1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "MCTargetDesc/ARMAddressingModes.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/IntrinsicLowering.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/PseudoSourceValue.h"
43 #include "llvm/CodeGen/SelectionDAG.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Target/TargetOptions.h"
46 #include "llvm/ADT/VectorExtras.h"
47 #include "llvm/ADT/StringExtras.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
59 // This option should go away when tail calls fully work.
61 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
62 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
66 EnableARMLongCalls("arm-long-calls", cl::Hidden,
67 cl::desc("Generate calls via indirect call instructions"),
71 ARMInterworking("arm-interworking", cl::Hidden,
72 cl::desc("Enable / disable ARM interworking (for debugging only)"),
76 class ARMCCState : public CCState {
78 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
79 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
80 LLVMContext &C, ParmContext PC)
81 : CCState(CC, isVarArg, MF, TM, locs, C) {
82 assert(((PC == Call) || (PC == Prologue)) &&
83 "ARMCCState users must specify whether their context is call"
84 "or prologue generation.");
90 // The APCS parameter registers.
91 static const unsigned GPRArgRegs[] = {
92 ARM::R0, ARM::R1, ARM::R2, ARM::R3
95 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
96 EVT PromotedBitwiseVT) {
97 if (VT != PromotedLdStVT) {
98 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
99 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
100 PromotedLdStVT.getSimpleVT());
102 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
103 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
104 PromotedLdStVT.getSimpleVT());
107 EVT ElemTy = VT.getVectorElementType();
108 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
109 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
110 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
111 if (ElemTy != MVT::i32) {
112 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
120 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
121 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
122 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
123 if (VT.isInteger()) {
124 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
125 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
126 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
127 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
128 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
129 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
130 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
131 setTruncStoreAction(VT.getSimpleVT(),
132 (MVT::SimpleValueType)InnerVT, Expand);
134 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
136 // Promote all bit-wise operations.
137 if (VT.isInteger() && VT != PromotedBitwiseVT) {
138 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
139 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
140 PromotedBitwiseVT.getSimpleVT());
141 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
142 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
143 PromotedBitwiseVT.getSimpleVT());
144 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
145 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
146 PromotedBitwiseVT.getSimpleVT());
149 // Neon does not support vector divide/remainder operations.
150 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
151 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
154 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
155 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
158 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
159 addRegisterClass(VT, ARM::DPRRegisterClass);
160 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
163 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
164 addRegisterClass(VT, ARM::QPRRegisterClass);
165 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
168 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
169 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
170 return new TargetLoweringObjectFileMachO();
172 return new ARMElfTargetObjectFile();
175 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
176 : TargetLowering(TM, createTLOF(TM)) {
177 Subtarget = &TM.getSubtarget<ARMSubtarget>();
178 RegInfo = TM.getRegisterInfo();
179 Itins = TM.getInstrItineraryData();
181 if (Subtarget->isTargetDarwin()) {
182 // Uses VFP for Thumb libfuncs if available.
183 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
184 // Single-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
186 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
187 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
188 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
190 // Double-precision floating-point arithmetic.
191 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
192 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
193 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
194 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
196 // Single-precision comparisons.
197 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
198 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
199 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
200 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
201 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
202 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
203 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
204 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
206 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
215 // Double-precision comparisons.
216 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
217 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
218 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
219 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
220 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
221 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
222 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
223 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
225 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
234 // Floating-point to integer conversions.
235 // i64 conversions are done via library routines even when generating VFP
236 // instructions, so use the same ones.
237 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
238 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
239 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
240 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
242 // Conversions between floating types.
243 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
244 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
246 // Integer to floating-point conversions.
247 // i64 conversions are done via library routines even when generating VFP
248 // instructions, so use the same ones.
249 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
250 // e.g., __floatunsidf vs. __floatunssidfvfp.
251 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
252 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
253 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
254 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
258 // These libcalls are not available in 32-bit.
259 setLibcallName(RTLIB::SHL_I128, 0);
260 setLibcallName(RTLIB::SRL_I128, 0);
261 setLibcallName(RTLIB::SRA_I128, 0);
263 if (Subtarget->isAAPCS_ABI()) {
264 // Double-precision floating-point arithmetic helper functions
265 // RTABI chapter 4.1.2, Table 2
266 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
267 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
268 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
269 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
270 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
271 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
272 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
273 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
275 // Double-precision floating-point comparison helper functions
276 // RTABI chapter 4.1.2, Table 3
277 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
278 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
279 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
280 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
281 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
282 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
284 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
285 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
286 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
287 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
288 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
289 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
290 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
291 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
292 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
293 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
302 // Single-precision floating-point arithmetic helper functions
303 // RTABI chapter 4.1.2, Table 4
304 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
305 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
306 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
307 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
308 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
309 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
310 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
311 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
313 // Single-precision floating-point comparison helper functions
314 // RTABI chapter 4.1.2, Table 5
315 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
316 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
317 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
318 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
319 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
320 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
322 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
323 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
324 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
325 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
326 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
327 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
328 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
329 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
330 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
331 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
340 // Floating-point to integer conversions.
341 // RTABI chapter 4.1.2, Table 6
342 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
343 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
345 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
346 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
347 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
348 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
349 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
350 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
359 // Conversions between floating types.
360 // RTABI chapter 4.1.2, Table 7
361 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
362 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
363 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
364 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
366 // Integer to floating-point conversions.
367 // RTABI chapter 4.1.2, Table 8
368 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
369 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
370 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
371 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
372 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
373 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
374 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
375 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385 // Long long helper functions
386 // RTABI chapter 4.2, Table 9
387 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
388 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
389 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
390 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
391 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
392 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
393 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
396 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
400 // Integer division functions
401 // RTABI chapter 4.3.1
402 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
403 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
404 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
405 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
406 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
407 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
408 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
416 // RTABI chapter 4.3.4
417 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
418 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
419 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
422 if (Subtarget->isThumb1Only())
423 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
425 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
426 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
427 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
428 if (!Subtarget->isFPOnlySP())
429 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
431 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
434 if (Subtarget->hasNEON()) {
435 addDRTypeForNEON(MVT::v2f32);
436 addDRTypeForNEON(MVT::v8i8);
437 addDRTypeForNEON(MVT::v4i16);
438 addDRTypeForNEON(MVT::v2i32);
439 addDRTypeForNEON(MVT::v1i64);
441 addQRTypeForNEON(MVT::v4f32);
442 addQRTypeForNEON(MVT::v2f64);
443 addQRTypeForNEON(MVT::v16i8);
444 addQRTypeForNEON(MVT::v8i16);
445 addQRTypeForNEON(MVT::v4i32);
446 addQRTypeForNEON(MVT::v2i64);
448 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
449 // neither Neon nor VFP support any arithmetic operations on it.
450 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
451 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
452 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
453 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
454 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
455 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
456 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
457 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
458 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
459 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
460 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
461 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
462 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
463 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
464 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
465 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
467 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
468 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
469 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
470 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
471 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
472 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
473 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
475 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
477 // Neon does not support some operations on v1i64 and v2i64 types.
478 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
479 // Custom handling for some quad-vector types to detect VMULL.
480 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
481 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
482 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
483 // Custom handling for some vector types to avoid expensive expansions
484 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
485 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
486 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
487 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
488 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
489 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
490 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
491 // a destination type that is wider than the source.
492 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
493 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
495 setTargetDAGCombine(ISD::INTRINSIC_VOID);
496 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
497 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
498 setTargetDAGCombine(ISD::SHL);
499 setTargetDAGCombine(ISD::SRL);
500 setTargetDAGCombine(ISD::SRA);
501 setTargetDAGCombine(ISD::SIGN_EXTEND);
502 setTargetDAGCombine(ISD::ZERO_EXTEND);
503 setTargetDAGCombine(ISD::ANY_EXTEND);
504 setTargetDAGCombine(ISD::SELECT_CC);
505 setTargetDAGCombine(ISD::BUILD_VECTOR);
506 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
507 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
508 setTargetDAGCombine(ISD::STORE);
509 setTargetDAGCombine(ISD::FP_TO_SINT);
510 setTargetDAGCombine(ISD::FP_TO_UINT);
511 setTargetDAGCombine(ISD::FDIV);
514 computeRegisterProperties();
516 // ARM does not have f32 extending load.
517 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
519 // ARM does not have i1 sign extending load.
520 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
522 // ARM supports all 4 flavors of integer indexed load / store.
523 if (!Subtarget->isThumb1Only()) {
524 for (unsigned im = (unsigned)ISD::PRE_INC;
525 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
526 setIndexedLoadAction(im, MVT::i1, Legal);
527 setIndexedLoadAction(im, MVT::i8, Legal);
528 setIndexedLoadAction(im, MVT::i16, Legal);
529 setIndexedLoadAction(im, MVT::i32, Legal);
530 setIndexedStoreAction(im, MVT::i1, Legal);
531 setIndexedStoreAction(im, MVT::i8, Legal);
532 setIndexedStoreAction(im, MVT::i16, Legal);
533 setIndexedStoreAction(im, MVT::i32, Legal);
537 // i64 operation support.
538 setOperationAction(ISD::MUL, MVT::i64, Expand);
539 setOperationAction(ISD::MULHU, MVT::i32, Expand);
540 if (Subtarget->isThumb1Only()) {
541 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
542 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
544 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
545 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
546 setOperationAction(ISD::MULHS, MVT::i32, Expand);
548 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
549 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
550 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
551 setOperationAction(ISD::SRL, MVT::i64, Custom);
552 setOperationAction(ISD::SRA, MVT::i64, Custom);
554 if (!Subtarget->isThumb1Only()) {
555 // FIXME: We should do this for Thumb1 as well.
556 setOperationAction(ISD::ADDC, MVT::i32, Custom);
557 setOperationAction(ISD::ADDE, MVT::i32, Custom);
558 setOperationAction(ISD::SUBC, MVT::i32, Custom);
559 setOperationAction(ISD::SUBE, MVT::i32, Custom);
562 // ARM does not have ROTL.
563 setOperationAction(ISD::ROTL, MVT::i32, Expand);
564 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
565 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
566 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
567 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
569 // Only ARMv6 has BSWAP.
570 if (!Subtarget->hasV6Ops())
571 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
573 // These are expanded into libcalls.
574 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
575 // v7M has a hardware divider
576 setOperationAction(ISD::SDIV, MVT::i32, Expand);
577 setOperationAction(ISD::UDIV, MVT::i32, Expand);
579 setOperationAction(ISD::SREM, MVT::i32, Expand);
580 setOperationAction(ISD::UREM, MVT::i32, Expand);
581 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
582 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
584 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
585 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
586 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
587 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
588 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
590 setOperationAction(ISD::TRAP, MVT::Other, Legal);
592 // Use the default implementation.
593 setOperationAction(ISD::VASTART, MVT::Other, Custom);
594 setOperationAction(ISD::VAARG, MVT::Other, Expand);
595 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
596 setOperationAction(ISD::VAEND, MVT::Other, Expand);
597 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
598 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
599 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
600 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
601 setExceptionPointerRegister(ARM::R0);
602 setExceptionSelectorRegister(ARM::R1);
604 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
605 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
606 // the default expansion.
607 // FIXME: This should be checking for v6k, not just v6.
608 if (Subtarget->hasDataBarrier() ||
609 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
610 // membarrier needs custom lowering; the rest are legal and handled
612 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
613 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
614 // Custom lowering for 64-bit ops
615 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
616 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
617 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
618 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
619 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
620 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
621 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
622 setInsertFencesForAtomic(true);
624 // Set them all for expansion, which will force libcalls.
625 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
626 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
627 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
628 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
629 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
630 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
631 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
632 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
633 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
634 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
635 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
636 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
637 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
638 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
639 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
640 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
641 // Since the libcalls include locking, fold in the fences
642 setShouldFoldAtomicFences(true);
645 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
647 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
648 if (!Subtarget->hasV6Ops()) {
649 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
650 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
652 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
654 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
655 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
656 // iff target supports vfp2.
657 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
658 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
661 // We want to custom lower some of our intrinsics.
662 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
663 if (Subtarget->isTargetDarwin()) {
664 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
665 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
666 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
667 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
670 setOperationAction(ISD::SETCC, MVT::i32, Expand);
671 setOperationAction(ISD::SETCC, MVT::f32, Expand);
672 setOperationAction(ISD::SETCC, MVT::f64, Expand);
673 setOperationAction(ISD::SELECT, MVT::i32, Custom);
674 setOperationAction(ISD::SELECT, MVT::f32, Custom);
675 setOperationAction(ISD::SELECT, MVT::f64, Custom);
676 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
677 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
678 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
680 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
681 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
682 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
683 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
684 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
686 // We don't support sin/cos/fmod/copysign/pow
687 setOperationAction(ISD::FSIN, MVT::f64, Expand);
688 setOperationAction(ISD::FSIN, MVT::f32, Expand);
689 setOperationAction(ISD::FCOS, MVT::f32, Expand);
690 setOperationAction(ISD::FCOS, MVT::f64, Expand);
691 setOperationAction(ISD::FREM, MVT::f64, Expand);
692 setOperationAction(ISD::FREM, MVT::f32, Expand);
693 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
694 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
695 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
697 setOperationAction(ISD::FPOW, MVT::f64, Expand);
698 setOperationAction(ISD::FPOW, MVT::f32, Expand);
700 setOperationAction(ISD::FMA, MVT::f64, Expand);
701 setOperationAction(ISD::FMA, MVT::f32, Expand);
703 // Various VFP goodness
704 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
705 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
706 if (Subtarget->hasVFP2()) {
707 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
708 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
709 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
710 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
712 // Special handling for half-precision FP.
713 if (!Subtarget->hasFP16()) {
714 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
715 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
719 // We have target-specific dag combine patterns for the following nodes:
720 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
721 setTargetDAGCombine(ISD::ADD);
722 setTargetDAGCombine(ISD::SUB);
723 setTargetDAGCombine(ISD::MUL);
725 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
726 setTargetDAGCombine(ISD::OR);
727 if (Subtarget->hasNEON())
728 setTargetDAGCombine(ISD::AND);
730 setStackPointerRegisterToSaveRestore(ARM::SP);
732 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
733 setSchedulingPreference(Sched::RegPressure);
735 setSchedulingPreference(Sched::Hybrid);
737 //// temporary - rewrite interface to use type
738 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
740 // On ARM arguments smaller than 4 bytes are extended, so all arguments
741 // are at least 4 bytes aligned.
742 setMinStackArgumentAlignment(4);
744 benefitFromCodePlacementOpt = true;
746 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
749 // FIXME: It might make sense to define the representative register class as the
750 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
751 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
752 // SPR's representative would be DPR_VFP2. This should work well if register
753 // pressure tracking were modified such that a register use would increment the
754 // pressure of the register class's representative and all of it's super
755 // classes' representatives transitively. We have not implemented this because
756 // of the difficulty prior to coalescing of modeling operand register classes
757 // due to the common occurrence of cross class copies and subregister insertions
759 std::pair<const TargetRegisterClass*, uint8_t>
760 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
761 const TargetRegisterClass *RRC = 0;
763 switch (VT.getSimpleVT().SimpleTy) {
765 return TargetLowering::findRepresentativeClass(VT);
766 // Use DPR as representative register class for all floating point
767 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
768 // the cost is 1 for both f32 and f64.
769 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
770 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
771 RRC = ARM::DPRRegisterClass;
772 // When NEON is used for SP, only half of the register file is available
773 // because operations that define both SP and DP results will be constrained
774 // to the VFP2 class (D0-D15). We currently model this constraint prior to
775 // coalescing by double-counting the SP regs. See the FIXME above.
776 if (Subtarget->useNEONForSinglePrecisionFP())
779 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
780 case MVT::v4f32: case MVT::v2f64:
781 RRC = ARM::DPRRegisterClass;
785 RRC = ARM::DPRRegisterClass;
789 RRC = ARM::DPRRegisterClass;
793 return std::make_pair(RRC, Cost);
796 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
799 case ARMISD::Wrapper: return "ARMISD::Wrapper";
800 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
801 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
802 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
803 case ARMISD::CALL: return "ARMISD::CALL";
804 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
805 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
806 case ARMISD::tCALL: return "ARMISD::tCALL";
807 case ARMISD::BRCOND: return "ARMISD::BRCOND";
808 case ARMISD::BR_JT: return "ARMISD::BR_JT";
809 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
810 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
811 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
812 case ARMISD::CMP: return "ARMISD::CMP";
813 case ARMISD::CMPZ: return "ARMISD::CMPZ";
814 case ARMISD::CMPFP: return "ARMISD::CMPFP";
815 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
816 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
817 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
818 case ARMISD::CMOV: return "ARMISD::CMOV";
820 case ARMISD::RBIT: return "ARMISD::RBIT";
822 case ARMISD::FTOSI: return "ARMISD::FTOSI";
823 case ARMISD::FTOUI: return "ARMISD::FTOUI";
824 case ARMISD::SITOF: return "ARMISD::SITOF";
825 case ARMISD::UITOF: return "ARMISD::UITOF";
827 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
828 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
829 case ARMISD::RRX: return "ARMISD::RRX";
831 case ARMISD::ADDC: return "ARMISD::ADDC";
832 case ARMISD::ADDE: return "ARMISD::ADDE";
833 case ARMISD::SUBC: return "ARMISD::SUBC";
834 case ARMISD::SUBE: return "ARMISD::SUBE";
836 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
837 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
839 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
840 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
841 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
843 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
845 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
847 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
849 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
850 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
852 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
854 case ARMISD::VCEQ: return "ARMISD::VCEQ";
855 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
856 case ARMISD::VCGE: return "ARMISD::VCGE";
857 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
858 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
859 case ARMISD::VCGEU: return "ARMISD::VCGEU";
860 case ARMISD::VCGT: return "ARMISD::VCGT";
861 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
862 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
863 case ARMISD::VCGTU: return "ARMISD::VCGTU";
864 case ARMISD::VTST: return "ARMISD::VTST";
866 case ARMISD::VSHL: return "ARMISD::VSHL";
867 case ARMISD::VSHRs: return "ARMISD::VSHRs";
868 case ARMISD::VSHRu: return "ARMISD::VSHRu";
869 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
870 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
871 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
872 case ARMISD::VSHRN: return "ARMISD::VSHRN";
873 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
874 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
875 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
876 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
877 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
878 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
879 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
880 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
881 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
882 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
883 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
884 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
885 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
886 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
887 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
888 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
889 case ARMISD::VDUP: return "ARMISD::VDUP";
890 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
891 case ARMISD::VEXT: return "ARMISD::VEXT";
892 case ARMISD::VREV64: return "ARMISD::VREV64";
893 case ARMISD::VREV32: return "ARMISD::VREV32";
894 case ARMISD::VREV16: return "ARMISD::VREV16";
895 case ARMISD::VZIP: return "ARMISD::VZIP";
896 case ARMISD::VUZP: return "ARMISD::VUZP";
897 case ARMISD::VTRN: return "ARMISD::VTRN";
898 case ARMISD::VTBL1: return "ARMISD::VTBL1";
899 case ARMISD::VTBL2: return "ARMISD::VTBL2";
900 case ARMISD::VMULLs: return "ARMISD::VMULLs";
901 case ARMISD::VMULLu: return "ARMISD::VMULLu";
902 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
903 case ARMISD::FMAX: return "ARMISD::FMAX";
904 case ARMISD::FMIN: return "ARMISD::FMIN";
905 case ARMISD::BFI: return "ARMISD::BFI";
906 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
907 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
908 case ARMISD::VBSL: return "ARMISD::VBSL";
909 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
910 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
911 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
912 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
913 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
914 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
915 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
916 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
917 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
918 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
919 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
920 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
921 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
922 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
923 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
924 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
925 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
926 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
927 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
928 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
932 /// getRegClassFor - Return the register class that should be used for the
933 /// specified value type.
934 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
935 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
936 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
937 // load / store 4 to 8 consecutive D registers.
938 if (Subtarget->hasNEON()) {
939 if (VT == MVT::v4i64)
940 return ARM::QQPRRegisterClass;
941 else if (VT == MVT::v8i64)
942 return ARM::QQQQPRRegisterClass;
944 return TargetLowering::getRegClassFor(VT);
947 // Create a fast isel object.
949 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
950 return ARM::createFastISel(funcInfo);
953 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
954 /// be used for loads / stores from the global.
955 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
956 return (Subtarget->isThumb1Only() ? 127 : 4095);
959 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
960 unsigned NumVals = N->getNumValues();
962 return Sched::RegPressure;
964 for (unsigned i = 0; i != NumVals; ++i) {
965 EVT VT = N->getValueType(i);
966 if (VT == MVT::Glue || VT == MVT::Other)
968 if (VT.isFloatingPoint() || VT.isVector())
969 return Sched::Latency;
972 if (!N->isMachineOpcode())
973 return Sched::RegPressure;
975 // Load are scheduled for latency even if there instruction itinerary
977 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
978 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
980 if (MCID.getNumDefs() == 0)
981 return Sched::RegPressure;
982 if (!Itins->isEmpty() &&
983 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
984 return Sched::Latency;
986 return Sched::RegPressure;
989 //===----------------------------------------------------------------------===//
991 //===----------------------------------------------------------------------===//
993 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
994 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
996 default: llvm_unreachable("Unknown condition code!");
997 case ISD::SETNE: return ARMCC::NE;
998 case ISD::SETEQ: return ARMCC::EQ;
999 case ISD::SETGT: return ARMCC::GT;
1000 case ISD::SETGE: return ARMCC::GE;
1001 case ISD::SETLT: return ARMCC::LT;
1002 case ISD::SETLE: return ARMCC::LE;
1003 case ISD::SETUGT: return ARMCC::HI;
1004 case ISD::SETUGE: return ARMCC::HS;
1005 case ISD::SETULT: return ARMCC::LO;
1006 case ISD::SETULE: return ARMCC::LS;
1010 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1011 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1012 ARMCC::CondCodes &CondCode2) {
1013 CondCode2 = ARMCC::AL;
1015 default: llvm_unreachable("Unknown FP condition!");
1017 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1019 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1021 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1022 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1023 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1024 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1025 case ISD::SETO: CondCode = ARMCC::VC; break;
1026 case ISD::SETUO: CondCode = ARMCC::VS; break;
1027 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1028 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1029 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1031 case ISD::SETULT: CondCode = ARMCC::LT; break;
1033 case ISD::SETULE: CondCode = ARMCC::LE; break;
1035 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1039 //===----------------------------------------------------------------------===//
1040 // Calling Convention Implementation
1041 //===----------------------------------------------------------------------===//
1043 #include "ARMGenCallingConv.inc"
1045 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1046 /// given CallingConvention value.
1047 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1049 bool isVarArg) const {
1052 llvm_unreachable("Unsupported calling convention");
1053 case CallingConv::Fast:
1054 if (Subtarget->hasVFP2() && !isVarArg) {
1055 if (!Subtarget->isAAPCS_ABI())
1056 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1057 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1058 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1061 case CallingConv::C: {
1062 // Use target triple & subtarget features to do actual dispatch.
1063 if (!Subtarget->isAAPCS_ABI())
1064 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1065 else if (Subtarget->hasVFP2() &&
1066 FloatABIType == FloatABI::Hard && !isVarArg)
1067 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1068 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1070 case CallingConv::ARM_AAPCS_VFP:
1071 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1072 case CallingConv::ARM_AAPCS:
1073 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1074 case CallingConv::ARM_APCS:
1075 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1079 /// LowerCallResult - Lower the result values of a call into the
1080 /// appropriate copies out of appropriate physical registers.
1082 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1083 CallingConv::ID CallConv, bool isVarArg,
1084 const SmallVectorImpl<ISD::InputArg> &Ins,
1085 DebugLoc dl, SelectionDAG &DAG,
1086 SmallVectorImpl<SDValue> &InVals) const {
1088 // Assign locations to each value returned by this call.
1089 SmallVector<CCValAssign, 16> RVLocs;
1090 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1091 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1092 CCInfo.AnalyzeCallResult(Ins,
1093 CCAssignFnForNode(CallConv, /* Return*/ true,
1096 // Copy all of the result registers out of their specified physreg.
1097 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1098 CCValAssign VA = RVLocs[i];
1101 if (VA.needsCustom()) {
1102 // Handle f64 or half of a v2f64.
1103 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1105 Chain = Lo.getValue(1);
1106 InFlag = Lo.getValue(2);
1107 VA = RVLocs[++i]; // skip ahead to next loc
1108 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1110 Chain = Hi.getValue(1);
1111 InFlag = Hi.getValue(2);
1112 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1114 if (VA.getLocVT() == MVT::v2f64) {
1115 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1116 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1117 DAG.getConstant(0, MVT::i32));
1119 VA = RVLocs[++i]; // skip ahead to next loc
1120 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1121 Chain = Lo.getValue(1);
1122 InFlag = Lo.getValue(2);
1123 VA = RVLocs[++i]; // skip ahead to next loc
1124 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1125 Chain = Hi.getValue(1);
1126 InFlag = Hi.getValue(2);
1127 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1128 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1129 DAG.getConstant(1, MVT::i32));
1132 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1134 Chain = Val.getValue(1);
1135 InFlag = Val.getValue(2);
1138 switch (VA.getLocInfo()) {
1139 default: llvm_unreachable("Unknown loc info!");
1140 case CCValAssign::Full: break;
1141 case CCValAssign::BCvt:
1142 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1146 InVals.push_back(Val);
1152 /// LowerMemOpCallTo - Store the argument to the stack.
1154 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1155 SDValue StackPtr, SDValue Arg,
1156 DebugLoc dl, SelectionDAG &DAG,
1157 const CCValAssign &VA,
1158 ISD::ArgFlagsTy Flags) const {
1159 unsigned LocMemOffset = VA.getLocMemOffset();
1160 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1161 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1162 return DAG.getStore(Chain, dl, Arg, PtrOff,
1163 MachinePointerInfo::getStack(LocMemOffset),
1167 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1168 SDValue Chain, SDValue &Arg,
1169 RegsToPassVector &RegsToPass,
1170 CCValAssign &VA, CCValAssign &NextVA,
1172 SmallVector<SDValue, 8> &MemOpChains,
1173 ISD::ArgFlagsTy Flags) const {
1175 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1176 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1177 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1179 if (NextVA.isRegLoc())
1180 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1182 assert(NextVA.isMemLoc());
1183 if (StackPtr.getNode() == 0)
1184 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1186 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1192 /// LowerCall - Lowering a call into a callseq_start <-
1193 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1196 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1197 CallingConv::ID CallConv, bool isVarArg,
1199 const SmallVectorImpl<ISD::OutputArg> &Outs,
1200 const SmallVectorImpl<SDValue> &OutVals,
1201 const SmallVectorImpl<ISD::InputArg> &Ins,
1202 DebugLoc dl, SelectionDAG &DAG,
1203 SmallVectorImpl<SDValue> &InVals) const {
1204 MachineFunction &MF = DAG.getMachineFunction();
1205 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1206 bool IsSibCall = false;
1207 // Temporarily disable tail calls so things don't break.
1208 if (!EnableARMTailCalls)
1211 // Check if it's really possible to do a tail call.
1212 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1213 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1214 Outs, OutVals, Ins, DAG);
1215 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1216 // detected sibcalls.
1223 // Analyze operands of the call, assigning locations to each operand.
1224 SmallVector<CCValAssign, 16> ArgLocs;
1225 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1226 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1227 CCInfo.AnalyzeCallOperands(Outs,
1228 CCAssignFnForNode(CallConv, /* Return*/ false,
1231 // Get a count of how many bytes are to be pushed on the stack.
1232 unsigned NumBytes = CCInfo.getNextStackOffset();
1234 // For tail calls, memory operands are available in our caller's stack.
1238 // Adjust the stack pointer for the new arguments...
1239 // These operations are automatically eliminated by the prolog/epilog pass
1241 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1243 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1245 RegsToPassVector RegsToPass;
1246 SmallVector<SDValue, 8> MemOpChains;
1248 // Walk the register/memloc assignments, inserting copies/loads. In the case
1249 // of tail call optimization, arguments are handled later.
1250 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1252 ++i, ++realArgIdx) {
1253 CCValAssign &VA = ArgLocs[i];
1254 SDValue Arg = OutVals[realArgIdx];
1255 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1256 bool isByVal = Flags.isByVal();
1258 // Promote the value if needed.
1259 switch (VA.getLocInfo()) {
1260 default: llvm_unreachable("Unknown loc info!");
1261 case CCValAssign::Full: break;
1262 case CCValAssign::SExt:
1263 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1265 case CCValAssign::ZExt:
1266 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1268 case CCValAssign::AExt:
1269 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1271 case CCValAssign::BCvt:
1272 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1276 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1277 if (VA.needsCustom()) {
1278 if (VA.getLocVT() == MVT::v2f64) {
1279 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1280 DAG.getConstant(0, MVT::i32));
1281 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1282 DAG.getConstant(1, MVT::i32));
1284 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1285 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1287 VA = ArgLocs[++i]; // skip ahead to next loc
1288 if (VA.isRegLoc()) {
1289 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1290 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1292 assert(VA.isMemLoc());
1294 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1295 dl, DAG, VA, Flags));
1298 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1299 StackPtr, MemOpChains, Flags);
1301 } else if (VA.isRegLoc()) {
1302 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1303 } else if (isByVal) {
1304 assert(VA.isMemLoc());
1305 unsigned offset = 0;
1307 // True if this byval aggregate will be split between registers
1309 if (CCInfo.isFirstByValRegValid()) {
1310 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1312 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1313 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1314 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1315 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1316 MachinePointerInfo(),
1318 MemOpChains.push_back(Load.getValue(1));
1319 RegsToPass.push_back(std::make_pair(j, Load));
1321 offset = ARM::R4 - CCInfo.getFirstByValReg();
1322 CCInfo.clearFirstByValReg();
1325 unsigned LocMemOffset = VA.getLocMemOffset();
1326 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1327 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1329 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1330 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1331 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1333 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1334 Flags.getByValAlign(),
1335 /*isVolatile=*/false,
1336 /*AlwaysInline=*/false,
1337 MachinePointerInfo(0),
1338 MachinePointerInfo(0)));
1340 } else if (!IsSibCall) {
1341 assert(VA.isMemLoc());
1343 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1344 dl, DAG, VA, Flags));
1348 if (!MemOpChains.empty())
1349 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1350 &MemOpChains[0], MemOpChains.size());
1352 // Build a sequence of copy-to-reg nodes chained together with token chain
1353 // and flag operands which copy the outgoing args into the appropriate regs.
1355 // Tail call byval lowering might overwrite argument registers so in case of
1356 // tail call optimization the copies to registers are lowered later.
1358 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1359 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1360 RegsToPass[i].second, InFlag);
1361 InFlag = Chain.getValue(1);
1364 // For tail calls lower the arguments to the 'real' stack slot.
1366 // Force all the incoming stack arguments to be loaded from the stack
1367 // before any new outgoing arguments are stored to the stack, because the
1368 // outgoing stack slots may alias the incoming argument stack slots, and
1369 // the alias isn't otherwise explicit. This is slightly more conservative
1370 // than necessary, because it means that each store effectively depends
1371 // on every argument instead of just those arguments it would clobber.
1373 // Do not flag preceding copytoreg stuff together with the following stuff.
1375 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1376 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1377 RegsToPass[i].second, InFlag);
1378 InFlag = Chain.getValue(1);
1383 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1384 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1385 // node so that legalize doesn't hack it.
1386 bool isDirect = false;
1387 bool isARMFunc = false;
1388 bool isLocalARMFunc = false;
1389 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1391 if (EnableARMLongCalls) {
1392 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1393 && "long-calls with non-static relocation model!");
1394 // Handle a global address or an external symbol. If it's not one of
1395 // those, the target's already in a register, so we don't need to do
1397 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1398 const GlobalValue *GV = G->getGlobal();
1399 // Create a constant pool entry for the callee address
1400 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1401 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1404 // Get the address of the callee into a register
1405 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1406 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1407 Callee = DAG.getLoad(getPointerTy(), dl,
1408 DAG.getEntryNode(), CPAddr,
1409 MachinePointerInfo::getConstantPool(),
1411 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1412 const char *Sym = S->getSymbol();
1414 // Create a constant pool entry for the callee address
1415 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1416 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1417 Sym, ARMPCLabelIndex, 0);
1418 // Get the address of the callee into a register
1419 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1420 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1421 Callee = DAG.getLoad(getPointerTy(), dl,
1422 DAG.getEntryNode(), CPAddr,
1423 MachinePointerInfo::getConstantPool(),
1426 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1427 const GlobalValue *GV = G->getGlobal();
1429 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1430 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1431 getTargetMachine().getRelocationModel() != Reloc::Static;
1432 isARMFunc = !Subtarget->isThumb() || isStub;
1433 // ARM call to a local ARM function is predicable.
1434 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1435 // tBX takes a register source operand.
1436 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1437 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1438 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1441 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1442 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1443 Callee = DAG.getLoad(getPointerTy(), dl,
1444 DAG.getEntryNode(), CPAddr,
1445 MachinePointerInfo::getConstantPool(),
1447 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1448 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1449 getPointerTy(), Callee, PICLabel);
1451 // On ELF targets for PIC code, direct calls should go through the PLT
1452 unsigned OpFlags = 0;
1453 if (Subtarget->isTargetELF() &&
1454 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1455 OpFlags = ARMII::MO_PLT;
1456 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1458 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1460 bool isStub = Subtarget->isTargetDarwin() &&
1461 getTargetMachine().getRelocationModel() != Reloc::Static;
1462 isARMFunc = !Subtarget->isThumb() || isStub;
1463 // tBX takes a register source operand.
1464 const char *Sym = S->getSymbol();
1465 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1466 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1467 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1468 Sym, ARMPCLabelIndex, 4);
1469 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1470 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1471 Callee = DAG.getLoad(getPointerTy(), dl,
1472 DAG.getEntryNode(), CPAddr,
1473 MachinePointerInfo::getConstantPool(),
1475 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1476 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1477 getPointerTy(), Callee, PICLabel);
1479 unsigned OpFlags = 0;
1480 // On ELF targets for PIC code, direct calls should go through the PLT
1481 if (Subtarget->isTargetELF() &&
1482 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1483 OpFlags = ARMII::MO_PLT;
1484 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1488 // FIXME: handle tail calls differently.
1490 if (Subtarget->isThumb()) {
1491 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1492 CallOpc = ARMISD::CALL_NOLINK;
1494 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1496 CallOpc = (isDirect || Subtarget->hasV5TOps())
1497 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1498 : ARMISD::CALL_NOLINK;
1501 std::vector<SDValue> Ops;
1502 Ops.push_back(Chain);
1503 Ops.push_back(Callee);
1505 // Add argument registers to the end of the list so that they are known live
1507 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1508 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1509 RegsToPass[i].second.getValueType()));
1511 if (InFlag.getNode())
1512 Ops.push_back(InFlag);
1514 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1516 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1518 // Returns a chain and a flag for retval copy to use.
1519 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1520 InFlag = Chain.getValue(1);
1522 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1523 DAG.getIntPtrConstant(0, true), InFlag);
1525 InFlag = Chain.getValue(1);
1527 // Handle result values, copying them out of physregs into vregs that we
1529 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1533 /// HandleByVal - Every parameter *after* a byval parameter is passed
1534 /// on the stack. Remember the next parameter register to allocate,
1535 /// and then confiscate the rest of the parameter registers to insure
1538 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1539 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1540 assert((State->getCallOrPrologue() == Prologue ||
1541 State->getCallOrPrologue() == Call) &&
1542 "unhandled ParmContext");
1543 if ((!State->isFirstByValRegValid()) &&
1544 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1545 State->setFirstByValReg(reg);
1546 // At a call site, a byval parameter that is split between
1547 // registers and memory needs its size truncated here. In a
1548 // function prologue, such byval parameters are reassembled in
1549 // memory, and are not truncated.
1550 if (State->getCallOrPrologue() == Call) {
1551 unsigned excess = 4 * (ARM::R4 - reg);
1552 assert(size >= excess && "expected larger existing stack allocation");
1556 // Confiscate any remaining parameter registers to preclude their
1557 // assignment to subsequent parameters.
1558 while (State->AllocateReg(GPRArgRegs, 4))
1562 /// MatchingStackOffset - Return true if the given stack call argument is
1563 /// already available in the same position (relatively) of the caller's
1564 /// incoming argument stack.
1566 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1567 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1568 const ARMInstrInfo *TII) {
1569 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1571 if (Arg.getOpcode() == ISD::CopyFromReg) {
1572 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1573 if (!TargetRegisterInfo::isVirtualRegister(VR))
1575 MachineInstr *Def = MRI->getVRegDef(VR);
1578 if (!Flags.isByVal()) {
1579 if (!TII->isLoadFromStackSlot(Def, FI))
1584 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1585 if (Flags.isByVal())
1586 // ByVal argument is passed in as a pointer but it's now being
1587 // dereferenced. e.g.
1588 // define @foo(%struct.X* %A) {
1589 // tail call @bar(%struct.X* byval %A)
1592 SDValue Ptr = Ld->getBasePtr();
1593 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1596 FI = FINode->getIndex();
1600 assert(FI != INT_MAX);
1601 if (!MFI->isFixedObjectIndex(FI))
1603 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1606 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1607 /// for tail call optimization. Targets which want to do tail call
1608 /// optimization should implement this function.
1610 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1611 CallingConv::ID CalleeCC,
1613 bool isCalleeStructRet,
1614 bool isCallerStructRet,
1615 const SmallVectorImpl<ISD::OutputArg> &Outs,
1616 const SmallVectorImpl<SDValue> &OutVals,
1617 const SmallVectorImpl<ISD::InputArg> &Ins,
1618 SelectionDAG& DAG) const {
1619 const Function *CallerF = DAG.getMachineFunction().getFunction();
1620 CallingConv::ID CallerCC = CallerF->getCallingConv();
1621 bool CCMatch = CallerCC == CalleeCC;
1623 // Look for obvious safe cases to perform tail call optimization that do not
1624 // require ABI changes. This is what gcc calls sibcall.
1626 // Do not sibcall optimize vararg calls unless the call site is not passing
1628 if (isVarArg && !Outs.empty())
1631 // Also avoid sibcall optimization if either caller or callee uses struct
1632 // return semantics.
1633 if (isCalleeStructRet || isCallerStructRet)
1636 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1637 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1638 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1639 // support in the assembler and linker to be used. This would need to be
1640 // fixed to fully support tail calls in Thumb1.
1642 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1643 // LR. This means if we need to reload LR, it takes an extra instructions,
1644 // which outweighs the value of the tail call; but here we don't know yet
1645 // whether LR is going to be used. Probably the right approach is to
1646 // generate the tail call here and turn it back into CALL/RET in
1647 // emitEpilogue if LR is used.
1649 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1650 // but we need to make sure there are enough registers; the only valid
1651 // registers are the 4 used for parameters. We don't currently do this
1653 if (Subtarget->isThumb1Only())
1656 // If the calling conventions do not match, then we'd better make sure the
1657 // results are returned in the same way as what the caller expects.
1659 SmallVector<CCValAssign, 16> RVLocs1;
1660 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1661 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1662 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1664 SmallVector<CCValAssign, 16> RVLocs2;
1665 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1666 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1667 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1669 if (RVLocs1.size() != RVLocs2.size())
1671 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1672 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1674 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1676 if (RVLocs1[i].isRegLoc()) {
1677 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1680 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1686 // If the callee takes no arguments then go on to check the results of the
1688 if (!Outs.empty()) {
1689 // Check if stack adjustment is needed. For now, do not do this if any
1690 // argument is passed on the stack.
1691 SmallVector<CCValAssign, 16> ArgLocs;
1692 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1693 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1694 CCInfo.AnalyzeCallOperands(Outs,
1695 CCAssignFnForNode(CalleeCC, false, isVarArg));
1696 if (CCInfo.getNextStackOffset()) {
1697 MachineFunction &MF = DAG.getMachineFunction();
1699 // Check if the arguments are already laid out in the right way as
1700 // the caller's fixed stack objects.
1701 MachineFrameInfo *MFI = MF.getFrameInfo();
1702 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1703 const ARMInstrInfo *TII =
1704 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1705 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1707 ++i, ++realArgIdx) {
1708 CCValAssign &VA = ArgLocs[i];
1709 EVT RegVT = VA.getLocVT();
1710 SDValue Arg = OutVals[realArgIdx];
1711 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1712 if (VA.getLocInfo() == CCValAssign::Indirect)
1714 if (VA.needsCustom()) {
1715 // f64 and vector types are split into multiple registers or
1716 // register/stack-slot combinations. The types will not match
1717 // the registers; give up on memory f64 refs until we figure
1718 // out what to do about this.
1721 if (!ArgLocs[++i].isRegLoc())
1723 if (RegVT == MVT::v2f64) {
1724 if (!ArgLocs[++i].isRegLoc())
1726 if (!ArgLocs[++i].isRegLoc())
1729 } else if (!VA.isRegLoc()) {
1730 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1742 ARMTargetLowering::LowerReturn(SDValue Chain,
1743 CallingConv::ID CallConv, bool isVarArg,
1744 const SmallVectorImpl<ISD::OutputArg> &Outs,
1745 const SmallVectorImpl<SDValue> &OutVals,
1746 DebugLoc dl, SelectionDAG &DAG) const {
1748 // CCValAssign - represent the assignment of the return value to a location.
1749 SmallVector<CCValAssign, 16> RVLocs;
1751 // CCState - Info about the registers and stack slots.
1752 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1753 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1755 // Analyze outgoing return values.
1756 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1759 // If this is the first return lowered for this function, add
1760 // the regs to the liveout set for the function.
1761 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1762 for (unsigned i = 0; i != RVLocs.size(); ++i)
1763 if (RVLocs[i].isRegLoc())
1764 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1769 // Copy the result values into the output registers.
1770 for (unsigned i = 0, realRVLocIdx = 0;
1772 ++i, ++realRVLocIdx) {
1773 CCValAssign &VA = RVLocs[i];
1774 assert(VA.isRegLoc() && "Can only return in registers!");
1776 SDValue Arg = OutVals[realRVLocIdx];
1778 switch (VA.getLocInfo()) {
1779 default: llvm_unreachable("Unknown loc info!");
1780 case CCValAssign::Full: break;
1781 case CCValAssign::BCvt:
1782 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1786 if (VA.needsCustom()) {
1787 if (VA.getLocVT() == MVT::v2f64) {
1788 // Extract the first half and return it in two registers.
1789 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1790 DAG.getConstant(0, MVT::i32));
1791 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1792 DAG.getVTList(MVT::i32, MVT::i32), Half);
1794 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1795 Flag = Chain.getValue(1);
1796 VA = RVLocs[++i]; // skip ahead to next loc
1797 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1798 HalfGPRs.getValue(1), Flag);
1799 Flag = Chain.getValue(1);
1800 VA = RVLocs[++i]; // skip ahead to next loc
1802 // Extract the 2nd half and fall through to handle it as an f64 value.
1803 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1804 DAG.getConstant(1, MVT::i32));
1806 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1808 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1809 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1810 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1811 Flag = Chain.getValue(1);
1812 VA = RVLocs[++i]; // skip ahead to next loc
1813 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1816 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1818 // Guarantee that all emitted copies are
1819 // stuck together, avoiding something bad.
1820 Flag = Chain.getValue(1);
1825 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1827 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1832 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1833 if (N->getNumValues() != 1)
1835 if (!N->hasNUsesOfValue(1, 0))
1838 unsigned NumCopies = 0;
1840 SDNode *Use = *N->use_begin();
1841 if (Use->getOpcode() == ISD::CopyToReg) {
1842 Copies[NumCopies++] = Use;
1843 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1844 // f64 returned in a pair of GPRs.
1845 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1847 if (UI->getOpcode() != ISD::CopyToReg)
1849 Copies[UI.getUse().getResNo()] = *UI;
1852 } else if (Use->getOpcode() == ISD::BITCAST) {
1853 // f32 returned in a single GPR.
1854 if (!Use->hasNUsesOfValue(1, 0))
1856 Use = *Use->use_begin();
1857 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1859 Copies[NumCopies++] = Use;
1864 if (NumCopies != 1 && NumCopies != 2)
1867 bool HasRet = false;
1868 for (unsigned i = 0; i < NumCopies; ++i) {
1869 SDNode *Copy = Copies[i];
1870 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1872 if (UI->getOpcode() == ISD::CopyToReg) {
1874 if (Use == Copies[0] || Use == Copies[1])
1878 if (UI->getOpcode() != ARMISD::RET_FLAG)
1887 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1888 if (!EnableARMTailCalls)
1891 if (!CI->isTailCall())
1894 return !Subtarget->isThumb1Only();
1897 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1898 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1899 // one of the above mentioned nodes. It has to be wrapped because otherwise
1900 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1901 // be used to form addressing mode. These wrapped nodes will be selected
1903 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1904 EVT PtrVT = Op.getValueType();
1905 // FIXME there is no actual debug info here
1906 DebugLoc dl = Op.getDebugLoc();
1907 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1909 if (CP->isMachineConstantPoolEntry())
1910 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1911 CP->getAlignment());
1913 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1914 CP->getAlignment());
1915 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1918 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1919 return MachineJumpTableInfo::EK_Inline;
1922 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1923 SelectionDAG &DAG) const {
1924 MachineFunction &MF = DAG.getMachineFunction();
1925 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1926 unsigned ARMPCLabelIndex = 0;
1927 DebugLoc DL = Op.getDebugLoc();
1928 EVT PtrVT = getPointerTy();
1929 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1930 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1932 if (RelocM == Reloc::Static) {
1933 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1935 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1936 ARMPCLabelIndex = AFI->createPICLabelUId();
1937 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1938 ARMCP::CPBlockAddress,
1940 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1942 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1943 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1944 MachinePointerInfo::getConstantPool(),
1946 if (RelocM == Reloc::Static)
1948 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1949 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1952 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1954 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1955 SelectionDAG &DAG) const {
1956 DebugLoc dl = GA->getDebugLoc();
1957 EVT PtrVT = getPointerTy();
1958 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1959 MachineFunction &MF = DAG.getMachineFunction();
1960 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1961 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1962 ARMConstantPoolValue *CPV =
1963 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1964 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1965 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1966 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1967 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1968 MachinePointerInfo::getConstantPool(),
1970 SDValue Chain = Argument.getValue(1);
1972 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1973 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1975 // call __tls_get_addr.
1978 Entry.Node = Argument;
1979 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
1980 Args.push_back(Entry);
1981 // FIXME: is there useful debug info available here?
1982 std::pair<SDValue, SDValue> CallResult =
1983 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
1984 false, false, false, false,
1985 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1986 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1987 return CallResult.first;
1990 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1991 // "local exec" model.
1993 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1994 SelectionDAG &DAG) const {
1995 const GlobalValue *GV = GA->getGlobal();
1996 DebugLoc dl = GA->getDebugLoc();
1998 SDValue Chain = DAG.getEntryNode();
1999 EVT PtrVT = getPointerTy();
2000 // Get the Thread Pointer
2001 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2003 if (GV->isDeclaration()) {
2004 MachineFunction &MF = DAG.getMachineFunction();
2005 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2006 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2007 // Initial exec model.
2008 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2009 ARMConstantPoolValue *CPV =
2010 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
2011 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, true);
2012 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2013 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2014 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2015 MachinePointerInfo::getConstantPool(),
2017 Chain = Offset.getValue(1);
2019 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2020 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2022 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2023 MachinePointerInfo::getConstantPool(),
2027 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMCP::TPOFF);
2028 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2029 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2030 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2031 MachinePointerInfo::getConstantPool(),
2035 // The address of the thread local variable is the add of the thread
2036 // pointer with the offset of the variable.
2037 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2041 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2042 // TODO: implement the "local dynamic" model
2043 assert(Subtarget->isTargetELF() &&
2044 "TLS not implemented for non-ELF targets");
2045 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2046 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2047 // otherwise use the "Local Exec" TLS Model
2048 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2049 return LowerToTLSGeneralDynamicModel(GA, DAG);
2051 return LowerToTLSExecModels(GA, DAG);
2054 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2055 SelectionDAG &DAG) const {
2056 EVT PtrVT = getPointerTy();
2057 DebugLoc dl = Op.getDebugLoc();
2058 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2059 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2060 if (RelocM == Reloc::PIC_) {
2061 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2062 ARMConstantPoolValue *CPV =
2063 new ARMConstantPoolValue(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2064 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2065 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2066 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2068 MachinePointerInfo::getConstantPool(),
2070 SDValue Chain = Result.getValue(1);
2071 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2072 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2074 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2075 MachinePointerInfo::getGOT(), false, false, 0);
2079 // If we have T2 ops, we can materialize the address directly via movt/movw
2080 // pair. This is always cheaper.
2081 if (Subtarget->useMovt()) {
2083 // FIXME: Once remat is capable of dealing with instructions with register
2084 // operands, expand this into two nodes.
2085 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2086 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2088 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2089 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2090 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2091 MachinePointerInfo::getConstantPool(),
2096 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2097 SelectionDAG &DAG) const {
2098 EVT PtrVT = getPointerTy();
2099 DebugLoc dl = Op.getDebugLoc();
2100 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2101 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2102 MachineFunction &MF = DAG.getMachineFunction();
2103 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2105 // FIXME: Enable this for static codegen when tool issues are fixed.
2106 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2108 // FIXME: Once remat is capable of dealing with instructions with register
2109 // operands, expand this into two nodes.
2110 if (RelocM == Reloc::Static)
2111 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2112 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2114 unsigned Wrapper = (RelocM == Reloc::PIC_)
2115 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2116 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2117 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2118 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2119 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2120 MachinePointerInfo::getGOT(), false, false, 0);
2124 unsigned ARMPCLabelIndex = 0;
2126 if (RelocM == Reloc::Static) {
2127 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2129 ARMPCLabelIndex = AFI->createPICLabelUId();
2130 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2131 ARMConstantPoolValue *CPV =
2132 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
2133 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2135 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2137 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2138 MachinePointerInfo::getConstantPool(),
2140 SDValue Chain = Result.getValue(1);
2142 if (RelocM == Reloc::PIC_) {
2143 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2144 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2147 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2148 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2154 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2155 SelectionDAG &DAG) const {
2156 assert(Subtarget->isTargetELF() &&
2157 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2158 MachineFunction &MF = DAG.getMachineFunction();
2159 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2160 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2161 EVT PtrVT = getPointerTy();
2162 DebugLoc dl = Op.getDebugLoc();
2163 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2164 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
2165 "_GLOBAL_OFFSET_TABLE_",
2166 ARMPCLabelIndex, PCAdj);
2167 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2168 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2169 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2170 MachinePointerInfo::getConstantPool(),
2172 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2173 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2177 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2179 DebugLoc dl = Op.getDebugLoc();
2180 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2181 Op.getOperand(0), Op.getOperand(1));
2185 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2186 DebugLoc dl = Op.getDebugLoc();
2187 SDValue Val = DAG.getConstant(0, MVT::i32);
2188 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
2189 Op.getOperand(1), Val);
2193 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2194 DebugLoc dl = Op.getDebugLoc();
2195 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2196 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2200 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2201 const ARMSubtarget *Subtarget) const {
2202 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2203 DebugLoc dl = Op.getDebugLoc();
2205 default: return SDValue(); // Don't custom lower most intrinsics.
2206 case Intrinsic::arm_thread_pointer: {
2207 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2208 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2210 case Intrinsic::eh_sjlj_lsda: {
2211 MachineFunction &MF = DAG.getMachineFunction();
2212 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2213 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2214 EVT PtrVT = getPointerTy();
2215 DebugLoc dl = Op.getDebugLoc();
2216 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2218 unsigned PCAdj = (RelocM != Reloc::PIC_)
2219 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2220 ARMConstantPoolValue *CPV =
2221 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
2222 ARMCP::CPLSDA, PCAdj);
2223 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2224 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2226 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2227 MachinePointerInfo::getConstantPool(),
2230 if (RelocM == Reloc::PIC_) {
2231 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2232 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2236 case Intrinsic::arm_neon_vmulls:
2237 case Intrinsic::arm_neon_vmullu: {
2238 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2239 ? ARMISD::VMULLs : ARMISD::VMULLu;
2240 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2241 Op.getOperand(1), Op.getOperand(2));
2246 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2247 const ARMSubtarget *Subtarget) {
2248 DebugLoc dl = Op.getDebugLoc();
2249 if (!Subtarget->hasDataBarrier()) {
2250 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2251 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2253 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2254 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2255 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2256 DAG.getConstant(0, MVT::i32));
2259 SDValue Op5 = Op.getOperand(5);
2260 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2261 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2262 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2263 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2265 ARM_MB::MemBOpt DMBOpt;
2266 if (isDeviceBarrier)
2267 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2269 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2270 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2271 DAG.getConstant(DMBOpt, MVT::i32));
2275 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2276 const ARMSubtarget *Subtarget) {
2277 // FIXME: handle "fence singlethread" more efficiently.
2278 DebugLoc dl = Op.getDebugLoc();
2279 if (!Subtarget->hasDataBarrier()) {
2280 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2281 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2283 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2284 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2285 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2286 DAG.getConstant(0, MVT::i32));
2289 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2290 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2293 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2294 const ARMSubtarget *Subtarget) {
2295 // ARM pre v5TE and Thumb1 does not have preload instructions.
2296 if (!(Subtarget->isThumb2() ||
2297 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2298 // Just preserve the chain.
2299 return Op.getOperand(0);
2301 DebugLoc dl = Op.getDebugLoc();
2302 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2304 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2305 // ARMv7 with MP extension has PLDW.
2306 return Op.getOperand(0);
2308 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2309 if (Subtarget->isThumb()) {
2311 isRead = ~isRead & 1;
2312 isData = ~isData & 1;
2315 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2316 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2317 DAG.getConstant(isData, MVT::i32));
2320 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2321 MachineFunction &MF = DAG.getMachineFunction();
2322 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2324 // vastart just stores the address of the VarArgsFrameIndex slot into the
2325 // memory location argument.
2326 DebugLoc dl = Op.getDebugLoc();
2327 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2328 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2329 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2330 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2331 MachinePointerInfo(SV), false, false, 0);
2335 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2336 SDValue &Root, SelectionDAG &DAG,
2337 DebugLoc dl) const {
2338 MachineFunction &MF = DAG.getMachineFunction();
2339 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2341 TargetRegisterClass *RC;
2342 if (AFI->isThumb1OnlyFunction())
2343 RC = ARM::tGPRRegisterClass;
2345 RC = ARM::GPRRegisterClass;
2347 // Transform the arguments stored in physical registers into virtual ones.
2348 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2349 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2352 if (NextVA.isMemLoc()) {
2353 MachineFrameInfo *MFI = MF.getFrameInfo();
2354 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2356 // Create load node to retrieve arguments from the stack.
2357 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2358 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2359 MachinePointerInfo::getFixedStack(FI),
2362 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2363 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2366 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2370 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2371 unsigned &VARegSize, unsigned &VARegSaveSize)
2374 if (CCInfo.isFirstByValRegValid())
2375 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2377 unsigned int firstUnalloced;
2378 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2379 sizeof(GPRArgRegs) /
2380 sizeof(GPRArgRegs[0]));
2381 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2384 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2385 VARegSize = NumGPRs * 4;
2386 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2389 // The remaining GPRs hold either the beginning of variable-argument
2390 // data, or the beginning of an aggregate passed by value (usuall
2391 // byval). Either way, we allocate stack slots adjacent to the data
2392 // provided by our caller, and store the unallocated registers there.
2393 // If this is a variadic function, the va_list pointer will begin with
2394 // these values; otherwise, this reassembles a (byval) structure that
2395 // was split between registers and memory.
2397 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2398 DebugLoc dl, SDValue &Chain,
2399 unsigned ArgOffset) const {
2400 MachineFunction &MF = DAG.getMachineFunction();
2401 MachineFrameInfo *MFI = MF.getFrameInfo();
2402 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2403 unsigned firstRegToSaveIndex;
2404 if (CCInfo.isFirstByValRegValid())
2405 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2407 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2408 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2411 unsigned VARegSize, VARegSaveSize;
2412 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2413 if (VARegSaveSize) {
2414 // If this function is vararg, store any remaining integer argument regs
2415 // to their spots on the stack so that they may be loaded by deferencing
2416 // the result of va_next.
2417 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2418 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2419 ArgOffset + VARegSaveSize
2422 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2425 SmallVector<SDValue, 4> MemOps;
2426 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2427 TargetRegisterClass *RC;
2428 if (AFI->isThumb1OnlyFunction())
2429 RC = ARM::tGPRRegisterClass;
2431 RC = ARM::GPRRegisterClass;
2433 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2434 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2436 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2437 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2439 MemOps.push_back(Store);
2440 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2441 DAG.getConstant(4, getPointerTy()));
2443 if (!MemOps.empty())
2444 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2445 &MemOps[0], MemOps.size());
2447 // This will point to the next argument passed via stack.
2448 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2452 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2453 CallingConv::ID CallConv, bool isVarArg,
2454 const SmallVectorImpl<ISD::InputArg>
2456 DebugLoc dl, SelectionDAG &DAG,
2457 SmallVectorImpl<SDValue> &InVals)
2459 MachineFunction &MF = DAG.getMachineFunction();
2460 MachineFrameInfo *MFI = MF.getFrameInfo();
2462 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2464 // Assign locations to all of the incoming arguments.
2465 SmallVector<CCValAssign, 16> ArgLocs;
2466 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2467 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2468 CCInfo.AnalyzeFormalArguments(Ins,
2469 CCAssignFnForNode(CallConv, /* Return*/ false,
2472 SmallVector<SDValue, 16> ArgValues;
2473 int lastInsIndex = -1;
2476 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2477 CCValAssign &VA = ArgLocs[i];
2479 // Arguments stored in registers.
2480 if (VA.isRegLoc()) {
2481 EVT RegVT = VA.getLocVT();
2483 if (VA.needsCustom()) {
2484 // f64 and vector types are split up into multiple registers or
2485 // combinations of registers and stack slots.
2486 if (VA.getLocVT() == MVT::v2f64) {
2487 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2489 VA = ArgLocs[++i]; // skip ahead to next loc
2491 if (VA.isMemLoc()) {
2492 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2493 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2494 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2495 MachinePointerInfo::getFixedStack(FI),
2498 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2501 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2502 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2503 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2504 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2505 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2507 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2510 TargetRegisterClass *RC;
2512 if (RegVT == MVT::f32)
2513 RC = ARM::SPRRegisterClass;
2514 else if (RegVT == MVT::f64)
2515 RC = ARM::DPRRegisterClass;
2516 else if (RegVT == MVT::v2f64)
2517 RC = ARM::QPRRegisterClass;
2518 else if (RegVT == MVT::i32)
2519 RC = (AFI->isThumb1OnlyFunction() ?
2520 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2522 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2524 // Transform the arguments in physical registers into virtual ones.
2525 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2526 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2529 // If this is an 8 or 16-bit value, it is really passed promoted
2530 // to 32 bits. Insert an assert[sz]ext to capture this, then
2531 // truncate to the right size.
2532 switch (VA.getLocInfo()) {
2533 default: llvm_unreachable("Unknown loc info!");
2534 case CCValAssign::Full: break;
2535 case CCValAssign::BCvt:
2536 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2538 case CCValAssign::SExt:
2539 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2540 DAG.getValueType(VA.getValVT()));
2541 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2543 case CCValAssign::ZExt:
2544 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2545 DAG.getValueType(VA.getValVT()));
2546 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2550 InVals.push_back(ArgValue);
2552 } else { // VA.isRegLoc()
2555 assert(VA.isMemLoc());
2556 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2558 int index = ArgLocs[i].getValNo();
2560 // Some Ins[] entries become multiple ArgLoc[] entries.
2561 // Process them only once.
2562 if (index != lastInsIndex)
2564 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2565 // FIXME: For now, all byval parameter objects are marked mutable.
2566 // This can be changed with more analysis.
2567 // In case of tail call optimization mark all arguments mutable.
2568 // Since they could be overwritten by lowering of arguments in case of
2570 if (Flags.isByVal()) {
2571 unsigned VARegSize, VARegSaveSize;
2572 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2573 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2574 unsigned Bytes = Flags.getByValSize() - VARegSize;
2575 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2576 int FI = MFI->CreateFixedObject(Bytes,
2577 VA.getLocMemOffset(), false);
2578 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2580 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2581 VA.getLocMemOffset(), true);
2583 // Create load nodes to retrieve arguments from the stack.
2584 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2585 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2586 MachinePointerInfo::getFixedStack(FI),
2589 lastInsIndex = index;
2596 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2601 /// isFloatingPointZero - Return true if this is +0.0.
2602 static bool isFloatingPointZero(SDValue Op) {
2603 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2604 return CFP->getValueAPF().isPosZero();
2605 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2606 // Maybe this has already been legalized into the constant pool?
2607 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2608 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2609 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2610 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2611 return CFP->getValueAPF().isPosZero();
2617 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2618 /// the given operands.
2620 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2621 SDValue &ARMcc, SelectionDAG &DAG,
2622 DebugLoc dl) const {
2623 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2624 unsigned C = RHSC->getZExtValue();
2625 if (!isLegalICmpImmediate(C)) {
2626 // Constant does not fit, try adjusting it by one?
2631 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2632 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2633 RHS = DAG.getConstant(C-1, MVT::i32);
2638 if (C != 0 && isLegalICmpImmediate(C-1)) {
2639 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2640 RHS = DAG.getConstant(C-1, MVT::i32);
2645 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2646 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2647 RHS = DAG.getConstant(C+1, MVT::i32);
2652 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2653 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2654 RHS = DAG.getConstant(C+1, MVT::i32);
2661 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2662 ARMISD::NodeType CompareType;
2665 CompareType = ARMISD::CMP;
2670 CompareType = ARMISD::CMPZ;
2673 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2674 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2677 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2679 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2680 DebugLoc dl) const {
2682 if (!isFloatingPointZero(RHS))
2683 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2685 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2686 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2689 /// duplicateCmp - Glue values can have only one use, so this function
2690 /// duplicates a comparison node.
2692 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2693 unsigned Opc = Cmp.getOpcode();
2694 DebugLoc DL = Cmp.getDebugLoc();
2695 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2696 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2698 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2699 Cmp = Cmp.getOperand(0);
2700 Opc = Cmp.getOpcode();
2701 if (Opc == ARMISD::CMPFP)
2702 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2704 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2705 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2707 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2710 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2711 SDValue Cond = Op.getOperand(0);
2712 SDValue SelectTrue = Op.getOperand(1);
2713 SDValue SelectFalse = Op.getOperand(2);
2714 DebugLoc dl = Op.getDebugLoc();
2718 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2719 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2721 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2722 const ConstantSDNode *CMOVTrue =
2723 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2724 const ConstantSDNode *CMOVFalse =
2725 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2727 if (CMOVTrue && CMOVFalse) {
2728 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2729 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2733 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2735 False = SelectFalse;
2736 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2741 if (True.getNode() && False.getNode()) {
2742 EVT VT = Op.getValueType();
2743 SDValue ARMcc = Cond.getOperand(2);
2744 SDValue CCR = Cond.getOperand(3);
2745 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2746 assert(True.getValueType() == VT);
2747 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2752 return DAG.getSelectCC(dl, Cond,
2753 DAG.getConstant(0, Cond.getValueType()),
2754 SelectTrue, SelectFalse, ISD::SETNE);
2757 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2758 EVT VT = Op.getValueType();
2759 SDValue LHS = Op.getOperand(0);
2760 SDValue RHS = Op.getOperand(1);
2761 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2762 SDValue TrueVal = Op.getOperand(2);
2763 SDValue FalseVal = Op.getOperand(3);
2764 DebugLoc dl = Op.getDebugLoc();
2766 if (LHS.getValueType() == MVT::i32) {
2768 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2769 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2770 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp);
2773 ARMCC::CondCodes CondCode, CondCode2;
2774 FPCCToARMCC(CC, CondCode, CondCode2);
2776 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2777 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2778 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2779 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2781 if (CondCode2 != ARMCC::AL) {
2782 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2783 // FIXME: Needs another CMP because flag can have but one use.
2784 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2785 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2786 Result, TrueVal, ARMcc2, CCR, Cmp2);
2791 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2792 /// to morph to an integer compare sequence.
2793 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2794 const ARMSubtarget *Subtarget) {
2795 SDNode *N = Op.getNode();
2796 if (!N->hasOneUse())
2797 // Otherwise it requires moving the value from fp to integer registers.
2799 if (!N->getNumValues())
2801 EVT VT = Op.getValueType();
2802 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2803 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2804 // vmrs are very slow, e.g. cortex-a8.
2807 if (isFloatingPointZero(Op)) {
2811 return ISD::isNormalLoad(N);
2814 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2815 if (isFloatingPointZero(Op))
2816 return DAG.getConstant(0, MVT::i32);
2818 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2819 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2820 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2821 Ld->isVolatile(), Ld->isNonTemporal(),
2822 Ld->getAlignment());
2824 llvm_unreachable("Unknown VFP cmp argument!");
2827 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2828 SDValue &RetVal1, SDValue &RetVal2) {
2829 if (isFloatingPointZero(Op)) {
2830 RetVal1 = DAG.getConstant(0, MVT::i32);
2831 RetVal2 = DAG.getConstant(0, MVT::i32);
2835 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2836 SDValue Ptr = Ld->getBasePtr();
2837 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2838 Ld->getChain(), Ptr,
2839 Ld->getPointerInfo(),
2840 Ld->isVolatile(), Ld->isNonTemporal(),
2841 Ld->getAlignment());
2843 EVT PtrType = Ptr.getValueType();
2844 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2845 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2846 PtrType, Ptr, DAG.getConstant(4, PtrType));
2847 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2848 Ld->getChain(), NewPtr,
2849 Ld->getPointerInfo().getWithOffset(4),
2850 Ld->isVolatile(), Ld->isNonTemporal(),
2855 llvm_unreachable("Unknown VFP cmp argument!");
2858 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2859 /// f32 and even f64 comparisons to integer ones.
2861 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2862 SDValue Chain = Op.getOperand(0);
2863 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2864 SDValue LHS = Op.getOperand(2);
2865 SDValue RHS = Op.getOperand(3);
2866 SDValue Dest = Op.getOperand(4);
2867 DebugLoc dl = Op.getDebugLoc();
2869 bool SeenZero = false;
2870 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2871 canChangeToInt(RHS, SeenZero, Subtarget) &&
2872 // If one of the operand is zero, it's safe to ignore the NaN case since
2873 // we only care about equality comparisons.
2874 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2875 // If unsafe fp math optimization is enabled and there are no other uses of
2876 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2877 // to an integer comparison.
2878 if (CC == ISD::SETOEQ)
2880 else if (CC == ISD::SETUNE)
2884 if (LHS.getValueType() == MVT::f32) {
2885 LHS = bitcastf32Toi32(LHS, DAG);
2886 RHS = bitcastf32Toi32(RHS, DAG);
2887 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2888 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2889 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2890 Chain, Dest, ARMcc, CCR, Cmp);
2895 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2896 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2897 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2898 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2899 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2900 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2901 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2907 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2908 SDValue Chain = Op.getOperand(0);
2909 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2910 SDValue LHS = Op.getOperand(2);
2911 SDValue RHS = Op.getOperand(3);
2912 SDValue Dest = Op.getOperand(4);
2913 DebugLoc dl = Op.getDebugLoc();
2915 if (LHS.getValueType() == MVT::i32) {
2917 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2918 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2919 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2920 Chain, Dest, ARMcc, CCR, Cmp);
2923 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2926 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2927 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2928 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2929 if (Result.getNode())
2933 ARMCC::CondCodes CondCode, CondCode2;
2934 FPCCToARMCC(CC, CondCode, CondCode2);
2936 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2937 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2938 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2939 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2940 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2941 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2942 if (CondCode2 != ARMCC::AL) {
2943 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2944 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2945 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2950 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2951 SDValue Chain = Op.getOperand(0);
2952 SDValue Table = Op.getOperand(1);
2953 SDValue Index = Op.getOperand(2);
2954 DebugLoc dl = Op.getDebugLoc();
2956 EVT PTy = getPointerTy();
2957 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2958 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2959 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2960 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2961 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2962 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2963 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2964 if (Subtarget->isThumb2()) {
2965 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2966 // which does another jump to the destination. This also makes it easier
2967 // to translate it to TBB / TBH later.
2968 // FIXME: This might not work if the function is extremely large.
2969 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2970 Addr, Op.getOperand(2), JTI, UId);
2972 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2973 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2974 MachinePointerInfo::getJumpTable(),
2976 Chain = Addr.getValue(1);
2977 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2978 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2980 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2981 MachinePointerInfo::getJumpTable(), false, false, 0);
2982 Chain = Addr.getValue(1);
2983 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2987 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2988 DebugLoc dl = Op.getDebugLoc();
2991 switch (Op.getOpcode()) {
2993 assert(0 && "Invalid opcode!");
2994 case ISD::FP_TO_SINT:
2995 Opc = ARMISD::FTOSI;
2997 case ISD::FP_TO_UINT:
2998 Opc = ARMISD::FTOUI;
3001 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3002 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3005 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3006 EVT VT = Op.getValueType();
3007 DebugLoc dl = Op.getDebugLoc();
3009 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3010 "Invalid type for custom lowering!");
3011 if (VT != MVT::v4f32)
3012 return DAG.UnrollVectorOp(Op.getNode());
3016 switch (Op.getOpcode()) {
3018 assert(0 && "Invalid opcode!");
3019 case ISD::SINT_TO_FP:
3020 CastOpc = ISD::SIGN_EXTEND;
3021 Opc = ISD::SINT_TO_FP;
3023 case ISD::UINT_TO_FP:
3024 CastOpc = ISD::ZERO_EXTEND;
3025 Opc = ISD::UINT_TO_FP;
3029 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3030 return DAG.getNode(Opc, dl, VT, Op);
3033 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3034 EVT VT = Op.getValueType();
3036 return LowerVectorINT_TO_FP(Op, DAG);
3038 DebugLoc dl = Op.getDebugLoc();
3041 switch (Op.getOpcode()) {
3043 assert(0 && "Invalid opcode!");
3044 case ISD::SINT_TO_FP:
3045 Opc = ARMISD::SITOF;
3047 case ISD::UINT_TO_FP:
3048 Opc = ARMISD::UITOF;
3052 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3053 return DAG.getNode(Opc, dl, VT, Op);
3056 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3057 // Implement fcopysign with a fabs and a conditional fneg.
3058 SDValue Tmp0 = Op.getOperand(0);
3059 SDValue Tmp1 = Op.getOperand(1);
3060 DebugLoc dl = Op.getDebugLoc();
3061 EVT VT = Op.getValueType();
3062 EVT SrcVT = Tmp1.getValueType();
3063 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3064 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3065 bool UseNEON = !InGPR && Subtarget->hasNEON();
3068 // Use VBSL to copy the sign bit.
3069 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3070 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3071 DAG.getTargetConstant(EncodedVal, MVT::i32));
3072 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3074 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3075 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3076 DAG.getConstant(32, MVT::i32));
3077 else /*if (VT == MVT::f32)*/
3078 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3079 if (SrcVT == MVT::f32) {
3080 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3082 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3083 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3084 DAG.getConstant(32, MVT::i32));
3085 } else if (VT == MVT::f32)
3086 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3087 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3088 DAG.getConstant(32, MVT::i32));
3089 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3090 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3092 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3094 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3095 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3096 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3098 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3099 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3100 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3101 if (VT == MVT::f32) {
3102 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3103 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3104 DAG.getConstant(0, MVT::i32));
3106 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3112 // Bitcast operand 1 to i32.
3113 if (SrcVT == MVT::f64)
3114 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3115 &Tmp1, 1).getValue(1);
3116 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3118 // Or in the signbit with integer operations.
3119 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3120 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3121 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3122 if (VT == MVT::f32) {
3123 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3124 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3125 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3126 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3129 // f64: Or the high part with signbit and then combine two parts.
3130 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3132 SDValue Lo = Tmp0.getValue(0);
3133 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3134 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3135 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3138 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3139 MachineFunction &MF = DAG.getMachineFunction();
3140 MachineFrameInfo *MFI = MF.getFrameInfo();
3141 MFI->setReturnAddressIsTaken(true);
3143 EVT VT = Op.getValueType();
3144 DebugLoc dl = Op.getDebugLoc();
3145 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3147 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3148 SDValue Offset = DAG.getConstant(4, MVT::i32);
3149 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3150 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3151 MachinePointerInfo(), false, false, 0);
3154 // Return LR, which contains the return address. Mark it an implicit live-in.
3155 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3156 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3159 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3160 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3161 MFI->setFrameAddressIsTaken(true);
3163 EVT VT = Op.getValueType();
3164 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3165 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3166 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3167 ? ARM::R7 : ARM::R11;
3168 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3170 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3171 MachinePointerInfo(),
3176 /// ExpandBITCAST - If the target supports VFP, this function is called to
3177 /// expand a bit convert where either the source or destination type is i64 to
3178 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3179 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3180 /// vectors), since the legalizer won't know what to do with that.
3181 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3182 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3183 DebugLoc dl = N->getDebugLoc();
3184 SDValue Op = N->getOperand(0);
3186 // This function is only supposed to be called for i64 types, either as the
3187 // source or destination of the bit convert.
3188 EVT SrcVT = Op.getValueType();
3189 EVT DstVT = N->getValueType(0);
3190 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3191 "ExpandBITCAST called for non-i64 type");
3193 // Turn i64->f64 into VMOVDRR.
3194 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3195 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3196 DAG.getConstant(0, MVT::i32));
3197 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3198 DAG.getConstant(1, MVT::i32));
3199 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3200 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3203 // Turn f64->i64 into VMOVRRD.
3204 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3205 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3206 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3207 // Merge the pieces into a single i64 value.
3208 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3214 /// getZeroVector - Returns a vector of specified type with all zero elements.
3215 /// Zero vectors are used to represent vector negation and in those cases
3216 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3217 /// not support i64 elements, so sometimes the zero vectors will need to be
3218 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3220 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3221 assert(VT.isVector() && "Expected a vector type");
3222 // The canonical modified immediate encoding of a zero vector is....0!
3223 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3224 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3225 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3226 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3229 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3230 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3231 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3232 SelectionDAG &DAG) const {
3233 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3234 EVT VT = Op.getValueType();
3235 unsigned VTBits = VT.getSizeInBits();
3236 DebugLoc dl = Op.getDebugLoc();
3237 SDValue ShOpLo = Op.getOperand(0);
3238 SDValue ShOpHi = Op.getOperand(1);
3239 SDValue ShAmt = Op.getOperand(2);
3241 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3243 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3245 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3246 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3247 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3248 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3249 DAG.getConstant(VTBits, MVT::i32));
3250 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3251 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3252 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3254 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3255 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3257 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3258 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3261 SDValue Ops[2] = { Lo, Hi };
3262 return DAG.getMergeValues(Ops, 2, dl);
3265 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3266 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3267 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3268 SelectionDAG &DAG) const {
3269 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3270 EVT VT = Op.getValueType();
3271 unsigned VTBits = VT.getSizeInBits();
3272 DebugLoc dl = Op.getDebugLoc();
3273 SDValue ShOpLo = Op.getOperand(0);
3274 SDValue ShOpHi = Op.getOperand(1);
3275 SDValue ShAmt = Op.getOperand(2);
3278 assert(Op.getOpcode() == ISD::SHL_PARTS);
3279 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3280 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3281 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3282 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3283 DAG.getConstant(VTBits, MVT::i32));
3284 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3285 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3287 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3288 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3289 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3291 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3292 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3295 SDValue Ops[2] = { Lo, Hi };
3296 return DAG.getMergeValues(Ops, 2, dl);
3299 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3300 SelectionDAG &DAG) const {
3301 // The rounding mode is in bits 23:22 of the FPSCR.
3302 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3303 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3304 // so that the shift + and get folded into a bitfield extract.
3305 DebugLoc dl = Op.getDebugLoc();
3306 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3307 DAG.getConstant(Intrinsic::arm_get_fpscr,
3309 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3310 DAG.getConstant(1U << 22, MVT::i32));
3311 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3312 DAG.getConstant(22, MVT::i32));
3313 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3314 DAG.getConstant(3, MVT::i32));
3317 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3318 const ARMSubtarget *ST) {
3319 EVT VT = N->getValueType(0);
3320 DebugLoc dl = N->getDebugLoc();
3322 if (!ST->hasV6T2Ops())
3325 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3326 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3329 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3330 const ARMSubtarget *ST) {
3331 EVT VT = N->getValueType(0);
3332 DebugLoc dl = N->getDebugLoc();
3337 // Lower vector shifts on NEON to use VSHL.
3338 assert(ST->hasNEON() && "unexpected vector shift");
3340 // Left shifts translate directly to the vshiftu intrinsic.
3341 if (N->getOpcode() == ISD::SHL)
3342 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3343 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3344 N->getOperand(0), N->getOperand(1));
3346 assert((N->getOpcode() == ISD::SRA ||
3347 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3349 // NEON uses the same intrinsics for both left and right shifts. For
3350 // right shifts, the shift amounts are negative, so negate the vector of
3352 EVT ShiftVT = N->getOperand(1).getValueType();
3353 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3354 getZeroVector(ShiftVT, DAG, dl),
3356 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3357 Intrinsic::arm_neon_vshifts :
3358 Intrinsic::arm_neon_vshiftu);
3359 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3360 DAG.getConstant(vshiftInt, MVT::i32),
3361 N->getOperand(0), NegatedCount);
3364 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3365 const ARMSubtarget *ST) {
3366 EVT VT = N->getValueType(0);
3367 DebugLoc dl = N->getDebugLoc();
3369 // We can get here for a node like i32 = ISD::SHL i32, i64
3373 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3374 "Unknown shift to lower!");
3376 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3377 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3378 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3381 // If we are in thumb mode, we don't have RRX.
3382 if (ST->isThumb1Only()) return SDValue();
3384 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3385 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3386 DAG.getConstant(0, MVT::i32));
3387 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3388 DAG.getConstant(1, MVT::i32));
3390 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3391 // captures the result into a carry flag.
3392 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3393 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3395 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3396 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3398 // Merge the pieces into a single i64 value.
3399 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3402 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3403 SDValue TmpOp0, TmpOp1;
3404 bool Invert = false;
3408 SDValue Op0 = Op.getOperand(0);
3409 SDValue Op1 = Op.getOperand(1);
3410 SDValue CC = Op.getOperand(2);
3411 EVT VT = Op.getValueType();
3412 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3413 DebugLoc dl = Op.getDebugLoc();
3415 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3416 switch (SetCCOpcode) {
3417 default: llvm_unreachable("Illegal FP comparison"); break;
3419 case ISD::SETNE: Invert = true; // Fallthrough
3421 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3423 case ISD::SETLT: Swap = true; // Fallthrough
3425 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3427 case ISD::SETLE: Swap = true; // Fallthrough
3429 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3430 case ISD::SETUGE: Swap = true; // Fallthrough
3431 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3432 case ISD::SETUGT: Swap = true; // Fallthrough
3433 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3434 case ISD::SETUEQ: Invert = true; // Fallthrough
3436 // Expand this to (OLT | OGT).
3440 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3441 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3443 case ISD::SETUO: Invert = true; // Fallthrough
3445 // Expand this to (OLT | OGE).
3449 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3450 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3454 // Integer comparisons.
3455 switch (SetCCOpcode) {
3456 default: llvm_unreachable("Illegal integer comparison"); break;
3457 case ISD::SETNE: Invert = true;
3458 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3459 case ISD::SETLT: Swap = true;
3460 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3461 case ISD::SETLE: Swap = true;
3462 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3463 case ISD::SETULT: Swap = true;
3464 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3465 case ISD::SETULE: Swap = true;
3466 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3469 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3470 if (Opc == ARMISD::VCEQ) {
3473 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3475 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3478 // Ignore bitconvert.
3479 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3480 AndOp = AndOp.getOperand(0);
3482 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3484 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3485 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3492 std::swap(Op0, Op1);
3494 // If one of the operands is a constant vector zero, attempt to fold the
3495 // comparison to a specialized compare-against-zero form.
3497 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3499 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3500 if (Opc == ARMISD::VCGE)
3501 Opc = ARMISD::VCLEZ;
3502 else if (Opc == ARMISD::VCGT)
3503 Opc = ARMISD::VCLTZ;
3508 if (SingleOp.getNode()) {
3511 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3513 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3515 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3517 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3519 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3521 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3524 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3528 Result = DAG.getNOT(dl, Result, VT);
3533 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3534 /// valid vector constant for a NEON instruction with a "modified immediate"
3535 /// operand (e.g., VMOV). If so, return the encoded value.
3536 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3537 unsigned SplatBitSize, SelectionDAG &DAG,
3538 EVT &VT, bool is128Bits, NEONModImmType type) {
3539 unsigned OpCmode, Imm;
3541 // SplatBitSize is set to the smallest size that splats the vector, so a
3542 // zero vector will always have SplatBitSize == 8. However, NEON modified
3543 // immediate instructions others than VMOV do not support the 8-bit encoding
3544 // of a zero vector, and the default encoding of zero is supposed to be the
3549 switch (SplatBitSize) {
3551 if (type != VMOVModImm)
3553 // Any 1-byte value is OK. Op=0, Cmode=1110.
3554 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3557 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3561 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3562 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3563 if ((SplatBits & ~0xff) == 0) {
3564 // Value = 0x00nn: Op=x, Cmode=100x.
3569 if ((SplatBits & ~0xff00) == 0) {
3570 // Value = 0xnn00: Op=x, Cmode=101x.
3572 Imm = SplatBits >> 8;
3578 // NEON's 32-bit VMOV supports splat values where:
3579 // * only one byte is nonzero, or
3580 // * the least significant byte is 0xff and the second byte is nonzero, or
3581 // * the least significant 2 bytes are 0xff and the third is nonzero.
3582 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3583 if ((SplatBits & ~0xff) == 0) {
3584 // Value = 0x000000nn: Op=x, Cmode=000x.
3589 if ((SplatBits & ~0xff00) == 0) {
3590 // Value = 0x0000nn00: Op=x, Cmode=001x.
3592 Imm = SplatBits >> 8;
3595 if ((SplatBits & ~0xff0000) == 0) {
3596 // Value = 0x00nn0000: Op=x, Cmode=010x.
3598 Imm = SplatBits >> 16;
3601 if ((SplatBits & ~0xff000000) == 0) {
3602 // Value = 0xnn000000: Op=x, Cmode=011x.
3604 Imm = SplatBits >> 24;
3608 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3609 if (type == OtherModImm) return SDValue();
3611 if ((SplatBits & ~0xffff) == 0 &&
3612 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3613 // Value = 0x0000nnff: Op=x, Cmode=1100.
3615 Imm = SplatBits >> 8;
3620 if ((SplatBits & ~0xffffff) == 0 &&
3621 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3622 // Value = 0x00nnffff: Op=x, Cmode=1101.
3624 Imm = SplatBits >> 16;
3625 SplatBits |= 0xffff;
3629 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3630 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3631 // VMOV.I32. A (very) minor optimization would be to replicate the value
3632 // and fall through here to test for a valid 64-bit splat. But, then the
3633 // caller would also need to check and handle the change in size.
3637 if (type != VMOVModImm)
3639 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3640 uint64_t BitMask = 0xff;
3642 unsigned ImmMask = 1;
3644 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3645 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3648 } else if ((SplatBits & BitMask) != 0) {
3654 // Op=1, Cmode=1110.
3657 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3662 llvm_unreachable("unexpected size for isNEONModifiedImm");
3666 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3667 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3670 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3671 bool &ReverseVEXT, unsigned &Imm) {
3672 unsigned NumElts = VT.getVectorNumElements();
3673 ReverseVEXT = false;
3675 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3681 // If this is a VEXT shuffle, the immediate value is the index of the first
3682 // element. The other shuffle indices must be the successive elements after
3684 unsigned ExpectedElt = Imm;
3685 for (unsigned i = 1; i < NumElts; ++i) {
3686 // Increment the expected index. If it wraps around, it may still be
3687 // a VEXT but the source vectors must be swapped.
3689 if (ExpectedElt == NumElts * 2) {
3694 if (M[i] < 0) continue; // ignore UNDEF indices
3695 if (ExpectedElt != static_cast<unsigned>(M[i]))
3699 // Adjust the index value if the source operands will be swapped.
3706 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3707 /// instruction with the specified blocksize. (The order of the elements
3708 /// within each block of the vector is reversed.)
3709 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3710 unsigned BlockSize) {
3711 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3712 "Only possible block sizes for VREV are: 16, 32, 64");
3714 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3718 unsigned NumElts = VT.getVectorNumElements();
3719 unsigned BlockElts = M[0] + 1;
3720 // If the first shuffle index is UNDEF, be optimistic.
3722 BlockElts = BlockSize / EltSz;
3724 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3727 for (unsigned i = 0; i < NumElts; ++i) {
3728 if (M[i] < 0) continue; // ignore UNDEF indices
3729 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3736 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3737 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3738 // range, then 0 is placed into the resulting vector. So pretty much any mask
3739 // of 8 elements can work here.
3740 return VT == MVT::v8i8 && M.size() == 8;
3743 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3744 unsigned &WhichResult) {
3745 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3749 unsigned NumElts = VT.getVectorNumElements();
3750 WhichResult = (M[0] == 0 ? 0 : 1);
3751 for (unsigned i = 0; i < NumElts; i += 2) {
3752 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3753 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3759 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3760 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3761 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3762 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3763 unsigned &WhichResult) {
3764 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3768 unsigned NumElts = VT.getVectorNumElements();
3769 WhichResult = (M[0] == 0 ? 0 : 1);
3770 for (unsigned i = 0; i < NumElts; i += 2) {
3771 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3772 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3778 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3779 unsigned &WhichResult) {
3780 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3784 unsigned NumElts = VT.getVectorNumElements();
3785 WhichResult = (M[0] == 0 ? 0 : 1);
3786 for (unsigned i = 0; i != NumElts; ++i) {
3787 if (M[i] < 0) continue; // ignore UNDEF indices
3788 if ((unsigned) M[i] != 2 * i + WhichResult)
3792 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3793 if (VT.is64BitVector() && EltSz == 32)
3799 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3800 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3801 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3802 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3803 unsigned &WhichResult) {
3804 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3808 unsigned Half = VT.getVectorNumElements() / 2;
3809 WhichResult = (M[0] == 0 ? 0 : 1);
3810 for (unsigned j = 0; j != 2; ++j) {
3811 unsigned Idx = WhichResult;
3812 for (unsigned i = 0; i != Half; ++i) {
3813 int MIdx = M[i + j * Half];
3814 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3820 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3821 if (VT.is64BitVector() && EltSz == 32)
3827 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3828 unsigned &WhichResult) {
3829 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3833 unsigned NumElts = VT.getVectorNumElements();
3834 WhichResult = (M[0] == 0 ? 0 : 1);
3835 unsigned Idx = WhichResult * NumElts / 2;
3836 for (unsigned i = 0; i != NumElts; i += 2) {
3837 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3838 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3843 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3844 if (VT.is64BitVector() && EltSz == 32)
3850 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3851 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3852 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3853 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3854 unsigned &WhichResult) {
3855 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3859 unsigned NumElts = VT.getVectorNumElements();
3860 WhichResult = (M[0] == 0 ? 0 : 1);
3861 unsigned Idx = WhichResult * NumElts / 2;
3862 for (unsigned i = 0; i != NumElts; i += 2) {
3863 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3864 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3869 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3870 if (VT.is64BitVector() && EltSz == 32)
3876 // If N is an integer constant that can be moved into a register in one
3877 // instruction, return an SDValue of such a constant (will become a MOV
3878 // instruction). Otherwise return null.
3879 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3880 const ARMSubtarget *ST, DebugLoc dl) {
3882 if (!isa<ConstantSDNode>(N))
3884 Val = cast<ConstantSDNode>(N)->getZExtValue();
3886 if (ST->isThumb1Only()) {
3887 if (Val <= 255 || ~Val <= 255)
3888 return DAG.getConstant(Val, MVT::i32);
3890 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3891 return DAG.getConstant(Val, MVT::i32);
3896 // If this is a case we can't handle, return null and let the default
3897 // expansion code take care of it.
3898 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3899 const ARMSubtarget *ST) const {
3900 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3901 DebugLoc dl = Op.getDebugLoc();
3902 EVT VT = Op.getValueType();
3904 APInt SplatBits, SplatUndef;
3905 unsigned SplatBitSize;
3907 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3908 if (SplatBitSize <= 64) {
3909 // Check if an immediate VMOV works.
3911 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3912 SplatUndef.getZExtValue(), SplatBitSize,
3913 DAG, VmovVT, VT.is128BitVector(),
3915 if (Val.getNode()) {
3916 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3917 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3920 // Try an immediate VMVN.
3921 uint64_t NegatedImm = (SplatBits.getZExtValue() ^
3922 ((1LL << SplatBitSize) - 1));
3923 Val = isNEONModifiedImm(NegatedImm,
3924 SplatUndef.getZExtValue(), SplatBitSize,
3925 DAG, VmovVT, VT.is128BitVector(),
3927 if (Val.getNode()) {
3928 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3929 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3934 // Scan through the operands to see if only one value is used.
3935 unsigned NumElts = VT.getVectorNumElements();
3936 bool isOnlyLowElement = true;
3937 bool usesOnlyOneValue = true;
3938 bool isConstant = true;
3940 for (unsigned i = 0; i < NumElts; ++i) {
3941 SDValue V = Op.getOperand(i);
3942 if (V.getOpcode() == ISD::UNDEF)
3945 isOnlyLowElement = false;
3946 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3949 if (!Value.getNode())
3951 else if (V != Value)
3952 usesOnlyOneValue = false;
3955 if (!Value.getNode())
3956 return DAG.getUNDEF(VT);
3958 if (isOnlyLowElement)
3959 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3961 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3963 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3964 // i32 and try again.
3965 if (usesOnlyOneValue && EltSize <= 32) {
3967 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3968 if (VT.getVectorElementType().isFloatingPoint()) {
3969 SmallVector<SDValue, 8> Ops;
3970 for (unsigned i = 0; i < NumElts; ++i)
3971 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3973 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
3974 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
3975 Val = LowerBUILD_VECTOR(Val, DAG, ST);
3977 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
3979 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
3981 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
3984 // If all elements are constants and the case above didn't get hit, fall back
3985 // to the default expansion, which will generate a load from the constant
3990 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
3992 SDValue shuffle = ReconstructShuffle(Op, DAG);
3993 if (shuffle != SDValue())
3997 // Vectors with 32- or 64-bit elements can be built by directly assigning
3998 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3999 // will be legalized.
4000 if (EltSize >= 32) {
4001 // Do the expansion with floating-point types, since that is what the VFP
4002 // registers are defined to use, and since i64 is not legal.
4003 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4004 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4005 SmallVector<SDValue, 8> Ops;
4006 for (unsigned i = 0; i < NumElts; ++i)
4007 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4008 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4009 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4015 // Gather data to see if the operation can be modelled as a
4016 // shuffle in combination with VEXTs.
4017 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4018 SelectionDAG &DAG) const {
4019 DebugLoc dl = Op.getDebugLoc();
4020 EVT VT = Op.getValueType();
4021 unsigned NumElts = VT.getVectorNumElements();
4023 SmallVector<SDValue, 2> SourceVecs;
4024 SmallVector<unsigned, 2> MinElts;
4025 SmallVector<unsigned, 2> MaxElts;
4027 for (unsigned i = 0; i < NumElts; ++i) {
4028 SDValue V = Op.getOperand(i);
4029 if (V.getOpcode() == ISD::UNDEF)
4031 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4032 // A shuffle can only come from building a vector from various
4033 // elements of other vectors.
4037 // Record this extraction against the appropriate vector if possible...
4038 SDValue SourceVec = V.getOperand(0);
4039 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4040 bool FoundSource = false;
4041 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4042 if (SourceVecs[j] == SourceVec) {
4043 if (MinElts[j] > EltNo)
4045 if (MaxElts[j] < EltNo)
4052 // Or record a new source if not...
4054 SourceVecs.push_back(SourceVec);
4055 MinElts.push_back(EltNo);
4056 MaxElts.push_back(EltNo);
4060 // Currently only do something sane when at most two source vectors
4062 if (SourceVecs.size() > 2)
4065 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4066 int VEXTOffsets[2] = {0, 0};
4068 // This loop extracts the usage patterns of the source vectors
4069 // and prepares appropriate SDValues for a shuffle if possible.
4070 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4071 if (SourceVecs[i].getValueType() == VT) {
4072 // No VEXT necessary
4073 ShuffleSrcs[i] = SourceVecs[i];
4076 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4077 // It probably isn't worth padding out a smaller vector just to
4078 // break it down again in a shuffle.
4082 // Since only 64-bit and 128-bit vectors are legal on ARM and
4083 // we've eliminated the other cases...
4084 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4085 "unexpected vector sizes in ReconstructShuffle");
4087 if (MaxElts[i] - MinElts[i] >= NumElts) {
4088 // Span too large for a VEXT to cope
4092 if (MinElts[i] >= NumElts) {
4093 // The extraction can just take the second half
4094 VEXTOffsets[i] = NumElts;
4095 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4097 DAG.getIntPtrConstant(NumElts));
4098 } else if (MaxElts[i] < NumElts) {
4099 // The extraction can just take the first half
4101 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4103 DAG.getIntPtrConstant(0));
4105 // An actual VEXT is needed
4106 VEXTOffsets[i] = MinElts[i];
4107 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4109 DAG.getIntPtrConstant(0));
4110 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4112 DAG.getIntPtrConstant(NumElts));
4113 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4114 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4118 SmallVector<int, 8> Mask;
4120 for (unsigned i = 0; i < NumElts; ++i) {
4121 SDValue Entry = Op.getOperand(i);
4122 if (Entry.getOpcode() == ISD::UNDEF) {
4127 SDValue ExtractVec = Entry.getOperand(0);
4128 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4129 .getOperand(1))->getSExtValue();
4130 if (ExtractVec == SourceVecs[0]) {
4131 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4133 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4137 // Final check before we try to produce nonsense...
4138 if (isShuffleMaskLegal(Mask, VT))
4139 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4145 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4146 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4147 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4148 /// are assumed to be legal.
4150 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4152 if (VT.getVectorNumElements() == 4 &&
4153 (VT.is128BitVector() || VT.is64BitVector())) {
4154 unsigned PFIndexes[4];
4155 for (unsigned i = 0; i != 4; ++i) {
4159 PFIndexes[i] = M[i];
4162 // Compute the index in the perfect shuffle table.
4163 unsigned PFTableIndex =
4164 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4165 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4166 unsigned Cost = (PFEntry >> 30);
4173 unsigned Imm, WhichResult;
4175 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4176 return (EltSize >= 32 ||
4177 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4178 isVREVMask(M, VT, 64) ||
4179 isVREVMask(M, VT, 32) ||
4180 isVREVMask(M, VT, 16) ||
4181 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4182 isVTBLMask(M, VT) ||
4183 isVTRNMask(M, VT, WhichResult) ||
4184 isVUZPMask(M, VT, WhichResult) ||
4185 isVZIPMask(M, VT, WhichResult) ||
4186 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4187 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4188 isVZIP_v_undef_Mask(M, VT, WhichResult));
4191 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4192 /// the specified operations to build the shuffle.
4193 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4194 SDValue RHS, SelectionDAG &DAG,
4196 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4197 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4198 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4201 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4210 OP_VUZPL, // VUZP, left result
4211 OP_VUZPR, // VUZP, right result
4212 OP_VZIPL, // VZIP, left result
4213 OP_VZIPR, // VZIP, right result
4214 OP_VTRNL, // VTRN, left result
4215 OP_VTRNR // VTRN, right result
4218 if (OpNum == OP_COPY) {
4219 if (LHSID == (1*9+2)*9+3) return LHS;
4220 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4224 SDValue OpLHS, OpRHS;
4225 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4226 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4227 EVT VT = OpLHS.getValueType();
4230 default: llvm_unreachable("Unknown shuffle opcode!");
4232 // VREV divides the vector in half and swaps within the half.
4233 if (VT.getVectorElementType() == MVT::i32 ||
4234 VT.getVectorElementType() == MVT::f32)
4235 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4236 // vrev <4 x i16> -> VREV32
4237 if (VT.getVectorElementType() == MVT::i16)
4238 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4239 // vrev <4 x i8> -> VREV16
4240 assert(VT.getVectorElementType() == MVT::i8);
4241 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4246 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4247 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4251 return DAG.getNode(ARMISD::VEXT, dl, VT,
4253 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4256 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4257 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4260 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4261 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4264 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4265 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4269 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4270 SmallVectorImpl<int> &ShuffleMask,
4271 SelectionDAG &DAG) {
4272 // Check to see if we can use the VTBL instruction.
4273 SDValue V1 = Op.getOperand(0);
4274 SDValue V2 = Op.getOperand(1);
4275 DebugLoc DL = Op.getDebugLoc();
4277 SmallVector<SDValue, 8> VTBLMask;
4278 for (SmallVectorImpl<int>::iterator
4279 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4280 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4282 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4283 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4284 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4287 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4288 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4292 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4293 SDValue V1 = Op.getOperand(0);
4294 SDValue V2 = Op.getOperand(1);
4295 DebugLoc dl = Op.getDebugLoc();
4296 EVT VT = Op.getValueType();
4297 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4298 SmallVector<int, 8> ShuffleMask;
4300 // Convert shuffles that are directly supported on NEON to target-specific
4301 // DAG nodes, instead of keeping them as shuffles and matching them again
4302 // during code selection. This is more efficient and avoids the possibility
4303 // of inconsistencies between legalization and selection.
4304 // FIXME: floating-point vectors should be canonicalized to integer vectors
4305 // of the same time so that they get CSEd properly.
4306 SVN->getMask(ShuffleMask);
4308 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4309 if (EltSize <= 32) {
4310 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4311 int Lane = SVN->getSplatIndex();
4312 // If this is undef splat, generate it via "just" vdup, if possible.
4313 if (Lane == -1) Lane = 0;
4315 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4316 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4318 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4319 DAG.getConstant(Lane, MVT::i32));
4324 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4327 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4328 DAG.getConstant(Imm, MVT::i32));
4331 if (isVREVMask(ShuffleMask, VT, 64))
4332 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4333 if (isVREVMask(ShuffleMask, VT, 32))
4334 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4335 if (isVREVMask(ShuffleMask, VT, 16))
4336 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4338 // Check for Neon shuffles that modify both input vectors in place.
4339 // If both results are used, i.e., if there are two shuffles with the same
4340 // source operands and with masks corresponding to both results of one of
4341 // these operations, DAG memoization will ensure that a single node is
4342 // used for both shuffles.
4343 unsigned WhichResult;
4344 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4345 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4346 V1, V2).getValue(WhichResult);
4347 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4348 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4349 V1, V2).getValue(WhichResult);
4350 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4351 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4352 V1, V2).getValue(WhichResult);
4354 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4355 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4356 V1, V1).getValue(WhichResult);
4357 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4358 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4359 V1, V1).getValue(WhichResult);
4360 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4361 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4362 V1, V1).getValue(WhichResult);
4365 // If the shuffle is not directly supported and it has 4 elements, use
4366 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4367 unsigned NumElts = VT.getVectorNumElements();
4369 unsigned PFIndexes[4];
4370 for (unsigned i = 0; i != 4; ++i) {
4371 if (ShuffleMask[i] < 0)
4374 PFIndexes[i] = ShuffleMask[i];
4377 // Compute the index in the perfect shuffle table.
4378 unsigned PFTableIndex =
4379 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4380 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4381 unsigned Cost = (PFEntry >> 30);
4384 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4387 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4388 if (EltSize >= 32) {
4389 // Do the expansion with floating-point types, since that is what the VFP
4390 // registers are defined to use, and since i64 is not legal.
4391 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4392 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4393 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4394 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4395 SmallVector<SDValue, 8> Ops;
4396 for (unsigned i = 0; i < NumElts; ++i) {
4397 if (ShuffleMask[i] < 0)
4398 Ops.push_back(DAG.getUNDEF(EltVT));
4400 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4401 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4402 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4405 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4406 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4409 if (VT == MVT::v8i8) {
4410 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4411 if (NewOp.getNode())
4418 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4419 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4420 SDValue Lane = Op.getOperand(1);
4421 if (!isa<ConstantSDNode>(Lane))
4424 SDValue Vec = Op.getOperand(0);
4425 if (Op.getValueType() == MVT::i32 &&
4426 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4427 DebugLoc dl = Op.getDebugLoc();
4428 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4434 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4435 // The only time a CONCAT_VECTORS operation can have legal types is when
4436 // two 64-bit vectors are concatenated to a 128-bit vector.
4437 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4438 "unexpected CONCAT_VECTORS");
4439 DebugLoc dl = Op.getDebugLoc();
4440 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4441 SDValue Op0 = Op.getOperand(0);
4442 SDValue Op1 = Op.getOperand(1);
4443 if (Op0.getOpcode() != ISD::UNDEF)
4444 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4445 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4446 DAG.getIntPtrConstant(0));
4447 if (Op1.getOpcode() != ISD::UNDEF)
4448 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4449 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4450 DAG.getIntPtrConstant(1));
4451 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4454 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4455 /// element has been zero/sign-extended, depending on the isSigned parameter,
4456 /// from an integer type half its size.
4457 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4459 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4460 EVT VT = N->getValueType(0);
4461 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4462 SDNode *BVN = N->getOperand(0).getNode();
4463 if (BVN->getValueType(0) != MVT::v4i32 ||
4464 BVN->getOpcode() != ISD::BUILD_VECTOR)
4466 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4467 unsigned HiElt = 1 - LoElt;
4468 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4469 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4470 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4471 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4472 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4475 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4476 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4479 if (Hi0->isNullValue() && Hi1->isNullValue())
4485 if (N->getOpcode() != ISD::BUILD_VECTOR)
4488 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4489 SDNode *Elt = N->getOperand(i).getNode();
4490 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4491 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4492 unsigned HalfSize = EltSize / 2;
4494 int64_t SExtVal = C->getSExtValue();
4495 if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
4498 if ((C->getZExtValue() >> HalfSize) != 0)
4509 /// isSignExtended - Check if a node is a vector value that is sign-extended
4510 /// or a constant BUILD_VECTOR with sign-extended elements.
4511 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4512 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4514 if (isExtendedBUILD_VECTOR(N, DAG, true))
4519 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4520 /// or a constant BUILD_VECTOR with zero-extended elements.
4521 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4522 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4524 if (isExtendedBUILD_VECTOR(N, DAG, false))
4529 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4530 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4531 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4532 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4533 return N->getOperand(0);
4534 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4535 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4536 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4537 LD->isNonTemporal(), LD->getAlignment());
4538 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4539 // have been legalized as a BITCAST from v4i32.
4540 if (N->getOpcode() == ISD::BITCAST) {
4541 SDNode *BVN = N->getOperand(0).getNode();
4542 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4543 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4544 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4545 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4546 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4548 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4549 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4550 EVT VT = N->getValueType(0);
4551 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4552 unsigned NumElts = VT.getVectorNumElements();
4553 MVT TruncVT = MVT::getIntegerVT(EltSize);
4554 SmallVector<SDValue, 8> Ops;
4555 for (unsigned i = 0; i != NumElts; ++i) {
4556 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4557 const APInt &CInt = C->getAPIntValue();
4558 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4560 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4561 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4564 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4565 unsigned Opcode = N->getOpcode();
4566 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4567 SDNode *N0 = N->getOperand(0).getNode();
4568 SDNode *N1 = N->getOperand(1).getNode();
4569 return N0->hasOneUse() && N1->hasOneUse() &&
4570 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4575 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4576 unsigned Opcode = N->getOpcode();
4577 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4578 SDNode *N0 = N->getOperand(0).getNode();
4579 SDNode *N1 = N->getOperand(1).getNode();
4580 return N0->hasOneUse() && N1->hasOneUse() &&
4581 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4586 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4587 // Multiplications are only custom-lowered for 128-bit vectors so that
4588 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4589 EVT VT = Op.getValueType();
4590 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4591 SDNode *N0 = Op.getOperand(0).getNode();
4592 SDNode *N1 = Op.getOperand(1).getNode();
4593 unsigned NewOpc = 0;
4595 bool isN0SExt = isSignExtended(N0, DAG);
4596 bool isN1SExt = isSignExtended(N1, DAG);
4597 if (isN0SExt && isN1SExt)
4598 NewOpc = ARMISD::VMULLs;
4600 bool isN0ZExt = isZeroExtended(N0, DAG);
4601 bool isN1ZExt = isZeroExtended(N1, DAG);
4602 if (isN0ZExt && isN1ZExt)
4603 NewOpc = ARMISD::VMULLu;
4604 else if (isN1SExt || isN1ZExt) {
4605 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4606 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4607 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4608 NewOpc = ARMISD::VMULLs;
4610 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4611 NewOpc = ARMISD::VMULLu;
4613 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4615 NewOpc = ARMISD::VMULLu;
4621 if (VT == MVT::v2i64)
4622 // Fall through to expand this. It is not legal.
4625 // Other vector multiplications are legal.
4630 // Legalize to a VMULL instruction.
4631 DebugLoc DL = Op.getDebugLoc();
4633 SDValue Op1 = SkipExtension(N1, DAG);
4635 Op0 = SkipExtension(N0, DAG);
4636 assert(Op0.getValueType().is64BitVector() &&
4637 Op1.getValueType().is64BitVector() &&
4638 "unexpected types for extended operands to VMULL");
4639 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4642 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4643 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4650 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4651 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4652 EVT Op1VT = Op1.getValueType();
4653 return DAG.getNode(N0->getOpcode(), DL, VT,
4654 DAG.getNode(NewOpc, DL, VT,
4655 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4656 DAG.getNode(NewOpc, DL, VT,
4657 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4661 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4663 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4664 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4665 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4666 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4667 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4668 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4669 // Get reciprocal estimate.
4670 // float4 recip = vrecpeq_f32(yf);
4671 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4672 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4673 // Because char has a smaller range than uchar, we can actually get away
4674 // without any newton steps. This requires that we use a weird bias
4675 // of 0xb000, however (again, this has been exhaustively tested).
4676 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4677 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4678 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4679 Y = DAG.getConstant(0xb000, MVT::i32);
4680 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4681 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4682 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4683 // Convert back to short.
4684 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4685 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4690 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4692 // Convert to float.
4693 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4694 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4695 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4696 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4697 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4698 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4700 // Use reciprocal estimate and one refinement step.
4701 // float4 recip = vrecpeq_f32(yf);
4702 // recip *= vrecpsq_f32(yf, recip);
4703 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4704 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4705 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4706 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4708 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4709 // Because short has a smaller range than ushort, we can actually get away
4710 // with only a single newton step. This requires that we use a weird bias
4711 // of 89, however (again, this has been exhaustively tested).
4712 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4713 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4714 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4715 N1 = DAG.getConstant(0x89, MVT::i32);
4716 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4717 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4718 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4719 // Convert back to integer and return.
4720 // return vmovn_s32(vcvt_s32_f32(result));
4721 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4722 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4726 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4727 EVT VT = Op.getValueType();
4728 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4729 "unexpected type for custom-lowering ISD::SDIV");
4731 DebugLoc dl = Op.getDebugLoc();
4732 SDValue N0 = Op.getOperand(0);
4733 SDValue N1 = Op.getOperand(1);
4736 if (VT == MVT::v8i8) {
4737 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4738 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4740 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4741 DAG.getIntPtrConstant(4));
4742 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4743 DAG.getIntPtrConstant(4));
4744 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4745 DAG.getIntPtrConstant(0));
4746 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4747 DAG.getIntPtrConstant(0));
4749 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4750 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4752 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4753 N0 = LowerCONCAT_VECTORS(N0, DAG);
4755 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4758 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4761 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4762 EVT VT = Op.getValueType();
4763 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4764 "unexpected type for custom-lowering ISD::UDIV");
4766 DebugLoc dl = Op.getDebugLoc();
4767 SDValue N0 = Op.getOperand(0);
4768 SDValue N1 = Op.getOperand(1);
4771 if (VT == MVT::v8i8) {
4772 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4773 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4775 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4776 DAG.getIntPtrConstant(4));
4777 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4778 DAG.getIntPtrConstant(4));
4779 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4780 DAG.getIntPtrConstant(0));
4781 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4782 DAG.getIntPtrConstant(0));
4784 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4785 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4787 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4788 N0 = LowerCONCAT_VECTORS(N0, DAG);
4790 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4791 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4796 // v4i16 sdiv ... Convert to float.
4797 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4798 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4799 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4800 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4801 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4802 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4804 // Use reciprocal estimate and two refinement steps.
4805 // float4 recip = vrecpeq_f32(yf);
4806 // recip *= vrecpsq_f32(yf, recip);
4807 // recip *= vrecpsq_f32(yf, recip);
4808 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4809 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4810 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4811 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4813 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4814 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4815 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4817 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4818 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4819 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4820 // and that it will never cause us to return an answer too large).
4821 // float4 result = as_float4(as_int4(xf*recip) + 2);
4822 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4823 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4824 N1 = DAG.getConstant(2, MVT::i32);
4825 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4826 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4827 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4828 // Convert back to integer and return.
4829 // return vmovn_u32(vcvt_s32_f32(result));
4830 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4831 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4835 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4836 EVT VT = Op.getNode()->getValueType(0);
4837 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4840 bool ExtraOp = false;
4841 switch (Op.getOpcode()) {
4842 default: assert(0 && "Invalid code");
4843 case ISD::ADDC: Opc = ARMISD::ADDC; break;
4844 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4845 case ISD::SUBC: Opc = ARMISD::SUBC; break;
4846 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4850 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4852 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4853 Op.getOperand(1), Op.getOperand(2));
4857 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
4858 SelectionDAG &DAG, unsigned NewOp) {
4859 EVT T = Node->getValueType(0);
4860 DebugLoc dl = Node->getDebugLoc();
4861 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
4863 SDValue Chain = Node->getOperand(0);
4864 SDValue In1 = Node->getOperand(1);
4865 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4866 Node->getOperand(2), DAG.getIntPtrConstant(0));
4867 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4868 Node->getOperand(2), DAG.getIntPtrConstant(1));
4869 SDValue Ops[] = { Chain, In1, In2L, In2H };
4870 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4872 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
4873 cast<MemSDNode>(Node)->getMemOperand());
4874 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
4875 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
4876 Results.push_back(Result.getValue(2));
4879 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4880 switch (Op.getOpcode()) {
4881 default: llvm_unreachable("Don't know how to custom lower this!");
4882 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4883 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4884 case ISD::GlobalAddress:
4885 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4886 LowerGlobalAddressELF(Op, DAG);
4887 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4888 case ISD::SELECT: return LowerSELECT(Op, DAG);
4889 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4890 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4891 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4892 case ISD::VASTART: return LowerVASTART(Op, DAG);
4893 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4894 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
4895 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4896 case ISD::SINT_TO_FP:
4897 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4898 case ISD::FP_TO_SINT:
4899 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4900 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4901 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4902 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4903 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4904 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4905 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4906 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4907 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4909 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4912 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4913 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4914 case ISD::SRL_PARTS:
4915 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4916 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4917 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
4918 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4919 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4920 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4921 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4922 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4923 case ISD::MUL: return LowerMUL(Op, DAG);
4924 case ISD::SDIV: return LowerSDIV(Op, DAG);
4925 case ISD::UDIV: return LowerUDIV(Op, DAG);
4929 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
4934 /// ReplaceNodeResults - Replace the results of node with an illegal result
4935 /// type with new values built out of custom code.
4936 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4937 SmallVectorImpl<SDValue>&Results,
4938 SelectionDAG &DAG) const {
4940 switch (N->getOpcode()) {
4942 llvm_unreachable("Don't know how to custom expand this!");
4945 Res = ExpandBITCAST(N, DAG);
4949 Res = Expand64BitShift(N, DAG, Subtarget);
4951 case ISD::ATOMIC_LOAD_ADD:
4952 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
4954 case ISD::ATOMIC_LOAD_AND:
4955 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
4957 case ISD::ATOMIC_LOAD_NAND:
4958 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
4960 case ISD::ATOMIC_LOAD_OR:
4961 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
4963 case ISD::ATOMIC_LOAD_SUB:
4964 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
4966 case ISD::ATOMIC_LOAD_XOR:
4967 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
4969 case ISD::ATOMIC_SWAP:
4970 ReplaceATOMIC_BINARY_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
4972 //case ISD::ATOMIC_CMP_SWAP:
4973 // ReplaceATOMIC_CMPXCHG_64(N, Results, DAG);
4976 Results.push_back(Res);
4979 //===----------------------------------------------------------------------===//
4980 // ARM Scheduler Hooks
4981 //===----------------------------------------------------------------------===//
4984 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
4985 MachineBasicBlock *BB,
4986 unsigned Size) const {
4987 unsigned dest = MI->getOperand(0).getReg();
4988 unsigned ptr = MI->getOperand(1).getReg();
4989 unsigned oldval = MI->getOperand(2).getReg();
4990 unsigned newval = MI->getOperand(3).getReg();
4991 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4992 DebugLoc dl = MI->getDebugLoc();
4993 bool isThumb2 = Subtarget->isThumb2();
4995 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
4997 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
4998 : ARM::GPRRegisterClass);
5001 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5002 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5003 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5006 unsigned ldrOpc, strOpc;
5008 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5010 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5011 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5014 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5015 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5018 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5019 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5023 MachineFunction *MF = BB->getParent();
5024 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5025 MachineFunction::iterator It = BB;
5026 ++It; // insert the new blocks after the current block
5028 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5029 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5030 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5031 MF->insert(It, loop1MBB);
5032 MF->insert(It, loop2MBB);
5033 MF->insert(It, exitMBB);
5035 // Transfer the remainder of BB and its successor edges to exitMBB.
5036 exitMBB->splice(exitMBB->begin(), BB,
5037 llvm::next(MachineBasicBlock::iterator(MI)),
5039 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5043 // fallthrough --> loop1MBB
5044 BB->addSuccessor(loop1MBB);
5047 // ldrex dest, [ptr]
5051 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5052 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5053 .addReg(dest).addReg(oldval));
5054 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5055 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5056 BB->addSuccessor(loop2MBB);
5057 BB->addSuccessor(exitMBB);
5060 // strex scratch, newval, [ptr]
5064 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
5066 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5067 .addReg(scratch).addImm(0));
5068 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5069 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5070 BB->addSuccessor(loop1MBB);
5071 BB->addSuccessor(exitMBB);
5077 MI->eraseFromParent(); // The instruction is gone now.
5083 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5084 unsigned Size, unsigned BinOpcode) const {
5085 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5086 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5088 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5089 MachineFunction *MF = BB->getParent();
5090 MachineFunction::iterator It = BB;
5093 unsigned dest = MI->getOperand(0).getReg();
5094 unsigned ptr = MI->getOperand(1).getReg();
5095 unsigned incr = MI->getOperand(2).getReg();
5096 DebugLoc dl = MI->getDebugLoc();
5097 bool isThumb2 = Subtarget->isThumb2();
5099 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5101 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5102 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5105 unsigned ldrOpc, strOpc;
5107 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5109 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5110 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5113 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5114 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5117 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5118 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5122 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5123 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5124 MF->insert(It, loopMBB);
5125 MF->insert(It, exitMBB);
5127 // Transfer the remainder of BB and its successor edges to exitMBB.
5128 exitMBB->splice(exitMBB->begin(), BB,
5129 llvm::next(MachineBasicBlock::iterator(MI)),
5131 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5133 TargetRegisterClass *TRC =
5134 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5135 unsigned scratch = MRI.createVirtualRegister(TRC);
5136 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5140 // fallthrough --> loopMBB
5141 BB->addSuccessor(loopMBB);
5145 // <binop> scratch2, dest, incr
5146 // strex scratch, scratch2, ptr
5149 // fallthrough --> exitMBB
5151 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5153 // operand order needs to go the other way for NAND
5154 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5155 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5156 addReg(incr).addReg(dest)).addReg(0);
5158 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5159 addReg(dest).addReg(incr)).addReg(0);
5162 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5164 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5165 .addReg(scratch).addImm(0));
5166 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5167 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5169 BB->addSuccessor(loopMBB);
5170 BB->addSuccessor(exitMBB);
5176 MI->eraseFromParent(); // The instruction is gone now.
5182 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5183 MachineBasicBlock *BB,
5186 ARMCC::CondCodes Cond) const {
5187 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5189 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5190 MachineFunction *MF = BB->getParent();
5191 MachineFunction::iterator It = BB;
5194 unsigned dest = MI->getOperand(0).getReg();
5195 unsigned ptr = MI->getOperand(1).getReg();
5196 unsigned incr = MI->getOperand(2).getReg();
5197 unsigned oldval = dest;
5198 DebugLoc dl = MI->getDebugLoc();
5199 bool isThumb2 = Subtarget->isThumb2();
5201 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5203 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5204 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5207 unsigned ldrOpc, strOpc, extendOpc;
5209 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5211 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5212 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5213 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5216 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5217 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5218 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5221 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5222 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5227 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5228 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5229 MF->insert(It, loopMBB);
5230 MF->insert(It, exitMBB);
5232 // Transfer the remainder of BB and its successor edges to exitMBB.
5233 exitMBB->splice(exitMBB->begin(), BB,
5234 llvm::next(MachineBasicBlock::iterator(MI)),
5236 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5238 TargetRegisterClass *TRC =
5239 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5240 unsigned scratch = MRI.createVirtualRegister(TRC);
5241 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5245 // fallthrough --> loopMBB
5246 BB->addSuccessor(loopMBB);
5250 // (sign extend dest, if required)
5252 // cmov.cond scratch2, dest, incr
5253 // strex scratch, scratch2, ptr
5256 // fallthrough --> exitMBB
5258 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
5260 // Sign extend the value, if necessary.
5261 if (signExtend && extendOpc) {
5262 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5263 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5268 // Build compare and cmov instructions.
5269 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5270 .addReg(oldval).addReg(incr));
5271 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5272 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5274 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
5276 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5277 .addReg(scratch).addImm(0));
5278 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5279 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5281 BB->addSuccessor(loopMBB);
5282 BB->addSuccessor(exitMBB);
5288 MI->eraseFromParent(); // The instruction is gone now.
5294 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5295 unsigned Op1, unsigned Op2,
5296 bool NeedsCarry) const {
5297 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5298 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5300 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5301 MachineFunction *MF = BB->getParent();
5302 MachineFunction::iterator It = BB;
5305 unsigned destlo = MI->getOperand(0).getReg();
5306 unsigned desthi = MI->getOperand(1).getReg();
5307 unsigned ptr = MI->getOperand(2).getReg();
5308 unsigned vallo = MI->getOperand(3).getReg();
5309 unsigned valhi = MI->getOperand(4).getReg();
5310 DebugLoc dl = MI->getDebugLoc();
5311 bool isThumb2 = Subtarget->isThumb2();
5313 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5315 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5316 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5317 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5320 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5321 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5323 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5324 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5325 MF->insert(It, loopMBB);
5326 MF->insert(It, exitMBB);
5328 // Transfer the remainder of BB and its successor edges to exitMBB.
5329 exitMBB->splice(exitMBB->begin(), BB,
5330 llvm::next(MachineBasicBlock::iterator(MI)),
5332 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5334 TargetRegisterClass *TRC =
5335 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5336 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5340 // fallthrough --> loopMBB
5341 BB->addSuccessor(loopMBB);
5344 // ldrexd r2, r3, ptr
5345 // <binopa> r0, r2, incr
5346 // <binopb> r1, r3, incr
5347 // strexd storesuccess, r0, r1, ptr
5348 // cmp storesuccess, #0
5350 // fallthrough --> exitMBB
5352 // Note that the registers are explicitly specified because there is not any
5353 // way to force the register allocator to allocate a register pair.
5355 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5356 // need to properly enforce the restriction that the two output registers
5357 // for ldrexd must be different.
5360 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5361 .addReg(ARM::R2, RegState::Define)
5362 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5363 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5364 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5365 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5367 // Perform binary operation
5368 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5369 .addReg(destlo).addReg(vallo))
5370 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5371 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5372 .addReg(desthi).addReg(valhi)).addReg(0);
5374 // Copy to physregs for strexd
5375 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5376 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5380 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5381 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5383 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5384 .addReg(storesuccess).addImm(0));
5385 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5386 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5388 BB->addSuccessor(loopMBB);
5389 BB->addSuccessor(exitMBB);
5395 MI->eraseFromParent(); // The instruction is gone now.
5401 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
5402 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
5403 E = MBB->succ_end(); I != E; ++I)
5406 llvm_unreachable("Expecting a BB with two successors!");
5410 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5411 MachineBasicBlock *BB) const {
5412 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5413 DebugLoc dl = MI->getDebugLoc();
5414 bool isThumb2 = Subtarget->isThumb2();
5415 switch (MI->getOpcode()) {
5418 llvm_unreachable("Unexpected instr type to insert");
5420 case ARM::STRi_preidx:
5421 case ARM::STRBi_preidx: {
5422 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
5423 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
5424 // Decode the offset.
5425 unsigned Offset = MI->getOperand(4).getImm();
5426 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
5427 Offset = ARM_AM::getAM2Offset(Offset);
5431 MachineMemOperand *MMO = *MI->memoperands_begin();
5432 BuildMI(*BB, MI, dl, TII->get(NewOpc))
5433 .addOperand(MI->getOperand(0)) // Rn_wb
5434 .addOperand(MI->getOperand(1)) // Rt
5435 .addOperand(MI->getOperand(2)) // Rn
5436 .addImm(Offset) // offset (skip GPR==zero_reg)
5437 .addOperand(MI->getOperand(5)) // pred
5438 .addOperand(MI->getOperand(6))
5439 .addMemOperand(MMO);
5440 MI->eraseFromParent();
5443 case ARM::STRr_preidx:
5444 case ARM::STRBr_preidx:
5445 case ARM::STRH_preidx: {
5447 switch (MI->getOpcode()) {
5448 default: llvm_unreachable("unexpected opcode!");
5449 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
5450 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
5451 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
5453 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
5454 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
5455 MIB.addOperand(MI->getOperand(i));
5456 MI->eraseFromParent();
5459 case ARM::ATOMIC_LOAD_ADD_I8:
5460 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5461 case ARM::ATOMIC_LOAD_ADD_I16:
5462 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5463 case ARM::ATOMIC_LOAD_ADD_I32:
5464 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
5466 case ARM::ATOMIC_LOAD_AND_I8:
5467 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5468 case ARM::ATOMIC_LOAD_AND_I16:
5469 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5470 case ARM::ATOMIC_LOAD_AND_I32:
5471 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
5473 case ARM::ATOMIC_LOAD_OR_I8:
5474 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5475 case ARM::ATOMIC_LOAD_OR_I16:
5476 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5477 case ARM::ATOMIC_LOAD_OR_I32:
5478 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
5480 case ARM::ATOMIC_LOAD_XOR_I8:
5481 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5482 case ARM::ATOMIC_LOAD_XOR_I16:
5483 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5484 case ARM::ATOMIC_LOAD_XOR_I32:
5485 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
5487 case ARM::ATOMIC_LOAD_NAND_I8:
5488 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5489 case ARM::ATOMIC_LOAD_NAND_I16:
5490 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5491 case ARM::ATOMIC_LOAD_NAND_I32:
5492 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
5494 case ARM::ATOMIC_LOAD_SUB_I8:
5495 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5496 case ARM::ATOMIC_LOAD_SUB_I16:
5497 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5498 case ARM::ATOMIC_LOAD_SUB_I32:
5499 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
5501 case ARM::ATOMIC_LOAD_MIN_I8:
5502 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
5503 case ARM::ATOMIC_LOAD_MIN_I16:
5504 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
5505 case ARM::ATOMIC_LOAD_MIN_I32:
5506 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
5508 case ARM::ATOMIC_LOAD_MAX_I8:
5509 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
5510 case ARM::ATOMIC_LOAD_MAX_I16:
5511 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
5512 case ARM::ATOMIC_LOAD_MAX_I32:
5513 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
5515 case ARM::ATOMIC_LOAD_UMIN_I8:
5516 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
5517 case ARM::ATOMIC_LOAD_UMIN_I16:
5518 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
5519 case ARM::ATOMIC_LOAD_UMIN_I32:
5520 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
5522 case ARM::ATOMIC_LOAD_UMAX_I8:
5523 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
5524 case ARM::ATOMIC_LOAD_UMAX_I16:
5525 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
5526 case ARM::ATOMIC_LOAD_UMAX_I32:
5527 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
5529 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
5530 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
5531 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
5533 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
5534 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
5535 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
5538 case ARM::ATOMADD6432:
5539 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
5540 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr, true);
5541 case ARM::ATOMSUB6432:
5542 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
5543 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, true);
5544 case ARM::ATOMOR6432:
5545 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
5546 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr, false);
5547 case ARM::ATOMXOR6432:
5548 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
5549 isThumb2 ? ARM::t2EORrr : ARM::EORrr, false);
5550 case ARM::ATOMAND6432:
5551 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
5552 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr, false);
5553 case ARM::ATOMSWAP6432:
5554 return EmitAtomicBinary64(MI, BB, 0, 0, false);
5556 case ARM::tMOVCCr_pseudo: {
5557 // To "insert" a SELECT_CC instruction, we actually have to insert the
5558 // diamond control-flow pattern. The incoming instruction knows the
5559 // destination vreg to set, the condition code register to branch on, the
5560 // true/false values to select between, and a branch opcode to use.
5561 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5562 MachineFunction::iterator It = BB;
5568 // cmpTY ccX, r1, r2
5570 // fallthrough --> copy0MBB
5571 MachineBasicBlock *thisMBB = BB;
5572 MachineFunction *F = BB->getParent();
5573 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5574 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5575 F->insert(It, copy0MBB);
5576 F->insert(It, sinkMBB);
5578 // Transfer the remainder of BB and its successor edges to sinkMBB.
5579 sinkMBB->splice(sinkMBB->begin(), BB,
5580 llvm::next(MachineBasicBlock::iterator(MI)),
5582 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5584 BB->addSuccessor(copy0MBB);
5585 BB->addSuccessor(sinkMBB);
5587 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
5588 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
5591 // %FalseValue = ...
5592 // # fallthrough to sinkMBB
5595 // Update machine-CFG edges
5596 BB->addSuccessor(sinkMBB);
5599 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5602 BuildMI(*BB, BB->begin(), dl,
5603 TII->get(ARM::PHI), MI->getOperand(0).getReg())
5604 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
5605 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5607 MI->eraseFromParent(); // The pseudo instruction is gone now.
5612 case ARM::BCCZi64: {
5613 // If there is an unconditional branch to the other successor, remove it.
5614 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
5616 // Compare both parts that make up the double comparison separately for
5618 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
5620 unsigned LHS1 = MI->getOperand(1).getReg();
5621 unsigned LHS2 = MI->getOperand(2).getReg();
5623 AddDefaultPred(BuildMI(BB, dl,
5624 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5625 .addReg(LHS1).addImm(0));
5626 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5627 .addReg(LHS2).addImm(0)
5628 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5630 unsigned RHS1 = MI->getOperand(3).getReg();
5631 unsigned RHS2 = MI->getOperand(4).getReg();
5632 AddDefaultPred(BuildMI(BB, dl,
5633 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5634 .addReg(LHS1).addReg(RHS1));
5635 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5636 .addReg(LHS2).addReg(RHS2)
5637 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
5640 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
5641 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
5642 if (MI->getOperand(0).getImm() == ARMCC::NE)
5643 std::swap(destMBB, exitMBB);
5645 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5646 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
5647 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2B : ARM::B))
5650 MI->eraseFromParent(); // The pseudo instruction is gone now.
5656 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
5657 SDNode *Node) const {
5658 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC,
5659 // RSB, RSC. Coming out of isel, they have an implicit CPSR def, but the
5660 // optional operand is not filled in. If the carry bit is used, then change
5661 // the optional operand to CPSR. Otherwise, remove the CPSR implicit def.
5662 const MCInstrDesc &MCID = MI->getDesc();
5663 if (Node->hasAnyUseOfValue(1)) {
5664 MachineOperand &MO = MI->getOperand(MCID.getNumOperands() - 2);
5665 MO.setReg(ARM::CPSR);
5668 for (unsigned i = MCID.getNumOperands(), e = MI->getNumOperands();
5670 const MachineOperand &MO = MI->getOperand(i);
5671 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
5672 MI->RemoveOperand(i);
5679 //===----------------------------------------------------------------------===//
5680 // ARM Optimization Hooks
5681 //===----------------------------------------------------------------------===//
5684 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
5685 TargetLowering::DAGCombinerInfo &DCI) {
5686 SelectionDAG &DAG = DCI.DAG;
5687 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5688 EVT VT = N->getValueType(0);
5689 unsigned Opc = N->getOpcode();
5690 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
5691 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
5692 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
5693 ISD::CondCode CC = ISD::SETCC_INVALID;
5696 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
5698 SDValue CCOp = Slct.getOperand(0);
5699 if (CCOp.getOpcode() == ISD::SETCC)
5700 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
5703 bool DoXform = false;
5705 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
5708 if (LHS.getOpcode() == ISD::Constant &&
5709 cast<ConstantSDNode>(LHS)->isNullValue()) {
5711 } else if (CC != ISD::SETCC_INVALID &&
5712 RHS.getOpcode() == ISD::Constant &&
5713 cast<ConstantSDNode>(RHS)->isNullValue()) {
5714 std::swap(LHS, RHS);
5715 SDValue Op0 = Slct.getOperand(0);
5716 EVT OpVT = isSlctCC ? Op0.getValueType() :
5717 Op0.getOperand(0).getValueType();
5718 bool isInt = OpVT.isInteger();
5719 CC = ISD::getSetCCInverse(CC, isInt);
5721 if (!TLI.isCondCodeLegal(CC, OpVT))
5722 return SDValue(); // Inverse operator isn't legal.
5729 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
5731 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
5732 Slct.getOperand(0), Slct.getOperand(1), CC);
5733 SDValue CCOp = Slct.getOperand(0);
5735 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
5736 CCOp.getOperand(0), CCOp.getOperand(1), CC);
5737 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
5738 CCOp, OtherOp, Result);
5743 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
5744 // (only after legalization).
5745 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
5746 TargetLowering::DAGCombinerInfo &DCI,
5747 const ARMSubtarget *Subtarget) {
5749 // Only perform optimization if after legalize, and if NEON is available. We
5750 // also expected both operands to be BUILD_VECTORs.
5751 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
5752 || N0.getOpcode() != ISD::BUILD_VECTOR
5753 || N1.getOpcode() != ISD::BUILD_VECTOR)
5756 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
5757 EVT VT = N->getValueType(0);
5758 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
5761 // Check that the vector operands are of the right form.
5762 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
5763 // operands, where N is the size of the formed vector.
5764 // Each EXTRACT_VECTOR should have the same input vector and odd or even
5765 // index such that we have a pair wise add pattern.
5767 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
5768 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5770 SDValue Vec = N0->getOperand(0)->getOperand(0);
5771 SDNode *V = Vec.getNode();
5772 unsigned nextIndex = 0;
5774 // For each operands to the ADD which are BUILD_VECTORs,
5775 // check to see if each of their operands are an EXTRACT_VECTOR with
5776 // the same vector and appropriate index.
5777 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
5778 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
5779 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
5781 SDValue ExtVec0 = N0->getOperand(i);
5782 SDValue ExtVec1 = N1->getOperand(i);
5784 // First operand is the vector, verify its the same.
5785 if (V != ExtVec0->getOperand(0).getNode() ||
5786 V != ExtVec1->getOperand(0).getNode())
5789 // Second is the constant, verify its correct.
5790 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
5791 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
5793 // For the constant, we want to see all the even or all the odd.
5794 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
5795 || C1->getZExtValue() != nextIndex+1)
5804 // Create VPADDL node.
5805 SelectionDAG &DAG = DCI.DAG;
5806 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5808 // Build operand list.
5809 SmallVector<SDValue, 8> Ops;
5810 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
5811 TLI.getPointerTy()));
5813 // Input is the vector.
5816 // Get widened type and narrowed type.
5818 unsigned numElem = VT.getVectorNumElements();
5819 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
5820 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
5821 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
5822 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
5824 assert(0 && "Invalid vector element type for padd optimization.");
5827 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
5828 widenType, &Ops[0], Ops.size());
5829 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
5832 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
5833 /// operands N0 and N1. This is a helper for PerformADDCombine that is
5834 /// called with the default operands, and if that fails, with commuted
5836 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
5837 TargetLowering::DAGCombinerInfo &DCI,
5838 const ARMSubtarget *Subtarget){
5840 // Attempt to create vpaddl for this add.
5841 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
5842 if (Result.getNode())
5845 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
5846 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
5847 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
5848 if (Result.getNode()) return Result;
5853 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
5855 static SDValue PerformADDCombine(SDNode *N,
5856 TargetLowering::DAGCombinerInfo &DCI,
5857 const ARMSubtarget *Subtarget) {
5858 SDValue N0 = N->getOperand(0);
5859 SDValue N1 = N->getOperand(1);
5861 // First try with the default operand order.
5862 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
5863 if (Result.getNode())
5866 // If that didn't work, try again with the operands commuted.
5867 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
5870 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
5872 static SDValue PerformSUBCombine(SDNode *N,
5873 TargetLowering::DAGCombinerInfo &DCI) {
5874 SDValue N0 = N->getOperand(0);
5875 SDValue N1 = N->getOperand(1);
5877 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
5878 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
5879 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
5880 if (Result.getNode()) return Result;
5886 /// PerformVMULCombine
5887 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
5888 /// special multiplier accumulator forwarding.
5894 static SDValue PerformVMULCombine(SDNode *N,
5895 TargetLowering::DAGCombinerInfo &DCI,
5896 const ARMSubtarget *Subtarget) {
5897 if (!Subtarget->hasVMLxForwarding())
5900 SelectionDAG &DAG = DCI.DAG;
5901 SDValue N0 = N->getOperand(0);
5902 SDValue N1 = N->getOperand(1);
5903 unsigned Opcode = N0.getOpcode();
5904 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5905 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
5906 Opcode = N1.getOpcode();
5907 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
5908 Opcode != ISD::FADD && Opcode != ISD::FSUB)
5913 EVT VT = N->getValueType(0);
5914 DebugLoc DL = N->getDebugLoc();
5915 SDValue N00 = N0->getOperand(0);
5916 SDValue N01 = N0->getOperand(1);
5917 return DAG.getNode(Opcode, DL, VT,
5918 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
5919 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
5922 static SDValue PerformMULCombine(SDNode *N,
5923 TargetLowering::DAGCombinerInfo &DCI,
5924 const ARMSubtarget *Subtarget) {
5925 SelectionDAG &DAG = DCI.DAG;
5927 if (Subtarget->isThumb1Only())
5930 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
5933 EVT VT = N->getValueType(0);
5934 if (VT.is64BitVector() || VT.is128BitVector())
5935 return PerformVMULCombine(N, DCI, Subtarget);
5939 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
5943 uint64_t MulAmt = C->getZExtValue();
5944 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
5945 ShiftAmt = ShiftAmt & (32 - 1);
5946 SDValue V = N->getOperand(0);
5947 DebugLoc DL = N->getDebugLoc();
5950 MulAmt >>= ShiftAmt;
5951 if (isPowerOf2_32(MulAmt - 1)) {
5952 // (mul x, 2^N + 1) => (add (shl x, N), x)
5953 Res = DAG.getNode(ISD::ADD, DL, VT,
5954 V, DAG.getNode(ISD::SHL, DL, VT,
5955 V, DAG.getConstant(Log2_32(MulAmt-1),
5957 } else if (isPowerOf2_32(MulAmt + 1)) {
5958 // (mul x, 2^N - 1) => (sub (shl x, N), x)
5959 Res = DAG.getNode(ISD::SUB, DL, VT,
5960 DAG.getNode(ISD::SHL, DL, VT,
5961 V, DAG.getConstant(Log2_32(MulAmt+1),
5968 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
5969 DAG.getConstant(ShiftAmt, MVT::i32));
5971 // Do not add new nodes to DAG combiner worklist.
5972 DCI.CombineTo(N, Res, false);
5976 static SDValue PerformANDCombine(SDNode *N,
5977 TargetLowering::DAGCombinerInfo &DCI) {
5979 // Attempt to use immediate-form VBIC
5980 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
5981 DebugLoc dl = N->getDebugLoc();
5982 EVT VT = N->getValueType(0);
5983 SelectionDAG &DAG = DCI.DAG;
5985 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
5988 APInt SplatBits, SplatUndef;
5989 unsigned SplatBitSize;
5992 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5993 if (SplatBitSize <= 64) {
5995 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
5996 SplatUndef.getZExtValue(), SplatBitSize,
5997 DAG, VbicVT, VT.is128BitVector(),
5999 if (Val.getNode()) {
6001 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6002 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6003 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6011 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6012 static SDValue PerformORCombine(SDNode *N,
6013 TargetLowering::DAGCombinerInfo &DCI,
6014 const ARMSubtarget *Subtarget) {
6015 // Attempt to use immediate-form VORR
6016 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6017 DebugLoc dl = N->getDebugLoc();
6018 EVT VT = N->getValueType(0);
6019 SelectionDAG &DAG = DCI.DAG;
6021 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6024 APInt SplatBits, SplatUndef;
6025 unsigned SplatBitSize;
6027 if (BVN && Subtarget->hasNEON() &&
6028 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6029 if (SplatBitSize <= 64) {
6031 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6032 SplatUndef.getZExtValue(), SplatBitSize,
6033 DAG, VorrVT, VT.is128BitVector(),
6035 if (Val.getNode()) {
6037 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6038 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6039 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6044 SDValue N0 = N->getOperand(0);
6045 if (N0.getOpcode() != ISD::AND)
6047 SDValue N1 = N->getOperand(1);
6049 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6050 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6051 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6053 unsigned SplatBitSize;
6056 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6058 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6059 HasAnyUndefs) && !HasAnyUndefs) {
6060 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6062 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6063 HasAnyUndefs) && !HasAnyUndefs &&
6064 SplatBits0 == ~SplatBits1) {
6065 // Canonicalize the vector type to make instruction selection simpler.
6066 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6067 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6068 N0->getOperand(1), N0->getOperand(0),
6070 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6075 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6078 // BFI is only available on V6T2+
6079 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6082 DebugLoc DL = N->getDebugLoc();
6083 // 1) or (and A, mask), val => ARMbfi A, val, mask
6084 // iff (val & mask) == val
6086 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6087 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6088 // && mask == ~mask2
6089 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6090 // && ~mask == mask2
6091 // (i.e., copy a bitfield value into another bitfield of the same width)
6096 SDValue N00 = N0.getOperand(0);
6098 // The value and the mask need to be constants so we can verify this is
6099 // actually a bitfield set. If the mask is 0xffff, we can do better
6100 // via a movt instruction, so don't use BFI in that case.
6101 SDValue MaskOp = N0.getOperand(1);
6102 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6105 unsigned Mask = MaskC->getZExtValue();
6109 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6110 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6112 unsigned Val = N1C->getZExtValue();
6113 if ((Val & ~Mask) != Val)
6116 if (ARM::isBitFieldInvertedMask(Mask)) {
6117 Val >>= CountTrailingZeros_32(~Mask);
6119 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6120 DAG.getConstant(Val, MVT::i32),
6121 DAG.getConstant(Mask, MVT::i32));
6123 // Do not add new nodes to DAG combiner worklist.
6124 DCI.CombineTo(N, Res, false);
6127 } else if (N1.getOpcode() == ISD::AND) {
6128 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6129 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6132 unsigned Mask2 = N11C->getZExtValue();
6134 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
6136 if (ARM::isBitFieldInvertedMask(Mask) &&
6138 // The pack halfword instruction works better for masks that fit it,
6139 // so use that when it's available.
6140 if (Subtarget->hasT2ExtractPack() &&
6141 (Mask == 0xffff || Mask == 0xffff0000))
6144 unsigned amt = CountTrailingZeros_32(Mask2);
6145 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6146 DAG.getConstant(amt, MVT::i32));
6147 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6148 DAG.getConstant(Mask, MVT::i32));
6149 // Do not add new nodes to DAG combiner worklist.
6150 DCI.CombineTo(N, Res, false);
6152 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
6154 // The pack halfword instruction works better for masks that fit it,
6155 // so use that when it's available.
6156 if (Subtarget->hasT2ExtractPack() &&
6157 (Mask2 == 0xffff || Mask2 == 0xffff0000))
6160 unsigned lsb = CountTrailingZeros_32(Mask);
6161 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
6162 DAG.getConstant(lsb, MVT::i32));
6163 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
6164 DAG.getConstant(Mask2, MVT::i32));
6165 // Do not add new nodes to DAG combiner worklist.
6166 DCI.CombineTo(N, Res, false);
6171 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
6172 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
6173 ARM::isBitFieldInvertedMask(~Mask)) {
6174 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
6175 // where lsb(mask) == #shamt and masked bits of B are known zero.
6176 SDValue ShAmt = N00.getOperand(1);
6177 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6178 unsigned LSB = CountTrailingZeros_32(Mask);
6182 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
6183 DAG.getConstant(~Mask, MVT::i32));
6185 // Do not add new nodes to DAG combiner worklist.
6186 DCI.CombineTo(N, Res, false);
6192 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
6193 /// the bits being cleared by the AND are not demanded by the BFI.
6194 static SDValue PerformBFICombine(SDNode *N,
6195 TargetLowering::DAGCombinerInfo &DCI) {
6196 SDValue N1 = N->getOperand(1);
6197 if (N1.getOpcode() == ISD::AND) {
6198 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6201 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
6202 unsigned LSB = CountTrailingZeros_32(~InvMask);
6203 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
6204 unsigned Mask = (1 << Width)-1;
6205 unsigned Mask2 = N11C->getZExtValue();
6206 if ((Mask & (~Mask2)) == 0)
6207 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
6208 N->getOperand(0), N1.getOperand(0),
6214 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
6215 /// ARMISD::VMOVRRD.
6216 static SDValue PerformVMOVRRDCombine(SDNode *N,
6217 TargetLowering::DAGCombinerInfo &DCI) {
6218 // vmovrrd(vmovdrr x, y) -> x,y
6219 SDValue InDouble = N->getOperand(0);
6220 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
6221 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
6223 // vmovrrd(load f64) -> (load i32), (load i32)
6224 SDNode *InNode = InDouble.getNode();
6225 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
6226 InNode->getValueType(0) == MVT::f64 &&
6227 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
6228 !cast<LoadSDNode>(InNode)->isVolatile()) {
6229 // TODO: Should this be done for non-FrameIndex operands?
6230 LoadSDNode *LD = cast<LoadSDNode>(InNode);
6232 SelectionDAG &DAG = DCI.DAG;
6233 DebugLoc DL = LD->getDebugLoc();
6234 SDValue BasePtr = LD->getBasePtr();
6235 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
6236 LD->getPointerInfo(), LD->isVolatile(),
6237 LD->isNonTemporal(), LD->getAlignment());
6239 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
6240 DAG.getConstant(4, MVT::i32));
6241 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
6242 LD->getPointerInfo(), LD->isVolatile(),
6243 LD->isNonTemporal(),
6244 std::min(4U, LD->getAlignment() / 2));
6246 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
6247 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
6248 DCI.RemoveFromWorklist(LD);
6256 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
6257 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
6258 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
6259 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
6260 SDValue Op0 = N->getOperand(0);
6261 SDValue Op1 = N->getOperand(1);
6262 if (Op0.getOpcode() == ISD::BITCAST)
6263 Op0 = Op0.getOperand(0);
6264 if (Op1.getOpcode() == ISD::BITCAST)
6265 Op1 = Op1.getOperand(0);
6266 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
6267 Op0.getNode() == Op1.getNode() &&
6268 Op0.getResNo() == 0 && Op1.getResNo() == 1)
6269 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6270 N->getValueType(0), Op0.getOperand(0));
6274 /// PerformSTORECombine - Target-specific dag combine xforms for
6276 static SDValue PerformSTORECombine(SDNode *N,
6277 TargetLowering::DAGCombinerInfo &DCI) {
6278 // Bitcast an i64 store extracted from a vector to f64.
6279 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6280 StoreSDNode *St = cast<StoreSDNode>(N);
6281 SDValue StVal = St->getValue();
6282 if (!ISD::isNormalStore(St) || St->isVolatile())
6285 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
6286 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
6287 SelectionDAG &DAG = DCI.DAG;
6288 DebugLoc DL = St->getDebugLoc();
6289 SDValue BasePtr = St->getBasePtr();
6290 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
6291 StVal.getNode()->getOperand(0), BasePtr,
6292 St->getPointerInfo(), St->isVolatile(),
6293 St->isNonTemporal(), St->getAlignment());
6295 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
6296 DAG.getConstant(4, MVT::i32));
6297 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
6298 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
6299 St->isNonTemporal(),
6300 std::min(4U, St->getAlignment() / 2));
6303 if (StVal.getValueType() != MVT::i64 ||
6304 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6307 SelectionDAG &DAG = DCI.DAG;
6308 DebugLoc dl = StVal.getDebugLoc();
6309 SDValue IntVec = StVal.getOperand(0);
6310 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6311 IntVec.getValueType().getVectorNumElements());
6312 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
6313 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6314 Vec, StVal.getOperand(1));
6315 dl = N->getDebugLoc();
6316 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
6317 // Make the DAGCombiner fold the bitcasts.
6318 DCI.AddToWorklist(Vec.getNode());
6319 DCI.AddToWorklist(ExtElt.getNode());
6320 DCI.AddToWorklist(V.getNode());
6321 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
6322 St->getPointerInfo(), St->isVolatile(),
6323 St->isNonTemporal(), St->getAlignment(),
6327 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
6328 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
6329 /// i64 vector to have f64 elements, since the value can then be loaded
6330 /// directly into a VFP register.
6331 static bool hasNormalLoadOperand(SDNode *N) {
6332 unsigned NumElts = N->getValueType(0).getVectorNumElements();
6333 for (unsigned i = 0; i < NumElts; ++i) {
6334 SDNode *Elt = N->getOperand(i).getNode();
6335 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
6341 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
6342 /// ISD::BUILD_VECTOR.
6343 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
6344 TargetLowering::DAGCombinerInfo &DCI){
6345 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
6346 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
6347 // into a pair of GPRs, which is fine when the value is used as a scalar,
6348 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
6349 SelectionDAG &DAG = DCI.DAG;
6350 if (N->getNumOperands() == 2) {
6351 SDValue RV = PerformVMOVDRRCombine(N, DAG);
6356 // Load i64 elements as f64 values so that type legalization does not split
6357 // them up into i32 values.
6358 EVT VT = N->getValueType(0);
6359 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
6361 DebugLoc dl = N->getDebugLoc();
6362 SmallVector<SDValue, 8> Ops;
6363 unsigned NumElts = VT.getVectorNumElements();
6364 for (unsigned i = 0; i < NumElts; ++i) {
6365 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
6367 // Make the DAGCombiner fold the bitcast.
6368 DCI.AddToWorklist(V.getNode());
6370 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
6371 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
6372 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
6375 /// PerformInsertEltCombine - Target-specific dag combine xforms for
6376 /// ISD::INSERT_VECTOR_ELT.
6377 static SDValue PerformInsertEltCombine(SDNode *N,
6378 TargetLowering::DAGCombinerInfo &DCI) {
6379 // Bitcast an i64 load inserted into a vector to f64.
6380 // Otherwise, the i64 value will be legalized to a pair of i32 values.
6381 EVT VT = N->getValueType(0);
6382 SDNode *Elt = N->getOperand(1).getNode();
6383 if (VT.getVectorElementType() != MVT::i64 ||
6384 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
6387 SelectionDAG &DAG = DCI.DAG;
6388 DebugLoc dl = N->getDebugLoc();
6389 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
6390 VT.getVectorNumElements());
6391 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
6392 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
6393 // Make the DAGCombiner fold the bitcasts.
6394 DCI.AddToWorklist(Vec.getNode());
6395 DCI.AddToWorklist(V.getNode());
6396 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
6397 Vec, V, N->getOperand(2));
6398 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
6401 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
6402 /// ISD::VECTOR_SHUFFLE.
6403 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
6404 // The LLVM shufflevector instruction does not require the shuffle mask
6405 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
6406 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
6407 // operands do not match the mask length, they are extended by concatenating
6408 // them with undef vectors. That is probably the right thing for other
6409 // targets, but for NEON it is better to concatenate two double-register
6410 // size vector operands into a single quad-register size vector. Do that
6411 // transformation here:
6412 // shuffle(concat(v1, undef), concat(v2, undef)) ->
6413 // shuffle(concat(v1, v2), undef)
6414 SDValue Op0 = N->getOperand(0);
6415 SDValue Op1 = N->getOperand(1);
6416 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
6417 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
6418 Op0.getNumOperands() != 2 ||
6419 Op1.getNumOperands() != 2)
6421 SDValue Concat0Op1 = Op0.getOperand(1);
6422 SDValue Concat1Op1 = Op1.getOperand(1);
6423 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
6424 Concat1Op1.getOpcode() != ISD::UNDEF)
6426 // Skip the transformation if any of the types are illegal.
6427 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6428 EVT VT = N->getValueType(0);
6429 if (!TLI.isTypeLegal(VT) ||
6430 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
6431 !TLI.isTypeLegal(Concat1Op1.getValueType()))
6434 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
6435 Op0.getOperand(0), Op1.getOperand(0));
6436 // Translate the shuffle mask.
6437 SmallVector<int, 16> NewMask;
6438 unsigned NumElts = VT.getVectorNumElements();
6439 unsigned HalfElts = NumElts/2;
6440 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
6441 for (unsigned n = 0; n < NumElts; ++n) {
6442 int MaskElt = SVN->getMaskElt(n);
6444 if (MaskElt < (int)HalfElts)
6446 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
6447 NewElt = HalfElts + MaskElt - NumElts;
6448 NewMask.push_back(NewElt);
6450 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
6451 DAG.getUNDEF(VT), NewMask.data());
6454 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
6455 /// NEON load/store intrinsics to merge base address updates.
6456 static SDValue CombineBaseUpdate(SDNode *N,
6457 TargetLowering::DAGCombinerInfo &DCI) {
6458 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6461 SelectionDAG &DAG = DCI.DAG;
6462 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
6463 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
6464 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
6465 SDValue Addr = N->getOperand(AddrOpIdx);
6467 // Search for a use of the address operand that is an increment.
6468 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
6469 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
6471 if (User->getOpcode() != ISD::ADD ||
6472 UI.getUse().getResNo() != Addr.getResNo())
6475 // Check that the add is independent of the load/store. Otherwise, folding
6476 // it would create a cycle.
6477 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
6480 // Find the new opcode for the updating load/store.
6482 bool isLaneOp = false;
6483 unsigned NewOpc = 0;
6484 unsigned NumVecs = 0;
6486 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
6488 default: assert(0 && "unexpected intrinsic for Neon base update");
6489 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
6491 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
6493 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
6495 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
6497 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
6498 NumVecs = 2; isLaneOp = true; break;
6499 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
6500 NumVecs = 3; isLaneOp = true; break;
6501 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
6502 NumVecs = 4; isLaneOp = true; break;
6503 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
6504 NumVecs = 1; isLoad = false; break;
6505 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
6506 NumVecs = 2; isLoad = false; break;
6507 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
6508 NumVecs = 3; isLoad = false; break;
6509 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
6510 NumVecs = 4; isLoad = false; break;
6511 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
6512 NumVecs = 2; isLoad = false; isLaneOp = true; break;
6513 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
6514 NumVecs = 3; isLoad = false; isLaneOp = true; break;
6515 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
6516 NumVecs = 4; isLoad = false; isLaneOp = true; break;
6520 switch (N->getOpcode()) {
6521 default: assert(0 && "unexpected opcode for Neon base update");
6522 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
6523 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
6524 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
6528 // Find the size of memory referenced by the load/store.
6531 VecTy = N->getValueType(0);
6533 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
6534 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
6536 NumBytes /= VecTy.getVectorNumElements();
6538 // If the increment is a constant, it must match the memory ref size.
6539 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
6540 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
6541 uint64_t IncVal = CInc->getZExtValue();
6542 if (IncVal != NumBytes)
6544 } else if (NumBytes >= 3 * 16) {
6545 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
6546 // separate instructions that make it harder to use a non-constant update.
6550 // Create the new updating load/store node.
6552 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
6554 for (n = 0; n < NumResultVecs; ++n)
6556 Tys[n++] = MVT::i32;
6557 Tys[n] = MVT::Other;
6558 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
6559 SmallVector<SDValue, 8> Ops;
6560 Ops.push_back(N->getOperand(0)); // incoming chain
6561 Ops.push_back(N->getOperand(AddrOpIdx));
6563 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
6564 Ops.push_back(N->getOperand(i));
6566 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
6567 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
6568 Ops.data(), Ops.size(),
6569 MemInt->getMemoryVT(),
6570 MemInt->getMemOperand());
6573 std::vector<SDValue> NewResults;
6574 for (unsigned i = 0; i < NumResultVecs; ++i) {
6575 NewResults.push_back(SDValue(UpdN.getNode(), i));
6577 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
6578 DCI.CombineTo(N, NewResults);
6579 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
6586 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
6587 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
6588 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
6590 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
6591 SelectionDAG &DAG = DCI.DAG;
6592 EVT VT = N->getValueType(0);
6593 // vldN-dup instructions only support 64-bit vectors for N > 1.
6594 if (!VT.is64BitVector())
6597 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
6598 SDNode *VLD = N->getOperand(0).getNode();
6599 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
6601 unsigned NumVecs = 0;
6602 unsigned NewOpc = 0;
6603 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
6604 if (IntNo == Intrinsic::arm_neon_vld2lane) {
6606 NewOpc = ARMISD::VLD2DUP;
6607 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
6609 NewOpc = ARMISD::VLD3DUP;
6610 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
6612 NewOpc = ARMISD::VLD4DUP;
6617 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
6618 // numbers match the load.
6619 unsigned VLDLaneNo =
6620 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
6621 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6623 // Ignore uses of the chain result.
6624 if (UI.getUse().getResNo() == NumVecs)
6627 if (User->getOpcode() != ARMISD::VDUPLANE ||
6628 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
6632 // Create the vldN-dup node.
6635 for (n = 0; n < NumVecs; ++n)
6637 Tys[n] = MVT::Other;
6638 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
6639 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
6640 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
6641 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
6642 Ops, 2, VLDMemInt->getMemoryVT(),
6643 VLDMemInt->getMemOperand());
6646 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
6648 unsigned ResNo = UI.getUse().getResNo();
6649 // Ignore uses of the chain result.
6650 if (ResNo == NumVecs)
6653 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
6656 // Now the vldN-lane intrinsic is dead except for its chain result.
6657 // Update uses of the chain.
6658 std::vector<SDValue> VLDDupResults;
6659 for (unsigned n = 0; n < NumVecs; ++n)
6660 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
6661 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
6662 DCI.CombineTo(VLD, VLDDupResults);
6667 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
6668 /// ARMISD::VDUPLANE.
6669 static SDValue PerformVDUPLANECombine(SDNode *N,
6670 TargetLowering::DAGCombinerInfo &DCI) {
6671 SDValue Op = N->getOperand(0);
6673 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
6674 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
6675 if (CombineVLDDUP(N, DCI))
6676 return SDValue(N, 0);
6678 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
6679 // redundant. Ignore bit_converts for now; element sizes are checked below.
6680 while (Op.getOpcode() == ISD::BITCAST)
6681 Op = Op.getOperand(0);
6682 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
6685 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
6686 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
6687 // The canonical VMOV for a zero vector uses a 32-bit element size.
6688 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6690 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
6692 EVT VT = N->getValueType(0);
6693 if (EltSize > VT.getVectorElementType().getSizeInBits())
6696 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
6699 // isConstVecPow2 - Return true if each vector element is a power of 2, all
6700 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
6701 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
6705 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
6707 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
6712 APFloat APF = C->getValueAPF();
6713 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
6714 != APFloat::opOK || !isExact)
6717 c0 = (I == 0) ? cN : c0;
6718 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
6725 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
6726 /// can replace combinations of VMUL and VCVT (floating-point to integer)
6727 /// when the VMUL has a constant operand that is a power of 2.
6729 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
6730 /// vmul.f32 d16, d17, d16
6731 /// vcvt.s32.f32 d16, d16
6733 /// vcvt.s32.f32 d16, d16, #3
6734 static SDValue PerformVCVTCombine(SDNode *N,
6735 TargetLowering::DAGCombinerInfo &DCI,
6736 const ARMSubtarget *Subtarget) {
6737 SelectionDAG &DAG = DCI.DAG;
6738 SDValue Op = N->getOperand(0);
6740 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
6741 Op.getOpcode() != ISD::FMUL)
6745 SDValue N0 = Op->getOperand(0);
6746 SDValue ConstVec = Op->getOperand(1);
6747 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
6749 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
6750 !isConstVecPow2(ConstVec, isSigned, C))
6753 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
6754 Intrinsic::arm_neon_vcvtfp2fxu;
6755 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6757 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
6758 DAG.getConstant(Log2_64(C), MVT::i32));
6761 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
6762 /// can replace combinations of VCVT (integer to floating-point) and VDIV
6763 /// when the VDIV has a constant operand that is a power of 2.
6765 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
6766 /// vcvt.f32.s32 d16, d16
6767 /// vdiv.f32 d16, d17, d16
6769 /// vcvt.f32.s32 d16, d16, #3
6770 static SDValue PerformVDIVCombine(SDNode *N,
6771 TargetLowering::DAGCombinerInfo &DCI,
6772 const ARMSubtarget *Subtarget) {
6773 SelectionDAG &DAG = DCI.DAG;
6774 SDValue Op = N->getOperand(0);
6775 unsigned OpOpcode = Op.getNode()->getOpcode();
6777 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
6778 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
6782 SDValue ConstVec = N->getOperand(1);
6783 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
6785 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
6786 !isConstVecPow2(ConstVec, isSigned, C))
6789 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
6790 Intrinsic::arm_neon_vcvtfxu2fp;
6791 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6793 DAG.getConstant(IntrinsicOpcode, MVT::i32),
6794 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
6797 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
6798 /// operand of a vector shift operation, where all the elements of the
6799 /// build_vector must have the same constant integer value.
6800 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
6801 // Ignore bit_converts.
6802 while (Op.getOpcode() == ISD::BITCAST)
6803 Op = Op.getOperand(0);
6804 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
6805 APInt SplatBits, SplatUndef;
6806 unsigned SplatBitSize;
6808 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
6809 HasAnyUndefs, ElementBits) ||
6810 SplatBitSize > ElementBits)
6812 Cnt = SplatBits.getSExtValue();
6816 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
6817 /// operand of a vector shift left operation. That value must be in the range:
6818 /// 0 <= Value < ElementBits for a left shift; or
6819 /// 0 <= Value <= ElementBits for a long left shift.
6820 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
6821 assert(VT.isVector() && "vector shift count is not a vector type");
6822 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6823 if (! getVShiftImm(Op, ElementBits, Cnt))
6825 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
6828 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
6829 /// operand of a vector shift right operation. For a shift opcode, the value
6830 /// is positive, but for an intrinsic the value count must be negative. The
6831 /// absolute value must be in the range:
6832 /// 1 <= |Value| <= ElementBits for a right shift; or
6833 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
6834 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
6836 assert(VT.isVector() && "vector shift count is not a vector type");
6837 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
6838 if (! getVShiftImm(Op, ElementBits, Cnt))
6842 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
6845 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
6846 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
6847 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
6850 // Don't do anything for most intrinsics.
6853 // Vector shifts: check for immediate versions and lower them.
6854 // Note: This is done during DAG combining instead of DAG legalizing because
6855 // the build_vectors for 64-bit vector element shift counts are generally
6856 // not legal, and it is hard to see their values after they get legalized to
6857 // loads from a constant pool.
6858 case Intrinsic::arm_neon_vshifts:
6859 case Intrinsic::arm_neon_vshiftu:
6860 case Intrinsic::arm_neon_vshiftls:
6861 case Intrinsic::arm_neon_vshiftlu:
6862 case Intrinsic::arm_neon_vshiftn:
6863 case Intrinsic::arm_neon_vrshifts:
6864 case Intrinsic::arm_neon_vrshiftu:
6865 case Intrinsic::arm_neon_vrshiftn:
6866 case Intrinsic::arm_neon_vqshifts:
6867 case Intrinsic::arm_neon_vqshiftu:
6868 case Intrinsic::arm_neon_vqshiftsu:
6869 case Intrinsic::arm_neon_vqshiftns:
6870 case Intrinsic::arm_neon_vqshiftnu:
6871 case Intrinsic::arm_neon_vqshiftnsu:
6872 case Intrinsic::arm_neon_vqrshiftns:
6873 case Intrinsic::arm_neon_vqrshiftnu:
6874 case Intrinsic::arm_neon_vqrshiftnsu: {
6875 EVT VT = N->getOperand(1).getValueType();
6877 unsigned VShiftOpc = 0;
6880 case Intrinsic::arm_neon_vshifts:
6881 case Intrinsic::arm_neon_vshiftu:
6882 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
6883 VShiftOpc = ARMISD::VSHL;
6886 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
6887 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
6888 ARMISD::VSHRs : ARMISD::VSHRu);
6893 case Intrinsic::arm_neon_vshiftls:
6894 case Intrinsic::arm_neon_vshiftlu:
6895 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
6897 llvm_unreachable("invalid shift count for vshll intrinsic");
6899 case Intrinsic::arm_neon_vrshifts:
6900 case Intrinsic::arm_neon_vrshiftu:
6901 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
6905 case Intrinsic::arm_neon_vqshifts:
6906 case Intrinsic::arm_neon_vqshiftu:
6907 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6911 case Intrinsic::arm_neon_vqshiftsu:
6912 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
6914 llvm_unreachable("invalid shift count for vqshlu intrinsic");
6916 case Intrinsic::arm_neon_vshiftn:
6917 case Intrinsic::arm_neon_vrshiftn:
6918 case Intrinsic::arm_neon_vqshiftns:
6919 case Intrinsic::arm_neon_vqshiftnu:
6920 case Intrinsic::arm_neon_vqshiftnsu:
6921 case Intrinsic::arm_neon_vqrshiftns:
6922 case Intrinsic::arm_neon_vqrshiftnu:
6923 case Intrinsic::arm_neon_vqrshiftnsu:
6924 // Narrowing shifts require an immediate right shift.
6925 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
6927 llvm_unreachable("invalid shift count for narrowing vector shift "
6931 llvm_unreachable("unhandled vector shift");
6935 case Intrinsic::arm_neon_vshifts:
6936 case Intrinsic::arm_neon_vshiftu:
6937 // Opcode already set above.
6939 case Intrinsic::arm_neon_vshiftls:
6940 case Intrinsic::arm_neon_vshiftlu:
6941 if (Cnt == VT.getVectorElementType().getSizeInBits())
6942 VShiftOpc = ARMISD::VSHLLi;
6944 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
6945 ARMISD::VSHLLs : ARMISD::VSHLLu);
6947 case Intrinsic::arm_neon_vshiftn:
6948 VShiftOpc = ARMISD::VSHRN; break;
6949 case Intrinsic::arm_neon_vrshifts:
6950 VShiftOpc = ARMISD::VRSHRs; break;
6951 case Intrinsic::arm_neon_vrshiftu:
6952 VShiftOpc = ARMISD::VRSHRu; break;
6953 case Intrinsic::arm_neon_vrshiftn:
6954 VShiftOpc = ARMISD::VRSHRN; break;
6955 case Intrinsic::arm_neon_vqshifts:
6956 VShiftOpc = ARMISD::VQSHLs; break;
6957 case Intrinsic::arm_neon_vqshiftu:
6958 VShiftOpc = ARMISD::VQSHLu; break;
6959 case Intrinsic::arm_neon_vqshiftsu:
6960 VShiftOpc = ARMISD::VQSHLsu; break;
6961 case Intrinsic::arm_neon_vqshiftns:
6962 VShiftOpc = ARMISD::VQSHRNs; break;
6963 case Intrinsic::arm_neon_vqshiftnu:
6964 VShiftOpc = ARMISD::VQSHRNu; break;
6965 case Intrinsic::arm_neon_vqshiftnsu:
6966 VShiftOpc = ARMISD::VQSHRNsu; break;
6967 case Intrinsic::arm_neon_vqrshiftns:
6968 VShiftOpc = ARMISD::VQRSHRNs; break;
6969 case Intrinsic::arm_neon_vqrshiftnu:
6970 VShiftOpc = ARMISD::VQRSHRNu; break;
6971 case Intrinsic::arm_neon_vqrshiftnsu:
6972 VShiftOpc = ARMISD::VQRSHRNsu; break;
6975 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6976 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
6979 case Intrinsic::arm_neon_vshiftins: {
6980 EVT VT = N->getOperand(1).getValueType();
6982 unsigned VShiftOpc = 0;
6984 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
6985 VShiftOpc = ARMISD::VSLI;
6986 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
6987 VShiftOpc = ARMISD::VSRI;
6989 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
6992 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
6993 N->getOperand(1), N->getOperand(2),
6994 DAG.getConstant(Cnt, MVT::i32));
6997 case Intrinsic::arm_neon_vqrshifts:
6998 case Intrinsic::arm_neon_vqrshiftu:
6999 // No immediate versions of these to check for.
7006 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
7007 /// lowers them. As with the vector shift intrinsics, this is done during DAG
7008 /// combining instead of DAG legalizing because the build_vectors for 64-bit
7009 /// vector element shift counts are generally not legal, and it is hard to see
7010 /// their values after they get legalized to loads from a constant pool.
7011 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7012 const ARMSubtarget *ST) {
7013 EVT VT = N->getValueType(0);
7015 // Nothing to be done for scalar shifts.
7016 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7017 if (!VT.isVector() || !TLI.isTypeLegal(VT))
7020 assert(ST->hasNEON() && "unexpected vector shift");
7023 switch (N->getOpcode()) {
7024 default: llvm_unreachable("unexpected shift opcode");
7027 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7028 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7029 DAG.getConstant(Cnt, MVT::i32));
7034 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7035 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7036 ARMISD::VSHRs : ARMISD::VSHRu);
7037 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7038 DAG.getConstant(Cnt, MVT::i32));
7044 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7045 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7046 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7047 const ARMSubtarget *ST) {
7048 SDValue N0 = N->getOperand(0);
7050 // Check for sign- and zero-extensions of vector extract operations of 8-
7051 // and 16-bit vector elements. NEON supports these directly. They are
7052 // handled during DAG combining because type legalization will promote them
7053 // to 32-bit types and it is messy to recognize the operations after that.
7054 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7055 SDValue Vec = N0.getOperand(0);
7056 SDValue Lane = N0.getOperand(1);
7057 EVT VT = N->getValueType(0);
7058 EVT EltVT = N0.getValueType();
7059 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7061 if (VT == MVT::i32 &&
7062 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7063 TLI.isTypeLegal(Vec.getValueType()) &&
7064 isa<ConstantSDNode>(Lane)) {
7067 switch (N->getOpcode()) {
7068 default: llvm_unreachable("unexpected opcode");
7069 case ISD::SIGN_EXTEND:
7070 Opc = ARMISD::VGETLANEs;
7072 case ISD::ZERO_EXTEND:
7073 case ISD::ANY_EXTEND:
7074 Opc = ARMISD::VGETLANEu;
7077 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7084 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7085 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7086 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7087 const ARMSubtarget *ST) {
7088 // If the target supports NEON, try to use vmax/vmin instructions for f32
7089 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
7090 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
7091 // a NaN; only do the transformation when it matches that behavior.
7093 // For now only do this when using NEON for FP operations; if using VFP, it
7094 // is not obvious that the benefit outweighs the cost of switching to the
7096 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7097 N->getValueType(0) != MVT::f32)
7100 SDValue CondLHS = N->getOperand(0);
7101 SDValue CondRHS = N->getOperand(1);
7102 SDValue LHS = N->getOperand(2);
7103 SDValue RHS = N->getOperand(3);
7104 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7106 unsigned Opcode = 0;
7108 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7109 IsReversed = false; // x CC y ? x : y
7110 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7111 IsReversed = true ; // x CC y ? y : x
7125 // If LHS is NaN, an ordered comparison will be false and the result will
7126 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
7127 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7128 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
7129 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7131 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
7132 // will return -0, so vmin can only be used for unsafe math or if one of
7133 // the operands is known to be nonzero.
7134 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
7136 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7138 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7147 // If LHS is NaN, an ordered comparison will be false and the result will
7148 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
7149 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7150 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7151 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7153 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
7154 // will return +0, so vmax can only be used for unsafe math or if one of
7155 // the operands is known to be nonzero.
7156 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
7158 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7160 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
7166 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
7169 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
7171 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
7172 SDValue Cmp = N->getOperand(4);
7173 if (Cmp.getOpcode() != ARMISD::CMPZ)
7174 // Only looking at EQ and NE cases.
7177 EVT VT = N->getValueType(0);
7178 DebugLoc dl = N->getDebugLoc();
7179 SDValue LHS = Cmp.getOperand(0);
7180 SDValue RHS = Cmp.getOperand(1);
7181 SDValue FalseVal = N->getOperand(0);
7182 SDValue TrueVal = N->getOperand(1);
7183 SDValue ARMcc = N->getOperand(2);
7184 ARMCC::CondCodes CC = (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
7202 /// FIXME: Turn this into a target neutral optimization?
7204 if (CC == ARMCC::NE && FalseVal == RHS) {
7205 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
7206 N->getOperand(3), Cmp);
7207 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
7209 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
7210 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
7211 N->getOperand(3), NewCmp);
7214 if (Res.getNode()) {
7215 APInt KnownZero, KnownOne;
7216 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
7217 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
7218 // Capture demanded bits information that would be otherwise lost.
7219 if (KnownZero == 0xfffffffe)
7220 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7221 DAG.getValueType(MVT::i1));
7222 else if (KnownZero == 0xffffff00)
7223 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7224 DAG.getValueType(MVT::i8));
7225 else if (KnownZero == 0xffff0000)
7226 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
7227 DAG.getValueType(MVT::i16));
7233 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
7234 DAGCombinerInfo &DCI) const {
7235 switch (N->getOpcode()) {
7237 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
7238 case ISD::SUB: return PerformSUBCombine(N, DCI);
7239 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
7240 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
7241 case ISD::AND: return PerformANDCombine(N, DCI);
7242 case ARMISD::BFI: return PerformBFICombine(N, DCI);
7243 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
7244 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
7245 case ISD::STORE: return PerformSTORECombine(N, DCI);
7246 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
7247 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
7248 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
7249 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
7250 case ISD::FP_TO_SINT:
7251 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
7252 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
7253 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
7256 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
7257 case ISD::SIGN_EXTEND:
7258 case ISD::ZERO_EXTEND:
7259 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
7260 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
7261 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
7262 case ARMISD::VLD2DUP:
7263 case ARMISD::VLD3DUP:
7264 case ARMISD::VLD4DUP:
7265 return CombineBaseUpdate(N, DCI);
7266 case ISD::INTRINSIC_VOID:
7267 case ISD::INTRINSIC_W_CHAIN:
7268 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
7269 case Intrinsic::arm_neon_vld1:
7270 case Intrinsic::arm_neon_vld2:
7271 case Intrinsic::arm_neon_vld3:
7272 case Intrinsic::arm_neon_vld4:
7273 case Intrinsic::arm_neon_vld2lane:
7274 case Intrinsic::arm_neon_vld3lane:
7275 case Intrinsic::arm_neon_vld4lane:
7276 case Intrinsic::arm_neon_vst1:
7277 case Intrinsic::arm_neon_vst2:
7278 case Intrinsic::arm_neon_vst3:
7279 case Intrinsic::arm_neon_vst4:
7280 case Intrinsic::arm_neon_vst2lane:
7281 case Intrinsic::arm_neon_vst3lane:
7282 case Intrinsic::arm_neon_vst4lane:
7283 return CombineBaseUpdate(N, DCI);
7291 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
7293 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
7296 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
7297 if (!Subtarget->allowsUnalignedMem())
7300 switch (VT.getSimpleVT().SimpleTy) {
7307 // FIXME: VLD1 etc with standard alignment is legal.
7311 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
7316 switch (VT.getSimpleVT().SimpleTy) {
7317 default: return false;
7332 if ((V & (Scale - 1)) != 0)
7335 return V == (V & ((1LL << 5) - 1));
7338 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
7339 const ARMSubtarget *Subtarget) {
7346 switch (VT.getSimpleVT().SimpleTy) {
7347 default: return false;
7352 // + imm12 or - imm8
7354 return V == (V & ((1LL << 8) - 1));
7355 return V == (V & ((1LL << 12) - 1));
7358 // Same as ARM mode. FIXME: NEON?
7359 if (!Subtarget->hasVFP2())
7364 return V == (V & ((1LL << 8) - 1));
7368 /// isLegalAddressImmediate - Return true if the integer value can be used
7369 /// as the offset of the target addressing mode for load / store of the
7371 static bool isLegalAddressImmediate(int64_t V, EVT VT,
7372 const ARMSubtarget *Subtarget) {
7379 if (Subtarget->isThumb1Only())
7380 return isLegalT1AddressImmediate(V, VT);
7381 else if (Subtarget->isThumb2())
7382 return isLegalT2AddressImmediate(V, VT, Subtarget);
7387 switch (VT.getSimpleVT().SimpleTy) {
7388 default: return false;
7393 return V == (V & ((1LL << 12) - 1));
7396 return V == (V & ((1LL << 8) - 1));
7399 if (!Subtarget->hasVFP2()) // FIXME: NEON?
7404 return V == (V & ((1LL << 8) - 1));
7408 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
7410 int Scale = AM.Scale;
7414 switch (VT.getSimpleVT().SimpleTy) {
7415 default: return false;
7424 return Scale == 2 || Scale == 4 || Scale == 8;
7427 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
7431 // Note, we allow "void" uses (basically, uses that aren't loads or
7432 // stores), because arm allows folding a scale into many arithmetic
7433 // operations. This should be made more precise and revisited later.
7435 // Allow r << imm, but the imm has to be a multiple of two.
7436 if (Scale & 1) return false;
7437 return isPowerOf2_32(Scale);
7441 /// isLegalAddressingMode - Return true if the addressing mode represented
7442 /// by AM is legal for this target, for a load/store of the specified type.
7443 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
7445 EVT VT = getValueType(Ty, true);
7446 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
7449 // Can never fold addr of global into load/store.
7454 case 0: // no scale reg, must be "r+i" or "r", or "i".
7457 if (Subtarget->isThumb1Only())
7461 // ARM doesn't support any R+R*scale+imm addr modes.
7468 if (Subtarget->isThumb2())
7469 return isLegalT2ScaledAddressingMode(AM, VT);
7471 int Scale = AM.Scale;
7472 switch (VT.getSimpleVT().SimpleTy) {
7473 default: return false;
7477 if (Scale < 0) Scale = -Scale;
7481 return isPowerOf2_32(Scale & ~1);
7485 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
7490 // Note, we allow "void" uses (basically, uses that aren't loads or
7491 // stores), because arm allows folding a scale into many arithmetic
7492 // operations. This should be made more precise and revisited later.
7494 // Allow r << imm, but the imm has to be a multiple of two.
7495 if (Scale & 1) return false;
7496 return isPowerOf2_32(Scale);
7503 /// isLegalICmpImmediate - Return true if the specified immediate is legal
7504 /// icmp immediate, that is the target has icmp instructions which can compare
7505 /// a register against the immediate without having to materialize the
7506 /// immediate into a register.
7507 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
7508 if (!Subtarget->isThumb())
7509 return ARM_AM::getSOImmVal(Imm) != -1;
7510 if (Subtarget->isThumb2())
7511 return ARM_AM::getT2SOImmVal(Imm) != -1;
7512 return Imm >= 0 && Imm <= 255;
7515 /// isLegalAddImmediate - Return true if the specified immediate is legal
7516 /// add immediate, that is the target has add instructions which can add
7517 /// a register with the immediate without having to materialize the
7518 /// immediate into a register.
7519 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
7520 return ARM_AM::getSOImmVal(Imm) != -1;
7523 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
7524 bool isSEXTLoad, SDValue &Base,
7525 SDValue &Offset, bool &isInc,
7526 SelectionDAG &DAG) {
7527 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7530 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
7532 Base = Ptr->getOperand(0);
7533 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7534 int RHSC = (int)RHS->getZExtValue();
7535 if (RHSC < 0 && RHSC > -256) {
7536 assert(Ptr->getOpcode() == ISD::ADD);
7538 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7542 isInc = (Ptr->getOpcode() == ISD::ADD);
7543 Offset = Ptr->getOperand(1);
7545 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
7547 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7548 int RHSC = (int)RHS->getZExtValue();
7549 if (RHSC < 0 && RHSC > -0x1000) {
7550 assert(Ptr->getOpcode() == ISD::ADD);
7552 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7553 Base = Ptr->getOperand(0);
7558 if (Ptr->getOpcode() == ISD::ADD) {
7560 ARM_AM::ShiftOpc ShOpcVal=
7561 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
7562 if (ShOpcVal != ARM_AM::no_shift) {
7563 Base = Ptr->getOperand(1);
7564 Offset = Ptr->getOperand(0);
7566 Base = Ptr->getOperand(0);
7567 Offset = Ptr->getOperand(1);
7572 isInc = (Ptr->getOpcode() == ISD::ADD);
7573 Base = Ptr->getOperand(0);
7574 Offset = Ptr->getOperand(1);
7578 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
7582 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
7583 bool isSEXTLoad, SDValue &Base,
7584 SDValue &Offset, bool &isInc,
7585 SelectionDAG &DAG) {
7586 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
7589 Base = Ptr->getOperand(0);
7590 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
7591 int RHSC = (int)RHS->getZExtValue();
7592 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
7593 assert(Ptr->getOpcode() == ISD::ADD);
7595 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
7597 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
7598 isInc = Ptr->getOpcode() == ISD::ADD;
7599 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
7607 /// getPreIndexedAddressParts - returns true by value, base pointer and
7608 /// offset pointer and addressing mode by reference if the node's address
7609 /// can be legally represented as pre-indexed load / store address.
7611 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7613 ISD::MemIndexedMode &AM,
7614 SelectionDAG &DAG) const {
7615 if (Subtarget->isThumb1Only())
7620 bool isSEXTLoad = false;
7621 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7622 Ptr = LD->getBasePtr();
7623 VT = LD->getMemoryVT();
7624 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7625 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7626 Ptr = ST->getBasePtr();
7627 VT = ST->getMemoryVT();
7632 bool isLegal = false;
7633 if (Subtarget->isThumb2())
7634 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7635 Offset, isInc, DAG);
7637 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
7638 Offset, isInc, DAG);
7642 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
7646 /// getPostIndexedAddressParts - returns true by value, base pointer and
7647 /// offset pointer and addressing mode by reference if this node can be
7648 /// combined with a load / store to form a post-indexed load / store.
7649 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7652 ISD::MemIndexedMode &AM,
7653 SelectionDAG &DAG) const {
7654 if (Subtarget->isThumb1Only())
7659 bool isSEXTLoad = false;
7660 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7661 VT = LD->getMemoryVT();
7662 Ptr = LD->getBasePtr();
7663 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
7664 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7665 VT = ST->getMemoryVT();
7666 Ptr = ST->getBasePtr();
7671 bool isLegal = false;
7672 if (Subtarget->isThumb2())
7673 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7676 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
7682 // Swap base ptr and offset to catch more post-index load / store when
7683 // it's legal. In Thumb2 mode, offset must be an immediate.
7684 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
7685 !Subtarget->isThumb2())
7686 std::swap(Base, Offset);
7688 // Post-indexed load / store update the base pointer.
7693 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
7697 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7701 const SelectionDAG &DAG,
7702 unsigned Depth) const {
7703 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
7704 switch (Op.getOpcode()) {
7706 case ARMISD::CMOV: {
7707 // Bits are known zero/one if known on the LHS and RHS.
7708 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
7709 if (KnownZero == 0 && KnownOne == 0) return;
7711 APInt KnownZeroRHS, KnownOneRHS;
7712 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
7713 KnownZeroRHS, KnownOneRHS, Depth+1);
7714 KnownZero &= KnownZeroRHS;
7715 KnownOne &= KnownOneRHS;
7721 //===----------------------------------------------------------------------===//
7722 // ARM Inline Assembly Support
7723 //===----------------------------------------------------------------------===//
7725 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
7726 // Looking for "rev" which is V6+.
7727 if (!Subtarget->hasV6Ops())
7730 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
7731 std::string AsmStr = IA->getAsmString();
7732 SmallVector<StringRef, 4> AsmPieces;
7733 SplitString(AsmStr, AsmPieces, ";\n");
7735 switch (AsmPieces.size()) {
7736 default: return false;
7738 AsmStr = AsmPieces[0];
7740 SplitString(AsmStr, AsmPieces, " \t,");
7743 if (AsmPieces.size() == 3 &&
7744 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
7745 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
7746 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
7747 if (Ty && Ty->getBitWidth() == 32)
7748 return IntrinsicLowering::LowerToByteSwap(CI);
7756 /// getConstraintType - Given a constraint letter, return the type of
7757 /// constraint it is for this target.
7758 ARMTargetLowering::ConstraintType
7759 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
7760 if (Constraint.size() == 1) {
7761 switch (Constraint[0]) {
7763 case 'l': return C_RegisterClass;
7764 case 'w': return C_RegisterClass;
7765 case 'h': return C_RegisterClass;
7766 case 'x': return C_RegisterClass;
7767 case 't': return C_RegisterClass;
7768 case 'j': return C_Other; // Constant for movw.
7769 // An address with a single base register. Due to the way we
7770 // currently handle addresses it is the same as an 'r' memory constraint.
7771 case 'Q': return C_Memory;
7773 } else if (Constraint.size() == 2) {
7774 switch (Constraint[0]) {
7776 // All 'U+' constraints are addresses.
7777 case 'U': return C_Memory;
7780 return TargetLowering::getConstraintType(Constraint);
7783 /// Examine constraint type and operand type and determine a weight value.
7784 /// This object must already have been set up with the operand type
7785 /// and the current alternative constraint selected.
7786 TargetLowering::ConstraintWeight
7787 ARMTargetLowering::getSingleConstraintMatchWeight(
7788 AsmOperandInfo &info, const char *constraint) const {
7789 ConstraintWeight weight = CW_Invalid;
7790 Value *CallOperandVal = info.CallOperandVal;
7791 // If we don't have a value, we can't do a match,
7792 // but allow it at the lowest weight.
7793 if (CallOperandVal == NULL)
7795 Type *type = CallOperandVal->getType();
7796 // Look at the constraint type.
7797 switch (*constraint) {
7799 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
7802 if (type->isIntegerTy()) {
7803 if (Subtarget->isThumb())
7804 weight = CW_SpecificReg;
7806 weight = CW_Register;
7810 if (type->isFloatingPointTy())
7811 weight = CW_Register;
7817 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
7819 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
7821 if (Constraint.size() == 1) {
7822 // GCC ARM Constraint Letters
7823 switch (Constraint[0]) {
7824 case 'l': // Low regs or general regs.
7825 if (Subtarget->isThumb())
7826 return RCPair(0U, ARM::tGPRRegisterClass);
7828 return RCPair(0U, ARM::GPRRegisterClass);
7829 case 'h': // High regs or no regs.
7830 if (Subtarget->isThumb())
7831 return RCPair(0U, ARM::hGPRRegisterClass);
7834 return RCPair(0U, ARM::GPRRegisterClass);
7837 return RCPair(0U, ARM::SPRRegisterClass);
7838 if (VT.getSizeInBits() == 64)
7839 return RCPair(0U, ARM::DPRRegisterClass);
7840 if (VT.getSizeInBits() == 128)
7841 return RCPair(0U, ARM::QPRRegisterClass);
7845 return RCPair(0U, ARM::SPR_8RegisterClass);
7846 if (VT.getSizeInBits() == 64)
7847 return RCPair(0U, ARM::DPR_8RegisterClass);
7848 if (VT.getSizeInBits() == 128)
7849 return RCPair(0U, ARM::QPR_8RegisterClass);
7853 return RCPair(0U, ARM::SPRRegisterClass);
7857 if (StringRef("{cc}").equals_lower(Constraint))
7858 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
7860 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
7863 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
7864 /// vector. If it is invalid, don't add anything to Ops.
7865 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
7866 std::string &Constraint,
7867 std::vector<SDValue>&Ops,
7868 SelectionDAG &DAG) const {
7869 SDValue Result(0, 0);
7871 // Currently only support length 1 constraints.
7872 if (Constraint.length() != 1) return;
7874 char ConstraintLetter = Constraint[0];
7875 switch (ConstraintLetter) {
7878 case 'I': case 'J': case 'K': case 'L':
7879 case 'M': case 'N': case 'O':
7880 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
7884 int64_t CVal64 = C->getSExtValue();
7885 int CVal = (int) CVal64;
7886 // None of these constraints allow values larger than 32 bits. Check
7887 // that the value fits in an int.
7891 switch (ConstraintLetter) {
7893 // Constant suitable for movw, must be between 0 and
7895 if (Subtarget->hasV6T2Ops())
7896 if (CVal >= 0 && CVal <= 65535)
7900 if (Subtarget->isThumb1Only()) {
7901 // This must be a constant between 0 and 255, for ADD
7903 if (CVal >= 0 && CVal <= 255)
7905 } else if (Subtarget->isThumb2()) {
7906 // A constant that can be used as an immediate value in a
7907 // data-processing instruction.
7908 if (ARM_AM::getT2SOImmVal(CVal) != -1)
7911 // A constant that can be used as an immediate value in a
7912 // data-processing instruction.
7913 if (ARM_AM::getSOImmVal(CVal) != -1)
7919 if (Subtarget->isThumb()) { // FIXME thumb2
7920 // This must be a constant between -255 and -1, for negated ADD
7921 // immediates. This can be used in GCC with an "n" modifier that
7922 // prints the negated value, for use with SUB instructions. It is
7923 // not useful otherwise but is implemented for compatibility.
7924 if (CVal >= -255 && CVal <= -1)
7927 // This must be a constant between -4095 and 4095. It is not clear
7928 // what this constraint is intended for. Implemented for
7929 // compatibility with GCC.
7930 if (CVal >= -4095 && CVal <= 4095)
7936 if (Subtarget->isThumb1Only()) {
7937 // A 32-bit value where only one byte has a nonzero value. Exclude
7938 // zero to match GCC. This constraint is used by GCC internally for
7939 // constants that can be loaded with a move/shift combination.
7940 // It is not useful otherwise but is implemented for compatibility.
7941 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
7943 } else if (Subtarget->isThumb2()) {
7944 // A constant whose bitwise inverse can be used as an immediate
7945 // value in a data-processing instruction. This can be used in GCC
7946 // with a "B" modifier that prints the inverted value, for use with
7947 // BIC and MVN instructions. It is not useful otherwise but is
7948 // implemented for compatibility.
7949 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
7952 // A constant whose bitwise inverse can be used as an immediate
7953 // value in a data-processing instruction. This can be used in GCC
7954 // with a "B" modifier that prints the inverted value, for use with
7955 // BIC and MVN instructions. It is not useful otherwise but is
7956 // implemented for compatibility.
7957 if (ARM_AM::getSOImmVal(~CVal) != -1)
7963 if (Subtarget->isThumb1Only()) {
7964 // This must be a constant between -7 and 7,
7965 // for 3-operand ADD/SUB immediate instructions.
7966 if (CVal >= -7 && CVal < 7)
7968 } else if (Subtarget->isThumb2()) {
7969 // A constant whose negation can be used as an immediate value in a
7970 // data-processing instruction. This can be used in GCC with an "n"
7971 // modifier that prints the negated value, for use with SUB
7972 // instructions. It is not useful otherwise but is implemented for
7974 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
7977 // A constant whose negation can be used as an immediate value in a
7978 // data-processing instruction. This can be used in GCC with an "n"
7979 // modifier that prints the negated value, for use with SUB
7980 // instructions. It is not useful otherwise but is implemented for
7982 if (ARM_AM::getSOImmVal(-CVal) != -1)
7988 if (Subtarget->isThumb()) { // FIXME thumb2
7989 // This must be a multiple of 4 between 0 and 1020, for
7990 // ADD sp + immediate.
7991 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
7994 // A power of two or a constant between 0 and 32. This is used in
7995 // GCC for the shift amount on shifted register operands, but it is
7996 // useful in general for any shift amounts.
7997 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8003 if (Subtarget->isThumb()) { // FIXME thumb2
8004 // This must be a constant between 0 and 31, for shift amounts.
8005 if (CVal >= 0 && CVal <= 31)
8011 if (Subtarget->isThumb()) { // FIXME thumb2
8012 // This must be a multiple of 4 between -508 and 508, for
8013 // ADD/SUB sp = sp + immediate.
8014 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8019 Result = DAG.getTargetConstant(CVal, Op.getValueType());
8023 if (Result.getNode()) {
8024 Ops.push_back(Result);
8027 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8031 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8032 // The ARM target isn't yet aware of offsets.
8036 int ARM::getVFPf32Imm(const APFloat &FPImm) {
8037 APInt Imm = FPImm.bitcastToAPInt();
8038 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
8039 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
8040 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
8042 // We can handle 4 bits of mantissa.
8043 // mantissa = (16+UInt(e:f:g:h))/16.
8044 if (Mantissa & 0x7ffff)
8047 if ((Mantissa & 0xf) != Mantissa)
8050 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
8051 if (Exp < -3 || Exp > 4)
8053 Exp = ((Exp+3) & 0x7) ^ 4;
8055 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
8058 int ARM::getVFPf64Imm(const APFloat &FPImm) {
8059 APInt Imm = FPImm.bitcastToAPInt();
8060 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
8061 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
8062 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
8064 // We can handle 4 bits of mantissa.
8065 // mantissa = (16+UInt(e:f:g:h))/16.
8066 if (Mantissa & 0xffffffffffffLL)
8069 if ((Mantissa & 0xf) != Mantissa)
8072 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
8073 if (Exp < -3 || Exp > 4)
8075 Exp = ((Exp+3) & 0x7) ^ 4;
8077 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
8080 bool ARM::isBitFieldInvertedMask(unsigned v) {
8081 if (v == 0xffffffff)
8083 // there can be 1's on either or both "outsides", all the "inside"
8085 unsigned int lsb = 0, msb = 31;
8086 while (v & (1 << msb)) --msb;
8087 while (v & (1 << lsb)) ++lsb;
8088 for (unsigned int i = lsb; i <= msb; ++i) {
8095 /// isFPImmLegal - Returns true if the target can instruction select the
8096 /// specified FP immediate natively. If false, the legalizer will
8097 /// materialize the FP immediate as a load from a constant pool.
8098 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8099 if (!Subtarget->hasVFP3())
8102 return ARM::getVFPf32Imm(Imm) != -1;
8104 return ARM::getVFPf64Imm(Imm) != -1;
8108 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8109 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
8110 /// specified in the intrinsic calls.
8111 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8113 unsigned Intrinsic) const {
8114 switch (Intrinsic) {
8115 case Intrinsic::arm_neon_vld1:
8116 case Intrinsic::arm_neon_vld2:
8117 case Intrinsic::arm_neon_vld3:
8118 case Intrinsic::arm_neon_vld4:
8119 case Intrinsic::arm_neon_vld2lane:
8120 case Intrinsic::arm_neon_vld3lane:
8121 case Intrinsic::arm_neon_vld4lane: {
8122 Info.opc = ISD::INTRINSIC_W_CHAIN;
8123 // Conservatively set memVT to the entire set of vectors loaded.
8124 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8125 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8126 Info.ptrVal = I.getArgOperand(0);
8128 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8129 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8130 Info.vol = false; // volatile loads with NEON intrinsics not supported
8131 Info.readMem = true;
8132 Info.writeMem = false;
8135 case Intrinsic::arm_neon_vst1:
8136 case Intrinsic::arm_neon_vst2:
8137 case Intrinsic::arm_neon_vst3:
8138 case Intrinsic::arm_neon_vst4:
8139 case Intrinsic::arm_neon_vst2lane:
8140 case Intrinsic::arm_neon_vst3lane:
8141 case Intrinsic::arm_neon_vst4lane: {
8142 Info.opc = ISD::INTRINSIC_VOID;
8143 // Conservatively set memVT to the entire set of vectors stored.
8144 unsigned NumElts = 0;
8145 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8146 Type *ArgTy = I.getArgOperand(ArgI)->getType();
8147 if (!ArgTy->isVectorTy())
8149 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8151 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8152 Info.ptrVal = I.getArgOperand(0);
8154 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8155 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8156 Info.vol = false; // volatile stores with NEON intrinsics not supported
8157 Info.readMem = false;
8158 Info.writeMem = true;
8161 case Intrinsic::arm_strexd: {
8162 Info.opc = ISD::INTRINSIC_W_CHAIN;
8163 Info.memVT = MVT::i64;
8164 Info.ptrVal = I.getArgOperand(2);
8168 Info.readMem = false;
8169 Info.writeMem = true;
8172 case Intrinsic::arm_ldrexd: {
8173 Info.opc = ISD::INTRINSIC_W_CHAIN;
8174 Info.memVT = MVT::i64;
8175 Info.ptrVal = I.getArgOperand(0);
8179 Info.readMem = true;
8180 Info.writeMem = false;