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/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineBasicBlock.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Target/TargetOptions.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(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::QPRRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
164 return new TargetLoweringObjectFileMachO();
166 return new ARMElfTargetObjectFile();
169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<ARMSubtarget>();
172 RegInfo = TM.getRegisterInfo();
173 Itins = TM.getInstrItineraryData();
175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
177 if (Subtarget->isTargetDarwin()) {
178 // Uses VFP for Thumb libfuncs if available.
179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
180 // Single-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
186 // Double-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
192 // Single-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
211 // Double-precision comparisons.
212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
230 // Floating-point to integer conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
238 // Conversions between floating types.
239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
242 // Integer to floating-point conversions.
243 // i64 conversions are done via library routines even when generating VFP
244 // instructions, so use the same ones.
245 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
246 // e.g., __floatunsidf vs. __floatunssidfvfp.
247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, 0);
256 setLibcallName(RTLIB::SRL_I128, 0);
257 setLibcallName(RTLIB::SRA_I128, 0);
259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point arithmetic helper functions
299 // RTABI chapter 4.1.2, Table 4
300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
309 // Single-precision floating-point comparison helper functions
310 // RTABI chapter 4.1.2, Table 5
311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
336 // Floating-point to integer conversions.
337 // RTABI chapter 4.1.2, Table 6
338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
355 // Conversions between floating types.
356 // RTABI chapter 4.1.2, Table 7
357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
362 // Integer to floating-point conversions.
363 // RTABI chapter 4.1.2, Table 8
364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 // Long long helper functions
382 // RTABI chapter 4.2, Table 9
383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
394 // Integer division functions
395 // RTABI chapter 4.3.1
396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
414 // RTABI chapter 4.3.4
415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
423 // Use divmod compiler-rt calls for iOS 5.0 and later.
424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430 if (Subtarget->isThumb1Only())
431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
433 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
435 !Subtarget->isThumb1Only()) {
436 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
437 if (!Subtarget->isFPOnlySP())
438 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
447 setTruncStoreAction((MVT::SimpleValueType)VT,
448 (MVT::SimpleValueType)InnerVT, Expand);
449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
456 if (Subtarget->hasNEON()) {
457 addDRTypeForNEON(MVT::v2f32);
458 addDRTypeForNEON(MVT::v8i8);
459 addDRTypeForNEON(MVT::v4i16);
460 addDRTypeForNEON(MVT::v2i32);
461 addDRTypeForNEON(MVT::v1i64);
463 addQRTypeForNEON(MVT::v4f32);
464 addQRTypeForNEON(MVT::v2f64);
465 addQRTypeForNEON(MVT::v16i8);
466 addQRTypeForNEON(MVT::v8i16);
467 addQRTypeForNEON(MVT::v4i32);
468 addQRTypeForNEON(MVT::v2i64);
470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
471 // neither Neon nor VFP support any arithmetic operations on it.
472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
473 // supported for v4f32.
474 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
477 // FIXME: Code duplication: FDIV and FREM are expanded always, see
478 // ARMTargetLowering::addTypeForNEON method for details.
479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
480 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
481 // FIXME: Create unittest.
482 // In another words, find a way when "copysign" appears in DAG with vector
484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
485 // FIXME: Code duplication: SETCC has custom operation action, see
486 // ARMTargetLowering::addTypeForNEON method for details.
487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
488 // FIXME: Create unittest for FNEG and for FABS.
489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
490 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
507 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
525 // Mark v2f32 intrinsics.
526 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
527 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
528 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
529 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
530 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
531 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
532 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
533 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
534 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
535 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
536 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
537 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
538 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
539 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
540 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
542 // Neon does not support some operations on v1i64 and v2i64 types.
543 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
544 // Custom handling for some quad-vector types to detect VMULL.
545 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
546 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
547 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
548 // Custom handling for some vector types to avoid expensive expansions
549 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
550 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
551 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
552 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
553 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
554 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
555 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
556 // a destination type that is wider than the source, and nor does
557 // it have a FP_TO_[SU]INT instruction with a narrower destination than
559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
561 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
562 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
564 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
565 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
567 // NEON does not have single instruction CTPOP for vectors with element
568 // types wider than 8-bits. However, custom lowering can leverage the
569 // v8i8/v16i8 vcnt instruction.
570 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
571 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
572 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
573 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
575 // NEON only has FMA instructions as of VFP4.
576 if (!Subtarget->hasVFP4()) {
577 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
578 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
581 setTargetDAGCombine(ISD::INTRINSIC_VOID);
582 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
583 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
584 setTargetDAGCombine(ISD::SHL);
585 setTargetDAGCombine(ISD::SRL);
586 setTargetDAGCombine(ISD::SRA);
587 setTargetDAGCombine(ISD::SIGN_EXTEND);
588 setTargetDAGCombine(ISD::ZERO_EXTEND);
589 setTargetDAGCombine(ISD::ANY_EXTEND);
590 setTargetDAGCombine(ISD::SELECT_CC);
591 setTargetDAGCombine(ISD::BUILD_VECTOR);
592 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
593 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
594 setTargetDAGCombine(ISD::STORE);
595 setTargetDAGCombine(ISD::FP_TO_SINT);
596 setTargetDAGCombine(ISD::FP_TO_UINT);
597 setTargetDAGCombine(ISD::FDIV);
599 // It is legal to extload from v4i8 to v4i16 or v4i32.
600 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
601 MVT::v4i16, MVT::v2i16,
603 for (unsigned i = 0; i < 6; ++i) {
604 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
605 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
606 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
610 // ARM and Thumb2 support UMLAL/SMLAL.
611 if (!Subtarget->isThumb1Only())
612 setTargetDAGCombine(ISD::ADDC);
615 computeRegisterProperties();
617 // ARM does not have f32 extending load.
618 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
620 // ARM does not have i1 sign extending load.
621 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
623 // ARM supports all 4 flavors of integer indexed load / store.
624 if (!Subtarget->isThumb1Only()) {
625 for (unsigned im = (unsigned)ISD::PRE_INC;
626 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
627 setIndexedLoadAction(im, MVT::i1, Legal);
628 setIndexedLoadAction(im, MVT::i8, Legal);
629 setIndexedLoadAction(im, MVT::i16, Legal);
630 setIndexedLoadAction(im, MVT::i32, Legal);
631 setIndexedStoreAction(im, MVT::i1, Legal);
632 setIndexedStoreAction(im, MVT::i8, Legal);
633 setIndexedStoreAction(im, MVT::i16, Legal);
634 setIndexedStoreAction(im, MVT::i32, Legal);
638 // i64 operation support.
639 setOperationAction(ISD::MUL, MVT::i64, Expand);
640 setOperationAction(ISD::MULHU, MVT::i32, Expand);
641 if (Subtarget->isThumb1Only()) {
642 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
643 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
645 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
646 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
647 setOperationAction(ISD::MULHS, MVT::i32, Expand);
649 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
650 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
651 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
652 setOperationAction(ISD::SRL, MVT::i64, Custom);
653 setOperationAction(ISD::SRA, MVT::i64, Custom);
655 if (!Subtarget->isThumb1Only()) {
656 // FIXME: We should do this for Thumb1 as well.
657 setOperationAction(ISD::ADDC, MVT::i32, Custom);
658 setOperationAction(ISD::ADDE, MVT::i32, Custom);
659 setOperationAction(ISD::SUBC, MVT::i32, Custom);
660 setOperationAction(ISD::SUBE, MVT::i32, Custom);
663 // ARM does not have ROTL.
664 setOperationAction(ISD::ROTL, MVT::i32, Expand);
665 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
666 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
667 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
668 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
670 // These just redirect to CTTZ and CTLZ on ARM.
671 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
672 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
674 // Only ARMv6 has BSWAP.
675 if (!Subtarget->hasV6Ops())
676 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
678 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
679 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
680 // These are expanded into libcalls if the cpu doesn't have HW divider.
681 setOperationAction(ISD::SDIV, MVT::i32, Expand);
682 setOperationAction(ISD::UDIV, MVT::i32, Expand);
684 setOperationAction(ISD::SREM, MVT::i32, Expand);
685 setOperationAction(ISD::UREM, MVT::i32, Expand);
686 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
687 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
689 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
690 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
691 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
692 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
693 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
695 setOperationAction(ISD::TRAP, MVT::Other, Legal);
697 // Use the default implementation.
698 setOperationAction(ISD::VASTART, MVT::Other, Custom);
699 setOperationAction(ISD::VAARG, MVT::Other, Expand);
700 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
701 setOperationAction(ISD::VAEND, MVT::Other, Expand);
702 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
703 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
705 if (!Subtarget->isTargetDarwin()) {
706 // Non-Darwin platforms may return values in these registers via the
707 // personality function.
708 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
709 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
710 setExceptionPointerRegister(ARM::R0);
711 setExceptionSelectorRegister(ARM::R1);
714 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
715 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
716 // the default expansion.
717 // FIXME: This should be checking for v6k, not just v6.
718 if (Subtarget->hasDataBarrier() ||
719 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
720 // membarrier needs custom lowering; the rest are legal and handled
722 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
723 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
724 // Custom lowering for 64-bit ops
725 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
726 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
727 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
728 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
729 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
730 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
731 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
732 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
733 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
734 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
735 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
736 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
737 setInsertFencesForAtomic(true);
739 // Set them all for expansion, which will force libcalls.
740 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
741 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
742 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
743 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
744 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
745 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
746 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
747 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
748 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
749 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
750 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
751 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
752 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
753 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
754 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
755 // Unordered/Monotonic case.
756 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
757 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
758 // Since the libcalls include locking, fold in the fences
759 setShouldFoldAtomicFences(true);
762 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
764 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
765 if (!Subtarget->hasV6Ops()) {
766 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
767 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
769 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
771 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
772 !Subtarget->isThumb1Only()) {
773 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
774 // iff target supports vfp2.
775 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
776 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
779 // We want to custom lower some of our intrinsics.
780 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
781 if (Subtarget->isTargetDarwin()) {
782 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
783 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
784 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
787 setOperationAction(ISD::SETCC, MVT::i32, Expand);
788 setOperationAction(ISD::SETCC, MVT::f32, Expand);
789 setOperationAction(ISD::SETCC, MVT::f64, Expand);
790 setOperationAction(ISD::SELECT, MVT::i32, Custom);
791 setOperationAction(ISD::SELECT, MVT::f32, Custom);
792 setOperationAction(ISD::SELECT, MVT::f64, Custom);
793 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
794 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
795 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
797 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
798 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
799 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
800 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
801 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
803 // We don't support sin/cos/fmod/copysign/pow
804 setOperationAction(ISD::FSIN, MVT::f64, Expand);
805 setOperationAction(ISD::FSIN, MVT::f32, Expand);
806 setOperationAction(ISD::FCOS, MVT::f32, Expand);
807 setOperationAction(ISD::FCOS, MVT::f64, Expand);
808 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
809 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
810 setOperationAction(ISD::FREM, MVT::f64, Expand);
811 setOperationAction(ISD::FREM, MVT::f32, Expand);
812 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
813 !Subtarget->isThumb1Only()) {
814 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
815 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
817 setOperationAction(ISD::FPOW, MVT::f64, Expand);
818 setOperationAction(ISD::FPOW, MVT::f32, Expand);
820 if (!Subtarget->hasVFP4()) {
821 setOperationAction(ISD::FMA, MVT::f64, Expand);
822 setOperationAction(ISD::FMA, MVT::f32, Expand);
825 // Various VFP goodness
826 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
827 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
828 if (Subtarget->hasVFP2()) {
829 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
830 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
831 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
832 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
834 // Special handling for half-precision FP.
835 if (!Subtarget->hasFP16()) {
836 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
837 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
841 // We have target-specific dag combine patterns for the following nodes:
842 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
843 setTargetDAGCombine(ISD::ADD);
844 setTargetDAGCombine(ISD::SUB);
845 setTargetDAGCombine(ISD::MUL);
846 setTargetDAGCombine(ISD::AND);
847 setTargetDAGCombine(ISD::OR);
848 setTargetDAGCombine(ISD::XOR);
850 if (Subtarget->hasV6Ops())
851 setTargetDAGCombine(ISD::SRL);
853 setStackPointerRegisterToSaveRestore(ARM::SP);
855 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
856 !Subtarget->hasVFP2())
857 setSchedulingPreference(Sched::RegPressure);
859 setSchedulingPreference(Sched::Hybrid);
861 //// temporary - rewrite interface to use type
862 MaxStoresPerMemset = 8;
863 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
864 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
865 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
866 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
867 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
869 // On ARM arguments smaller than 4 bytes are extended, so all arguments
870 // are at least 4 bytes aligned.
871 setMinStackArgumentAlignment(4);
873 BenefitFromCodePlacementOpt = true;
875 // Prefer likely predicted branches to selects on out-of-order cores.
876 PredictableSelectIsExpensive = Subtarget->isLikeA9();
878 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
881 // FIXME: It might make sense to define the representative register class as the
882 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
883 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
884 // SPR's representative would be DPR_VFP2. This should work well if register
885 // pressure tracking were modified such that a register use would increment the
886 // pressure of the register class's representative and all of it's super
887 // classes' representatives transitively. We have not implemented this because
888 // of the difficulty prior to coalescing of modeling operand register classes
889 // due to the common occurrence of cross class copies and subregister insertions
891 std::pair<const TargetRegisterClass*, uint8_t>
892 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
893 const TargetRegisterClass *RRC = 0;
895 switch (VT.SimpleTy) {
897 return TargetLowering::findRepresentativeClass(VT);
898 // Use DPR as representative register class for all floating point
899 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
900 // the cost is 1 for both f32 and f64.
901 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
902 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
903 RRC = &ARM::DPRRegClass;
904 // When NEON is used for SP, only half of the register file is available
905 // because operations that define both SP and DP results will be constrained
906 // to the VFP2 class (D0-D15). We currently model this constraint prior to
907 // coalescing by double-counting the SP regs. See the FIXME above.
908 if (Subtarget->useNEONForSinglePrecisionFP())
911 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
912 case MVT::v4f32: case MVT::v2f64:
913 RRC = &ARM::DPRRegClass;
917 RRC = &ARM::DPRRegClass;
921 RRC = &ARM::DPRRegClass;
925 return std::make_pair(RRC, Cost);
928 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
931 case ARMISD::Wrapper: return "ARMISD::Wrapper";
932 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
933 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
934 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
935 case ARMISD::CALL: return "ARMISD::CALL";
936 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
937 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
938 case ARMISD::tCALL: return "ARMISD::tCALL";
939 case ARMISD::BRCOND: return "ARMISD::BRCOND";
940 case ARMISD::BR_JT: return "ARMISD::BR_JT";
941 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
942 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
943 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
944 case ARMISD::CMP: return "ARMISD::CMP";
945 case ARMISD::CMN: return "ARMISD::CMN";
946 case ARMISD::CMPZ: return "ARMISD::CMPZ";
947 case ARMISD::CMPFP: return "ARMISD::CMPFP";
948 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
949 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
950 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
952 case ARMISD::CMOV: return "ARMISD::CMOV";
954 case ARMISD::RBIT: return "ARMISD::RBIT";
956 case ARMISD::FTOSI: return "ARMISD::FTOSI";
957 case ARMISD::FTOUI: return "ARMISD::FTOUI";
958 case ARMISD::SITOF: return "ARMISD::SITOF";
959 case ARMISD::UITOF: return "ARMISD::UITOF";
961 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
962 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
963 case ARMISD::RRX: return "ARMISD::RRX";
965 case ARMISD::ADDC: return "ARMISD::ADDC";
966 case ARMISD::ADDE: return "ARMISD::ADDE";
967 case ARMISD::SUBC: return "ARMISD::SUBC";
968 case ARMISD::SUBE: return "ARMISD::SUBE";
970 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
971 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
973 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
974 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
976 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
978 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
980 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
982 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
983 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
985 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
987 case ARMISD::VCEQ: return "ARMISD::VCEQ";
988 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
989 case ARMISD::VCGE: return "ARMISD::VCGE";
990 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
991 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
992 case ARMISD::VCGEU: return "ARMISD::VCGEU";
993 case ARMISD::VCGT: return "ARMISD::VCGT";
994 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
995 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
996 case ARMISD::VCGTU: return "ARMISD::VCGTU";
997 case ARMISD::VTST: return "ARMISD::VTST";
999 case ARMISD::VSHL: return "ARMISD::VSHL";
1000 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1001 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1002 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
1003 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
1004 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
1005 case ARMISD::VSHRN: return "ARMISD::VSHRN";
1006 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1007 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1008 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1009 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1010 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1011 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1012 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1013 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1014 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1015 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1016 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1017 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1018 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1019 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1020 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1021 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1022 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1023 case ARMISD::VDUP: return "ARMISD::VDUP";
1024 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1025 case ARMISD::VEXT: return "ARMISD::VEXT";
1026 case ARMISD::VREV64: return "ARMISD::VREV64";
1027 case ARMISD::VREV32: return "ARMISD::VREV32";
1028 case ARMISD::VREV16: return "ARMISD::VREV16";
1029 case ARMISD::VZIP: return "ARMISD::VZIP";
1030 case ARMISD::VUZP: return "ARMISD::VUZP";
1031 case ARMISD::VTRN: return "ARMISD::VTRN";
1032 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1033 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1034 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1035 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1036 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1037 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1038 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1039 case ARMISD::FMAX: return "ARMISD::FMAX";
1040 case ARMISD::FMIN: return "ARMISD::FMIN";
1041 case ARMISD::BFI: return "ARMISD::BFI";
1042 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1043 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1044 case ARMISD::VBSL: return "ARMISD::VBSL";
1045 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1046 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1047 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1048 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1049 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1050 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1051 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1052 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1053 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1054 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1055 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1056 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1057 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1058 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1059 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1060 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1061 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1062 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1063 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1064 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1068 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1069 if (!VT.isVector()) return getPointerTy();
1070 return VT.changeVectorElementTypeToInteger();
1073 /// getRegClassFor - Return the register class that should be used for the
1074 /// specified value type.
1075 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1076 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1077 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1078 // load / store 4 to 8 consecutive D registers.
1079 if (Subtarget->hasNEON()) {
1080 if (VT == MVT::v4i64)
1081 return &ARM::QQPRRegClass;
1082 if (VT == MVT::v8i64)
1083 return &ARM::QQQQPRRegClass;
1085 return TargetLowering::getRegClassFor(VT);
1088 // Create a fast isel object.
1090 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1091 const TargetLibraryInfo *libInfo) const {
1092 return ARM::createFastISel(funcInfo, libInfo);
1095 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1096 /// be used for loads / stores from the global.
1097 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1098 return (Subtarget->isThumb1Only() ? 127 : 4095);
1101 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1102 unsigned NumVals = N->getNumValues();
1104 return Sched::RegPressure;
1106 for (unsigned i = 0; i != NumVals; ++i) {
1107 EVT VT = N->getValueType(i);
1108 if (VT == MVT::Glue || VT == MVT::Other)
1110 if (VT.isFloatingPoint() || VT.isVector())
1114 if (!N->isMachineOpcode())
1115 return Sched::RegPressure;
1117 // Load are scheduled for latency even if there instruction itinerary
1118 // is not available.
1119 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1120 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1122 if (MCID.getNumDefs() == 0)
1123 return Sched::RegPressure;
1124 if (!Itins->isEmpty() &&
1125 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1128 return Sched::RegPressure;
1131 //===----------------------------------------------------------------------===//
1133 //===----------------------------------------------------------------------===//
1135 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1136 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1138 default: llvm_unreachable("Unknown condition code!");
1139 case ISD::SETNE: return ARMCC::NE;
1140 case ISD::SETEQ: return ARMCC::EQ;
1141 case ISD::SETGT: return ARMCC::GT;
1142 case ISD::SETGE: return ARMCC::GE;
1143 case ISD::SETLT: return ARMCC::LT;
1144 case ISD::SETLE: return ARMCC::LE;
1145 case ISD::SETUGT: return ARMCC::HI;
1146 case ISD::SETUGE: return ARMCC::HS;
1147 case ISD::SETULT: return ARMCC::LO;
1148 case ISD::SETULE: return ARMCC::LS;
1152 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1153 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1154 ARMCC::CondCodes &CondCode2) {
1155 CondCode2 = ARMCC::AL;
1157 default: llvm_unreachable("Unknown FP condition!");
1159 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1161 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1163 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1164 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1165 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1166 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1167 case ISD::SETO: CondCode = ARMCC::VC; break;
1168 case ISD::SETUO: CondCode = ARMCC::VS; break;
1169 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1170 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1171 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1173 case ISD::SETULT: CondCode = ARMCC::LT; break;
1175 case ISD::SETULE: CondCode = ARMCC::LE; break;
1177 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1181 //===----------------------------------------------------------------------===//
1182 // Calling Convention Implementation
1183 //===----------------------------------------------------------------------===//
1185 #include "ARMGenCallingConv.inc"
1187 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1188 /// given CallingConvention value.
1189 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1191 bool isVarArg) const {
1194 llvm_unreachable("Unsupported calling convention");
1195 case CallingConv::Fast:
1196 if (Subtarget->hasVFP2() && !isVarArg) {
1197 if (!Subtarget->isAAPCS_ABI())
1198 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1199 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1200 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1203 case CallingConv::C: {
1204 // Use target triple & subtarget features to do actual dispatch.
1205 if (!Subtarget->isAAPCS_ABI())
1206 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1207 else if (Subtarget->hasVFP2() &&
1208 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1210 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1211 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1213 case CallingConv::ARM_AAPCS_VFP:
1215 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1217 case CallingConv::ARM_AAPCS:
1218 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1219 case CallingConv::ARM_APCS:
1220 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1221 case CallingConv::GHC:
1222 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1226 /// LowerCallResult - Lower the result values of a call into the
1227 /// appropriate copies out of appropriate physical registers.
1229 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1230 CallingConv::ID CallConv, bool isVarArg,
1231 const SmallVectorImpl<ISD::InputArg> &Ins,
1232 DebugLoc dl, SelectionDAG &DAG,
1233 SmallVectorImpl<SDValue> &InVals) const {
1235 // Assign locations to each value returned by this call.
1236 SmallVector<CCValAssign, 16> RVLocs;
1237 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1238 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1239 CCInfo.AnalyzeCallResult(Ins,
1240 CCAssignFnForNode(CallConv, /* Return*/ true,
1243 // Copy all of the result registers out of their specified physreg.
1244 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1245 CCValAssign VA = RVLocs[i];
1248 if (VA.needsCustom()) {
1249 // Handle f64 or half of a v2f64.
1250 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1252 Chain = Lo.getValue(1);
1253 InFlag = Lo.getValue(2);
1254 VA = RVLocs[++i]; // skip ahead to next loc
1255 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1257 Chain = Hi.getValue(1);
1258 InFlag = Hi.getValue(2);
1259 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1261 if (VA.getLocVT() == MVT::v2f64) {
1262 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1263 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1264 DAG.getConstant(0, MVT::i32));
1266 VA = RVLocs[++i]; // skip ahead to next loc
1267 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1268 Chain = Lo.getValue(1);
1269 InFlag = Lo.getValue(2);
1270 VA = RVLocs[++i]; // skip ahead to next loc
1271 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1272 Chain = Hi.getValue(1);
1273 InFlag = Hi.getValue(2);
1274 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1275 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1276 DAG.getConstant(1, MVT::i32));
1279 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1281 Chain = Val.getValue(1);
1282 InFlag = Val.getValue(2);
1285 switch (VA.getLocInfo()) {
1286 default: llvm_unreachable("Unknown loc info!");
1287 case CCValAssign::Full: break;
1288 case CCValAssign::BCvt:
1289 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1293 InVals.push_back(Val);
1299 /// LowerMemOpCallTo - Store the argument to the stack.
1301 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1302 SDValue StackPtr, SDValue Arg,
1303 DebugLoc dl, SelectionDAG &DAG,
1304 const CCValAssign &VA,
1305 ISD::ArgFlagsTy Flags) const {
1306 unsigned LocMemOffset = VA.getLocMemOffset();
1307 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1308 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1309 return DAG.getStore(Chain, dl, Arg, PtrOff,
1310 MachinePointerInfo::getStack(LocMemOffset),
1314 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1315 SDValue Chain, SDValue &Arg,
1316 RegsToPassVector &RegsToPass,
1317 CCValAssign &VA, CCValAssign &NextVA,
1319 SmallVector<SDValue, 8> &MemOpChains,
1320 ISD::ArgFlagsTy Flags) const {
1322 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1323 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1324 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1326 if (NextVA.isRegLoc())
1327 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1329 assert(NextVA.isMemLoc());
1330 if (StackPtr.getNode() == 0)
1331 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1333 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1339 /// LowerCall - Lowering a call into a callseq_start <-
1340 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1343 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1344 SmallVectorImpl<SDValue> &InVals) const {
1345 SelectionDAG &DAG = CLI.DAG;
1346 DebugLoc &dl = CLI.DL;
1347 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1348 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1349 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1350 SDValue Chain = CLI.Chain;
1351 SDValue Callee = CLI.Callee;
1352 bool &isTailCall = CLI.IsTailCall;
1353 CallingConv::ID CallConv = CLI.CallConv;
1354 bool doesNotRet = CLI.DoesNotReturn;
1355 bool isVarArg = CLI.IsVarArg;
1357 MachineFunction &MF = DAG.getMachineFunction();
1358 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1359 bool IsSibCall = false;
1360 // Disable tail calls if they're not supported.
1361 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1364 // Check if it's really possible to do a tail call.
1365 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1366 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1367 Outs, OutVals, Ins, DAG);
1368 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1369 // detected sibcalls.
1376 // Analyze operands of the call, assigning locations to each operand.
1377 SmallVector<CCValAssign, 16> ArgLocs;
1378 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1379 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1380 CCInfo.AnalyzeCallOperands(Outs,
1381 CCAssignFnForNode(CallConv, /* Return*/ false,
1384 // Get a count of how many bytes are to be pushed on the stack.
1385 unsigned NumBytes = CCInfo.getNextStackOffset();
1387 // For tail calls, memory operands are available in our caller's stack.
1391 // Adjust the stack pointer for the new arguments...
1392 // These operations are automatically eliminated by the prolog/epilog pass
1394 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1396 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1398 RegsToPassVector RegsToPass;
1399 SmallVector<SDValue, 8> MemOpChains;
1401 // Walk the register/memloc assignments, inserting copies/loads. In the case
1402 // of tail call optimization, arguments are handled later.
1403 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1405 ++i, ++realArgIdx) {
1406 CCValAssign &VA = ArgLocs[i];
1407 SDValue Arg = OutVals[realArgIdx];
1408 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1409 bool isByVal = Flags.isByVal();
1411 // Promote the value if needed.
1412 switch (VA.getLocInfo()) {
1413 default: llvm_unreachable("Unknown loc info!");
1414 case CCValAssign::Full: break;
1415 case CCValAssign::SExt:
1416 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1418 case CCValAssign::ZExt:
1419 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1421 case CCValAssign::AExt:
1422 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1424 case CCValAssign::BCvt:
1425 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1429 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1430 if (VA.needsCustom()) {
1431 if (VA.getLocVT() == MVT::v2f64) {
1432 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1433 DAG.getConstant(0, MVT::i32));
1434 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1435 DAG.getConstant(1, MVT::i32));
1437 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1438 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1440 VA = ArgLocs[++i]; // skip ahead to next loc
1441 if (VA.isRegLoc()) {
1442 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1443 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1445 assert(VA.isMemLoc());
1447 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1448 dl, DAG, VA, Flags));
1451 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1452 StackPtr, MemOpChains, Flags);
1454 } else if (VA.isRegLoc()) {
1455 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1456 } else if (isByVal) {
1457 assert(VA.isMemLoc());
1458 unsigned offset = 0;
1460 // True if this byval aggregate will be split between registers
1462 if (CCInfo.isFirstByValRegValid()) {
1463 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1465 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1466 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1467 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1468 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1469 MachinePointerInfo(),
1470 false, false, false, 0);
1471 MemOpChains.push_back(Load.getValue(1));
1472 RegsToPass.push_back(std::make_pair(j, Load));
1474 offset = ARM::R4 - CCInfo.getFirstByValReg();
1475 CCInfo.clearFirstByValReg();
1478 if (Flags.getByValSize() - 4*offset > 0) {
1479 unsigned LocMemOffset = VA.getLocMemOffset();
1480 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1481 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1483 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1484 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1485 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1487 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1489 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1490 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1491 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1492 Ops, array_lengthof(Ops)));
1494 } else if (!IsSibCall) {
1495 assert(VA.isMemLoc());
1497 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1498 dl, DAG, VA, Flags));
1502 if (!MemOpChains.empty())
1503 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1504 &MemOpChains[0], MemOpChains.size());
1506 // Build a sequence of copy-to-reg nodes chained together with token chain
1507 // and flag operands which copy the outgoing args into the appropriate regs.
1509 // Tail call byval lowering might overwrite argument registers so in case of
1510 // tail call optimization the copies to registers are lowered later.
1512 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1513 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1514 RegsToPass[i].second, InFlag);
1515 InFlag = Chain.getValue(1);
1518 // For tail calls lower the arguments to the 'real' stack slot.
1520 // Force all the incoming stack arguments to be loaded from the stack
1521 // before any new outgoing arguments are stored to the stack, because the
1522 // outgoing stack slots may alias the incoming argument stack slots, and
1523 // the alias isn't otherwise explicit. This is slightly more conservative
1524 // than necessary, because it means that each store effectively depends
1525 // on every argument instead of just those arguments it would clobber.
1527 // Do not flag preceding copytoreg stuff together with the following stuff.
1529 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1530 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1531 RegsToPass[i].second, InFlag);
1532 InFlag = Chain.getValue(1);
1537 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1538 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1539 // node so that legalize doesn't hack it.
1540 bool isDirect = false;
1541 bool isARMFunc = false;
1542 bool isLocalARMFunc = false;
1543 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1545 if (EnableARMLongCalls) {
1546 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1547 && "long-calls with non-static relocation model!");
1548 // Handle a global address or an external symbol. If it's not one of
1549 // those, the target's already in a register, so we don't need to do
1551 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1552 const GlobalValue *GV = G->getGlobal();
1553 // Create a constant pool entry for the callee address
1554 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1555 ARMConstantPoolValue *CPV =
1556 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1558 // Get the address of the callee into a register
1559 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1560 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1561 Callee = DAG.getLoad(getPointerTy(), dl,
1562 DAG.getEntryNode(), CPAddr,
1563 MachinePointerInfo::getConstantPool(),
1564 false, false, false, 0);
1565 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1566 const char *Sym = S->getSymbol();
1568 // Create a constant pool entry for the callee address
1569 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1570 ARMConstantPoolValue *CPV =
1571 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1572 ARMPCLabelIndex, 0);
1573 // Get the address of the callee into a register
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);
1581 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1582 const GlobalValue *GV = G->getGlobal();
1584 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1585 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1586 getTargetMachine().getRelocationModel() != Reloc::Static;
1587 isARMFunc = !Subtarget->isThumb() || isStub;
1588 // ARM call to a local ARM function is predicable.
1589 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1590 // tBX takes a register source operand.
1591 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1592 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1593 ARMConstantPoolValue *CPV =
1594 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1595 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1596 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1597 Callee = DAG.getLoad(getPointerTy(), dl,
1598 DAG.getEntryNode(), CPAddr,
1599 MachinePointerInfo::getConstantPool(),
1600 false, false, false, 0);
1601 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1602 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1603 getPointerTy(), Callee, PICLabel);
1605 // On ELF targets for PIC code, direct calls should go through the PLT
1606 unsigned OpFlags = 0;
1607 if (Subtarget->isTargetELF() &&
1608 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1609 OpFlags = ARMII::MO_PLT;
1610 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1612 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1614 bool isStub = Subtarget->isTargetDarwin() &&
1615 getTargetMachine().getRelocationModel() != Reloc::Static;
1616 isARMFunc = !Subtarget->isThumb() || isStub;
1617 // tBX takes a register source operand.
1618 const char *Sym = S->getSymbol();
1619 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1620 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1621 ARMConstantPoolValue *CPV =
1622 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1623 ARMPCLabelIndex, 4);
1624 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1625 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1626 Callee = DAG.getLoad(getPointerTy(), dl,
1627 DAG.getEntryNode(), CPAddr,
1628 MachinePointerInfo::getConstantPool(),
1629 false, false, false, 0);
1630 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1631 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1632 getPointerTy(), Callee, PICLabel);
1634 unsigned OpFlags = 0;
1635 // On ELF targets for PIC code, direct calls should go through the PLT
1636 if (Subtarget->isTargetELF() &&
1637 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1638 OpFlags = ARMII::MO_PLT;
1639 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1643 // FIXME: handle tail calls differently.
1645 bool HasMinSizeAttr = MF.getFunction()->getAttributes().
1646 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
1647 if (Subtarget->isThumb()) {
1648 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1649 CallOpc = ARMISD::CALL_NOLINK;
1651 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1653 if (!isDirect && !Subtarget->hasV5TOps())
1654 CallOpc = ARMISD::CALL_NOLINK;
1655 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1656 // Emit regular call when code size is the priority
1658 // "mov lr, pc; b _foo" to avoid confusing the RSP
1659 CallOpc = ARMISD::CALL_NOLINK;
1661 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1664 std::vector<SDValue> Ops;
1665 Ops.push_back(Chain);
1666 Ops.push_back(Callee);
1668 // Add argument registers to the end of the list so that they are known live
1670 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1671 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1672 RegsToPass[i].second.getValueType()));
1674 // Add a register mask operand representing the call-preserved registers.
1675 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1676 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1677 assert(Mask && "Missing call preserved mask for calling convention");
1678 Ops.push_back(DAG.getRegisterMask(Mask));
1680 if (InFlag.getNode())
1681 Ops.push_back(InFlag);
1683 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1685 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1687 // Returns a chain and a flag for retval copy to use.
1688 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1689 InFlag = Chain.getValue(1);
1691 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1692 DAG.getIntPtrConstant(0, true), InFlag);
1694 InFlag = Chain.getValue(1);
1696 // Handle result values, copying them out of physregs into vregs that we
1698 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1702 /// HandleByVal - Every parameter *after* a byval parameter is passed
1703 /// on the stack. Remember the next parameter register to allocate,
1704 /// and then confiscate the rest of the parameter registers to insure
1707 ARMTargetLowering::HandleByVal(
1708 CCState *State, unsigned &size, unsigned Align) const {
1709 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1710 assert((State->getCallOrPrologue() == Prologue ||
1711 State->getCallOrPrologue() == Call) &&
1712 "unhandled ParmContext");
1713 if ((!State->isFirstByValRegValid()) &&
1714 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1715 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1716 unsigned AlignInRegs = Align / 4;
1717 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1718 for (unsigned i = 0; i < Waste; ++i)
1719 reg = State->AllocateReg(GPRArgRegs, 4);
1722 State->setFirstByValReg(reg);
1723 // At a call site, a byval parameter that is split between
1724 // registers and memory needs its size truncated here. In a
1725 // function prologue, such byval parameters are reassembled in
1726 // memory, and are not truncated.
1727 if (State->getCallOrPrologue() == Call) {
1728 unsigned excess = 4 * (ARM::R4 - reg);
1729 assert(size >= excess && "expected larger existing stack allocation");
1734 // Confiscate any remaining parameter registers to preclude their
1735 // assignment to subsequent parameters.
1736 while (State->AllocateReg(GPRArgRegs, 4))
1740 /// MatchingStackOffset - Return true if the given stack call argument is
1741 /// already available in the same position (relatively) of the caller's
1742 /// incoming argument stack.
1744 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1745 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1746 const TargetInstrInfo *TII) {
1747 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1749 if (Arg.getOpcode() == ISD::CopyFromReg) {
1750 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1751 if (!TargetRegisterInfo::isVirtualRegister(VR))
1753 MachineInstr *Def = MRI->getVRegDef(VR);
1756 if (!Flags.isByVal()) {
1757 if (!TII->isLoadFromStackSlot(Def, FI))
1762 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1763 if (Flags.isByVal())
1764 // ByVal argument is passed in as a pointer but it's now being
1765 // dereferenced. e.g.
1766 // define @foo(%struct.X* %A) {
1767 // tail call @bar(%struct.X* byval %A)
1770 SDValue Ptr = Ld->getBasePtr();
1771 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1774 FI = FINode->getIndex();
1778 assert(FI != INT_MAX);
1779 if (!MFI->isFixedObjectIndex(FI))
1781 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1784 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1785 /// for tail call optimization. Targets which want to do tail call
1786 /// optimization should implement this function.
1788 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1789 CallingConv::ID CalleeCC,
1791 bool isCalleeStructRet,
1792 bool isCallerStructRet,
1793 const SmallVectorImpl<ISD::OutputArg> &Outs,
1794 const SmallVectorImpl<SDValue> &OutVals,
1795 const SmallVectorImpl<ISD::InputArg> &Ins,
1796 SelectionDAG& DAG) const {
1797 const Function *CallerF = DAG.getMachineFunction().getFunction();
1798 CallingConv::ID CallerCC = CallerF->getCallingConv();
1799 bool CCMatch = CallerCC == CalleeCC;
1801 // Look for obvious safe cases to perform tail call optimization that do not
1802 // require ABI changes. This is what gcc calls sibcall.
1804 // Do not sibcall optimize vararg calls unless the call site is not passing
1806 if (isVarArg && !Outs.empty())
1809 // Also avoid sibcall optimization if either caller or callee uses struct
1810 // return semantics.
1811 if (isCalleeStructRet || isCallerStructRet)
1814 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1815 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1816 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1817 // support in the assembler and linker to be used. This would need to be
1818 // fixed to fully support tail calls in Thumb1.
1820 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1821 // LR. This means if we need to reload LR, it takes an extra instructions,
1822 // which outweighs the value of the tail call; but here we don't know yet
1823 // whether LR is going to be used. Probably the right approach is to
1824 // generate the tail call here and turn it back into CALL/RET in
1825 // emitEpilogue if LR is used.
1827 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1828 // but we need to make sure there are enough registers; the only valid
1829 // registers are the 4 used for parameters. We don't currently do this
1831 if (Subtarget->isThumb1Only())
1834 // If the calling conventions do not match, then we'd better make sure the
1835 // results are returned in the same way as what the caller expects.
1837 SmallVector<CCValAssign, 16> RVLocs1;
1838 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1839 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1840 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1842 SmallVector<CCValAssign, 16> RVLocs2;
1843 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1844 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1845 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1847 if (RVLocs1.size() != RVLocs2.size())
1849 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1850 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1852 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1854 if (RVLocs1[i].isRegLoc()) {
1855 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1858 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1864 // If Caller's vararg or byval argument has been split between registers and
1865 // stack, do not perform tail call, since part of the argument is in caller's
1867 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1868 getInfo<ARMFunctionInfo>();
1869 if (AFI_Caller->getVarArgsRegSaveSize())
1872 // If the callee takes no arguments then go on to check the results of the
1874 if (!Outs.empty()) {
1875 // Check if stack adjustment is needed. For now, do not do this if any
1876 // argument is passed on the stack.
1877 SmallVector<CCValAssign, 16> ArgLocs;
1878 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1879 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1880 CCInfo.AnalyzeCallOperands(Outs,
1881 CCAssignFnForNode(CalleeCC, false, isVarArg));
1882 if (CCInfo.getNextStackOffset()) {
1883 MachineFunction &MF = DAG.getMachineFunction();
1885 // Check if the arguments are already laid out in the right way as
1886 // the caller's fixed stack objects.
1887 MachineFrameInfo *MFI = MF.getFrameInfo();
1888 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1889 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1890 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1892 ++i, ++realArgIdx) {
1893 CCValAssign &VA = ArgLocs[i];
1894 EVT RegVT = VA.getLocVT();
1895 SDValue Arg = OutVals[realArgIdx];
1896 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1897 if (VA.getLocInfo() == CCValAssign::Indirect)
1899 if (VA.needsCustom()) {
1900 // f64 and vector types are split into multiple registers or
1901 // register/stack-slot combinations. The types will not match
1902 // the registers; give up on memory f64 refs until we figure
1903 // out what to do about this.
1906 if (!ArgLocs[++i].isRegLoc())
1908 if (RegVT == MVT::v2f64) {
1909 if (!ArgLocs[++i].isRegLoc())
1911 if (!ArgLocs[++i].isRegLoc())
1914 } else if (!VA.isRegLoc()) {
1915 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1927 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1928 MachineFunction &MF, bool isVarArg,
1929 const SmallVectorImpl<ISD::OutputArg> &Outs,
1930 LLVMContext &Context) const {
1931 SmallVector<CCValAssign, 16> RVLocs;
1932 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
1933 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
1938 ARMTargetLowering::LowerReturn(SDValue Chain,
1939 CallingConv::ID CallConv, bool isVarArg,
1940 const SmallVectorImpl<ISD::OutputArg> &Outs,
1941 const SmallVectorImpl<SDValue> &OutVals,
1942 DebugLoc dl, SelectionDAG &DAG) const {
1944 // CCValAssign - represent the assignment of the return value to a location.
1945 SmallVector<CCValAssign, 16> RVLocs;
1947 // CCState - Info about the registers and stack slots.
1948 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1949 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1951 // Analyze outgoing return values.
1952 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1956 SmallVector<SDValue, 4> RetOps;
1957 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1959 // Copy the result values into the output registers.
1960 for (unsigned i = 0, realRVLocIdx = 0;
1962 ++i, ++realRVLocIdx) {
1963 CCValAssign &VA = RVLocs[i];
1964 assert(VA.isRegLoc() && "Can only return in registers!");
1966 SDValue Arg = OutVals[realRVLocIdx];
1968 switch (VA.getLocInfo()) {
1969 default: llvm_unreachable("Unknown loc info!");
1970 case CCValAssign::Full: break;
1971 case CCValAssign::BCvt:
1972 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1976 if (VA.needsCustom()) {
1977 if (VA.getLocVT() == MVT::v2f64) {
1978 // Extract the first half and return it in two registers.
1979 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1980 DAG.getConstant(0, MVT::i32));
1981 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1982 DAG.getVTList(MVT::i32, MVT::i32), Half);
1984 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1985 Flag = Chain.getValue(1);
1986 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1987 VA = RVLocs[++i]; // skip ahead to next loc
1988 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1989 HalfGPRs.getValue(1), Flag);
1990 Flag = Chain.getValue(1);
1991 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1992 VA = RVLocs[++i]; // skip ahead to next loc
1994 // Extract the 2nd half and fall through to handle it as an f64 value.
1995 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1996 DAG.getConstant(1, MVT::i32));
1998 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2000 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2001 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
2002 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
2003 Flag = Chain.getValue(1);
2004 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2005 VA = RVLocs[++i]; // skip ahead to next loc
2006 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
2009 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2011 // Guarantee that all emitted copies are
2012 // stuck together, avoiding something bad.
2013 Flag = Chain.getValue(1);
2014 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2017 // Update chain and glue.
2020 RetOps.push_back(Flag);
2022 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
2023 RetOps.data(), RetOps.size());
2026 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2027 if (N->getNumValues() != 1)
2029 if (!N->hasNUsesOfValue(1, 0))
2032 SDValue TCChain = Chain;
2033 SDNode *Copy = *N->use_begin();
2034 if (Copy->getOpcode() == ISD::CopyToReg) {
2035 // If the copy has a glue operand, we conservatively assume it isn't safe to
2036 // perform a tail call.
2037 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2039 TCChain = Copy->getOperand(0);
2040 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2041 SDNode *VMov = Copy;
2042 // f64 returned in a pair of GPRs.
2043 SmallPtrSet<SDNode*, 2> Copies;
2044 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2046 if (UI->getOpcode() != ISD::CopyToReg)
2050 if (Copies.size() > 2)
2053 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2055 SDValue UseChain = UI->getOperand(0);
2056 if (Copies.count(UseChain.getNode()))
2063 } else if (Copy->getOpcode() == ISD::BITCAST) {
2064 // f32 returned in a single GPR.
2065 if (!Copy->hasOneUse())
2067 Copy = *Copy->use_begin();
2068 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2070 Chain = Copy->getOperand(0);
2075 bool HasRet = false;
2076 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2078 if (UI->getOpcode() != ARMISD::RET_FLAG)
2090 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2091 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2094 if (!CI->isTailCall())
2097 return !Subtarget->isThumb1Only();
2100 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2101 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2102 // one of the above mentioned nodes. It has to be wrapped because otherwise
2103 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2104 // be used to form addressing mode. These wrapped nodes will be selected
2106 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2107 EVT PtrVT = Op.getValueType();
2108 // FIXME there is no actual debug info here
2109 DebugLoc dl = Op.getDebugLoc();
2110 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2112 if (CP->isMachineConstantPoolEntry())
2113 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2114 CP->getAlignment());
2116 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2117 CP->getAlignment());
2118 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2121 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2122 return MachineJumpTableInfo::EK_Inline;
2125 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2126 SelectionDAG &DAG) const {
2127 MachineFunction &MF = DAG.getMachineFunction();
2128 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2129 unsigned ARMPCLabelIndex = 0;
2130 DebugLoc DL = Op.getDebugLoc();
2131 EVT PtrVT = getPointerTy();
2132 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2133 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2135 if (RelocM == Reloc::Static) {
2136 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2138 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2139 ARMPCLabelIndex = AFI->createPICLabelUId();
2140 ARMConstantPoolValue *CPV =
2141 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2142 ARMCP::CPBlockAddress, PCAdj);
2143 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2145 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2146 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2147 MachinePointerInfo::getConstantPool(),
2148 false, false, false, 0);
2149 if (RelocM == Reloc::Static)
2151 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2152 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2155 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2157 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2158 SelectionDAG &DAG) const {
2159 DebugLoc dl = GA->getDebugLoc();
2160 EVT PtrVT = getPointerTy();
2161 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2162 MachineFunction &MF = DAG.getMachineFunction();
2163 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2164 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2165 ARMConstantPoolValue *CPV =
2166 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2167 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2168 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2169 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2170 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2171 MachinePointerInfo::getConstantPool(),
2172 false, false, false, 0);
2173 SDValue Chain = Argument.getValue(1);
2175 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2176 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2178 // call __tls_get_addr.
2181 Entry.Node = Argument;
2182 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2183 Args.push_back(Entry);
2184 // FIXME: is there useful debug info available here?
2185 TargetLowering::CallLoweringInfo CLI(Chain,
2186 (Type *) Type::getInt32Ty(*DAG.getContext()),
2187 false, false, false, false,
2188 0, CallingConv::C, /*isTailCall=*/false,
2189 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2190 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2191 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2192 return CallResult.first;
2195 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2196 // "local exec" model.
2198 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2200 TLSModel::Model model) const {
2201 const GlobalValue *GV = GA->getGlobal();
2202 DebugLoc dl = GA->getDebugLoc();
2204 SDValue Chain = DAG.getEntryNode();
2205 EVT PtrVT = getPointerTy();
2206 // Get the Thread Pointer
2207 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2209 if (model == TLSModel::InitialExec) {
2210 MachineFunction &MF = DAG.getMachineFunction();
2211 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2212 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2213 // Initial exec model.
2214 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2215 ARMConstantPoolValue *CPV =
2216 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2217 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2219 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2220 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2221 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2222 MachinePointerInfo::getConstantPool(),
2223 false, false, false, 0);
2224 Chain = Offset.getValue(1);
2226 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2227 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2229 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2230 MachinePointerInfo::getConstantPool(),
2231 false, false, false, 0);
2234 assert(model == TLSModel::LocalExec);
2235 ARMConstantPoolValue *CPV =
2236 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2237 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2238 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2239 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2240 MachinePointerInfo::getConstantPool(),
2241 false, false, false, 0);
2244 // The address of the thread local variable is the add of the thread
2245 // pointer with the offset of the variable.
2246 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2250 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2251 // TODO: implement the "local dynamic" model
2252 assert(Subtarget->isTargetELF() &&
2253 "TLS not implemented for non-ELF targets");
2254 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2256 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2259 case TLSModel::GeneralDynamic:
2260 case TLSModel::LocalDynamic:
2261 return LowerToTLSGeneralDynamicModel(GA, DAG);
2262 case TLSModel::InitialExec:
2263 case TLSModel::LocalExec:
2264 return LowerToTLSExecModels(GA, DAG, model);
2266 llvm_unreachable("bogus TLS model");
2269 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2270 SelectionDAG &DAG) const {
2271 EVT PtrVT = getPointerTy();
2272 DebugLoc dl = Op.getDebugLoc();
2273 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2274 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2275 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2276 ARMConstantPoolValue *CPV =
2277 ARMConstantPoolConstant::Create(GV,
2278 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2279 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2280 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2281 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2283 MachinePointerInfo::getConstantPool(),
2284 false, false, false, 0);
2285 SDValue Chain = Result.getValue(1);
2286 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2287 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2289 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2290 MachinePointerInfo::getGOT(),
2291 false, false, false, 0);
2295 // If we have T2 ops, we can materialize the address directly via movt/movw
2296 // pair. This is always cheaper.
2297 if (Subtarget->useMovt()) {
2299 // FIXME: Once remat is capable of dealing with instructions with register
2300 // operands, expand this into two nodes.
2301 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2302 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2304 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2305 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2306 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2307 MachinePointerInfo::getConstantPool(),
2308 false, false, false, 0);
2312 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2313 SelectionDAG &DAG) const {
2314 EVT PtrVT = getPointerTy();
2315 DebugLoc dl = Op.getDebugLoc();
2316 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2317 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2319 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2320 // update ARMFastISel::ARMMaterializeGV.
2321 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2323 // FIXME: Once remat is capable of dealing with instructions with register
2324 // operands, expand this into two nodes.
2325 if (RelocM == Reloc::Static)
2326 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2327 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2329 unsigned Wrapper = (RelocM == Reloc::PIC_)
2330 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2331 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2332 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2333 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2334 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2335 MachinePointerInfo::getGOT(),
2336 false, false, false, 0);
2340 unsigned ARMPCLabelIndex = 0;
2342 if (RelocM == Reloc::Static) {
2343 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2345 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2346 ARMPCLabelIndex = AFI->createPICLabelUId();
2347 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2348 ARMConstantPoolValue *CPV =
2349 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2351 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2353 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2355 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2356 MachinePointerInfo::getConstantPool(),
2357 false, false, false, 0);
2358 SDValue Chain = Result.getValue(1);
2360 if (RelocM == Reloc::PIC_) {
2361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2362 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2365 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2366 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2367 false, false, false, 0);
2372 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2373 SelectionDAG &DAG) const {
2374 assert(Subtarget->isTargetELF() &&
2375 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2376 MachineFunction &MF = DAG.getMachineFunction();
2377 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2378 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2379 EVT PtrVT = getPointerTy();
2380 DebugLoc dl = Op.getDebugLoc();
2381 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2382 ARMConstantPoolValue *CPV =
2383 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2384 ARMPCLabelIndex, PCAdj);
2385 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2386 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2387 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2388 MachinePointerInfo::getConstantPool(),
2389 false, false, false, 0);
2390 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2391 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2395 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2396 DebugLoc dl = Op.getDebugLoc();
2397 SDValue Val = DAG.getConstant(0, MVT::i32);
2398 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2399 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2400 Op.getOperand(1), Val);
2404 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2405 DebugLoc dl = Op.getDebugLoc();
2406 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2407 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2411 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2412 const ARMSubtarget *Subtarget) const {
2413 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2414 DebugLoc dl = Op.getDebugLoc();
2416 default: return SDValue(); // Don't custom lower most intrinsics.
2417 case Intrinsic::arm_thread_pointer: {
2418 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2419 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2421 case Intrinsic::eh_sjlj_lsda: {
2422 MachineFunction &MF = DAG.getMachineFunction();
2423 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2424 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2425 EVT PtrVT = getPointerTy();
2426 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2428 unsigned PCAdj = (RelocM != Reloc::PIC_)
2429 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2430 ARMConstantPoolValue *CPV =
2431 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2432 ARMCP::CPLSDA, PCAdj);
2433 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2434 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2436 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2437 MachinePointerInfo::getConstantPool(),
2438 false, false, false, 0);
2440 if (RelocM == Reloc::PIC_) {
2441 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2442 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2446 case Intrinsic::arm_neon_vmulls:
2447 case Intrinsic::arm_neon_vmullu: {
2448 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2449 ? ARMISD::VMULLs : ARMISD::VMULLu;
2450 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2451 Op.getOperand(1), Op.getOperand(2));
2456 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2457 const ARMSubtarget *Subtarget) {
2458 DebugLoc dl = Op.getDebugLoc();
2459 if (!Subtarget->hasDataBarrier()) {
2460 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2461 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2463 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2464 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2465 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2466 DAG.getConstant(0, MVT::i32));
2469 SDValue Op5 = Op.getOperand(5);
2470 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2471 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2472 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2473 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2475 ARM_MB::MemBOpt DMBOpt;
2476 if (isDeviceBarrier)
2477 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2479 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2480 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2481 DAG.getConstant(DMBOpt, MVT::i32));
2485 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2486 const ARMSubtarget *Subtarget) {
2487 // FIXME: handle "fence singlethread" more efficiently.
2488 DebugLoc dl = Op.getDebugLoc();
2489 if (!Subtarget->hasDataBarrier()) {
2490 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2491 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2493 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2494 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2495 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2496 DAG.getConstant(0, MVT::i32));
2499 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2500 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2503 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2504 const ARMSubtarget *Subtarget) {
2505 // ARM pre v5TE and Thumb1 does not have preload instructions.
2506 if (!(Subtarget->isThumb2() ||
2507 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2508 // Just preserve the chain.
2509 return Op.getOperand(0);
2511 DebugLoc dl = Op.getDebugLoc();
2512 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2514 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2515 // ARMv7 with MP extension has PLDW.
2516 return Op.getOperand(0);
2518 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2519 if (Subtarget->isThumb()) {
2521 isRead = ~isRead & 1;
2522 isData = ~isData & 1;
2525 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2526 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2527 DAG.getConstant(isData, MVT::i32));
2530 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2531 MachineFunction &MF = DAG.getMachineFunction();
2532 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2534 // vastart just stores the address of the VarArgsFrameIndex slot into the
2535 // memory location argument.
2536 DebugLoc dl = Op.getDebugLoc();
2537 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2538 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2539 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2540 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2541 MachinePointerInfo(SV), false, false, 0);
2545 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2546 SDValue &Root, SelectionDAG &DAG,
2547 DebugLoc dl) const {
2548 MachineFunction &MF = DAG.getMachineFunction();
2549 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2551 const TargetRegisterClass *RC;
2552 if (AFI->isThumb1OnlyFunction())
2553 RC = &ARM::tGPRRegClass;
2555 RC = &ARM::GPRRegClass;
2557 // Transform the arguments stored in physical registers into virtual ones.
2558 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2559 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2562 if (NextVA.isMemLoc()) {
2563 MachineFrameInfo *MFI = MF.getFrameInfo();
2564 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2566 // Create load node to retrieve arguments from the stack.
2567 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2568 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2569 MachinePointerInfo::getFixedStack(FI),
2570 false, false, false, 0);
2572 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2573 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2576 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2580 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2581 unsigned &VARegSize, unsigned &VARegSaveSize)
2584 if (CCInfo.isFirstByValRegValid())
2585 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2587 unsigned int firstUnalloced;
2588 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2589 sizeof(GPRArgRegs) /
2590 sizeof(GPRArgRegs[0]));
2591 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2594 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2595 VARegSize = NumGPRs * 4;
2596 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2599 // The remaining GPRs hold either the beginning of variable-argument
2600 // data, or the beginning of an aggregate passed by value (usually
2601 // byval). Either way, we allocate stack slots adjacent to the data
2602 // provided by our caller, and store the unallocated registers there.
2603 // If this is a variadic function, the va_list pointer will begin with
2604 // these values; otherwise, this reassembles a (byval) structure that
2605 // was split between registers and memory.
2607 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2608 DebugLoc dl, SDValue &Chain,
2609 const Value *OrigArg,
2610 unsigned OffsetFromOrigArg,
2612 bool ForceMutable) const {
2613 MachineFunction &MF = DAG.getMachineFunction();
2614 MachineFrameInfo *MFI = MF.getFrameInfo();
2615 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2616 unsigned firstRegToSaveIndex;
2617 if (CCInfo.isFirstByValRegValid())
2618 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2620 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2621 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2624 unsigned VARegSize, VARegSaveSize;
2625 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2626 if (VARegSaveSize) {
2627 // If this function is vararg, store any remaining integer argument regs
2628 // to their spots on the stack so that they may be loaded by deferencing
2629 // the result of va_next.
2630 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2631 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2632 ArgOffset + VARegSaveSize
2635 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2638 SmallVector<SDValue, 4> MemOps;
2639 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2640 const TargetRegisterClass *RC;
2641 if (AFI->isThumb1OnlyFunction())
2642 RC = &ARM::tGPRRegClass;
2644 RC = &ARM::GPRRegClass;
2646 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2647 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2649 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2650 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2652 MemOps.push_back(Store);
2653 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2654 DAG.getConstant(4, getPointerTy()));
2656 if (!MemOps.empty())
2657 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2658 &MemOps[0], MemOps.size());
2660 // This will point to the next argument passed via stack.
2661 AFI->setVarArgsFrameIndex(
2662 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable));
2666 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2667 CallingConv::ID CallConv, bool isVarArg,
2668 const SmallVectorImpl<ISD::InputArg>
2670 DebugLoc dl, SelectionDAG &DAG,
2671 SmallVectorImpl<SDValue> &InVals)
2673 MachineFunction &MF = DAG.getMachineFunction();
2674 MachineFrameInfo *MFI = MF.getFrameInfo();
2676 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2678 // Assign locations to all of the incoming arguments.
2679 SmallVector<CCValAssign, 16> ArgLocs;
2680 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2681 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2682 CCInfo.AnalyzeFormalArguments(Ins,
2683 CCAssignFnForNode(CallConv, /* Return*/ false,
2686 SmallVector<SDValue, 16> ArgValues;
2687 int lastInsIndex = -1;
2689 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2690 unsigned CurArgIdx = 0;
2691 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2692 CCValAssign &VA = ArgLocs[i];
2693 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2694 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2695 // Arguments stored in registers.
2696 if (VA.isRegLoc()) {
2697 EVT RegVT = VA.getLocVT();
2699 if (VA.needsCustom()) {
2700 // f64 and vector types are split up into multiple registers or
2701 // combinations of registers and stack slots.
2702 if (VA.getLocVT() == MVT::v2f64) {
2703 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2705 VA = ArgLocs[++i]; // skip ahead to next loc
2707 if (VA.isMemLoc()) {
2708 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2709 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2710 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2711 MachinePointerInfo::getFixedStack(FI),
2712 false, false, false, 0);
2714 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2717 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2718 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2719 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2720 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2721 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2723 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2726 const TargetRegisterClass *RC;
2728 if (RegVT == MVT::f32)
2729 RC = &ARM::SPRRegClass;
2730 else if (RegVT == MVT::f64)
2731 RC = &ARM::DPRRegClass;
2732 else if (RegVT == MVT::v2f64)
2733 RC = &ARM::QPRRegClass;
2734 else if (RegVT == MVT::i32)
2735 RC = AFI->isThumb1OnlyFunction() ?
2736 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2737 (const TargetRegisterClass*)&ARM::GPRRegClass;
2739 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2741 // Transform the arguments in physical registers into virtual ones.
2742 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2743 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2746 // If this is an 8 or 16-bit value, it is really passed promoted
2747 // to 32 bits. Insert an assert[sz]ext to capture this, then
2748 // truncate to the right size.
2749 switch (VA.getLocInfo()) {
2750 default: llvm_unreachable("Unknown loc info!");
2751 case CCValAssign::Full: break;
2752 case CCValAssign::BCvt:
2753 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2755 case CCValAssign::SExt:
2756 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2757 DAG.getValueType(VA.getValVT()));
2758 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2760 case CCValAssign::ZExt:
2761 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2762 DAG.getValueType(VA.getValVT()));
2763 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2767 InVals.push_back(ArgValue);
2769 } else { // VA.isRegLoc()
2772 assert(VA.isMemLoc());
2773 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2775 int index = ArgLocs[i].getValNo();
2777 // Some Ins[] entries become multiple ArgLoc[] entries.
2778 // Process them only once.
2779 if (index != lastInsIndex)
2781 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2782 // FIXME: For now, all byval parameter objects are marked mutable.
2783 // This can be changed with more analysis.
2784 // In case of tail call optimization mark all arguments mutable.
2785 // Since they could be overwritten by lowering of arguments in case of
2787 if (Flags.isByVal()) {
2788 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2789 if (!AFI->getVarArgsFrameIndex()) {
2790 VarArgStyleRegisters(CCInfo, DAG,
2791 dl, Chain, CurOrigArg,
2792 Ins[VA.getValNo()].PartOffset,
2793 VA.getLocMemOffset(),
2794 true /*force mutable frames*/);
2795 int VAFrameIndex = AFI->getVarArgsFrameIndex();
2796 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy()));
2798 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
2799 VA.getLocMemOffset(), false);
2800 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2803 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2804 VA.getLocMemOffset(), true);
2806 // Create load nodes to retrieve arguments from the stack.
2807 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2808 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2809 MachinePointerInfo::getFixedStack(FI),
2810 false, false, false, 0));
2812 lastInsIndex = index;
2819 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2820 CCInfo.getNextStackOffset());
2825 /// isFloatingPointZero - Return true if this is +0.0.
2826 static bool isFloatingPointZero(SDValue Op) {
2827 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2828 return CFP->getValueAPF().isPosZero();
2829 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2830 // Maybe this has already been legalized into the constant pool?
2831 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2832 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2833 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2834 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2835 return CFP->getValueAPF().isPosZero();
2841 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2842 /// the given operands.
2844 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2845 SDValue &ARMcc, SelectionDAG &DAG,
2846 DebugLoc dl) const {
2847 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2848 unsigned C = RHSC->getZExtValue();
2849 if (!isLegalICmpImmediate(C)) {
2850 // Constant does not fit, try adjusting it by one?
2855 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2856 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2857 RHS = DAG.getConstant(C-1, MVT::i32);
2862 if (C != 0 && isLegalICmpImmediate(C-1)) {
2863 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2864 RHS = DAG.getConstant(C-1, MVT::i32);
2869 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2870 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2871 RHS = DAG.getConstant(C+1, MVT::i32);
2876 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2877 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2878 RHS = DAG.getConstant(C+1, MVT::i32);
2885 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2886 ARMISD::NodeType CompareType;
2889 CompareType = ARMISD::CMP;
2894 CompareType = ARMISD::CMPZ;
2897 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2898 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2901 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2903 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2904 DebugLoc dl) const {
2906 if (!isFloatingPointZero(RHS))
2907 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2909 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2910 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2913 /// duplicateCmp - Glue values can have only one use, so this function
2914 /// duplicates a comparison node.
2916 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2917 unsigned Opc = Cmp.getOpcode();
2918 DebugLoc DL = Cmp.getDebugLoc();
2919 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2920 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2922 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2923 Cmp = Cmp.getOperand(0);
2924 Opc = Cmp.getOpcode();
2925 if (Opc == ARMISD::CMPFP)
2926 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2928 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2929 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2931 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2934 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2935 SDValue Cond = Op.getOperand(0);
2936 SDValue SelectTrue = Op.getOperand(1);
2937 SDValue SelectFalse = Op.getOperand(2);
2938 DebugLoc dl = Op.getDebugLoc();
2942 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2943 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2945 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2946 const ConstantSDNode *CMOVTrue =
2947 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2948 const ConstantSDNode *CMOVFalse =
2949 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2951 if (CMOVTrue && CMOVFalse) {
2952 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2953 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2957 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2959 False = SelectFalse;
2960 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2965 if (True.getNode() && False.getNode()) {
2966 EVT VT = Op.getValueType();
2967 SDValue ARMcc = Cond.getOperand(2);
2968 SDValue CCR = Cond.getOperand(3);
2969 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2970 assert(True.getValueType() == VT);
2971 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2976 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2977 // undefined bits before doing a full-word comparison with zero.
2978 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2979 DAG.getConstant(1, Cond.getValueType()));
2981 return DAG.getSelectCC(dl, Cond,
2982 DAG.getConstant(0, Cond.getValueType()),
2983 SelectTrue, SelectFalse, ISD::SETNE);
2986 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2987 EVT VT = Op.getValueType();
2988 SDValue LHS = Op.getOperand(0);
2989 SDValue RHS = Op.getOperand(1);
2990 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2991 SDValue TrueVal = Op.getOperand(2);
2992 SDValue FalseVal = Op.getOperand(3);
2993 DebugLoc dl = Op.getDebugLoc();
2995 if (LHS.getValueType() == MVT::i32) {
2997 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2998 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2999 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
3002 ARMCC::CondCodes CondCode, CondCode2;
3003 FPCCToARMCC(CC, CondCode, CondCode2);
3005 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3006 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3007 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3008 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3010 if (CondCode2 != ARMCC::AL) {
3011 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3012 // FIXME: Needs another CMP because flag can have but one use.
3013 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3014 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3015 Result, TrueVal, ARMcc2, CCR, Cmp2);
3020 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3021 /// to morph to an integer compare sequence.
3022 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3023 const ARMSubtarget *Subtarget) {
3024 SDNode *N = Op.getNode();
3025 if (!N->hasOneUse())
3026 // Otherwise it requires moving the value from fp to integer registers.
3028 if (!N->getNumValues())
3030 EVT VT = Op.getValueType();
3031 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3032 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3033 // vmrs are very slow, e.g. cortex-a8.
3036 if (isFloatingPointZero(Op)) {
3040 return ISD::isNormalLoad(N);
3043 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3044 if (isFloatingPointZero(Op))
3045 return DAG.getConstant(0, MVT::i32);
3047 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3048 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3049 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3050 Ld->isVolatile(), Ld->isNonTemporal(),
3051 Ld->isInvariant(), Ld->getAlignment());
3053 llvm_unreachable("Unknown VFP cmp argument!");
3056 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3057 SDValue &RetVal1, SDValue &RetVal2) {
3058 if (isFloatingPointZero(Op)) {
3059 RetVal1 = DAG.getConstant(0, MVT::i32);
3060 RetVal2 = DAG.getConstant(0, MVT::i32);
3064 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3065 SDValue Ptr = Ld->getBasePtr();
3066 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3067 Ld->getChain(), Ptr,
3068 Ld->getPointerInfo(),
3069 Ld->isVolatile(), Ld->isNonTemporal(),
3070 Ld->isInvariant(), Ld->getAlignment());
3072 EVT PtrType = Ptr.getValueType();
3073 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3074 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3075 PtrType, Ptr, DAG.getConstant(4, PtrType));
3076 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3077 Ld->getChain(), NewPtr,
3078 Ld->getPointerInfo().getWithOffset(4),
3079 Ld->isVolatile(), Ld->isNonTemporal(),
3080 Ld->isInvariant(), NewAlign);
3084 llvm_unreachable("Unknown VFP cmp argument!");
3087 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3088 /// f32 and even f64 comparisons to integer ones.
3090 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3091 SDValue Chain = Op.getOperand(0);
3092 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3093 SDValue LHS = Op.getOperand(2);
3094 SDValue RHS = Op.getOperand(3);
3095 SDValue Dest = Op.getOperand(4);
3096 DebugLoc dl = Op.getDebugLoc();
3098 bool LHSSeenZero = false;
3099 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3100 bool RHSSeenZero = false;
3101 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3102 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3103 // If unsafe fp math optimization is enabled and there are no other uses of
3104 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3105 // to an integer comparison.
3106 if (CC == ISD::SETOEQ)
3108 else if (CC == ISD::SETUNE)
3111 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3113 if (LHS.getValueType() == MVT::f32) {
3114 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3115 bitcastf32Toi32(LHS, DAG), Mask);
3116 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3117 bitcastf32Toi32(RHS, DAG), Mask);
3118 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3119 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3120 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3121 Chain, Dest, ARMcc, CCR, Cmp);
3126 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3127 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3128 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3129 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3130 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3131 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3132 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3133 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3134 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3140 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3141 SDValue Chain = Op.getOperand(0);
3142 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3143 SDValue LHS = Op.getOperand(2);
3144 SDValue RHS = Op.getOperand(3);
3145 SDValue Dest = Op.getOperand(4);
3146 DebugLoc dl = Op.getDebugLoc();
3148 if (LHS.getValueType() == MVT::i32) {
3150 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3151 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3152 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3153 Chain, Dest, ARMcc, CCR, Cmp);
3156 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3158 if (getTargetMachine().Options.UnsafeFPMath &&
3159 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3160 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3161 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3162 if (Result.getNode())
3166 ARMCC::CondCodes CondCode, CondCode2;
3167 FPCCToARMCC(CC, CondCode, CondCode2);
3169 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3170 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3171 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3172 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3173 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3174 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3175 if (CondCode2 != ARMCC::AL) {
3176 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3177 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3178 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3183 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3184 SDValue Chain = Op.getOperand(0);
3185 SDValue Table = Op.getOperand(1);
3186 SDValue Index = Op.getOperand(2);
3187 DebugLoc dl = Op.getDebugLoc();
3189 EVT PTy = getPointerTy();
3190 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3191 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3192 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3193 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3194 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3195 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3196 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3197 if (Subtarget->isThumb2()) {
3198 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3199 // which does another jump to the destination. This also makes it easier
3200 // to translate it to TBB / TBH later.
3201 // FIXME: This might not work if the function is extremely large.
3202 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3203 Addr, Op.getOperand(2), JTI, UId);
3205 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3206 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3207 MachinePointerInfo::getJumpTable(),
3208 false, false, false, 0);
3209 Chain = Addr.getValue(1);
3210 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3211 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3213 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3214 MachinePointerInfo::getJumpTable(),
3215 false, false, false, 0);
3216 Chain = Addr.getValue(1);
3217 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3221 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3222 EVT VT = Op.getValueType();
3223 DebugLoc dl = Op.getDebugLoc();
3225 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3226 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3228 return DAG.UnrollVectorOp(Op.getNode());
3231 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3232 "Invalid type for custom lowering!");
3233 if (VT != MVT::v4i16)
3234 return DAG.UnrollVectorOp(Op.getNode());
3236 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3237 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3240 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3241 EVT VT = Op.getValueType();
3243 return LowerVectorFP_TO_INT(Op, DAG);
3245 DebugLoc dl = Op.getDebugLoc();
3248 switch (Op.getOpcode()) {
3249 default: llvm_unreachable("Invalid opcode!");
3250 case ISD::FP_TO_SINT:
3251 Opc = ARMISD::FTOSI;
3253 case ISD::FP_TO_UINT:
3254 Opc = ARMISD::FTOUI;
3257 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3258 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3261 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3262 EVT VT = Op.getValueType();
3263 DebugLoc dl = Op.getDebugLoc();
3265 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3266 if (VT.getVectorElementType() == MVT::f32)
3268 return DAG.UnrollVectorOp(Op.getNode());
3271 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3272 "Invalid type for custom lowering!");
3273 if (VT != MVT::v4f32)
3274 return DAG.UnrollVectorOp(Op.getNode());
3278 switch (Op.getOpcode()) {
3279 default: llvm_unreachable("Invalid opcode!");
3280 case ISD::SINT_TO_FP:
3281 CastOpc = ISD::SIGN_EXTEND;
3282 Opc = ISD::SINT_TO_FP;
3284 case ISD::UINT_TO_FP:
3285 CastOpc = ISD::ZERO_EXTEND;
3286 Opc = ISD::UINT_TO_FP;
3290 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3291 return DAG.getNode(Opc, dl, VT, Op);
3294 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3295 EVT VT = Op.getValueType();
3297 return LowerVectorINT_TO_FP(Op, DAG);
3299 DebugLoc dl = Op.getDebugLoc();
3302 switch (Op.getOpcode()) {
3303 default: llvm_unreachable("Invalid opcode!");
3304 case ISD::SINT_TO_FP:
3305 Opc = ARMISD::SITOF;
3307 case ISD::UINT_TO_FP:
3308 Opc = ARMISD::UITOF;
3312 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3313 return DAG.getNode(Opc, dl, VT, Op);
3316 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3317 // Implement fcopysign with a fabs and a conditional fneg.
3318 SDValue Tmp0 = Op.getOperand(0);
3319 SDValue Tmp1 = Op.getOperand(1);
3320 DebugLoc dl = Op.getDebugLoc();
3321 EVT VT = Op.getValueType();
3322 EVT SrcVT = Tmp1.getValueType();
3323 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3324 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3325 bool UseNEON = !InGPR && Subtarget->hasNEON();
3328 // Use VBSL to copy the sign bit.
3329 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3330 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3331 DAG.getTargetConstant(EncodedVal, MVT::i32));
3332 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3334 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3335 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3336 DAG.getConstant(32, MVT::i32));
3337 else /*if (VT == MVT::f32)*/
3338 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3339 if (SrcVT == MVT::f32) {
3340 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3342 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3343 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3344 DAG.getConstant(32, MVT::i32));
3345 } else if (VT == MVT::f32)
3346 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3347 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3348 DAG.getConstant(32, MVT::i32));
3349 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3350 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3352 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3354 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3355 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3356 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3358 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3359 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3360 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3361 if (VT == MVT::f32) {
3362 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3363 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3364 DAG.getConstant(0, MVT::i32));
3366 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3372 // Bitcast operand 1 to i32.
3373 if (SrcVT == MVT::f64)
3374 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3375 &Tmp1, 1).getValue(1);
3376 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3378 // Or in the signbit with integer operations.
3379 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3380 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3381 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3382 if (VT == MVT::f32) {
3383 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3384 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3385 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3386 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3389 // f64: Or the high part with signbit and then combine two parts.
3390 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3392 SDValue Lo = Tmp0.getValue(0);
3393 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3394 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3395 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3398 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3399 MachineFunction &MF = DAG.getMachineFunction();
3400 MachineFrameInfo *MFI = MF.getFrameInfo();
3401 MFI->setReturnAddressIsTaken(true);
3403 EVT VT = Op.getValueType();
3404 DebugLoc dl = Op.getDebugLoc();
3405 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3407 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3408 SDValue Offset = DAG.getConstant(4, MVT::i32);
3409 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3410 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3411 MachinePointerInfo(), false, false, false, 0);
3414 // Return LR, which contains the return address. Mark it an implicit live-in.
3415 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3416 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3419 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3420 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3421 MFI->setFrameAddressIsTaken(true);
3423 EVT VT = Op.getValueType();
3424 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3425 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3426 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3427 ? ARM::R7 : ARM::R11;
3428 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3430 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3431 MachinePointerInfo(),
3432 false, false, false, 0);
3436 /// ExpandBITCAST - If the target supports VFP, this function is called to
3437 /// expand a bit convert where either the source or destination type is i64 to
3438 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3439 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3440 /// vectors), since the legalizer won't know what to do with that.
3441 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3442 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3443 DebugLoc dl = N->getDebugLoc();
3444 SDValue Op = N->getOperand(0);
3446 // This function is only supposed to be called for i64 types, either as the
3447 // source or destination of the bit convert.
3448 EVT SrcVT = Op.getValueType();
3449 EVT DstVT = N->getValueType(0);
3450 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3451 "ExpandBITCAST called for non-i64 type");
3453 // Turn i64->f64 into VMOVDRR.
3454 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3455 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3456 DAG.getConstant(0, MVT::i32));
3457 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3458 DAG.getConstant(1, MVT::i32));
3459 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3460 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3463 // Turn f64->i64 into VMOVRRD.
3464 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3465 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3466 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3467 // Merge the pieces into a single i64 value.
3468 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3474 /// getZeroVector - Returns a vector of specified type with all zero elements.
3475 /// Zero vectors are used to represent vector negation and in those cases
3476 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3477 /// not support i64 elements, so sometimes the zero vectors will need to be
3478 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3480 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3481 assert(VT.isVector() && "Expected a vector type");
3482 // The canonical modified immediate encoding of a zero vector is....0!
3483 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3484 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3485 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3486 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3489 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3490 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3491 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3492 SelectionDAG &DAG) const {
3493 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3494 EVT VT = Op.getValueType();
3495 unsigned VTBits = VT.getSizeInBits();
3496 DebugLoc dl = Op.getDebugLoc();
3497 SDValue ShOpLo = Op.getOperand(0);
3498 SDValue ShOpHi = Op.getOperand(1);
3499 SDValue ShAmt = Op.getOperand(2);
3501 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3503 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3505 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3506 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3507 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3508 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3509 DAG.getConstant(VTBits, MVT::i32));
3510 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3511 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3512 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3514 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3515 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3517 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3518 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3521 SDValue Ops[2] = { Lo, Hi };
3522 return DAG.getMergeValues(Ops, 2, dl);
3525 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3526 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3527 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3528 SelectionDAG &DAG) const {
3529 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3530 EVT VT = Op.getValueType();
3531 unsigned VTBits = VT.getSizeInBits();
3532 DebugLoc dl = Op.getDebugLoc();
3533 SDValue ShOpLo = Op.getOperand(0);
3534 SDValue ShOpHi = Op.getOperand(1);
3535 SDValue ShAmt = Op.getOperand(2);
3538 assert(Op.getOpcode() == ISD::SHL_PARTS);
3539 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3540 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3541 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3542 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3543 DAG.getConstant(VTBits, MVT::i32));
3544 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3545 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3547 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3548 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3549 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3551 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3552 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3555 SDValue Ops[2] = { Lo, Hi };
3556 return DAG.getMergeValues(Ops, 2, dl);
3559 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3560 SelectionDAG &DAG) const {
3561 // The rounding mode is in bits 23:22 of the FPSCR.
3562 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3563 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3564 // so that the shift + and get folded into a bitfield extract.
3565 DebugLoc dl = Op.getDebugLoc();
3566 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3567 DAG.getConstant(Intrinsic::arm_get_fpscr,
3569 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3570 DAG.getConstant(1U << 22, MVT::i32));
3571 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3572 DAG.getConstant(22, MVT::i32));
3573 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3574 DAG.getConstant(3, MVT::i32));
3577 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3578 const ARMSubtarget *ST) {
3579 EVT VT = N->getValueType(0);
3580 DebugLoc dl = N->getDebugLoc();
3582 if (!ST->hasV6T2Ops())
3585 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3586 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3589 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
3590 /// for each 16-bit element from operand, repeated. The basic idea is to
3591 /// leverage vcnt to get the 8-bit counts, gather and add the results.
3593 /// Trace for v4i16:
3594 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3595 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
3596 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
3597 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
3598 /// [b0 b1 b2 b3 b4 b5 b6 b7]
3599 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
3600 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
3601 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
3602 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
3603 EVT VT = N->getValueType(0);
3604 DebugLoc DL = N->getDebugLoc();
3606 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
3607 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
3608 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
3609 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
3610 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
3611 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
3614 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
3615 /// bit-count for each 16-bit element from the operand. We need slightly
3616 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
3617 /// 64/128-bit registers.
3619 /// Trace for v4i16:
3620 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
3621 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
3622 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
3623 /// v4i16:Extracted = [k0 k1 k2 k3 ]
3624 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
3625 EVT VT = N->getValueType(0);
3626 DebugLoc DL = N->getDebugLoc();
3628 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
3629 if (VT.is64BitVector()) {
3630 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
3631 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
3632 DAG.getIntPtrConstant(0));
3634 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
3635 BitCounts, DAG.getIntPtrConstant(0));
3636 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
3640 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
3641 /// bit-count for each 32-bit element from the operand. The idea here is
3642 /// to split the vector into 16-bit elements, leverage the 16-bit count
3643 /// routine, and then combine the results.
3645 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
3646 /// input = [v0 v1 ] (vi: 32-bit elements)
3647 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
3648 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
3649 /// vrev: N0 = [k1 k0 k3 k2 ]
3651 /// N1 =+[k1 k0 k3 k2 ]
3653 /// N2 =+[k1 k3 k0 k2 ]
3655 /// Extended =+[k1 k3 k0 k2 ]
3657 /// Extracted=+[k1 k3 ]
3659 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
3660 EVT VT = N->getValueType(0);
3661 DebugLoc DL = N->getDebugLoc();
3663 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
3665 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
3666 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
3667 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
3668 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
3669 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
3671 if (VT.is64BitVector()) {
3672 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
3673 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
3674 DAG.getIntPtrConstant(0));
3676 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
3677 DAG.getIntPtrConstant(0));
3678 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
3682 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
3683 const ARMSubtarget *ST) {
3684 EVT VT = N->getValueType(0);
3686 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
3687 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
3688 VT == MVT::v4i16 || VT == MVT::v8i16) &&
3689 "Unexpected type for custom ctpop lowering");
3691 if (VT.getVectorElementType() == MVT::i32)
3692 return lowerCTPOP32BitElements(N, DAG);
3694 return lowerCTPOP16BitElements(N, DAG);
3697 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3698 const ARMSubtarget *ST) {
3699 EVT VT = N->getValueType(0);
3700 DebugLoc dl = N->getDebugLoc();
3705 // Lower vector shifts on NEON to use VSHL.
3706 assert(ST->hasNEON() && "unexpected vector shift");
3708 // Left shifts translate directly to the vshiftu intrinsic.
3709 if (N->getOpcode() == ISD::SHL)
3710 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3711 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3712 N->getOperand(0), N->getOperand(1));
3714 assert((N->getOpcode() == ISD::SRA ||
3715 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3717 // NEON uses the same intrinsics for both left and right shifts. For
3718 // right shifts, the shift amounts are negative, so negate the vector of
3720 EVT ShiftVT = N->getOperand(1).getValueType();
3721 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3722 getZeroVector(ShiftVT, DAG, dl),
3724 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3725 Intrinsic::arm_neon_vshifts :
3726 Intrinsic::arm_neon_vshiftu);
3727 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3728 DAG.getConstant(vshiftInt, MVT::i32),
3729 N->getOperand(0), NegatedCount);
3732 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3733 const ARMSubtarget *ST) {
3734 EVT VT = N->getValueType(0);
3735 DebugLoc dl = N->getDebugLoc();
3737 // We can get here for a node like i32 = ISD::SHL i32, i64
3741 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3742 "Unknown shift to lower!");
3744 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3745 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3746 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3749 // If we are in thumb mode, we don't have RRX.
3750 if (ST->isThumb1Only()) return SDValue();
3752 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3753 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3754 DAG.getConstant(0, MVT::i32));
3755 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3756 DAG.getConstant(1, MVT::i32));
3758 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3759 // captures the result into a carry flag.
3760 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3761 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3763 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3764 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3766 // Merge the pieces into a single i64 value.
3767 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3770 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3771 SDValue TmpOp0, TmpOp1;
3772 bool Invert = false;
3776 SDValue Op0 = Op.getOperand(0);
3777 SDValue Op1 = Op.getOperand(1);
3778 SDValue CC = Op.getOperand(2);
3779 EVT VT = Op.getValueType();
3780 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3781 DebugLoc dl = Op.getDebugLoc();
3783 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3784 switch (SetCCOpcode) {
3785 default: llvm_unreachable("Illegal FP comparison");
3787 case ISD::SETNE: Invert = true; // Fallthrough
3789 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3791 case ISD::SETLT: Swap = true; // Fallthrough
3793 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3795 case ISD::SETLE: Swap = true; // Fallthrough
3797 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3798 case ISD::SETUGE: Swap = true; // Fallthrough
3799 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3800 case ISD::SETUGT: Swap = true; // Fallthrough
3801 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3802 case ISD::SETUEQ: Invert = true; // Fallthrough
3804 // Expand this to (OLT | OGT).
3808 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3809 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3811 case ISD::SETUO: Invert = true; // Fallthrough
3813 // Expand this to (OLT | OGE).
3817 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3818 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3822 // Integer comparisons.
3823 switch (SetCCOpcode) {
3824 default: llvm_unreachable("Illegal integer comparison");
3825 case ISD::SETNE: Invert = true;
3826 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3827 case ISD::SETLT: Swap = true;
3828 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3829 case ISD::SETLE: Swap = true;
3830 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3831 case ISD::SETULT: Swap = true;
3832 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3833 case ISD::SETULE: Swap = true;
3834 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3837 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3838 if (Opc == ARMISD::VCEQ) {
3841 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3843 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3846 // Ignore bitconvert.
3847 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3848 AndOp = AndOp.getOperand(0);
3850 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3852 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3853 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3860 std::swap(Op0, Op1);
3862 // If one of the operands is a constant vector zero, attempt to fold the
3863 // comparison to a specialized compare-against-zero form.
3865 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3867 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3868 if (Opc == ARMISD::VCGE)
3869 Opc = ARMISD::VCLEZ;
3870 else if (Opc == ARMISD::VCGT)
3871 Opc = ARMISD::VCLTZ;
3876 if (SingleOp.getNode()) {
3879 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3881 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3883 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3885 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3887 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3889 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3892 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3896 Result = DAG.getNOT(dl, Result, VT);
3901 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3902 /// valid vector constant for a NEON instruction with a "modified immediate"
3903 /// operand (e.g., VMOV). If so, return the encoded value.
3904 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3905 unsigned SplatBitSize, SelectionDAG &DAG,
3906 EVT &VT, bool is128Bits, NEONModImmType type) {
3907 unsigned OpCmode, Imm;
3909 // SplatBitSize is set to the smallest size that splats the vector, so a
3910 // zero vector will always have SplatBitSize == 8. However, NEON modified
3911 // immediate instructions others than VMOV do not support the 8-bit encoding
3912 // of a zero vector, and the default encoding of zero is supposed to be the
3917 switch (SplatBitSize) {
3919 if (type != VMOVModImm)
3921 // Any 1-byte value is OK. Op=0, Cmode=1110.
3922 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3925 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3929 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3930 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3931 if ((SplatBits & ~0xff) == 0) {
3932 // Value = 0x00nn: Op=x, Cmode=100x.
3937 if ((SplatBits & ~0xff00) == 0) {
3938 // Value = 0xnn00: Op=x, Cmode=101x.
3940 Imm = SplatBits >> 8;
3946 // NEON's 32-bit VMOV supports splat values where:
3947 // * only one byte is nonzero, or
3948 // * the least significant byte is 0xff and the second byte is nonzero, or
3949 // * the least significant 2 bytes are 0xff and the third is nonzero.
3950 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3951 if ((SplatBits & ~0xff) == 0) {
3952 // Value = 0x000000nn: Op=x, Cmode=000x.
3957 if ((SplatBits & ~0xff00) == 0) {
3958 // Value = 0x0000nn00: Op=x, Cmode=001x.
3960 Imm = SplatBits >> 8;
3963 if ((SplatBits & ~0xff0000) == 0) {
3964 // Value = 0x00nn0000: Op=x, Cmode=010x.
3966 Imm = SplatBits >> 16;
3969 if ((SplatBits & ~0xff000000) == 0) {
3970 // Value = 0xnn000000: Op=x, Cmode=011x.
3972 Imm = SplatBits >> 24;
3976 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3977 if (type == OtherModImm) return SDValue();
3979 if ((SplatBits & ~0xffff) == 0 &&
3980 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3981 // Value = 0x0000nnff: Op=x, Cmode=1100.
3983 Imm = SplatBits >> 8;
3988 if ((SplatBits & ~0xffffff) == 0 &&
3989 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3990 // Value = 0x00nnffff: Op=x, Cmode=1101.
3992 Imm = SplatBits >> 16;
3993 SplatBits |= 0xffff;
3997 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3998 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3999 // VMOV.I32. A (very) minor optimization would be to replicate the value
4000 // and fall through here to test for a valid 64-bit splat. But, then the
4001 // caller would also need to check and handle the change in size.
4005 if (type != VMOVModImm)
4007 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4008 uint64_t BitMask = 0xff;
4010 unsigned ImmMask = 1;
4012 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4013 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4016 } else if ((SplatBits & BitMask) != 0) {
4022 // Op=1, Cmode=1110.
4025 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4030 llvm_unreachable("unexpected size for isNEONModifiedImm");
4033 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4034 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4037 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4038 const ARMSubtarget *ST) const {
4039 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
4042 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4043 assert(Op.getValueType() == MVT::f32 &&
4044 "ConstantFP custom lowering should only occur for f32.");
4046 // Try splatting with a VMOV.f32...
4047 APFloat FPVal = CFP->getValueAPF();
4048 int ImmVal = ARM_AM::getFP32Imm(FPVal);
4050 DebugLoc DL = Op.getDebugLoc();
4051 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4052 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4054 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4055 DAG.getConstant(0, MVT::i32));
4058 // If that fails, try a VMOV.i32
4060 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
4061 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
4063 if (NewVal != SDValue()) {
4064 DebugLoc DL = Op.getDebugLoc();
4065 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4067 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4069 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4070 DAG.getConstant(0, MVT::i32));
4073 // Finally, try a VMVN.i32
4074 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
4076 if (NewVal != SDValue()) {
4077 DebugLoc DL = Op.getDebugLoc();
4078 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4079 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4081 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4082 DAG.getConstant(0, MVT::i32));
4088 // check if an VEXT instruction can handle the shuffle mask when the
4089 // vector sources of the shuffle are the same.
4090 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4091 unsigned NumElts = VT.getVectorNumElements();
4093 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4099 // If this is a VEXT shuffle, the immediate value is the index of the first
4100 // element. The other shuffle indices must be the successive elements after
4102 unsigned ExpectedElt = Imm;
4103 for (unsigned i = 1; i < NumElts; ++i) {
4104 // Increment the expected index. If it wraps around, just follow it
4105 // back to index zero and keep going.
4107 if (ExpectedElt == NumElts)
4110 if (M[i] < 0) continue; // ignore UNDEF indices
4111 if (ExpectedElt != static_cast<unsigned>(M[i]))
4119 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4120 bool &ReverseVEXT, unsigned &Imm) {
4121 unsigned NumElts = VT.getVectorNumElements();
4122 ReverseVEXT = false;
4124 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4130 // If this is a VEXT shuffle, the immediate value is the index of the first
4131 // element. The other shuffle indices must be the successive elements after
4133 unsigned ExpectedElt = Imm;
4134 for (unsigned i = 1; i < NumElts; ++i) {
4135 // Increment the expected index. If it wraps around, it may still be
4136 // a VEXT but the source vectors must be swapped.
4138 if (ExpectedElt == NumElts * 2) {
4143 if (M[i] < 0) continue; // ignore UNDEF indices
4144 if (ExpectedElt != static_cast<unsigned>(M[i]))
4148 // Adjust the index value if the source operands will be swapped.
4155 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4156 /// instruction with the specified blocksize. (The order of the elements
4157 /// within each block of the vector is reversed.)
4158 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4159 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4160 "Only possible block sizes for VREV are: 16, 32, 64");
4162 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4166 unsigned NumElts = VT.getVectorNumElements();
4167 unsigned BlockElts = M[0] + 1;
4168 // If the first shuffle index is UNDEF, be optimistic.
4170 BlockElts = BlockSize / EltSz;
4172 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4175 for (unsigned i = 0; i < NumElts; ++i) {
4176 if (M[i] < 0) continue; // ignore UNDEF indices
4177 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4184 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4185 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4186 // range, then 0 is placed into the resulting vector. So pretty much any mask
4187 // of 8 elements can work here.
4188 return VT == MVT::v8i8 && M.size() == 8;
4191 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4192 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4196 unsigned NumElts = VT.getVectorNumElements();
4197 WhichResult = (M[0] == 0 ? 0 : 1);
4198 for (unsigned i = 0; i < NumElts; i += 2) {
4199 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4200 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4206 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4207 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4208 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4209 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4210 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4214 unsigned NumElts = VT.getVectorNumElements();
4215 WhichResult = (M[0] == 0 ? 0 : 1);
4216 for (unsigned i = 0; i < NumElts; i += 2) {
4217 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4218 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4224 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4225 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4229 unsigned NumElts = VT.getVectorNumElements();
4230 WhichResult = (M[0] == 0 ? 0 : 1);
4231 for (unsigned i = 0; i != NumElts; ++i) {
4232 if (M[i] < 0) continue; // ignore UNDEF indices
4233 if ((unsigned) M[i] != 2 * i + WhichResult)
4237 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4238 if (VT.is64BitVector() && EltSz == 32)
4244 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4245 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4246 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4247 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4248 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4252 unsigned Half = VT.getVectorNumElements() / 2;
4253 WhichResult = (M[0] == 0 ? 0 : 1);
4254 for (unsigned j = 0; j != 2; ++j) {
4255 unsigned Idx = WhichResult;
4256 for (unsigned i = 0; i != Half; ++i) {
4257 int MIdx = M[i + j * Half];
4258 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4264 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4265 if (VT.is64BitVector() && EltSz == 32)
4271 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4272 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4276 unsigned NumElts = VT.getVectorNumElements();
4277 WhichResult = (M[0] == 0 ? 0 : 1);
4278 unsigned Idx = WhichResult * NumElts / 2;
4279 for (unsigned i = 0; i != NumElts; i += 2) {
4280 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4281 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4286 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4287 if (VT.is64BitVector() && EltSz == 32)
4293 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4294 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4295 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4296 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4297 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4301 unsigned NumElts = VT.getVectorNumElements();
4302 WhichResult = (M[0] == 0 ? 0 : 1);
4303 unsigned Idx = WhichResult * NumElts / 2;
4304 for (unsigned i = 0; i != NumElts; i += 2) {
4305 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4306 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4311 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4312 if (VT.is64BitVector() && EltSz == 32)
4318 /// \return true if this is a reverse operation on an vector.
4319 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4320 unsigned NumElts = VT.getVectorNumElements();
4321 // Make sure the mask has the right size.
4322 if (NumElts != M.size())
4325 // Look for <15, ..., 3, -1, 1, 0>.
4326 for (unsigned i = 0; i != NumElts; ++i)
4327 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4333 // If N is an integer constant that can be moved into a register in one
4334 // instruction, return an SDValue of such a constant (will become a MOV
4335 // instruction). Otherwise return null.
4336 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4337 const ARMSubtarget *ST, DebugLoc dl) {
4339 if (!isa<ConstantSDNode>(N))
4341 Val = cast<ConstantSDNode>(N)->getZExtValue();
4343 if (ST->isThumb1Only()) {
4344 if (Val <= 255 || ~Val <= 255)
4345 return DAG.getConstant(Val, MVT::i32);
4347 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4348 return DAG.getConstant(Val, MVT::i32);
4353 // If this is a case we can't handle, return null and let the default
4354 // expansion code take care of it.
4355 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4356 const ARMSubtarget *ST) const {
4357 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4358 DebugLoc dl = Op.getDebugLoc();
4359 EVT VT = Op.getValueType();
4361 APInt SplatBits, SplatUndef;
4362 unsigned SplatBitSize;
4364 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4365 if (SplatBitSize <= 64) {
4366 // Check if an immediate VMOV works.
4368 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4369 SplatUndef.getZExtValue(), SplatBitSize,
4370 DAG, VmovVT, VT.is128BitVector(),
4372 if (Val.getNode()) {
4373 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4374 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4377 // Try an immediate VMVN.
4378 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4379 Val = isNEONModifiedImm(NegatedImm,
4380 SplatUndef.getZExtValue(), SplatBitSize,
4381 DAG, VmovVT, VT.is128BitVector(),
4383 if (Val.getNode()) {
4384 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4385 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4388 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4389 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4390 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4392 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4393 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4399 // Scan through the operands to see if only one value is used.
4401 // As an optimisation, even if more than one value is used it may be more
4402 // profitable to splat with one value then change some lanes.
4404 // Heuristically we decide to do this if the vector has a "dominant" value,
4405 // defined as splatted to more than half of the lanes.
4406 unsigned NumElts = VT.getVectorNumElements();
4407 bool isOnlyLowElement = true;
4408 bool usesOnlyOneValue = true;
4409 bool hasDominantValue = false;
4410 bool isConstant = true;
4412 // Map of the number of times a particular SDValue appears in the
4414 DenseMap<SDValue, unsigned> ValueCounts;
4416 for (unsigned i = 0; i < NumElts; ++i) {
4417 SDValue V = Op.getOperand(i);
4418 if (V.getOpcode() == ISD::UNDEF)
4421 isOnlyLowElement = false;
4422 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4425 ValueCounts.insert(std::make_pair(V, 0));
4426 unsigned &Count = ValueCounts[V];
4428 // Is this value dominant? (takes up more than half of the lanes)
4429 if (++Count > (NumElts / 2)) {
4430 hasDominantValue = true;
4434 if (ValueCounts.size() != 1)
4435 usesOnlyOneValue = false;
4436 if (!Value.getNode() && ValueCounts.size() > 0)
4437 Value = ValueCounts.begin()->first;
4439 if (ValueCounts.size() == 0)
4440 return DAG.getUNDEF(VT);
4442 if (isOnlyLowElement)
4443 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4445 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4447 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4448 // i32 and try again.
4449 if (hasDominantValue && EltSize <= 32) {
4453 // If we are VDUPing a value that comes directly from a vector, that will
4454 // cause an unnecessary move to and from a GPR, where instead we could
4455 // just use VDUPLANE.
4456 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4457 // We need to create a new undef vector to use for the VDUPLANE if the
4458 // size of the vector from which we get the value is different than the
4459 // size of the vector that we need to create. We will insert the element
4460 // such that the register coalescer will remove unnecessary copies.
4461 if (VT != Value->getOperand(0).getValueType()) {
4462 ConstantSDNode *constIndex;
4463 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4464 assert(constIndex && "The index is not a constant!");
4465 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4466 VT.getVectorNumElements();
4467 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4468 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4469 Value, DAG.getConstant(index, MVT::i32)),
4470 DAG.getConstant(index, MVT::i32));
4472 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4473 Value->getOperand(0), Value->getOperand(1));
4477 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4479 if (!usesOnlyOneValue) {
4480 // The dominant value was splatted as 'N', but we now have to insert
4481 // all differing elements.
4482 for (unsigned I = 0; I < NumElts; ++I) {
4483 if (Op.getOperand(I) == Value)
4485 SmallVector<SDValue, 3> Ops;
4487 Ops.push_back(Op.getOperand(I));
4488 Ops.push_back(DAG.getConstant(I, MVT::i32));
4489 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4494 if (VT.getVectorElementType().isFloatingPoint()) {
4495 SmallVector<SDValue, 8> Ops;
4496 for (unsigned i = 0; i < NumElts; ++i)
4497 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4499 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4500 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4501 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4503 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4505 if (usesOnlyOneValue) {
4506 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4507 if (isConstant && Val.getNode())
4508 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4512 // If all elements are constants and the case above didn't get hit, fall back
4513 // to the default expansion, which will generate a load from the constant
4518 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4520 SDValue shuffle = ReconstructShuffle(Op, DAG);
4521 if (shuffle != SDValue())
4525 // Vectors with 32- or 64-bit elements can be built by directly assigning
4526 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4527 // will be legalized.
4528 if (EltSize >= 32) {
4529 // Do the expansion with floating-point types, since that is what the VFP
4530 // registers are defined to use, and since i64 is not legal.
4531 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4532 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4533 SmallVector<SDValue, 8> Ops;
4534 for (unsigned i = 0; i < NumElts; ++i)
4535 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4536 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4537 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4543 // Gather data to see if the operation can be modelled as a
4544 // shuffle in combination with VEXTs.
4545 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4546 SelectionDAG &DAG) const {
4547 DebugLoc dl = Op.getDebugLoc();
4548 EVT VT = Op.getValueType();
4549 unsigned NumElts = VT.getVectorNumElements();
4551 SmallVector<SDValue, 2> SourceVecs;
4552 SmallVector<unsigned, 2> MinElts;
4553 SmallVector<unsigned, 2> MaxElts;
4555 for (unsigned i = 0; i < NumElts; ++i) {
4556 SDValue V = Op.getOperand(i);
4557 if (V.getOpcode() == ISD::UNDEF)
4559 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4560 // A shuffle can only come from building a vector from various
4561 // elements of other vectors.
4563 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4564 VT.getVectorElementType()) {
4565 // This code doesn't know how to handle shuffles where the vector
4566 // element types do not match (this happens because type legalization
4567 // promotes the return type of EXTRACT_VECTOR_ELT).
4568 // FIXME: It might be appropriate to extend this code to handle
4569 // mismatched types.
4573 // Record this extraction against the appropriate vector if possible...
4574 SDValue SourceVec = V.getOperand(0);
4575 // If the element number isn't a constant, we can't effectively
4576 // analyze what's going on.
4577 if (!isa<ConstantSDNode>(V.getOperand(1)))
4579 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4580 bool FoundSource = false;
4581 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4582 if (SourceVecs[j] == SourceVec) {
4583 if (MinElts[j] > EltNo)
4585 if (MaxElts[j] < EltNo)
4592 // Or record a new source if not...
4594 SourceVecs.push_back(SourceVec);
4595 MinElts.push_back(EltNo);
4596 MaxElts.push_back(EltNo);
4600 // Currently only do something sane when at most two source vectors
4602 if (SourceVecs.size() > 2)
4605 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4606 int VEXTOffsets[2] = {0, 0};
4608 // This loop extracts the usage patterns of the source vectors
4609 // and prepares appropriate SDValues for a shuffle if possible.
4610 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4611 if (SourceVecs[i].getValueType() == VT) {
4612 // No VEXT necessary
4613 ShuffleSrcs[i] = SourceVecs[i];
4616 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4617 // It probably isn't worth padding out a smaller vector just to
4618 // break it down again in a shuffle.
4622 // Since only 64-bit and 128-bit vectors are legal on ARM and
4623 // we've eliminated the other cases...
4624 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4625 "unexpected vector sizes in ReconstructShuffle");
4627 if (MaxElts[i] - MinElts[i] >= NumElts) {
4628 // Span too large for a VEXT to cope
4632 if (MinElts[i] >= NumElts) {
4633 // The extraction can just take the second half
4634 VEXTOffsets[i] = NumElts;
4635 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4637 DAG.getIntPtrConstant(NumElts));
4638 } else if (MaxElts[i] < NumElts) {
4639 // The extraction can just take the first half
4641 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4643 DAG.getIntPtrConstant(0));
4645 // An actual VEXT is needed
4646 VEXTOffsets[i] = MinElts[i];
4647 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4649 DAG.getIntPtrConstant(0));
4650 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4652 DAG.getIntPtrConstant(NumElts));
4653 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4654 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4658 SmallVector<int, 8> Mask;
4660 for (unsigned i = 0; i < NumElts; ++i) {
4661 SDValue Entry = Op.getOperand(i);
4662 if (Entry.getOpcode() == ISD::UNDEF) {
4667 SDValue ExtractVec = Entry.getOperand(0);
4668 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4669 .getOperand(1))->getSExtValue();
4670 if (ExtractVec == SourceVecs[0]) {
4671 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4673 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4677 // Final check before we try to produce nonsense...
4678 if (isShuffleMaskLegal(Mask, VT))
4679 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4685 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4686 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4687 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4688 /// are assumed to be legal.
4690 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4692 if (VT.getVectorNumElements() == 4 &&
4693 (VT.is128BitVector() || VT.is64BitVector())) {
4694 unsigned PFIndexes[4];
4695 for (unsigned i = 0; i != 4; ++i) {
4699 PFIndexes[i] = M[i];
4702 // Compute the index in the perfect shuffle table.
4703 unsigned PFTableIndex =
4704 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4705 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4706 unsigned Cost = (PFEntry >> 30);
4713 unsigned Imm, WhichResult;
4715 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4716 return (EltSize >= 32 ||
4717 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4718 isVREVMask(M, VT, 64) ||
4719 isVREVMask(M, VT, 32) ||
4720 isVREVMask(M, VT, 16) ||
4721 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4722 isVTBLMask(M, VT) ||
4723 isVTRNMask(M, VT, WhichResult) ||
4724 isVUZPMask(M, VT, WhichResult) ||
4725 isVZIPMask(M, VT, WhichResult) ||
4726 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4727 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4728 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
4729 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
4732 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4733 /// the specified operations to build the shuffle.
4734 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4735 SDValue RHS, SelectionDAG &DAG,
4737 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4738 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4739 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4742 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4751 OP_VUZPL, // VUZP, left result
4752 OP_VUZPR, // VUZP, right result
4753 OP_VZIPL, // VZIP, left result
4754 OP_VZIPR, // VZIP, right result
4755 OP_VTRNL, // VTRN, left result
4756 OP_VTRNR // VTRN, right result
4759 if (OpNum == OP_COPY) {
4760 if (LHSID == (1*9+2)*9+3) return LHS;
4761 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4765 SDValue OpLHS, OpRHS;
4766 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4767 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4768 EVT VT = OpLHS.getValueType();
4771 default: llvm_unreachable("Unknown shuffle opcode!");
4773 // VREV divides the vector in half and swaps within the half.
4774 if (VT.getVectorElementType() == MVT::i32 ||
4775 VT.getVectorElementType() == MVT::f32)
4776 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4777 // vrev <4 x i16> -> VREV32
4778 if (VT.getVectorElementType() == MVT::i16)
4779 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4780 // vrev <4 x i8> -> VREV16
4781 assert(VT.getVectorElementType() == MVT::i8);
4782 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4787 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4788 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4792 return DAG.getNode(ARMISD::VEXT, dl, VT,
4794 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4797 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4798 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4801 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4802 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4805 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4806 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4810 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4811 ArrayRef<int> ShuffleMask,
4812 SelectionDAG &DAG) {
4813 // Check to see if we can use the VTBL instruction.
4814 SDValue V1 = Op.getOperand(0);
4815 SDValue V2 = Op.getOperand(1);
4816 DebugLoc DL = Op.getDebugLoc();
4818 SmallVector<SDValue, 8> VTBLMask;
4819 for (ArrayRef<int>::iterator
4820 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4821 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4823 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4824 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4825 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4828 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4829 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4833 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
4834 SelectionDAG &DAG) {
4835 DebugLoc DL = Op.getDebugLoc();
4836 SDValue OpLHS = Op.getOperand(0);
4837 EVT VT = OpLHS.getValueType();
4839 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
4840 "Expect an v8i16/v16i8 type");
4841 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
4842 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
4843 // extract the first 8 bytes into the top double word and the last 8 bytes
4844 // into the bottom double word. The v8i16 case is similar.
4845 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
4846 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
4847 DAG.getConstant(ExtractNum, MVT::i32));
4850 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4851 SDValue V1 = Op.getOperand(0);
4852 SDValue V2 = Op.getOperand(1);
4853 DebugLoc dl = Op.getDebugLoc();
4854 EVT VT = Op.getValueType();
4855 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4857 // Convert shuffles that are directly supported on NEON to target-specific
4858 // DAG nodes, instead of keeping them as shuffles and matching them again
4859 // during code selection. This is more efficient and avoids the possibility
4860 // of inconsistencies between legalization and selection.
4861 // FIXME: floating-point vectors should be canonicalized to integer vectors
4862 // of the same time so that they get CSEd properly.
4863 ArrayRef<int> ShuffleMask = SVN->getMask();
4865 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4866 if (EltSize <= 32) {
4867 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4868 int Lane = SVN->getSplatIndex();
4869 // If this is undef splat, generate it via "just" vdup, if possible.
4870 if (Lane == -1) Lane = 0;
4872 // Test if V1 is a SCALAR_TO_VECTOR.
4873 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4874 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4876 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4877 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4879 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4880 !isa<ConstantSDNode>(V1.getOperand(0))) {
4881 bool IsScalarToVector = true;
4882 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4883 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4884 IsScalarToVector = false;
4887 if (IsScalarToVector)
4888 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4890 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4891 DAG.getConstant(Lane, MVT::i32));
4896 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4899 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4900 DAG.getConstant(Imm, MVT::i32));
4903 if (isVREVMask(ShuffleMask, VT, 64))
4904 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4905 if (isVREVMask(ShuffleMask, VT, 32))
4906 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4907 if (isVREVMask(ShuffleMask, VT, 16))
4908 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4910 if (V2->getOpcode() == ISD::UNDEF &&
4911 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
4912 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
4913 DAG.getConstant(Imm, MVT::i32));
4916 // Check for Neon shuffles that modify both input vectors in place.
4917 // If both results are used, i.e., if there are two shuffles with the same
4918 // source operands and with masks corresponding to both results of one of
4919 // these operations, DAG memoization will ensure that a single node is
4920 // used for both shuffles.
4921 unsigned WhichResult;
4922 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4923 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4924 V1, V2).getValue(WhichResult);
4925 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4926 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4927 V1, V2).getValue(WhichResult);
4928 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4929 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4930 V1, V2).getValue(WhichResult);
4932 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4933 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4934 V1, V1).getValue(WhichResult);
4935 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4936 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4937 V1, V1).getValue(WhichResult);
4938 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4939 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4940 V1, V1).getValue(WhichResult);
4943 // If the shuffle is not directly supported and it has 4 elements, use
4944 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4945 unsigned NumElts = VT.getVectorNumElements();
4947 unsigned PFIndexes[4];
4948 for (unsigned i = 0; i != 4; ++i) {
4949 if (ShuffleMask[i] < 0)
4952 PFIndexes[i] = ShuffleMask[i];
4955 // Compute the index in the perfect shuffle table.
4956 unsigned PFTableIndex =
4957 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4958 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4959 unsigned Cost = (PFEntry >> 30);
4962 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4965 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4966 if (EltSize >= 32) {
4967 // Do the expansion with floating-point types, since that is what the VFP
4968 // registers are defined to use, and since i64 is not legal.
4969 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4970 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4971 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4972 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4973 SmallVector<SDValue, 8> Ops;
4974 for (unsigned i = 0; i < NumElts; ++i) {
4975 if (ShuffleMask[i] < 0)
4976 Ops.push_back(DAG.getUNDEF(EltVT));
4978 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4979 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4980 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4983 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4984 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4987 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
4988 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
4990 if (VT == MVT::v8i8) {
4991 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4992 if (NewOp.getNode())
4999 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5000 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5001 SDValue Lane = Op.getOperand(2);
5002 if (!isa<ConstantSDNode>(Lane))
5008 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5009 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5010 SDValue Lane = Op.getOperand(1);
5011 if (!isa<ConstantSDNode>(Lane))
5014 SDValue Vec = Op.getOperand(0);
5015 if (Op.getValueType() == MVT::i32 &&
5016 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5017 DebugLoc dl = Op.getDebugLoc();
5018 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5024 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5025 // The only time a CONCAT_VECTORS operation can have legal types is when
5026 // two 64-bit vectors are concatenated to a 128-bit vector.
5027 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5028 "unexpected CONCAT_VECTORS");
5029 DebugLoc dl = Op.getDebugLoc();
5030 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5031 SDValue Op0 = Op.getOperand(0);
5032 SDValue Op1 = Op.getOperand(1);
5033 if (Op0.getOpcode() != ISD::UNDEF)
5034 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5035 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5036 DAG.getIntPtrConstant(0));
5037 if (Op1.getOpcode() != ISD::UNDEF)
5038 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5039 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5040 DAG.getIntPtrConstant(1));
5041 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5044 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5045 /// element has been zero/sign-extended, depending on the isSigned parameter,
5046 /// from an integer type half its size.
5047 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5049 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5050 EVT VT = N->getValueType(0);
5051 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5052 SDNode *BVN = N->getOperand(0).getNode();
5053 if (BVN->getValueType(0) != MVT::v4i32 ||
5054 BVN->getOpcode() != ISD::BUILD_VECTOR)
5056 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5057 unsigned HiElt = 1 - LoElt;
5058 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5059 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5060 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5061 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5062 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5065 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5066 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5069 if (Hi0->isNullValue() && Hi1->isNullValue())
5075 if (N->getOpcode() != ISD::BUILD_VECTOR)
5078 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5079 SDNode *Elt = N->getOperand(i).getNode();
5080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5081 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5082 unsigned HalfSize = EltSize / 2;
5084 if (!isIntN(HalfSize, C->getSExtValue()))
5087 if (!isUIntN(HalfSize, C->getZExtValue()))
5098 /// isSignExtended - Check if a node is a vector value that is sign-extended
5099 /// or a constant BUILD_VECTOR with sign-extended elements.
5100 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5101 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5103 if (isExtendedBUILD_VECTOR(N, DAG, true))
5108 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5109 /// or a constant BUILD_VECTOR with zero-extended elements.
5110 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5111 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5113 if (isExtendedBUILD_VECTOR(N, DAG, false))
5118 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5119 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5120 /// We insert the required extension here to get the vector to fill a D register.
5121 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5124 unsigned ExtOpcode) {
5125 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5126 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5127 // 64-bits we need to insert a new extension so that it will be 64-bits.
5128 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5129 if (OrigTy.getSizeInBits() >= 64)
5132 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5133 MVT::SimpleValueType OrigSimpleTy = OrigTy.getSimpleVT().SimpleTy;
5135 switch (OrigSimpleTy) {
5136 default: llvm_unreachable("Unexpected Orig Vector Type");
5145 return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N);
5148 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5149 /// does not do any sign/zero extension. If the original vector is less
5150 /// than 64 bits, an appropriate extension will be added after the load to
5151 /// reach a total size of 64 bits. We have to add the extension separately
5152 /// because ARM does not have a sign/zero extending load for vectors.
5153 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5154 SDValue NonExtendingLoad =
5155 DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(),
5156 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5157 LD->isNonTemporal(), LD->isInvariant(),
5158 LD->getAlignment());
5160 switch (LD->getExtensionType()) {
5161 default: llvm_unreachable("Unexpected LoadExtType");
5163 case ISD::SEXTLOAD: ExtOp = ISD::SIGN_EXTEND; break;
5164 case ISD::ZEXTLOAD: ExtOp = ISD::ZERO_EXTEND; break;
5166 MVT::SimpleValueType MemType = LD->getMemoryVT().getSimpleVT().SimpleTy;
5167 MVT::SimpleValueType ExtType = LD->getValueType(0).getSimpleVT().SimpleTy;
5168 return AddRequiredExtensionForVMULL(NonExtendingLoad, DAG,
5169 MemType, ExtType, ExtOp);
5172 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5173 /// extending load, or BUILD_VECTOR with extended elements, return the
5174 /// unextended value. The unextended vector should be 64 bits so that it can
5175 /// be used as an operand to a VMULL instruction. If the original vector size
5176 /// before extension is less than 64 bits we add a an extension to resize
5177 /// the vector to 64 bits.
5178 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5179 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5180 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5181 N->getOperand(0)->getValueType(0),
5185 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5186 return SkipLoadExtensionForVMULL(LD, DAG);
5188 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5189 // have been legalized as a BITCAST from v4i32.
5190 if (N->getOpcode() == ISD::BITCAST) {
5191 SDNode *BVN = N->getOperand(0).getNode();
5192 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5193 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5194 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5195 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
5196 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5198 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5199 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5200 EVT VT = N->getValueType(0);
5201 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5202 unsigned NumElts = VT.getVectorNumElements();
5203 MVT TruncVT = MVT::getIntegerVT(EltSize);
5204 SmallVector<SDValue, 8> Ops;
5205 for (unsigned i = 0; i != NumElts; ++i) {
5206 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5207 const APInt &CInt = C->getAPIntValue();
5208 // Element types smaller than 32 bits are not legal, so use i32 elements.
5209 // The values are implicitly truncated so sext vs. zext doesn't matter.
5210 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5212 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5213 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
5216 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5217 unsigned Opcode = N->getOpcode();
5218 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5219 SDNode *N0 = N->getOperand(0).getNode();
5220 SDNode *N1 = N->getOperand(1).getNode();
5221 return N0->hasOneUse() && N1->hasOneUse() &&
5222 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5227 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5228 unsigned Opcode = N->getOpcode();
5229 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5230 SDNode *N0 = N->getOperand(0).getNode();
5231 SDNode *N1 = N->getOperand(1).getNode();
5232 return N0->hasOneUse() && N1->hasOneUse() &&
5233 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5238 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5239 // Multiplications are only custom-lowered for 128-bit vectors so that
5240 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5241 EVT VT = Op.getValueType();
5242 assert(VT.is128BitVector() && VT.isInteger() &&
5243 "unexpected type for custom-lowering ISD::MUL");
5244 SDNode *N0 = Op.getOperand(0).getNode();
5245 SDNode *N1 = Op.getOperand(1).getNode();
5246 unsigned NewOpc = 0;
5248 bool isN0SExt = isSignExtended(N0, DAG);
5249 bool isN1SExt = isSignExtended(N1, DAG);
5250 if (isN0SExt && isN1SExt)
5251 NewOpc = ARMISD::VMULLs;
5253 bool isN0ZExt = isZeroExtended(N0, DAG);
5254 bool isN1ZExt = isZeroExtended(N1, DAG);
5255 if (isN0ZExt && isN1ZExt)
5256 NewOpc = ARMISD::VMULLu;
5257 else if (isN1SExt || isN1ZExt) {
5258 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5259 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5260 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5261 NewOpc = ARMISD::VMULLs;
5263 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5264 NewOpc = ARMISD::VMULLu;
5266 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5268 NewOpc = ARMISD::VMULLu;
5274 if (VT == MVT::v2i64)
5275 // Fall through to expand this. It is not legal.
5278 // Other vector multiplications are legal.
5283 // Legalize to a VMULL instruction.
5284 DebugLoc DL = Op.getDebugLoc();
5286 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5288 Op0 = SkipExtensionForVMULL(N0, DAG);
5289 assert(Op0.getValueType().is64BitVector() &&
5290 Op1.getValueType().is64BitVector() &&
5291 "unexpected types for extended operands to VMULL");
5292 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5295 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5296 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5303 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5304 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5305 EVT Op1VT = Op1.getValueType();
5306 return DAG.getNode(N0->getOpcode(), DL, VT,
5307 DAG.getNode(NewOpc, DL, VT,
5308 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5309 DAG.getNode(NewOpc, DL, VT,
5310 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5314 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5316 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5317 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5318 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5319 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5320 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5321 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5322 // Get reciprocal estimate.
5323 // float4 recip = vrecpeq_f32(yf);
5324 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5325 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5326 // Because char has a smaller range than uchar, we can actually get away
5327 // without any newton steps. This requires that we use a weird bias
5328 // of 0xb000, however (again, this has been exhaustively tested).
5329 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5330 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5331 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5332 Y = DAG.getConstant(0xb000, MVT::i32);
5333 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5334 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5335 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5336 // Convert back to short.
5337 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5338 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5343 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5345 // Convert to float.
5346 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5347 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5348 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5349 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5350 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5351 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5353 // Use reciprocal estimate and one refinement step.
5354 // float4 recip = vrecpeq_f32(yf);
5355 // recip *= vrecpsq_f32(yf, recip);
5356 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5357 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5358 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5359 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5361 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5362 // Because short has a smaller range than ushort, we can actually get away
5363 // with only a single newton step. This requires that we use a weird bias
5364 // of 89, however (again, this has been exhaustively tested).
5365 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5366 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5367 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5368 N1 = DAG.getConstant(0x89, MVT::i32);
5369 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5370 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5371 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5372 // Convert back to integer and return.
5373 // return vmovn_s32(vcvt_s32_f32(result));
5374 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5375 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5379 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5380 EVT VT = Op.getValueType();
5381 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5382 "unexpected type for custom-lowering ISD::SDIV");
5384 DebugLoc dl = Op.getDebugLoc();
5385 SDValue N0 = Op.getOperand(0);
5386 SDValue N1 = Op.getOperand(1);
5389 if (VT == MVT::v8i8) {
5390 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5391 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5393 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5394 DAG.getIntPtrConstant(4));
5395 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5396 DAG.getIntPtrConstant(4));
5397 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5398 DAG.getIntPtrConstant(0));
5399 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5400 DAG.getIntPtrConstant(0));
5402 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5403 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5405 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5406 N0 = LowerCONCAT_VECTORS(N0, DAG);
5408 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5411 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5414 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5415 EVT VT = Op.getValueType();
5416 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5417 "unexpected type for custom-lowering ISD::UDIV");
5419 DebugLoc dl = Op.getDebugLoc();
5420 SDValue N0 = Op.getOperand(0);
5421 SDValue N1 = Op.getOperand(1);
5424 if (VT == MVT::v8i8) {
5425 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5426 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5428 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5429 DAG.getIntPtrConstant(4));
5430 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5431 DAG.getIntPtrConstant(4));
5432 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5433 DAG.getIntPtrConstant(0));
5434 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5435 DAG.getIntPtrConstant(0));
5437 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5438 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5440 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5441 N0 = LowerCONCAT_VECTORS(N0, DAG);
5443 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5444 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5449 // v4i16 sdiv ... Convert to float.
5450 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5451 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5452 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5453 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5454 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5455 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5457 // Use reciprocal estimate and two refinement steps.
5458 // float4 recip = vrecpeq_f32(yf);
5459 // recip *= vrecpsq_f32(yf, recip);
5460 // recip *= vrecpsq_f32(yf, recip);
5461 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5462 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5463 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5464 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5466 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5467 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5468 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5470 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5471 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5472 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5473 // and that it will never cause us to return an answer too large).
5474 // float4 result = as_float4(as_int4(xf*recip) + 2);
5475 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5476 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5477 N1 = DAG.getConstant(2, MVT::i32);
5478 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5479 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5480 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5481 // Convert back to integer and return.
5482 // return vmovn_u32(vcvt_s32_f32(result));
5483 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5484 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5488 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5489 EVT VT = Op.getNode()->getValueType(0);
5490 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5493 bool ExtraOp = false;
5494 switch (Op.getOpcode()) {
5495 default: llvm_unreachable("Invalid code");
5496 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5497 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5498 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5499 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5503 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5505 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5506 Op.getOperand(1), Op.getOperand(2));
5509 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5510 // Monotonic load/store is legal for all targets
5511 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5514 // Aquire/Release load/store is not legal for targets without a
5515 // dmb or equivalent available.
5521 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5522 SelectionDAG &DAG, unsigned NewOp) {
5523 DebugLoc dl = Node->getDebugLoc();
5524 assert (Node->getValueType(0) == MVT::i64 &&
5525 "Only know how to expand i64 atomics");
5527 SmallVector<SDValue, 6> Ops;
5528 Ops.push_back(Node->getOperand(0)); // Chain
5529 Ops.push_back(Node->getOperand(1)); // Ptr
5531 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5532 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5533 // High part of Val1
5534 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5535 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5536 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5537 // High part of Val1
5538 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5539 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5540 // High part of Val2
5541 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5542 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5544 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5546 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5547 cast<MemSDNode>(Node)->getMemOperand());
5548 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5549 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5550 Results.push_back(Result.getValue(2));
5553 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5554 switch (Op.getOpcode()) {
5555 default: llvm_unreachable("Don't know how to custom lower this!");
5556 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5557 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5558 case ISD::GlobalAddress:
5559 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5560 LowerGlobalAddressELF(Op, DAG);
5561 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5562 case ISD::SELECT: return LowerSELECT(Op, DAG);
5563 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5564 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5565 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5566 case ISD::VASTART: return LowerVASTART(Op, DAG);
5567 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5568 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5569 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5570 case ISD::SINT_TO_FP:
5571 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5572 case ISD::FP_TO_SINT:
5573 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5574 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5575 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5576 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5577 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5578 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5579 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5580 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5582 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5585 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5586 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5587 case ISD::SRL_PARTS:
5588 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5589 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5590 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
5591 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5592 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5593 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5594 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5595 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5596 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5597 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5598 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5599 case ISD::MUL: return LowerMUL(Op, DAG);
5600 case ISD::SDIV: return LowerSDIV(Op, DAG);
5601 case ISD::UDIV: return LowerUDIV(Op, DAG);
5605 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5606 case ISD::ATOMIC_LOAD:
5607 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5611 /// ReplaceNodeResults - Replace the results of node with an illegal result
5612 /// type with new values built out of custom code.
5613 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5614 SmallVectorImpl<SDValue>&Results,
5615 SelectionDAG &DAG) const {
5617 switch (N->getOpcode()) {
5619 llvm_unreachable("Don't know how to custom expand this!");
5621 Res = ExpandBITCAST(N, DAG);
5625 Res = Expand64BitShift(N, DAG, Subtarget);
5627 case ISD::ATOMIC_LOAD_ADD:
5628 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5630 case ISD::ATOMIC_LOAD_AND:
5631 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5633 case ISD::ATOMIC_LOAD_NAND:
5634 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5636 case ISD::ATOMIC_LOAD_OR:
5637 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5639 case ISD::ATOMIC_LOAD_SUB:
5640 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5642 case ISD::ATOMIC_LOAD_XOR:
5643 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5645 case ISD::ATOMIC_SWAP:
5646 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5648 case ISD::ATOMIC_CMP_SWAP:
5649 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5651 case ISD::ATOMIC_LOAD_MIN:
5652 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG);
5654 case ISD::ATOMIC_LOAD_UMIN:
5655 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG);
5657 case ISD::ATOMIC_LOAD_MAX:
5658 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG);
5660 case ISD::ATOMIC_LOAD_UMAX:
5661 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG);
5665 Results.push_back(Res);
5668 //===----------------------------------------------------------------------===//
5669 // ARM Scheduler Hooks
5670 //===----------------------------------------------------------------------===//
5673 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5674 MachineBasicBlock *BB,
5675 unsigned Size) const {
5676 unsigned dest = MI->getOperand(0).getReg();
5677 unsigned ptr = MI->getOperand(1).getReg();
5678 unsigned oldval = MI->getOperand(2).getReg();
5679 unsigned newval = MI->getOperand(3).getReg();
5680 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5681 DebugLoc dl = MI->getDebugLoc();
5682 bool isThumb2 = Subtarget->isThumb2();
5684 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5685 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5686 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5687 (const TargetRegisterClass*)&ARM::GPRRegClass);
5690 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5691 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5692 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5695 unsigned ldrOpc, strOpc;
5697 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5699 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5700 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5703 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5704 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5707 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5708 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5712 MachineFunction *MF = BB->getParent();
5713 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5714 MachineFunction::iterator It = BB;
5715 ++It; // insert the new blocks after the current block
5717 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5718 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5719 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5720 MF->insert(It, loop1MBB);
5721 MF->insert(It, loop2MBB);
5722 MF->insert(It, exitMBB);
5724 // Transfer the remainder of BB and its successor edges to exitMBB.
5725 exitMBB->splice(exitMBB->begin(), BB,
5726 llvm::next(MachineBasicBlock::iterator(MI)),
5728 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5732 // fallthrough --> loop1MBB
5733 BB->addSuccessor(loop1MBB);
5736 // ldrex dest, [ptr]
5740 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5741 if (ldrOpc == ARM::t2LDREX)
5743 AddDefaultPred(MIB);
5744 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5745 .addReg(dest).addReg(oldval));
5746 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5747 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5748 BB->addSuccessor(loop2MBB);
5749 BB->addSuccessor(exitMBB);
5752 // strex scratch, newval, [ptr]
5756 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5757 if (strOpc == ARM::t2STREX)
5759 AddDefaultPred(MIB);
5760 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5761 .addReg(scratch).addImm(0));
5762 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5763 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5764 BB->addSuccessor(loop1MBB);
5765 BB->addSuccessor(exitMBB);
5771 MI->eraseFromParent(); // The instruction is gone now.
5777 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5778 unsigned Size, unsigned BinOpcode) const {
5779 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5780 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5782 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5783 MachineFunction *MF = BB->getParent();
5784 MachineFunction::iterator It = BB;
5787 unsigned dest = MI->getOperand(0).getReg();
5788 unsigned ptr = MI->getOperand(1).getReg();
5789 unsigned incr = MI->getOperand(2).getReg();
5790 DebugLoc dl = MI->getDebugLoc();
5791 bool isThumb2 = Subtarget->isThumb2();
5793 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5795 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5796 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5799 unsigned ldrOpc, strOpc;
5801 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5803 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5804 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5807 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5808 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5811 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5812 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5816 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5817 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5818 MF->insert(It, loopMBB);
5819 MF->insert(It, exitMBB);
5821 // Transfer the remainder of BB and its successor edges to exitMBB.
5822 exitMBB->splice(exitMBB->begin(), BB,
5823 llvm::next(MachineBasicBlock::iterator(MI)),
5825 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5827 const TargetRegisterClass *TRC = isThumb2 ?
5828 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5829 (const TargetRegisterClass*)&ARM::GPRRegClass;
5830 unsigned scratch = MRI.createVirtualRegister(TRC);
5831 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5835 // fallthrough --> loopMBB
5836 BB->addSuccessor(loopMBB);
5840 // <binop> scratch2, dest, incr
5841 // strex scratch, scratch2, ptr
5844 // fallthrough --> exitMBB
5846 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5847 if (ldrOpc == ARM::t2LDREX)
5849 AddDefaultPred(MIB);
5851 // operand order needs to go the other way for NAND
5852 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5853 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5854 addReg(incr).addReg(dest)).addReg(0);
5856 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5857 addReg(dest).addReg(incr)).addReg(0);
5860 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5861 if (strOpc == ARM::t2STREX)
5863 AddDefaultPred(MIB);
5864 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5865 .addReg(scratch).addImm(0));
5866 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5867 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5869 BB->addSuccessor(loopMBB);
5870 BB->addSuccessor(exitMBB);
5876 MI->eraseFromParent(); // The instruction is gone now.
5882 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5883 MachineBasicBlock *BB,
5886 ARMCC::CondCodes Cond) const {
5887 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5889 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5890 MachineFunction *MF = BB->getParent();
5891 MachineFunction::iterator It = BB;
5894 unsigned dest = MI->getOperand(0).getReg();
5895 unsigned ptr = MI->getOperand(1).getReg();
5896 unsigned incr = MI->getOperand(2).getReg();
5897 unsigned oldval = dest;
5898 DebugLoc dl = MI->getDebugLoc();
5899 bool isThumb2 = Subtarget->isThumb2();
5901 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5903 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5904 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5907 unsigned ldrOpc, strOpc, extendOpc;
5909 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5911 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5912 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5913 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5916 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5917 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5918 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5921 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5922 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5927 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5928 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5929 MF->insert(It, loopMBB);
5930 MF->insert(It, exitMBB);
5932 // Transfer the remainder of BB and its successor edges to exitMBB.
5933 exitMBB->splice(exitMBB->begin(), BB,
5934 llvm::next(MachineBasicBlock::iterator(MI)),
5936 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5938 const TargetRegisterClass *TRC = isThumb2 ?
5939 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5940 (const TargetRegisterClass*)&ARM::GPRRegClass;
5941 unsigned scratch = MRI.createVirtualRegister(TRC);
5942 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5946 // fallthrough --> loopMBB
5947 BB->addSuccessor(loopMBB);
5951 // (sign extend dest, if required)
5953 // cmov.cond scratch2, incr, dest
5954 // strex scratch, scratch2, ptr
5957 // fallthrough --> exitMBB
5959 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5960 if (ldrOpc == ARM::t2LDREX)
5962 AddDefaultPred(MIB);
5964 // Sign extend the value, if necessary.
5965 if (signExtend && extendOpc) {
5966 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5967 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5972 // Build compare and cmov instructions.
5973 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5974 .addReg(oldval).addReg(incr));
5975 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5976 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5978 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5979 if (strOpc == ARM::t2STREX)
5981 AddDefaultPred(MIB);
5982 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5983 .addReg(scratch).addImm(0));
5984 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5985 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5987 BB->addSuccessor(loopMBB);
5988 BB->addSuccessor(exitMBB);
5994 MI->eraseFromParent(); // The instruction is gone now.
6000 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
6001 unsigned Op1, unsigned Op2,
6002 bool NeedsCarry, bool IsCmpxchg,
6003 bool IsMinMax, ARMCC::CondCodes CC) const {
6004 // This also handles ATOMIC_SWAP, indicated by Op1==0.
6005 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6007 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6008 MachineFunction *MF = BB->getParent();
6009 MachineFunction::iterator It = BB;
6012 unsigned destlo = MI->getOperand(0).getReg();
6013 unsigned desthi = MI->getOperand(1).getReg();
6014 unsigned ptr = MI->getOperand(2).getReg();
6015 unsigned vallo = MI->getOperand(3).getReg();
6016 unsigned valhi = MI->getOperand(4).getReg();
6017 DebugLoc dl = MI->getDebugLoc();
6018 bool isThumb2 = Subtarget->isThumb2();
6020 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
6022 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
6023 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
6024 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
6027 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6028 MachineBasicBlock *contBB = 0, *cont2BB = 0;
6029 if (IsCmpxchg || IsMinMax)
6030 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
6032 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
6033 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6035 MF->insert(It, loopMBB);
6036 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB);
6037 if (IsCmpxchg) MF->insert(It, cont2BB);
6038 MF->insert(It, exitMBB);
6040 // Transfer the remainder of BB and its successor edges to exitMBB.
6041 exitMBB->splice(exitMBB->begin(), BB,
6042 llvm::next(MachineBasicBlock::iterator(MI)),
6044 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6046 const TargetRegisterClass *TRC = isThumb2 ?
6047 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6048 (const TargetRegisterClass*)&ARM::GPRRegClass;
6049 unsigned storesuccess = MRI.createVirtualRegister(TRC);
6053 // fallthrough --> loopMBB
6054 BB->addSuccessor(loopMBB);
6057 // ldrexd r2, r3, ptr
6058 // <binopa> r0, r2, incr
6059 // <binopb> r1, r3, incr
6060 // strexd storesuccess, r0, r1, ptr
6061 // cmp storesuccess, #0
6063 // fallthrough --> exitMBB
6068 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD))
6069 .addReg(destlo, RegState::Define)
6070 .addReg(desthi, RegState::Define)
6073 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6074 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD))
6075 .addReg(GPRPair0, RegState::Define).addReg(ptr));
6076 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
6077 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo)
6078 .addReg(GPRPair0, 0, ARM::gsub_0);
6079 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi)
6080 .addReg(GPRPair0, 0, ARM::gsub_1);
6083 unsigned StoreLo, StoreHi;
6086 for (unsigned i = 0; i < 2; i++) {
6087 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
6089 .addReg(i == 0 ? destlo : desthi)
6090 .addReg(i == 0 ? vallo : valhi));
6091 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6092 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6093 BB->addSuccessor(exitMBB);
6094 BB->addSuccessor(i == 0 ? contBB : cont2BB);
6095 BB = (i == 0 ? contBB : cont2BB);
6098 // Copy to physregs for strexd
6099 StoreLo = MI->getOperand(5).getReg();
6100 StoreHi = MI->getOperand(6).getReg();
6102 // Perform binary operation
6103 unsigned tmpRegLo = MRI.createVirtualRegister(TRC);
6104 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo)
6105 .addReg(destlo).addReg(vallo))
6106 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
6107 unsigned tmpRegHi = MRI.createVirtualRegister(TRC);
6108 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi)
6109 .addReg(desthi).addReg(valhi))
6110 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax));
6115 // Copy to physregs for strexd
6120 // Compare and branch to exit block.
6121 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6122 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR);
6123 BB->addSuccessor(exitMBB);
6124 BB->addSuccessor(contBB);
6132 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess)
6133 .addReg(StoreLo).addReg(StoreHi).addReg(ptr));
6135 // Marshal a pair...
6136 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6137 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6138 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
6139 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair);
6140 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1)
6143 .addImm(ARM::gsub_0);
6144 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair)
6147 .addImm(ARM::gsub_1);
6150 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess)
6151 .addReg(StorePair).addReg(ptr));
6154 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6155 .addReg(storesuccess).addImm(0));
6156 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6157 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6159 BB->addSuccessor(loopMBB);
6160 BB->addSuccessor(exitMBB);
6166 MI->eraseFromParent(); // The instruction is gone now.
6171 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6172 /// registers the function context.
6173 void ARMTargetLowering::
6174 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6175 MachineBasicBlock *DispatchBB, int FI) const {
6176 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6177 DebugLoc dl = MI->getDebugLoc();
6178 MachineFunction *MF = MBB->getParent();
6179 MachineRegisterInfo *MRI = &MF->getRegInfo();
6180 MachineConstantPool *MCP = MF->getConstantPool();
6181 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6182 const Function *F = MF->getFunction();
6184 bool isThumb = Subtarget->isThumb();
6185 bool isThumb2 = Subtarget->isThumb2();
6187 unsigned PCLabelId = AFI->createPICLabelUId();
6188 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6189 ARMConstantPoolValue *CPV =
6190 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6191 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6193 const TargetRegisterClass *TRC = isThumb ?
6194 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6195 (const TargetRegisterClass*)&ARM::GPRRegClass;
6197 // Grab constant pool and fixed stack memory operands.
6198 MachineMemOperand *CPMMO =
6199 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6200 MachineMemOperand::MOLoad, 4, 4);
6202 MachineMemOperand *FIMMOSt =
6203 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6204 MachineMemOperand::MOStore, 4, 4);
6206 // Load the address of the dispatch MBB into the jump buffer.
6208 // Incoming value: jbuf
6209 // ldr.n r5, LCPI1_1
6212 // str r5, [$jbuf, #+4] ; &jbuf[1]
6213 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6214 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6215 .addConstantPoolIndex(CPI)
6216 .addMemOperand(CPMMO));
6217 // Set the low bit because of thumb mode.
6218 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6220 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6221 .addReg(NewVReg1, RegState::Kill)
6223 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6224 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6225 .addReg(NewVReg2, RegState::Kill)
6227 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6228 .addReg(NewVReg3, RegState::Kill)
6230 .addImm(36) // &jbuf[1] :: pc
6231 .addMemOperand(FIMMOSt));
6232 } else if (isThumb) {
6233 // Incoming value: jbuf
6234 // ldr.n r1, LCPI1_4
6238 // add r2, $jbuf, #+4 ; &jbuf[1]
6240 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6241 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6242 .addConstantPoolIndex(CPI)
6243 .addMemOperand(CPMMO));
6244 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6245 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6246 .addReg(NewVReg1, RegState::Kill)
6248 // Set the low bit because of thumb mode.
6249 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6250 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6251 .addReg(ARM::CPSR, RegState::Define)
6253 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6254 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6255 .addReg(ARM::CPSR, RegState::Define)
6256 .addReg(NewVReg2, RegState::Kill)
6257 .addReg(NewVReg3, RegState::Kill));
6258 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6259 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6261 .addImm(36)); // &jbuf[1] :: pc
6262 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6263 .addReg(NewVReg4, RegState::Kill)
6264 .addReg(NewVReg5, RegState::Kill)
6266 .addMemOperand(FIMMOSt));
6268 // Incoming value: jbuf
6271 // str r1, [$jbuf, #+4] ; &jbuf[1]
6272 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6273 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6274 .addConstantPoolIndex(CPI)
6276 .addMemOperand(CPMMO));
6277 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6278 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6279 .addReg(NewVReg1, RegState::Kill)
6280 .addImm(PCLabelId));
6281 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6282 .addReg(NewVReg2, RegState::Kill)
6284 .addImm(36) // &jbuf[1] :: pc
6285 .addMemOperand(FIMMOSt));
6289 MachineBasicBlock *ARMTargetLowering::
6290 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6291 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6292 DebugLoc dl = MI->getDebugLoc();
6293 MachineFunction *MF = MBB->getParent();
6294 MachineRegisterInfo *MRI = &MF->getRegInfo();
6295 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6296 MachineFrameInfo *MFI = MF->getFrameInfo();
6297 int FI = MFI->getFunctionContextIndex();
6299 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6300 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6301 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6303 // Get a mapping of the call site numbers to all of the landing pads they're
6305 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6306 unsigned MaxCSNum = 0;
6307 MachineModuleInfo &MMI = MF->getMMI();
6308 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6310 if (!BB->isLandingPad()) continue;
6312 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6314 for (MachineBasicBlock::iterator
6315 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6316 if (!II->isEHLabel()) continue;
6318 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6319 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6321 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6322 for (SmallVectorImpl<unsigned>::iterator
6323 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6324 CSI != CSE; ++CSI) {
6325 CallSiteNumToLPad[*CSI].push_back(BB);
6326 MaxCSNum = std::max(MaxCSNum, *CSI);
6332 // Get an ordered list of the machine basic blocks for the jump table.
6333 std::vector<MachineBasicBlock*> LPadList;
6334 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6335 LPadList.reserve(CallSiteNumToLPad.size());
6336 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6337 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6338 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6339 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6340 LPadList.push_back(*II);
6341 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6345 assert(!LPadList.empty() &&
6346 "No landing pad destinations for the dispatch jump table!");
6348 // Create the jump table and associated information.
6349 MachineJumpTableInfo *JTI =
6350 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6351 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6352 unsigned UId = AFI->createJumpTableUId();
6353 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6355 // Create the MBBs for the dispatch code.
6357 // Shove the dispatch's address into the return slot in the function context.
6358 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6359 DispatchBB->setIsLandingPad();
6361 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6362 unsigned trap_opcode;
6363 if (Subtarget->isThumb())
6364 trap_opcode = ARM::tTRAP;
6366 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6368 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6369 DispatchBB->addSuccessor(TrapBB);
6371 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6372 DispatchBB->addSuccessor(DispContBB);
6375 MF->insert(MF->end(), DispatchBB);
6376 MF->insert(MF->end(), DispContBB);
6377 MF->insert(MF->end(), TrapBB);
6379 // Insert code into the entry block that creates and registers the function
6381 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6383 MachineMemOperand *FIMMOLd =
6384 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6385 MachineMemOperand::MOLoad |
6386 MachineMemOperand::MOVolatile, 4, 4);
6388 MachineInstrBuilder MIB;
6389 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6391 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6392 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6394 // Add a register mask with no preserved registers. This results in all
6395 // registers being marked as clobbered.
6396 MIB.addRegMask(RI.getNoPreservedMask());
6398 unsigned NumLPads = LPadList.size();
6399 if (Subtarget->isThumb2()) {
6400 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6401 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6404 .addMemOperand(FIMMOLd));
6406 if (NumLPads < 256) {
6407 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6409 .addImm(LPadList.size()));
6411 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6412 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6413 .addImm(NumLPads & 0xFFFF));
6415 unsigned VReg2 = VReg1;
6416 if ((NumLPads & 0xFFFF0000) != 0) {
6417 VReg2 = MRI->createVirtualRegister(TRC);
6418 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6420 .addImm(NumLPads >> 16));
6423 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6428 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6433 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6434 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6435 .addJumpTableIndex(MJTI)
6438 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6441 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6442 .addReg(NewVReg3, RegState::Kill)
6444 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6446 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6447 .addReg(NewVReg4, RegState::Kill)
6449 .addJumpTableIndex(MJTI)
6451 } else if (Subtarget->isThumb()) {
6452 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6453 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6456 .addMemOperand(FIMMOLd));
6458 if (NumLPads < 256) {
6459 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6463 MachineConstantPool *ConstantPool = MF->getConstantPool();
6464 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6465 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6467 // MachineConstantPool wants an explicit alignment.
6468 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6470 Align = getDataLayout()->getTypeAllocSize(C->getType());
6471 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6473 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6474 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6475 .addReg(VReg1, RegState::Define)
6476 .addConstantPoolIndex(Idx));
6477 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6482 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6487 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6488 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6489 .addReg(ARM::CPSR, RegState::Define)
6493 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6494 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6495 .addJumpTableIndex(MJTI)
6498 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6499 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6500 .addReg(ARM::CPSR, RegState::Define)
6501 .addReg(NewVReg2, RegState::Kill)
6504 MachineMemOperand *JTMMOLd =
6505 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6506 MachineMemOperand::MOLoad, 4, 4);
6508 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6509 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6510 .addReg(NewVReg4, RegState::Kill)
6512 .addMemOperand(JTMMOLd));
6514 unsigned NewVReg6 = NewVReg5;
6515 if (RelocM == Reloc::PIC_) {
6516 NewVReg6 = MRI->createVirtualRegister(TRC);
6517 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6518 .addReg(ARM::CPSR, RegState::Define)
6519 .addReg(NewVReg5, RegState::Kill)
6523 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6524 .addReg(NewVReg6, RegState::Kill)
6525 .addJumpTableIndex(MJTI)
6528 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6529 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6532 .addMemOperand(FIMMOLd));
6534 if (NumLPads < 256) {
6535 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6538 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6539 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6540 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6541 .addImm(NumLPads & 0xFFFF));
6543 unsigned VReg2 = VReg1;
6544 if ((NumLPads & 0xFFFF0000) != 0) {
6545 VReg2 = MRI->createVirtualRegister(TRC);
6546 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6548 .addImm(NumLPads >> 16));
6551 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6555 MachineConstantPool *ConstantPool = MF->getConstantPool();
6556 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6557 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6559 // MachineConstantPool wants an explicit alignment.
6560 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6562 Align = getDataLayout()->getTypeAllocSize(C->getType());
6563 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6565 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6566 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6567 .addReg(VReg1, RegState::Define)
6568 .addConstantPoolIndex(Idx)
6570 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6572 .addReg(VReg1, RegState::Kill));
6575 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6580 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6582 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6584 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6585 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6586 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6587 .addJumpTableIndex(MJTI)
6590 MachineMemOperand *JTMMOLd =
6591 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6592 MachineMemOperand::MOLoad, 4, 4);
6593 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6595 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6596 .addReg(NewVReg3, RegState::Kill)
6599 .addMemOperand(JTMMOLd));
6601 if (RelocM == Reloc::PIC_) {
6602 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6603 .addReg(NewVReg5, RegState::Kill)
6605 .addJumpTableIndex(MJTI)
6608 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6609 .addReg(NewVReg5, RegState::Kill)
6610 .addJumpTableIndex(MJTI)
6615 // Add the jump table entries as successors to the MBB.
6616 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6617 for (std::vector<MachineBasicBlock*>::iterator
6618 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6619 MachineBasicBlock *CurMBB = *I;
6620 if (SeenMBBs.insert(CurMBB))
6621 DispContBB->addSuccessor(CurMBB);
6624 // N.B. the order the invoke BBs are processed in doesn't matter here.
6625 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6626 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6627 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6628 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6629 MachineBasicBlock *BB = *I;
6631 // Remove the landing pad successor from the invoke block and replace it
6632 // with the new dispatch block.
6633 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6635 while (!Successors.empty()) {
6636 MachineBasicBlock *SMBB = Successors.pop_back_val();
6637 if (SMBB->isLandingPad()) {
6638 BB->removeSuccessor(SMBB);
6639 MBBLPads.push_back(SMBB);
6643 BB->addSuccessor(DispatchBB);
6645 // Find the invoke call and mark all of the callee-saved registers as
6646 // 'implicit defined' so that they're spilled. This prevents code from
6647 // moving instructions to before the EH block, where they will never be
6649 for (MachineBasicBlock::reverse_iterator
6650 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6651 if (!II->isCall()) continue;
6653 DenseMap<unsigned, bool> DefRegs;
6654 for (MachineInstr::mop_iterator
6655 OI = II->operands_begin(), OE = II->operands_end();
6657 if (!OI->isReg()) continue;
6658 DefRegs[OI->getReg()] = true;
6661 MachineInstrBuilder MIB(*MF, &*II);
6663 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6664 unsigned Reg = SavedRegs[i];
6665 if (Subtarget->isThumb2() &&
6666 !ARM::tGPRRegClass.contains(Reg) &&
6667 !ARM::hGPRRegClass.contains(Reg))
6669 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6671 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6674 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6681 // Mark all former landing pads as non-landing pads. The dispatch is the only
6683 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6684 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6685 (*I)->setIsLandingPad(false);
6687 // The instruction is gone now.
6688 MI->eraseFromParent();
6694 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6695 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6696 E = MBB->succ_end(); I != E; ++I)
6699 llvm_unreachable("Expecting a BB with two successors!");
6702 MachineBasicBlock *ARMTargetLowering::
6703 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6704 // This pseudo instruction has 3 operands: dst, src, size
6705 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6706 // Otherwise, we will generate unrolled scalar copies.
6707 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6708 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6709 MachineFunction::iterator It = BB;
6712 unsigned dest = MI->getOperand(0).getReg();
6713 unsigned src = MI->getOperand(1).getReg();
6714 unsigned SizeVal = MI->getOperand(2).getImm();
6715 unsigned Align = MI->getOperand(3).getImm();
6716 DebugLoc dl = MI->getDebugLoc();
6718 bool isThumb2 = Subtarget->isThumb2();
6719 MachineFunction *MF = BB->getParent();
6720 MachineRegisterInfo &MRI = MF->getRegInfo();
6721 unsigned ldrOpc, strOpc, UnitSize = 0;
6723 const TargetRegisterClass *TRC = isThumb2 ?
6724 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6725 (const TargetRegisterClass*)&ARM::GPRRegClass;
6726 const TargetRegisterClass *TRC_Vec = 0;
6729 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6730 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6732 } else if (Align & 2) {
6733 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6734 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6737 // Check whether we can use NEON instructions.
6738 if (!MF->getFunction()->getAttributes().
6739 hasAttribute(AttributeSet::FunctionIndex,
6740 Attribute::NoImplicitFloat) &&
6741 Subtarget->hasNEON()) {
6742 if ((Align % 16 == 0) && SizeVal >= 16) {
6743 ldrOpc = ARM::VLD1q32wb_fixed;
6744 strOpc = ARM::VST1q32wb_fixed;
6746 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6748 else if ((Align % 8 == 0) && SizeVal >= 8) {
6749 ldrOpc = ARM::VLD1d32wb_fixed;
6750 strOpc = ARM::VST1d32wb_fixed;
6752 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6755 // Can't use NEON instructions.
6756 if (UnitSize == 0) {
6757 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6758 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6763 unsigned BytesLeft = SizeVal % UnitSize;
6764 unsigned LoopSize = SizeVal - BytesLeft;
6766 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6767 // Use LDR and STR to copy.
6768 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6769 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6770 unsigned srcIn = src;
6771 unsigned destIn = dest;
6772 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6773 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6774 unsigned srcOut = MRI.createVirtualRegister(TRC);
6775 unsigned destOut = MRI.createVirtualRegister(TRC);
6776 if (UnitSize >= 8) {
6777 AddDefaultPred(BuildMI(*BB, MI, dl,
6778 TII->get(ldrOpc), scratch)
6779 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6781 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6782 .addReg(destIn).addImm(0).addReg(scratch));
6783 } else if (isThumb2) {
6784 AddDefaultPred(BuildMI(*BB, MI, dl,
6785 TII->get(ldrOpc), scratch)
6786 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6788 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6789 .addReg(scratch).addReg(destIn)
6792 AddDefaultPred(BuildMI(*BB, MI, dl,
6793 TII->get(ldrOpc), scratch)
6794 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6797 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6798 .addReg(scratch).addReg(destIn)
6799 .addReg(0).addImm(UnitSize));
6805 // Handle the leftover bytes with LDRB and STRB.
6806 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6807 // [destOut] = STRB_POST(scratch, destIn, 1)
6808 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6809 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6810 for (unsigned i = 0; i < BytesLeft; i++) {
6811 unsigned scratch = MRI.createVirtualRegister(TRC);
6812 unsigned srcOut = MRI.createVirtualRegister(TRC);
6813 unsigned destOut = MRI.createVirtualRegister(TRC);
6815 AddDefaultPred(BuildMI(*BB, MI, dl,
6816 TII->get(ldrOpc),scratch)
6817 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6819 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6820 .addReg(scratch).addReg(destIn)
6821 .addReg(0).addImm(1));
6823 AddDefaultPred(BuildMI(*BB, MI, dl,
6824 TII->get(ldrOpc),scratch)
6825 .addReg(srcOut, RegState::Define).addReg(srcIn)
6826 .addReg(0).addImm(1));
6828 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6829 .addReg(scratch).addReg(destIn)
6830 .addReg(0).addImm(1));
6835 MI->eraseFromParent(); // The instruction is gone now.
6839 // Expand the pseudo op to a loop.
6842 // movw varEnd, # --> with thumb2
6844 // ldrcp varEnd, idx --> without thumb2
6845 // fallthrough --> loopMBB
6847 // PHI varPhi, varEnd, varLoop
6848 // PHI srcPhi, src, srcLoop
6849 // PHI destPhi, dst, destLoop
6850 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6851 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6852 // subs varLoop, varPhi, #UnitSize
6854 // fallthrough --> exitMBB
6856 // epilogue to handle left-over bytes
6857 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6858 // [destOut] = STRB_POST(scratch, destLoop, 1)
6859 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6860 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6861 MF->insert(It, loopMBB);
6862 MF->insert(It, exitMBB);
6864 // Transfer the remainder of BB and its successor edges to exitMBB.
6865 exitMBB->splice(exitMBB->begin(), BB,
6866 llvm::next(MachineBasicBlock::iterator(MI)),
6868 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6870 // Load an immediate to varEnd.
6871 unsigned varEnd = MRI.createVirtualRegister(TRC);
6873 unsigned VReg1 = varEnd;
6874 if ((LoopSize & 0xFFFF0000) != 0)
6875 VReg1 = MRI.createVirtualRegister(TRC);
6876 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6877 .addImm(LoopSize & 0xFFFF));
6879 if ((LoopSize & 0xFFFF0000) != 0)
6880 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6882 .addImm(LoopSize >> 16));
6884 MachineConstantPool *ConstantPool = MF->getConstantPool();
6885 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6886 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6888 // MachineConstantPool wants an explicit alignment.
6889 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6891 Align = getDataLayout()->getTypeAllocSize(C->getType());
6892 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6894 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6895 .addReg(varEnd, RegState::Define)
6896 .addConstantPoolIndex(Idx)
6899 BB->addSuccessor(loopMBB);
6901 // Generate the loop body:
6902 // varPhi = PHI(varLoop, varEnd)
6903 // srcPhi = PHI(srcLoop, src)
6904 // destPhi = PHI(destLoop, dst)
6905 MachineBasicBlock *entryBB = BB;
6907 unsigned varLoop = MRI.createVirtualRegister(TRC);
6908 unsigned varPhi = MRI.createVirtualRegister(TRC);
6909 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6910 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6911 unsigned destLoop = MRI.createVirtualRegister(TRC);
6912 unsigned destPhi = MRI.createVirtualRegister(TRC);
6914 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6915 .addReg(varLoop).addMBB(loopMBB)
6916 .addReg(varEnd).addMBB(entryBB);
6917 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6918 .addReg(srcLoop).addMBB(loopMBB)
6919 .addReg(src).addMBB(entryBB);
6920 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6921 .addReg(destLoop).addMBB(loopMBB)
6922 .addReg(dest).addMBB(entryBB);
6924 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6925 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6926 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6927 if (UnitSize >= 8) {
6928 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6929 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6931 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6932 .addReg(destPhi).addImm(0).addReg(scratch));
6933 } else if (isThumb2) {
6934 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6935 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6937 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6938 .addReg(scratch).addReg(destPhi)
6941 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6942 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6945 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6946 .addReg(scratch).addReg(destPhi)
6947 .addReg(0).addImm(UnitSize));
6950 // Decrement loop variable by UnitSize.
6951 MachineInstrBuilder MIB = BuildMI(BB, dl,
6952 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6953 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6954 MIB->getOperand(5).setReg(ARM::CPSR);
6955 MIB->getOperand(5).setIsDef(true);
6957 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6958 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6960 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6961 BB->addSuccessor(loopMBB);
6962 BB->addSuccessor(exitMBB);
6964 // Add epilogue to handle BytesLeft.
6966 MachineInstr *StartOfExit = exitMBB->begin();
6967 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6968 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6970 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6971 // [destOut] = STRB_POST(scratch, destLoop, 1)
6972 unsigned srcIn = srcLoop;
6973 unsigned destIn = destLoop;
6974 for (unsigned i = 0; i < BytesLeft; i++) {
6975 unsigned scratch = MRI.createVirtualRegister(TRC);
6976 unsigned srcOut = MRI.createVirtualRegister(TRC);
6977 unsigned destOut = MRI.createVirtualRegister(TRC);
6979 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6980 TII->get(ldrOpc),scratch)
6981 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6983 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6984 .addReg(scratch).addReg(destIn)
6987 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6988 TII->get(ldrOpc),scratch)
6989 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6991 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6992 .addReg(scratch).addReg(destIn)
6993 .addReg(0).addImm(1));
6999 MI->eraseFromParent(); // The instruction is gone now.
7004 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7005 MachineBasicBlock *BB) const {
7006 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7007 DebugLoc dl = MI->getDebugLoc();
7008 bool isThumb2 = Subtarget->isThumb2();
7009 switch (MI->getOpcode()) {
7012 llvm_unreachable("Unexpected instr type to insert");
7014 // The Thumb2 pre-indexed stores have the same MI operands, they just
7015 // define them differently in the .td files from the isel patterns, so
7016 // they need pseudos.
7017 case ARM::t2STR_preidx:
7018 MI->setDesc(TII->get(ARM::t2STR_PRE));
7020 case ARM::t2STRB_preidx:
7021 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7023 case ARM::t2STRH_preidx:
7024 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7027 case ARM::STRi_preidx:
7028 case ARM::STRBi_preidx: {
7029 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7030 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7031 // Decode the offset.
7032 unsigned Offset = MI->getOperand(4).getImm();
7033 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7034 Offset = ARM_AM::getAM2Offset(Offset);
7038 MachineMemOperand *MMO = *MI->memoperands_begin();
7039 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7040 .addOperand(MI->getOperand(0)) // Rn_wb
7041 .addOperand(MI->getOperand(1)) // Rt
7042 .addOperand(MI->getOperand(2)) // Rn
7043 .addImm(Offset) // offset (skip GPR==zero_reg)
7044 .addOperand(MI->getOperand(5)) // pred
7045 .addOperand(MI->getOperand(6))
7046 .addMemOperand(MMO);
7047 MI->eraseFromParent();
7050 case ARM::STRr_preidx:
7051 case ARM::STRBr_preidx:
7052 case ARM::STRH_preidx: {
7054 switch (MI->getOpcode()) {
7055 default: llvm_unreachable("unexpected opcode!");
7056 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7057 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7058 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7060 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7061 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7062 MIB.addOperand(MI->getOperand(i));
7063 MI->eraseFromParent();
7066 case ARM::ATOMIC_LOAD_ADD_I8:
7067 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7068 case ARM::ATOMIC_LOAD_ADD_I16:
7069 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7070 case ARM::ATOMIC_LOAD_ADD_I32:
7071 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
7073 case ARM::ATOMIC_LOAD_AND_I8:
7074 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7075 case ARM::ATOMIC_LOAD_AND_I16:
7076 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7077 case ARM::ATOMIC_LOAD_AND_I32:
7078 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7080 case ARM::ATOMIC_LOAD_OR_I8:
7081 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7082 case ARM::ATOMIC_LOAD_OR_I16:
7083 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7084 case ARM::ATOMIC_LOAD_OR_I32:
7085 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7087 case ARM::ATOMIC_LOAD_XOR_I8:
7088 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7089 case ARM::ATOMIC_LOAD_XOR_I16:
7090 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7091 case ARM::ATOMIC_LOAD_XOR_I32:
7092 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7094 case ARM::ATOMIC_LOAD_NAND_I8:
7095 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7096 case ARM::ATOMIC_LOAD_NAND_I16:
7097 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7098 case ARM::ATOMIC_LOAD_NAND_I32:
7099 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
7101 case ARM::ATOMIC_LOAD_SUB_I8:
7102 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7103 case ARM::ATOMIC_LOAD_SUB_I16:
7104 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7105 case ARM::ATOMIC_LOAD_SUB_I32:
7106 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
7108 case ARM::ATOMIC_LOAD_MIN_I8:
7109 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
7110 case ARM::ATOMIC_LOAD_MIN_I16:
7111 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
7112 case ARM::ATOMIC_LOAD_MIN_I32:
7113 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
7115 case ARM::ATOMIC_LOAD_MAX_I8:
7116 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
7117 case ARM::ATOMIC_LOAD_MAX_I16:
7118 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
7119 case ARM::ATOMIC_LOAD_MAX_I32:
7120 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
7122 case ARM::ATOMIC_LOAD_UMIN_I8:
7123 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
7124 case ARM::ATOMIC_LOAD_UMIN_I16:
7125 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
7126 case ARM::ATOMIC_LOAD_UMIN_I32:
7127 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
7129 case ARM::ATOMIC_LOAD_UMAX_I8:
7130 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
7131 case ARM::ATOMIC_LOAD_UMAX_I16:
7132 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
7133 case ARM::ATOMIC_LOAD_UMAX_I32:
7134 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
7136 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
7137 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
7138 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
7140 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
7141 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
7142 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
7145 case ARM::ATOMADD6432:
7146 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
7147 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
7148 /*NeedsCarry*/ true);
7149 case ARM::ATOMSUB6432:
7150 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7151 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7152 /*NeedsCarry*/ true);
7153 case ARM::ATOMOR6432:
7154 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
7155 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
7156 case ARM::ATOMXOR6432:
7157 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
7158 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
7159 case ARM::ATOMAND6432:
7160 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
7161 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
7162 case ARM::ATOMSWAP6432:
7163 return EmitAtomicBinary64(MI, BB, 0, 0, false);
7164 case ARM::ATOMCMPXCHG6432:
7165 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7166 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7167 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
7168 case ARM::ATOMMIN6432:
7169 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7170 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7171 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7172 /*IsMinMax*/ true, ARMCC::LT);
7173 case ARM::ATOMMAX6432:
7174 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7175 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7176 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7177 /*IsMinMax*/ true, ARMCC::GE);
7178 case ARM::ATOMUMIN6432:
7179 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7180 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7181 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7182 /*IsMinMax*/ true, ARMCC::LO);
7183 case ARM::ATOMUMAX6432:
7184 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
7185 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
7186 /*NeedsCarry*/ true, /*IsCmpxchg*/false,
7187 /*IsMinMax*/ true, ARMCC::HS);
7189 case ARM::tMOVCCr_pseudo: {
7190 // To "insert" a SELECT_CC instruction, we actually have to insert the
7191 // diamond control-flow pattern. The incoming instruction knows the
7192 // destination vreg to set, the condition code register to branch on, the
7193 // true/false values to select between, and a branch opcode to use.
7194 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7195 MachineFunction::iterator It = BB;
7201 // cmpTY ccX, r1, r2
7203 // fallthrough --> copy0MBB
7204 MachineBasicBlock *thisMBB = BB;
7205 MachineFunction *F = BB->getParent();
7206 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7207 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7208 F->insert(It, copy0MBB);
7209 F->insert(It, sinkMBB);
7211 // Transfer the remainder of BB and its successor edges to sinkMBB.
7212 sinkMBB->splice(sinkMBB->begin(), BB,
7213 llvm::next(MachineBasicBlock::iterator(MI)),
7215 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7217 BB->addSuccessor(copy0MBB);
7218 BB->addSuccessor(sinkMBB);
7220 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7221 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7224 // %FalseValue = ...
7225 // # fallthrough to sinkMBB
7228 // Update machine-CFG edges
7229 BB->addSuccessor(sinkMBB);
7232 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7235 BuildMI(*BB, BB->begin(), dl,
7236 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7237 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7238 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7240 MI->eraseFromParent(); // The pseudo instruction is gone now.
7245 case ARM::BCCZi64: {
7246 // If there is an unconditional branch to the other successor, remove it.
7247 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
7249 // Compare both parts that make up the double comparison separately for
7251 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7253 unsigned LHS1 = MI->getOperand(1).getReg();
7254 unsigned LHS2 = MI->getOperand(2).getReg();
7256 AddDefaultPred(BuildMI(BB, dl,
7257 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7258 .addReg(LHS1).addImm(0));
7259 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7260 .addReg(LHS2).addImm(0)
7261 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7263 unsigned RHS1 = MI->getOperand(3).getReg();
7264 unsigned RHS2 = MI->getOperand(4).getReg();
7265 AddDefaultPred(BuildMI(BB, dl,
7266 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7267 .addReg(LHS1).addReg(RHS1));
7268 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7269 .addReg(LHS2).addReg(RHS2)
7270 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7273 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7274 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7275 if (MI->getOperand(0).getImm() == ARMCC::NE)
7276 std::swap(destMBB, exitMBB);
7278 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7279 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7281 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7283 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7285 MI->eraseFromParent(); // The pseudo instruction is gone now.
7289 case ARM::Int_eh_sjlj_setjmp:
7290 case ARM::Int_eh_sjlj_setjmp_nofp:
7291 case ARM::tInt_eh_sjlj_setjmp:
7292 case ARM::t2Int_eh_sjlj_setjmp:
7293 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7294 EmitSjLjDispatchBlock(MI, BB);
7299 // To insert an ABS instruction, we have to insert the
7300 // diamond control-flow pattern. The incoming instruction knows the
7301 // source vreg to test against 0, the destination vreg to set,
7302 // the condition code register to branch on, the
7303 // true/false values to select between, and a branch opcode to use.
7308 // BCC (branch to SinkBB if V0 >= 0)
7309 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7310 // SinkBB: V1 = PHI(V2, V3)
7311 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7312 MachineFunction::iterator BBI = BB;
7314 MachineFunction *Fn = BB->getParent();
7315 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7316 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7317 Fn->insert(BBI, RSBBB);
7318 Fn->insert(BBI, SinkBB);
7320 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7321 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7322 bool isThumb2 = Subtarget->isThumb2();
7323 MachineRegisterInfo &MRI = Fn->getRegInfo();
7324 // In Thumb mode S must not be specified if source register is the SP or
7325 // PC and if destination register is the SP, so restrict register class
7326 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7327 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7328 (const TargetRegisterClass*)&ARM::GPRRegClass);
7330 // Transfer the remainder of BB and its successor edges to sinkMBB.
7331 SinkBB->splice(SinkBB->begin(), BB,
7332 llvm::next(MachineBasicBlock::iterator(MI)),
7334 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7336 BB->addSuccessor(RSBBB);
7337 BB->addSuccessor(SinkBB);
7339 // fall through to SinkMBB
7340 RSBBB->addSuccessor(SinkBB);
7342 // insert a cmp at the end of BB
7343 AddDefaultPred(BuildMI(BB, dl,
7344 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7345 .addReg(ABSSrcReg).addImm(0));
7347 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7349 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7350 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7352 // insert rsbri in RSBBB
7353 // Note: BCC and rsbri will be converted into predicated rsbmi
7354 // by if-conversion pass
7355 BuildMI(*RSBBB, RSBBB->begin(), dl,
7356 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7357 .addReg(ABSSrcReg, RegState::Kill)
7358 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7360 // insert PHI in SinkBB,
7361 // reuse ABSDstReg to not change uses of ABS instruction
7362 BuildMI(*SinkBB, SinkBB->begin(), dl,
7363 TII->get(ARM::PHI), ABSDstReg)
7364 .addReg(NewRsbDstReg).addMBB(RSBBB)
7365 .addReg(ABSSrcReg).addMBB(BB);
7367 // remove ABS instruction
7368 MI->eraseFromParent();
7370 // return last added BB
7373 case ARM::COPY_STRUCT_BYVAL_I32:
7375 return EmitStructByval(MI, BB);
7379 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7380 SDNode *Node) const {
7381 if (!MI->hasPostISelHook()) {
7382 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7383 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7387 const MCInstrDesc *MCID = &MI->getDesc();
7388 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7389 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7390 // operand is still set to noreg. If needed, set the optional operand's
7391 // register to CPSR, and remove the redundant implicit def.
7393 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7395 // Rename pseudo opcodes.
7396 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7398 const ARMBaseInstrInfo *TII =
7399 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7400 MCID = &TII->get(NewOpc);
7402 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7403 "converted opcode should be the same except for cc_out");
7407 // Add the optional cc_out operand
7408 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7410 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7412 // Any ARM instruction that sets the 's' bit should specify an optional
7413 // "cc_out" operand in the last operand position.
7414 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7415 assert(!NewOpc && "Optional cc_out operand required");
7418 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7419 // since we already have an optional CPSR def.
7420 bool definesCPSR = false;
7421 bool deadCPSR = false;
7422 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7424 const MachineOperand &MO = MI->getOperand(i);
7425 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7429 MI->RemoveOperand(i);
7434 assert(!NewOpc && "Optional cc_out operand required");
7437 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7439 assert(!MI->getOperand(ccOutIdx).getReg() &&
7440 "expect uninitialized optional cc_out operand");
7444 // If this instruction was defined with an optional CPSR def and its dag node
7445 // had a live implicit CPSR def, then activate the optional CPSR def.
7446 MachineOperand &MO = MI->getOperand(ccOutIdx);
7447 MO.setReg(ARM::CPSR);
7451 //===----------------------------------------------------------------------===//
7452 // ARM Optimization Hooks
7453 //===----------------------------------------------------------------------===//
7455 // Helper function that checks if N is a null or all ones constant.
7456 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7457 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7460 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7463 // Return true if N is conditionally 0 or all ones.
7464 // Detects these expressions where cc is an i1 value:
7466 // (select cc 0, y) [AllOnes=0]
7467 // (select cc y, 0) [AllOnes=0]
7468 // (zext cc) [AllOnes=0]
7469 // (sext cc) [AllOnes=0/1]
7470 // (select cc -1, y) [AllOnes=1]
7471 // (select cc y, -1) [AllOnes=1]
7473 // Invert is set when N is the null/all ones constant when CC is false.
7474 // OtherOp is set to the alternative value of N.
7475 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7476 SDValue &CC, bool &Invert,
7478 SelectionDAG &DAG) {
7479 switch (N->getOpcode()) {
7480 default: return false;
7482 CC = N->getOperand(0);
7483 SDValue N1 = N->getOperand(1);
7484 SDValue N2 = N->getOperand(2);
7485 if (isZeroOrAllOnes(N1, AllOnes)) {
7490 if (isZeroOrAllOnes(N2, AllOnes)) {
7497 case ISD::ZERO_EXTEND:
7498 // (zext cc) can never be the all ones value.
7502 case ISD::SIGN_EXTEND: {
7503 EVT VT = N->getValueType(0);
7504 CC = N->getOperand(0);
7505 if (CC.getValueType() != MVT::i1)
7509 // When looking for an AllOnes constant, N is an sext, and the 'other'
7511 OtherOp = DAG.getConstant(0, VT);
7512 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7513 // When looking for a 0 constant, N can be zext or sext.
7514 OtherOp = DAG.getConstant(1, VT);
7516 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7522 // Combine a constant select operand into its use:
7524 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7525 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7526 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7527 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7528 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7530 // The transform is rejected if the select doesn't have a constant operand that
7531 // is null, or all ones when AllOnes is set.
7533 // Also recognize sext/zext from i1:
7535 // (add (zext cc), x) -> (select cc (add x, 1), x)
7536 // (add (sext cc), x) -> (select cc (add x, -1), x)
7538 // These transformations eventually create predicated instructions.
7540 // @param N The node to transform.
7541 // @param Slct The N operand that is a select.
7542 // @param OtherOp The other N operand (x above).
7543 // @param DCI Context.
7544 // @param AllOnes Require the select constant to be all ones instead of null.
7545 // @returns The new node, or SDValue() on failure.
7547 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7548 TargetLowering::DAGCombinerInfo &DCI,
7549 bool AllOnes = false) {
7550 SelectionDAG &DAG = DCI.DAG;
7551 EVT VT = N->getValueType(0);
7552 SDValue NonConstantVal;
7555 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7556 NonConstantVal, DAG))
7559 // Slct is now know to be the desired identity constant when CC is true.
7560 SDValue TrueVal = OtherOp;
7561 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7562 OtherOp, NonConstantVal);
7563 // Unless SwapSelectOps says CC should be false.
7565 std::swap(TrueVal, FalseVal);
7567 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7568 CCOp, TrueVal, FalseVal);
7571 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7573 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7574 TargetLowering::DAGCombinerInfo &DCI) {
7575 SDValue N0 = N->getOperand(0);
7576 SDValue N1 = N->getOperand(1);
7577 if (N0.getNode()->hasOneUse()) {
7578 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7579 if (Result.getNode())
7582 if (N1.getNode()->hasOneUse()) {
7583 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7584 if (Result.getNode())
7590 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7591 // (only after legalization).
7592 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7593 TargetLowering::DAGCombinerInfo &DCI,
7594 const ARMSubtarget *Subtarget) {
7596 // Only perform optimization if after legalize, and if NEON is available. We
7597 // also expected both operands to be BUILD_VECTORs.
7598 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7599 || N0.getOpcode() != ISD::BUILD_VECTOR
7600 || N1.getOpcode() != ISD::BUILD_VECTOR)
7603 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7604 EVT VT = N->getValueType(0);
7605 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7608 // Check that the vector operands are of the right form.
7609 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7610 // operands, where N is the size of the formed vector.
7611 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7612 // index such that we have a pair wise add pattern.
7614 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7615 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7617 SDValue Vec = N0->getOperand(0)->getOperand(0);
7618 SDNode *V = Vec.getNode();
7619 unsigned nextIndex = 0;
7621 // For each operands to the ADD which are BUILD_VECTORs,
7622 // check to see if each of their operands are an EXTRACT_VECTOR with
7623 // the same vector and appropriate index.
7624 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7625 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7626 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7628 SDValue ExtVec0 = N0->getOperand(i);
7629 SDValue ExtVec1 = N1->getOperand(i);
7631 // First operand is the vector, verify its the same.
7632 if (V != ExtVec0->getOperand(0).getNode() ||
7633 V != ExtVec1->getOperand(0).getNode())
7636 // Second is the constant, verify its correct.
7637 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7638 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7640 // For the constant, we want to see all the even or all the odd.
7641 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7642 || C1->getZExtValue() != nextIndex+1)
7651 // Create VPADDL node.
7652 SelectionDAG &DAG = DCI.DAG;
7653 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7655 // Build operand list.
7656 SmallVector<SDValue, 8> Ops;
7657 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7658 TLI.getPointerTy()));
7660 // Input is the vector.
7663 // Get widened type and narrowed type.
7665 unsigned numElem = VT.getVectorNumElements();
7666 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7667 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7668 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7669 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7671 llvm_unreachable("Invalid vector element type for padd optimization.");
7674 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7675 widenType, &Ops[0], Ops.size());
7676 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7679 static SDValue findMUL_LOHI(SDValue V) {
7680 if (V->getOpcode() == ISD::UMUL_LOHI ||
7681 V->getOpcode() == ISD::SMUL_LOHI)
7686 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7687 TargetLowering::DAGCombinerInfo &DCI,
7688 const ARMSubtarget *Subtarget) {
7690 if (Subtarget->isThumb1Only()) return SDValue();
7692 // Only perform the checks after legalize when the pattern is available.
7693 if (DCI.isBeforeLegalize()) return SDValue();
7695 // Look for multiply add opportunities.
7696 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7697 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7698 // a glue link from the first add to the second add.
7699 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7700 // a S/UMLAL instruction.
7703 // \ / \ [no multiline comment]
7709 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7710 SDValue AddcOp0 = AddcNode->getOperand(0);
7711 SDValue AddcOp1 = AddcNode->getOperand(1);
7713 // Check if the two operands are from the same mul_lohi node.
7714 if (AddcOp0.getNode() == AddcOp1.getNode())
7717 assert(AddcNode->getNumValues() == 2 &&
7718 AddcNode->getValueType(0) == MVT::i32 &&
7719 AddcNode->getValueType(1) == MVT::Glue &&
7720 "Expect ADDC with two result values: i32, glue");
7722 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7723 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7724 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7725 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7726 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7729 // Look for the glued ADDE.
7730 SDNode* AddeNode = AddcNode->getGluedUser();
7731 if (AddeNode == NULL)
7734 // Make sure it is really an ADDE.
7735 if (AddeNode->getOpcode() != ISD::ADDE)
7738 assert(AddeNode->getNumOperands() == 3 &&
7739 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7740 "ADDE node has the wrong inputs");
7742 // Check for the triangle shape.
7743 SDValue AddeOp0 = AddeNode->getOperand(0);
7744 SDValue AddeOp1 = AddeNode->getOperand(1);
7746 // Make sure that the ADDE operands are not coming from the same node.
7747 if (AddeOp0.getNode() == AddeOp1.getNode())
7750 // Find the MUL_LOHI node walking up ADDE's operands.
7751 bool IsLeftOperandMUL = false;
7752 SDValue MULOp = findMUL_LOHI(AddeOp0);
7753 if (MULOp == SDValue())
7754 MULOp = findMUL_LOHI(AddeOp1);
7756 IsLeftOperandMUL = true;
7757 if (MULOp == SDValue())
7760 // Figure out the right opcode.
7761 unsigned Opc = MULOp->getOpcode();
7762 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7764 // Figure out the high and low input values to the MLAL node.
7765 SDValue* HiMul = &MULOp;
7766 SDValue* HiAdd = NULL;
7767 SDValue* LoMul = NULL;
7768 SDValue* LowAdd = NULL;
7770 if (IsLeftOperandMUL)
7776 if (AddcOp0->getOpcode() == Opc) {
7780 if (AddcOp1->getOpcode() == Opc) {
7788 if (LoMul->getNode() != HiMul->getNode())
7791 // Create the merged node.
7792 SelectionDAG &DAG = DCI.DAG;
7794 // Build operand list.
7795 SmallVector<SDValue, 8> Ops;
7796 Ops.push_back(LoMul->getOperand(0));
7797 Ops.push_back(LoMul->getOperand(1));
7798 Ops.push_back(*LowAdd);
7799 Ops.push_back(*HiAdd);
7801 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7802 DAG.getVTList(MVT::i32, MVT::i32),
7803 &Ops[0], Ops.size());
7805 // Replace the ADDs' nodes uses by the MLA node's values.
7806 SDValue HiMLALResult(MLALNode.getNode(), 1);
7807 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7809 SDValue LoMLALResult(MLALNode.getNode(), 0);
7810 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7812 // Return original node to notify the driver to stop replacing.
7813 SDValue resNode(AddcNode, 0);
7817 /// PerformADDCCombine - Target-specific dag combine transform from
7818 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7819 static SDValue PerformADDCCombine(SDNode *N,
7820 TargetLowering::DAGCombinerInfo &DCI,
7821 const ARMSubtarget *Subtarget) {
7823 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7827 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7828 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7829 /// called with the default operands, and if that fails, with commuted
7831 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7832 TargetLowering::DAGCombinerInfo &DCI,
7833 const ARMSubtarget *Subtarget){
7835 // Attempt to create vpaddl for this add.
7836 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7837 if (Result.getNode())
7840 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7841 if (N0.getNode()->hasOneUse()) {
7842 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7843 if (Result.getNode()) return Result;
7848 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7850 static SDValue PerformADDCombine(SDNode *N,
7851 TargetLowering::DAGCombinerInfo &DCI,
7852 const ARMSubtarget *Subtarget) {
7853 SDValue N0 = N->getOperand(0);
7854 SDValue N1 = N->getOperand(1);
7856 // First try with the default operand order.
7857 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7858 if (Result.getNode())
7861 // If that didn't work, try again with the operands commuted.
7862 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7865 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7867 static SDValue PerformSUBCombine(SDNode *N,
7868 TargetLowering::DAGCombinerInfo &DCI) {
7869 SDValue N0 = N->getOperand(0);
7870 SDValue N1 = N->getOperand(1);
7872 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7873 if (N1.getNode()->hasOneUse()) {
7874 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7875 if (Result.getNode()) return Result;
7881 /// PerformVMULCombine
7882 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7883 /// special multiplier accumulator forwarding.
7889 static SDValue PerformVMULCombine(SDNode *N,
7890 TargetLowering::DAGCombinerInfo &DCI,
7891 const ARMSubtarget *Subtarget) {
7892 if (!Subtarget->hasVMLxForwarding())
7895 SelectionDAG &DAG = DCI.DAG;
7896 SDValue N0 = N->getOperand(0);
7897 SDValue N1 = N->getOperand(1);
7898 unsigned Opcode = N0.getOpcode();
7899 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7900 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7901 Opcode = N1.getOpcode();
7902 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7903 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7908 EVT VT = N->getValueType(0);
7909 DebugLoc DL = N->getDebugLoc();
7910 SDValue N00 = N0->getOperand(0);
7911 SDValue N01 = N0->getOperand(1);
7912 return DAG.getNode(Opcode, DL, VT,
7913 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7914 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7917 static SDValue PerformMULCombine(SDNode *N,
7918 TargetLowering::DAGCombinerInfo &DCI,
7919 const ARMSubtarget *Subtarget) {
7920 SelectionDAG &DAG = DCI.DAG;
7922 if (Subtarget->isThumb1Only())
7925 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7928 EVT VT = N->getValueType(0);
7929 if (VT.is64BitVector() || VT.is128BitVector())
7930 return PerformVMULCombine(N, DCI, Subtarget);
7934 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7938 int64_t MulAmt = C->getSExtValue();
7939 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7941 ShiftAmt = ShiftAmt & (32 - 1);
7942 SDValue V = N->getOperand(0);
7943 DebugLoc DL = N->getDebugLoc();
7946 MulAmt >>= ShiftAmt;
7949 if (isPowerOf2_32(MulAmt - 1)) {
7950 // (mul x, 2^N + 1) => (add (shl x, N), x)
7951 Res = DAG.getNode(ISD::ADD, DL, VT,
7953 DAG.getNode(ISD::SHL, DL, VT,
7955 DAG.getConstant(Log2_32(MulAmt - 1),
7957 } else if (isPowerOf2_32(MulAmt + 1)) {
7958 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7959 Res = DAG.getNode(ISD::SUB, DL, VT,
7960 DAG.getNode(ISD::SHL, DL, VT,
7962 DAG.getConstant(Log2_32(MulAmt + 1),
7968 uint64_t MulAmtAbs = -MulAmt;
7969 if (isPowerOf2_32(MulAmtAbs + 1)) {
7970 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7971 Res = DAG.getNode(ISD::SUB, DL, VT,
7973 DAG.getNode(ISD::SHL, DL, VT,
7975 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7977 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7978 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7979 Res = DAG.getNode(ISD::ADD, DL, VT,
7981 DAG.getNode(ISD::SHL, DL, VT,
7983 DAG.getConstant(Log2_32(MulAmtAbs-1),
7985 Res = DAG.getNode(ISD::SUB, DL, VT,
7986 DAG.getConstant(0, MVT::i32),Res);
7993 Res = DAG.getNode(ISD::SHL, DL, VT,
7994 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7996 // Do not add new nodes to DAG combiner worklist.
7997 DCI.CombineTo(N, Res, false);
8001 static SDValue PerformANDCombine(SDNode *N,
8002 TargetLowering::DAGCombinerInfo &DCI,
8003 const ARMSubtarget *Subtarget) {
8005 // Attempt to use immediate-form VBIC
8006 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8007 DebugLoc dl = N->getDebugLoc();
8008 EVT VT = N->getValueType(0);
8009 SelectionDAG &DAG = DCI.DAG;
8011 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8014 APInt SplatBits, SplatUndef;
8015 unsigned SplatBitSize;
8018 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8019 if (SplatBitSize <= 64) {
8021 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8022 SplatUndef.getZExtValue(), SplatBitSize,
8023 DAG, VbicVT, VT.is128BitVector(),
8025 if (Val.getNode()) {
8027 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8028 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8029 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8034 if (!Subtarget->isThumb1Only()) {
8035 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8036 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8037 if (Result.getNode())
8044 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8045 static SDValue PerformORCombine(SDNode *N,
8046 TargetLowering::DAGCombinerInfo &DCI,
8047 const ARMSubtarget *Subtarget) {
8048 // Attempt to use immediate-form VORR
8049 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8050 DebugLoc dl = N->getDebugLoc();
8051 EVT VT = N->getValueType(0);
8052 SelectionDAG &DAG = DCI.DAG;
8054 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8057 APInt SplatBits, SplatUndef;
8058 unsigned SplatBitSize;
8060 if (BVN && Subtarget->hasNEON() &&
8061 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8062 if (SplatBitSize <= 64) {
8064 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8065 SplatUndef.getZExtValue(), SplatBitSize,
8066 DAG, VorrVT, VT.is128BitVector(),
8068 if (Val.getNode()) {
8070 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8071 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8072 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8077 if (!Subtarget->isThumb1Only()) {
8078 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8079 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8080 if (Result.getNode())
8084 // The code below optimizes (or (and X, Y), Z).
8085 // The AND operand needs to have a single user to make these optimizations
8087 SDValue N0 = N->getOperand(0);
8088 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8090 SDValue N1 = N->getOperand(1);
8092 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8093 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8094 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8096 unsigned SplatBitSize;
8099 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8101 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8102 HasAnyUndefs) && !HasAnyUndefs) {
8103 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8105 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8106 HasAnyUndefs) && !HasAnyUndefs &&
8107 SplatBits0 == ~SplatBits1) {
8108 // Canonicalize the vector type to make instruction selection simpler.
8109 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8110 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8111 N0->getOperand(1), N0->getOperand(0),
8113 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8118 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8121 // BFI is only available on V6T2+
8122 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8125 DebugLoc DL = N->getDebugLoc();
8126 // 1) or (and A, mask), val => ARMbfi A, val, mask
8127 // iff (val & mask) == val
8129 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8130 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8131 // && mask == ~mask2
8132 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8133 // && ~mask == mask2
8134 // (i.e., copy a bitfield value into another bitfield of the same width)
8139 SDValue N00 = N0.getOperand(0);
8141 // The value and the mask need to be constants so we can verify this is
8142 // actually a bitfield set. If the mask is 0xffff, we can do better
8143 // via a movt instruction, so don't use BFI in that case.
8144 SDValue MaskOp = N0.getOperand(1);
8145 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8148 unsigned Mask = MaskC->getZExtValue();
8152 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8153 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8155 unsigned Val = N1C->getZExtValue();
8156 if ((Val & ~Mask) != Val)
8159 if (ARM::isBitFieldInvertedMask(Mask)) {
8160 Val >>= CountTrailingZeros_32(~Mask);
8162 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8163 DAG.getConstant(Val, MVT::i32),
8164 DAG.getConstant(Mask, MVT::i32));
8166 // Do not add new nodes to DAG combiner worklist.
8167 DCI.CombineTo(N, Res, false);
8170 } else if (N1.getOpcode() == ISD::AND) {
8171 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8172 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8175 unsigned Mask2 = N11C->getZExtValue();
8177 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8179 if (ARM::isBitFieldInvertedMask(Mask) &&
8181 // The pack halfword instruction works better for masks that fit it,
8182 // so use that when it's available.
8183 if (Subtarget->hasT2ExtractPack() &&
8184 (Mask == 0xffff || Mask == 0xffff0000))
8187 unsigned amt = CountTrailingZeros_32(Mask2);
8188 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8189 DAG.getConstant(amt, MVT::i32));
8190 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8191 DAG.getConstant(Mask, MVT::i32));
8192 // Do not add new nodes to DAG combiner worklist.
8193 DCI.CombineTo(N, Res, false);
8195 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8197 // The pack halfword instruction works better for masks that fit it,
8198 // so use that when it's available.
8199 if (Subtarget->hasT2ExtractPack() &&
8200 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8203 unsigned lsb = CountTrailingZeros_32(Mask);
8204 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8205 DAG.getConstant(lsb, MVT::i32));
8206 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8207 DAG.getConstant(Mask2, MVT::i32));
8208 // Do not add new nodes to DAG combiner worklist.
8209 DCI.CombineTo(N, Res, false);
8214 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8215 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8216 ARM::isBitFieldInvertedMask(~Mask)) {
8217 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8218 // where lsb(mask) == #shamt and masked bits of B are known zero.
8219 SDValue ShAmt = N00.getOperand(1);
8220 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8221 unsigned LSB = CountTrailingZeros_32(Mask);
8225 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8226 DAG.getConstant(~Mask, MVT::i32));
8228 // Do not add new nodes to DAG combiner worklist.
8229 DCI.CombineTo(N, Res, false);
8235 static SDValue PerformXORCombine(SDNode *N,
8236 TargetLowering::DAGCombinerInfo &DCI,
8237 const ARMSubtarget *Subtarget) {
8238 EVT VT = N->getValueType(0);
8239 SelectionDAG &DAG = DCI.DAG;
8241 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8244 if (!Subtarget->isThumb1Only()) {
8245 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8246 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8247 if (Result.getNode())
8254 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8255 /// the bits being cleared by the AND are not demanded by the BFI.
8256 static SDValue PerformBFICombine(SDNode *N,
8257 TargetLowering::DAGCombinerInfo &DCI) {
8258 SDValue N1 = N->getOperand(1);
8259 if (N1.getOpcode() == ISD::AND) {
8260 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8263 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8264 unsigned LSB = CountTrailingZeros_32(~InvMask);
8265 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
8266 unsigned Mask = (1 << Width)-1;
8267 unsigned Mask2 = N11C->getZExtValue();
8268 if ((Mask & (~Mask2)) == 0)
8269 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
8270 N->getOperand(0), N1.getOperand(0),
8276 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8277 /// ARMISD::VMOVRRD.
8278 static SDValue PerformVMOVRRDCombine(SDNode *N,
8279 TargetLowering::DAGCombinerInfo &DCI) {
8280 // vmovrrd(vmovdrr x, y) -> x,y
8281 SDValue InDouble = N->getOperand(0);
8282 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8283 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8285 // vmovrrd(load f64) -> (load i32), (load i32)
8286 SDNode *InNode = InDouble.getNode();
8287 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8288 InNode->getValueType(0) == MVT::f64 &&
8289 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8290 !cast<LoadSDNode>(InNode)->isVolatile()) {
8291 // TODO: Should this be done for non-FrameIndex operands?
8292 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8294 SelectionDAG &DAG = DCI.DAG;
8295 DebugLoc DL = LD->getDebugLoc();
8296 SDValue BasePtr = LD->getBasePtr();
8297 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8298 LD->getPointerInfo(), LD->isVolatile(),
8299 LD->isNonTemporal(), LD->isInvariant(),
8300 LD->getAlignment());
8302 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8303 DAG.getConstant(4, MVT::i32));
8304 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8305 LD->getPointerInfo(), LD->isVolatile(),
8306 LD->isNonTemporal(), LD->isInvariant(),
8307 std::min(4U, LD->getAlignment() / 2));
8309 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8310 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8311 DCI.RemoveFromWorklist(LD);
8319 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8320 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8321 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8322 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8323 SDValue Op0 = N->getOperand(0);
8324 SDValue Op1 = N->getOperand(1);
8325 if (Op0.getOpcode() == ISD::BITCAST)
8326 Op0 = Op0.getOperand(0);
8327 if (Op1.getOpcode() == ISD::BITCAST)
8328 Op1 = Op1.getOperand(0);
8329 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8330 Op0.getNode() == Op1.getNode() &&
8331 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8332 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
8333 N->getValueType(0), Op0.getOperand(0));
8337 /// PerformSTORECombine - Target-specific dag combine xforms for
8339 static SDValue PerformSTORECombine(SDNode *N,
8340 TargetLowering::DAGCombinerInfo &DCI) {
8341 StoreSDNode *St = cast<StoreSDNode>(N);
8342 if (St->isVolatile())
8345 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8346 // pack all of the elements in one place. Next, store to memory in fewer
8348 SDValue StVal = St->getValue();
8349 EVT VT = StVal.getValueType();
8350 if (St->isTruncatingStore() && VT.isVector()) {
8351 SelectionDAG &DAG = DCI.DAG;
8352 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8353 EVT StVT = St->getMemoryVT();
8354 unsigned NumElems = VT.getVectorNumElements();
8355 assert(StVT != VT && "Cannot truncate to the same type");
8356 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8357 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8359 // From, To sizes and ElemCount must be pow of two
8360 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8362 // We are going to use the original vector elt for storing.
8363 // Accumulated smaller vector elements must be a multiple of the store size.
8364 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8366 unsigned SizeRatio = FromEltSz / ToEltSz;
8367 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8369 // Create a type on which we perform the shuffle.
8370 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8371 NumElems*SizeRatio);
8372 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8374 DebugLoc DL = St->getDebugLoc();
8375 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8376 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8377 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
8379 // Can't shuffle using an illegal type.
8380 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8382 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8383 DAG.getUNDEF(WideVec.getValueType()),
8385 // At this point all of the data is stored at the bottom of the
8386 // register. We now need to save it to mem.
8388 // Find the largest store unit
8389 MVT StoreType = MVT::i8;
8390 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8391 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8392 MVT Tp = (MVT::SimpleValueType)tp;
8393 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8396 // Didn't find a legal store type.
8397 if (!TLI.isTypeLegal(StoreType))
8400 // Bitcast the original vector into a vector of store-size units
8401 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8402 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8403 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8404 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8405 SmallVector<SDValue, 8> Chains;
8406 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8407 TLI.getPointerTy());
8408 SDValue BasePtr = St->getBasePtr();
8410 // Perform one or more big stores into memory.
8411 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8412 for (unsigned I = 0; I < E; I++) {
8413 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8414 StoreType, ShuffWide,
8415 DAG.getIntPtrConstant(I));
8416 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8417 St->getPointerInfo(), St->isVolatile(),
8418 St->isNonTemporal(), St->getAlignment());
8419 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8421 Chains.push_back(Ch);
8423 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8427 if (!ISD::isNormalStore(St))
8430 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8431 // ARM stores of arguments in the same cache line.
8432 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8433 StVal.getNode()->hasOneUse()) {
8434 SelectionDAG &DAG = DCI.DAG;
8435 DebugLoc DL = St->getDebugLoc();
8436 SDValue BasePtr = St->getBasePtr();
8437 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8438 StVal.getNode()->getOperand(0), BasePtr,
8439 St->getPointerInfo(), St->isVolatile(),
8440 St->isNonTemporal(), St->getAlignment());
8442 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8443 DAG.getConstant(4, MVT::i32));
8444 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8445 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8446 St->isNonTemporal(),
8447 std::min(4U, St->getAlignment() / 2));
8450 if (StVal.getValueType() != MVT::i64 ||
8451 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8454 // Bitcast an i64 store extracted from a vector to f64.
8455 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8456 SelectionDAG &DAG = DCI.DAG;
8457 DebugLoc dl = StVal.getDebugLoc();
8458 SDValue IntVec = StVal.getOperand(0);
8459 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8460 IntVec.getValueType().getVectorNumElements());
8461 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8462 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8463 Vec, StVal.getOperand(1));
8464 dl = N->getDebugLoc();
8465 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8466 // Make the DAGCombiner fold the bitcasts.
8467 DCI.AddToWorklist(Vec.getNode());
8468 DCI.AddToWorklist(ExtElt.getNode());
8469 DCI.AddToWorklist(V.getNode());
8470 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8471 St->getPointerInfo(), St->isVolatile(),
8472 St->isNonTemporal(), St->getAlignment(),
8476 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8477 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8478 /// i64 vector to have f64 elements, since the value can then be loaded
8479 /// directly into a VFP register.
8480 static bool hasNormalLoadOperand(SDNode *N) {
8481 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8482 for (unsigned i = 0; i < NumElts; ++i) {
8483 SDNode *Elt = N->getOperand(i).getNode();
8484 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8490 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8491 /// ISD::BUILD_VECTOR.
8492 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8493 TargetLowering::DAGCombinerInfo &DCI){
8494 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8495 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8496 // into a pair of GPRs, which is fine when the value is used as a scalar,
8497 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8498 SelectionDAG &DAG = DCI.DAG;
8499 if (N->getNumOperands() == 2) {
8500 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8505 // Load i64 elements as f64 values so that type legalization does not split
8506 // them up into i32 values.
8507 EVT VT = N->getValueType(0);
8508 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8510 DebugLoc dl = N->getDebugLoc();
8511 SmallVector<SDValue, 8> Ops;
8512 unsigned NumElts = VT.getVectorNumElements();
8513 for (unsigned i = 0; i < NumElts; ++i) {
8514 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8516 // Make the DAGCombiner fold the bitcast.
8517 DCI.AddToWorklist(V.getNode());
8519 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8520 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8521 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8524 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8525 /// ISD::INSERT_VECTOR_ELT.
8526 static SDValue PerformInsertEltCombine(SDNode *N,
8527 TargetLowering::DAGCombinerInfo &DCI) {
8528 // Bitcast an i64 load inserted into a vector to f64.
8529 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8530 EVT VT = N->getValueType(0);
8531 SDNode *Elt = N->getOperand(1).getNode();
8532 if (VT.getVectorElementType() != MVT::i64 ||
8533 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8536 SelectionDAG &DAG = DCI.DAG;
8537 DebugLoc dl = N->getDebugLoc();
8538 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8539 VT.getVectorNumElements());
8540 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8541 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8542 // Make the DAGCombiner fold the bitcasts.
8543 DCI.AddToWorklist(Vec.getNode());
8544 DCI.AddToWorklist(V.getNode());
8545 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8546 Vec, V, N->getOperand(2));
8547 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8550 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8551 /// ISD::VECTOR_SHUFFLE.
8552 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8553 // The LLVM shufflevector instruction does not require the shuffle mask
8554 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8555 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8556 // operands do not match the mask length, they are extended by concatenating
8557 // them with undef vectors. That is probably the right thing for other
8558 // targets, but for NEON it is better to concatenate two double-register
8559 // size vector operands into a single quad-register size vector. Do that
8560 // transformation here:
8561 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8562 // shuffle(concat(v1, v2), undef)
8563 SDValue Op0 = N->getOperand(0);
8564 SDValue Op1 = N->getOperand(1);
8565 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8566 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8567 Op0.getNumOperands() != 2 ||
8568 Op1.getNumOperands() != 2)
8570 SDValue Concat0Op1 = Op0.getOperand(1);
8571 SDValue Concat1Op1 = Op1.getOperand(1);
8572 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8573 Concat1Op1.getOpcode() != ISD::UNDEF)
8575 // Skip the transformation if any of the types are illegal.
8576 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8577 EVT VT = N->getValueType(0);
8578 if (!TLI.isTypeLegal(VT) ||
8579 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8580 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8583 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8584 Op0.getOperand(0), Op1.getOperand(0));
8585 // Translate the shuffle mask.
8586 SmallVector<int, 16> NewMask;
8587 unsigned NumElts = VT.getVectorNumElements();
8588 unsigned HalfElts = NumElts/2;
8589 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8590 for (unsigned n = 0; n < NumElts; ++n) {
8591 int MaskElt = SVN->getMaskElt(n);
8593 if (MaskElt < (int)HalfElts)
8595 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8596 NewElt = HalfElts + MaskElt - NumElts;
8597 NewMask.push_back(NewElt);
8599 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8600 DAG.getUNDEF(VT), NewMask.data());
8603 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8604 /// NEON load/store intrinsics to merge base address updates.
8605 static SDValue CombineBaseUpdate(SDNode *N,
8606 TargetLowering::DAGCombinerInfo &DCI) {
8607 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8610 SelectionDAG &DAG = DCI.DAG;
8611 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8612 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8613 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8614 SDValue Addr = N->getOperand(AddrOpIdx);
8616 // Search for a use of the address operand that is an increment.
8617 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8618 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8620 if (User->getOpcode() != ISD::ADD ||
8621 UI.getUse().getResNo() != Addr.getResNo())
8624 // Check that the add is independent of the load/store. Otherwise, folding
8625 // it would create a cycle.
8626 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8629 // Find the new opcode for the updating load/store.
8631 bool isLaneOp = false;
8632 unsigned NewOpc = 0;
8633 unsigned NumVecs = 0;
8635 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8637 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8638 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8640 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8642 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8644 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8646 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8647 NumVecs = 2; isLaneOp = true; break;
8648 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8649 NumVecs = 3; isLaneOp = true; break;
8650 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8651 NumVecs = 4; isLaneOp = true; break;
8652 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8653 NumVecs = 1; isLoad = false; break;
8654 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8655 NumVecs = 2; isLoad = false; break;
8656 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8657 NumVecs = 3; isLoad = false; break;
8658 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8659 NumVecs = 4; isLoad = false; break;
8660 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8661 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8662 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8663 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8664 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8665 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8669 switch (N->getOpcode()) {
8670 default: llvm_unreachable("unexpected opcode for Neon base update");
8671 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8672 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8673 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8677 // Find the size of memory referenced by the load/store.
8680 VecTy = N->getValueType(0);
8682 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8683 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8685 NumBytes /= VecTy.getVectorNumElements();
8687 // If the increment is a constant, it must match the memory ref size.
8688 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8689 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8690 uint64_t IncVal = CInc->getZExtValue();
8691 if (IncVal != NumBytes)
8693 } else if (NumBytes >= 3 * 16) {
8694 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8695 // separate instructions that make it harder to use a non-constant update.
8699 // Create the new updating load/store node.
8701 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8703 for (n = 0; n < NumResultVecs; ++n)
8705 Tys[n++] = MVT::i32;
8706 Tys[n] = MVT::Other;
8707 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8708 SmallVector<SDValue, 8> Ops;
8709 Ops.push_back(N->getOperand(0)); // incoming chain
8710 Ops.push_back(N->getOperand(AddrOpIdx));
8712 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8713 Ops.push_back(N->getOperand(i));
8715 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8716 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8717 Ops.data(), Ops.size(),
8718 MemInt->getMemoryVT(),
8719 MemInt->getMemOperand());
8722 std::vector<SDValue> NewResults;
8723 for (unsigned i = 0; i < NumResultVecs; ++i) {
8724 NewResults.push_back(SDValue(UpdN.getNode(), i));
8726 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8727 DCI.CombineTo(N, NewResults);
8728 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8735 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8736 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8737 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8739 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8740 SelectionDAG &DAG = DCI.DAG;
8741 EVT VT = N->getValueType(0);
8742 // vldN-dup instructions only support 64-bit vectors for N > 1.
8743 if (!VT.is64BitVector())
8746 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8747 SDNode *VLD = N->getOperand(0).getNode();
8748 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8750 unsigned NumVecs = 0;
8751 unsigned NewOpc = 0;
8752 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8753 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8755 NewOpc = ARMISD::VLD2DUP;
8756 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8758 NewOpc = ARMISD::VLD3DUP;
8759 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8761 NewOpc = ARMISD::VLD4DUP;
8766 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8767 // numbers match the load.
8768 unsigned VLDLaneNo =
8769 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8770 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8772 // Ignore uses of the chain result.
8773 if (UI.getUse().getResNo() == NumVecs)
8776 if (User->getOpcode() != ARMISD::VDUPLANE ||
8777 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8781 // Create the vldN-dup node.
8784 for (n = 0; n < NumVecs; ++n)
8786 Tys[n] = MVT::Other;
8787 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8788 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8789 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8790 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8791 Ops, 2, VLDMemInt->getMemoryVT(),
8792 VLDMemInt->getMemOperand());
8795 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8797 unsigned ResNo = UI.getUse().getResNo();
8798 // Ignore uses of the chain result.
8799 if (ResNo == NumVecs)
8802 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8805 // Now the vldN-lane intrinsic is dead except for its chain result.
8806 // Update uses of the chain.
8807 std::vector<SDValue> VLDDupResults;
8808 for (unsigned n = 0; n < NumVecs; ++n)
8809 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8810 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8811 DCI.CombineTo(VLD, VLDDupResults);
8816 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8817 /// ARMISD::VDUPLANE.
8818 static SDValue PerformVDUPLANECombine(SDNode *N,
8819 TargetLowering::DAGCombinerInfo &DCI) {
8820 SDValue Op = N->getOperand(0);
8822 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8823 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8824 if (CombineVLDDUP(N, DCI))
8825 return SDValue(N, 0);
8827 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8828 // redundant. Ignore bit_converts for now; element sizes are checked below.
8829 while (Op.getOpcode() == ISD::BITCAST)
8830 Op = Op.getOperand(0);
8831 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8834 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8835 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8836 // The canonical VMOV for a zero vector uses a 32-bit element size.
8837 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8839 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8841 EVT VT = N->getValueType(0);
8842 if (EltSize > VT.getVectorElementType().getSizeInBits())
8845 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8848 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8849 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8850 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8854 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8856 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8861 APFloat APF = C->getValueAPF();
8862 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8863 != APFloat::opOK || !isExact)
8866 c0 = (I == 0) ? cN : c0;
8867 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8874 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8875 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8876 /// when the VMUL has a constant operand that is a power of 2.
8878 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8879 /// vmul.f32 d16, d17, d16
8880 /// vcvt.s32.f32 d16, d16
8882 /// vcvt.s32.f32 d16, d16, #3
8883 static SDValue PerformVCVTCombine(SDNode *N,
8884 TargetLowering::DAGCombinerInfo &DCI,
8885 const ARMSubtarget *Subtarget) {
8886 SelectionDAG &DAG = DCI.DAG;
8887 SDValue Op = N->getOperand(0);
8889 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8890 Op.getOpcode() != ISD::FMUL)
8894 SDValue N0 = Op->getOperand(0);
8895 SDValue ConstVec = Op->getOperand(1);
8896 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8898 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8899 !isConstVecPow2(ConstVec, isSigned, C))
8902 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8903 Intrinsic::arm_neon_vcvtfp2fxu;
8904 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8906 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8907 DAG.getConstant(Log2_64(C), MVT::i32));
8910 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8911 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8912 /// when the VDIV has a constant operand that is a power of 2.
8914 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8915 /// vcvt.f32.s32 d16, d16
8916 /// vdiv.f32 d16, d17, d16
8918 /// vcvt.f32.s32 d16, d16, #3
8919 static SDValue PerformVDIVCombine(SDNode *N,
8920 TargetLowering::DAGCombinerInfo &DCI,
8921 const ARMSubtarget *Subtarget) {
8922 SelectionDAG &DAG = DCI.DAG;
8923 SDValue Op = N->getOperand(0);
8924 unsigned OpOpcode = Op.getNode()->getOpcode();
8926 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8927 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8931 SDValue ConstVec = N->getOperand(1);
8932 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8934 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8935 !isConstVecPow2(ConstVec, isSigned, C))
8938 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8939 Intrinsic::arm_neon_vcvtfxu2fp;
8940 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8942 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8943 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8946 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8947 /// operand of a vector shift operation, where all the elements of the
8948 /// build_vector must have the same constant integer value.
8949 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8950 // Ignore bit_converts.
8951 while (Op.getOpcode() == ISD::BITCAST)
8952 Op = Op.getOperand(0);
8953 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8954 APInt SplatBits, SplatUndef;
8955 unsigned SplatBitSize;
8957 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8958 HasAnyUndefs, ElementBits) ||
8959 SplatBitSize > ElementBits)
8961 Cnt = SplatBits.getSExtValue();
8965 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8966 /// operand of a vector shift left operation. That value must be in the range:
8967 /// 0 <= Value < ElementBits for a left shift; or
8968 /// 0 <= Value <= ElementBits for a long left shift.
8969 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8970 assert(VT.isVector() && "vector shift count is not a vector type");
8971 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8972 if (! getVShiftImm(Op, ElementBits, Cnt))
8974 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8977 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8978 /// operand of a vector shift right operation. For a shift opcode, the value
8979 /// is positive, but for an intrinsic the value count must be negative. The
8980 /// absolute value must be in the range:
8981 /// 1 <= |Value| <= ElementBits for a right shift; or
8982 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8983 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8985 assert(VT.isVector() && "vector shift count is not a vector type");
8986 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8987 if (! getVShiftImm(Op, ElementBits, Cnt))
8991 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8994 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8995 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8996 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8999 // Don't do anything for most intrinsics.
9002 // Vector shifts: check for immediate versions and lower them.
9003 // Note: This is done during DAG combining instead of DAG legalizing because
9004 // the build_vectors for 64-bit vector element shift counts are generally
9005 // not legal, and it is hard to see their values after they get legalized to
9006 // loads from a constant pool.
9007 case Intrinsic::arm_neon_vshifts:
9008 case Intrinsic::arm_neon_vshiftu:
9009 case Intrinsic::arm_neon_vshiftls:
9010 case Intrinsic::arm_neon_vshiftlu:
9011 case Intrinsic::arm_neon_vshiftn:
9012 case Intrinsic::arm_neon_vrshifts:
9013 case Intrinsic::arm_neon_vrshiftu:
9014 case Intrinsic::arm_neon_vrshiftn:
9015 case Intrinsic::arm_neon_vqshifts:
9016 case Intrinsic::arm_neon_vqshiftu:
9017 case Intrinsic::arm_neon_vqshiftsu:
9018 case Intrinsic::arm_neon_vqshiftns:
9019 case Intrinsic::arm_neon_vqshiftnu:
9020 case Intrinsic::arm_neon_vqshiftnsu:
9021 case Intrinsic::arm_neon_vqrshiftns:
9022 case Intrinsic::arm_neon_vqrshiftnu:
9023 case Intrinsic::arm_neon_vqrshiftnsu: {
9024 EVT VT = N->getOperand(1).getValueType();
9026 unsigned VShiftOpc = 0;
9029 case Intrinsic::arm_neon_vshifts:
9030 case Intrinsic::arm_neon_vshiftu:
9031 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9032 VShiftOpc = ARMISD::VSHL;
9035 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9036 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9037 ARMISD::VSHRs : ARMISD::VSHRu);
9042 case Intrinsic::arm_neon_vshiftls:
9043 case Intrinsic::arm_neon_vshiftlu:
9044 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
9046 llvm_unreachable("invalid shift count for vshll intrinsic");
9048 case Intrinsic::arm_neon_vrshifts:
9049 case Intrinsic::arm_neon_vrshiftu:
9050 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9054 case Intrinsic::arm_neon_vqshifts:
9055 case Intrinsic::arm_neon_vqshiftu:
9056 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9060 case Intrinsic::arm_neon_vqshiftsu:
9061 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9063 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9065 case Intrinsic::arm_neon_vshiftn:
9066 case Intrinsic::arm_neon_vrshiftn:
9067 case Intrinsic::arm_neon_vqshiftns:
9068 case Intrinsic::arm_neon_vqshiftnu:
9069 case Intrinsic::arm_neon_vqshiftnsu:
9070 case Intrinsic::arm_neon_vqrshiftns:
9071 case Intrinsic::arm_neon_vqrshiftnu:
9072 case Intrinsic::arm_neon_vqrshiftnsu:
9073 // Narrowing shifts require an immediate right shift.
9074 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9076 llvm_unreachable("invalid shift count for narrowing vector shift "
9080 llvm_unreachable("unhandled vector shift");
9084 case Intrinsic::arm_neon_vshifts:
9085 case Intrinsic::arm_neon_vshiftu:
9086 // Opcode already set above.
9088 case Intrinsic::arm_neon_vshiftls:
9089 case Intrinsic::arm_neon_vshiftlu:
9090 if (Cnt == VT.getVectorElementType().getSizeInBits())
9091 VShiftOpc = ARMISD::VSHLLi;
9093 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
9094 ARMISD::VSHLLs : ARMISD::VSHLLu);
9096 case Intrinsic::arm_neon_vshiftn:
9097 VShiftOpc = ARMISD::VSHRN; break;
9098 case Intrinsic::arm_neon_vrshifts:
9099 VShiftOpc = ARMISD::VRSHRs; break;
9100 case Intrinsic::arm_neon_vrshiftu:
9101 VShiftOpc = ARMISD::VRSHRu; break;
9102 case Intrinsic::arm_neon_vrshiftn:
9103 VShiftOpc = ARMISD::VRSHRN; break;
9104 case Intrinsic::arm_neon_vqshifts:
9105 VShiftOpc = ARMISD::VQSHLs; break;
9106 case Intrinsic::arm_neon_vqshiftu:
9107 VShiftOpc = ARMISD::VQSHLu; break;
9108 case Intrinsic::arm_neon_vqshiftsu:
9109 VShiftOpc = ARMISD::VQSHLsu; break;
9110 case Intrinsic::arm_neon_vqshiftns:
9111 VShiftOpc = ARMISD::VQSHRNs; break;
9112 case Intrinsic::arm_neon_vqshiftnu:
9113 VShiftOpc = ARMISD::VQSHRNu; break;
9114 case Intrinsic::arm_neon_vqshiftnsu:
9115 VShiftOpc = ARMISD::VQSHRNsu; break;
9116 case Intrinsic::arm_neon_vqrshiftns:
9117 VShiftOpc = ARMISD::VQRSHRNs; break;
9118 case Intrinsic::arm_neon_vqrshiftnu:
9119 VShiftOpc = ARMISD::VQRSHRNu; break;
9120 case Intrinsic::arm_neon_vqrshiftnsu:
9121 VShiftOpc = ARMISD::VQRSHRNsu; break;
9124 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9125 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9128 case Intrinsic::arm_neon_vshiftins: {
9129 EVT VT = N->getOperand(1).getValueType();
9131 unsigned VShiftOpc = 0;
9133 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9134 VShiftOpc = ARMISD::VSLI;
9135 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9136 VShiftOpc = ARMISD::VSRI;
9138 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9141 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
9142 N->getOperand(1), N->getOperand(2),
9143 DAG.getConstant(Cnt, MVT::i32));
9146 case Intrinsic::arm_neon_vqrshifts:
9147 case Intrinsic::arm_neon_vqrshiftu:
9148 // No immediate versions of these to check for.
9155 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9156 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9157 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9158 /// vector element shift counts are generally not legal, and it is hard to see
9159 /// their values after they get legalized to loads from a constant pool.
9160 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9161 const ARMSubtarget *ST) {
9162 EVT VT = N->getValueType(0);
9163 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9164 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9165 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9166 SDValue N1 = N->getOperand(1);
9167 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9168 SDValue N0 = N->getOperand(0);
9169 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9170 DAG.MaskedValueIsZero(N0.getOperand(0),
9171 APInt::getHighBitsSet(32, 16)))
9172 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
9176 // Nothing to be done for scalar shifts.
9177 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9178 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9181 assert(ST->hasNEON() && "unexpected vector shift");
9184 switch (N->getOpcode()) {
9185 default: llvm_unreachable("unexpected shift opcode");
9188 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9189 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
9190 DAG.getConstant(Cnt, MVT::i32));
9195 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9196 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9197 ARMISD::VSHRs : ARMISD::VSHRu);
9198 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
9199 DAG.getConstant(Cnt, MVT::i32));
9205 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9206 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9207 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9208 const ARMSubtarget *ST) {
9209 SDValue N0 = N->getOperand(0);
9211 // Check for sign- and zero-extensions of vector extract operations of 8-
9212 // and 16-bit vector elements. NEON supports these directly. They are
9213 // handled during DAG combining because type legalization will promote them
9214 // to 32-bit types and it is messy to recognize the operations after that.
9215 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9216 SDValue Vec = N0.getOperand(0);
9217 SDValue Lane = N0.getOperand(1);
9218 EVT VT = N->getValueType(0);
9219 EVT EltVT = N0.getValueType();
9220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9222 if (VT == MVT::i32 &&
9223 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9224 TLI.isTypeLegal(Vec.getValueType()) &&
9225 isa<ConstantSDNode>(Lane)) {
9228 switch (N->getOpcode()) {
9229 default: llvm_unreachable("unexpected opcode");
9230 case ISD::SIGN_EXTEND:
9231 Opc = ARMISD::VGETLANEs;
9233 case ISD::ZERO_EXTEND:
9234 case ISD::ANY_EXTEND:
9235 Opc = ARMISD::VGETLANEu;
9238 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
9245 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9246 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9247 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9248 const ARMSubtarget *ST) {
9249 // If the target supports NEON, try to use vmax/vmin instructions for f32
9250 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9251 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9252 // a NaN; only do the transformation when it matches that behavior.
9254 // For now only do this when using NEON for FP operations; if using VFP, it
9255 // is not obvious that the benefit outweighs the cost of switching to the
9257 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9258 N->getValueType(0) != MVT::f32)
9261 SDValue CondLHS = N->getOperand(0);
9262 SDValue CondRHS = N->getOperand(1);
9263 SDValue LHS = N->getOperand(2);
9264 SDValue RHS = N->getOperand(3);
9265 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9267 unsigned Opcode = 0;
9269 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9270 IsReversed = false; // x CC y ? x : y
9271 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9272 IsReversed = true ; // x CC y ? y : x
9286 // If LHS is NaN, an ordered comparison will be false and the result will
9287 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9288 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9289 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9290 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9292 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9293 // will return -0, so vmin can only be used for unsafe math or if one of
9294 // the operands is known to be nonzero.
9295 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9296 !DAG.getTarget().Options.UnsafeFPMath &&
9297 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9299 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9308 // If LHS is NaN, an ordered comparison will be false and the result will
9309 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9310 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9311 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9312 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9314 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9315 // will return +0, so vmax can only be used for unsafe math or if one of
9316 // the operands is known to be nonzero.
9317 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9318 !DAG.getTarget().Options.UnsafeFPMath &&
9319 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9321 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9327 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
9330 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9332 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9333 SDValue Cmp = N->getOperand(4);
9334 if (Cmp.getOpcode() != ARMISD::CMPZ)
9335 // Only looking at EQ and NE cases.
9338 EVT VT = N->getValueType(0);
9339 DebugLoc dl = N->getDebugLoc();
9340 SDValue LHS = Cmp.getOperand(0);
9341 SDValue RHS = Cmp.getOperand(1);
9342 SDValue FalseVal = N->getOperand(0);
9343 SDValue TrueVal = N->getOperand(1);
9344 SDValue ARMcc = N->getOperand(2);
9345 ARMCC::CondCodes CC =
9346 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9364 /// FIXME: Turn this into a target neutral optimization?
9366 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9367 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9368 N->getOperand(3), Cmp);
9369 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9371 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9372 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9373 N->getOperand(3), NewCmp);
9376 if (Res.getNode()) {
9377 APInt KnownZero, KnownOne;
9378 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
9379 // Capture demanded bits information that would be otherwise lost.
9380 if (KnownZero == 0xfffffffe)
9381 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9382 DAG.getValueType(MVT::i1));
9383 else if (KnownZero == 0xffffff00)
9384 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9385 DAG.getValueType(MVT::i8));
9386 else if (KnownZero == 0xffff0000)
9387 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9388 DAG.getValueType(MVT::i16));
9394 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9395 DAGCombinerInfo &DCI) const {
9396 switch (N->getOpcode()) {
9398 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9399 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9400 case ISD::SUB: return PerformSUBCombine(N, DCI);
9401 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9402 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9403 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9404 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9405 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9406 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9407 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9408 case ISD::STORE: return PerformSTORECombine(N, DCI);
9409 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9410 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9411 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9412 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9413 case ISD::FP_TO_SINT:
9414 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9415 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9416 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9419 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9420 case ISD::SIGN_EXTEND:
9421 case ISD::ZERO_EXTEND:
9422 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9423 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9424 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9425 case ARMISD::VLD2DUP:
9426 case ARMISD::VLD3DUP:
9427 case ARMISD::VLD4DUP:
9428 return CombineBaseUpdate(N, DCI);
9429 case ISD::INTRINSIC_VOID:
9430 case ISD::INTRINSIC_W_CHAIN:
9431 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9432 case Intrinsic::arm_neon_vld1:
9433 case Intrinsic::arm_neon_vld2:
9434 case Intrinsic::arm_neon_vld3:
9435 case Intrinsic::arm_neon_vld4:
9436 case Intrinsic::arm_neon_vld2lane:
9437 case Intrinsic::arm_neon_vld3lane:
9438 case Intrinsic::arm_neon_vld4lane:
9439 case Intrinsic::arm_neon_vst1:
9440 case Intrinsic::arm_neon_vst2:
9441 case Intrinsic::arm_neon_vst3:
9442 case Intrinsic::arm_neon_vst4:
9443 case Intrinsic::arm_neon_vst2lane:
9444 case Intrinsic::arm_neon_vst3lane:
9445 case Intrinsic::arm_neon_vst4lane:
9446 return CombineBaseUpdate(N, DCI);
9454 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9456 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9459 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
9460 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9461 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9463 switch (VT.getSimpleVT().SimpleTy) {
9469 // Unaligned access can use (for example) LRDB, LRDH, LDR
9470 if (AllowsUnaligned) {
9472 *Fast = Subtarget->hasV7Ops();
9479 // For any little-endian targets with neon, we can support unaligned ld/st
9480 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9481 // A big-endian target may also explictly support unaligned accesses
9482 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9492 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9493 unsigned AlignCheck) {
9494 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9495 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9498 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9499 unsigned DstAlign, unsigned SrcAlign,
9500 bool IsMemset, bool ZeroMemset,
9502 MachineFunction &MF) const {
9503 const Function *F = MF.getFunction();
9505 // See if we can use NEON instructions for this...
9506 if ((!IsMemset || ZeroMemset) &&
9507 Subtarget->hasNEON() &&
9508 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9509 Attribute::NoImplicitFloat)) {
9512 (memOpAlign(SrcAlign, DstAlign, 16) ||
9513 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) {
9515 } else if (Size >= 8 &&
9516 (memOpAlign(SrcAlign, DstAlign, 8) ||
9517 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) {
9522 // Lowering to i32/i16 if the size permits.
9528 // Let the target-independent logic figure it out.
9532 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9533 if (Val.getOpcode() != ISD::LOAD)
9536 EVT VT1 = Val.getValueType();
9537 if (!VT1.isSimple() || !VT1.isInteger() ||
9538 !VT2.isSimple() || !VT2.isInteger())
9541 switch (VT1.getSimpleVT().SimpleTy) {
9546 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9553 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9558 switch (VT.getSimpleVT().SimpleTy) {
9559 default: return false;
9574 if ((V & (Scale - 1)) != 0)
9577 return V == (V & ((1LL << 5) - 1));
9580 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9581 const ARMSubtarget *Subtarget) {
9588 switch (VT.getSimpleVT().SimpleTy) {
9589 default: return false;
9594 // + imm12 or - imm8
9596 return V == (V & ((1LL << 8) - 1));
9597 return V == (V & ((1LL << 12) - 1));
9600 // Same as ARM mode. FIXME: NEON?
9601 if (!Subtarget->hasVFP2())
9606 return V == (V & ((1LL << 8) - 1));
9610 /// isLegalAddressImmediate - Return true if the integer value can be used
9611 /// as the offset of the target addressing mode for load / store of the
9613 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9614 const ARMSubtarget *Subtarget) {
9621 if (Subtarget->isThumb1Only())
9622 return isLegalT1AddressImmediate(V, VT);
9623 else if (Subtarget->isThumb2())
9624 return isLegalT2AddressImmediate(V, VT, Subtarget);
9629 switch (VT.getSimpleVT().SimpleTy) {
9630 default: return false;
9635 return V == (V & ((1LL << 12) - 1));
9638 return V == (V & ((1LL << 8) - 1));
9641 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9646 return V == (V & ((1LL << 8) - 1));
9650 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9652 int Scale = AM.Scale;
9656 switch (VT.getSimpleVT().SimpleTy) {
9657 default: return false;
9666 return Scale == 2 || Scale == 4 || Scale == 8;
9669 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9673 // Note, we allow "void" uses (basically, uses that aren't loads or
9674 // stores), because arm allows folding a scale into many arithmetic
9675 // operations. This should be made more precise and revisited later.
9677 // Allow r << imm, but the imm has to be a multiple of two.
9678 if (Scale & 1) return false;
9679 return isPowerOf2_32(Scale);
9683 /// isLegalAddressingMode - Return true if the addressing mode represented
9684 /// by AM is legal for this target, for a load/store of the specified type.
9685 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9687 EVT VT = getValueType(Ty, true);
9688 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9691 // Can never fold addr of global into load/store.
9696 case 0: // no scale reg, must be "r+i" or "r", or "i".
9699 if (Subtarget->isThumb1Only())
9703 // ARM doesn't support any R+R*scale+imm addr modes.
9710 if (Subtarget->isThumb2())
9711 return isLegalT2ScaledAddressingMode(AM, VT);
9713 int Scale = AM.Scale;
9714 switch (VT.getSimpleVT().SimpleTy) {
9715 default: return false;
9719 if (Scale < 0) Scale = -Scale;
9723 return isPowerOf2_32(Scale & ~1);
9727 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9732 // Note, we allow "void" uses (basically, uses that aren't loads or
9733 // stores), because arm allows folding a scale into many arithmetic
9734 // operations. This should be made more precise and revisited later.
9736 // Allow r << imm, but the imm has to be a multiple of two.
9737 if (Scale & 1) return false;
9738 return isPowerOf2_32(Scale);
9744 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9745 /// icmp immediate, that is the target has icmp instructions which can compare
9746 /// a register against the immediate without having to materialize the
9747 /// immediate into a register.
9748 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9749 // Thumb2 and ARM modes can use cmn for negative immediates.
9750 if (!Subtarget->isThumb())
9751 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9752 if (Subtarget->isThumb2())
9753 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9754 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9755 return Imm >= 0 && Imm <= 255;
9758 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9759 /// *or sub* immediate, that is the target has add or sub instructions which can
9760 /// add a register with the immediate without having to materialize the
9761 /// immediate into a register.
9762 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9763 // Same encoding for add/sub, just flip the sign.
9764 int64_t AbsImm = llvm::abs64(Imm);
9765 if (!Subtarget->isThumb())
9766 return ARM_AM::getSOImmVal(AbsImm) != -1;
9767 if (Subtarget->isThumb2())
9768 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9769 // Thumb1 only has 8-bit unsigned immediate.
9770 return AbsImm >= 0 && AbsImm <= 255;
9773 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9774 bool isSEXTLoad, SDValue &Base,
9775 SDValue &Offset, bool &isInc,
9776 SelectionDAG &DAG) {
9777 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9780 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9782 Base = Ptr->getOperand(0);
9783 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9784 int RHSC = (int)RHS->getZExtValue();
9785 if (RHSC < 0 && RHSC > -256) {
9786 assert(Ptr->getOpcode() == ISD::ADD);
9788 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9792 isInc = (Ptr->getOpcode() == ISD::ADD);
9793 Offset = Ptr->getOperand(1);
9795 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9797 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9798 int RHSC = (int)RHS->getZExtValue();
9799 if (RHSC < 0 && RHSC > -0x1000) {
9800 assert(Ptr->getOpcode() == ISD::ADD);
9802 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9803 Base = Ptr->getOperand(0);
9808 if (Ptr->getOpcode() == ISD::ADD) {
9810 ARM_AM::ShiftOpc ShOpcVal=
9811 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9812 if (ShOpcVal != ARM_AM::no_shift) {
9813 Base = Ptr->getOperand(1);
9814 Offset = Ptr->getOperand(0);
9816 Base = Ptr->getOperand(0);
9817 Offset = Ptr->getOperand(1);
9822 isInc = (Ptr->getOpcode() == ISD::ADD);
9823 Base = Ptr->getOperand(0);
9824 Offset = Ptr->getOperand(1);
9828 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9832 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9833 bool isSEXTLoad, SDValue &Base,
9834 SDValue &Offset, bool &isInc,
9835 SelectionDAG &DAG) {
9836 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9839 Base = Ptr->getOperand(0);
9840 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9841 int RHSC = (int)RHS->getZExtValue();
9842 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9843 assert(Ptr->getOpcode() == ISD::ADD);
9845 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9847 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9848 isInc = Ptr->getOpcode() == ISD::ADD;
9849 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9857 /// getPreIndexedAddressParts - returns true by value, base pointer and
9858 /// offset pointer and addressing mode by reference if the node's address
9859 /// can be legally represented as pre-indexed load / store address.
9861 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9863 ISD::MemIndexedMode &AM,
9864 SelectionDAG &DAG) const {
9865 if (Subtarget->isThumb1Only())
9870 bool isSEXTLoad = false;
9871 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9872 Ptr = LD->getBasePtr();
9873 VT = LD->getMemoryVT();
9874 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9875 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9876 Ptr = ST->getBasePtr();
9877 VT = ST->getMemoryVT();
9882 bool isLegal = false;
9883 if (Subtarget->isThumb2())
9884 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9885 Offset, isInc, DAG);
9887 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9888 Offset, isInc, DAG);
9892 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9896 /// getPostIndexedAddressParts - returns true by value, base pointer and
9897 /// offset pointer and addressing mode by reference if this node can be
9898 /// combined with a load / store to form a post-indexed load / store.
9899 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9902 ISD::MemIndexedMode &AM,
9903 SelectionDAG &DAG) const {
9904 if (Subtarget->isThumb1Only())
9909 bool isSEXTLoad = false;
9910 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9911 VT = LD->getMemoryVT();
9912 Ptr = LD->getBasePtr();
9913 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9914 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9915 VT = ST->getMemoryVT();
9916 Ptr = ST->getBasePtr();
9921 bool isLegal = false;
9922 if (Subtarget->isThumb2())
9923 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9926 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9932 // Swap base ptr and offset to catch more post-index load / store when
9933 // it's legal. In Thumb2 mode, offset must be an immediate.
9934 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9935 !Subtarget->isThumb2())
9936 std::swap(Base, Offset);
9938 // Post-indexed load / store update the base pointer.
9943 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9947 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9950 const SelectionDAG &DAG,
9951 unsigned Depth) const {
9952 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9953 switch (Op.getOpcode()) {
9955 case ARMISD::CMOV: {
9956 // Bits are known zero/one if known on the LHS and RHS.
9957 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9958 if (KnownZero == 0 && KnownOne == 0) return;
9960 APInt KnownZeroRHS, KnownOneRHS;
9961 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9962 KnownZero &= KnownZeroRHS;
9963 KnownOne &= KnownOneRHS;
9969 //===----------------------------------------------------------------------===//
9970 // ARM Inline Assembly Support
9971 //===----------------------------------------------------------------------===//
9973 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9974 // Looking for "rev" which is V6+.
9975 if (!Subtarget->hasV6Ops())
9978 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9979 std::string AsmStr = IA->getAsmString();
9980 SmallVector<StringRef, 4> AsmPieces;
9981 SplitString(AsmStr, AsmPieces, ";\n");
9983 switch (AsmPieces.size()) {
9984 default: return false;
9986 AsmStr = AsmPieces[0];
9988 SplitString(AsmStr, AsmPieces, " \t,");
9991 if (AsmPieces.size() == 3 &&
9992 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9993 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9994 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9995 if (Ty && Ty->getBitWidth() == 32)
9996 return IntrinsicLowering::LowerToByteSwap(CI);
10004 /// getConstraintType - Given a constraint letter, return the type of
10005 /// constraint it is for this target.
10006 ARMTargetLowering::ConstraintType
10007 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10008 if (Constraint.size() == 1) {
10009 switch (Constraint[0]) {
10011 case 'l': return C_RegisterClass;
10012 case 'w': return C_RegisterClass;
10013 case 'h': return C_RegisterClass;
10014 case 'x': return C_RegisterClass;
10015 case 't': return C_RegisterClass;
10016 case 'j': return C_Other; // Constant for movw.
10017 // An address with a single base register. Due to the way we
10018 // currently handle addresses it is the same as an 'r' memory constraint.
10019 case 'Q': return C_Memory;
10021 } else if (Constraint.size() == 2) {
10022 switch (Constraint[0]) {
10024 // All 'U+' constraints are addresses.
10025 case 'U': return C_Memory;
10028 return TargetLowering::getConstraintType(Constraint);
10031 /// Examine constraint type and operand type and determine a weight value.
10032 /// This object must already have been set up with the operand type
10033 /// and the current alternative constraint selected.
10034 TargetLowering::ConstraintWeight
10035 ARMTargetLowering::getSingleConstraintMatchWeight(
10036 AsmOperandInfo &info, const char *constraint) const {
10037 ConstraintWeight weight = CW_Invalid;
10038 Value *CallOperandVal = info.CallOperandVal;
10039 // If we don't have a value, we can't do a match,
10040 // but allow it at the lowest weight.
10041 if (CallOperandVal == NULL)
10043 Type *type = CallOperandVal->getType();
10044 // Look at the constraint type.
10045 switch (*constraint) {
10047 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10050 if (type->isIntegerTy()) {
10051 if (Subtarget->isThumb())
10052 weight = CW_SpecificReg;
10054 weight = CW_Register;
10058 if (type->isFloatingPointTy())
10059 weight = CW_Register;
10065 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10067 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10069 if (Constraint.size() == 1) {
10070 // GCC ARM Constraint Letters
10071 switch (Constraint[0]) {
10072 case 'l': // Low regs or general regs.
10073 if (Subtarget->isThumb())
10074 return RCPair(0U, &ARM::tGPRRegClass);
10075 return RCPair(0U, &ARM::GPRRegClass);
10076 case 'h': // High regs or no regs.
10077 if (Subtarget->isThumb())
10078 return RCPair(0U, &ARM::hGPRRegClass);
10081 return RCPair(0U, &ARM::GPRRegClass);
10083 if (VT == MVT::f32)
10084 return RCPair(0U, &ARM::SPRRegClass);
10085 if (VT.getSizeInBits() == 64)
10086 return RCPair(0U, &ARM::DPRRegClass);
10087 if (VT.getSizeInBits() == 128)
10088 return RCPair(0U, &ARM::QPRRegClass);
10091 if (VT == MVT::f32)
10092 return RCPair(0U, &ARM::SPR_8RegClass);
10093 if (VT.getSizeInBits() == 64)
10094 return RCPair(0U, &ARM::DPR_8RegClass);
10095 if (VT.getSizeInBits() == 128)
10096 return RCPair(0U, &ARM::QPR_8RegClass);
10099 if (VT == MVT::f32)
10100 return RCPair(0U, &ARM::SPRRegClass);
10104 if (StringRef("{cc}").equals_lower(Constraint))
10105 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10107 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10110 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10111 /// vector. If it is invalid, don't add anything to Ops.
10112 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10113 std::string &Constraint,
10114 std::vector<SDValue>&Ops,
10115 SelectionDAG &DAG) const {
10116 SDValue Result(0, 0);
10118 // Currently only support length 1 constraints.
10119 if (Constraint.length() != 1) return;
10121 char ConstraintLetter = Constraint[0];
10122 switch (ConstraintLetter) {
10125 case 'I': case 'J': case 'K': case 'L':
10126 case 'M': case 'N': case 'O':
10127 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10131 int64_t CVal64 = C->getSExtValue();
10132 int CVal = (int) CVal64;
10133 // None of these constraints allow values larger than 32 bits. Check
10134 // that the value fits in an int.
10135 if (CVal != CVal64)
10138 switch (ConstraintLetter) {
10140 // Constant suitable for movw, must be between 0 and
10142 if (Subtarget->hasV6T2Ops())
10143 if (CVal >= 0 && CVal <= 65535)
10147 if (Subtarget->isThumb1Only()) {
10148 // This must be a constant between 0 and 255, for ADD
10150 if (CVal >= 0 && CVal <= 255)
10152 } else if (Subtarget->isThumb2()) {
10153 // A constant that can be used as an immediate value in a
10154 // data-processing instruction.
10155 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10158 // A constant that can be used as an immediate value in a
10159 // data-processing instruction.
10160 if (ARM_AM::getSOImmVal(CVal) != -1)
10166 if (Subtarget->isThumb()) { // FIXME thumb2
10167 // This must be a constant between -255 and -1, for negated ADD
10168 // immediates. This can be used in GCC with an "n" modifier that
10169 // prints the negated value, for use with SUB instructions. It is
10170 // not useful otherwise but is implemented for compatibility.
10171 if (CVal >= -255 && CVal <= -1)
10174 // This must be a constant between -4095 and 4095. It is not clear
10175 // what this constraint is intended for. Implemented for
10176 // compatibility with GCC.
10177 if (CVal >= -4095 && CVal <= 4095)
10183 if (Subtarget->isThumb1Only()) {
10184 // A 32-bit value where only one byte has a nonzero value. Exclude
10185 // zero to match GCC. This constraint is used by GCC internally for
10186 // constants that can be loaded with a move/shift combination.
10187 // It is not useful otherwise but is implemented for compatibility.
10188 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10190 } else if (Subtarget->isThumb2()) {
10191 // A constant whose bitwise inverse can be used as an immediate
10192 // value in a data-processing instruction. This can be used in GCC
10193 // with a "B" modifier that prints the inverted value, for use with
10194 // BIC and MVN instructions. It is not useful otherwise but is
10195 // implemented for compatibility.
10196 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10199 // A constant whose bitwise inverse can be used as an immediate
10200 // value in a data-processing instruction. This can be used in GCC
10201 // with a "B" modifier that prints the inverted value, for use with
10202 // BIC and MVN instructions. It is not useful otherwise but is
10203 // implemented for compatibility.
10204 if (ARM_AM::getSOImmVal(~CVal) != -1)
10210 if (Subtarget->isThumb1Only()) {
10211 // This must be a constant between -7 and 7,
10212 // for 3-operand ADD/SUB immediate instructions.
10213 if (CVal >= -7 && CVal < 7)
10215 } else if (Subtarget->isThumb2()) {
10216 // A constant whose negation can be used as an immediate value in a
10217 // data-processing instruction. This can be used in GCC with an "n"
10218 // modifier that prints the negated value, for use with SUB
10219 // instructions. It is not useful otherwise but is implemented for
10221 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10224 // A constant whose negation can be used as an immediate value in a
10225 // data-processing instruction. This can be used in GCC with an "n"
10226 // modifier that prints the negated value, for use with SUB
10227 // instructions. It is not useful otherwise but is implemented for
10229 if (ARM_AM::getSOImmVal(-CVal) != -1)
10235 if (Subtarget->isThumb()) { // FIXME thumb2
10236 // This must be a multiple of 4 between 0 and 1020, for
10237 // ADD sp + immediate.
10238 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10241 // A power of two or a constant between 0 and 32. This is used in
10242 // GCC for the shift amount on shifted register operands, but it is
10243 // useful in general for any shift amounts.
10244 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10250 if (Subtarget->isThumb()) { // FIXME thumb2
10251 // This must be a constant between 0 and 31, for shift amounts.
10252 if (CVal >= 0 && CVal <= 31)
10258 if (Subtarget->isThumb()) { // FIXME thumb2
10259 // This must be a multiple of 4 between -508 and 508, for
10260 // ADD/SUB sp = sp + immediate.
10261 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10266 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10270 if (Result.getNode()) {
10271 Ops.push_back(Result);
10274 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10278 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10279 // The ARM target isn't yet aware of offsets.
10283 bool ARM::isBitFieldInvertedMask(unsigned v) {
10284 if (v == 0xffffffff)
10286 // there can be 1's on either or both "outsides", all the "inside"
10287 // bits must be 0's
10288 unsigned int lsb = 0, msb = 31;
10289 while (v & (1 << msb)) --msb;
10290 while (v & (1 << lsb)) ++lsb;
10291 for (unsigned int i = lsb; i <= msb; ++i) {
10298 /// isFPImmLegal - Returns true if the target can instruction select the
10299 /// specified FP immediate natively. If false, the legalizer will
10300 /// materialize the FP immediate as a load from a constant pool.
10301 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10302 if (!Subtarget->hasVFP3())
10304 if (VT == MVT::f32)
10305 return ARM_AM::getFP32Imm(Imm) != -1;
10306 if (VT == MVT::f64)
10307 return ARM_AM::getFP64Imm(Imm) != -1;
10311 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10312 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10313 /// specified in the intrinsic calls.
10314 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10316 unsigned Intrinsic) const {
10317 switch (Intrinsic) {
10318 case Intrinsic::arm_neon_vld1:
10319 case Intrinsic::arm_neon_vld2:
10320 case Intrinsic::arm_neon_vld3:
10321 case Intrinsic::arm_neon_vld4:
10322 case Intrinsic::arm_neon_vld2lane:
10323 case Intrinsic::arm_neon_vld3lane:
10324 case Intrinsic::arm_neon_vld4lane: {
10325 Info.opc = ISD::INTRINSIC_W_CHAIN;
10326 // Conservatively set memVT to the entire set of vectors loaded.
10327 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10328 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10329 Info.ptrVal = I.getArgOperand(0);
10331 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10332 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10333 Info.vol = false; // volatile loads with NEON intrinsics not supported
10334 Info.readMem = true;
10335 Info.writeMem = false;
10338 case Intrinsic::arm_neon_vst1:
10339 case Intrinsic::arm_neon_vst2:
10340 case Intrinsic::arm_neon_vst3:
10341 case Intrinsic::arm_neon_vst4:
10342 case Intrinsic::arm_neon_vst2lane:
10343 case Intrinsic::arm_neon_vst3lane:
10344 case Intrinsic::arm_neon_vst4lane: {
10345 Info.opc = ISD::INTRINSIC_VOID;
10346 // Conservatively set memVT to the entire set of vectors stored.
10347 unsigned NumElts = 0;
10348 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10349 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10350 if (!ArgTy->isVectorTy())
10352 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10354 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10355 Info.ptrVal = I.getArgOperand(0);
10357 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10358 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10359 Info.vol = false; // volatile stores with NEON intrinsics not supported
10360 Info.readMem = false;
10361 Info.writeMem = true;
10364 case Intrinsic::arm_strexd: {
10365 Info.opc = ISD::INTRINSIC_W_CHAIN;
10366 Info.memVT = MVT::i64;
10367 Info.ptrVal = I.getArgOperand(2);
10371 Info.readMem = false;
10372 Info.writeMem = true;
10375 case Intrinsic::arm_ldrexd: {
10376 Info.opc = ISD::INTRINSIC_W_CHAIN;
10377 Info.memVT = MVT::i64;
10378 Info.ptrVal = I.getArgOperand(0);
10382 Info.readMem = true;
10383 Info.writeMem = false;