM_AM::getFP64Imm(Imm) != -1;
9565 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9566 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9567 /// specified in the intrinsic calls.
9568 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9570 unsigned Intrinsic) const {
9571 switch (Intrinsic) {
9572 case Intrinsic::arm_neon_vld1:
9573 case Intrinsic::arm_neon_vld2:
9574 case Intrinsic::arm_neon_vld3:
9575 case Intrinsic::arm_neon_vld4:
9576 case Intrinsic::arm_neon_vld2lane:
9577 case Intrinsic::arm_neon_vld3lane:
9578 case Intrinsic::arm_neon_vld4lane: {
9579 Info.opc = ISD::INTRINSIC_W_CHAIN;
9580 // Conservatively set memVT to the entire set of vectors loaded.
9581 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9582 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9583 Info.ptrVal = I.getArgOperand(0);
9585 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9586 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9587 Info.vol = false; // volatile loads with NEON intrinsics not supported
9588 Info.readMem = true;
9589 Info.writeMem = false;
9592 case Intrinsic::arm_neon_vst1:
9593 case Intrinsic::arm_neon_vst2:
9594 case Intrinsic::arm_neon_vst3:
9595 case Intrinsic::arm_neon_vst4:
9596 case Intrinsic::arm_neon_vst2lane:
9597 case Intrinsic::arm_neon_vst3lane:
9598 case Intrinsic::arm_neon_vst4lane: {
9599 Info.opc = ISD::INTRINSIC_VOID;
9600 // Conservatively set memVT to the entire set of vectors stored.
9601 unsigned NumElts = 0;
9602 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9603 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9604 if (!ArgTy->isVectorTy())
9606 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9608 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9609 Info.ptrVal = I.getArgOperand(0);
9611 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9612 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9613 Info.vol = false; // volatile stores with NEON intrinsics not supported
9614 Info.readMem = false;
9615 Info.writeMem = true;
9618 case Intrinsic::arm_strexd: {
9619 Info.opc = ISD::INTRINSIC_W_CHAIN;
9620 Info.memVT = MVT::i64;
9621 Info.ptrVal = I.getArgOperand(2);
9625 Info.readMem = false;
9626 Info.writeMem = true;
9629 case Intrinsic::arm_ldrexd: {
9630 Info.opc = ISD::INTRINSIC_W_CHAIN;
9631 Info.memVT = MVT::i64;
9632 Info.ptrVal = I.getArgOperand(0);
9636 Info.readMem = true;
9637 Info.writeMem = false;